Sampling Techniques

hird edition

WILLIAM G. COCHRAN I

Professor of Statistics, Emeritl4s

Harvard University

@J
WILEY EASTERN LIMITED
New Delhi Bangaloro Bom ;y
"Tlrird U .S. Edition, 1m I

Tbird Wiley Eastern Reprint February~

. Fourth WHey Eastern Repring, October, 1985

Authorized reprint of the editIOn pub1iabed by John Wiley &: Sons" Inc.,
New York, Chichester, Brisbane and Toroato
Copyright C 1m, John Wiley &: Sons, Inc.
All rights reserved. No part of this book may be reproduced in any form
without tbe written permission of Wiley-lnterscience, Inc.

This book is not to be sold outside the country

to which it is consigned by Wiley Eastern Umited

Sales Area: India

10 9 8 7 6 5 4

ISBN 0 85226 121 7

Published by Mohinder Singh Sejwal for Wiley Eastern Lunited,

4835/24 Ansari Road, Daryapnj, New Delhi 110002 and printed by
Prlimodh P. Kapur at Raj'Bandhu lnd.-Co., C-61 Mayapuri Phuc-D.
New Delhi·ll0064. Printed inlDitia.
Preface

As did the previous editions, this text\-.~ - ' mts a comprehensive account of
,sampling theory as it has been devr _ in sample surveys. It contains

lillustrations to show how the t f 1 apptred in practice, and exercises to be

worked by the student. The bo . NiH be useful both as a text for a course on
sample surveys in which the maj . r emphasis is on theory and for individual
reading by the student.
The minimum mathematical equipment necessary to follow the great bulk of
the material is a familiarity with algebra, especially relatively complicated algeb-
raic expressions, plus a knowledge of probability for finite sample spacef includ-
ing combinatorial probabilities. The book presupposes an introductory statistics
course that covers means and standard deviations, the normal , binomial,
hypergeometric, and multinomial distributions, the central limit theorem, linear
regression, and the simpler types of analyses of variance . Since much of classical
sample survey theory deals with the distributions of estimators over the set of
randomizations provided by the sampling plan, some knowledge of nonparamet-
ric methods is helpful.
The topics in this edition are presented in essentially the same order as in earlier
editions. New sections have been included, or sections rewritten, primarily for one
of three reasons : (1) to present introductions to topics (sampling plans or methods
of estimation) relatively new in the field ; (2) to cover further work done during the
last 15 years on older methods, intended either to improve them or to learn more
about the performance of rival methods; and (3) to shorten, clarify, or simplify
proofs given in previous editions.
New topics in this edition include the approximate methods developed for the
difficult problem of attaching standard errors or confidence limits to nonlinear
estimates made from the results of surveys with complex plans. These methods
will be more and more needed as statistical analyses (e.g. , regressions) are
performed on the results. For surveys containing sensitive questions that some
respondents are unlikely to be willing to answer truthfully, a new device is to
present the respondent with eitber the sensitive question or an innocuous ques-
tion; the specific choice, made by randomization, is unknown to the interviewer.
In some sampling problems it may seem economically attractive, or essential in
countries without full sampling resources, to use two overlapping lists (or frames,
as they are called) to cover the complete population. The method of double
sampling has been extended to cases where the- objective is to compare the means
vii
viii PREFACE

of a number of subgroups within the population. There has been interesting

work on the attractive properties that the ratio and regression estimators have if it
can be assumed that the finite population is itself a random sample from an infinite
superpopulation in which a mathematical model appropriate to the ratio or
regression estimator holds. This kind of assumption is not new-I noticed
recently that Laplace used it around 1800 in a samplin~ problem-but it clarifies
the relation between sample survey theory and standard statistical theory.
An example of further work on topics included in previous editions is Chapter
9A, which has been written partly from material previously in Chapter 9; this was
done mainly to give a more adequate account of what seem to me the principal
methods produced for sampling with unequal probabilities without replacement.
These include the similar methods given independently by Brewer, J. N. K. Rao,
and Durbin, Murthy'S method, the Rao, Hartley, Cochran method, and Madow's
method related to systematic sampling, with comparisons of the performances of
the methods on natural populations. New studies have been done of the sizes of
components of errors of measurement in surveys by repeat measurements by
different interviewers, by interpenetrating subsamples, and by a combination of
the two approaches . For the ratio estimator, data from natural populations have
been used to appraise the small-sample biases in the standard large-sample
formulas for the variance and the estimated variance. Attempts have also been
made to create less biased variants of the ratio estimator itself and of the formula
for estimating its sampling variance. In stratified sampling there has been addi-
tional work on allocating sample sizes to strata when more than one item is of
importance and on estimating sample errors when only one unit is to be selected
per stratum. Some new systematic sampling methods for handling populations
having linear trends are also of interest.
Alva L. Finkner and Emil H. Jebe prepared a large part of the lecture notes
frol11 which the first edition of this book was written. Some investigations that
provided background material were supported by the Office of Naval Research,
Navy Department. From discussions of recent developments in sampling or
suggestions about this edition, I have been greatly helped by Tore DaJenius,
David J . Finney, Daniel G . Horvitz, Leslie Kish, P. S. R. Sambasiva Rao, Martin
Sandelius, Joseph Sedransk, Amode R. Sen, and especially Jon N. K. Rao, whose
painstaking reading of the new and revised sections of this edition resulted in
many constructive suggestions about gaps, weaknesses, obscurities, and selection
of topics. For typing and other work involved in production of a typescript I am
indebted to Rowena Foss, Holly Grano, and Edith Klotz. My thanks to all.

William G. Cochran
South Orleans, Massachusetts
February, 1977
Contents
CHAPTER PAGE

1 INTRODUCTION
1.1 Advantages of the Sampling Method )
1.2 Some Uses of Sample Surveys . . . 2
1.3 The Principal Steps in a Sample Survey 4
1.4 The Role of Sampling Theory . . . 8
1.5 Probability Sampling .. . . . . 9
1.6 Alternatives to Probability Sampling J(}
1.7 Use of the Normal Distribution 11
1.8 Bias and lts Effects 12
1.9 The Mean Square Error IS
EUTcises 16

CHAPTER
2 SIMPLE RANDOM SAMPLING 18

2.1 Simple Random Sampling . . . . . .. 18

2.2 Selection of a Simple Random Sample - 19
2.3 Definitions and Notation . 20
2.4 Properties ofthe Estima~s . . . 21
2.5 Variances of the Estimates 23
2.6 The Finite Population Correction 24
2.7 Estimation of the Standard Error from a Sample 2S
2.8 Confidence Limits .. ... . . 27
2.9 An Alternative Method of Proof . . 28
2.10 Random Sampling with Replacement 29
2.11 Estimation of a Ratio . .. . . . 30
2.12 Estimates of Means Over Subpopulations 34
2.13 Estimates of Totals Over Subpopulations 35
2.14 Comparisons Between Domain Means . 39
2.15 Validity of thr.Normal Approximiltion . 39
?-.16 Linear Estimators of the Population Mean 44
Exercises 45
ill
x CONTENTS

CHAPTER
3 SAMPLING PROPORTIONS
AND PERCENTAGES 50

3.I Qualitative Characteristics . . . . . 50

3.2 Variances of the Sample Estimates . . 50
) .3 The Effect of Pon the Standard Errors 53
3.4 The Binomial Distribution 55
3.5 The Hypergeometric Distribution . . 55
3.6 Confidence Limits .. . ,.... 57
3.7 Classification into More than Two Classes 60
3.8 Confidence Limits with More than Two Classes 60
3.9 The Conditional Distribution of p . . . , . 61
'3.10 Proportions and T-otals Over Subpo pulations 63
3. 11 Comparisons Between Different Domains 64
3. 12 Estimation of Proportions in Cluster Sampling M
Exercises 68

CHAPTER
4 THE ESTIMATION OF SAMPLE SIZE 72

4.1 A Hypothetical Example . . 72

4.2 Analysi~ of the Prohlem 73
4 .3 Thc Specification of Preci~ion 74
4 .-1 Th e Formula for n in Sampling for Proportions 75
4 .5 Rare Items- Inverse Sampling . . . . . 76
-I .n The Formula (or n with Continuous Data 77
-1 .7 Advance Estimate, of Population Variance, 78
-I.X Sample Sil.t: with Mmc than One (tem 81
4 .\1 Sample SI ZI! when EMimates Are Wanted for Subdivisions of the
Population . . . . . . . . . . H2
4 . 10 Sample Size in Deci~ion Problems 83
4 . II rhe Dc,ign Effect (De/!) 85
Exercises 116

CHAPTER
5 STRATIFIED RANDOM SAMPLING 89

5.1 Description """ 89

5.2 Notation ,.,.,. 90
5.3 Properties of the Estimates 91
5.4 The Estimated Variance and Confidence Limits . 95
5.5 Optimum Allocation . . •. . . . . . . . . 96
 CONTENTS xi
5.6 Relative Precision of Stratified Random and Simple Random
Sampling .. . . . .... .. . . . .. . .. 99
5.7 When Does Stratification Produce Large Gains in Precision? J0 1
5.8 Allocation Requiring More rhan 100 Per ent Sampling 104
5.9 Estimation of Sample Si:rewith Continuous Data . . . 105
5.10 Stratified Sampling for Proportions ........ 107
5.11 Gains in Precision in Stratified Sampling for Proportions 109
5. J 2" Estimation of Sample Size with Proporlions . . . . . 110
Exercises 1 11

CHAPTER
5A FURTHER ASPECTS OF
STRATIFIED SAMPLING 11 5
SA. I Effects of Deviations from the Optimum Allocation 115
5A.2 Effecrs of Errors in the Srratum Sizes . '.... 117
SA ,3 The Problem of Allocation with More than One Item 119
SA.4 Other Methods of Allocation with More than One Item 12 1
SA.S Two-Way Stratification with Small Samples 124
SA.6 Controlled Selection 126
5A.7 The Construction of Strata . . . . . . . 127
SA.S Number of Strata . . . . . . , . . . . 132
5A 9 Stratification After Selection of the Sample (Poststratification) 134
SAIO Quota Sampling . ' . . . . . . . . . . . . . . . . 13,5
SA .11 Estimation from a Sample of the Gain Due to Stratification 136
5A.12 Estimation of Variance with One Unit per Stratum 138
5A.13 Strata as Domains of Study , . . . . . . . . . 140
SA.14 Estimating Totals and Means Over Subpopulations . . . . . . , 142
5AlS SamplingfromTwoFrames . . . . . _ . . . . . . . . . . 144
Exercises 146

CHAPTER
6 RATIO ESTIMATORS 150

6,1 Methods of Estimation , . 150

6.2 The Ratio Estimate ISO
6.3 Approximate Variance of the Ratio Estimate 153
6.4 Estimation of the Variance from a Sample . . 155
6.5 Confidence Limits . . . . . , .. .. . 156
6.6 Comparison of the Ratio Estimllte with Mean per Unit 157
6.7 Conditions Under Which the Ratio Estimate Is 8 Best Linear
Unbia ed Estimator . . . IS8
6.8 aias of the Ratio Estimate . . . . . . 160
xii CONTENTS

6.9 Accuracy of the Formulas for the Variance and Estimated

Variance . . . . . . . . . 162
6.10 Ratio Estimates in Stratified Random Sampling 164
f).11 The Combined Ratio Estimate . . . . . . . 165
.12 Comparison of the Combined and Separate Estimates 167
6.13 Short-Cut Computation ofthe Estimated Variance 169
6.14 Optimum Allocation with a Ratio Estimate 172
6.15 Unbiased Ratio-type Estimates 174
6.16 Comparison of the Methods 177
6.17 Improved Estimation of Variance 178
6.18 Comparison of Two Ratios 180
6.19 Ratio of Two Ratios . . . . 183
6.20 Multivariate Ratio Estimates 184
6.21 Product Estimators 186
EXercises 186

CHAPTER
7 REGRESSION ESTIMATORS 189

7.1 The Linear Regression Estimate . . . . . . . . . . . . 1,89

7.2 Regression Estimates with Preassigned b . . . . . . . . 190
7.3 Regression Estimates when b Is Computed from the Sample 193
7.4 Sample Estimate of Variance . . . . . . . . . . . . . 195
7.5 Large-Sample Comparison with the Ratio Estimate and the Mean
per Unit . . . . . . . . . . . . . . . . . . . . . . 195
7.6 Accuracy of the Large-Sample Formulas for V{y/,) and v(y/,) . . . . 197
7.7 Bias of the Linear Regression Estimate . . . . . . . . . . . 198
7.8 The Linear Regression Estimator Under a Linear Regression
Model . . . . . . . . . . . . . . . . 199
7.9 Regression Estimates in Stratified Sampling 200
7.10 Regression Coefficients Estimated from the Sample . 201
7.1 I Comparison of the Two Types of Regre sian Estimate 203
Exercises 203

CHAPTER
8 SYSTEMATIC SAMPLING 205

8.1 Description . . . . . . . . 205

8.2 Relation to Cluster Sampling 207
8.3 Variance of the Estimated Mean 207
8.4 Comparison of Systematic with Stratified Random Sampling 212
8.5 Populations in " Random" Order . . . . . . . . . . . 212
 CONTENTS !iii
8.6 Populations with Linear Trend 214
8.7 Method~ for Populations with Linear Trends 216
8.8 Populations with Periodic Variation 217
8.9 Autocorrelated Populations . . . . . . 219
8.10 Natural Populations . . . . . . . . . . 221
8.1 J Estimation of the Variance from a Single Sample 223
8.12 StratifiedSystematicSampling 226
8.13 Systematic Sampling in Two Dimensions 227
8.14 Summary . . . . . . . . . . . . . 229
Exercises 231

CHAPTER
9 SINGLE-STAGE CLUSTER SAMPLING:
CLUSTERS OF EQUAL SIZES 233

9.1 ReasonsforClusterSampling . . . . . "233

9.2 A Simple Rule . . . . . . . . . . . . 234
9.3 Comparisons of Precision Made from Survey Data 238
9.4 Variance in Terms of Intrac!uster Correlation 240
9.5 Variance Functions . . . . . . 243
9.6 A Cost Function . . ..... 244
9."7 Cluster Sampling for Proportion ~ 246
Exercises 247

CHAPTER
9A SINGLE-STAGE CLUSTER SAMPLING:
CLUSTERS OF UNEQUAL SIZES 249

9AI Ouster Units of Unequal Sizes . . . . . . . . . . 249

9A.2 Sampling with Probability Proportional to Si7.e 250
9A3 Selection with Unequal Probabilities with Replacement 252
9A.4 The Optimum Measure of Size . . . . . . . . . . 255
9A.5 Relative Accuracies of Three Techniques . . ... 255
9 A.6 Sampling with Unequal Probabilities Without Replacement 258
9A7 The Horvitz-Thompson Estimator 259
9A.8 Brewer's Method . . . . . . . . . . 261
9A.9 Murthy'sMethod . . . . . . . . . . 263
9AI0 Methods Related to Systematic Sampling 265
9All The Rao, Hartley, Cochran Method 266
9A.12 NumedcalComparisons . . . 267
9Al3 Stratified and Ratio Estimates 270
Exercises 272
xiv CONTENTS

CHAPTER
10 SUBSAMPLlNG WITH UNITS OF EQUAL SIZE 274

10. 1 Two-Stage Sampling . . . . . . . . .. .... 274

10.2 Finding Means and Variances in Tw<'- tage Sampling . 275
J 0.3 Variance of the Estimated Mean in Two-Stage Sampling 276
10.4 Sample Estimation 01 the Vllriance ..... 27M
10.5 The Estimation of Proportio ns .. .. . . . 27')
10.6 Optimum Sampling and Subsampling Fractions 2HO
10.7 Estimation of m.,. from a Pilot Survey 2X3
10J! Three-Stage Sampling . . . . . . . . . . 2XS
10.9 Stratified Sampling of the Unib . . . ... . 2XX
10.10 Optimum Allocation with Stratified Sampling 2X')

Exercises 2')0

CHAPTER
11 SUBSAMPLlNG WITH UNITS OF UNEQUAL
SiZES 2')2

11 .1 Introduction . . . . . . . . . . . . . . . . . . . 2,}2
11 .2 Sampling Methods when n = I .... .. . . . . . 2');1
11 .3 Sampling with Probability Proportional to Est imat ed Size 2IJ7
II <I Summary of Methods for If = I 2')')
II .."! Sampling Methods When If I . . . . . .. . .. . 300
11 .0 Two Useful Result. . . .. . . . .. ... . ... 300
11 .7 Units Selected with Equal Probabilities : Unbiased Estimator 303
11 .8 Unit · Selected with Equal Probabilities: Ratio to Sile Estimate .. 03
J 1.9 Units Selected with Unequal Probahilitil's with Replacement :
Unbiased Estimator . . . . . . . 306
11. 10 Unit:. elected Without Replacement 301l
11.1 1 Comparison of the Methods ] I0
11 . 12 RlIlios to Another Variable 311
11. 13 hoice of Sampling and ubsampling Fractions. Equal Prob-
IIbilities . . . . . . . . . . . . . . . . . . . . . . . . . 3 13
11.14 Optimum Selection Prohabilities and Sampling and Subsampling
Rates . . . . . . . . . . . . . . . 314
1l.IS Stratified Sampling. Unbia~ed Estimators 316
11 .16 Stratified Sampling. Ratio Estimates . . 317
J 1.17 Nonli near Estimators in Complex Surveys .. I H
1 1.1R Taylor Series xpansion :; II}
11.19 Balanced Repeated Replications .120
11.20 The Jackknife Method . . . . ;121
11.21 Comparison of the Three Approaches ;1.:!2
Exerdses .12~
 CONTENTS xv
C HAPTER
J2 DOUBLE SAMPLING 327

12 . 1 De~cription of th~ Technique 327

12.2 Double Sampling for Stratification 327
12 . ~ Optimum Alloca(ion ..... 331
12 .4 Esti ma ted Vari a nc~ in Double Sampling for Stratification 333
12.5 Douhle Sampling for Analytical Compar i s('ln~ . . . . . 335
12 .6 Regres&lon E~ tim a t or~ . . . . . . . . . . . . . . . 338
12.7 Optimum A lloc(ltion and omparison with Si ngl e Sampli ng 341
12.1:1 Estimated Variam;e in Douhle Sampling for Regression 343
12 .9 Ratio l~~tim d tor~ . . . . . . . . . . . 343
12 . 1() Ro;:pcaled Sampling of the Same Population 344
12. 11 Sampling on Two (kca~ llIn ~ . . .... 346
12 . 12 Samplin)!onM{)felh"nT\\"(kca~iun~ 348
12 . 1.1 Simp li ficiltlon~ and Furtlll:r f)c·\,..: lopmcnh 351
(:'xerc;sp.\· 355

:HAPTER.
13 SOURCES OF ERROR IN SURVEYS 359

13 . 1 Introduction 359
11 .2 EIT..:ct~of N(lnre~pnr"c 359
1.1 ..' Type" of Nonr\:#ponsc 364
11.-1 Call-backs 365
1.1.5 A Mathematical Modcl of the liect~ of Call-backs 367
13 .1\ Optimum Sam plin g ruction Amung the Nonrcspondents 370
1.1 .7 Adj ustm.:nth for Biu~ Without Call-hacks . . . . 374
1.1 .X A Mathematical Model for Errors of Mea~uremcnt . . . 377
IJ.t) Effects of Constan t Bias ... .. . . 379
D . IO EtTects of Errors that Are Uncorrelated Within the Sample 380
D . II tfcct~ of fnt rasample Corre lation Betwc.:n Errors of Measure -
ment . . . . . . . . . . . . . . 383
13 . 12 Summary of the Etfectsof Errors of Mea ~ urem<.!nt 384
13 .13 The Study of E rrors of Mea~urement . 3114
13.14 Repeated Measurement of Suh~amples . . . . . 3116
1.1 . 15 Interpenetrati ng Subsamples . . . . . . . . . 388
1.1 . 16 Combination of Interpenetration and Repeated Measurement 391
13 . 17 Sensitive Questions; Randomized R e~po n ses 392
13.ltl The Unrelated Second Question 393
13.19 Summary . . . . . . . .. . . . . . . 395
Exercises 396
xvi CONTENTS

References 400
Answers to Exercises 41 2
Author Index 419

Subject Index 422

CHAPTER 1

Introduction

1.1 ADV ANT AGES OF THE SAMPLING METHOD

Our knowledge, our attitudes, and our actions are based tc a '~ery large extent
on samples. This is equally true in everyday life and in scientific research. A
person's opinion of an institution that conducts thousands of transactions every
day is often determined by the one or two encounters he has had with the
Institution in the course of several year~, Travelers who spend 10 days in a foreign
l'o llntry and then proceed to write a book telling the inhabitants how to revive
their industries , reform their political system, balance their budget, and improve
thl; food in their hoteh are a familiar figure of fun. But in a real sense they differ
from the political scientist who devotes 20 years to living and studying in the
t'O untry only in that they b:1"e their conclusions on a much smaller sample of
experience and are less lik(" Y to be aware of the extent of their ignorance. I"
!-.cience and human affairs alik e we lack the resources to study more than a
fragment of the phenomena that might advance our knowledge.
This book contains an account of the body of theory that has been built up to
provide a hackground for good sampling methods. In most of the applications for
which this theory was constructed , the aggregate about which information is
d<! ~II..:J is finite and delimited- the inhabitants of a town, the machines in a
factory, the fish in a lake. In some cases it may seem feasible to obtain the
Information by taking a complete enumeration or census of the aggregate .
Administrators accustomed to dealing with censuses were at first inclined to be
' u~p ic i()u~ of samples and reluctant to u, e them in place of censuses. Although this
attitude no longer persists, it may be well to list the principal advantages of
sampling as compared with complete enumeration .
Reduced Cost
If data are secured from only (t small fraction of the aggregate , expenditures are
smaller than if a complete censu~ is attempted . With large popUlations, results
accurate enough to be useful can be obtained from samples that represent only a
~ mall fraction of the population. In the United States the most important
rl: ~urrent surH' Ys taken by the government use samples of around 105,000

I
2 SAMPUNG TECHNIQUES

persons, or about one person in 1240. Surveys used to provide facts bearing on
sales and advertising policy in market research may employ samples of only a few
thousand.

Greater Speed
For the same reason, the data can be collected and summarized more quickly
with a sample than with a complete count. This is a vital consideration when the
information is urgently needed.
Greater Scope
In certain types of inquiry highly trained personnel or specialized equipment,
limited in availability, must be used to obtain the data. A complete census is
impracticable: the choice lies between obtaining the information by sampling or
not at all. Thus surveys that rely on sampling have more scope and flexibility
regarding the types of information that can be obtained. On the other hand, if
accurate information is wanted for many subdivisions of the population, the size of
sample needed to do the job is sometimes so large that a complete enumeration
offers the best solution.
Greater Accuracy
Because personnel of higher quality can be employed and given intensive
training and becau e more careful supervision of the field work and processing of
results becomes feasible when the volume of work is reduced, a sample may
produce more accurate results than the kind of complete enumeration that can be
taken .

1.2 SOME USES OF SAMPLE SURVEYS

To an observer of developments in sampling over the last 25 years the most
striking feature is the rapid increase in the number and types of surveys taken by
sampling. The Statistical Office of the United Nations publishes reports from time
to time on "Sample Surveys of Current Interest" conducted by member countries.
The 1968 report lists survey.s from 46 countries. Many of these surveys seek
information of obvious importance to national planning on topics such as agricul-
tural production and land use , unemployment and the size of the labor force,
industrial production , wholesale and retail prices, health status of the people, and
family incomes and expenditures. But more specialized inquiries can also be
found: for example, annual leave arrangements (Australia), causes of divorce
(Hungary), rural debt and investment (India), hou ehold water consumption
(Israel), radio listening (Malaysia), holiday spending (Netherlands), age structure
of cows (Czechoslovakia), and job vacancies (United States).
Sampling has come to playa prominent part in national decennial censuses. In
the United States a 5% sample was introduced into the 1940 Census by asking
 INTRODUCfION 3

extra questions about occupation, parentage, fertility, and the like, of those
persons whose names fell on two of the 40 lines on each page of the schedule. The
use of sampling was greatly extended in 1950. From a 20% sample (every fifth
line) information was obtained on items uch as income, years in school, mjgra-
tion , and service in armed forces . By taking every sixth person in the 20% sample,
a further sample of 3~% was created to give information on marriage and fertility.
A series of questions dealing with the condition and age of housing was split into
five sets, each set being filled in at every fifth house. Sampling was also employed
to speed up publication of the results. Preliminary tabulations { r many important
items, made on a sample basis, appeared more than a year and half before the final
reports.
This process continued in the 1960 and] 970 Censuses. Except for certain basic
information required from every person for constitutional or legal reasons, the
whole census was shifted to a sample basi . This change. accompanied by greatly
increased mechanization , resulted in much earlier publication and substantial
savings.
In addition to their use in censuses, continuing samples are employed by
government bureaus to obtain current information. In the United States, exam-
ples are the Current Population Survey, which provides monthly data on the size
and composition of the labor force and on the number of unemployed, the
National Health Survey, and the series of samples needed for the calculation of
the monthly Consumer Price Index.
On a smaller scale , local governments-city, state. and county- are making
increased use of sample surveys to obtain information needed for future planning
and for meeting pressing problems. In the United States most large cities have
commercial agencies that make a business of planning and conducting sample
surveys fnr clients.
Market research is heavily dependent on the sampling approach . Estimates of
the sizes of television and radio audiences for different programs and of news-
paper and magazine reader hip (including the advertisements) are kept continu-
ally under scrutiny. Manufacturers and retailers want to know the reactions of
people to new products or new methods of packaging, their complaints about old
products, and their reasons for preferring one product to another.
Business and industry have many uses for sampling in attempting to increase the
efficiency of their internal operations. The important areas of quality control and
acceptance sampling are outside the scope of this book. But, obviously, decisions
taken with respect to level or change of quality or to acceptance or rejection of
batches are well grounded only if results obtained from the sample data are valid
(within a reasonable tolerance) for the whole batch. The sampling of records of
business transaction (accounts, payrolls, stock, personnel)-u ually much easier
than the sampling of people- can provide serviceable information quickly and
economically. Savings can also be made through sampling in the estimation of
inventories, in studies of the condition and length of the life of equipment, in the
4 SAMPUNG TECHNIQUES

inspection of the accuracy and rate of output of clerical work, in investigating how
key personnel distribute their working time among different tasks, and, more
generally,.in the field known as operations research. The books by Deming (1960)
and Slonim (1960) contain many interesting examples showing the range of
applications of the sampling method in business.
Opinion, attitude, and election polls, which did much to bring the technique of
sampling before the public eye, continue to be a popular feature of newspapers. In
the field of accounting and auditing, which has employed sampling for many years,
a new interest has arisen in adapting modern developments to the particular
problems of this field . Thus, Neter (1972) describes how airlines and railways save
money by using samples of records to apportion income from freight and
passenger service. The status of sample surveys as evidence in lawsuits has also
been subject to lively discussion. Gallup (1972) has noted the major contribution
that sample surveys can make to the process of informed government by deter-
mining quickly people's opinions on proposed or new government programs and
has stressed their role as sources of information in social science.
Sample surveys can be classified broadly into two types- descriptive and
analytical. In a descriptive survey the objective is simply to obtain certain
information about large groups: for example, the numbers of men, women, and
children who view a television program. In an analytical survey, comparisons are
made between different subgroups of the population, in order to discover whether
differences exist among them and to form or to verify hypotheses about the
reasons for the e differences. The Indianapolis fertility survey, for instance, was
an attempt to determine the extent to which married couples plan the number and
spacing of children, the husband's and wife's attitudes toward this planning, the
reasons for these atti1lldes, and the degree of success attained (Kiser and Whelp-
ton, 1953).
The distinction between descriptive and analytical surveys is not, of course,
clear-cut. Many surveys provide data that serve both purposes. Along with the rise
in the number of descriptive surveys, there has, however, been a noticeable
increase in surveys taken primarily for analytical purposes, particularly in the
study of human behavior and health. Surveys of the teeth of school children before
and after fluoridation of water, of the death rates and causes of death of people
who smoke different amounts, and the huge study of the effectiveness of the Salk
polio vaccine may be cited. The study by Coleman (1966) on equality of
educational opportunity, conducted on a national sample of schools, contained
many egression analyses that estimated the relative contributions of school
characteristics, home b.:lckground, and the child's outlook to variations in exam
results.

1.3 THE PRINCIPAL STEPS IN A SAMPLE SURVEY

As a preliminary to a liiscussion of the role that theory plays in a sample survey,
it is useful to describe briefly the steps involved in the planning and execution of a
 INTRODUCTION 5
survey. Surveys vary greatly in their complexity. To take a sample from 5000
cards, neatly arranged and numbered in a file, is an easy task. It is another matter
to sample the inhabitants of a region where transport is by water through the
forests, where there are 'no maps, where 15 different dialects are spoken, and
where the inhabitants are very suspicious of an inquisitve stranger. Problems that
are baffling in one survey may be trivial or nonexistent in another.
The principal step in a survey are grouped somewhat arbitrarily under 11
headings.

Objectives of the Survey

A lucid statement of the objectives is most helpful. Without this, it is easy in a
complex survey to forget the objectives when engrossed in the details of planning,
and to make decisions that are at variance with the objectives.

Population to be Sampled
The word population is used to denote the aggregate frorn which the sample is
chosen. The definition of the population may present no problem, as when
sampling a batch of electric light bulbs in order to estimate the average length of
life of a bulb. In sampling a population of farms, on the other hand, rules must be
set up to define a farm , and borderline cases arise . These rules must be usable in
practice: the enumerator must be able to decide in the field, without much
hesitation, whether or not a doubtful case belongs to the population.
The population to be sampled (the sampled population) should coincide with
the population about which information ill wanted (the target population). Some-
times, for reasons of practicability or convenience. the sampled population is
more restricted than the target population. If so, it should be remembered that
conclusions drawn from the sample apply to the sampled population. Judgment
about the extent to which these conclusions will also apply to the target population
must depend on other sources of information. Any supplementary information
that can be gathered about the nature of the differences between sampled and
target population may be helpful.

Data to be CoUected
It is well to verify that all the data are relevant to the purposes of the survey and
that no essential data are omitted. There is frequently a tendency, particularly
with human populations, to ask too many questions, some of which are never
subseq uently analyzed. An overlong questionnaire lowers the quality of the
answers to important as well as unimportant questions.
Degree of Precision Desired
The results of sample surveys are always subject to some uncertainty because
only part of the population has been measured and because of errors of measure-
ment. This uncertainty can be reduced by taking larger samples and by using
6 SAMPUNO TECHNIQUES

superior instruments of measurement. But this usually costs time and muney.
Consequently, the specification of the degree of precision wanted in the results is
an important step. This step is the responsibility of the person who is going to use
the data. It may present difficulties, since many administrators are unaccustomed
to thinking in terms of the amount of error that can be tolerated in estimates,
consistent with making good decisions. The statistician can often help at this stage.

Methods of Measurement
There may be a choice of measuring instrument and of method of approach to
the population. Data about a person 's state of health may be obtained from
statements that he or she makes or from a medical examina,tion. The survey may
employ a self-administered questionnaire, an interviewer who reads a standard
set of questions with no discretion, or an interviewing process that allows much
latitude in the form and ordering of the questions. The approach may be by mail,
by telephone, by personal visit, or by a combination of the three. Much study has
been made of interviewing methods and problems (see, e.g., Hyman, 1954 and
Payne, 1951).
A major part of the preliminary work is the construction of record forms on
which the questions and answers are to be entered. With simple questionnaires,
the answers can sometimes be precoded-that is, entered in a manner in which
they can be routinely transferred to mechanical equipment. In fact , for the
construction of good record forms, it is necessary to visualize the structure of the
final summary tables that will be used for drawing conclusions.

The Frame
Before selecting the sample, the population must be divided into parts that are
caJled sampling units, or units. These units must cover the whole of the population
and they must not overlap, in the sense that every element in the population
belongs to one and only one unit. Sometimes the appropriate unit is obvious, as in
a population of light bulbs, in which the unit is the single bulb. Sometimes there is
a choice of unit. In sampling the people in a town, the unit might be an individual
person, the members of a family, or all persons living in the same city block. In
sampling an agricultural crop, the unit might be a field, a farm, or an area of land
whose shape and dimensions are at our disposal.
The construction of this list of sampling units, called a frame , is often one of the
major practical problems. From bitter experience, samplers have acquired a
critical attitude toward lists that have been routinely collected for some purpose.
Despite assurances to the contrary, such lists are often found to be incomplete, or
partly illegible, or to contain an unknown amount of duplication. A good frame
may be hard to come by when the population is specialized, as in populations of
bookmakers or of people who keep turkeys. Jessen (1955) presents an interesting
method of constructing a frame from the branches of a fruit tree.
 INTRODUCTJON 7
Selection of the Sample
There is now a variety of plans by which the samplc may be selected. For each
plan that is considered, rough estimates of the !.Ize of sample can be made from a
knowledge of the degree of precision desired. The relative costs and time involved
for each plan are also compared before making a decision .

The Pretest
It has been found useful to tryout the questionnaire and the field methods on a
small scale. This nearly always results in improvements in the questionnaire and
may reveal other troubles that will be serious on a large scale, for example, that the
cost will be much greater than expected.

Organization of the Field Work

In extensive surveys many problem. of business administration are met. The
personnel must receive training in the purpose of the survey and in the methods of
measurement to be employed and must be adequately surervised in their work . A
procedure for early chieking of the quality of the returns is invaluable. Plans must
be made for handling nonresponse. that is, the failure of the enumerator to obtain
information from certain of the units in the sample.

Summary and Analysis of the Data

The first step is to edit the completed questionnaires, in the hope of amending
recording errors, or at least of deleting data that are obviously erroneous.
Decisions about computing procedure are needed in cases in which answers to
certain questions were omitted by some respondents or were dele ted in the editing
process. Thereafter, the computations that lead to the estimates are performed .
Different methods of estimation may be available for the same data .
In the presentation of results it is good practice to rerort the amount of error I<l
be expected in the most important estimates. One of the advantages of probability
sampling is that such statements can be made, altho ugh they have to be severely
qualified if the amount of nonresponse is substantial.

Information Gained for Future Surveys

The more information we have initially about a population, the easier it is to
devise a sample that will give accurate estimates. Any completed s(lmple is
potentially a guide to improved future sampling, in the data that it supplies (Jbout
the means, standard deviations, and nature of the variability of the princir' I
measurements and about the costs involved in getting the data . Sampling practi ce
advances more rapidly when provisions are made to assemble and record informa -
tion of this type.
There is another important respect in which any completed samrle facilitates
fut ure samples. Things never go exactly as planned in a complex survey. The alert
8 SAMPLING TECHNIQUES

sampler learns to recognize mistakes in execution and to see that they do not occur
in future surveys.

1.4 THE ROLE OF SAMPLING THEORY

This list of the steps in a sample survey has been given in order to emphasize that
sampling is a practical business, which calls for several different types of skill. In
some of the steps-the definition of the population, the determination of the data
to be collected and of the methods of measurement, and the organization of the
field work-sampling theory plays at most a minor role. Although these topics are
not discussed further in this book, their importance should be realized. Sampling
demands attention to all phases of the activity: poor work in one phase may ruin a
survey in which everything else is done well.
The purpose of sampling theory is to make sampling more efficient. It attempts
to develop methods of sample selection and of estimation that provide, at the
lowest possible cost, estimates that are precise enough for our purpose. This
principle of specified precision at minimum cost recurs repeatedly in the presenta-
tion of theory.
[n order to apply this principle, we must be anle to predict, for any sampling
procedure that is under consideration, the precision and the cost to be expected.
So far as precision is concerned, we cannot foretell exactly how large an error will
be present in an estimate in any specific situation, for this would require a
knowledge of the true value for the population. Instead, the precision of a
sampling procedure is judged by examining the frequency distribution generated
for the estimate if the procedure is applied again and again to the same population.
This iSI of course, the standard technique by which precision is judged in statistical
theory.
A further simplification is introduced. With samples of the sizes that are
common in practice, there is often good reason to suppose that the sample
estimates are approximately normally distributed. Witli a normally distributed
estimate, the whole shape of the frequency distribution is known if we know the
mean and the standard deviation (or the variance). A considerable part of sample
survey theory is concerned with finding formulas for these means and variances.
There are two differences between standard sample survey theory and the
classical sampling theory as taught in books on mathematical statistics. In classical
theory the measurements that are made on the sampling units in the population
are usually assumed to follow a frequency distribution, for example, the normal
distribution, of known mathematical form cipart from certain population parame-
ters such as the mean and variance whose values have to be estimated from the
sample data. In sample survey theory, on the other hand, the attitude has been to
assume only very limited information about this frequency distribution. In
particular, its mathematical form is not assumed known, so that the approach
might be described as model-free or distribution-free. This attitude is natural for
 INTRODUCTION 9
largc survey!> in which many different measurements with differing frequency
distributions are made on the units. In surveys in which only a few measurements
per unit are made, studies of their frequency distributions may justify the
a~s umption of known mathematical forms, permitting the results from classical
theory to be applied .
A second difference is that the populations in survey work contain a finite
number of units. Re ~ ults are slightly more complicated when sampling is from a
finite instead of an infinite population. For practical purposes these differences in
result for finite and infinite populations can often be ignored. Cases in which this
is not so will be pointed out.

1.5 PROBABILITY SAMPLING

The sampling procedures considered in this book have the following mathemat-
ical properties in common.
I . We are able to define the set of distinct samples, 5 1,52 , • •• , 5u , which the
rrocedure is capable of se lecting if applied to a specific population .. This mean!>
tha t we can say precisely what sampling units belong to 51, to 52, and so on. For
example, suppo!>c that the population contains six units, numbered I to 6. A
common procedure for choosing a sample of sIze 2 gives three possible
ca ndidates-5 1 - ( 1. 4); 5 2 - (2.5); 5 3 - (3,6). Note that not all possible sa mples
of size 2 need be included .
2. Each possible sample 5, has assigned to it a known probability of selection
1T,•
3. We se lecl one of the 5, by a random process in which each 5, receives its
appropriate probability 1TI of being selected. In the example we might assign equal
probabilities to the three samples. Then the draw itself can be made by choosing a
random number belween 1 and 3. lt this number is j , 5j is the sample Ihat i lake n.
4. The method for computing the estimate from the sample must be stated and
must lead to a unique estimate for any specific sample. We may declare, for
example, that the estimate is to be the average of the measurements on the
individual units in the sample.
For any sampling procedure that satisfies these properties, we are in a position
to calculate the frequency distribution of the estimates it generates if repeatedly
applied to the same population. We know how frequently any particular sample 51
will be selected , and we know how to calculate the estimate from the data in 5" It is
clear, therefore, that a sampling t. eory can be developed for any procedure of this
type, although the details of the development may be intricate. The term
prvbability sampling refers to a method of this type .
In practice we seldom draw a probability sample by writing down the 51 and 1Tj
as outlined above. This is intolerably laborious with a large population, where a
sampling procedure may produce billions of possible samples. The draw i most
10 SAMPUNG TECHNIQUES

commonly made by specifying probabilities of inclusion for the individual units

and drawing units, one by one or in groups until the sample of desired size and type
is constructed. For the purposes of a theory it is sufficient to know that we could
write down the S, and 'TTl if we wanted to and had unlimited time.

1.6 ALTERNA TIVES TO PROBABILITY SAMPLING

The following are some common types of nonprobability sampling.
1. The sample is restricted to a part of the population that is readily accessible.
A sample of coal from an open wagon may be taken from the top 6 to 9 in.
2. The sample is selected haphazardly. In picking 10 rabbit from a large cage
in a laboratory, the investigator may take those that his hands rest on, without
conscious planning.
3. With a smaJl but heterogeneous population, the sampler inspects the whole
of it and selects a small sample of " typical " units- that is, units t!-tat are close to his
impression of the average of the population.
4. The sample consists essentiaJly of volunteirs, in studies in which the
measuring process is unpleasant or troublesome to the person being measured.
Under the right conditions. any of these methods can give useful results. They
are not, however. amenable to the development of a sampling theory that is
model-free , since no element of random selection is involved. About the only way
of examining how good one of them may be is to find a situation in which the
results are known , either for the whole population or for a probability sample, and
make comparisons. Even if a method appears to do well in one such comparison,
this does not guarantee that it will do well under different circumstances.
In this connection. some of the earliest uses of sampling by country and city
governments from ] 850 onward were intended to save money in making esti-
mates from the results of a Census. For the most important items in the Census,
the country or city totals were calculated from the complete Census data. For
other items a sample of say 15 or 25% of the Census returns was selected in order
to lighten the work of estimating country or city totals for these items. Two rival
methods of sample selection came into use. One, called random selection, was an
application of probability sampling in which each unit in the population (e.g., each
Census return) had an equal chance of being included in the sample. For this
method it was realized that by use (If sampling theory and the normal distribution,
as noted previously, the sampler could predict approximately from the sample
data the amount of error to be expected in the estimates made from the sample.
Moreover, for the most important items for which complete Census data were
available, he could check to some extent the accuracy of the predictions.
The other method was purposive selection. This was not specifically defined in
detail but usually had two common features. The sampling unit consisted of
groups of returns, often relatively large groups. For exarpple, in the 1921 Italian
 INTRODUCfION 11
Census the country consisted of 8354 communes grouped into 214 districts. In
dra wing a 14% sample, the Italian statisticians Gini and GaJvani selected 29
districts purposively rather than 1250 communes. Seco nd, the 29 districts were
chosen so that the sample gave aCCurate estimates for 7 important control
va ria bles for which results were known for the whole country. The hope was that
this sa mple would give good estimates for other variables highly correl ated with
the control variables .
In the J 920s the Inte rnational Statistical Institute appointed a commission to
report on th e advantages and disadva ntages of the two me thods. The re port , by
Jense n ( 1926), seemed on balance to favor purposive selection . However , pur-
posive selection was abandoned re latively soo n as a method of sampling for
obtaining national estimates in surveys in which many items were measured. It
lacked the flex ibility that late r developments of proba bility sampling produced, it
was un able to predict from the sa mple the accuracy to be ex pected in the
estimates, and it used sampling units that were too large. Gini and Galvani
concluded that the probability method called st·ratified random sa mpling (Cha pter
5), with the com mune as a sa mpling unit. would have given better results than
their method.

1.7 USE OF THE NORMAL DISTRlBUTION

It is sometimes useful to empl oy the word estimator to denote the rule by which
an es timate of some population characteristic IJ. is calculated from the sampl e
r es ult~ . the word estimate being ap pli ed to the value o btained from a specific
sa mple . An es timator P- of IJ. give n by a sa mpling plan is ca lled unbi ased if the
mC(ln valu e or p-. taken ove r all possible sa mples provided by the pl an, is eq ual to
IJ.. In the notation of section 1.5. this condition may be writte n
l

E(p- ) = L 'TT,p-, = IJ.

where p-, is the estimate give n by the ith sample. The symbol E. which stands for
" the ex pected va lue of." is used frequently.
As me ntioned in section 1.4, the samples in surveys are often large e nough so
that estimates made fro m them are approximately normally distributed . Further-
more , with plobability sampling, we have formulas that give the mean and
va riance of the estimates. Suppose that we have taken a sample by a procedure
known to give an unbiased estimator and have computed the sample estimate p-
and its standard deviation CT,;_ (often called, alternatively, its standard error) . How
good is the estimate? We cannot know the exact value of the error of estimate
(p- - IJ.) but, from the properties of the normal curve, the chances are
0 .32 (about 1 in 3) that the absolute error ip- - IJ. i exceeds CT,;_
0.05 (1 in 20) that the absolute error ip- - IJ.i exceeds 1.96CT,;_ == 2CT,;_
0 .0 I (1 in 100) that the absolute error ip- - Ii- exceeds 2.58CT,;_
12 SAMPUNG TECHNIQUES

For example, if a probability sample of the records of batteries in routine use in

a large factory shows an average life P- ::: 394 days, with a standard error 0'". ::: 4.6
days, the chances are 99 in 100 that the average life in the population of batteries
lies between
P- ::: 394 - (2.58)(4.6) = 382 days
and
P-u = 394 + (2.58)(4.6) = 406 days
The limits, 382 aays and 406 days, are called lower and upper confidence limits.
With a single estimate from a single survey, the statement "IJ. lies between 382 and
406 days" is not certain to be correct. The "99% confidence" figure implies that if
the same sampling plan were used many times in a population, a confidence
statement being made from each sample, about 99% of these statements would be
correct and 1% wrong. When sampling is being introduced into an operation in
which complete censuses have previously been used, a demonstration of this
property is sometimes made by drawing repeated samples of the type proposed
from a population for which complete records exist, so that p- is known (see, e.g.,
Trueblood and Cyert, 1957). The practical verification that approximately the
stated proportion of statements is correct does much to educate and reassure .
administrators about the nature of sampling. Similarly, when a single sample is
taken from each of a series of different populations, about 95% of the 95%
confidence statements are correct.
The preceding discussion assumes that 0'"., as computed from the sample, is
known exactly. Actually, O'ii' like p-, is subject to a sampling error. With a normally
distributed variable, tables of Student's f distribution are used instead of the
normal tables to calculate confidence limits for f.L when the sample is small.
Replacement of the normal table by the f table makes almost no difference if the
number of degrees of freedom in 0',;. exceeds 50. With certain types of stratified
sampling and with the method of replicated sampling (section 11 .19) the degrees
of freedom are small and the t table is needed.

1.8 BIAS AND ITS EFFECTS

In sample survey theory it is necessary to consider biased estimators for two
reasons.
1. In some of the most common problems, particularly in the estimation of
ratios, estimators that are otherwise convenient and suitable are found to be
biased.
2. Even with estimators that are unbiased in probability sampling, errors of
measurement and nonreponse may produce biases in the numbers that we are able
to compute from the data. This happens, for instance, if the persons who refuse
to be interviewed are almost all opposed to some expenditure of public funds,
whereas those who are interviewed are split evenly for and against.
 INTRODUcnON 13

0.4

0.2

01

o
m

Fig. 1.1 Effect of bias on errors of estimation .

To examine the effect of bias, suppose that the estimate P- is normally

di~trib utedabout a mean m that is a distance B from the true population value iJ..
as shown in Fig. 1. 1. The amount of bias is B = m - iJ.. Suppose that we do not
know that any bias is present. We compute the standard deviation a of the
freque ncy distribution of the estimate- this will , of course, be the standard
devia tion about the mean m of the distribution, not about the true mean iJ.. We are
u~ing a in place of aiL' As a statement about the accuracy of the estimate, we
declare that the probability is 0.05 that the estimate P- is in error by more than
1.96a.
We will consider how the presence of bias distorts this probability. To do this.
we calculate the true probability that the estimate is in error by more than J .96a.
whe re error is measured from the true mean iJ.. The two tails of the distribution
must be examined separately. For the upper tail, the probability of an error of
more than + 1.96a is the shaded area above 0 in Fig. 1.1. This area is given by

1
---- J"" e - (iL - m) 2/2u d iJ.•
2

a.JZ; '" + I 9617

Put P- - m = at. The lower limit of the range of integration for t i~

iJ. - m + 1.96 = 1.96 - ~

u a
Thus the area is

1
--- J"" e - ,2/ 2 d t
& 1.96- (8/u)
14 SAMPLING TECHNIQUES

Similarly, the lower tail, that is, the shaded area below P, has an area

--
1 f
- 1.96- (8/0')
e- t2 / 2 dt
& -00

From the form of the integrals it is clear that the amount of disturbance depends
solely on the ratio of the bias to the standard deviation . The results are shown in
Table 1.1.

TABLE 1.1
EFFECT OF A BIAS B ON THE PROBABILITY OF AN ERROR
GREATER THAN 1.9611
Probability of Error
B/11 <- 1.9611 > 1.9611 Total
0.02 0.0238 0.0262 0.0500
0.04 . 0.0228 0.0274 0.0502
0.06 0.0217 0.0287 0.0504
0.08 0.0207 0.0301 0.0508
0. 10 0.0197 0.0314 0.0511
0.20 0.0154 0.0392 0.0546
0.40 0.009 1 0.0594 0.0685
0.60 0.0052 0 .0869 0.0921
0.80 0.0029 0. 1230 0.1259
1.00 0.0015 0.1685 .0. 1700
1.50 0.0003 0.3228 0.3231

For the total probability of an error of more than 1.96CT, the bias ha little effect
provided that it is less than one tenth of the standard deviation. At this point the
total probability is 0.0511 instead of the 0.05 that we think it is. As the bias
increases further, the disturbance becomes more serious. At B = CT, the total
probability of error is 0.17, more than three times the presumed value.
The two tails are affected differently. With a positive bias, as in this example, the
probability of an underestimate by more than 1.96CT shrinks rapidly from the
presumed 0.025 to become negligible when B = CT. The probability of the corres-
ponding overestimate mounts steadily. In most applications the total error is the
primary interest, but occasionally we are particularly interested in errors in one
direction.
As a working rule, the effect of bias on the accuracy of an estimate is negligible if
the bias is less than one tenth of the standard deviation of the estimate. If we have
a biased method of estimation for which B/CT < 0.1 , where B is the absolute value
of the bias, it can be claimed that the bias is not an appreciable disadvantage of the
 INTRODUCTION 15
method . Even with BI u == 0.2, the disturbance in the probability of the total error
is modest.
In using these results, a distinction must be made between the two sources of
bias mentioned at the beginning of this section. With biases of the type that arise in
estimating ratios, an upper limit to the ratio BI u can be found mathematically. If
the sample is large enough, we can be confident that BIu will not exceed 0.1. With
biases caused by errors of measurement or nonresponse, on the other hand, it is
usu<:llly impossible to find a guaranteed upper limit to Bl u that is small. This
tro ublesome problem is discussed in Chapter 13 .

1.9 THE MEAN SQUARE ERROR

tn order to compare a biased estimator with an unbiased estimator, or two
estimators with different amounts of bias, a useful criterion is the mean square
error (MSE) of the estimate, measured from the population value that is being
estimated. Formally,
MSEIJ1) = E(J1 - JA.) 2 = E[(J1 - rn) + (rn - JA.')f
= E (J1 - rn) 2+ 2(rn - JA.)E(J1 - rn) +(rn - JA.) 2
= (variance of J1) + (bias) 2
the cross-product term vanishing since E(J1- rn ) = O.
Use of the MSE as a criterion of the accuracy of an estimator amounts to
regarding two estimates that have the same MSE a~ equivalent. Thi is not strictly
correct because the frequency distributions of errors (J1 - iJ.) of different sizes will
not be the same for the two estimates if they have different amounts of bias. It has
heen shown , however, by Hansen, Hurwitz, and Madow (1953) that if Blu is less
than about one half, the two frequency distributIOns arc almost identical in regard
\0 absolute errors 1J1- JA.I of different sizes . Tahle 1.2 illustrates this result.

TABLE 1.2
PROBABILITY Of AN AOSOlUTE ERROR ~ 1 \ MS[ ,
1.96 MS AND 2.576 V MSE
Probability

Birr 1 \MSE 1.96 MS 2.576 \ MSE

0 0 .317 0 .0500 0.0100

0.2 0.317 0 .0499 0.0100
0.4 0.319 0.0495 0.0095
0 .6 0.324 0.0479 0.0083
- ----
16 SAMPUNG TECHNIQUES

Even at B/ u =0.6, the changes in the probabilities as compared with those for
B/u = 0 are slight.
Because of the difficulty of ensuring that no unsuspected bias enters into
estimates, we will usually speak of the precision of an estimate instead of its
accuracy. Accuracy refers to the size of deviations from the true mean p" whereas
precision refers to the size of deviations from the flean m obtained by repeated
application of the sampling procedure.

EXERCISES
1.1 SUPPOSt ' hat you were using sampling to estimate the total number of words in a
book that contains jJlustrations.
(a) Is there an y problem of definition of the population? (b) What are the pros and cons
of (1) the page, (2) the line, as a sampling unit?
1.2 A sample is to be taken from a list of names that are on cards (one name to a card)
numbered consecutively in a file . Each name is to have an equal chance of being drawn in
the sample. What problems arise in the following common situations? (a) Some of the
names do not belong to the target population, although this fact cannot be verified for any
name until it has been drawn . (b) Some names appear on more than one card. All cards with
the same name bear consecutive numbers and therefore appear together in the file . (c)
Some names appear on more than one card, but cards bearing the same name may be
scattered anywhere about the file.
1.3 The problem of finding a frame that is complete and enables the sample to be drawn
is often an obst~c1e. What kinds of frames might be tried for the following surveys? Have
the frames any serious weaknesses? (a) A survey of stores that sell luggage in a large city.
(b) A survey of the kinds of articles left behind in subways or buses. (c) A survey of persons
bitten by snakes during the last year. (d) A survey to estimate the number of hours per week
spent by family members in watching television.
1.4 A city directory, 4 years old, lists the addresses in order along each street, and gives
the names of the persons living at each address. For a current interview survey of the people
in the city, what are the deficiencies of this frame? Can they be remedied by the
interviewers during the course of the field work? In using the directory, would you draw a
list of addresses (dwelling places) or a list of persons?
1.5 In estimating by sampling the actual value of the small items in the inventory of a
large firm, the actual and the book value were recorded for each item in the sample. For the
total sample, the ratio of actual to book value was 1.021, this estimate being approximately
normally distributed with a standard error of 0.0082. If the book value of the inventory is
S80,OOO, compute 95% confidence limits for the actual value.
1.6 Frequently data must be treated as a sample, although at first sight they appear to
be a complete enumeration. A proprietor of a parking lot finds that business is poor on
Sunday mornings. After 26 Sundays in operation, his average receipts per Sunday morning
are exactly SID. The standard error of this figure, computed from week-to-week variations,
is 11.2. The attendant costs S7 each Sunday. The proprietor is willing to keep the lot open at
this time if his expected future profit is S5 per Sunday morning. What is the confidence
probability that the long-term profit rate will be at least S5? What assumption must be
made in order to answer this question?
 INTRODUcnON 17
1. 7 In Table 1.2, what happens to the probability of exceeding ]./MSE. 1. 96./MSE,
and 2.576./MSE·when Blu tends to infinity, that i5, when the MSE is due entirely to bias?
Do your results agree with the directions of the cbanges noted in Table 1.2 as Blu moves
from 0 to 0.67
1.8 When it is necessary to compare two estimates that have different frequency
distributions of errors (,1 - iJ.) , it is occasionally possible, in specialized problems, to
compute the cost or loss that will result from an error (,1 - iJ.) of any given size. The estimate
that gives the smaller expected loss is preferrod, other things being equal. Show that if the
loss is a quadratic function A(,1 - iJ.)2 of the error, we should choose the estimate with the
smaller mean square error.
CHAPTER 2

Simple Random Sampling

2.1 SIMPLE RANDOM SAMPLING

Simple random sampling is a method of selecting n units out of the N such that
everyone of the ~n distinct samples has an equal chance of being drawll . In
practice a simple random sample is drawn unit by unit. The units in the population
are numbered from 1 to N. A series of random numbers between I and N is then
drawn, either by means of a table of random numbers or by means of a computer
program that produces such a table . At any draw the process used must give an
equal chance of selection to any number in the population nOI already drawn . The
units that bear the e n numbers constitute the sample.
It is easily verified that all NCn distinct samples have an equal chance of being
selected by this method. Consider one distinct sample , that is, one set of n
specified units . At the first draw the probability that some one of the 11 specified
units is selected is n/ N. At the second draw the probability that some one of the
remaining (n -1) specified units is drawn is (n - l) / (N -1), and so on . Hence the
the probability that all n specified units are selected in n draws is

.!!... (n - ]) . (n - 2) . . . 1 n!(N-n )!
=-- (2 . 1)
N (N - l) (N - 2) (N - n + 1) (N)! NCn

Since a number that has been. drawn is removed from the population for all
subsequent draws, this method is also called random sampling without replace -
ment. Random sampling with replacement is entirely feasible: at any draw, all.
N members of the population are given an equal chance of being drawn, no matter
how often they have already been drawn . The formulas for the variances and
estimated variances of estimates made from the sample are often simpler when
sampling is with replacement than when it is without replacement. For this reason
sampling with replacement is sometimes used in the more complex sampling
plans, although at first sight there seems little point in having the same unit two or
more times in the sample.
 SIMPLE RANDOM SAMPLING 19

2.2 SELEcnON OF A SIMPLE RANDOM SAMPLE

Tables of random numbers are tables of the digits 0, 1,2, . . . 9, each digit
having an equal chance of selection at any draw. Among the larger tables are those
published by the Rand Corporation (1955)- 1 million digits-and by Kendall and
Smith (1 938)-100,000 digits. Numerous tables are available , many in standard
statistical texts. Table 2. 1 shows 1000 random digits for iJlustration, from
Snedecor and Cochran (1967).

TABLE 2. 1
O NE THOUSAND R AN DOM DIGITS

0()....04 05- 09 10- 14 15- 19 2()""24 25-29 3()""34 35- 39 40-44 45-49

00 54463 22662 65905 70639 79365 67382 ~~085 6983 1 47058 08186
OJ 15389 8 205 18850 39226 42249 90669 9~325 23248 60933 26927
02 85941 40756 824 14 02015 13858 78030 16269 65978 01385 15345
O~ 6 1149 69440 11286 8gZ t8 58925 03638 S2862 62733 33451 77455
04 05219 8J619 10651 67079 9251 1 5988k ~1502 72095 83463 75577

05 414 17 98326 87719 92294 46614 50948 64886 20002 97365 30976
06 21:057 94070 20652 35774 16249 75019 211 4505217 47286 76305
07 17783 00015 10806 8309 1 91530 36466 6248 1
3'998 1 49 J77 75779
08 40950 R4820 29881 85966 .62800 70326 84740 62660 77379 90279
09 82995 64 157 66 J64 411 80 100R9 41757 7825R 96488 88629 3723 1

10 96754 17676 55659 44 105 47361 34833 86679 23930 53249 27083
11 34357 88040 53364 71726 45690 66334 60332 22554 90600 711 13
12 063 18 37403 49927 57715 50423 (,7372 63 116 48888 21505 80182
13 621 11 52820 07243 79931 89292 84767 85693 73947 22278 11 551
14 47534 09243 67879 00544 234 10 12740 02540 54440 32949 13491

15 986 14 75993 84460 62846 59844 14922 48730 73443 48167 34770
16 24856 0364 ~44898 0935 1 98795 18644 39765 71058 90368 44104
17 \16887 12479 8062 1 66223 860115 78285 02432 53342 42846 94771
18 90801 21472 42815 77408 37390 76766 52615 321 41 30261S 18 106
19 55 165 77312 83666 36028 28420 70219 81369 41943 47366 4 1067

In using these tables to select a simple random sample, the first step is to number
the units in the population from 1 to N. If the first digit of N is a number between 5
and 9, the following method of selection is adequate. Suppose N = 528, and we
want n = 10. Select three columns (rom Table 2. 1, say columns 25 to 27. Go down
the three columns, e lecting the first 10 distinct numbers betwee n 001 and 528.
These are 36, 509, 364, 417, 348, 127, 149, 186,290, and 162. For the last two
numbers we jumped to columns 30 to 32. In repeated selecti ons it is advisable to
va ry the starting point in the table.
20 SAMPUNO TECHNIOUES

The disadvantage of this method is that the three-digit numbers 000 and 529 to
999 are not used, although skipping numbers does not waste much time. When the
first digit of N is less than 5, some may still prefer this method if n is small and a
large table of random digits is available.
With N = 128, for example, a second method that involves less rejection and i
easily applied is as follows. In a series of three-digit numbers, subtract 200 from all
numbers between 201 and 400, 400 from all numbers between 401 and 600, 600
from aU numbers between 601 and 800, 800 from all numbers between 801 and
999 and, of course, 000 from all numbers between 000 and 200. All remainders
greater than 129 and the numbers 000, 200, and so forth, are rejected. USing
columns 05 to07 in Table 2.1, we get 26,52,7,94,16,48,41,80,128, and 92, the
draw requiring 15 three-digit numbers for n = 10. In this sample the rejection rate
5/15 = 33% is close to the probability of rejection 72/200 = 36% for this method.
In using this method with a number N like 384, note that one subtracts 400 from a
number between 401 and 800, but automatically rejects all numbers greater than
800. Subtraction of 800 from numbers between 801 and 999 would give a higher
probability of acceptance to remainders between 001 and 199 than to remainders
between 200 and 384.
Other methods of sampling are often preferable to simple random sampling on
the grounds of convenience or of increased precision. Simple random sampling
serves best to introduce sampling theory.

2.3 DEFINITIONS AND NOTATION

In a sample survey we decide on certain properties that we attempt to measure
and record for every unit that comes into the sample. These properties of the units
are referred to as characteristics or, more simply, as items.
The values obtained for any specific item in the N units that comprise the
population are denoted by Yl> Y2, . . . , YN' The corresponding values for the units
in the sample are denoted by Yh Y2, ... ,Yn, or, if we wish to refer to a typical
sample member, by yt(i = I, 2, . .. , n). Note thatthe sample will not consist of the
first n units in the population. except in the instance, usually rare, in which these
units happen to be drawn. If this point is kept in mind, my experience has been that
no confusion need result.
Capital letters refer to characteristics of the population and lowercase letters to
those of the sample. For totals and means we have the following definitions.
Population Sample

Total: Y=LY,
N
=YI +Yl+" '+ YN
"
r y, = YI +Y2+" . + Y.

y = YI+Yl +' " +YN

N

LY, _ Y,+Y2+" ·+Y.

.
LY,
Mean : y=
N N n n
 SIMPLE RANDOM SAMPUNO 21
Although sampling is undertaken for many purposes, interest centers most
frequently on four characteristics of the population.

1. Mean == Y (e.g. , the average number of children per school).

2. Total == Y (e.g., the total number of acres of wheat in a resion).
3. Ratio of two totals or means R = Y/ X == Y/ X (e.g. , ratio of liquid assets to
total assets in a group of families) . '
4. Proportion of units that fall into some defined class (e.g., proportion of
people with false teeth).

Estimation of the first three quantities is discussed in this chapter.

The symbol denotes an estimate of a population characteristic made from a
A

sample. In this chapter only the simplest estimators are considered.

Estimator

Population mean Y y= y =sample mean

Population total Y 9'= Ny == Nry,/n
Population ratio R y/
R == YI i = £ £x,

In Y the factor N / n by which the sample total is multiplied is sometimes called

the expansion or raising or inflation factor . Its inverse n/N, the ratio of tbe size of .
the sample to that of the population, is called the sampling fraction and is denoted
by the letter f.

1.4 PROPER11ES OF TIlE ESTIMATES

The precision of any estimate made from a sample depends both on the method
by which the estimate is calculated from the sample data and on the plan of
sampling. To save space we sometimes write of " the precision of the sample
mean" or " the precision of simple random sampling," without specifically men-
tioning the other fundamental factor. This has been done, we hope, only in
instances in which it is clear from the context what the missing factor is. When
studying any formula that is presented, the reader should make sure that he or she
knows the specific method of sampling and method of estimation for which the
formula has been established.
In this book a method of estimation is called consistent if the estimate becomes
exactly equal to the population value when n == N, that is, when the sample
consists of the whole population. For simple random sampling it is obvious that y
and Ny are consistent estimates of the population mean and total, respectively.
Consistency is a desirable property of estimators. On the other band, an inconsis-
tent estimator is not necessarily useless, since it may give satisfactory precision
When n is small compared to N. Its utility is likely to be confined to .s situation.
22 SAMPUNG TECHNIQUES

Hansen, Hurwitz, and Madow (1953) and Murthy (1967) give an alternative
definition of consistency, similar to that in classical statistics. An estimator is
consistent if the probability that it is in error by more than any given amount tends
to zero as the sample becomes large. Exact statement of this definition requires
care with complex survey plans.
As we have seen, a method of estimation is unbiased if the average value of the
estimate, taken over all possible samples of given size n, is exactly equal to the true
population value. If the method is to be unbiased without qualification, this result
must hold for any population of finite values Yi and for any n. To investigate
whether y is unbiased with simple random sampling, we calculate the value of 9
for all NC" samples and find the average of the estimates. The symbol E denotes
this average over all 'possible samples.
1beorem 2.1. The sample mean y is an unbiased estimate of Y.
Proof. By its definition
E -- L9 _L(Yt+Y2+ " '+y,,)
(2.2)
Y - ~"- n[N!/n!(N - n)!]
where the sum extends over all ",c" samples. To evaluate this sum, we find out in
how many samples any specific value Yi appears. Since there are (N -1) other
units available for the rest of the sample and (n - 1) other places to fill in the
sample, the number of samples containing Yi is
(N-l)!
(2.3)

Hence
(N - 1)1
L (Yt +Y2+' . . +y") = (n -l)!(N~ n)!(Y' + Y2 + ' .. + YN)

From (2 .2) this gives

_ (N-l)! n!(N-n)!
EY = (n - l)!(N - n)! nN! (YJ+Y 2+ " '+YN)

= (YI+Y 2+" ' +YN) = y (2.4)

N
Corollary. Y = Ny is an unbiased estimate of the population total Y.
A less cumbersome proof of theorem 2.1 is obtained as follows. Since every unit
appears in the same number of samples, it is clear that
E(Yl + Y2 + ... + Y") must be some multiple of Yl + Yz + ... + YN (2.5)
The multiplier must be n/ N, since the expression on the left has n terms and that
on the right has N terms. This leads to the result.
 SIMPLE RANDOM SAMPLING 23
2.5 VARIANCES OF THE ESTIMATES
The variance of the y, in a finite population is usually defined as

(2.6)

As a matter of notation , results are presented In terms of a slightly different

express ion, in which the divisor (N -1) is used instead of N. We take
N
L (y, - Y)2
S :=..:..' - - - (2.7)
N- l
This convention has been usea by those who approach sampling theory by means
of the analysis of variance. Its advantage is that most results take a slightly simpler
form . Provided that the same notation is maintained consistently, all results are
eq ui valent in either notation .
Wc now consider the variance of y. By this we mean E(y - Y)2 taken over all
NC, samples.
Theorem 2.2. The variance of the mean y from a simple random sample is
- 2 S2 (N - n ) S2
V(y) = E(y-y) = - - - = - ( l - j) (2.8)
n N n
where f = n/ N is the sampling fraction .
Proof.
(2.9)

By the argument of symmetry used in relation (2 .5), it follows that

and also that

E[(Y' - Y)(Y 2-Y}+(Y'-Y)(YJ-Y)+ .,. +(Yn - ,-Y)(Yn-Y)]

n(n - 1) - - - -
= N(N - l)[(Y' - Y)(Y 2 - y) + (y,- Y)(Y3 - Y)

(2.11)

In (2.1 1) the sums of products extend over all pairs of units in the sample and
population. respecti ve ly. The sum on the left contains n (n - 1)/2 terms and that
on the right contains N(N - I )/2 terms .
24 SAMPUNO TECHNIQUES

Now square (2.9) and average over all simple random samples. Using (2.10) and
(2.11), we obtain
2 .:. 2 n{ - 2
n E(y- Y) = N (~I- Y) + .. ·+(YN- Y)
- 2

~( n -1) - - - - }
+ N-l [(YI-Y)(Y2-Y)+·· ·+(YN-I-Y)(YN-Y)]
Completing the square on the cr6ss-product term, we have

2
n E(y-Y) = N
-2 n{( I-N_l
n-1) [(YI-Y) + .. ·+(YN-Y)] -2 -2

n -1 -
+N-l [(YI- Y)+ .. ·+(YN- Y)]
- 2}
The second term inside the curly bracket vanishes, since the sum of the Y, equals
NY. Division by n 2 gives
V(y) = E(y- Y)2= N-n
nN(N-l) I-I
f (YI_ Y)2= S2n (N-N n)
This completes the proof.
CoroUary 1. The standard error of y is
S S
Uj=~=:;;;n-t (2.12)

CGroIIary 2. The variance of Y = Ny, as an estimate of the population total Y, is

2 2 2 2
V(Y> = E(Y - Y)2= N S (N-n) = N S (1- f) (2.13)
n N n
CoroUary 3. The standard error of Y is
NS NS
(T~=~=i;!8 (2.14)

2.6 TIlE FINITE POPULA110N CORREC110N

For a random sample of size n from an infinite population, it is well known that
the variance of the mean is u 2 / n. The only change in this result when the
population is finite is the introduction of the factor (N-n)/N. The factors
(N - n)/N for the variance and J(N -n)/Nfor the standard error are called the
finite population corrections (fpc). They are given with a divisor (N -1) in place of
N by writers who present results in terms of (T. Provided that the sampling fraction
n/ N remains low, these factors are close to unity, and the size of the population as
 SI~LE RANDOM SAMPUNG 25
such has no direct effect on the standard error of the sample mean. For instance, if
S is the same in the two populations, a sample of 500 from a population of 200,000
gives almost as pQ!cise an estimate of the population mean as a sample of 500 from
a population of 10.,,000. Persons unfamiliar with sampling often find this result
difficult to believe and, indeed, it is remarkable. To them it seems intuitively
obvious that if information has been obtained about only a very small fraction of
the population, the sample mean cannot be accurate. It is instructive for the reader
to consider why this point of view is erroneous.
In practice the fpc can be ignored whenever the sampling fraction does not
exceed 5% and for many purposes even if'it is as high as 10% . The effect of
ignoring the correction is to overestimate the standard error of the estimate y.
The following theorem, which is an extension of theorem 2.2, is not required for
the discussion in this chapter, but it is proved here for later reference .
Theorem 2.3. If Yj, Xj are a pair of variates defined on every unit in the
population and y, i are the corresponding means from a simple random sample of
size n, then their covariance
_ _ N- n 1 N _ _
E (y - Y)(i - X) = nN N -11~1 (YI - y)(x j - X) (2.15)

This theorem reduces to theorem 2.2 if the variates YI, XI are equal on every unit.
I7oo/:. ~ply theorem 2.2 to the variate Uj = YI + Xj ' The population mean of UI
is U :c y + X, and theorem 2.2 gives

E(u - D'l = N - n _1_

nN N ..... I I
r
_1
(Uj - 0)2

that is
- - 2 N- 1 1 ~ - - 2
E[(y - y)+(i - X)) =N- N - l t... [(YI - y)+(Xj - X)] (2.16)
n ,_I
Expand tl1e quadratic terms on both sides. By theorem 2.2,
- 2 N- n 1 2
E(y- Y) = --
nN N- l LN (YI -
j _
-
Y)
1

with a similar relation for E(i - x'l Hence these two terms cancel on the left and
right sides of (2.16). The result of the theorem (equation 2.15) follows from the
cross-product terms.

2.7 ES11MADON OF 11IE STANDARD ERROR

FROM A SAMPLE
The formulas for the standard errors of the estimated population mean and
total are used primarily for three purposes: (1) to compare the precision obtained
by simple random sampling with that given by other methods of sampling, (2) to
26 SAMPLING TECHNIQUES

estimate the size of the sample needed in a survey that is being planned, and (3) to
estimate the precision actually attained in a survey that has been completed. The
formulas involve 5 2, the population variance. In practice this ~ill not be known,
but it can be estimated from the sample data. The relevant result is stated in
theorem 2.4.
Theorem 2.4. For a simple random sample

t (Yi - y)2
S2=..;.I_ _-
n -l
is an unbiased estimate of
N
L (Yi - Y) 2
5 2 =..;.1_ __
N- l
Proof. We may write
2 1" - -2
5 =- L [(Yi - Y)-(y- Y)] (2.1 7)
n - 1i~ 1

1["L
=-
n-l ,- I
-2
(y, - Y) - n(y- Y) -2] (2.18)

Now average over all simple random samples of size n. By the argument of
symmetry used in theorem 2'.2,

E[ i:. (y,- Y)2] =!!. I (Yi - Y)2 =n(N -. 1) 5 2

i- I N i=J N

by the definition of 5 2. Furthermore, by theorem 2.2,

- N- n
E[n(y- Y)2] = ~52

Hence
52
E(5 2 ) =-~n (N - l)-(N-n)]=5 2 (2.19)
(n - l)~

CoroUary. Unbiased estimates of the variances of y and Y = Ny are

S2(N-n) 5 2
V(Y) = 5/=- - - =-(1-fJ (2.20)
n N n
2
V(Y) = 5y = -n-
N
2 2
5 (N -
N n) =N-n-(
s l - f)
2 2
(2.21)
 SIMPLE RANDOM SAMPLING 27
For the standard errors we take
Ns
Sj1= ~, S~. =~ (2.22)

These estimates are slightly biased : (or most applications the bias is unimportant.
The reader should note the symbols employed for true and estimated variances
of the estimates. Thus, for y, we write
True variance: V(y) = u/
Estimated variance: v(y) = s/
1.8 CONFIDENCE LIMITS
It is usually assumed that the estimates y and Y.are normally distributed about
the corresponding population values. The reasons for this assumption and its
limitations are considered in section 2.15 . If the assumption holds, lower and
uppyr confidence limits for the population mean and total are.as follows :
~ean : .

(2.23)

Total :

(2.24)

11le symbol I is the value of the normal deviate corresponding to the desired
confidence probability. The most common values are
Confidence probability (%) 50 80 90 95 99
0.67 1.28 1.64 1.96 2.58
If the sample size is less than 50, the percentage points may be taken from
Student's I table with (n - 1) degrees of freedom , these being the degrees of
freedom in the estimated variance S2 . The t distribution holds exactly only if the
observations Yi are themselves normally distributed and N is infinite. ~oderate
departures from normality do not affect it greatly. For small samples with very
skew distributions, special methods are needed.
Exilmple. Signatures to a petition were collected on 676 sheets. Each sheet had enough
space for 42 signatures, but on many sheets a smaller number of signatures haH been
collected. The numbers of signatures per sheet were counted on a random sample of 50
sheets (about a 7% sample), with the results shown in Table 2.2 .
Estimate the total number of signatures to the petition and the 80% confidence limits.
28 SAMPUNG TECHNIQUES

The sampling unit is a sh,et, and the observations ,. are the numbers of signatures per
sheet. Since about half the sheets had the maximum number of signatures, 42, the data are
presented as a frequency distribution. Note that the original distribution appears to be far
from normal, the greatest frequency being at the upper end. Nevertheless, there is reason to
believe from experience that the means of samples of 50 are approximately normally
distributed.
We find
n =1..,. = 50, y = r.f,y, = 1471, 1.. f,y/ = 54,497
Hence the estimated total number of signatures is
Y=Ny = (676)(1471) 19,888
50
For the sample variance $ 2 we have

$2 ~
== n 1 a:f,<Y, - y)2] =n ~ J
r."y/- ("[. l;,YJ
'" "!"[54 497 - (1471)2] = 2290
49' 50 .
From (2.22) the 80% confidence limits are

INs r-. V· 28)(676)(15 .13)v'1 - 0.0740

19,888± r"l-f=19,888± r=
"n "50
This gives 18,107 and 21 ,669 for the 80% limits. A complete count showed 2i,045
signatures.
TABLE 2.2
REsULTS FOR A SAMPLE OF 50 PETfl10N SHEETS "I,
FREQUENCY
= NUMBER OF SIGNATURES: ,=
y, 42 41 36 32 29 27 23 19 16 15
f, 23 4 1 1 1 2 1 1 2 2

y, 14 11 10 9 7 6 5 4 3 Total
f, 1 1 1 1 1 3 2 1 1 50

2.9 AN ALTERNATIVE MEmOD OF PROOF

Cornfield (1944) suggested a method of proving the principal results for simple
raodom sampling without replacement that enables us to use standard results
from infinite population theory. Let 4, be a random variate that takes the value 1 if
the ith unit is in the sample and the value 0 otherwise. The sample mean j may be
written

(2.25)
 SlMPU! RANDOM SAMPUNG 29
where the sum extends over all N units in the pop.ulation. In this expression the a,
are random variables and the y, are a set of fixed numbers.
Oearly
n
Pr (a, = 0) = 1--
N
Thus aj is distributed as a binomial variate in a single trial, with P = nl N. Hence
n
E(aj)=P= N' (2.26)

To find V(y) we need also the covariance of al and a,. The product aja, is 1 if the
ith and jth unit are both in the sample and is zero otherwise. The probability that
two specific units are both in the sample is easily found to be n(n -1)1N(N -1).
Hence
Cov (a,a,) =E(ala,) - E(a,)E(aJ)
= n(n -1)
N(N-1)
(.!!.)2 =
N
n (
N(N -l) 1- N
n) (2.27)

Applying this approach to find V(y), we have, from (2.25),

(2.28)

(2.29)

using (2.26) and (2.27). Completing the square on the cross-product term gives

V(y) = ~1(: 1 r.. y/ - N ~ 1 y2) (2.30)

= 1 - f r.. ( 1- Y')2 = (1- f)S2 (2.31)

n(N-1) Y n
The method gives easy proofs of theorems 2 .3 and 2.4. It may be used to find
higber moments of the distribution of 9, although for this purpose a methOd given
by Tukey (1950), with further development by Wishart (1952), is more powerful.

2.10 RANDOM SAMPLING WI'I1I REPLACEMENT

A similar approach applies when sampling is with replacement. In this event the
ith unit may appear 0, 1,2, . . . , n times in the sample. Let t, be the number of
times that the ith unit appears in the sample. Then
1 N
Y=- L tlYI (2.32)
n I-I
30 SAMPLING TECHNIQUES

Since the probability that the 'ith unit is drawn is 1/N at each draw, the variate
distributed as a binomial number of successes out of n trials with p = 1/N.
tl
Hence

(2.33)

Jointly, the variates tl follow a multinomial distribution. For this ,

(2.34)

Using (2.32), (2.33), and (2.34), we have, for sampling with replacement,

_ 1 [N l n(N - 1) N n]
V(Y) =2 LYI N2 2 LYIYiN2 (2.35)
n 1- 1 1<1

1 N _ 2 (j2 N - 1 52
=- L (YI- Y) = - : : ; - - (2.36)
nN I _ I n N n

Consequently, V(y) in sampling without replacement is only (N - n)j(N -1)

times its value in samplil1g with replacement. If instead of y the mean Yd of the
different or distinct units in the sample is used as an estimate when sampling is with
replacement, Murthy (1967) has shown that the leading term in the average
variance of Yd is (1- 1/2)5 2 / n, following work by Basu (1958) and Des Raj and
Khamis (1 958). In some applications the cost of measuring the distinct units in the
sample may be predominating, so that the cost of the sample is proportional to the
number of distinct units. In this situation, Seth and J. N. K. Rao (1964) showed
that for given average cost, V(y) in sampling without replacement is less than
V(Yd) in sampling with replacement. They also prove the more general result that
if Yd' = I(p )Yd/ Ef(p), where p is the number of distinct units in the sample andl(p)
is a function of P, then V(y) < V(Yd') if 5 2< Ny2, a condition satisfied by nearly all
populations encountered in sample s\Jrveys.

2.11 ESTIMATION OF A RATIO

Frequently the quantity that is to be estimated from a simple random sample
is the ratio of two variables both of which vary from unit to unit. In a household
survey examples are the average number of suits of clothes per adult male, the
average expenditure 00 cosmetics per adult female, and the average oumber of
hours per week spent watching television per child aged 10 to 15. In order to
estimate the first of these items, we would record for the ith household (i =
1, 2, ... , n) the number of adult males XI who live there and the total number of
 SIMPLE RANDOM SAMPLING 31
sui ts Yi that they possess. The population parameter to be estimated is the ratio
N

R ,). total number of suits

rYi
1
(2.37)
, total number of adult males N
LXI
1

The corresponding sample estimate is

A
tYI Y-
1
R=-=- (2.38)
" i
IXi
1

Examples of this kind occur frequently when the sampling unit (the household)
comprises a group or cluster of elements (adult males) and Our interest is in the
population mean per element. Ratios also appear in many other applications, for
example, the ratio of loans for building purposes to total loans in a bank or the
ratio of acre~ of wheat to total acres on a farm .
The sampling distribution of R is more complicated than that of 9 because both
the numerator y and the denominator i vary from sample to sample. In small
samples the distribution of R is skew and R is usually a slightly biased estimate of
R. In large samples the distribution of R tends to normality and the bias becomes
ncgiigi ble. The folla.wing approximate result will serve for most purposes: the
distribution of'R is studied in more detail in Chapter 6.
Theorem 2.5. If variates y" Xi are measured on each unit of a simple random
sample of size n, assumed large, the MSE and variance of R = y/i are each
approximately
N
L (Yi - RX,) 2
MSE(R ) == V(R) == 1-::: f ,,-i ~_ _
- ...:....I (2.39)
~ nX2 N-l
where R = Y/ X is the ratio of the population means and f = n/ N.
Proof.
A Y y-Ri
R-R= -- R = - - (2.40)
:i :i
If n is large, i should not differ greatly from X. In order to avoid having to work
out the distribution of the ratio of two random variables (9 - R:i) and i. we replace
x by X in the denominator of (2.40) as an approximatiofl. This gives
R - R == i..:::.!!.
X
(2.41)
32 SAMPUNG 'I'ECHMlQUES

Now average over all simple random samples of size n.

.E(R _ R) == E(y - R£) Y-RX

X X
o (2.42)

since R = Y/ X. This. shows that to the order of approximation used here R is an

unbiased estimate of R.
From (2.41) we also obtain the result

(2.43)

The quantity y -JU is the sample mean of the variate d j = Yi - Rx j , whose

population mean 15 = Y- RX = O. Hence we can find VCR) by applying theorem
2.2 for the variance of the mean of a simple random sample to the variate d j and
dividing by X2 • This gives
A 1• _ _ 2 1 Si(
V(R) =-V2E(y-Ri) =""=2 - I-f) (2.44)
A X n
N -2
1- f j~J (dj - D)
= ,&2 (N-l)
(2.45)

This completes the proof.

The way in which theorem 2.5 was proved is worth noting-It was shown that the
formula in theorem 2.2 for the variance of the sample mean y gives the formula for
the approximate variance of the ratio j/£, if the variate Yj is replaced by the
x..
variate (Yj - Rx,)/ The same result, or its natural extension, holds also in more
complex sampling situations and is lIsed frequently later in this book.
As a sample estimate of

N-l
it is customary to take

n-l
This estim.ate can be shown to have a bias of order 1/ n.
For the estimated standard error of R, this gives

(2.46)

l
 SIMPLE RANDOM SAMPLING 33
If X is not known, the sample estimate i is s.ubstituted in the denomina!or.
One way to compute s(R) is to express it as

(2.47)

Example. Table 2.3 shows the number of persons (XI). the weekly family income (x,).
and the weekly expenditure on food (y) in a simple random sample of 33 low-income
fa milies. Since the sample is small, the data are intended only to illustrate the calculations.
Estimate from the sample (a) the mean weekly expenditure on food per family. (b) the
mean weekly expenditure on food per person, and (c) the percentage of the income that is
spent on food . Compute the standard errors of these estimates.
Weekly Expenditure on Food per Family. This is the ordinary sample mean
907.2
y =33= S27 .49
By theorem 2.2 (ignoring the fpc). its standard error is

Sv =2_JI
.r;,
(YI _y)2
n- 1
1
,)n (n -1)
Jr. y,'_CIy;)2
- n
1
=Ji33)(32)./28,224 - (907 .2)2/33 = S1.76
(33)(32)
(The uncorrected sum of squares 28,224 is given underneath Table 2.3.)
Weekly Expenditure on Food per Person. Since the size of family varies. the estimate is a
ratio of two variables,
R r. y 907.2
1=r.X = 123 =S7.38per person
I

The sums of squares and products needed to compute S(R) by (2 .47) are found under
Table 2.3. We need in addition
2RI = 14.7512, R12= 54.3996, i,=3.7273
Extra decimals are carried in RI> 2R" RI ' to preserve accuracy.
Hence, from (2.47),

s(R I) 1 ),28,224) - (14.7512)(3595.5) + (54.3996)(533)

m(3.7273) 32
= SO.534
Percentage of Income Spent on Food. This again is a ratio of two variables
R = 10oJ.-t = (100)(907.2) = 37 90;'
2 r. X2 2394 . 0

By (2.47) the reader may verify that the standard error is 2.38%.
34 SAMPLING TECHNIQUES

TABLE 2.3
SIZE, WEEKLY INCOME, AND FOOD COST OF 33 FAMILIES
Food Food
Family Size Income Cost Family Size Income Cost
Number xI X2 y Number Xl X2 Y

I 2 62 14.3 18 4 83 36.0
2 3 62 20.8 19 2 85 20.6
3 3 87 22.7 20 4 73 27.7 '
4 5 65 30.5 21 2 66 25.9
5 4 58 41.2 22 5 58 23.3
6 7 92 28.2 23 3 77 39.8
7 2 88 24.2 24 4 69 16.8
8 4 79 30.0 25 7 65 37.8
9 2 83 24.2 26 3 77 34.8
10 ~ 62 44.4 27 3 69 28.7
II 3 63 13.4 28 6 95 63 .0
12 6 62 19.8 29 2 77 19.5
13 4 60 29.4 30 ~ 69 21.6
14 4 75 27.1 31 6 69 18.2
15 2 90 22.2 32 4 67 20. 1
16 5 75 37.7 33 2 63 20.7
17 3 69 22.6
Total 123 2394 907.2

Lx ,2 = 533. LX/= 177,254, L y2= 28,224

I x, )' = 3595.5, L X 2Y = 66,678

1.12 ESTIMATES OF MEANS OVER SUBPOPULATlONS

In many surveys, estimates are made for each of a number of classes into which
the population is subdivided. In a household survey separate estimates might be
wanted for families with 0, 1, 2, .. , children, for owners and renters, or for
families in different occupation groups. The term domains of study has been given
to these subpopulations by the U.N . Subcommission on Sampling (1950) .
In the simplest situation each unit in the population falls into one of the
domains. Let the jth domain contain N j units, and let n, be the number of units in a
simple random sample of size n that bappen to fall in this domain. If Yjk
(k = 1, 2, ... , nj) are the measurements on these units, the population mean Yi
fo r the jth domain i estimated by

(2.48)
 SIMPLE RANDOM SAMPLING 35
At first sight YI seems to be a ratio estimate as in section 2.11. Although n is
fixed. nj will vary from one sample of size n to another. The complication of a ratio
estimate can be avoided by considering the distribution of Yi over samples in
which both nand nj are fixed . We assume nj > O.
n,
In the totality of samples with given nand the probability that any specific set
of ni units from the N j units in domain j is drawn is •
N- N;Cn - n, 1
N-N,Cn - n, • N,Cn, N,Cn,
Since each specific set of n, units from domain j can appear with all selections of
(n -nj) units from the (N - nj) that are not in domainj, the numerator above is the
number of samples containing a specified set of nt, and the denominator is the total
number of samples. It follows that theorems 2.1,2.2, and 2.4 apply to the Yjk if we
put n; for n and IV, for N.
From theorem 2.1: Y; is an unbiased estimate of Y; (2.49)
s·
From theorem 2.2: the standard error of y, i s~ (2.50)
v n,
where
5,2 =
"'I
~
(Yjk - 9y
""-'-_--'-'- (2 .51)
k~J IV,·-l
From theorem 2.4: An estimate of the standard error of Yi is

(2.52)

where
"I ( - )2
S2= ~ Y,k - Yj (2.53)
, k- I n;-l
If the value of Nj is not known, the quantity n/ N may be used in place of ni N;
When computing the fpc. (With simple random sampling, nt/ N j is an unbiased
estimate of n/ N. )

2.13 ESTIMATES OF TOTALS OVER SUBPOPULATIONS

In a firm 's list of accounts receivable, in which some accounts have been paid
and some not, we might wish to estimate by a sample the total dollar amount of
un paid bills. If ~ (the number of unpaid bills in the population) is known, there is
no problem. The sample .estimate is NjYj and its conditional standard error is IV,
time.$ expression (2.50).
36 SAMPUNG TECHNIQUES

Alternatively, if the total amount receivable in the list is known, a ratio estimate
can be used. The sample gives an estimate of the ratio (total amount of unpaid
bills)/(total amount of all biJIs). This is multiplied by the known total amount
receivable in the list.
If neither N, nor the total receivables is known, these estimates cannot be made.
Instead, we multiply the sample total of the y 's over units falling in the jth domain
by the raising factor N/n . This gives the estimate

(2.54)

We will show that Yi is unbiased and obtain its standard error over repeated
samples of size n . The device of keeping nj fixed as well as n does not help in this
problem.
In presenting the proof we revert to the original notation, in which Yi is the
measurement on the ith unit in the population. Define fo r every unit in the
population a new variate y;" where

,_{y, if the unit is in the jth domain,

Yi - .
o otherwIse
The population total of the y/ is
N
L Y;' = L Y, => Y, (2.55)
;- 1 jthdom

In a simple random sample of size n, Yi' = Yi for each of the n, units that lie in the
jth domain; Y;' = 0 for each of the remaining n - nj units. If Y' is the ordinary
sample mean of the y;" the quantity

(2.56)

This result shows that the estimate Yi as defined in equation (2.54) is N tim~s
the sample mean of the Yi'.
In repeated samples of size n we can clearly apply theorems 2.1, 2.2, and 2.4 to
the variates y/. These show that Yi is an unbiased estimate of Yj with standard
error
NS'
u(Yi) = J;, h - (n/ N) (2.57)

where S' is the population standard deviation of the YI'. In order to compute S', we
regard the populatior as consisting of the M values Yt that are in the jth domain
 SIMPLE RANDOM SAMPUNG 37
and of N - Nt zero values. Thus
~,2 _
" _--
1 (~ 2
l.. YI--
Yj2) (2.58)
N-l jthdom N
From theorem 2.4 a sample estimate of the standard error of Y is
j

Ns'
s(Yj)=-Jl-(n/N) (2.5!1)
.rn
In computing s', any unit not in the jth domain is given a zero value. Some students '
seem to have a psychological objection to doing this, but the method is sound.
The methods of this and the preceding section also apply to surveys in which the
frame used contains units that do not belong to the population as it has been
defined. An example illustrates this application.

&ample. From a list of 2422 minor household expenditures a simple random sample
of 180 items was drawn in order to estimate the total spent for operation of the household.
Certain types of expenditure (on clothing and car upkeep) were not considered relevant. Of
the 180 sample items, 152 were relevant. The sum and uncorrected sum of squares of the
re levant amounts (in dollars) were as follows .
L y,'= 343.5, Ly,'2= 1491 .38
Estimate the total expenditure for household operation and give the standard error of the
estimate.
y.=!:!. £y,' =(2422)(343.5) $4622
, n '-I 180
From (2.59)

In computing s' we regard our sample of 180 items as having 28 zeros. Hence

S
'2 1 ['t".2
=(179) ... y, -~
(L y,')2]

1 [ (343.5)2]
= - - 1491.38-- - = 4670
(179) 180 '
Finally,

Sy, = (2422)
J4.670( 180) 7
180 1- 2422 = $3 5

The estimate is not precise, its coefficient of variation 375/4622 being about 8% .
.
In this example expenditures on car upkeep and clothing were excluded as not
relevant and therefore were scored as zeros in the sample. In some applications it
38 SAMPLING TECHNIQUES

is known in advance that certain units in the population contribute nothing to the
total that is being estimated. For instance, in a survey of stores to estimate total
sales of luggage, some stores do not handle luggage; certain area sampling units
for farm studies contain no farms . Sometimes it is possible, by expenditure of
effort, to identify and count the units that e<;>ntribute nothing, so that in our
notation (N - N,), hence N j , is known.
Consequently it is worth examining by how much V(~) is reduced when N j is
known. If N , is not known, (2.57) gives
2
N S 12 (
A

V(Y;.)=-n- 1- N
n)
If Yt and S, are the mean and standard deviation in the domain of interest (i.e.,
among the nonzero units) the reader may verify that

(N - l )S,2= (~-1)S7+~ Y}( 1-~) (2.60)

Since terms in 1/ N/ and 1/ N are nearly always negligible,

S'2 = PjS/ + PjO, Y/ (2.61)
where Pi = N;/ Nand 0, = 1-~. This gives

V(Y/) A • N
2

= -;;-(p(S,2 -2 (
+PiO,y;) 1- n) (2 .62)
N
If nonzero units are ide ntified, we draw a sample of size n, from them. The
estimate of the domain total is N;y, with variance

N;2
V(Ny ) =-S 2(1--ni)=N-2p .2S 2(I - -ni) (2.63)
I I n, I N, ni I I N;
The comparable variances are (2.62) and (2. 63), In (2.62) the average number of
nonzero units in the sample of size n is nP,.If we take nl = nP, in (2. 63), so that the
number of nonzeros to be measured is about the same wirh both methods. (2 .63)
becomes
_ N
2

V(N(y,) = -;;- PtSJ 1- N

2( 11) (2.64)

The ratio of the variances (2.64) to (2.62) is

V(N, known)
2
S/ 2=-2--
c/ (2.65)
V(Nj not known) SI + 0 /Y, C, + 0 /
where C, = S;/ ,Y/ is the coefficient of variation among the nonzeros. As might be
expected, the redttction in variance due to a knowledge of N;. is greater when the
proportion of zero units is large and when Yj varies relatively little among the
nonzero units. For further study of this problem, see Jessen and Houseman
(1944). to
 SIMPLE RANDOM SAMPLING 39
2.14 COMPARISONS BElWEEN DOMAIN MEANS
Let Yi' Yk be the sample means in the jth and kth of a set of domains into which
the units in a simple random sample are classified. The variance of t}leir difference
is
(2.66)
This formula applies also to the difference between two ratios RI and Rk .
One point should be noted . It is seldom of scientific interest to ask whether
91 = Yk , because these means would not be exactly equal in a finite population,
except by a rare chance, even if the data in both domains were drawn at random
fro m the same infinite population. Instead, we test the null hypothesis that the two
domai ns were drawn from infinite populations having the same mean. Conse-
quently we omit the fpc when computing V(y,) and V (Yk), using the formula
2
V(y, _ Yk) = S/ + Sk (2.67)
n, nk
A formula similar to (2 .67) is obtained for tests of significance if one frames the
question: Could the samples from the two domains have been drawn at random
from the same finite population?
Under this null hypothesis it may be proved (see exercise 2.16) that
_ _ 2(1-+-1)
V(Y'-Yk)=S'k
n j nk
where Slk is the variance of the finite population consisting of the combined
domair.s.

2.15 VALIDITY OF THE NORMAL APPROXIMATION

Confidence that the normal approximation is adequate in most practical
situations comes from a variety of sources. In the theory of probability much study
has been made of the distribution of means of random samples. It has been proved
that for any population that has a finite standard deviation the distribution of the
sample mean tends to normality as n increases (see, e.g., Feller, 1957). This work
relates to infinite populations.
For sampling without replacement from finite populations, Hajek (1960) has
given necessary and sufficient conditions under which the distribution of the
sample mean tends to normality, following work by Erdos and Renyi (1959) and
Madow (1948). Hajek assumes a sequence of values n", Nv tending to infinity in
such a way that (Nv - nv) also tends to infinity. The measurements in the Jlth
population are denoted by y.;, (i = 1, 2, .. . , N v). For thi population, let S'" be the
set of units in the population for which

Iyv;- Yvl > ,.,Jn.(1- tV)Sv

40 SAMPLING TECHNIQUES

wher.e Y", S",f~ are the population mean, s.d., and fpc, and T is a number >0. Then
the Lindeberg-type condition
,
I(Yvl- Y~f
Ii ::!s,.,~ _ _~ o
•. ~ (Nv - l)S/

is necessary and sufficient to ensure that y~ tends to normality with the mean and
variance given in theorems 2.1 and 2.2.
This imposing body of knowledge leaves something to be desired . It is not easy
to answer the direct question: "For this population, how large must n be so that
the normal approximation is accurate enough?" Non-normal distributions vary
greatly both in the nature and in the degree of their departure from normality. The
distributions of many types of economic enterprise (stores, chicken farms, towns)
exhibit a marked positive skewness, with a few large units and many small units.
The same kind of skewness is displayed by some biological populations (e.g., the
number of rats or flies per city block).

140

130 t"-

120

110 '-

100

90

~80

*
~ 70
.:: 60

40
~
I
30

20
...__
10

00
t:
100 200
l
I
300 400 500 600
r--1
700
City size (thousands)
800
-
900 1000 1100

Fla. 2.1. Frequency distribution of sizes of 196 United States Cities in 1920.
 SIMPLE RANDOM SAMPUNG 41
As an illustration of a positively skewed distribution, Fig. 2.1 shows the
frequency distribution of the numbers of inhabitants in 196 large United States
cities in 1920. (The four largest cities, New York, Chicago, Philadelphia, and
Detroit, were omitted. Their inclusion would extend the horizontal scale to more
than five times the length shown and WOUld, of course, greatly accentuate the
skewness.) Figure 2.2 shows the frequency distribution of the total number of

50

40
r--
,
r-- .-..

.r-:- n
rL
,~.

_j
4 5 6 7 8 9
Millions

Fig.2.2 Frequency distribution of totals of 200 simple random samples with " = 49.

inhabitants in each of 200 simple random samples, with n = 49, drawn from this
population. The distribution of the sample totals, and likewise of the means, is
much more similar to a normal curve but still displays some positive skewness.
From statistical theory and from the results of sampling experiments on skewed
populations, some statements can be made about what usually happens to
confidence probabilities when we sample from positively skew populations, as
follows :
1. The frequency with which the assertion
y - 1.96sj1 < 9 <y + 1.96sj1
is wrong is usually higher than 5% .
2. The frequency with which
9 > y+1.96s9
is greater than 2.5%.
3. The frequency with which

is less than 2.5% .

As an illustration, consider a variate y that is essentially binomially distributed,

so that the exa~ distribution of y can be read from the binomial tables. The variate
42 SAMPLING TECHNlQUES

y takes only two values-the value h with probability P and the value 0 with
probability Q. The population mean is Y = Ph. A simple random sample of size n
shows a units that have the value hand n - a units that have the valu.e O. For the
sample,
_ ah
y =-
n
i a 2h 2
(n - 1)s =LY 2 - -2 2
ny =ah - - -
n

Hence 95% normal confidence limits for Yare estimated as

(2.68)

Let n = 400, P = 0.1. Then Y"" O.lh . By trial we find that if a = 29 in expression
(2.68) the upper confidence limit is 39. 18h/400 = 0.098h, whereas a = 30 gives
40.34h/400 = O. lOlh. Hence any value of a $ 29 gives an upper confidence limit
that is too low . Similarly we find that if a 2:: 54 the lower limit is too high .
The variate a follows the binomial distribution with n == 400, P = 0.1. The
tables (Harvard Computation Laboratory, 1955) show that

Pr (stated upper limit too low) = Pr (a $ 29) .= 0.0357

Pr (stated lower limit too high = Pr (a 2:: 54) = 0.0217

Pr (confidence statement wrong) = 0.0574

The total probability of being wrong is not far from 0.05. In more than 60% of
the wrong statements, the true mean is higher than the stated upper limit.
There is no safe general rule as to how large n must be for use of the normal
approximation in computing confidence limits. For populations in which the
principal deviation from normality consists of marked positive skewness, a crude
rule that I have occasionally found useful is
n > 250\ 2 (2.69)
where G 1 IS Fisher'S measure of skewness (Fisher, 1932).
- E(y; - Yf _ 1 ~ ( y)- 3
(2.70)
O1- 3 - - - 3 I.. Yi-
U Nu i - I
This rule is designed so that a 95% confidence probability statement will be
wrong not more than 6% of the time. It is derived mathematically by assuming
 SIMPLE RANDOM SAMPLING 43
that any disturbance due to moments of the distribution of ji higher than the third
is negligible. The rule attempts to control only the total frequency of wrong
statements, ignoring the direction of the error of estimate.
By calculating Gj, or an estimate, for a specific population, we can obtain a
rough idea of the sample size needed for application of the normal approximation
to compute confidence limits. The result should be checked by sampling experi-
ments whenever possible.

TABLE 2.4
FREQUENCY DISTRIBUTION OF AcRES IN · CROPS ON 556 FARMS
Class Coded
Intervals Scale Frequency jiYi fly? !iyl
(acres) Yi j;
0-29 -0.9 47 -42.3 38.1 -34.3
30-63 0 143 0 0 0
64;-97 1 154 154 154 154
98:_131 2 82 164 328 656
132- 165 3 62 186 558 1,674
166-199 4 33 132 528 2,112
200-233 5 13 65 325 1,625
234-267 6 6 36 216 1,296
268- 301 7 4 28 196 1,372
302- 335 8 6 48 384 3,072
336-369 9 2 18 162 1,458
370-403 10 0 0 0 0
404-437 II 2 22 241 2,662
438-471 12 0 0 0 0
472- 505 13 2 26 338 4,394

Totals 556 836.7 3,469.1 20,440.7

E (y,) = Y =836. 7 = 1.50486

556
3469.1
E(y,l) =~ = 6.23939

20,440.7
E(y/ ) == == 36.76385
556
u 2 = E(y/ ) - ji"2 = 3.97479
K3 =E(y, - 9)3=E(y/ ) - 3E(y/ ) Y + 2 y l

=15.411
G ",,~_lS.411 _19
I r 7.925 .

111111111111111
10486
44 SAMPUNG TECHNIQUES

Example. The data in Table 2.4 show the numbers of acres devoted to crops on 556
farms in Seneca County, New York. The data come from a series of studies by West (1951),
who drew repeated samples of size 100 from this population and examined the frequency
distributions of y, s, and Student's I for several items of interest in farm management
surveys.
The computation of 0, is shown under the table. The computations are made on a coded
scale, and, since 0, is a pure number, there is no need to return to the original scale. Note
that the first class-interval was slightly different from the others.
Since 0, = 1.9, we take as a suggested minimum n
n"'(25)(1.W=90
For samples of size 100, West found with this item (acres in crops) that neither the
distribution of ji nor that of Student's I differed significantly from the corresponding
theoretical normal distributions.

Good sampling practice tends to make the normal approximation more valid.
Failure of the normal approximation occurs mostly when the population contains
some extreme individuals who dominate the sample average when they are
present. However, these extremes also have a much more serious effect of
increasing the variance of the sample and decreasing the precision. Consequently,
it is wise to segregate them and make separate plans for coping with them, perhaps
by taking a complete enumeration of them if they are not numerous. This removal
of the extremes from the main body of the population reduces the skewness and
improves the normal approximation. This technique is an example of stratified
sampling, which is discussed in Chapter 5.

2.16 LINEAR ESTIMATORS OF THE POPULATION MEAN

Under simple random sampling, is the sample mean y the best estimator of Y?
This question has naturally attracted a good deal of work. The answer depends on
the set of competitors to y that are allowed and on the definition of "best." For
sampling, the units in the population are usually numbered in some way from 1 to
N These numbers that identify the units are often called the labels attached to the
units.
Early results proved by Horvitz and Thompson (1952) for linear estimators in
simple random sampling are as follows. If any YI always receives the same weight
w, whenever unit labeled i is drawn, the sample mean y is the only unbiased
n
estimator of Y of the form ~ W'Yi. Since every unit appears in a fraction nlN of
simple random samples, E(f
WIY') = n(f
WIYI)/ N-
Yonly if every WI = If,lin.
alternatively, the weight depends only on the order in which the unit is drawn into
the sample, then y has minimum variance among unbiased linear estimators of the
form f WdYCd)t where YCd) is the y-value on the unit that turns up at the dth draw.
d
 SIMPLE RANDOM SAMPUNG 45
n
A wider class of competitors is the set L WIsYi, where Wi, may depend on the
other units that fall into the sample as well as on i. Godambe (1955) showed that in
this class no unbiased estimate of Y exists with minimum variance for all
populations.
Further properties of y have been developed by Hartley and J. N. K. Rao (1968,
1969), Royall (1968), and C. R. Rao (1971). In any finite popuJation there will be
at most Ts N distinct values of the Yi. In the sample, let there be n, values equal to
y" where L nr = n. Hartley and Rao (1968) show that y has minimum variance
among unbiased estimators of Y that are functions only of the nr and Yr' For
random sampling with replacement, they show that the mean of the distinct values
in the sample is tt.e maximum likelihood estimator of Y, although it does not have
minimum variance in all populations.
C. R. Rao (1971), following work by Kempthorne (1969), considered unbiased
A

estimators Y = L WisYi discussed by Godambe. In order to represent the case in

which the labels l supply no information about the values YI, he calculated the
average of V( Y) over all N! permutations of the labels attached to the values, and
showed that over these permutations, y has minimum average variance. Royall
(1970b) has given a more general result.
Godambe's (1955) work has stimulated numerous investigations on sampling
design and estimation, including topics such as criteria by which to judge
estimators, the role of maximum likelihood, the use of auxiliary information that
the labels may carry about the Yi, Bayesian estimators, and methods of estimation
when assumptions can be made about the frequency distribution of the Yf. The
influence of this work on sampling practice has been limited thus far, but should
steadily increase. Some reference to it will be made from time to time. For
reviews, see J. N. K. Rao* (1975a) and Smith (1976).

EXERCISES
2.1 In a population with N=6 the values of y, are 8, 3, I, 11,4, and 7. Calculate"the
sample mean_j for all possible simple random samples of size 2. Verify that y is an unbiased
estimate of Y and that its variance is as given in theorem 2.2.
2.2 For the same population, calculate S2 for all simple random samples of size 3 and
verify that E(S2) = S 2.
2.3 If random samples of size 2 are drawn with replacement from this population, show
by finding all possible samples that V(y) satifies the equation
V(Y) = = S2 (N -1)
(7'2

n 11 N
2.4 A simple random sample of 30 households was draw'n from a city area containing
14,848 households. The numbers of persons per household in the sample were as follows.
5, 6.3,3,2, 3,3,3, 4, 4, 3,2,7,4, 3,5, 4, 4,3,3, 4,3,3,1,2, 4,3, 4, 2,4

.. Henceforth in this book. the surname Rao will refer to 1. N. K. Rao unless otherwise noted.
 46 SAMPLING TECHNIQUES

Estimate the total number of people in the area and compute the probability that this
estimate is within ± 10% of the true value.
2.5 In a study of the possible use of sampling to cut down the work in taking inventory
in a stock room, a count is made of the value of the articles on each of 36 shelves in the
room. The values to the nearest dollar are as follows.
29,38,42,44,45,47,51,53,53,54,56,56,56, 58,58,59,60,60,
60,60,61,61,61,62,64,65,65,67,67,68,69,71, 74, 77,82,85 .
The estimate of total value made from a sample is to be correct within S200, apart from a
1 in 20 chance. An advisor suggests that a simple random sample of 12 shelves will meet the
requirements. Do you agree?
Ly=2138, Ly2=131,682
2.6 After the sample in Table 2.2 (p. 28) was taken, the number of completely filled
sheets (with 42 signatures each) was counted and found to be 326. Use this information to
make an improved estimate of the total number of signatures and find the :;tandard error of
your estimate.
2.7 From a listof 468 sma1l2-yearcolleges a simple random sample of 100 colleges was
drawn. Th~ sample contained 54 public and 46 private colleges. Data for number of
students (y) and number of teachers (x) are shown below.
n L (y) rex)
Public 54 31,281 2,024
Private 46 13,707 1,075
L (y2) L (yx) L (x 2)
Public 29,881,219 1,729,349 111 ,090
Private 6,366,785 431,041 33,119

(a) For each type of college in the population, estimate the ratio (number of
s~dents)/(numbe~ of teachers). (b) Compute the standard errors of your estimates. (c) For
the public colleges; find 90% confidence limits for the student/teacher ratio in the whole
population.
2.8 In the preceding example test at the 5% level whether the student/teacher ratio is
significantly different in the two types of colleges.
2.9 For the public colleges, e3timate th~ total number of teachers (a) given that the
total number of public colleges in the population is 251, (b) without knowing this figure . In
each case compute the standard error of your estimate .
2.10 The table below shows the numbers of inhabitants in each of the 197 United
States cities that had populations over 50,000 in 1940. Calculate the st~ndard error of the
estimated total. number of inhabitant in all 197 cities for the following methods of
sampling: (a) a simple random sample of size 50, (b) a sample that includes the five largest
cities and is a simple random sample of size 45 from the remaining 192 cities (c) a sample
that includes the nine largest cities and is a simple random sample of size 41 from the
remaining cities.
 SIMPLE RANDOM SAMPLING 47
FREQUENCY DISTRIBUTION OF CITY SIZES
Size Class Size C lass Size Class
(1000's) f (1000's} f (1000's) f
50-100 105 550--600 2
100-150 36 600-650 I 1500-1550
150-200 13 650-700 2
200-250 '6 700-750 0 1600-1650
250-300 7 750-800
300-350 8 800-850 1900-1950
350-400 4 850-900 2
400-450 1 900-950 0 3350-3400
450-500 3 950-1000 0
500-550 0 1000-1050 0 7450- 7500

Gaps in the intervals are indicated by .. ..

2.11 Calculate the coefficient of skewness G I for the original population and the
population remaining after removing (a) the five largest cities, (b) the nine largest cities.
2.12 A small survey is to be taken to compare home-owners with renters. In the
population about 75% are owners, 25% are renters. For one item the variance is thought to
be about 15 for both owners and renters. The standard error of the difference between the
two domain means is not to exceed 1. How large a sample is needed (a) if owners and
renters can be identified in advance of drawing the sample, (b) if not? (An approximate
answer will do in (b); an exact discussion requires binomial tables.)
2.13 A simple random sample of size 3 is drawn from a population of size N with
replacement. Show that the probabilities that the sample contains 1,2, and 3 different units
(for example, aaa , aab, abc, respectively) are
1 3(N - 1) p = (N - 1)(N - 2)
P'= N 2 ' P2=~ ' J N2

As an estimate of Y we take y', the unweighted mean over the different units in the
sa mple. Show that the average variance of :Y' is

V(Y') = (2N - ~~ - 1)S2 : (1 - f/2)S 2/3

One way to do this is to show that

, (N - 1 N- 2 N - 3)
V(y ) = S 2 --;::;- P, + 2N P2 + 3N P3

Hence show that V(ji') < V(y) , where ji is the ordinary mean of the n observations
in the sample. The result that V(ji') < V(ji) for any n > 2 was proved by Des Raj and
Kbamis (1958).
2.14 Two dentists A and B make a survey of the state of the teeth of 200 children in a
village. Dr. A selects a simple random sample of 20 children and counts the number of
decayed t~eth for each child, with the following results.
 48 SAMPUNO TECHNIQUES

Number of decayed 0 1 2 3 4 5 6 7 8 9 10
teeth/child
Number of children 8 4 2 2 1 0 0 0 1 1

Dr. B , using the same dental techniques, examines all 200 children, recording merely
those who have no decayed teeth. He finds 60 children with no decayed teeth.
Estimate the total number of decayed teeth in the village children, (a) using A 's results
only, (b) using both A's and B's results. (c) Are the estimates unbiased? (d) Which
estimate do you expect to be more precise?
2.15 A company intends to interview a simple random sample of employees who have
been with it more than 5 years. The company has UOOO to spend, and each interview costs
$ 10. There is no separate list of employees with more than 5 years service, but a list can be
compiled from the files at a cost of $200. The company can either (a) compile the list and
interview a simple random sample drawn from the eligible employees or (b ) draw a simple
random sample of all employees, interviewing only those eligible. The cost of rejecting
those not eligible in the sample is assumed negligible .
Show that for estimating a total over the poP'ulation of eligible employees, plan (a ) gives
a smaller variance than plan (b ) only if Cj < 2~, where Cj is the coefficient of variation of
the item among eligible employees and OJis the proportion of noneligi bles in the company.
Ignore the fpc.
2.16 A simple random sample of size n = '" + n2 with mean y is drawn from a finite
population, and a simple random subsample of size n. is drawn from it with mean y, . Show
that (a) V(YI - Y2) =S2[(I/n, ) +(I/n2)J. where Y2is the mean of the remaining units in"2
the sample, (b ) V(y, - y) = S2[(1/n,) - (l/n)J, (c ) Cov {y, y, - y} = O. Repeated sampling
implies repetition of the drawing of both the sample and the subsample.
2.17 The number of distinct simple random samples of size n is of course N!I n I(N -
n)1. There has been some interest in finding smaller sets of samples of size n that have the
same properties as the set of simple random samples. One set is that of balanced incomplete
block (bib) designs. These I1re samples of " distinct units out of N such that (i) every unit
appears in the same number (r) of samples, (ii) every pair of units appears together in A
samples.
Verify that A = r(n - 1)/(N - 1) and that the number of distinct samples in the set b rN/ n.
Over the set of bib samples, prove in the uS!lal notation that if ji is the mean of a sample, (a )
V(j) =(1 - f)S 2/ n and (b) v(y) = (1-f)L(YI -ji)2/ n (n - l) is an unbiased estimate of
~~ . .
Note. There is no general method for finding the smallest r for which a bib can be
constructed. Sometimes the smallest known r provides N!/ n !(N - n)! samples, bringing us
back to simple random samples. But for N = 91, n '" 10, the smallest bib set has 91 samples
as against over 6 United States trillion SRS. Avadhani and Sukhatme (1973) have shown
how bib designs may be used in attempting to reduce travel costs between sampling units.
2.18 The following is an illustration by Royall (1968) of the fact that in simple random
sampling the sample m~an ji does not have uniformly minimum variance in the class of
estimators of the form L W ..y, considered by Godambe (1955), where the weight W,. may
depend on the other units that faU in the sample. For N = 3, n = 2, consider the estimator
~ A ~
YI2=iy,+ h2; Y 13 =h ,+h3; Y23 =h2+!Y3
~ A
where >';, is the estimatqr for the sample that has units (i, j). Prove Royall 's results that YI}
is unbiased and that V( Y,,) < V(y) if Y3(3Y 2- 3y, - .)'3) > O. The illustration is taken from an
earlier example by Roy and Chakravarti (1960).
 SIMPLE RANDOM SA.MPLING 49
2.19 This exercise is another example of estimators geared to particular features of
populations. After the decision to take a simple random sample had been made, it was
realized that y, would be unusually low and YN would be unusually high . For this situation,
Sarndai (1972) examined the following unbiased estimator of Y.
Ys "" y+c if the sample contains y, but not YN
=y- c if the sample contains YN but not y,
=y for all other amples
where c is a constant. Prove SarndaJ's result that Ys is unbiased with
.:! [S2
2c
V(Ys)=(l-f) -;-(N _ l)(YN -y ,-nc)
]

so that VCY)< V(.y) ifO<c«YN-y,)/n.

2.20 For a population with N = 8 and values y, = 1, 4, 5, S, 6, 6. 8, 13, show that with
n = 4, V(y) == 1.5, while V( Ys ) = 0.214 when c = 1.5 (its best value), and 0.357 when c = 1
or 2.
Given the information in exercise 2.19, an alternative sampling plan is to include both ." 1
and Ys in every sa mple. dra~ing a simple random sample of size 2 from y~, .. .• y" with
mean Y2' The estimate of Y is
Y" = (y, +6Y2+ YH)/8
Sho}" th at Y., is unbiased, with variance 9V(Y2)/16. For this population . show that
V( f l l ) = 0.350. This estimator is an example of stratified sampling (Chapter 5) with three
strata : y,; Y2 ... y,; and YH'
CHAPTER 3

Sampling Proportions
and Percentages

3.1 QUALITATIVE CHARACTERISTICS

Sometimes we wish to estimate the total number, the proportion, or the
percentage of units in the population that possess some characteristic or attribute
or fall into some defined class. Many of the results regularly published from
censuses or surveys are of this form, for example, numbers of unemployed
persons, the percentage of the population that is native-born. The classification
may be introduced directly into the questionnaire, as in questions that are
answered by a simple " yes" or "no." In other cases the original measurements are
more or less continuous, and the classification is introduced in the tabulation of
results. Thus we may record the respondents' ages to the nearest year but publish
the percentage of the population aged 60 and over.
Notation. We suppose that every unit in the population falls into one of the
two classe C and C . The notation is as follows:
Number of units in C in Proportion of units in C in
Population Sample Population Sample
A a P = A/N p = a/n
The sample estimate of P is p, and the sample estimate of A is Np or Na/ n.
In statistical work the binomial distribution is often applied to estimates like a
and p. As will be seen, the correct distribution for finite populations is the
hypergeometric, although the binomial is usually a satisfactory approximation.

3.2 VARIANCES OF THE SAMPLE ESTIMATES

By means of a simple device it is possible to apply the theorems established in
Chapter 2 to this situation. For any unit in the sample or population, define YI as 1
if the unit is in C and as 0 if it is in C. For this population of values Ylo it is clear that
N
Y=LYI=A (3.1)
50
 SAMPLING PROPORTIONS AND PERCENTAGES 51
N

- fYi
A
Y=N=N=P . (3.2)
Also, for the sample,
n

_
IYi
1 a
y=-=-=p (3.3)
n n
Consequently the problem of estimating A and P can be regarded as that of
t'stimating the total and mean of a population in which every YI is either 1 or O. In
order to use the theorems in Chapter 2, we first express 5 2and S 2 in terms of P and
p. Note th:lt
N
Iy?=A =NP, fy ;2= a =np
1 1

Hence
N _ 2 N 2 -2
I
5 2 =1
(YI - Y) I YI - NY
_1_______

N -l N- l
1 2 N
= N_l(NP-NP)= N-I PO (3.4)

where 0 = 1 - P. Similarly

S2 =
f (YI- y)2 = _n_
_1_ _ _ _
pq (3.5)
n- l n -l
Application of theorems 2.1,2.2, and 2.4 to this population gives the following
results for simple random sampling of the units that are being classified.

Theorem 3.1. The sample proportion p = al n is an unbiased estimate of the

population proportion P = AIN .

Theorem 3.2. The variance of p is

V(P) =E(P - p)2= 5

2
(N-n) n
= PO{ti- ) (3.6)
n N n W-1
using (3.4).

CoroDary 1. If p and P are the sample and population percentages, respec-

tively, falling into class C, (3.6) continues to hold for the variance of p.
 52 SAMPLING TECHNIQUES

Corollary 2. The variance of A = Np, the estimated total number of units in

class C, is
V(A) = N 2PQffi - n) (3.7)
n W- l
lbeorem 3.3. An unbiased estimate of the variance of p, derived from the
sample, is
. N-n
v(P) = s/ =(n -l)NfXi (3.8)

Proof. In the corollary of theorem 2.4 it was shown that for a variate Yi an
unbiased estimate of the variance of the sample mean y is
v(y) = S2 (N-n) (3.9)
n N
For proportions, P takes the place of y, and in (3.5) we showed that
2 n
S =--pq (3.10) .
n- 1
Hence
2 N- n
v(P)=Sp = (n-l)NfXi (3.11)

It follows that if N is very lar.ge relative to n, so that the fpc is negligible, an

unbiased estimate of the variance of p is

..J!!L
n -1
The result may appear puzzling to some readers, since the expression pq/ n is
almost invariably used in practice for the estimated variance. The fact is that pq/ n
is not unbiased even with an infinite population.

CoroUary. An unbiased estimate of the variance of A= Np, the estimated

total number of units in class C in the population, is
A) 2 N(N-n)
v (A = SNp = pq (3.12)
n-l

Exll"'ple. From a list of 3042 names and addresses, a simple random sample of 200
names showed on investigation 38 wrong addresses. Estimate the total number of
addresses needing correction in the list and find the standard error of this estimate. We have
N = 3042, n =200, Q =38, p =0.19
 SAMPLING PROPORTIONS AND PERCENTAGES 53
The estimated total number of wrong addresses is
A =< Np =(3042)(0.19) =< 578

S~ =< J[(3042)(2842)(0.19)(0.81)/199] = J6686 = 81.8

Since the sampling ratio is under 7%, the fpc makes little difference. To remove it. replace
th e term N - n by N. If, in addition, we replace n - 1 by n, we have the simpler formula
,
SNp = NJpq/n = (3042)J(0.19)(0.81)/200 = 84.4
This is in fairly close agreement with the previous result, 81.8 .

The preceding formulas for the variance and the estimated variance of p hold
only if the units are classified into Cor C so that p is the ratio of the number of
units in C in the sample to the total number of units in the sample. In many surveys
each unit i composed of a group of elements, and it is the elements that are
classified. A few examples are as follows:

Sampling Unit Elements

Family Members of the family

Restaurant Employees
Crate of eggs Individual eggs
Peach tree Individual peaches

If a simple random sample of units is drawn in order to estimate the proportion P

of elements in the population that belong to class C, the preceding formulas do not
apply. Appropriate methods are given in section 3.12.

3.3 mE EFFECT OF P ON mE STANDARD ERRORS

Equation (3 .6) shows how the variance of the estimated percentage changes
with P, for fixed nand N. If the fpc is ignored, we have'

v(P)=PO
n
The function PO and its square root are shown in Table 3.1. These functions
may be regarded as the variance and standard deviation, respectively, for a sample
of size 1.
The functions have their greatest values when the population is equally divided
between the two classes, and are symmetrical about this point. The standard error
of p changes relatively little when P lies anywhere between 30 and 70%. At the
maximum value of .,[jiQ, 50, a sample size of 100 is needed to reduce tbe standard
54 SAMPLING TECHNIQUES

TABLE 3.1
V ALVI;S m PQ AND v'PQ
P = Population percentage in class C
P 0 10 20 30 40 50 60 70 80 90 100

PQ 0 900 1600 2100 2400 2500 2400 21 00 1600 900 0

v'PQ 0 30 40 46 49 50 49 46 40 30 0

error of the estimate to 5%. To attain a 1% standard error requires a sample size
of 2500.
This approach is not appropriate when interest lies in the total number of units
in the population that are in class C. In this event it is more natural to ask : Is the
estimate likely to be correct to within, say, 7% of the true total'! Thus we tenj to
think of the standard error expressed as a fraction or percentage of the true value,
NP. The fraction is

(3.13)

This quantity is called the coefficient of variation of the estimate. If the fpc is
ignored, the coefficient is J OJ nP. The ratio J OJP, which might be considered the
coefficient of variation for a sample of size 1, is shown in Table 3.2.

TABLE 3.2
VALUES OF Y Qf P FOR DIFFERENT VALVES OF P
P = Population percentage in class C

p 0 0.1 0.5 5 10 20
v' QfP 00 31.6 14.1 9.9 4.4 3.0 2.0

P 30 40 50 60 70 80 90
v' QfP 1.5 1.2 1.0 0.8 0.7 0.5 0.3

For a fixed sample size, the coefficient of variation of the estimated total in class
C decreases stea(llly as the true percentage in C increases. The coefficient is high
when P is less than 5 % . Very large samples are needed for precise estimates of the
total number PQssessing any attribute that is rare in the population. For P = 1 %,
we must have ~ =99 in order to reduce the coefficient of variation of the estimate
 SAMPLING PROPORTIONS AND PERCENTAGES 55
to 0.1 or 10%. This gives a sample size of 9801. Simple random sampling, or any
method of sampling that is adapted for general purposes, is an expensive method
of estimating the total number of units of a scarce type.

3.4 THE BINOMIAL DISTRIBUTION

Since the population is of a particularly simple type, in which the Yi are either 1
or 0, we -:an find the actual frequency distribution of the estimate p and not merely
its mean and variance.
The population contains A units that are in class C and N - A units in C, where
P = A/ N. If the first unit that is drawn happens to be in C, there will remain in the
population A _: 1 units in C and N - A in C. Thus the proportion of units in C,
after the first draw, changes slightly to (A -l)/(N -1). Alternatively, if the first
unit drawn is in C, the proportion in C changes to A/(N -1). In sampling without
replacement, the proportion keeps changing in this way throughout the draw. In
the present section these variations are ignored, that is, P is assumed constant.
This amounts to assuming that A and N ....: A are both large relative to the sample
size n, or that sampling is with replacement.
With this assumption, the process of drawing the sample consists of a series of n
trials, in each of which the probability that the unit drawn is in C is P. This
situation gives rise to the familiar binomial frequency distribution for the number
of units in C in the sample. The probability that the sample contains a units in Cis

( ) n! IJQ" - a (314)
Pra=a!(n_a)t .

From this expression we may tabulate the frequency distribution of a, of

p = a/ n, or of the estimated total Np.
There are three comprehensive sets of tables. All give P by intervals of 0.01.
The ranges for n are as follows .
U.S. Bureau of Standards (1950):
n = 1(1)49, (i.e., goes from 1 to 49 by intervals of 1).
Romig (1952): n = 50(5)100.
Harvard Computation Laboratory (1955) :
n = 1(1)50(2)100(10)200(20)500(50)1000

3.5 THE HYPERGEOMETRIC DISTRIBUTION

The distribution of p can be found without the assumption that the population is
large in relation to the sample. The numbers of units in the two classes C and C in
the population are A and A I , respectively. We will calculate the probability that
56 SAMPUNG TECHNIQUES

the corresponding numbers in the sample are a and a', where

a+a'=n, A+A'=N

In simple random sampling each of the (~ different selections of n units out of

N has an equal chance of being drawn. To find the probability wanted, we count
how many of these samples contain exactly a units from C and a ' from C. The
number of different selections of a units among the ·A that ~re in Cis (;),
whereas the number of different selections of a I among A I is (~, ) . Each selection
of the first type can be combined with anyone of the second to give a different
sample of the required type . The total number of samples of the required type is
therefore

Hence, if a simple random sample of size n is drawn, the probability that it is of

the required type is

Pr(a, a'iA, A') = ( ; ) . (~,')/ tJ (3.1S)

This is the frequency distribution of a or np, from which that of p is immediately

derivable. The distribution is called the hypergeometric distribution.
For computing purposes the hypergeometric probability (3.15) may be written
as follows.
n! .A (A - 1) .. . (A - a + 1)(A ')(A 1) .... (A a + 1)
I - I - I
(3.16)
a!(n - a)! N(N-l) .. . (N-n+l)

Example. A family of eight contains three males and five females. Find the frequency
distribution of the number of males in a simple random sample of size 4. In this case
A = 3; A' =5, N = 8; n=4
From (3.16) the distribution of the number of males, a, is as follows :
a Probability
4! 5.4.3.2 1
o - .- - = -
0!4! 8.7.6.5 14
4! 3.5.4.3 6
1!3!·8.7.6.5 "" 14
4! 3.2.5.4 6
2
2!2!·8.7.6.5 = 14
41 3.2.1.5 1
3 3!1I · 8.7.6.5 = 14
4 Impossible = 0
 SAMPLING PROPORTIONS AND PERCENTAGES 57
The reader may verify that the mean number of males is ~ and the variance is H. The e
results agree with the formulas previously established in section 3.2, which give

E(np) = nP = nA = (4)(3) =~
N 8 2
N-n 354 15
V(np) = nPO N - 1 = 4'8'8'7 = 28

3.6 CONFIDENCE LIMITS

We first discuss the meaning of confidence limits in the case of qualitative
characteristics. In the sample, a out of n fall in class C. Suppose that inferences are
to be made about the number A in the population that fall in class C. For an
upper confidence limit to A , we compute a value Au such that for this value the
probability of getting a or less falling in C in the sample is some small quantity au,
for example, 0.025. Formally, Au satisfies the equation

f
;- 0
Pr(j, n - iiA u, N - Au) = auU (3.17)

where Pr is the probability t~ rm for the hypergeome tric distribution, as defined in

(3.15).
When au is chosen ir advance, (3 .17) requires in general a nonintegral value of
Au to satisfy it, whereas conceptually Au should be a whole number. In practice
we choose A u as the smallest integral value of A such that the left side of (3.17) is
less than or equal to au. Similarly, the lower confidence limit AL is the largest
integral value such that

(3.18)
i-a

Confidence limits for P are then found by taking Pu = Au/ N , A

= AdN.
Numerous methods are available for computing confidence limits.
Exact Methods
Chung and DeLury (1950) present charts of the 90, 95, and 99% limits for P for
N = 500,2500, and 10,000. Values for intermediate population sizes are obtain-
able by interpolation. Ueberman and Owen (1961) give tables of individual and
cumulative terms of the hypergeometric distribution, but N extends only to 100.
The Normal Approximation ...
From (3.8) for the estimated variance of p, one form of the normal approxima-
tion to the confidence limits for P is

(3.19)
58 SAMPLING TECHNIOUES

where f = n/Nand t is the normal deviate corresponding to the confidence

probability. Use of the more familiar term Jpq/n seldom makes an appreciable
difference. The last term on the right is a correction for continuity. This produces
only a slight improvement in the approximation. However, without the correc-
tion, the normal approximation usually gives too narrow a confidence interval.

TABLE 3.3
SMALLEST VALUES OF np fOR liSE OF THE NORMAL
ApPROXIMATION

np = Number Observed n=
p in the Smaller Class Sample Size

O.S IS 30
0.4 20 50
0.3 24 80
0.2 40 200
0.1 60 600
0.05 70 1400
-0. 80 <Xl

.. This means that p is extremely small, so that np follows

the Poisson distribution.

The errOr in the normal approximation depends on all the quantities n, p, N, au,
and aL. The quantity to which the error is most sensitive is np or more specifically
the number observed in the smaller class. Table 3.3 gives working rules for
deciding when the normal approximation (3.19) may be used .
The rules in Table 3.3 are constructed so that with 95% confidence limits the
true frequency with which the limits fail to enclose P is not greater than 5.5% .
Furthermore, the probability that the upper limit is below P is between 2.5 and
3.5%, and the probability that the lower limit exceeds P is between 2.5 and 1.5%.
Example 1. In a simple random sample o{ size 100, from a population of size 500, there
are 37 units in class C. Find the 95% confidence limits for the proportion and for the total
number in class C in the population. In thjs example
n = 100, N=SOO, p=0.37
The example lies in the range in which the normal approximation is recommended. The
estimated standard. error of p is

J(l- /)pq/(n - 1) = J(0.8)(0.37)(0.63)/99 =0.0434

The correction for continuity, 1/2n, equals 0.005. Hence the 95% limits for P are
 SAMPLING PROPORTIONS AND PERCENTAGES 59
estimated as
0.37±(1.96 x 0.0434+0.005) = 0.37±0.090
PL =0.280, Pu = 0.460
The limits as read from the charts by Chung and DeLury are 0.285 and 0.462, respectively.
To find limits for the total number in class C in the population, we mUltiply by N,
obtaining 140 and 230, respectively.

Binomial Approximations
When the normal approximation does not apply, limits for P may be found from
the binomial tables (section 3.4) and adjusted, if necessary, to take account of the
fpc. Table VIm in Fisher and Yates' Statistical Tables (1957) gives binomial
confidence limits for P for any value of n, and is a useful alternative to the ordinary
binomial tables. Example 2 shows how the binomial approximation is computed.

Example 2. For another item in the sample in example 1, nine of the 100 units fall in
class C. From Romig's table for n = 100 the 95% limits for P are found to be 0.041 and
0.165. (The Fisher-Yates tables give 0.042 and 0.164 .) Iff, the sampling fraction, is less
than 5% , limits found in this way are close enough for most purposes. In this example,
f= 0.2 and adjustment is needed .
To apP!y"!'he adjustment, we shorten the interval between p and each limit by the factor
J1 - f = -/0.8 = 0.894. The adjusted limits are as follows :
PL = 0.090 - (0.894)(0.090 - 0.041) = 0.046
Pu = 0.090 + (0.894)(0.165 - 0.090) = 0.157
The limits read from the charts by Chung and DeLury are 0.045 and 0.157, respectively.
Burstein (1975) has produced a variant of this calculation that is slightly more accurate.
Suppose that a umts out of n arc in class C (in this example, a = 9, n = 100). In PL , replace
a/n = 0.090 by (a - 0.5 )/n ;; 0.085. In Pu, replace a/ n by (a + a/ n) / n = 0.0909. Also,
(I - f) is taken as (N - n)/(N - l). Thus, by Burstein's method,

PL = 0.085 - (0.895)(0.085 - 0.041) = 0.046

Pu ;; 0.0909 + (0.895)(0.165 - 0.0909) = 0.157
there being no change in 'the limits in this example.

Example J. In auditing records in which a very low error rate is demanded, the upper
confidence limit for A is primarily of interest . Suppose that 200 of 1000 records are verified
and that the batch of 1000 is accepted if no errors are found . Special tables have been
constructed to give the upper confidence limit for the number of errors in the batch. A good
approximation results from the following relation. The probability that no errors are found
in n when A errors are present in N is, from the hypergeometric distribution,

(N-A)(N-A -1) .. . (N-A -n +1) . (N - A -U)'

N(N -1) . .. (N - n + 1) N- U
62 SAMPLING TECHNIQUES

Example. Consider a population that consists of the five units, b, c. d, e, t, that fall in
three classes.
Class A, Units Denoted By

1 1 b
2 2 c,d
3 2 e,f

With random samples of size 3, we wish to estimate P = A ,/(A, +A 2 ) or, in this case,!.
Thus N =5 and N ' =' 3.
There are 10 possiblt> samples of size 3, all with equal initial probabilities. These are
grouped according to the value of n '.
n'= 1

Conditional
Sample a, a2 P Probability (P - p)

bel 1 0 ] i
_1
~
eel or de! 0 1 0 3 3

If samples are specified by the values of 0" 02, only two types are obtainable : 0, == 1,
02=0 ; a, = 0, a l == 1. Their conditional probabilities, ~ and 5. respectively, agree with the
general expression (3.24). Furthermore,
E(P) =i

cr/ = G)(~)+ (~) (~) = 267 = ~

The estimate p is unbiased. and its variance agrees with the general formula

cr 2 =
p
(N' - n'\PQ = (3 - 1)(!)(~) =~
N'-1)R" 3- 1 3 3 9
For n ' == 2 there Bre six possible sBmples, which give only two sets of values of 0" 0 2.

n' =2
Conditional
Sample a, a2 P Probability (P - P)

bee, be!, Ixh. or bdf 1 1 ! •

)
1
6
-t
ede or cd! 0 2 0 i
The estimate is again unbiased and its variance is
 SAMPLrNG PROPORTIONS AND PERCENTAGES 63
which may be verified from the general formula. Note that the variance is only one fourth of
that obtained when n' = 1. In a conditional approach the variance changes with the
configuration of the sample that was drawn.
For n' = 3, there is only one possible sample, bcd. This gives the correct population
fraction, l. The conditional variance of p is zero, as indicated by the general formula, which
reduces to zero when N ' = n'.

3.10 PROPORTIONS AND TOTALS OVER SUBPOPULAll0NS

If separate estimates are to be made for each of a number of subpopuiations or
domains of study to which the units in the sample are allotted, the results in
sections 3.8 and 3.9 are applicable . The sample data may be presented as follows.

Domain! Domain2 Domain k Total

Oass CCC C C C

Number of units n

Of the n units, (a I + a 1') are found to fall in domain 1 and of these a I fall in class
C The proportion falling in class Coin domain 1 is estimated by PI = at/(al + al').
The frequency distribution and confidence limits for PI were discussed under Case
2 in sections 3.8 and 3.9.
For estimating the total number A J of units in class C in domain 1, there are two
possibilities. If Nh the total number of units in domain 1 in the population, is
known, we may use the conditional estimate
A NlaJ
AI=NIPl= - --, (3.25)
al +a1
Its standard error is computed as

(3.26)

where nl = al +at
If NI is not known, the estimate is
A,_Na)
A 1 - -- (3.27)
n
with estimated standard error

sCAt') =N-h -(n/N)Jpq/(n -1) (3.28)

where P = ad n.
64 SAMPLING TECHNIQUES

3.11 COMPARI~ONS BETWEEN DIFFERENT DOMAINS

Since proportions are estimated independently in different domains, compari-
sons between such proportions are made by standard elementary methods. For
example, to test whether the proportion PI = al/(al +al') differs significantly
from the proportion P2 = a 2/(a2 + a2')' we form the usual 2 x 2 table .
Domain
1 2

C a, a2
C a,' a2'

Total ", ",'

The ordinary J! test (Fisher, 1958) or the normal approximation to the
distribution of (PI - P2) is appropriate. Similarly, comparisons among proportions
for more than two domains are made by the methods for a 2 x k contingency table.
Occasionally it is desired to test whether a I differs significantly from a 2; for
example, whether the number of Republicans who favor some proposal is greater
than the number of Democrats in favor. On the null hypothesis that these two
numbers are equal in the population, the total n' ;:: a I +a 2 in the two classes in
question should divide with equal probability between the two classes. Conse-
quently we may regard al as a binomial number of successes in n' trials, with
probability of success t on the null hypothesis. It may be verified that the normal
devi&te (corrected for continuity) is
2(l al-!n'I -!)
J;i

3.12 ESTIMATION OF PROPORTIONS IN CLUSTER

SAMPLING
As mentioned in section 3.2, the preceding methods are not valid if each unit is.a
cluster of elements and we are estimating the proportion of elements that fall into
class C.
If each unit contains the same number m of elements, let PI = at! m be the
proporti n of elements in the ith unit that fall into class C. The proportion falling
in C in the sample is
La,
p=-=-I Pi
1"
nm n
that is, the estimate P is the unweighted mean of the quantities Pi' Consequently, if
Yi is replaced by Pi. the formulas in Chapter 2 may be applied directly to give the
 SAMPLING PROPORTIONS AND PERCENTAGES 65
true and estimated variance of p.
N
V(P) = 1- n:. (p, - p)2 (3.29)
n N -l
An unbiased sample estimate of this variance is

(3 .30)

Example 1. A group of 61 leprosy patie nts were treated with a drug for 48 weeks. To
measure the effect of the drug on the leprosy bacilli, the presence of bacilli at six sites on the
body of each patient was tested bacteriologically. Among the 366 sites, 153, or 41.8% ,
were negative. What is the standard error of this percentage?
This example comes from a co ntrolled experiment rather than a survey, but it illustrates
how erroneous the binomial formula may be. By the binomial formula , we have n = 366,
and
s.e. (p) = J pq/( n - 1) = J(41.8 )(58.2)/365 = 2.58%
Each patient is a cluster unit with m = 6 elements (sites). To find the standard error by the
correct formula , we need the frequency distribution of the 61 values of P" It is more
conve nient to tabulate the distribution of Yi' the number of negative sites per patient. With
p, expressed in percents, Pi = 100y,/6. From the distribution in Table 3.4 we find I fy 2=
669 and

= JIf, (y, - 9)2 _ J669 - [(153? / 61)

s.e. (ji) ,, (n - l ) (6 1)(60) 0.279
Hence
S.e . (p) = 1~0 s.e. (9) = 4.65%
This figure is about 1.8 times the value given by the binomial formula . The binomial
fo rmula requires the assumption that results at different sites on the same patient are
independent, although actually they have a strong positive correlation. The last line of
Table 3.4 shows the expected number of patients with O. 1,2, . .. negative sites, computed
from the binomial (0.58 + 0.42)10. Note the marked excesses of observed frequencies f of
patient~ with zero negatives and with five and six negatives.

TABLE 3.4
N U MB ER OF NEGATIVE SITES PER PATI ENT

Yi = 6p{/IOO 0 2 3 4 5 6 Total

f 17 II 4 4 7 14 4 61
fYI 0 II 8 12 28 70 24 153
/.xP 2.3 10. 1 18.3 17.6 9.6 2.8 0.3 61.0
66 SAMPLING TECHNIQUES

If the size of cluster is not constant, let mi be the number of elements in the ith
cluster unit and let Pi = all mi. The proportion of units falling in class C in the
sample is

(3.31)

Structurally, this is a typical ratio estimate, discussed in section 2. 11 and later in

Chapter 6. It is slightly biased, although the bias is seldom likely to be of practical
importance.
If we put a i for YI and mi for X i in (2.39), the approximate variance of P is

V(P )= 1-_1 L (al- Pmi)2

(3.32)
nM2 N-l
_ N
where P is the proportion of elements in C in the population and M = L mJ N is
the average number of ele ments per cluster. An alternative expression is

V(p) = 1 - If ("!..i)2(Pi - p)2 (3.33)

n M N- l

This form shows that the approximate variance involves a weighted sum of
squares of deviations of the PI from the population value P.
For the estimated variance we have
2
v(p ) = 1- 1 .2: a?- 2p.2: aim; +p 2 L m ,
(3.34)
nm 2 n- 1

where m = L mJ n is the average number of elements per cluster in the sample.

Ezample 2. A simple random sample of 30 households was drawn from a census taken
in 1947 in wards 6 and 7 of the Eastern Health District of Baltimore. The population
contains about 15,000 households. In Table 3.5 the persons in each household are classified
(a) according to whether they had consulted a doctor in the last 12 months, (b) according
to sex.
Our purpose is to contrast the ratio formula with the inappropriate binomial formula.
Consider first the proportion of people who had consulted a doctor. For the binomial
formula, we would take
30
n = 104, p = 104 =0.2885

Hence
_ /Xl _ (0.2885)(0.7115) 0.00197
_VbI.
(p) - n - 104
 SAMPLING PROPORTIONS AND -PERCENTAGES 67

TABLE 3.5
DATA FOR A SIMPLE RANDOM SAMPLE OF 30 HOUSEHOLDS
Doctor Seen in
Number of Last Year
Number of
Household Persons Males Females Yes No
Number mj Qj
°i
1 5 1 4 5 0
2 6 3 3 0 6
3 3 2 2
4 3 2 3 0
5 2 I 0 2
6 3 2 0 3
7 3 2 0 3
8 3 1 2 0 3
9 4 2 2 0 4
10 4 3 0 4
II 3 2 0 3
12 2 1 1 0 2
13 7 3 4 0 7
14 4 3 4 0
15 3 2 1 I 2
16 5 3 2 2 3
17 4 3 0 4
18 4 3 I 0 4
19 3 2 1 I 2
20 3 I 2 3 0
21 4 I 3 2 2
22 3 2 0 3
23 3 2 0 3
24 1 0 0 J
25 2 1 1 2 0
26 4 3 1 2 2
27 3 1 2 0 3
28 4 2 2 2 2
29 2 1 1 0 2
30 4 2 2 1 3

Totals 104 53 51 30 74
68 SAMPLING TECHN IQUES

For the ratio formula, we note that there are 30 clusters and take
n =30
m, = total number in ith household
a, = number in ith household who had seen a doctOr
p = 0.2885, as before
104
m =-=3 .4667
30
L a,2=86; L m,' =404; L aim, = 113
The fpc may be ignored . Hence, from (3.34),
u ) = (86) - 2(0.2885)(113) + (0.2885)2( 404 ) = 0.00520
(p (30)(29)(3.4667)'
The variance given by the ratio method, 0.00520, is much larger thall that given by the
binomial formula , 0.00197. for various reasons. families differ in the frequency with which
their members co nsult a doctor. For the sample as a whole, the proportion who con.ult a
doctor is only a little more than one in four. but there are several families in which every
member has seen a doctor. Similar results would be obtained for any characteristic in which
the members of the sa me family tend to act in the same way .
[n estimating the proportion of males in the population, the results are different. By the
same type of calculation, we find
binomial formula : u(p) = 0 .00240
ratio formula u(p) = 0.00114
Here the binomial formula ouerestimates the variance. The reason is interesting. Most
households are set up as a result of a marriage, hence contain at least one male and one
female . Consequently the proportion of males per family varies less from one half than
would be expected from the binomia! formula . None of the 30 families. except one with
only one member, is compo;ed enti rely of males, or entirely of fema les. If the binomial
distribution were applicah le, with a true P of approximately o ne half. households with all
members of the same sex would constitute o ne quarter o[ the households of size 3 and one
eighth of the households of size 4. This property of the sex ratio has been discussed by
Hansen and Hurwitz (1942) . Other illustrations of the error committed by improper use of
the binomial formula in sociological investigations have been given by Kish (1957).

EXERCISES
3.1 For a population with N = 6, A = 4 , A' = 2, work out the value of a for all possible
simple random samples of size 3. Verify the theorems given for the mean and variance of
p = a/ II . Verify that
N-n
- --pq
(n-!)N
is an unbiased estimate of the variance of p.
3.2 In a simple random sample of 200 from a population of 2000 colleges, 120 colleges
wert in favor of a proposal: 57 were opposed, and 23 had no opinion . Estimate 95%
 SAMPLING PROPORTIONS AND PERCENTAGES 69
confidence limits for the number of colleges in the population that favored the
proposal.
3.3 Do the results of the previous sample furnish conclusive evidence that the majority
of the colleges in the population favored this proposal?
3.4 A, population with N =7 consists of the elements BI> C" C 2 , C J , D" D 2 , and D J • A
simple random sample of size 4 is taken in order to estimate the proportion of C's to
C's + D's. Work out the conditional distributions of this proportion, p, and verify the
formula for its conditional variance.
3.5 In the preceding exercise. what is the probability that a sample of size 4 contains
8, ? Find (he average variance of p in exercise 3.4 over aI/ simple random samples of size 4.
This is 0.0393 as against 0.025 with N = 6. n = 4. and B, absent.
3.6 A simple random sample of 290 households was chosen from a city area containing
14,828 households. Each family was asked whether it owned or rented the house and also
whether it had the exclusive use of an indoor toilet. Results were as follows .
Owned Rented Total

Exclusive use of toilet Yes No Yes No

141 6 109 34 290
(a) For families who rent, estimate the percentage in the area with exclusive use of an
indoor toilet and give the standard error of your estimate; (b) estimate the total number of
renting families in the area who do not have exclusive indoor toilet facilities and give the
standard error of this estimate.
3. 7 If. in example 3.6, the total number of renting families in the city area is 7526, make
a new estimate of the number of renters without exclusive toilet facilities and give the
standard error of this estimate.
3.8 For estimating the total number of units in class C in domain I (section 3.10), the
estimateA, = N,p, was recommended if N, were known, as against A,' = Na,/n if N, ~ere
not known. Ignoring the fpc, show that in large samples the ratio of the variance of AJ to
that of AJ' is approximately 0 ,/( OJ + P,7T) , where 7T is the proportion of the population
that is not in domain I, and PI> as in section 3.10, is the proportion of the units in domain 1
that fall in class C. State the conditions under which knowledge of N, produces large
reductions in variance.
'3.9 In a simple random sample of size 5 from a population of size 30, no units in the
sample were in class C. By the hypergeometric distribution, find the upper limit to the
number A of units in class C in the population, corresponding to a one-tailed confidence
probability of 95% . Find also the approximation to Au obtained by computing the upper
95% binomial limit Pu and shortening the interval as described in section 3.6. Try also the
method on p. 59, Example 3.
3.10 A student health service has a record of the total number of eligible students N
and of the total number of visits Y made by students during a year. Some students made no
visits. The service wishes to estimate the mean number of visits Y/ Nt for tbe N, students
who made at least one visit, but does not know the value of N, . A simple random sample of
n eligible students is taken. In it n, students out of the n made at least one visit and their
total number of visits was y. Ignore the fpc in this question. (a) Show that y/n, is an
unbiased estimate of Y/N J and that its conditional variance is S2/n" where S2 is the
variance of the number of visits among students making at least one visit. (b) A second
method of estimating Yj N, is to use N, = Nnl/n as an estimate of NJ and hence Yn/ Nnl as
an estimate of Y/ N I . Show that this estimate is biased and that the ratio of the bias to the
70 SAMPLING TECHNIQUES

true value Y/ Nt is approximately (N - NJ)/nN J. Find an approximatCf expression for the

variance of the esti!l1ate Y./ Nn I and show that the estimate in (0) has a higher variance if

Hint. If p is a binomial estimate of P, based on n trials, then approximately

3.11 Which of the two previous estimates seems more precise in the following
circumstances? N =2004, Y = 3011 . The sample with n = 100 showed that 73 students
made at least one visit. Their total number of visits was 152 and the estimated variance S2
was 1.55.
3.12 A simple random sample of n cluster units, each with m elements, is taken from a
population in which the proportion of elements in class C is P. A~ the intracluster
correlation varies, what are the highest and lowest possible values of the true variance of p
(the sample estimate of P ) and how do they compare with the binomial variance? Ignore the
fpc.
3.] 3 For thc sample of 30 households in Table 3.5, the dllta shown below refer to visits
to the dentist in the last year. Estimate the variance of the proportion of persons who saw a
dentist, and compare this with the binomial estimate of the variance.
3.14 In sampling for a rare attribute, one method is to continue drawing a simple
random sample until m units that possess the rare attribute have been found (Haldane,
1945) where m is chosen in advance. If the fpcis ignored, prove that the probability that the

Number Dentist Seen Number Dentist Seen

of of
Persons Yes No Persons Yes No

5 1 4 5 4
6 0 6 4 4 0
3 I 2 4 3
3 2 I 3 J 2
2 0 2 3 0 3
3 0 3 4 J 3
3 2 3 0 3
3 2 3 1 2
4 I 3 1 0 I
4 0 4 2 0 2
3 1 2 4 0 4
2 0 2 3 1 2
7 2 5 4 1 3
4 1 3 2 0 2
3 0 3 4 0 4
 SAMPLING PROPORTIONS AND PERCENTAGES 71
total sample required is of size n is
(n -I)! pmQ"- m
(n c: m )
(m-l)!(n - m )!
where P is the frequency of the rare attribute. Find the average size of the total sample and
show that if m > 1, p = (m - l)/ (n - 1) is an unbiased estimate of P. (For further discussion.
see Finney, 1949. and Sandelius. 1951. who considers a plan in which sampling continues
until e ither m have been found or the total sample size has reached a preassigned limit no.)
See also section 4.5.
 74 SAMPLING TECHNIQUES

4. It often happens that data are to be published for certain major subdivisions
of the population and that desired limits of error are set up for each subdivision. A
separate calculation is made for the n in each subdivision, and the total n is found
by addition .
5. More than one item or characteristic is usually measured in a sample survey:
sometimes the number of items is large. If a desired degree of precision is
prescribed for each item, the calculations lead to a series of conflicting values of n,
one for each item. Some method must be found for reconciling these values.
6. Finally, the chosen value of n must be appraised to see whether it is
consistent with the resources available to take the sample. This demands an
estimation of the cost, labor, time, and materials required to obtain the proposed
size of sample. It sometimes becomes apparent that n will have to be drastically
reduced. A hard decision must then be faced-whether to proceed with a much
smaller sample size, thus reducing precision, 'or to abandon efforts until more
resources can be found.
In succeeding sections some of these questions are examined in more detail.
*

4.3 THE SPECIFICATION OF PRECISION

The statement of precision desired may be made by giving the amount of error
that we are willing to tolerate in the sample estimates. This amount is determined,
as best we can, in the light of the uses to which the sample results are to be put.
Sometimes it is difficult to decide how much error should be tolerated, particularly
when the results have severa) different uses. Suppose that we asked the
anthropologist why he wished the percentage with blood group 0 to be correct to
5% instead of, say, 4 or 6% . He might reply that the blood group data are to be
used primarily for racial classification. He strongly suspects that the islanders
belong either to a racial type with a P of about 35% or to one with a P of about
50% , An error limit of 5% in the estimate seemed to him small enough to permit
classification Into one of these types. He WOUld, however, have no violent
objection to 4 or 6% limits of error.
Thus the choice of a 5%limit of error by the anthropologist was to some extent
arbitrary. In this respect the example is typical of the way in which a limit of error
is often decided on. In fact, the anthropologist was more certain of what he wanted
than many other scientists and administrators will be found to be. When the
question of desired degree of precision is first raised, such persons may confess
that they have never thOUght about it and have no idea of the answer. My
experience has been, however, that after discussion they can frequently indicate at
least roughly the size of a limit of error that appears reasonable to them.
Further than this we may not be able to go in many practical situations. Part of
the difficulty is that not enough is known about the consequences of errors of
different sizes as they affect the wisdom of practical decisions that are made from
 TIre ES1lMAnON OF SAMPLE SIZE 75
survey results. Even when these consequences are known, however, the results of
many important surveys are used by different people for different purposes, and
some of the purposes are not foreseen at the time when the survey is planned.
Therefore, an element of guesswork is likely to be prominent in the specification
of precision for some time to come.
If the sample is taken for a very specific purpose, (e.g. , for making a single "yes"
or " no" decision or for deciding how much money to spend on a certain venture),
the precision needed can usually be stated in a more definite manner, in terms of
the consequences of errors in the decision. A general approach to problems of this
type is given in section 4 .10, which, although in need of amplification, offers a
logical start on a solution.

4.4 THE FORMULA FOR n IN SAMPLING FOR PROPORTIONS

The units are classified into two classes, C and C. Some margin of error d in the
estimated proportion p of units in class C has been agreed on, and there is a small
risk a that we are willing to incur that the actual error is larger than d; that is, we
want
Pr (I p - pi 2: d) =a
Simple random sampling is assumed, and p is taken as normally distributed.
From theorem 3.2, section 3.2,

CT
p
= [i:i::;,
VN=1Y~
!PO
Hence the formula that connects n with the desired degree of precision is

d=t ~ fPQ
V~V-;;=-
where t is the abscissa of the normal curve that cuts off an area of a at the tails.
Solving for n, we find

n= 2 (4.1)
1+!(t PO-I)
2
N d
For practical use, an advance estimate p of P is substituted in this formula. If N is
large, a first approximation is

(4.2)
76 SAMPLING TECHNIQUES

where
V = pq = desired variance of the sample proportion
no
In practice we first calculate no. If no/ N is negligible, no is a satisfactory
approximation to the n of (4.1) . If not, it is apparent on comparison of (4.1) and
(4.2) that n is obtained as
no
no: =. no
(4.3)
1+(no-1)/N l+(no/N)
Example. In the hypothetical blood groups example we had
d = 0.05 , p = 0.5, (l' = 0.05 , t =2
Thus
_ (4)(0.5)(0.5) 00
no - (0.0025) 4
Let us assume that there are only 3200 people on the island . The fpc is needed, and we
find
no 400
n- - - = 356
1+(no- l)/ N L+~)
The formula for no holds also if d, p. and q_ are all expressed as percentages instead of
proportions. Since the product pq increases. p moves toward t or 50%, a conservative
estimate of n is obtained by choosing for p the value nearest to ~ in the range in which pis
thought likely to lie. If p seems likely to lie between 5 and 9%, for instance. we assume 9%
for the estimation of n.
Sometimes, particularly when estimating the total number NP of units in class
C, we wish to control the relative error r instead of the absolute error in Np ; for
example, we may wish to estimate NP with an error not exceeding 10% . That is,
we want
'NP - NPI )
Pr ( NP >r =Pr('p-PI~rP)=a

For this specification, we substitute rP or rp for d in formulas (4.1) and (4.2). From
(4.2) we get

(4.2),

Formula (4.3) is unchanged.

4.5 RARE ITEMS-INVERSE SAMPLING

In estimating n from formulas (4.1), (4.2) and (4.2)" the sampler inserts his best
advance estimate of the population proportion P. If P is known to be between 30
 THE ESTIMATION OF SAMPLE SIZE 77
and 70%, as in the example in Section 4.1, accurate estimation of P is not crucial.
But with a rare item (e.g., PslO%), the necessary n for a specified relative error r
is 11 times as large when P = 1% as when P = 10%. In this situation (P small but
not well known in advance), Haldane's (1945) method of continuing sampling
until m of the rare items have been found in the sample has one important
advantage . The method is usually called inverse sampling.
If n is the sample size at which the mth rare item appears, (m > 1), an unbiased
estimate of P isp = (m -l)/(n -1). For Nvery large, Psmall, and m ~ 10, a good
approximation to V(P) may be shown to be mP 2 0/(m - 0 2 . Hence.
cv( p) == (mQ)I /2 /(m -1) <Fm/(m -1), which will be a close upper limit if P is
,mall. Thus, by fixing m in advance, we can control the value of cv(P) without
advance knowledge of P. The value m = 27 gives cv(P) < 20%, but m = 102 is
needed for cv(P) < 10%. The value of n with this method is a random variable, but
will be large if P is small.

4.6 THE FORMULA FOR n WITH CONTINUOUS DATA

Most commonly, we wish to ~ontro l the relative error r in the estimated
population total or mean. With a simple random sample having mean 9, we want
9 - YI)
~ r ==Pr (INY-NYI) --
Pr (1 y-- NY 2:. r =Pr(!9-Y!2:.ry)=a

where a is a small probability. We assume that 9 is normally distributed: from

theorem 2.2. corollary 1. its standard error is

Uy= ~~
Hence
-
r Y = tU9 == tV~S
-:;--- r (4.4)
N vn
Solving for n gives

Note that the population characteristic on which n depends is its coefficient of

variation S/ Y. This is often more stable and easier to guess in advance than S
Itself.
As a first approximation we take

(4.5)
78 SAMPLING TECHNIQUES

substituting an advance estimate of (SI n.

The quantity C is the desired (CV)2 of
the sample estimate.
If nolN is appreciable we compute n as in (4.3)

n = _-,.n.. . ::o ..,-_ (4.3)

1 + (nolN)
If instead of the relative error r we wish to control the absolute error d in y, we
take no = (2 S 2I d 2 = S2 I V, where V is the desired variance of Y.

Example. In nurseries that produce young trees for sale it is advisable to estimate, in
late winter or early spring. how many healthy young trees are likely to be on hand, since this
determines policy toward the solicitation and acceptance of orders. A study of sampling
methods for the estimation of the total numbers of seedlings was undertaken by Johnson
(1943). The data that follow were obtained from a bed of silver maple seedlings 1 ft wide
and 430 ft long. The sampling unit was 1 it of the length of the bed. so that N = 430. By
complete enumeration of the bed if was found that Y = 19, S2 = 85.6, these being the true
population values.
With simple random sampling, how many units must be taken to estimate Y within 10%.
apart from a chance of 1 in 20? From (4.5) we obtain
12 S2 (4)(85.6)
no= r2f2= (1.9)2 95

Since no/ N is not negligible, we take

95
n=--= 78
1+&0
Almost 20% of the bed has to be counted in order to attain the precision desired.
The formulas for n given here apply only to simple random sampling in which
the sample mean is used as the estimate of Y. The appropriate formulas for other
methods of sampling and estimation are presented with the discussion of these
techniques.

4.7 ADVANCE ESTIMATES OF POPULATION VARIANCES

The nursery example is atypical in that the population variance S2 was known.
In practice, there are four ways of estimating population variances tor sample size
determinations : (1) by taking the sample in two steps, the first being a simple
random sample of size nl from which estimates s/ or PI of S2 or P and the
required n will be obtained; (2) by the results of a pilot survey; (3) by previous
sampling of the same or a similar population ; and (4) by guesswork about the
structure of the population, assisted by some mathematical results.
Method 1 gives the most reliable estimates of S 2 or P, but it is not often used,
since it slows up the oompletion of the survey. When the method is feasible, Cox
(1952), following work by Stein (1945), shows how to compute n from A 12 or PI so
that the final estimate y or P will have a preassigned variance V, a preassi~ed
 THE ESTIMATION OF SAMPLE SIZE 79
limit of error d, or a preassigned coefficient of variation. The first sample is
assumed large enough to neglect terms of order 1/ n J 2. A few results are quoted.
The results given here assume n 1 ~ n, the size of the final sample. When this is
not so, see Cox (1952).
Estimation of Y with given cv = JC
The results assume y, normally distributed. If s I 2 is the estimated variance from
the first sample, take additional units to make the final sample size

n = s/
- -2 ( 1 +8C+--+-
s/ 2) (4 .6)
CYI nlY/ nl
The mean Yof the final sample is slightly biased. Take :- = y(1 - 2C).
Estimation of Y with Variance V
Take additional units to make the total sample size

n = _!_ 2)
S 2 ( 1+-
(4 .7)
V nl

If S were known exactly, the required sample size would be S2/ V. The effect of
not knowing S is to increase tile average size by the factor (1 + 2/ n I) .
Estimation of P with Variance V
Let PI be the estimate of P from the first sample. The combined size of the first
two samples should be
Plql 3 - 8plql 1 - 3plql
n = --+ + - --'--'-'''':' (4.8)
V Plql Vnl
The first term on the right is the size required if P is known to be equal to PI. With
this method, the ordinary binomial estimate P made from the complete sample of
size n is slightly biased. To correct for bias, take
~ 1_
_V_:_( - __;2p'-'-)
P=p+
pq

Estimation of P with given cv =JC

Take
ql 3 1
n= - +--+-- (4.9)
CPI Plql CPlnJ
The estimate is P= P- CP/ q. In all results given above the fpc is ignored.

Example. A sampler wishes to estimate P with a coefficient of variation of O.l(1u%) .

He guesses that P will lie somewhere between 5 nd 20% . This range is too wide to give a
80 SAMPLING TECHNIQUES -

good initial estimate of the required n. Since the cv of P is JOlnP, it is easily verified that
n = 400 is adequate for P 0= 20%, but n = 1900 will be needed if P is only 5%.
Accordingly, he tak.es an initial sample with n , 0= 396 and finds P, =0.101. Since
JC=O. J, C = 0.01. Equation 4.9 gives
n= (0.899) + _ _3_ + I = 926
(0.01)(0.101) (0.0908 ) (0 .01)(40)

The combined sample gives tip = R8 ; p = 8R/ 926 = 0.0950. The correction (or bias, Cpl q,
amounts to 0.001 I. giving a final estimate of 0.094 or 9.4% .

The second method, a small pilot survey, serves many purposes, especially if the
feasibility of the main survey is in doubt. If the pilot survey it itself a simple
random sample, the preceding methods apply. But often the pilot work is
restricted to a part of the population that is convenient to handle or that will reveal
the magnitude of certain problems. Allowance must be made for the selective
nature of the pilot when using its results to estimate 5 2 or P. For instance, a
common practice is to confine the pilot work to a few clusters of units. Thus the
computed .\.2 measures mostly the variation within a cluster and may be an
underestimate of the relevant 5 2. The relation between intra- and intercluster
variation is discussed in Chapter 9. The same problem arises in cluster sampling
for proportions, in which the formula pq/ n may underestimate the effect of
variation among clusters. Cornfield (1951) gives a good illustration of the estima-
tion of sample size in cluster sampling for proportions.
Method 3-the use of results from previous surveys- points to the value of
making available, or at least keeping accessible, any data on standard deviations
obtained in previous surveys. Unfortunately, the cost of computing standard
deviations in complex surveys is high, even with electronic machines, and fre-
quently only those S.d.'s needed to give a rough idea of the precision of the
principal estimates are computed and recorded. If suitable past data are found, the
value of S2 may require adjustment for time changes. With skew data in which Vis
changing with time, 5 2is often found to change at a rate lying somewhere between
k Yand k y2, where k is a constant. Thus, if Y is thought to have increased by 10%
in the time interval since the previous survey, we might increase our initial
estimate of 5 ~ by 10 to 20% . .
Finally, a serviceable estimate of 5 2 can sometimes be made from relatively
little information about the nature of the population. In early studies of the
numbers of wireworms in soils, a tool was used to take a sample (9 x 9 x 5 in.) of
the tOPSOIl. For estimating n, the sampler needed to know the tandard deviation
of the number of wireworms found in a boring with the tool. If wireworms were
distributed at random over the topsoil, the number found in a small volume would
follow the Poisson distribution, for which 52 = Y. Since there might be some
tendency for wireworms to congregate, it was decided to assume 5 2 = 1.2 Y, the
factor 1.2 being an arbitrary safety factor. Although Y was not known, the values
 TIlE ESTIMATION OF SAMPLE SIZE 81
of Y that are of economic importance with respect to crop damage could be
delineated. These two pieces of information made it possible to determine sample
sizes that proved satisfactory.
Deming (1960) shows how some simple mathematical distributions may be
used to estimate S2 from a knowledge of the range and a general idea of the shape
of the distribution. If the distribution is like a binomial, with a proportion p of the
observations at one end of the range and a proportion q at the other end,
52= pqh 2 , where h is the range . When p = q = t the value of S2= 0.2Sh 2 is the
maximum possible for a given range h. Other useful relations are that S2=
=
O.083h 2 for a rectangular distribution, S2 O.OS6h 2 for a distribution shaped like
a right triangle, and S2= O.042h 2 for an isosceles triangle.
These relations do not help much if h is large or poorly known . However, if h is
large, good sampling practice is to stratify the population (Chapter 5) so that
within any stratum the range is much reduced. Usually the shape also becomes
simpler (closer to rectan~ular) within a stratum. Consequently, these relations are
effective in predicting S , hence n, within individual strata.

4.8 SAMPLE SIZE WITH MORE THAN ONE ITEM

In most surveys information is collected on more than one item. One method of
determining sample size is to specify margins of error for the items that are
regarded as most vital to the survey. An estimation of the sample size needed is
first made separately for each of these important items.
When the single item estimations of n have been completed, it is time to take
stock of the si tuation. It may happen that the n 's required are all reasonably close.
If the largest of the n 's falls within the limits of the budget, this n is selected. More
commonly, there is a sufficient variation among the n's so that we are reluctant to
choose the largest, either from budgetary considerations or because this will give
an over-all standard of precision substantially higher than originaHy contem-
plated . In this event the desired standard of precision may be relaxed for certain of
the items, in order to permit the use of a smaller value of n.
In some cases the n 's required for different items are so discordant that certain
of them must be dropped from the inquiry; with the resources available the
precision expected for these items is totally inadequate. The difficulty may not be
merely one of sample size. Some items call for a different type of sampling from
others. With populations that are sampled repeatedly, it is useful to amass
information about those items that can be combined economically in a general
survey and those that necessitate special methods. As an example, a classification
of items into four types, suggested by experience in regional agricultural surveys,
is shown in Table 4.1. In this classification, a general survey means one in wllk h
the units are fairly evenlY distributed over some region as, for example, by a
simple random sample.
82 SAMPLING TECHNIQUES

TABLE 4.1
AN EXAMPLE Of DIFfERENT TYPES OF ITEM IN
REGIONAL SURVEYS

Type Characteristics of Item Type of Sampling Needed

Widespread throughout the region. occur- A general survey w~th low

ring with reasonable frequency in all sampling ratio.
parts.
2 Widespread throughout the region but with A general survey. but with
low frequency. a higher sampling ratio.
3 Occurring with reasonable frequency in For best results, a stratified
most parts of the region. but with more sample with different in-
sporadic distribution, being absent in tensities in different parts
some parts and highly concentrated in of the region (Chapter 5).
others. Ca n sometimes be in-
cluded in a general survey
with supplementary sam-
pling.
4 Distribution very sporadic or concentrated Not suitable for a gen-
in a small part of the region. eral survey. Requires a
~ample geared to its dis-
tribution.

4.9 SAMPLE SIZE WHEN ESTIMATES ARE WANTED

FOR SUBDIVISIONS OF THE POPULATION
It is often planned to present estimates not only for the population as a whole
but for certain subdivisions. If these can be identified in advance, as with different
geographical regions, a separate calculation of n is made for each region . Suppose
that the mean of each subdivision is to be estimated with a specified variance V.
For the ith subdivision, we have nj = S,2/ V, so that the total sample size n =
L S? / V. The individual S/ will, on the average, be smaller than S2, the population
variance, but often they are only slightly smaller. Thus, if there are k subdivisions,
n == kS 2 / V, whereas if only the estimate for the population as a whole were
wanted we would take n = S2/ V.
Thus if estimates with variance V are wanted for each of k subdivisions the
sample size may approach k times the n needed for an over-all estimate of the
same precision. This point tends to be overlooked in calculations of sample size by
persons inexperienced in survey methods.
If the subdivisions represent classifications by variables such as age, sex,
income, and years of schooling, the subdivision to which a person belongs is not
known until the sample has been taken. Advance sample size e timates can still be
 THE ESTIMAnON OF SAMPLE SIZE 83
made if the proportions l7j of the units that belong to the various subdivisions are
known. If a simple random sample of size" is selected, the expected size of sample
from the ith subdivision ~s "171, The average variance of the mean from this
subdivision is
(4.10)

if nl7j is large. Hence we require" == S? / l7j V in order to make V(YI) = v. If this is

to hold for every subdivision,

n == max ·( S? ) (4.11)
l7j V
If the subdivisions are into classes like age, income , S ; /171 may be less than S 2 for
central classes, but may be large for an extreme class with smalll7j. In this event,
we may either have to increase the value qf V in this subdivision or find some way
of identifying units in this subdivision in advance so that they can be
sampled at a higher rate. The method of double sampling (Chapter 12) is
sometimes useful for this purpose.
The demands on sample size are still greater in analytical studies in which the
specifications are
V(y, - Yi):S V (4.12)
fo r every pair of subdivisions (domains). In this case

n == max-
1 (S2 S.2)
~+ ...L (4.l3)
1./ V 171 17/
2
If the S/ are not very different from S 2, " will be 2kS / Vwhen the k domains are
of equal size, and still greater otherwise. The effect of fpc terms, neglected in this
discussion, is to reduce the required" 's to some extent.

4.10 SAMPLE SIZE IN DECISION PROBLEMS

A more logical approach to the determination of sample size can sometimes be
developed when a practical decision is to be made from the results of the sample.
The decision will presumably be more soundly based if the sample estimate has a
low error than if it has a high error. We may be able to calculate, in monetary
terms, the loss I(z) that will be incurred in a decision through an error of amount z
in the estimate. Although the actual value of z is not predictable in advance,
sampling theory enables us to find the frequency distribution fez , n) of z which,
for a specified sampling method will depend on the sample size n. Hence the
expected loss for a given size of sample is

L(n) = J I(z)f(z, n) dz (4.14)

84 SAMPLING TECHN IQUES

The purpose in taking the sample is to diminish this loss. If C(n) is the cost of a
sample of size n, a reasonable procedure is to choose n to minimize
C(n )+ L(n) (4.15)
since this is the total cost involved in taking the sample and in making decisions
from its results. The choice of n determines both the optimum size of sample and
the most advantageous degree of precision .
Alternatively. th e same approach can be presented in terms of the monetary
gain that accrues from having the sample information, instead of in terms of the
loss that arises from errors in the sample information. If monetary gain is used, we
construct an expected gain G (n ) from a sample of size n. where G (n ) is zero if no
sample is taken . We maximize
G(n) - C(n)
In this form the principle is equivalent to the rule in classical eco lomics that profit
is to be maximized .
The simplest application occurs when the loss function , l (z), is Az 2, where A is a
constant. It follows that
L(n) = AE(z2) (4.16)
For instance, if Yis the sample estimate of Y, and z = Y- Y,
!! AS 2 AS 2
L(n)=A V(y) = - - - (4.17)
n N
if simple ra ndom sampling is used .
The simplest type of cost function for the sa mple is
C(n )=co+c,n (4.18)
where Cl) is the overhead cost. By differentiation. the value of n that minimizes
cost plus loss i
n = J AS 2/c, (4.19)
A mo re general form of this result is given by Yates (1960). The same analysis
applies to any method of sampling and estimation in which the variance of the
estimate is inversely proportional to n and the cost is a linear function of n.
Blythe (1945) descri bes the application of this principle to the estimation of the
volume of timber in a lot for selling purposes (see exercise 4 .11). Nordin (1944)
discusses the optimum size of sample for estimating potential sales in a market
that a manufacturer intends to enter. If the sales can be forecast accurately, the
amount of fixed equipment and the production per unit period can be allocated to
maximize the manufacturer'S expected profit. Grundy et al. (1954. 1956) consider
the optimum size of a second sample when the resul ts of a first sample are already
known .
 TIlE ESTIMATION OF SAMPLE SIZE 85
This approach has received substantial further development from workers on
statistical decision theory. Generalizations include the substitution of utility for
money value as a scale on which to measure c sts and losses, the explicit use of
subjective prior information about unknown parameters by expressing this infor-
mation as " prior" probability distributions of the unknown parameters, and the
investigation of different types of cOSJ and loss functions and of qualitative as well
as quantitative data . For a comprehensive account of the method, see Raiffa and
Schlaifer (1961). Although it is still not evident how frequently decision prob-
lems will be amenable to complete solution by this approach, the method has
value in stimulating clear thinking about the important factors in a good decision .
One area that appears suitable for applications is the sampling of lots of articles in
a mass-production process in order to decide whether to accept or reject the lot on
the basis of its estimated quality. Sittig (1951) considers the economics of
sample-size determination, taking account of costs of inspection and the costs
incurred through defective articles in accepted lots and good articles in rejected
lots.

4.11 THE DESIGN EFFECf (De//)

With the more complex sampling plans described later in this book, a useful
quantity is the design effecr (deff) of the plan (Kish, 1965). He describes this as the
ratio of the variance of the estimate obtained from the (more complex) sample to
the variance of the estimate obtained from a simple random sample of the same
number of units. The design effect has two primary uses-in sample-size estima-
tion and in appraising the efficiency of more complex plans. For instance, in
estimating the proportion of people who possess some attribute, it i often
convenient to use the household instead of the person as a sampling unit. As noted
in Chapter 3, the formula POln cannot be used with these plans. For estimating
the proportion who had seen a doctor (section 3.12), a simple random sample of
households gave v(p) = 0.00520 as against pql n = 0.00197 for an equal-sized
simple random sample of persons. An estimate of the deff for this cluster sample
and this variate is 520/197 = 2.6. When the sampling fractions are small, we can
therefore estimate sample size by calculating the n (number of persons) needed
with a simple random sample of persons and mUltiplying by 2.6. By noting deff
ratios in this way for the important variates with a complex plan, we can use the
simple formulas in this chapter for estimating the sample size with the complex
plan and also judge whether the complex plan is advantageous in efficiency
relative to its cost and complexity.
Estimating the deff from the results of a complex sample may require some
algebra. We need to show how these results provide, if possible, unbiased
estimates of S2 or of PQ. Examples of these calculations are given for stratified
random sampling in section 5A.l1 and for cluster sampling with clusters of equal
sizes in section 9.3.
 86 SAMPLING TECHNIQUES

EXERCISES
4.1 In a district containing 4000 houses the perc;entage of owned houses is to be
estimated with a s.e. of not more than 2% and the percentage of two-car households with a
s.e. of not more than 1% . (The figures 2 and 1% are the absolute values, not the cv's.) The
true percentage of owners is tbought to lie between 45 and 65% and the percentage of
two-car households between 5 and 10% . How large a sample is necessary to satisfy both
aims?
4.2 In the population of 676 petition sheets (Table 2.2. page 28) how large must the
sample be if the tota l numher of signatures is to be estimated with a margin of error of 1000.
apart from a 1 in 20 chance? Assume that the value of S 2 given on page 28 is the population
S' .
4.3 A survey is to be made of the prevalence of the common diseases in a large
population. For any disea e that affects at least 1% of the individuals in the population, it is
desired to estimate the total number of cases. with a coefficient of variato n of not more than
. 20% . (a) What size of si mple random ample is needed, assuming that the presence of the
disease can be recognized without mistakes? (b) What size is needed if total cases are
wanted separately for males and females, with the same precision?
4.4 In a wireworm survey the number of wireworms per acre is to be estimated with
a limit of error of 30% , at the 95 % ptobability level, in any field in which wireworm
dens it y exceeds 200.000 per acre in the top 5 in . of soil. The sampling tool measures
Y x 9 x 5 in . deep. Assuming that the number of wireworms in a sin_gle sample follows a
distribution slightly more variable th an the Poisson , we take 52= 1.2 Y . What size of simple
ran dom sample is needed ? ( I acre = 43 ,560 sq ft.)
4.5 The following coefficients of variation per unit were obtained in a farm survey in
Iowa, the unit being an area 1 mile square (data of R. J . Jessen):

Estimated cv
Item (% )

Acres in farms 38
Acres in corn 39
Acres in oats 44
Number of famil y workers 100
Number of hired workers 110
Number of unemployed 317

A survey is planned to estimate acreage items with a cv of 2!% and numbers of workers
(excl uding unemployed) with a cv of 5% . With simple random sampling, how many units
are needed? How well would this sample be expected to estimate the number of unem-
ployed?
4.6 By experimental sampling, the mean value of a random variate is to be estimated
with variance V == 0.0005 . The values of the random variate for the first 20 samples drawn
are shown on p. 81. How many more samples are needed? (Use equation 4.7.)
4.7 A household survey is designed to estimate the proportion of families possessing
certain attributes. For the principal items of interest, the value of P is expected to lie
between 30 and 70% . With simple random sampling, how large are the values of n
necessary to estimate the following means with a standard error not exceeding 3%? (a) The
 TIlE ESTIMATION OF SAMPLE SIZE 87
Sample Value of Sample Value of
Number Random Variate Number Random Variate

I 0.0725 II 0.071 2
2 0.0755 12 0.0748
3 0.0759 13 0.0878
4 0.0739 14 0.0710
5 0.0732 15 0.0754
6 0.0843 16 0.071 2
7 0.0727 17 0.0757
8 0.0769 18 0.0737
9 0.0730 19 0.0704
10 0.0727 20 0.0723

over-all mean P7 (b) The individual means Pi for the income classes-under '5000; $5000
to SI0,000; over SI0,000. (i = 1, 2, 3)7 (c) The differences between the means (PI - PI) for
every pair of the classes in (b)? Give a separate answer for (a), (b), and (c) . Income statistics
indicate that the proportions of families with incomes in t"e three classes above are 50, 38,
and 12% .
4 .8 The 4-year colleges in the United States were divided into cla. ses of four different
sizes according to their 1952-1953 enrollments. The standard deviations within each class
are shown below.
Class

2 3 4

Number of students < 1000 1000--3000 3000--10,000 over 10,000

SI 236 625 2008 10,023

If you know the.class boundaries but not the values of SI, how wen can you guess the SI
values by using simple mathematical figures (section 4.7)7 No college has less than 200
students and the largest has about 50,000 stude nts.
4.9 With a quadratic loss function and a linear cost function, as in section 4.10, S 2 is
reduced to S'2 by a superior sampling plan, Co. CI> and A remainiPg unchanged. If n ', V'
denote the new optimum sample size and the accompanying V { y), show that n ' < n and
that V'< V.
4.10 If the loss function due to an error in yisA Iy - Y1 and if the cost C= co+c ,n, show
that with simple random sampling, ignoring the fpc, the most economical value of n is
(~\2/3
c,.n;;,J
4.11 (Adapted from Blythe, 1945). The selling price of a lot of standing timber is UW,
where U is the price per unit volume and W is the volume of timber on the lot. The number
N of logs on the lot is counted, and the average volume per log is estimated from a simple
random sample of n logs. The estimate is made and paid for by the seller and is
 88 SAMPLING TECHNIQUES

provisionally accepted by the buyer. Later, the buyer finds out the exact volume purchased,
and the seller reimburses him if he has paid for more than was delivered. If he has paid for
less than was delivered, the buyer does not mention the fact.
Construct th e seller's loss function . As~'uming that the cost of measuring n logs is en, find
the optimum value of n. The standard deviation of the volume per log may be denoted by S
and the fpc ignored.
4 .12 ·(a) The presence or absence of each of two characteristics is to be measured on
each unit in a simple random sample from a large population. If PI. P, are the percentages
of units in the population that possess characteristics 1 and 2, a client wishe to estimate
(PI - P, ) with a standard error not exceeding two percentage points. What sample size do
you suggest if the client thinks that PI a nd P, both lie between 40 and 60% and that the
characteristics are independently distributed on the units?
(b ) Suppose that in (a) the client thinks that the characteristics are positively correlated,
but does not know the correlation. You suggest an initial sample of 200, with the following
results.

Chal dcteristics
1 2 Number of units

Yes Yes 72
Yes No 44
No Yes 14
No No 70

200

What sample size do you now recommend to estimate (PI - P,) with a standard error s 2%?
4.13 (a) Suppose Y0U are estimating the sex ratio. which IS close to equality, and could
sample households of four persons father, mother, two child ren. Ignoring the small
proportion of families with identical twins, find the deff factor for a simple random sample
of /I households ver~us one of the 4n persons.
(b) Would identittal twin families lower or raise the deff factor?
 C HAPTER 5

Stratified Random Sampling

5.1 DESCRIPTION
In stratified sampling the population of N units is first divided into subpopula-
tions of N J, N 1 • .. . , N L units, respectively. These subpopulations are nonover-
lapping, and togethe r they comprise the whole of the population , so that
N 1 +N2 + " '+NL=N
The subpopulations are called strata . To obtain the full benefit from stratification,
the values of the Nh must be known . When the strata have been determined , a
sample is drawn from each. the drawings bei ng made independently in different
strata . The sample sizes within the strata arc denoted by n I, n 2, ... .n,..respec-
tively .
If a simple random sample is taken in each stratum, the whole procedure is
described as stratified random sampling.
Stratification is a ommon technique . There are many reason for this; the
principal ones are the following.

I.. If data of known precision are wanted for certain subdivisions of the
popu lation. it is advisable to treat each subdivision as a "population" in its own
right.
2. Administrative convenience may dictate the use of stratification; for exam-
ple , the agency conducting the survey may have field offices, each of which can
supervise the survey for a part of the population .
J . Sampling problems may differ markedly in different parts of the population .
With human popUlations, people living in institutions (e.g. , hotels. hospitals,
prisons) are often placed in a different stratum from people living in ordinary
homes because a different approach to the sampling i ' appropriate for the two
situations. In sampling businesses we may possess a list of the large firms, which
are placed in a separate stratum. Some type of area sampling may have to be used
for the smaller firms .
4. Stratification may produce a gain in precision in the estimates of characteris-
tics of the whole population . It may possible to divide a heterogeneous population
90 SAMPLING TECHNIOUES

into subpopulations, each of which is internally homogeneous. This is suggested

by the name strata, with its implication of a division into layers. If each stratum is
homogeneous, in that the measurements vary little from one unit to another, a
precise estimate of any stratum meap can be obtained from a small sample in that
stratum. These estimates can then be combinecl into a precise estimate for the
whole population.

The theory of stratified sampling deals with the properties of the estimates from
a stratified sample and with the best choice of the sample sizes nh to obtain
maximum precision. In this development it is taken for granted that the strata
have already been constructed. The problems of how to construct strata and of
how many strata there should be are postponed to a later stage (section SA.7).

S.2 NOTATION
The suffix h denotes the stratum and i the unit within the stratum. The notation
is a natural extension of that previously used . The following symbols all refer to
stratum h.

total number of units

number of units in sample
value obtained for the ith unit

stratum weight

sampling fraction in the stratum

true mean
II,

_
L
;- 1
Ylti
y,,==-- sample mean
nit

true variance

Note that the divisor for the variance is (Nit -1).

 STRA TlFlED RANDOM SAMPLING 91
5.3 PROPERnES OF THE ESTIMATES
For the population mean per unit, the estimate used in stratified sampling is Y.fI
(st for stratified ), where

(5 . J)

where N = N, +N2 +· .. +NL .

The estimate Y., is not in general the same as the sample mean. The sample
mean, y, can be written as
L

_
L nhYh
h- J
y=- - (5.2)
n
The difference is that in y" the estimates from the individual stra ta receive their
correct weights Nh! N. It is evident that Y coincides with y" proviqed that in every
stratum
nh
- =-
Ni,
or or
n N
This mean that the sampling fraction is the same in all strata. Thi stratification is
described as stratification with proportional allocation of the nh' It gives a
self-weighting sample. If numerous estimates have to be made, a self-weighting
sample is time-saving.
The principal properties of the estimate rtf
are outlined in the following
theorems. The first two theorems apply to stratified sampling in general and are
not restricted to stratified random sampling; that is, the sample from any stratum
need not be a simple random sample.
Theorem 5.1. If in every stratum the sample estimate Yh is unbiased, -then Y." is
an unbiased estimate of the population mean Y.
Proof.
L L
E (Y.,) = E L WhYh = L Wh Yh
h- J h- I

since the estimates are unbiased in the individual strata. But the population mean
Y may be written

N
This completes the proof.
92 SAMPLING TECHNIQUES

Theorem 5.2. If the samples are drawn independently in dlfierent strata,

L
V(Yst) = r W h V(Yh)
h- I
2
(5.3)

where V(Yh) is the variance of Yh over repeated samples from stratum h.

Proof. Since
I.
Ys, = r
h= 1
WhYh (5.4)

Ys, is a linear function of the jih with fixed weights Wh o Hence we may quote the
result in statistics for the variance of a linear function .
L L L
V(y.,) = r "'1/ V(YII) + 2 h-r I I r> h W W; Cov (YIIM
II - I
h (5.5)

But since samples are drawn independently in different strata, all covariance
terms vanish . This gives the result (5.3) .
To summarize theorems 5.1 and 5.2: if Yh is an unbiased estimate of Yh in every
stratum, and sample selection is independent in different strata, then Y." is an
unbiased estimate of Y with variance W,/ V(Yh). r
The important point about this result is that the variance of Y., depends only on
the variances of the estimates of the individual stratum means Yh • If it were
possible to divide a highly variable population into strata such that all items had
the same value within a stratum, we could estimate Y without any error. Equation
(5.4) shows that it is the use of the correct stratum weights N h / N in making the
estimate Y., that leads to this result.

Theorem 5.3. For stratified random sampling, the variance of the estimate jill
is

Proof. Since Yh is an unbiased estimate of YII , theorem 5.2 can be applied .

Furthermore, by theorem 2.2, applied to an individual stratum,
2
-) Sh Nfl - nil
V( Yh = - - - -
nh Nh

By substitution into the result of theorem 5.2, we obtain

2 2
1 1 Sh 2 Sit
V(Y") = N 2 rL

h= 1
2
Nit V(Yh) = N2 r L

11 - 1
Nh(Nh-nll)-=r Wh -(1 - 111)
nil nil

Some particular cases of this formula are given in the following corollaries.
 STRATIFIED RANDOM SAMPLING

Corollary 1. If the sampling fractions n,,/ NI, are negligible in all ~trata.

I " N,, 2SI,2 \~ Wh SIo

2 2
V( - )
YS( = N 2L.---=L. - - - (:').7 )
nil nil

This is the appropriate formula when finite population c()rrection~ ca n bc ignored

Corollary 2. With proportional allocation, we substitute

in (S.6) . Thc variance reduces to

V( -.
y"
)="NN SIon (NN-n) = 1-["
L.
I,
2

n
w L. 10 10
2
(S .H)

Corollary 3. If sampling is proportio nal a nd the variances in all strata have th e

same value, S..2. we obtain the simple res ult

V(ysr) = S:\N~n)
Theorem 5.4. If ¥s( = Ny" is th e estimate of the popul a tion IOtal )', then
2
S"
V( Y,,) = L NIo(NI, -
A

/11') -- (5 . I () )
/I"
This follows at o nee from theo re m S.3.

Example. Table 5 . 1 s h ow~ th e I l)20 and I <J3() number of inhabitants. in th o u ~and,. 01

M large ci ti es in th e U nited Sta tes. T he da ta were ttbtained by ta king the ci ti e~ wlll l'1l
ra nked fifth to sixt y-e ighth in th~ U nil ed States in lot al numher of inh abitants in 1920 . The
cities are arranged in two strat a, th e first cont ainin g th c 111 larges t eille, and th c 't:cond th c
re mai nin g 4K ci ti es.
The to ta l number of inhabitants in all 64 citi es in 1930 is to be cstima tcd from a sa mpl e of
sIze 24. Fi nd the standard error of the estim ated tot al for ( I) a si mpl e random ~a mpl c. (2) ~
stratified rando m sa mpl e with propo rti o nal a lloca tion , (3) a stratified random ~a mpl e w1l h
12 units drawn from each stratum .
This population resembles th e popul a tions of ma ny types of busin ess e nterprise in th at
some units-the large cities-contribute very substantiall y to th e to ta l a nd displa y mu ch
grea te r variability tha n the re mainder.
The stratum total s and sums of squares are given unde r Table 5. I. Only the 11)30 data lIrc
u. cd in thi example : the 1920 data appear in a later exa mpl e.
For the complete populatio n in 1930, we find
Y = 19.568, S' = 52,448
The three estimates of Yare denoted by Y,"., Y.,.op. a nd Y..,u.('
I . For simple random sampling
2 2
V(Y. ) = N S N - n = (64?(52,448)(40) = 5 594453
'". n N 2~ 64 ' ,
 94 SAMPLING TECHNIQUES

TABLE 5.1
SIZES OF64 CITIES (IN 1000's) IN 1920 AND 1930
1920 Size (xIl1) 1930 Size (VIII)

Stratum Stratum
h = I 2 I 2

797 314 172 121 900 364 209 113

773 298 172 120 822 317 183 115
748 296 163 119 781 328 163 123
734 258 162 118 805 302 253 154
588 256 161 118 670 288 232 140
577 243 159 116 1238 291 260 119
507 238 153 116 573 253 201 130
507 237 144 113 634 291 147 127
457 235 138 113 578 308 292 100
438 235 138 110 487 272 164 107
415 216 138 110 442 284 143 114
401 208 138 108 451 255 169 III
387 201 136 106 459 270 139 163
381 192 132 104 464 214 170 116
324 180 130 101 400 195 150 122
315 179 126 100 366 260 143 134

Note. Cities are arranged in the same order in both years.

Totals and sums of squares

1920 I 1930

Stratum Z(XM) Z (x/li2) Z (Yil') Z ('!1M2)

I 8,349 4,756,619 10,070 7,.145,450
2 7,941 1,474,871 9,498 2,141,720

from theorem 2.2, corollary 2. 'The standard error is

uCy..... ) =2365
2, For the individual strata the variances are

5,2= 53,843, S/= 5581

Note that the stratum with the largest cities has a variance nearly 10 times that of the other
stratum.
 . STRATIFIED RANDOM SAMPLING 95
In proportional allocation, we have II, = 6, "2 = 18. From (5 .7), multipJying by WZ, we
have ;
N-"
V( f',...p) = - - L N~Sh 2
"
= m(16)(53,843) +(48)(5581)] = 1,882,293
u(f'_) = 1372
3. For", ="2= 12 we use the general formula (5.9):
.
V( Y",,,,,,) = L Nh (Nh - "h) -
S/
"h
(16)(4)(53,843) (48)(36)(5581)
= 12 + 12 1,090,827

u(Y",,,,,,) =1044
In this example equal sample sizes in the two strata are more precise than proportional
allocation. Both are greatly superior to simple random sampling.

5.4 THE ES11MAlED VARIANCE AND

CONFIDENCE LIMITS
If a simple random sample is taken within each stratum, an unbiased estimate of
Sh 2 (from theorem 2.4) is
2
Sh =--1
1 n. (Yh;-Yh)
L
_ 2
(5.11)
"h - ;- 1

Hence we obtain the following.

Theorem 5.S. With stratified random sampling, an unbiased estimate of the
variance of in is

(5.12)

An alternative form for computing purposes is

L w; 2 2 L W; 2
s 2(9,,)= L ~- L __!!!!!_ (5.13)
h-I "h h- I N
The second term on the right represents the reduction due to the fpc.
In order to compute this estimate, there must .be at least two units drawn from
every stratum. Estimation of the variance when stratification is carried to the point
at which only one unit is chosen per stratum is discussed in section 5A.12.
The formulas for confidence limits are as follows.
Popu/~tio" mean: y., ± ts(9,,) (5 .14)
Population total: Ny., ± tNs(j.,) (5.15)
 96 SAMPLING TECHNIQUES

These formulas assume that y., is normally distributed and that s(y.,,) is well
determined, so that the multiplier t can be read from tables of the normal
distri bution.
If only a few degrees of freedom are provided by each stratum, the usual
procedure for taking account of the sampling error attached to a quantity like
s(y.,) is to read the t-value from the tables of Student's t instead of from the
normal table. The distribution of s(ys,) is in general too complex to allow a strict
application of this method. An approximate method 01 assigning an effective
number of degrees of freedom to s(y.,) is as follows (Satterthwaite, 1946).
We may write

The effective number of degrees of freedom n. is

n. =
(L gh Sh
2
Y
2

4 (5 .16)
L gh 5h
nh - 1

The value of n. always lies between the smallest of the values (nh - 1) and their
sum . The approximation takes account of the fact that Sh 2 may vary from stratum
to stratum. It requires the assumption that the YIII are normal , since it depends on
tht" result that the variance of s/ is 2ah 4/(nh - 1). If the distribution of Yhi has
positive kurtosis, the variance of Sl, 2 will be larger than this and formula 5.16
overestimates the effective degrees of freedom .

5.5 OPTIMUM ALLOCATION

In stratified sampling the values of the sample sizes nh in the respective strata
are chosen by the sampler. They may be selected to minimize V(YsI) for a specified
cost of tciking the sample or to minimize the cost for a specified value of V(y., ).
The simplest cost function is of the form
(5 .17)

Within any stratum the cost is proportional to the size of sample, but the cost per
unit Ch may vary from stratum to stratum. The term Co represents an overhead
co t. This costfunction is appropriate when the major item of cost is that of taking
the measurements on each unit. If travel costs between units are substantial,
empirical and mathematical studies suggest that travel costs are better rep-
resented by the expression !.th.r;;;,
where th is the travel cost per unit (Beardwood
et aI., 1959). Only the linear cost function (5.17) i considered here .
 STRATIFIED RANDOM SAMPLrNG 97
Theorem 5.6. In stratified random sampling with a linear cost function of the
form (5.17), Ithe variance of the estimated mean y" is a minimum for a specified
co. t C. and t,he cost is a minimum for a specified variance V(y.,) , when is "h
proportional t,o WhShlJC".
Proof. We have
L
C = co + l: Chnh (5 .]7)
11 = 1

(5.18)

Our problems are \ !ither (1) to choose the"h soas to minimize V for specified C, or
(2) to choose the"h so as to minimize C for specified V. It happens that apart from
their flnal steps, t he problem have othe same solution . Choosing the nh to
mi nimize V for fix( :d C or C for fixed V are both equivalent to minimizing the
product

V'C' = ( V+l: ~~~h)C- CQ)

= (r W~:S/)(r Ch"h) (5.] 9)

Stuart (] 954) has noted that (5 .19) may be minimized neatly by use of the
Cauchy- Schwarz ineqLlality. If aI!, bl! are two sets of L positive numbers, this
inequality comes from tthe identity

(5.20)

It follows from (5.20) that

(5.21 )

equality occurring if and only if bhl ah is constant for all h. In (5 .] 9) take

The inequality (5.2]) gives

Thus, no choice of the "h can make V'C smaller than (r WIlShJC"y. The
98 SAMPLING TECHNIQUES

minimum value occurs when

(5.22)

as stated in the theorem.

In terms of the total sample size nh in a stratum, we have

nh W h Sh l./0. N h Sh l./0.
(5 .23)
; '"" I (Wh s l./0.) ::: L (N S l./0.)
h h h

This theorem leads to the following rules of conduct. In a given stratum, take a
larger sample if

1. The stratum is larger.

2. The stratum is more variable internally.
3. Sampling is cheaper in the stratum.

One further step is needed to complete the allocation . Equation (5.23) gives the
nilin terms of n, but we do not yet know what value n has. The solution depends on
whether the sample is chosen to meet a specified total cost C or to give a specified
variance V for y." If cost is fixed, substitute the optimum values of nil in the cost
function (5.17) and solve for n. This gives

(C - co) L (NII SII /./0.)

n= (5.24)
L (NII Sh ./0.)
If V is fixed , substitute the optimum nh in the formula for V{ji.,) . We find

n ::=
(I WkS"Jc:)I W"SklJc:
(5.25)
V+(llN) L WhS,,2
where WII = NIII N.
An important special case arises if CII = e, that is, if the cost per unit is the same in
all strata. The cost becomes C = Co + en, and optimum allocation for fixed cost
reduce to optimum allocation for fixed sample size. The result in this special case
is as follows.
TIteorem S.7. In stratified random sampling VCr.,) is minimized for a fixed
total size of sample n if

(5.26)
 STRATIFIED RANDOM SAMPLING 99
This allocation is sometimes called Neyman allocation, after Neyman (1934),
whose proof gave the result prominence. An earlier proof by Tschuprow (1923)
was later discovered .
A formula for the minimum variance with fixed n is obtained by substituting the
value of nh in (5.26) into the general formula for Verst) . The result is

_ (I WSh) 2 I WhSh2
Vmin(y.,) = n N (5.27)

The second term on the right represents the fpc.

5.6 RELA nVE PRECISION OF STRA11FIED

RANDOM AND SIMPLE RANDOM SAMPLING
If intelligently used , stratification nearly always results in a smaller variance for
the estimated mean or total than is given by a comparable simple r.andom sample.
It is not true, however, that any stratified random sample gives a smailer variance
than a simple random sample. If the values of the nh are far frpm optimum,
stratified sampling may have a higher variance. In fact, even stratification with
optimum allocation for fixed total sample size may give a higher variance,
although this result is an academic curiosity rather than something likely to
happen in practice.
In this section a comparison is made between simple random sampling and
stratified random sampling with proportional and optimum allocation. This
comparison shows how the gain due to stratification is achieved .
The variances of the estimated means are denoted by v,."n> VfNOP ' and Vop"
respectively.
Theorem 5.S. If terms in 1/ Nh are ignored relative to unity,
Vop, $ Vprop $ V,an (5.28)
where the optimllm allocation is for fixed n, that is, with nh OCNhSh •
Proof.
(5.29)

(5.30)

[from equation (5.8), section 5.3)

(I WkSk)2
VoP, = - - - - (5.31)
n
[from equation (5.27), section 5.5)
 100 SAMPLING TECHNIQUES

, From the standard algebraic identity for the analysis of variance of a stratified
population, we have .
2 ~ - 2
(N - l)S = '" I (Yhi - Y)
/I i

= II
(Yhi - h )2+ y I Nh(Yh - 5')2
hilt

= I (Nh - 1)Sh 2+ I Nh (Yh - 5')2 (5.32)

h h

If terms in 1/ Nh are negligible and hence also in 1/ N, (5.32) gives

S 2::= I WhS/ + I Wh(Y,,- 5') 2 (5.33)
Hence

(5.34)

(I - I) ~ - - 2
= Vprop + - - " , Wh(Yh - Y) (5.35)
n
By the definition of V o"" we must have V prop <:!: V OPI' By (5.30) and (5 .31) their
difference is

(5 .36)

where S == I WhSh is a weighted mean of the Sh.

From (5 .35) and (5.36), with terms in 1/Nh negligible,
1 -2 (I - f) --2
Vran = V OPI +-n I Wh(Sh -S) +- n - I Wh (Y h - Y) (5 .37)

To summarize, in equation (5.37) there are two components of the decrease in

variance as we change from simple random sampling to optimum allocation. The
first component (term on the extreme right) comes from the elimination of
differences among the stratum means; the second (middle term on the right)
comes from elimination of the effect of differences among the stratum standard
deviations. The second component represents the difference in variance between
optimum and proportional allocation.
If terms in 1/Nh are not negligible, substitution for S 2 from (5.32) leads to
(1 - f) [ - - 2 1 2]
V",,. = Vprop + n(N - l) IN"(Y,, - Y) - NI (N - Nh)S" (5.38)

instead of to (5.35).
 STRATIFIED RANDOM SAMPLING 101
It fol:ows that proportional stratification gives a higher variance than simple
random sampling if

(5.39)

Mathematically, this can happen. Suppose that the Sh 2 are all equal to S.} , so that
proportional allocation is optimum in the sense of Neyman. Then (5.39) bcwme~
2: Nh(Yh - Y)2< (L - 1)Sw2
or

(5.40)

Those familiar with the analysis of variance will recognize this relation as implying
that the mean square among strata is smaller than the mean square within strata,
that is, that the F -ratio is less than 1.

5.7 WHEN DOES STRATIFICATION PRODUCE

LARGE GAINS IN PRECISION?
The ideal variate for stratification is the value of y itself- the quantity to be
measured in the survey. If we could stratify by the values of y, there would be no
overlap between strata, and the variance within strata would be much smaller than
the over-all variance, particularly if there were many strata. This situation is
illustrated by the example in section 5.3, page 94. The population consisted of the
sizes (numbers of inhabitants) of 64 cities in 1930, stratified by size. Although
there were only two strata, proportional stratification reduced the s.e. (y) from
2365 to 1372. Stratification with n 1 = n2 = 12, which is optimum under Neyman
allocation, produced a further reduction to 1044.
In practice, of course, we cannot stratify by the values of y. But some important
applications come close to this situation, and therefore give large gains in
precision, by satisfying the following three conditions.

1. The population is composed of institutions varying widely in size.

2. The principal variables to be measured are closely related to the sizes of the
institutions.
3. A good measure of size is available for setting up the strata.

Examples are businesses of a specific kind, for example, groceries (in surveys
dealing with the volume of business or number of employees), schools (in surveys
related to'numbers of pupils), hospitals (in studies of patient load), and income tax
returns (for items highly correlated with taxable income). In the United States
farms also vary greatly in size as measured by total acreage or gross income, but
102 SAMPLING TECHNIQUES

common farm items, such as the production of particular crops or types of

livestock, often exhibit only a moderate correlation with farm size, so that the
gains from stratification by farm size are not huge.
If the size of the institution remains stable through time, at least for short
periods, then its best practical measure is usual!y the size of the institution on some
recent occasion when a census was taken. The example in section 5.3 illustrates
the situation in which good previous data are available. Table 5.2 shows the Sh and
the resulting optimum nh oc NSh , when the allocation is made from 1920 and
1930 data, respectively.
The 1920 data indicate an n, of 11.56, as against a "true" optimum of 12.21 for
the 1930 data . When rounded to integers, both sets of data give the same
allocation- a sample size of 12 from each stratum .

TABLE 5.2
C"LCUL,tI.TION OF TH~ OPTIMUM ALLO ,tI.T!ON

1920 Data 1930 Data

Stratum Nh Sh NhS n" S" N"Sh nh

"
I 16 163 .30 2612.80 11.56 232.04 371 2. 64 12.:! 1
2 48 58 .55 2810.40 12.44 74 .71 3586.08 11.79

Totals 64 5423 .20 24.00 7298 .72 24.00

Note that the optimum sampling fraction is 75% in stratum 1 but only 25% in
stratum 2. It is often found that because of the high variability of the stratum
consisting of the largest institutions, the formula calls for 100% sampling in this
stratum. Indeed, the allocation may call for more than 100% sampling (see section
5.8). Note also that the Sh are smaller in 1920 than in 1930. The] 920 data give an
overoptimistic impression of the precision to be obtained in a 1930 survey. As
mentioned in section 4.7, the possibility of a change in the levels of the Sh should
always be considered when using past data, even though an allowance for change
may have to be something of guess.
Geographic stratification, in which the strata are compact areas such as counties
or neighbourhoods in a city, is common-often for administrative convenience or
because eparate data are wanted for each stratum. It is usually accompanied by
some increase in precision because many factors operate to make people living or
crops growing in the same area show similarities in their principal characteristics.
The gains from geographic stratification, however, are generally modest. For
example, Table 5.3 shows data published by Jessen (1942) and Jessen and
Houseman (1944) on the effectiveness of geographic tratification for a number of
typical -farm economic items.
 STRATIFIED RANDOM SAMPLING 103
Four sizes of stratum are represented- the township, the county, the "type of
farmi ng" area, and the state. To give some idea of the relative sizes of the strata,
there are about 1600 townships, 100 counties, and 5 areas in Iowa.
In the table the precision of a method of stratification is taken as inversely
proportional to the value of V(Y$I) given by the method. Thus the relative
precision of method 1 to method 2 is the ratio V2 (Y.,)/ V 1(y.,) , expressed as a
percentage. The data shown are averages over the numbers of items given in the
second column. The county is taken as a standard in each case. As indicated, the
gai ns in precision are moderate. In Iowa the use of 1600 strata (townships)
compared with no stratification (state) increases the precision by about 30% ; that
'5, it reduces the variance by about 25%.

TABLE 5.3
RELATIVE PRECISION OF DIFFERENT KINDS Of GEOORAPHIC
STRATIFICATION (IN PER CENT)
Stratum

Type of
No. of Farming
State Items Township County Area State

Iowa, 1938 18 115 100 96 91

Iowa, 1939 19 121 100 97 91
Florida, 1942
Citrus fruit area 14 144 100
Truck farming area 15 III 100
California, 1942 17 113 100 97

As regards proportional versus optimum stratification, there are two situations

in which optimum stratification wins handsomely. The first is the case, already
discussed, in which the population conSIsts of large and small institutions,
stratified by some measure of size . The variances Sk 2 are usually much greater for
the large institutions than for the small, making proportional stratification ineffi -
cient. The second situation is found in surveys in which some strata are much more
expensive to sample than others. The influence of the factor J0. may make
proportional allocation poor.
When planning an allocation in which the estimated nk do not differ greatly
from proportionality, it is worthwhile to estimate how much larger V(Y.,) or
V( 'Y.,) become if proportional allocation is used. The optima in the allocation
problem are rather flat (see section 5A.2) and the increa e in variance may turn
out surprisingly small . Moreover, the superiority of the optimum, as computed
from estimated values of the Sk, is always exaggerated because of the errors in the
estimated Sk' The simplicity and the self-weighting feature of proportional
allocation are probably worth a 10 to 20% increase in variance.
104 SAMPLING TECHN1QUES

5.8 ALLOCATION REQUlRIN(; MOllE THAN 100

PER CENT SAMPLING
As mentioned in section 5.7, the formula fOl'the optimum may produce an in "It
some stratum that is larger than the corresponding iNk' Consider the example on
city sizes in section 5.3. A sample of 24 cities, distributed between two strata,
called for 12 cities out of 16 in the first stratum and 12 out of 48 in the second . Had
the sample size been 48, the allocation would demand 24 cities out of 16 in the first
stratum. The best that can be done is to take all cities in the stratum, leaving 32
cities for the second stratum instead of the 24 postulated by the formula . This
problem arises only when the over-all sampling fraction is substantial and some
strata are much more variable than others. It has occurred in practice on several
occasions.
If the original allocation gives nl > NI when there are more than two strata, the
optimum revised allocation is:

(h 2:: 2) (5.41)

provided that ti h S Nh for h 2:: 2. If it should happen that n, > N 2 , we change the
allocation to
_ WhSh
nh = (n - N j - N2)-L - - ' (h ~ 3) (5.41 )'
L WhSh
3

provided that nh s Nh for h 2:: 3. We continue this process until every nh s N h. The
resulting allocation may be hown to be optimum for given n, as would I)e
expected.
Care must be taken to use the correct formula for V(y.,) . The general formula
(5.6) in section 5.3 is correct if the nh given by the revised optimum allocation
are sub tituted. Formula (5 .27) for V",in(YSI)'

(5 .27)

no longer holds. If I' denotes summation over the strata in which fi h < N h , an
alternative correct formula is

(5.42)

where" is the revised total sample size in these strata.

I
 STRATIFIED RAND OM SAMPLING 105
5.9 ESTIMATION OF ! 5AMPLE SIZE WITH
CONTINUO! US DATA

FormuLas for the determinatiotl' of n \ md r an estimated optimum allocation

were given in section 5.5. The present SI !ction presents formulas for any alloca-
tion, with some useful special cases. It is; lssumed that the estimate has a specified
variance V. If, instead , the margin of e 'rror d (section 4 .4) has been specified,
V = (d/ l f, where t i. the normal deviate corresponding to the allowable probabil-
ity that the error will exceed the desirel j margin.
Estimation of the Population Men Y
Let 5" be the estimate of Sh and let nh ,= w"n where the w" have been chosen. In
these terms the anticipated V(y.,) (fro m theorem 5.3, section 5.3) is
I~Wh 2S,, 2 1 2
V =- t... - ---- I WhS (5.43)
n \ v" N "
with W" = N,,/ N. This gives, as a gen leral formula for n,
tl I2 2
I~
w" (5.44)

If the fpc is ignored, we have, as a first approximation,

1 ~ 25 2
n, O=_I_h_h (5.45)
V Wh

If no/ N is not negligible, we ma~ I calculate n as

n =----''---- n (l
(5.46)

In particular cases the form ' alas take various forms that may be more conve-
nient for computation. A few are given.
Presumed optimum allocation (for fixed n): Wh ex:: Whs" .
(2:: WhSh )2
(5.47)

Proportional allocation: w" = W" = Nh / N.

I Wh S/t.2
n =' - -
no
(5.48)
no= V
1 + no
N
106 SAMPLING TECHNIQUES

Estimation of the Population Total

If V is the desired VCY.,), the principal formulas are as follows
General:

n (5.49)

Presumed optimum (for fixed n) :

(t Nhs h ) 2
n= 2 (5.50)
V+L Nhs h
Proportional:

no
n =-- (5 .51)
1+ nu
N

Example. This example comes from a paper by Cornell (1947), which describes :!
sample of United States colleges and universities drawn in 1946 by the U.S. Office of
Education in order to estimate enrollments lor the 1946- 1947 academic year. The
illustration is for the population of 196 teachers' colleges and normal schools . These were
arranged in seven strata, of which one small stratum will be ignored . The first five strata
were constructed by size of institution: the sixth contained colleges for women only.
Estimates Sh of the Sh were computed from results for the 1943- 1944 academic year. An
"optimum," stratification based on these Sh was employed .
The objective was a coefficient of variation of 5% in the estimated total enrollment. In
1943 the total enrollment for this group of colleges was 56,472. Thus the desired standard
error is
(0.05)(56,472) = 2824

so that the desired variance is

V = (2824)2 == 7,974,976

It may be objected that enrollments will be greater in 1946 than in 1943 and that
allowance should be made for this increase . Actually, the calculation assumes only that the
cv per college remains the same in J 943 and 1946-an assumption that may not be
unreasonable.
Table 5.4 shows the values of N h , Sh ' and Nhsh, which were known before determining n.
The appropriate formula for n is (5 .50), which applies to an "optimum" allocation for
estimating a total. With only 196 units in this population, it is improbable that the fpc will be
negligible. However, for purposes of illustration, a first approximation ignoring the fpc will
be sought. This is
 STRATIFIED RANDOM SAMPLING 107
TABLE 5.4
DATA FOR ESTIMATING SAMPLE SIZE
Stratum Nh Sh Nhs h "h
I 13 325 4,225 9
2 18 190 3,420 7
3 26 189 4,914 10
4 42 82 3,444 7
5 73 86 6.278 13
6 24 190 4,560 10

Totals 196 26.841 S6

Ad justment is obviously needed. For the correct n in (5.50), we have

n= no 90.34 57.1
1 +_!_" N s 2 1 +....;4.'--64_0...;...3_8_7
VL h h 7,974,976
A sample size of 56 was chosen ." Thr. nh for individual strata appear in the right-hand
column of Table 5.4.

5.10 STRA TlFlED SAMPLING FOR PROPORTIONS

If we wish to estimate the proportion of units in the population that fall into
some defi'led class C, the ideal stratification is attained if we can place in the first
stratum every unit that fall in C. and in the second every unit that does not.
Failing this, we try to construct strata such that the proportion in class C varies as
much as possible from stratum to stratum.
Let

be the proportions of units in C in the hth stratum and in the sample from that
stratum, respectively. For the proportion in the whole population, the estimate
appropriate to stratified random sampling is
~NhPh
(5.52)
P" =/...N
Theorem 5.9. With stratified random sampling, the variance of Pst is
V( )= J... ~ Nh 2(Nh - nh) PhQh (5.53)
P.. N 2 /... Nh - 1 nh
• The arithmetical results differ slightly from those given by Cornell (1947).
108 SAMP1..1NG TECHNIOUES

Proof. This is a particular case of the general theorem for the variance of the
estimated mean . From theorem 5.3
1 S 2
V(y.,) = N 2I Nh (Nh - nh )_h_ (5.54)
nh
Let Yhi be a variate which has the value 1 when the unit is in C, and zero otherwise.
In section 3.2, equation 3.4, it was shown that or this variate
2 Nh
Sh = N" _] P"O" (5.55)

This gives the result.

Note. In nearly all applications, even if the fpc is not negligible, terms in II Nh
will be negligible, and the slightly simpler formula

V(p.,) = N12I N" (N" - n"/"O,, =I w/ P"O" (l - I,,) (5.56)

11" nh
can be used.

Corollary 1. When the fpc can be ignored,

V(p., ) = I Wh2PhO" (5.57)

nIl

Coronary 2. With proportional allocation,

Vip )= N - n _1_I N"2P"O,, (5.58)
., N nN Nh - l
. t-f
=n-I W"PhO" (5.59)

For the sample estimate of the variance, substitute p"q"/(nh -1) for the
unknown P1,Oh/n" in any of tbe formulas above.
The best choice of the nIl in order to minimize V(p.,) follows from the general
theory in section 5.5.
Minimum Variance lor FixetfTotal Sample Size.

Thus

(5.60)
 STRATIFIED RANDOM SAMP ING 109
Mini mum Variance for Fixed Cost, where Cost:; Co + L Ch nh'

nh
Nh J PhOhl CII
=. n=-"--,:=::==:i;;;:: (5.61)
L Nh JPhOhlch
The value of n is found as in section 5.5.

5.11 GAINS IN PRECISION IN STRATIFfED

SAMPLING FOR PROPORTIONS
If the costs per unit are the same in all strata, two usef ul working rules are that
(a) the gain in precision from stratified random over simple random sampling is
small or modest unless the P" vary greatly from stratum to stratum. and (b)
optimum allocation for fixed n gains little over proportional all ocation if all Ph lie
between 0.1 and 0.9.

tABLE 5.5
R ELATI VE PR ECI SION OF STRAT IFI ED AND SIMPLE RANDOM SAMPLING

Simple Stratified

" V(p )i( 1 - /) n V(p , t)!(l - I) Relati ve

Ph = PQ = 1 IPhQh Precision ( %)

0.4, 0.5, 0.6 2500 2433 J03

0.3, 0.5, 0.7 2500 2233 112
0.2, 0.5 , 0 .8 2500 1900 J32
0. 1,0.5,0.9 2500 1433 174

To illustrate the first result, Table 5.5 compares stratified random samplinE
(proportional allocation) with simple random sa mpling for three strata of equal
sizes (Wh = !). Four cases are included , the fir t having Ph = 0.4, 0:5, and 0:6 in the
three strata and the last (a nd most extreme) having Ph = 0.1,0.5, and 0.9. The next
two columns show the variances of the estimated proportion , multiplied by
nl(1 - f), and the last gives the relative precisions of stratified to simple random
sampling. The gain in precision is large only in the last two cases.
To compare proportional with optimum allocation for fixed n, it will be found
that apa rt from the multiplier (I -f),

(5.62)
110 SAMPLING TECHNIQUES

The relative precision of proportional to optimum allocation is therefore

VapI _ a: W .JP;Q;,)2
h
(5.63)
Vprap - L WhPhOh
If all Ph lie between the two values Po and (1 - Po), we are interested in the
smallest value the relative precision will take. For simplicity, we consider two
strata of equal size (WI = W2 ). TIle minimum relative precision is attained when
t
P l = and P2= Po· The relative precision then becomes
VapI _ (0 .5+JP;Ool
(5.64)
Vprap - 2(0.25 + Po 00)
Some values of this function are given in table 5.6. Even with Po equal to 0.1, or as
high as 0.9, the relative precision is 94%. In most cases the :implicity and the
self-weighting feature of proportional stratification more than compensate for this
slight loss in precision.
The limitations of the example should be noted . It does not take account of
differential costs of sampling in different strata. In some surveys the Ph are very·
small, but they range from , say, 0.001 to 0.05 in different strata. Here there would
be a more substantial gain from optimum stratification.

TABLE 5.6
RELATIVE PRECISION Of PROPORTIONAL TO OPTIMUM ALLOCATION

0.4 or 0.6 C.l or 0.7 0.2 or 0.8 0.1 or 0.9 0.05 or 0.95

R P( ~~) 100.0 99.8 98.8 94.1 86.6

S.12 ESTIMATION OF SAMPLE SIZE WITIf

PROPORTIONS
Formulas can be deduced from the more general formulas in section 5.9. Let V
be the desired variance in the estimate of the proportion P for the whole
population. The formulas for the two principal types of allocation are as follows :
Proportional:
no
n = -- (5.65)
1+ flO
N
Presumed optimum :
_( LW"~) 2 (5.66)
no - V '
 STRATIFIED RANDOM SAMPLING 111
where no i the first approximation , which ignores the fpc, and n is the corrected
value taking account of the fpc. In the development of these formulas, the factors
Nh/(N h - 1) have been taken as unity.
These results apply to the estimate of a proportion. If it is preferable to think in
terms of percentages, the same formulas apply if Ph, qh, V, and so forth, are
expressed as percentages. For the estimation of the total number in the population
in class C, that is, of NP, all variances are multiplied by N 2 •

EXERCISES
5.1 In a population with N = 6 and L == 2 the values of y", are 0,1,2 in stratum 1 and 4 ,
Ii, II in stratum 2. A sample with n == 4 is to be taken . (a) Show that the optimum n" under
Neyman allocation, when rounded to integers, are nh = 1 in stratum 1 and nh == 3 in stratum
2. (b) Compute the estimate y" for every possible sample that can be drawn under
optimum allocation and under proportional allocation . Verify that the estimates arc
unbiased. Hence find Vapt(.Y,,) and Vprop(Y,,) directly. (c) Verify that Vu",(y,,) agrees with the
formula given in equation (5.6) and that Vprap(Y,,) agrees with the formula given in equation
(5.8) , page 93 . (d) Use of formula (5 .27), page 99 , to compute V"p,(y,,) is slightly incorrect
because it does not allow for the fact that the n h were rounded to integers. How well does it
agree with the corrected value?
5.2 The households in a town are to be sampled in oruer to estimate the average
amount of assets per household that are readily convertible into cash . The households are
stra tified into a high -rent and a low-rent stratum. A house in the high-rent stratum is
thought to have about nine times as much assets as one in the low-rent stratum, and Sh is
expected to be proportional to the square root of the stratum mean.
There are 4000 households in the high-rent stratum and 20,000 in the low -rent stratum .
(a) How would you distribute a sample of 1000 house holds between the two strata? (b) If
the object i to estimate the difference between assets per household in the two strata, how
should the sample be distributed?
5.3 The following data show the stratification of all the farms in a county by farm size
and the average acres of corn (ma;ze) per farm in each stratum. For a sample of 100 farms ,
compute the same sizes in each stratum under (a) proportional allocation. (b) optimum
allocation . Compare the precisions of these methods with that of simple random sampling.

Number of Average Standard

Farm Size Farms Corn Acres Deviation
(acres) Nh Ph Sh

0-40 394 5.4 8.3

41 - 80 461 16.3 13.3
81 - 120 391 24.3 IS.1
121 - 160 334 34.5 19.8
161 - 200 169 42.1 24.5
201-240 113 50.1 26.0
241 - 148 63 .8 35.2

Total or mean 2010 26.3

112 SAMPLlt 'G TECHNIQUES

5.4 Prove th.e result tated in formul a (5.38), section 5.6:

- J..!..:1l[~ N
V,a~ - Vp'aP+ n(N - I) r.... h
(Y - y)- 2-
h
_!_ ~ (N - Nh )5h 2J
NL...

5.5 A samlpler haS(wo strata with relative sizes W I' W2 • He believes that 5" 52 can be
taken as equa' l but thinks that C2 may be between 2c, and 4c, . He would prefer to use
proportional :allocaticm but does not wish to incur a substantial increase in variance
compared witl~ optimum allocation. For a given cost C= C,n, +c 2 n 2 , ignoring the fpc, show
that
V""",(y,,) = Wlc, + W2C2
Va,., (y,,) (WIFc; + W2~)2
If W, = W z, compute th e relative increases in variance from using proportional alloca-
tion wh en C2 / e, = 2. 4 .
5.6 A sa mple r proposes to take a stratified random sample . He expects that his field
costs will be of the form L Ch nh' His advance estimates of relevant quan tities for the two
strata are as follows .
Stratum

I 0.4 10 $4
2 0.6 20 $9

(a) Find the values of" 1/11 and "2/ " that minimize the total field cost for a given value of
V(Y"j . (b) Fi nd the sa mpl e size required, under this optimum allocation, to make
V(y,,) = J. Ignore th e fpc . (c) How much will the total field cost be?
5.7 After the sa mple in exercise 5.6 is taken, the sampler finds tHat his field costs were
actually $ 2 per unit in stratum 1 and $ 12 in stratum 2. (a) How much greater is the field .cost
than anticipated? (b) If he had known the correct field costs in advance, could he have
attained V(}',,) = I for the original estimated field cost in exercise 5.6? (Hint. The
Cauchy- chwarz inequality, pagc 97, with V' = I, gives the answer to this question without
finding the new allocation .)
5.8 In a stratification with two strata , the values of the Wh and 5h are as follows.

Stratum

I 0.8 2
2 0.2 4

mpute the ample sizes 11 , . tl 2 in the two strata needed to satisfy the following conditions.
Each cllse requires a separate computation. (Ignore the fpc.) (a) The standard error of the
estimated populatiotl mean 9., is to be 0.1 and the total sample size 11 = n, + fJ 2 is to be
minimized . (b) The standard error of the;: estimated mean of each stratum is to be 0.1. (c)
The standard error of the difference between the two estimated stratum means is to be 0.1 ,
again minimizing the total size of sample .
5.9 With two strata, a sampler would like to have n, = n 2 for administrative conveni-
ence, instead of using the values given by the Neyman allocation. If V(Y.. ), V.,.,(Y,,} denote
 STRA TIFlED RANDOM SAMPLING 113
(he variances given by the 1I J = n 2 and the Neyman allocations, respectively, show that the
fractional increase in variance '
V(y.,)- v.".(y,,) _ (r-l)2
V.".(y..,) r+1
where r = nl/ n2 as given by Neyman allocation. For the strata in exercise 5.8, case G, what
"I "2
would the fractional increase in variance be by using = instead of the optimum?
5. 10 If the cost function is of the form C = Co + L th rn:.,
where Co and the 'h are known
numbers, show that in order to minimize V(y.,) for fixed total cost nh must be proportional
to

(W::S/Y"
Fjnd the nh for a sample of size 1000 under the following conditions.
Stratum

1 0.4 1
2 0.3 2
3 0.2 4

5.11 If Vpn>p(y,,) is the variance of the estimated mean from a stratified random sample
of size " with proportional allocation and V(y) is the variance of the mean of a simple
random sample of size n, show that the ratio
VP"'P (ji.. )
V(y)
does not depend on the size of sample but that the ratio
VmI.(y.,)
V_(ji.,)
decreases as " increases. (This implies that optimum allocation for fixed n becomes more
effective in relation to proportional allocation as " increases.) [Use formulas (5 .8 and
5.27) .]
5.12 Compare the values obtained for V(P.,) under proportional allocation and
optimum allocation for fixed sample size in the following two populations. Each stratum is
of equal size. The fpc may be ignored.

Population 1 Population 2

Stratum Ph Stratum Ph

1 0.1 1 0.01
2 0.5 2 0.05
3 0.9 3 0.10

What general result is iUustrated by these two populations?

114 SAMPLING TECHNIQUES

5.13 Show that in the estimation of proportions the results corresponding to theorem
5.8 are as follows .
(1- f) 2
.V",. = V,..op +--I
n
W.(p. - P)

v '" V +I W. (./P;Q;, - .j"i''lJ;y

prop opt n

where

Jp.O. =I W.Jp.O•.
5.14 In a firm , 62% of the employees are skilled or unskilled males, 31 % are clerical
females, and 7% are supervisory. From a sample of 400 employees the firm wishes to
estimate the proportion that uses certain recreational facilities. Rough guesses are that the
facilities are used by 40 to 50% of the males, 20 to 30% of the females , and 5 to 10% of the
supervisors. (a) How would you allocate the sample among the three gr ups? (b) If the true
proportions of users were 48. 21 , and 4% , respectively, what would the s.e. of the estimated
proportion p" be? (c) What would the s.e. of p be from a simple random sample with
n = 400?
5.15 Formula (5 .27) for the minimum variance of y" under Neyman allocation reads as
follows .

A student comments : "Since I W.S. 2 > (:L WS. )2 unless all the S. are equal, the formula
must be wrong because as n approaches N it will give a negative value for V( y,, ). " Is the
formula or the student wrong?
5.16 By formula (5.26) for Neyman allocation , the sampling fraction in stratum h is
f. = ".1 N. = "S.I N:L ~Sl. The situations in which this formula calls for more than 100%
sampling in a stratum (f. > 1) are therefore likely to be those in which the overall samplKtg
fraction nl N is fairly substantial and one stratum has unusually high variability. The
following is an example for a small population , with N = 100, n = 40.
Optimum
Stratum N. S. n.

1 60 2 15
2 30 4 15
3 10 15 10

100 40

(a) Verify that the optimum n. aTe as shown in the right column. (b) Calculate V(.y,,) by
formula (5.6) and by formula (5.42) and show that both give V(y... ) = 0. 12.
C HAPTER SA

Further Aspects
of Stratified Sampling

SA.1 EFFECI'S OF DEVIATIONS FROM TIlE OPTIMUM

ALLOCATION
This chapter discusses a number of special topics in the practical use of stratified
sampling. Sections 5Al to 5A8, 5AIO, and 5A15 deal with problems that may
come up in the planning of the sample; the remaining sections deal with techni-
ques of analysis of results. The present section considers the loss in precision by
failure to achieve an optimum allocation of the sample.
Suppose that it is intended to use optimum allocation for given n. The sample
size nh' in stratum h should be
, n(WhSh)
nit = L WhSh (SAl)

From equation (5 .27), page 99 , the resulting minimum variance is

_ 1 (~ )2 1 ~ 2 (5A2)
Vmin (Yst) = -;;- L. WhSh - N L. WItSIt

In practice, since the Sit are not known, we can only approximate this allocation.
If nit is the sample size used in stratum h, the variance actually attained, from
equation (5.6), page 92, is

(5A3)

The increase in variance caused by the imperfect allocation is

__ (_)_~W/Sb2_.!.(~ )2
V (Y.,) V",," Y., - L. ~ L. WItSIt (5AA)
n~ n
In the first term on the right substitute for WitS" in terms of n,,' from (5Al). This
115
116 SAMPLING TECHNIQUES

gives the interesting result

_ _ _0: WSh)2(
V(y.,)- Vlnin(Y,,)- 2
"),,2
I-A--n
)

. n ~

= (I WkSk)~ I (nk - nk ,)2

(SA.S)
n2 nh

Reverting to equation (5A2), if the fpc (last term on the right) is negligible, we see
that
Vmu.(Y.rr) (I Wh Sh ) 2
(SA6)
n n2
Hence the proportional increase in variance resulting from deviations from the
optimum allocation is
V(,y,,) - Vmin(Y,,) .!_}: (nlo - nlo')2
(5A7)
Vlni" (,y.,) n h=I nh

where nil is the actual and nh' the optimum sample size in stratum h. If the fpc i
not negligible, the = sign in (5A.7) becomes 2::.
Let gh = lrih - "h '1/"10 be the absolute difference in the sample sizes in stratum h,
expressed as a fraction of the actual sample size nh. Then (5A7) becomes

(5AB)

a weighted mean of the g/. A conservative upper limit to (V- Vmin)/Vmln is

therefore g2, where g is the largest proportional difference in any stratum. Thus. if
8 == 0.2 or 20%, the proportional increase in variance cannot exceed (0.2)2 or 4%.
If g = 30% the proportional increase in variance is at most 9%. In this sense the
optimum can be described as Hat.

TABLE SA.l
EFFECTS OF DEVIATIONS FROM OPTIMUM ALLOCATION

Stratum
"h ' li h IliA - "h'l (ii A - "A')'
(opt) (act) Ii" IiA

I 200 150 0.33 16.7

2 100 120 0.17 3.3
3 40 70 0.43 12.9

Total 340 340 32.9

 FURTHER ASPECt'S OF STRATtPl£D SA)o4PLINO 117
Furthermore, (5A.8) suggests that the upper limit g2 win often overatimate tIM-. '
proportional increase in variance by a substantial amount. Table 5A.l gives an
example with three strata for n = 340. Optimum allocation requires sample sizes
of 200, 100, and 40, whereas the sizes actually used are 150, 120, and 70.
Since the value of g is 0.43 (stratum 3), the rough rule gives 18% as tho
proportional increase in variance. From t~e column on the right, the actual
increase is seen to be 32.9/340 = 9.7% . .
E vans (1951) examined the same question in terms of the effects of errors in the
estimated S" and developed an approximate rule showing whether an estimated
optimum is likely to be more precise than proportional allocation. He supp«*s
that the coefficient of variation of the estimated SIt is the same in all strata. This
assumption is appropriate when the Sit have been estimated from a preliminary
sample of the same size in each stratum. He shows how to compute the size of a
preliminary sample needed to make an "optimum" allocation better, on the
average, than proportional allocation. Previously, Sukhatme (1953-) showed' that a
smalJ initial sample usually gives a high probability that " optimum" allocation wiD
be superior to proportional stratification. See Sukhatme aDd Sukhatm,e (1910), p.
88. .

SA.2 EFFEcrs OF ERRORS IN TIlE STRATUM SIZES

For a desirable type of stratification, the stratum totals N~ may not be known
exactly, being derived from census data that are out of date. Instead of the true
stratum proportions WIt , we have estimates WIt . The sample estimate of Y is
Lw"y".
In general terms, the consequenoes of using weights that are in error are as
follows.

1. The sample estimate is biased. Because of the bias, we measure the accuracy
of the estimate by its mean square error about Y rather than by its variance aoout
its own mean (see section 1.9).
2. 'The bias remains constant as the sample size increases. Consequently, a size
of sample is always reached for which the estimate is less accurate than simple
random sampling, and all the gain in precision from stratification # lost.
3. The usual estimate s(Y.,) underestimates the true error of y.,. since it does not
oontain the contribution of the bias to the error. ' .

To justify these statements, note that in repeated sampling the mean valye of
r
the estimate is w" Y". The bias therefore amounts to
r(w" - W,,)Y,,
It is independent of the size of the sample. In finding the mean square error (MSE)
of the estimate, it is easy to verify that the varianoe term is given by the usual
118 SAMPLING TECHNIQUEf

formula , with Wh in place of Wh' Hence

2S 2
Wh h - 2
MSE(.Ysr ) = I - - (1- !,,)+ [I (Wh - Wh ) Yh ] (5A.9)
nh
This expression was given by Stephan (1941). Finally, the usual formula for
s\y.,) is clearly an unbiased estimate of the first term in (5A.9) but takes no
account of the second term. .

Example. This illustrates the loss of precision from incorrect weights when stratifica-
tion is (a) slightly effective, (b) highly effective. Consider a large population with S l = I,
divisi ble into two strat a with WI = 0.9, W2 ,., 0.1. We will assume SI = S2 = S~ . Then,
neglecting terms in 1/ N~ ,
S2 == I W"S,,2+ I W"(Y,, - Y)l (SA. IO)
= S,,2+ WI W 2(Y I - Y2)2
that is,
1= 5,,2 "'0 .09(Y I - y 2)2
In (a) tak(.; 9, - 9, = I. Then S'/ =0.91 , and proportional stratifica tion with correct
weights reduces th e var iance by 9% , compared with simple random sampling.
In (b) take 9, - Y z = 3, giving S" 2 = 0.19, a reduction in variance of more than 80%.
With two strata and incorrect weights, the bias may be written
(w l - WI )(91 - 92 )
since (w) - WI) = -( W 2 - W 2) . Suppose that the estimated weights are WI = 0.92 and
W 2 = 0.08 . The bia& amounts to (0 .02)( I) = 0.02 in (a) and to 0.06 in (b) . Hence we have the
following comparable MSE's for a sa mple of size II.

Simple random sa mpling: V(ji) = .!._

n
Stratified random sampling :
0.91
(a ): MSE(ji,,)= - +0 .0004
n
0.19
(b) : MSE(,ji,,) = - + 0.0036
n
As Table SA .2 shows, simple random sampling begins to win relative to (a) at n = 300.
There is little to choose be tween the two methods, however, up to n = 1000.
In (b). with more at stake, stratification is superior up to n = 200, although most of the
potenti al gain has already bee n lost at this sample size. Beyond n = 300 stratification
becomes markedly inferior to simple random sampling. Accurate estimation of the WI> is
particularly important when stratification is highly effective or when the sample size is
large.

In some surveys a large preliminary sample of size n' can be taken in order to
estimate the Wh oThis techClitIue, known as double sampling or two -phase sampl-
ing. has numerous applications and is discussed in Chapter 12. It will be shown
 FURTIfER ASPECTS OF STRATIFIED SAMPLING 119
TABLE 5A.2
COMPARABLE VALUES OF MSE(y)
Stratified Random
S'Impe
n Random (a) (b)

SO 0.0200 0.0186 0.0074

100 0.0100 0.0095 0.0055
200 0.0050 0.0049 0.0045
300 0.0033 0.0034 0.0042
400 0.0025 0.0027 0.0041
1000 0.0010 0.0013 0.0038

that with double sampling the mean square error of y., is approximately
L WhS/ L Wh(Yh - Y) 2
n + n' (5A.Il)

By comparing this MSE with S2/ n, as given by equation (5 A. ] 0), we see that most
of the gain from stratification is retained provided that n' is much greater than n.
To put it more generally, a set of estimated weights preserves most of the potential
gain from stratification if the weights are much more accurately estimated than
they would be from a simple random sample of size n.

SA.3 TIlE PROBLEM OF ALLOCATION WITIf MORE

THAN ONE ITEM
Since the best allocation for one item will not in general be best for another,
some compromise must be reached in a survey with numerous items. The first step
is to reduce the items considere~ in the allocation to a relatively small number
thought to be most important. If good previous data are available, we can then
compute the optimum allocation for each item separately and see to what extent
there is disagreement. In a survey of a specialized type the correlations among the
items may be high and the allocations may differ relatively little.

Example. Data given by Jessen (1942) illustrate a farm survey of this kind. The state of
Iowa was divided into five geographic regions, each denoted by its major agricultural
enterprise. Suppose that these regions are to be used as strata in a survey on dairy farming.
The three items of most interest are the number of cows milked per day, the number of
gallons of milk per day, and the total annual cash receipts from dairy products. From a
survey made in 1938, the estimated standard deviations 3~ within strata are shown in table
SA.3. In Table 5A.4 the optimum Neyman allocations based on these 3_ are given for the
individual items in a sample of 1000 farms.
 120 SAMPLING TECHNIQUES

TABLE 5A.3
STANDARD' DEVIATIONS WITHIN STRATA
S"
Receipts
$" s" for Dairy
CoWS Gallons Products
Stratum Milked of Milk (I)

Northeast dairy 0.197 4.6 11.7 332

Cash grain 0.191 3.4 9.8 357
Western livestock 0.219 3.3 7.0 246
Southern pasture 0.184 2.8 6.5 173
Eastern livestock 0.208 3.7 9.8 279

TABLE 5AA
SAMPLE SIZES WITHIN STRATA (n = 10(0)
Allocation

Optimum for
Average
Stratum Proportional Cows Gallons Receipts m"
Northeast dairy 197 254 258 236 250
Cash grain 191 182 209 246 212
Weste.rn livestock 219 203 171 194 189
Southern pasture 184 ]45 134 115 131
Eastern livestock 208 2]6 228 209 218

TABLE 5A.5
ExPECTED VARIANCES OF THE EsTIMATED MEAN
Type of allocation Cows Gallons Receipts

Optimum 0.0127 0.0800 76.9

Compromise 0.0128 0.0802 77.6
Proportional 0.0131 0.0837 80.9
I •

'The individual optimum allocations differ only moderately from each other. With one
exception, all three deviate in the same direction from a proportional allocation. Thus, in
the first stratum, proportional allocation suggests 197 fllfOlS. and the individual allocations
lead to numbers between 236 and 258. The average of the optimum sample sizes for tbe
three items, shown in the right-hand column, provides a satisfactory compromise alloca-
tion.
 FURTHER ASPECTS OF STRATIFIED SAMPLING 121
Table 5A.5 shows the expected sampling variances of y", as given by the individual
optima, the compror.lise, and the proportional allocations. The formulas are as follows.
a:: WhSh) ~ V =L (WhSh) l () = L W~h~
v.", = n ' comp mh ' prop n
The compromise allocation gives results almost as precise as if it were possible to use
~epara te optimum allocations for each item. What is more noteworthy is that proportional
allocation is only slightly less precise than the compromise or the individual optima.
Furthermore, Table 5A.5 overestimates the precision of the optima and of the compro-
mise, since these allocations were made from estimated variances. This result is another
ill ustration of the flatness of the optimum mentioned in section 5A.l.

SA.4 OmER MEmODS OF ALLOCATION WITH

MORE DIAN ONE ITEM
An alternative compromise allocation suggested by Chatterjee (1967) is to
choose the nIt that minimize the average of the proportional increases in variance
from (5 A. 7), taken over the variables. If j denotes a variable this amounts to
choosing

(SA.12)

where n lit is the optimum sample size in stratum h for variable j. For the data in
Table 5A.4, where the individual optima differ only slightly, Chatterjee's nit vary
from the average mIt in Table 5A.4 by, at most, one unit in any stratum.
In some surveys the optimum allocations for individual variates differ so much
that there is no obvious compromise. Some principle is needed to determine the
allocation to be used. Two useful ones suggested by Yates (1960) are presented.
The first applies to surveys with a specialized objective, in which the loss due to
an error of given size in an estimate can be measured in terms of money or utility,
as discussed in section 4.10. With k variates and quadratic loss functions, it may be
reasonable to express the total expected loss as a linear function of the variances of
the estimated population means or totals. For the means,

L =
k _ Ie
LJ aJV(YJ") = LJaJ"L WIt
L 22 (1 1)
SJ" - - '" (5A.13)
nit Hit

where Sf" is the variance of the jth variate in stratum h. Interchange of the order of
summation gives

L = L W/(}: a/Sf,,) -Nt}: W,,(}: a,9;,,) (5A.I4)

"n" j J
With a linear function for the costs of sampling, we have
C = co+}: c"n" (5A.I5)
122 SAMPLING TECHNIQUES

Minimizing the product of (C - co) and the first term in L (the term depending On
the n,,) gives, by the Cauchy- Schwarz inequality,
Wh ~
nh ex: r- V L a/Si" (5A.16)
vc" /
The constant of proportionality is found by satisfying the consrraint given for Lor
C. For instance, suppose that the value of L is specified and that the fpc term may
be ignored. We have

n(W"A"/£) (SA.l?)
n" = L (W"A"/ £ )

where A" = JL a;S;". The required total sample size is, from (SA. 14),
j

n =2.(L W~h)(L WltA,,£) (5A.18)

L h vc" h

In the second approach we specify the desired variance \') for each variate. For
population means this implies that

(j = 1, 2, ... , k ) (SA.19)

Inequality signs are used because the most economical allocation may supply
variances smaller than the desired \tj for some items.
!~ this approach the cost C [equation (SA.IS)] is minimized subject to the
tolerances "i and the conditions O:s n" :S N" . The problem is one in nonlinear
programming. Algorithms for its solution have been given by Hartley and
Hocking (1963), Chatterjee (1966), Zukhovitsky and Avdeyeva (1966), and
Huddleston ef al. (1970). Earlier, Dalenius (1957) gave an ingenious graphical
solution, while Yates (1960) and Kokan (1963) developed methods of successive
approximation , illustrated in the second edition of this book.
A useful first step is, of course, to work out the optimum allocation for each
variate separately and find the cost of satisfying its tolerance. Take the variate, say
Yh for which the cost C 1 is highest and examine whether the optimum nit values
for YI satisfy all the other (k - 1) tolerances. If so, we use this allocation and the
problem is solved, because no other allocation will satisfy the tolerance VI for Yt
at a cost as low as C I •
By working a series of examples in a related problem, Booth and Sedransk
(1969) have pointed out that in default of a computer program a good approxima-
tion to the solution of Yates' second problem can often be obtained by solving the
easier first problem. Specify that L in (SA.l3) shall have the value V* =L aJ "i,
where the \'} are the desired individual tolerances and the oJ are made inversely
 FURTHER ASPECl'S OF STRATIFLED SAMPLING 123
proportional to the \tj. Thus with two variates, al = V2 /(V. + V2 ) , a2 =
V1/( V1 + V 2 ) , and

(5A.20)

Example. (Follr strata, two variates.) The data and the application of the approximate
method are shown in columns (1) to (6) of Table 5A.6. The problem ino find the smallest n
[or which
V(y",) ;:;; 0.04, V(Y2U) s; 0.01

TABLE 5A.6
ARTIFI CIAL DATA FOR FOUR STRATA, Two VARIATES

C-e/umn (1) (2) (3) (4) (5) (6) (7)

Stratum W" S~" S~" A,,2= La,S:' WhA" n. n.

I

1 0 .4 25 5.8 0.963 206 194

2 0.3 25 4 8.2 0.859 183 180
3 0.2 25 16 17.8 0 .844 180 187
4 0.1 25 64 56.2 0 .750 160 171

Totals 3.416 729 732

By working out the optimum allocation for each variate separately it is easily verified thar '
n =625 is needed to satisfy the first constraint and n = 676 is needed to satisfy the second.
However, n = 676 with its allocation does not satisfy the fin>t constraint giving 0.0589
instead of 0 .04 for V" An iterative solution to satisfy both constraints (presented in the
n
second edition),_gave = 732, with the nh
shown in column (7) of Table 5A.6.
To use the Booth and Sedransk approach with V, = 0.04, V2 = 0.01 , we specify
2(0.04)(0.01)
L = 0.2 V(YI,,) +0.8 V(Y2,,) (0.05) 0.016

From (5A.18), with c. = I, we have

(L w~"r
n = - -L--
(3 .416)2 _ 9
0.016 - 72
using column (5) of Table 5A.6. From (SA.17), column (S) also leads to the nIl values,
shown in column (6) of Table SA.6. As columns (6) and (7) show, the two solutions nIl and
n. agree well.
As Booth and Sedransk note, n sri in all problems of this type, since n satisfies the single
constraint L = V*, but it need not satisfy the constraint on every variate, whereas the ri
allocation satisfies L as well as the individual constraints.
 124 SAMPLING TECHNIQUES

SA.5 1WO·WAY STRADnCADON wrm SMALL SAMPLES

Suppose tbat there are two criteria of stratilieation, say by R rows and C
columns, making RC cells. If n ~ RC, every cell can be represented in the sample.
A problem arises when n < Re, and we woukllike the sample to give proportional
representation to each criterion of stratification. In a simple method developed by
Bryant, Hartley, and Jessen (1960) the technique requires only that n exceed tbe
greater of R and C.
To illustrate this method, suppose that a small population of 165 school has
been stratified by size of city into five classes and by average expenditure per pupil
into four classes. The numbers of schools m" and the proportions of schools
P" = m,/165 in each of the 20 cells are shown in Table SA.7.

TABLE SA.7
NUMBI!R AND PROPOJlTlON Of ScHOOLS IN EACH Ceu.
Size: Expenditure per Pupil
of
City A B C D Totals nl.

mIl 15 21 17 9 mI . 62
I
PI I 0.091 0.127 0.103 o.O!s Pl . 0.376 4

m 21 10 8 13 7 nit. 38
II
PI I 0.061 0.049 0.079 0.042 PI. 0.231 2
.,
m,1 6 9 5 8 mao 28
III (1.055
Pal 0.036 0.030 0.049 Pa. 0.170 2

m.1 4 3 6 6 m•. 19
IV
p. J 0.024 0.018 0.036 0.036 p•. 0. 114 1

mas 3 2 S 8 ma. 18
V
P 6I 0.018 0.012 0.030 0.049 PI. 0. 109 1

Totals m.1 38 43 46 38 \65

p .I 0.230 0.261 0.278 0.231 1.000
n .1 2 3 3 2

The objective is.to give each school an approximately equal chance of selection
while giving each marginal class its proportional (epresentation. In this illustration
n = 10. Compute the numbers n,. ::II~. and n., .., nP./t where these products are
rounded to the nearest integers (with a further minor adjustment, if needed, so
that the nt. and the nJ both add to n). These numben are shown in Table SA.7.
 FUltniER AiP£CTS OF STRA11FIED SAMPLING 12S
The next step is to draw n ~10 cells with probability nl.n);'2 for the ijtb cell.
This is done by constructing an" x n square (Table SA.8). In row 1 one column is
drawn at random. In row 2 one of the remaining columns is drawn at random, and
so on. At the end, each row and column contains one unit. (This draw is most
quickly made by a random pennutation of the numbers 1 to 10.) The results of one
draw are indicated by X's in Table SA.8 .

TABLE SA.B
10 x 10 SQuARE FOR DRAWlNG THE SAMPLE

Column
I 2 3 4 S 6 7 8 9 10
Row A B C D

I x
2 I x
3 x
4 x

5 x
II
6 x

7 III x
8 x

9 IV x

10 V x

Note that columns 1 and 2 are assigned to marginal stratum A, since " .1= 2.
Similarly, rows 1 through 4 are assigned to marginal stratum I, since n I. =4, and so
on. This completes the allocation of the sample to the 20 cells. The allocation
appears in more compact form in Table SA.9. Two schools are drawn at random •
from the 15 schools in cell IA. and so on. The probability that a school in row i,
columnj is drawn is proportional to n,.n./ Pij • Thus the probabilities are not equal.
although they will be approximately so if P" =nl.n./n2.
An unbiased estimate of the mean per school is
_ 1 n 2 Pii
Yu=-t--YIJ
" "1."./
=
where YI/ is the sample total in the ijth cell. If, however. Pi! nl.n./n2. the sample
mean y is probably preferable. since its bias should be negligible. A sample
estimate of variance is avaHable for both the unbiased and biased estimates.
126 SAMPLING TECHNIQUES

TABLE SA.9
ALLOCATION OF THE SAMPLE 1'0 THE 20 CELLS
A B C D Total

I 2 I I 0 4
11 0 0 2 0 2
III 0 I 0 I 2
IV 0 I 0 0 1
V 0 0 0 I I

Total 2 3 3 2 10

provided that n is at least twice the greater of Rand C and that at least two units
are drawn in every row and column.
If Pi/ differs markedly from ni.n.;/n 2 in some cells. an extra step keeps the
probabilities of selection of schools more nearly constant. After computing the nl.
and n.b examine the quantities Dij = nPI! - nl.nj n, after rounding them to inte-
gers. If, in any cell. D lj is a positive integer, automatically assign D./ units to this
cell. Reduce n, the ni.• and the n.j by the amounts required by this fixed allocation
and carry out the remaining allocation as before.

SA.6 CONTROLLED SELECflON

Another technique for this problem with small samples was named controlled
selection by Goodman and Kish (1950). A simple illustration given by Hess,
Riedel, and Fitzpatrick (1976), who applied the method for sampling hospitals,
shows the basic idea. The principal stratification is by size of hospital (two strata).
Representation of each of two types of ownership of the hospitals is also desirable,
but only one unit (hospital) is to be drawn from each principal stratum. In

TABLE 5A.IO
ORDERING OF UNITS WITHJN STRATA FOR CONTROLLED SELEcnON
Original Order Revised Order
Stratum Stratum

I-Large Hospital II-Small Hospital Large Hospital Small Hospital

1 1 1 3'
2 2 2 4'
3' 3' 3' 5'
4' 4' 4' 1
5' 2
 f'UR11ffiR ASPBCTS OF STRATIFIED SAMPLING 127
numbering the units within strata, a prime (') indicates one ownership type, and
absence of a prime indicates the other type. Table 5A.l 0 (left side) shows the units
in the two strata.
If unit 1 or 2 is drawn from stratum I, we would like to draw unit 3',4', or 5'
from stratum II, so that both types of stratification are present with n = 2.
Similarly, 3' or 4' (stratum I) is desired with 1 or 2 (stratum II) . Controlled
selection makes the probability of these desired combinations as high as is
mathematically possible, while retaining equal probability selection within strata
and therefore unbiased estimates by the usual formulas for stratif}.ed sampling.
The purpose is either increased accuracy for given n or a saving in field costs.
With stratified random ampling the probability of a desired combination is
(.5)(.6)+(.5)(.4)=.5. This probability can be increased to .9 by two simple
changes in sample selection. Rearrange the units in stratum II so that the desired
combinations (3',4',5') with 1 and 2 in stratum 1 come first, as on the right in Table
SA.) O. Then draw a random number, between 1 and 100 and use it to select the
units from both strata. In stratum I, 1 s , s 25 selects unit 1, 26 s , s 50 selects
unit 2, and so on, so as to give each unit the desired one fourth probability of being
chosen. Similarly, in stratum II, 1 s , s 20 selects 3', 21 s , s 40 selects 4', and so
on. Hence, if 1 s , s20, we select (1,3'), if20 s , s 25, we elect (1, 4') and soon.
The joint selections and their probabilities are as foHows .
Pair: (1,3') (1,4') (2,4') (2,5') (3',5') (3',1) (4', 1) (4',2)
Probability .20 .05 .15 .10 .10 .15 .05 .20
The only nondesired combination is (3',5'). Thus the total probability of the
desired combinations is .90.
Since sampling is not independent in the two strata, the formulas for V(Y .. ) and
v(y,) do not apply. Hess, Riedel, and Fitzpatrick (1976) , give approximate
formulas . This monograph also gives an algorithm for the application of control-
led selection in problems with more strata, larger n, and more complex controls.
For another approach using balanced incomplete block designs, see Avadhani
and Sukhatme (1973) .

SA.' THE CONSTRUcnON OF STRATA

This topic raises several questions. What is the best characteristic for the
construction of strata? How should the boundaries between the trata be deter-
mined? How many strata should there be? For a single item or variable y the best
characteristic is clearly the frequency distribution of y itself. The next best is
presumably the frequency distribution of some other quantity highly correlated
with y. Given the number of strata, the equations for determining the be t stratum
boundaries under proportional and Neyman allocation have been worked out by
Dalenius (1957), and quicker approximat methods by several workers. We will
consider Neyman allocation, since it is usually superior to proportional allocation
128 SAMPLING TECHNIQUES

in populations in which gains from stratification are greatest. It is assumed at first

that the strata are set up by using the value of )' itself.
Let )'0, YL be the smallest and largest values of Yin the population. The problem
is to find intermediate stratum boundaries Y10 )'2, •• • , YL-l such that

(5A.21)

is a minimum. If the fpc is ignored, it is sufficient to minimize L W"SIo . Since )'It

appears in this sum only in the terms W"SIo and WIo + 1S" +h we have
o 0 0
-0 (L WS,, ) =-;-( W"S,,)+-(W" +l S" +I)
Ylt UYIo aYh
Now if f(y) is the frequency function of y,

W" = f Y.
>'h - I
f (t) dt, (5A.22)

Further,

(5A.23)

Diffe~entjation of (5A.23) gives

aw" as"
2
Sh -"- + 2 WSh-a = y" I(ylo) -
2
2)'1oJi-"f(Yh) + Ji-" 2f( y" ) (5A.24)
VYIt Ylt
where Ji-h is the mean of Y in stratum h. Add S" 2 iJ W"I ay" to the left side, and the
equal quantity S/ f (Yh ) to the right side. This gives, on dividing by 2SIt ,
a(WltS,,} S aWh Ivas,, _ l f( )(Y"_Ji-,,}2+ S,,2
h- - + I ' t ' h - - - y" ..:.:..:;.__;...;.;..:.._-,,- (5A.25)
0YIt aYh 0YIt 2 Sit
Similarly we find
0( W h + 1Sh + 1) _ -!f(YIt} (Ylt - Ji-,, +t)2+ S~+l (5A.26)
0YIt 2 S" +1
Hence the calculus equations for )lh are
(Ylt -1-',,)2+S/ (Ylt - Ji-" + 1)2 +S~ +~
(h =' I , 2, .. . , L - 1) (5A.27)
Sh Sh +J

Unfortunately, these equations are ill adapted to practical Computation, since

both Ji-h and Sit depend on )'It . A quick approximate method, due to Dalenius and
 FURTHER ASPECI'S OF STRATIflED SAMPLING 129
Hodges (1959), is presented for minimizing L W"S". Let

Z(y) == r Yo
.Jj{i) dt (5A.28)

If the strata are numerous and narrow, f(y) should be approximately constant
(rectangular) within a given stratum. Hence. .

I
y·
W h == f(t) dt == fh(YIt - Yh - t) (5A.29)
Yh - I

1
Sit == r.::(Yh - Yh - l) (5A.30)
v12
Y.

Zit - ZIt - l ==
J'
)'h - l
.Jj{i) dt == JJ,.(Yh - Y"-l) (5A.31)

where fh is the "constant" value of fey) in stratum h. By substituting these

approximations. we find
L L L
..m h-IL WhSh == ItL- I fh(Yh-Yh _I)2== L
11 - 1
(Zh- Z " _1)2 (SA.32)

Since (ZL - Zo) is fixed, it is easy to verify that the sum on the right is minimized by
making (Zit - Zit - I) constant.
Given f(y) . the rule is to form the cumulative of.JfCy) and choose the Yit SO that
they create equal intervals on the cum.Jf(y) scale. Table SA.II illustrates the use
of the rule.

TABLE SA.ll
CALCULATION OF STRATUM BoUNDARiES BY 1llE CUM ...;f<tl> Ruu
lndu,trial Loans Cum Industrial Loans Cum
Total Loans % fM "';f(y) Total Loans % fM VIM
0-5 3464 58.9 50-55 126 340.3
5- 10 2516 109.1 55-60 107 350.6
10-15 2157 155.5 60-65 82 359.7
15-20 1581 195.3 65-70 50 366.8
20-25 1142 229.1 70-75 39 373.0
25-30 746 256.4 75-80 25 378.0
30-35 512 279.0 80-85 16 382.0
35-40 376 298.4 85-90 19 386.4
4G-45 265 314.7 90-95 2 387.8
45-50 207 329.1 95-100 3 389.5
130 SAMPLING TE HNIQUES

Example. The data show the frequency distribution of the percentage of bank loans
devoted to industrial loans in a population of 13,435 banks of the United States
(McEvoy, 1956). The distribution is Skej with its mode at the lower end. In the cum.Jr
column, 58.9 = J3436. 109.1 = J3464 + 2516, and so on.
Suppose that we want five strata. Since the total of cum.Jr is 389.5, the division points
should be at 77.9. 155 .8,233.7, and 311.6 on this cale . The nearest available points are as
follows :
Stratum

2 3 4 5

Boundaries 0-5 0 0 5- 1S ~~ 15- 25· 0 2S-45 ~~ 45- 100 °:

Interval on cum vI 58 .9 96.6 73 .6 85.6 74 .8

The first two intervals. 58.9 and 96.6. are rather unequal. but cannot be improved on
without a finer subdivision of the original classes.
If the class intervals in the original distribution of yare of unequal length, a
slight change is needed. When the interval changes from one of length d to one of
length ud, the value of ..Jjfor the second interval is multiplied by.fU when forming
cum ..Jj.
Another method, propo ed by Sethi (1963), is to work out the boundaries given
by the calculus equations (5A.27) for a standard continuous distribution resem-
bling the study population. For the normal and various x 2 djstributions, Sethi has
tabulated the optimum boundaries for Neyman, equal, and proportional alloca-
tion fcit L s 6. If one of these distributions seems to approximate that in the study
population, the boundaries can be read from Sethi's tables.
Two further approximate methods require ome trial and error. From relatIOns
(5A.32), the Dalenius-Hodges rule is roughly equivalent to making WIrS/r con-
stant, as conjectured earlier by Dalenius and Gurney (1951). A similar rule is that
of Ekman (1959), who makes Wh(Yh - Yh - J) constant.
In comparisons on some theoretical and eight study populations, Cochran
(1961) found that the cum . ..Jjrule and the Ekman rule worked consistently well
(the Sethi method was not tried). In a study of United States hospital bed capacity,
whose distribution resembles X 2 with 1 degree of freedom, Hess, Sethi and
Balakrishnan (1966) found the Ekman method slightly superior to cum.lt and
Sethi's for L > 2, while Murthy (1967) also reports good performance by Ekman's
method.
The relations (5A.32) have an interesting consequence. If "",Sir is constant,
Neyman allocation gives a constant sample size nh = n/L in all strata. For the
approximate methods, the comparisons that have been made suggest that the
simple rule nil = n/ L is satisfactory.
Thus far we have made the unrealistic assumption that stratification can be
based on the values of Y itself. In practice, some other variable x is used (perhaps
 FURTHER ASPECT'S OF STRATIFlED SAMPLING 131
the value of y at a recent census). Dalenius (1957) develops equations for the
boundaries of x that minimize L WhS yh , given a knowledge 0( the regression of y
on x.If this regression is nonlinear, these boundaries may differ considerably from
those that are optimum when x itself is the variable to be measured. The equations
indicate, however, that if the regression of y on x is linear and the correlation
between y and x is high within all strata the two sets of boundaries should be
nearly the same. Let

y = a +f3x+e

where E(e) = 0 for all x and e, x are ul1correlated. The variance of e within
stratum his S'/' Then the x-boundaries that make V(y,,) a minimum satisfy the
equations (Dalenius, 1957).

f3 2[(Xh - /-LxIY + SXI>2] + 2S./ f32[(Xh -/-Lx.h+l)2+ S;.h+ l] + 2S;.h+l

f3S xh Jl +S./I f3 2S-x/ f3Sx.h+JJl +S;.h+)1 f3 2 S ;,h+l

If S;hl/32S ;h is small for all h, these equations reduce to the fofm (5A.27) that
gives optimum boundaries for x. But S ;hl/3 2S ;h = (1- Ph 2)1 p/ where Ph is the
correlation between y and x within stratum h.
Although more investigation is needed, this result suggests that the cum.Jj rule
applied to x should give an efficient stratification for another variable y that has a
linear regression on x with high correlation. Some numerical results by Cochran
(1961) support this conjecture. Moreover, if the Ph are only moderate, as will
happen when the number of strata is increased, failure to use the optimum
x-boundaries should have a less deleterious effect on y.
The preceding discussion is, of course, mainly relevant to the sampling of
institutions stratified by'some measure of size. The situation is different when one
set of variables is closely related to Y I and another set, with a markedly different
frequency distribution, is closely related to Y2' One possibility is to seek compro-
mise stratum boundaries that meet the desired tolerances on V(Ylst) and V(Y2.,),
following a general approach given in section 5A.4, but computational methods
have not been worked out.
In geographical stratification the problem is less amenable to a mathematical
approach, since there are so many different ways in which stratum boundaries may
be formed . The usual procedure is to select a few variables that have high
correlations with the principal items in the survey and to use a combination of
judgment and trial and error to construct boundaries That are good for these
selected variables. Since the gains in precision from stratification are likely to be
modest, it is not worthwhile to expend a great deal of effort in improving
boundaries. Bases of stratification for economic items have been discussed by
Stephan (1941) and Hagood and Bernert (1945) and for farm items by King and
McCarty (1941).
132 SAMPUNG TECHNIOUES

SA.S NUMBER OF STRATA

The two questions relevant to a decision about the number of strata L are (a) at
what rate does the variance of Y., decrease as L is increased? (b) How is the cost of
the survey affected by an increase in L?
As regards (a), suppose first that strata are constructed by the values of y. To
take the simplest case, let the distribution of y be rectangular in the interval
2
(a, a + d). Then S/, before stratification, is d /12, so that with a simple random
sample of size n, V(y) = d 2 /12n. If L strata of equal size are created, the variance
within any stratum is S;h = d 2 /12L 2. Hence, for a stratified sample, with W" =
l/L and nIl = n/ L,

V(y.,) =.!. (}: W"Sy,,)2 =.!. ( ~ .!.. ~ )2== ~ = V(~) (SA.33)

n " - I n II - I L v12L 12nL L
Thus with a rectangular distribution the variance of y., decreases inversely as the
square of the number of strata. Rather remarkably, this relation continues to hold,
roughly, when actual skew distributions with finite range are stratified with the
optimum choice of boundaries for Neyman allocation. In eight distributions of
data of the type likely to occur in practice, Cochran (1961) found that the average
values of V(Y.,)! V(y) were 0.232, 0.098, and 0.053 for L =2, 3, 4, as compared
with 0.2S0, 0.111, and 0.0(;2 for the rectangular distribution.
These results, which suggest that multiplicaton of strata is profitable, give a
misleading picture of what happens when some other variable x is used to
conStruct the strata. If q,(x) =E(ylx) is the regression of yon x, we may write
y =q,(x)+e (SA.34)
where q, and e are uncorrelated. Hence
S/=S,/+S/ (5A.3S)
By the preceding results, creation of L optimal strata for x may reduce S.2 to
S.2/ L 2 if q,(x) is linear or at a smaller rate if q,(x) is nonlinear. But the term S.2 is
not reduced by stratification on x. As L increases, a value is reached sooner or
later at which the term S/ dominates. Further increases in L will produce only a
trivial proportional reduction in V(y,,).
How quickly the point of diminishing returns is reached depends on a number of
factors-particularly the relative sizes of S/ and S.2 and the nature of q,(x). Only
a few examples from actual data are available in the literature. To supplement
them, a simple theoretical approach is used. Suppose that the optimum choice of
stratum boundaries by means of x, with samples of equal size n/L in each stra um,
reduces V(.i,,) at a rate proportional to 1/ L 2. Thus
L ~ 2 2 s/
V (i.,)=- '-' Wit SIC"=nL i (SA.36)
n "-I
 FURTImR ASPEcrs OF STRATIFIED SAMPLING 133
Suppose also that the regression of y on x is linear, that is,
y=a+fJx+e (5A.37)
2
where S. is constant. Then,

V(Y")=~
n
t
h- I
Wh2S~h=LfJ2
n
t
It - I
LS
W h2S;h+_ _e_2t
n It -I
W h2 (5A.38)

(5A.39)

where p is the correlation between y and x in the unstratified population.

With this model, Table 5A.12 shows V(YI')/ V(y) for p = 0.99,0.95,0.90, and
0.85 and L == 2 to 6, assuming that relation (5A.39) is an equality. The right-hand
columns of the table give V(YIl)/ V(ji) for three sets of actual data, described
under the table, in which x is the value·of y at some earlier time.
The results for the regression model indicate that unless p exceeds 0.95, little
reduction in variance is to be expected byond L = 6. Data sets 2 and 3 support this
conclusion, although some further increase in L might be profitable with the

TABLE SA.12
V(Yat)! V(ff) AS A FUNcnON OF L FOR THE LINEAR REGIlESSION MODEL AND FOR
SoME ACTUAl. DATA
Linear Regression Model Data, Set

p-
L 0.99 0.95 0.90 0.85 2 3

2 0.265 0.323 0.392 0.458 0.197 0.295 0.500

3 0.129 0.198 0.280 0.358 0.108 0.178 0.375
4 0.081 0.154 0.241 0.323 0.075 0.142 0.244
5 0.059 0.134 0.222 0.306 0.065 0.105 0.241
6 0.047 0.123 0.212 0.298 0.050 0.104 0.212
ex> 0.020 0.098 0.190 0.277

Type of Data
Set Data x '!I Source

1 College enrollments 1952 1958 Cochran (1961)

2 City sizes 194Q 1950 Cochran (1961)
3 Family incomes 1929 1933 Dalenius and Gurney (1951)
134 SAMPLING TECHNIOUES

college enrollment data (set 1). In two compari ons on survey data, Hess, Sethi,
and Balakri hnan (1966) found that V(Ysr) decreased faster with L than (5A.39)
predicts, which suggests that model (5A.37) is oversimplified.
To complete this analysis, we require a cost function that shows how the cost
depends on L. Dalenius (1957) suggests the relation C = Les + nen . The cost ratio
e.1 en will vary with the type of survey. An increase in the number of strata
involves extra work in planning and drawing the ample and increases the number
of weights used in computing the estimates, unless they are self-weighting. In
orne surveys almost no change is required in the organization of the field work; in
others a separate field unit is set up in each stratum . Whatever the form of the cost
function, the results in Table 5A. 12 suggest that if an increase in L beyond 6
necessitates any substantial decrease in n in order to keep the cost constant the
increase will seldom bc profitable.
The discussion in this section is confined to surveys in which only over-all
estimates are to be made . If estimates are wanted also for geographic subdivisions
of the population, the argument fOl a larger number of strata is stronger.

SA.9 STRATIFICATION AFTER SELECTION OF THE SAMPLE

(POSTSTRATIFICATION)
With some variables that are suitable for stratifiration, the stratum to which a
unit belongs is not known until the data have been collected. Personal characteris-
tics such as age, sex, race, and educational level are common examples. The
stratum sizes Nit may be obtainable fairly accurately from official statistics, but the
units can be classified into the strata only after the sample data are known . We
assume here that Wit , Nit are known .
One procedure is to take a simple random sample of size n and classify the units.
Instead of the sample mean Y, we use the estimate Yw = L WhYh, where Yh is the
mean of the sample units that fall in stratum h, and Wit = Nltl N. This method is
almost as precise as proportional stratified sampling, provided that (a) the sample
is reasonably large, say >20, in every stratum, and (b) the effects of errors in the
weights W h can be ignored (see section 5A.2).
To show this, let mit be the number of units in the sample that fall in stratum h,
where mit will vary from sample to sample. For samples in which the mit are fixed
and all mit exceed zero,

(5A.40)

The average value of V(Yw) in repeated samples of size n must now be

calculated . This requires a little care, since one or more of the mit could be zero. If
this happened, two or more strata would have to be combined before making the
estimate, and a less precise estimate would be produced. With increasing n, the
 FUR1HER ASPECTS Of STRATIfIED SAMPLING 135
probability that any mh is zero becomes so small that the contribution to the
variance from this source is negligible.
If the case in which mh is zero is ignored, Stephan (1945) has shown that to
terms of order n - 2

(5A.41)

Hence

(5A.42)

The first term is the value of V(Y.,,) for proportional stratification. The second
represents the increase in variance that arises because the mh do not distribute
themselves proportionally. But
1 _ ,
'2 L(1 - Wh )Sh-=- -
1 (L) Sh'
- - 21 L WhSh =-_I-Sh- -'2L
~ 2 1
2
Wh5 k
2
(5A.43)
n n n n nnh n

where Sh 2 is the average of the Sh 2 and fih = n/ L is the average number of units per
stratum. Thus, if the Sh 2 do not differ greatly, the increase is about (L - 1)/Ui h
times the variance for proportional stratification, ignoring the fpc . The increase
will be small if fin is reasonably large.
This method can also be applied to a sample that is already stratified by another
factor, for example, into five geographjc regions, provided that the! W h are known
separately within each regjon. Thi twofold stratification is widely employed in
U .S. National Survey: see Bean (1970) for a description of the estimation
formulas in the Health Interview Survey of the National Center for Health
Statistics.

SA.tO QUOTA SAMPLING

[n another method that has been u ed in opinion and market research surveys
the nh required in each stratum is computed in advance so that stratification is
proportional. The enumerator is instructed to continue sampling until the neces-
sary " quota" has been obtained in each stratum. The most common variables for
stratification are geographic area, age, sex, race, and some measure of economic
level. If the enumerator were to choose persons at random within the geographic
areas and assign each to his appropriate stratum, the method would be identical
with tratified random sampling. A considerable amount of field work would be
required to fill all quotas, however, since in the later stages most of the per ons
approached would fall in quota already filled.
To expedite the filling of quotas, some latitude is allowed to the enumerator
regarding the persons or households to be included. The amount of latitude varies
136 SAMPLlNG TECHNIQUES

with the agency but, in general, quota sampling may be described as stratified
sampling with a more or less nonrandom selection of units within strata. For this
reason, sampling-error formulas cannot be applied with confidence to the results
of quota samples. A number of comparisons betw~n the results of quota and
probability samples are summarized by Stephan and McCarthy (1958), who give
an excellent critique of the performance of both types of survey. The quota
method seems likely to produce samples that are biased on characteristics such as
income, education, and occupation, although it often agrees well with the proba-
bility samples on questions of opinion and attitude.

SA.ll ESTIMATION FROM A SAMPLE OF THE GAIN DUE

TO STRATIF1CATION
When a stratified random sample has been taken, it may be of interest, as a
guide to the conduct of future surveys, to appraise the gain in precision relative to
simple random sampling.
The data available from the sample are the values of N h, nh, Yh, and Sh 2. From
section 5.4, the estimated variance of the weighted mean from the stratified
sample is, by formula (5.13),
-) 't' w: 2
h Sh
2
't'"' hSh
II I 2
V ( Y.. =£... -- - -£... - N
nk
The problem is to compare this variance with an estimate of the variance of the
mea" that would have been obtained from a simple random sample. One
procedure sometimes used calculates the familiar mean square deviation from the
sample mean,
2 L (Yro - yfl
S = n- 1
where the strata are ignored. This is taken as an estimate of S 2, so that "(iran =
(N - n)sl / Nn for the mean of a simple random sample. This method works well
enough if the allocation is proportional, since a simple random sample distributes
itself approximately proportionally among strata. But if an allocation far from
proportional has been adopted, the sample actually taken does not resemble a
simple random sample, and this S 2 may be a poor estimator. A general procedure
is given, the proof being due to J. N. K. Rao (1962).
Theorem SA.t. Given the results of a stratified random sample, an unbiased
estimator of V",", the variance of the mean of a simple random sample from the
same population is
(N - n) [1
l.. Nk 2 -2
tn.
V"," =n(N - l) N~ nh Yhf- Y"+V(Y,,)
_ ]
(5A.44)

where v(y.,) is the usual unbiased estimator of V(Y,,) .

 FURTHER ASPECfS OF STRAnFIED SAMPLING 137
Proof·

(SA.4S)

Now -

(SA.46)

Also, since v(Y$1) and y" are unbiased estimators of V(y,,) and Y, respectively,
Ev(y,,) = V(y.,)=E(Y;,)- y2 (SA.47)

and hence an unbiased estimator of y2 in (SA.4S) is

y;,-v(y,,) (SA.48)
From (SA.46) and (SA.48) it follows that an unbiased sample estimator of V,an in
(SA.4S) is
(N - n)
V,an = n(N-I)
[1N~ Nh n.~Y~i-y;r+V(Y") ]
I.
nh (SA.44)

This proves Theorem SA.l.

With proportional allocation, (Nhl nh = Nln), the first two terms inside the
square brackets become (II n) times the within-sample sum of squares =
(n -1)s2 I n. Formula (SA.44) then reduces to

(N - n) [(n -l) 2 - ]
V'an = n(N-I) - n- S +v(y,,) (SA.49)

If n i large, (n - 1) ,; n, (N - 1) ,; N. and the term in v(y,,) is of order lin

relative to the term in S2 in (SA.49). Hence

. (N-n) 2
V,an =~S (SA.SO)

for proportional allocation. In the general case the corresponding simplification (n

large) is

(SA.Sl)

Example. The calculations are illustrated from the first three strata in the sample of
teachers' colleges (section 5.9). The data in Table 5A.13 are for the later 1946 sample. The
means represent enrollment per college in thousands. The S. 2 values are slightly higher than
in the second edition, owing to a correction.
138 SAMPLING TECHNIQUES

TABLE SA.13
BASIC DATA FROM A STRATIFIED SAMPLE OF TEACHERS'
COLLEGES

Stratum N. n. y. S.2 :·(r y~.)

1 13 9 2.200 1.8173 83.920
2 18 7 1.638 0.0735 49.429
3 26 10 0.992 0.0859 27.596

57 26 160.945

With n small we u e formula (5A.44). We find y" =- 1.4715. The values of the for the Y.,
sample were not reported, but the figures in the right-hand column can be obtained from
preceding columns of Table 5A.13. The formulas work out as follows.
_ ] N.(N. -n.)
v(y,,)=- ~r2L S. 2 =0.00497
1'< "h

31 [160.945 ]
V,on = (26)(56) - 5-7 --(1 .4715)2+0.00497 = 0.01412

Stratification appears to have reduced the variance to about one third of the value tor a
simple random sample, the estimated del! factor (section 4.11) being 0.00497/0.01412 =
0.35. ' f

SA.12 ESTIMATION OF VARIANCE WITH ONE UNIT

PER STRATUM
If the population is highly variable and many effective criteria for stratification
are known, stratification may be carried to the point at which the sample contains
only one unit in each stratum. In this event the formulas previously given for
estimating VO~:'r) and V(y,,) cannot be used . With L even, an estimate may be
attempted by grouping the strata in pairs thought beforehand to have roughly
equal true stratum totals. The allocation into pairs should be made before seeing
the sample results, for reasons that will become evident.
Let the sample observations in a typical pair be Y/h Yj2. where j goes from 1 to
L/2. Let itl = Nj1Y/l. Yiz = N j2Yi2 be the estimated stratum totals. Now
Yi1 - Yi2 = (Yjl - Y/ z) + (YiJ - Y/I) - (Yi2 - y/2) (5A.52)
Hence, averaging over all samples from this pair,
 f1JRlHER ASPECTS OF STRATIFfED SAMPLING 139
For VCYII ) consider the estimate

(5A.54)

By (5A.53) the expected value of this quantity is

_ L Ln
I +I (Y/l-Y/ 2f
2
BVl(Y")= N h (Nh - I)Sh (5A.55)
h- l ;- 1

The first term on the right is the correct variance (by theorem 5.4 with nh = 1).
The second term represents a positive bias, whose size depends on the success
attained in selecting pairs of strata whose true totals differ little. The form of the
estimate (5A.54) warns that construction of pairs by making the sample estimated
totals differ as little as possible can give a serious underestimate . The technique is
called the method of "collapsed strata."
With L odd, at least one group must of course, be of size different from 2. The
extension of the estimate (5A.54) to G groups of any chosen sizes L j 2: 2 is
A G Lj L, A ,. 2
Vl(Y.,)=I-1 I (Y/k-y//L j ) (5A.S6)
/ - 1 L / - k- I

where Y/ is the estimated total for group j. For L, = 2, when Y/ = }fl + Y,2, this
form agrees with (SA.S4). As with (S A.S4), the expectation of this VI(Y") gives
the correct variance V( Y.,), plus a positive bias foune by substituting Y/k and Y/
for Y/ k and Y, in (5A.56) .
When an auxiliary variate Ah is known for each stratum that predicts the
stratum total Y h , Hansen, Hurwltz, and Madow (1953) suggested the alternative
variance estimator

- 9. L, !:t -jk - 2
V2(Y.,) =j~l L /- I k~1 ( Y - A/kY;/A,) (SA.S7)

If An is a good predictor, the positive bias term in V2, coming from the
deviations (Y'k - Ajk Y;/ AY, is likely to be smaller than the corresponding term in
VI, although unlike Vb V2 also gives a biased estimate of the term in the S/ in
V(YII ). Hartley, Rao, and G. Kiefer (1969) found v21ess biased than VI in two of
three populations, with liftle difference in the third . .
These authors developed a method that does not involve the collapsing of
strata. This method uses one or more auxiliary variates Xlh, X2h, and so forth. 01.
which the true stratum means Yn are thought to have a linear regression. If Yh is
the sample value in stratum h, the method use the deviations

dh =Yh - Y- I b;(Xlh - .iJ (5A.58)

i

The variance-covariance matrix of the d h can be expressed as a linear function

of the 0-/, plus certain bias terms. By inverting this relation, estimates rJ/ are
140 SAMPLING TECHNIQUES

obtained, where u" = (N" - 1)Sh 2 / Nit, giving

2

(5A.59)

The method appears promising and extends to ratio estimates, but the authors
warn that additional comparisons with " collapsed strata" methods are needed.
Using a different approach, Fuller (1970) developed a method of stratum
construction that provides an unbiased sample estimate of V(Y,,) with one unit
per stratum. For simplicity suppose that N/n =N/L = k (an integer). Select a
random number r between 1 and k. The first stratum consists of the units
numbered from (r + 1) up to (r + k) , the second those numbered from (r + k + 1) up
to (r+2k) , and so on, the last (Lth and nth stratum) those numbered from
r + (n - I)k + 1 to N = nk and those from 1 to r. At first sight, this last stratum may
look a poor choice. As Fuller notes, however, this method can work well in
geographic stratification with areal units. Here, stratification usually leans on the
notion that units near one another tend to be similar. By numbering units in
serpentine fashion , one can have YN near YI> so that the stratum that includes both
YN and Yl an also be internally homogeneous. The estimate v(Y,,) is a weighted
sum of the differences (Yh - Yh +I) 2.
The circular method would be less effective for a population showing a rising
trend from YI to YN, in which the stratum including both YI and YN would have
large internal variability. For this situation Fuller gives a second plan, slightly
more complex, which should give good precision with a rising trend and also
furnishes an unbiased estimate of V( Y.,).
"

SA.13 STRATA AS DOMAINS OF S11JDY

This section deals with surveys in which the primary purpose is to make
comparisons between different strata, assumed to be identifiable in advance. The
rules for allocating the sample sizes to the strata are different from those that apply
when the objective is to make over-all population estimates. If there are only two
strata, we might choose nl, n2 to minimize the variance of the difference (ji) - ji2)
between the estimated strata means. Omitting the fpc's for reasons given in
section 2.14, we have
(SA. 60)

With a linear cost function

(SA.61)
V is minimized when

(5A.62)
 FURTII.E R ASPECTS OF STRAnFIEO SAMPLING 141
With L strata, L > 2, the optimum allocation depends on the amounts of precision
desired for different comparisons. For instance, the cost might be minimized
subject to the set of L(L -1)/2 conditions that V(Yh - y;)s V hh where the values
of V hl are chosen according to the precision considered necessary for a satisfac-
tory comparison of strata hand i.
Frequently a simpler method of allocation is adequate, especiaUy if the Sh and
Ch do not differ greatly. One approach is to minimize the average variance of the
difference between all L(L -1)/2 pairs of strata, that is, to minimize
2 (S
V- =- / S/ ... +-
-+-+ SL2) (SA.63)
L nl n2 nL
V is minimized, for fixed C, by the rule in (5A.62),
Sh
nhOC- (5A.64)
.£
This rule may result in certain pairs of strata being more precisely compared and
others less precisely than is felt appropriate. An alternative is to select the nh so
that the s.e. of the difference is the same, say ./V, for every pair of strata. This
amounts to making Sh 21nh = VI2 for every stratum. For a fixed cost this method
gives less over-all precision than the first method. The reader may verify that the
two optimum allocations give
2('[. Sh JZ.)2
v- =-=,-::,,-___..:.::_ v::= 2('[. S/Ch) (SA.6S)
L(C-co) , (C -co)
It follows from the Cauchy-Schwarz inequality that V is always greater than ii
unless Sh£. = constant. If V is substantially greater than ii, a compromise
allocation can sometimes be found, after a little trial and error, that will give an
average variance close to ii and also keep V(Yh - y,) reasonably constant.
Sometimes the objective is to obtain estimates for each stratum as well as
over-all estimates for the whole population. In planning the survey, we might
specify the following conditions.
S2
V(Yh) = _ h (I - In) s V h ,
nh

The fpc terms are now included, since the purpose is to specify the precision with
which the means in the finite population are to be estimated. The conditions on the
V(Yh) determine lower limits to the values of the nh. If these lower limits are found
to satisfy the condition on V(Y.,), the allocation problem is solved. When the
condition on V(y,,) is not satisfied, Dalenius (1957) has indicated a graphical
approach.
More complex problems arise when the L =2k strata represent all combina-
tions of k factors each at two levels, and the objective is to estimate the average
effects of the factors. If the stratum or cell to whicltany member of the population
142 SAMPLING TECHNIQUES

belongs is known in advance of sampling, a sample of desired size nIl can be

drawn from stratum h. For 2, 3, or 4 factors Sedransk (1967) has given methods
for finding the nh that minimize the, cost under different specifications about the
variances of the estimated average et,fects of the factors and the desired power of a
test for interactions.

SA.14 ESTIMATING TOTALS AND MEANS OVER

SUBPOPULATIONS
Frequently the subpopulations or domains of study are represented in all strata.
If stratification is geographic, for example, separate estimates may be wanted,
over the whole population, for males and females, for different age groups, for
users and nonusers of Blank's toothpaste, and the like. The problem presents
orne complications. The basic formulas were given by Yates (1953) with further
discussion and proofs by Durbin (1958) and Hartley (1959). Methods applicable
to a single stratum are discussed in sections 2.10 and 2.11.
The following notation applies to the units in stratum h that lie in domain j.
Notation.

Number in sample: nhl' I nloi = nlo

Measur,ement on individual unit: Yh"

Sample mean :

N,(
Domain mean : i\ = L Y}u,
1=1 Nhl
The population total and mean for domain j over all strata are, respectivel)\,
- - Y,
Yi = INhjYIt" Yj =-
h N,
where N; =I N h /.
II
The complication arises because the nh, are random variable . If the N,,/ were
known, the problem would be simple. As estimates of Yi and 9" we could use
... y'
'9:.'=-'
, N;
By the method in ection 2.12, the ordinary formula for V(Y',,) is still valid,
provided all nil, > O. Thus
 FURTHER ASPECTS OF STRATIfIED SAMPLING 143

VCY/) = L N~IS~f (l _ nhf ) (5A.66)

h nh/ Nhf
wher~ / the variance among units in domain j within stratum h. In applica-
tions,: ho ever, the Nhf are rarely known .

Estimating Domain Totals

In default of the N h" each stratum total of the domain is estimated as in section
2.13 . These totals are added to obtain an estimated domain total , that i ,

(5 A.67)

The true and estimated variance of ~ are found hy the device used in section
2.13. A variate Y~, is introduced that equals Yhl, for all units in domain j and equals
zero for all other units in the population. As shown in section 2.13 thi s gives for the
estimated variance

(5A.68)

Estimating Domain Means

In order to estimate the domain mean Y/ N,. a ample estimate of N, is
required. An unbiased estimate is

(5A.69)

Hence we take

(5A.70)

With proportional tratification, Yf reduces to the ordinary sample mean of the

units that fall in domain j . In the general case, this estimate is known as a combined
ralio estimate, discus oed later in ection 6.11. To show it, introduce another
dummy variate x~{ which equals I for every unit in domain j and 0 for all other
units, where i now goes from} to N k • Clearly,

(5A.?})
 144 SAMPLING TECHNIOUES

SO that the estimated domain mean may be written

r. (Nni nit) r. Yltij r. N"y,,'
'9, = " I -=0"_ _
(SA72)
I r. (N,,/nlt)nltj r. N"i,,'
h h

This is the formula for the combined ratio estimate for the two variables Yh/ and
Xh/' From section 6.11, the estimated variance may be expressed approximately
as
(SA.73)

The second summation may be written

~ 2 "~ 2
~
r.l
(Yhl - ljx"d - nlt(Yh - Y;x,,) = r. (Yltii -
nitro , .0.

i
, 2 _, _ , .0. 2 nh · _
lj) -:.::!!L(YltJ - Yj)
n"
2
(SA.74)

using (SA71). Furthermore, the first term in (5A74) can be expressed alterna-
tivelyas

Inserting these results in (SA73) gives! finally , for the estimated variance,
~) ' 1 '\' Nh 2(1_ fh) ['\'
V(Yj =N.~2t..
_ 2 (
nhi) _ ~
( -1) t..(Yhl/ -Yhi) +nhi 1 - - (Yh/-Yi )
2] (SA.7S)
j h nit nit I nit
The .•term on the right represents a between-stratum contribution to the
variance. Differences among strata means are not entirely e:liminated from the
variance of the estimated mean of any subpopulation. The between-stratum
contribution is small if the terms 1- nit;! nh are small, that is, if the subpopulation
is almost as large as the complete population.
As Durbin (1958) has pointed out, (5A75) applies also to means estimated for
the whole population, if the sample is incomplete for any reason such as 110n-
response, provided, of course, that Yj is the estimate used. In this event Yj is
interpreted as the estimated mean for the part of the population that would give a
response under the methods of data collection employed. There is, however, an
additional complication, in that the "nonreponse" p¥t of the population often has
a different mean from the "response" part. Thus Yj is a biased estimate of the
mean of the whole population, and this bias contribution is not included in
(SA75).

SA.IS SAMPLING FROM 1WO FRAMES

An early example of the use of a sample from a list B of large businesses in
combination with a sample from an areal frame A that covers the complete
population is the 1949 sample survey of retail stores taken by the Census Bureau
 FURTHER ASPECTS OF STRATIFIED SAMPLING 145
and described in Hansen, Hurwitz, and Madow (1953, p. 516 ff.). The objectives
. in this combined use of an incomplete list frame and a complete areal frame were
to gain increased accuracy and save money. Large businesses can sometimes be
sampled cheaply by a combination of mail and telephone; moreover, being large,
they are often the businesses that have the largest variance for the y variables
being measured. Placing them in a eparate stratum with optimum allocation (or
100 per cent sampling if this seems close to optimum) can produce substantial
increases in accuracy. In the retail stores survey, businesses in the list frame that
were present in the area sample were identified and removed from the area
sample, so that the population being sampled fell into two distinct strata. This
process, called unduplication, was performed by the field supervisor in advance of
sampling. Unduplication is sometimes complicated and subject to errors. Some of
the practical difficulties are discussed by Hansen, Hurwitz and labine (1963), as
well a:; various ways in which incomplete Ii ts can be helpful.
Identification of the sample members from the frame A sample that belong to
the list frame B sometimes requires measurement of the y variables for these
members. The sampler then has three samples at his disposal in which y has been
measured-a sample from stratum a (the part of A that does not belong to B) and
two independent samples from stratum B. One, denoted by the suffix ab, is the
sample obtained from A and identified as belonging to B, and one is obtained by
direct sampling of frame B. With frame A complete and simple random sampling
from both frames, Hartley (1962) proposed that both B samples be used in the
poststratified estimate
(5A.76)

where Ya, Yab, YB denote the respective sample means. The weighting factors p and
q for the two samples that belong to frame B, with P + q = 1, are chosen to
minimize V( Y) under a cost function of the form

(5A.77)

In (5A. 76) the stratum sizes Nab = NB and Na :; (N.II. - N B ) will, of course, be
known.
With SB 2 >S/ and CB CA , Hartley showed that this method can give large
reductions in V(n as compared with sampling from frame A only, even if the
frame A sample is poststratified into the two strata a and B =abo
The problem becomes more difficult if frame A is also incomplete, two frames
A and B, with some duplication, being required to obtain complete coverage of
the population. For poststratification, there are three distinct strata: 4 (units in A
alone); ab (units in both A and B); and b (units in B alone). The three strata
cannot .be sampled directly, samples of sizes nA, nB having to be drawn from
frames A and B. Furthermore, the strata izes N a, Nab, Nb will not usually be
known. For simple random sampling from frames A and B, Hartley (1962)
146 SAMPLING TECHNIQUES

suggested the estimate

A NA NA N8 No
Y= - Ya +P-Yab +q-Yb<r +-Yb (SA.78)
nA nA nB nB

where, as before, P + q = 1 and the y' are sample totals in the strata, the suffixes
ab and ba denoting the samples in the duplicate stratum found in nA and nB.
Hartley determined P and q to minimize V( y) for fixed cost. Improvements in
Hartley's estimate (SA.78) have been given by Lund (1968) and Fuller and
Burmeister (1972), essentially by using better estimates of Na , Nab, Nb than are
implied in Hartley'S estimate. Fuller and Burmeister (1972) also dealt with the
case in which frame A i areal , with subsampling of the areal units. Hartley (1974)
gives a general approach to two-frame sampling, applicable to any sample design
in the two frames.

" EXERCISES
5A.I In planning a survey of sales in a certain type of store, with n =550. good
estimates of Sh are available from a previous survey in two of the three strata. The third
stratum consists of new stores and stores that had no sales in the previous survey, so that a
value for S3 has to be guessed. If S3 is actually 10, compute V (,y,,) as given by an estimated
Neyman allocation when S, is guessed as (a) 5, (b) 20. Show that in both cases the
proportional increase in variance over the true optimum is slightly over 2% .
True Estimated SA
Stratum WA SA (0) (b )

I 0.3 30 30 30
2 0.6 20 20 20
3 0.1 10 5 20

~A.2 Show that if all Sh, except St.. are correctly estimated and St. is estimated as
5t. = SL(I + A), tfte proportional increase in V.",(Y,,) , using 5, instead of the true 5 L for
Neyman allocation, is
A 2ni.(n - nU
U+A)n 2
where hi is the sample size in stratum L under true Neyman allocation. Verify that this
formula agrees with the results in exercise 5A.I. (The agreement is not exact because of the
rounding of the Ilh to integers.) Hence show that aSO% underestimation of SL has the same
effect as 100% overestimation.
SA.3 If there are two strata and if 41 is the ratio of the actual nl/n 2 to the Neyman
optimum n J 1l 2' show that whatever the values of N .. N2• S .. and 52' the ratio
Voni.(Y,, )/ V(Y.. ) is never less than 441(1 + 41 )2 when the fpc's are negligible .
5A.4 The results of a simple random sample with n ... 1000 can be classified into three
"strata," with Yh=10.2 . 12.6, and 17.1. s/=10.82 (the same in each stratum), and
J~ - 17 .66. The estimat d strlltum weight are Wh -O.S, 0.3, 0.2, respectively. Th~
 FURnlER ASPECJ'S OF STRATIFIED SAMPLING 147
weights are known to be inexact, but it is thought that all are correct within 5%, 80 that the
worst cases are either W~ =0.525, 0.285,and 0.190 or Wh =0.475, 0.315, and 0.210. By
the methods of section 5A2, would you recommend stratification1 (Where needed, assume
that y~ = y~ and s~ 2 = S~ 2.)
5A5 In a stratified random sample with two variates the objective is to satisfy the
specifications

for minimum cost C= r c~nh' The fpc's can be ignored. (a) Prove the result by Chatterjee
(1972) that a compromise aUocation is necessary if

r W~S2AJc;; V r
Wh(Si,jSlk)Jc;;
-"'-----'--"'-'----=5 - 2 5=--~~:::...;...:.:;;,....;;..::
r W.. (S~,jS2~>Jc: VI r
W"SI~Jc;; .
(b) If V 2 / VI equals or exceeds the upper limit, the optimum aUocation for y, satisfies
both tolerances, with a corresponding re ult about the lower limit.
SA6 A survey with three strata is planned to estimate the percentage of families who
have accounts in saving banks and the average amount invested per family. Advance
estimates of the percentages Ph and the within-stratum S~ for tbe amount invested are as
foUows.
Stratum

1 0.6 20 90
2 0 .3 40 180
3 0.1 70 520

Compute the smallest sample sizes n and the nIt that satisfy the {oUowing requirements: (a)
The percentage of families is to be estimated with s.e. = 2 and the average amount invested
with s.e. = $S . (b) The percentage of families is to be estimated with s.e. = 1.S and the
average amount invested with s.e. = $S. •
Part (b) requires a compromise allocation, either by a computer program or the method
in the second edition, p. 123. The aUocation nh '" 371,344,315, with.n '" 1030 satisfies both
tolerances. Show that the 8ooth-Sedrarls~ method (section SA4) gives n,jn =0.431,
0.326, 0.243. This allocation would require n = 1073 to meet both tolerances.
5A.7 The table at top of p. 148 shows the frequency distribution of a population of 911
city sizes (or cities from 10,000 to 60000, arranged in classes of 2000. To shorten the
calculations, a coded y' and values oC JJ.
cum . ./i, cum. f,{y', andrfy'2 are given. Apply the
Dalenius-Hodges rule to create two strata for optimum allocation in the sense of Neyman.
Find the values of W.. and S.. for each of your strata. Verify (a) that the optimum sample
sizes are almost the same in the two ,trata and (b) by finding S2 for the whole population,
that
V(y) == 4 8
V.... (y,,) .
5A8 The right triangular distribution f(y) "'2(I-y). O<y<l. is divided into two
strata ~t the point Q . (a) Show that
W, = a(2 -a), W:z..... (1-a)2
s 2= a (6-6a+a
2 2
) :r. (l .... a?
I 18(2 - a)2 ' 2""~ ,.
 148 SAMPUNG TECHNIQUES

(b) Show that under the cum ../frule the best choice of a is 1 - 1/~ - 0 .37 and that with
this boundary the optimum ntln, is aboutH and V(y.,) is about 27% of the value given by
simple random sampling.
SA.9 In both exercises SA. 7 and SA .8, show that the Ekman rule Wit (Ylt - Y"_I) -
constant agrees very closely with the cum . ./irule in determining the stratum boundaries .

f y' .;/ Cumf Cum .;/ [Y'

10- 20S 0 14.3 20S 14.3 0

12- 135 1 11.6 340 25.9 135
14- 106 2 10.3 446 36.2 212
16- 82 3 9.1 528 4.5.3 246
18- 61 4 7.8 .589 53 .1 244
20- 42 5 6.S 631 59.6 210
22- 32 6 S.7 663 6S .3 192
24 - 30 7 5.5 693 70.8 210
26- 27 8 5.2 720 76.0 216
28- 18 9 4.2 738 80.2 162
30- 22 10 4.7 760 84.9 .. 220
32- 21 11 4.6 781 89.S 231
34- 19 12 4.4 800 93 .9 228
36 - 16 13 4.0 816 97.9 208
38- 14 14 3.7 830 101.6 196
40- 17 15 4.1 847 10.5.7 2S5
42- 9 16 3.0 8.56 108.7 144
.. 44 -
46-
8
11
17
18
2.8
3.3
864
87S
111.5
114.8
136
198
48- 9 19 3.0 884 117.8 171
50- 7 20 2.6 891 120.4 140
.52- 4 21 2.0 89S 122.4 84
.54- 5 22 2.2 900 124.6 110
.56- .5 23 2.2 90S 126.8 11.5
58- 6 24 2.4 911 129.2 144

Totals 911 129.2 4407

Ify·· - .50,39.5

SA.I0 A sum of $SOOO is available for a stratifted .ample. In the notation of section
SA.8 the cost function is thought to be, roughly, C ... 200L + IOn and
S2[p2
V(y.,) = -;- L1+(1 - p') ]

where p is the correlation between the variate used to construct the strata and the variate to
be measured in the survey. Compute the optimum L for p - 0.9S, 0.9, and 0.8. What is •
good compromise number of strata to' use for all three v.alues of p?
SA. 11 The following data are derived from a stratified sample of tire dealers taken in
March 1945 (Deming and Simmons, 1946). The dealers were aSlianed to Itrata accordln,
 FURlHER ASPECT'S OF STRATIPlSD SAMPLING 149
fO tbe number of new tires held at a previous census. The sample means Yh are the mean
numbers of new tires per dealer. (a) Estimate the gain in precision due to tbe stratification.
(b) Compare this result with the gain that would have been attained from prooortional
allocation.
Stratum
Boundaries Nh Wh s_t n"
!i"
1-9 19,850 0.8032 4 .1 34.8 3000
10-19 3,250 0.1315 13.0 92.2 600
20-29 1,007 0.0407 25.0 174.2 340
30-39 606 0.0245 38.2 320.4 230

Totals 24,713 0.9999 4170

SA.12 A population has two strata of relative sizes W. = 0.8, W2 = 0.2 and within-
stratum variance~ S. 2= 100, S2 2 ,. 400. A stratified random sample is to be taken to satisfy
the following requirements: (i) the means of each stratum are to be estimated with variance
:s1; (U) V(Y.. ) :sO.S. lanoring the fpc. find the values of n lo n2 that satisfy all three
requirements for minimum n ... n. +112'
Hint. Note that -oV(y.. )/on 1 >-0\'(9.,)/0" 2 if ". <2" 2' Fuller (1966) has discussed
various methods of handling this problem.
SA.13 In an example due to Nordbotten (1956) and worked by Kokan (1963), a survey
is planned to estimate total employment Y. and the value of production Y1 in establish-
ments manufacturing furniture. When establishments ate stratified by size, the N~ and
l'Ou&h estimate of the S~~ are as follows.

Stratum

Large 600 200 500,000

Small 1,000 10 4,000

1,600

The requirement that estimates of Y. and Y2 not be in error by more than 6% (P ... 0.95)
amounts to tolerances
VI - V(YI.):S 0.03S1 : V1 - V(YlM):s S6.2S
Show that the optimum allocation for YI with nl - 450, nz -167, II '" 617 satisfies both
tolerances. Note that in this problem the fpc cannot be ianored.
SA.14 In stratified random samplin, with one unit per stratum, assume that the strata
can be grouped into pairs with ~ . .. N", - ~ (/ - 1. 2, . ..• L/2). An alternative sampling
method' draws two units at random from each pair of strata. Show that fOT this method

V(1).. J - (~-
,-
li) ,-I
'f [2N/(~-1) I-If S~+(Y,I - y",)2 ]
Hence show that the expected value of the "collapsed ,trata" estimate VI(t.) in formula
(SA. 54), sec.'tion SA.12, overestimates V(1)..J), tbe variance that would apply if strata twice
as larac were used.
 CHAPTER 6
..
Ratio Estimators

6.1 METHODS OF ESTIMATION

One feature of theoretical statistics is the creatior. of a large body of theory that
discusses how to make good estimates from data. In the development of theory
specifically for sample surveys, relatively little use has been made of this know-
ledge. I think there are two principlll reasons. First, in surveys that contain a large
number of items, there is a great advantage, even with computers, in estimation
procedures that require little more than simple addition, whereas the superior
methods of estimation in statistical theory, such as maximum likelihood, may
necessitate a series of successive approximations before the estimate can be
found. Second, as noted in section 1.4, there has been a difference in attitude in
the two lines of research. Most of the estimation methods in theoretical statistics
assume that we know the functional form of the frequency distribution followed
by t~e data in the sample, and the method of estimation is carefully geared to this
type ~f distribution. The preference in sample survey theory has been to make, at
most, limited assumptions about this frequency distribution (that it is very skew or
rather symmetrical) and to leave its specific functional form out of the discussion.
This preference leads to the use of simple methods of estimation that work well
under a range of types of frequency distributions. This attitude is a reasonable one
for handling surveys in which the type of distribution may change from one item to
another and when we do not wish to stop and examine all of them before deciding
how to make each estimate.
Consequently, estimation techniques for sample survey work are at present
restricted in scope. Two techniques will now be considered-the ratio method in
this chapter and the linear regression method in Chapter 7.

6.2 THE RATIO ESDMATOR

In the ratio method an auxiliary variate XI, correlated with YI, is obtained for
each unit in the sample. The population total X of the XI must be known. In
practice, XI is often the value of YI at some previous time when a complete census
was taken. The aim in this method is to obtain increased precision by taking
~50
 RATIO ESTIMATORS 151
advantage of the correlation between YI and X i' At present we assume simple
random sampling.
The ratio estimate of Y, the population total of the Y.. is

Yn. =IX=~X
x X
(6.1)

where y, x are the sample totals of the Y, and Xi, respectively.

If X, is the value of Y, at some previous time the ratio method uses the sample to
estimate the relative change Y / X that has occurred since that time. The esti mated
relative change Y/ x is multiplied by the known population total X on the previous
occasion to provide an estimate of the current population total. If the ratio Y'/ X, is
nearly the same on all sampling units, the values of y/ x vary little from one sample
to another, and the ratio estimate is of high precision. In another application X,
may be the total acreage of a farm and Y, the number of acres sown to some crop.
The ratio estimate will be successful in this case if all farmers de\'ol!! about the
same percentage of their total acreage to this crop.
If the quantity to be estimated is Y, the population mean value of y" the ratio
estimate is

Frequently we wish to estimate a ratio rather than a to tal or mean, for example,
the ratio of corn acres to wheat acres, the ratio of expenditures on labor to total
ex.penditures, or the ratio of liquid assets to total assets . The sample estimate is
R = y/x. In this case X need not be known. The use of ratio estimates for this
purpo e has already been discus ed in section 2.11 and (with cluster sampling for
proportions) 3.12.
Example. Table 6 .1 shows the number of inhabitants (in 1000's) in each of a simple
random sample of 49 cioes drawn from the population of 196 large cities discussed in
section 2.15 . The problem is to estimate the rotal number of Inhabitants in the 196 cities in
1930. The true 1920 total, X, is assumed to be known. Its value i 22,919.
The example is a suitable One for the ratio e timate. The majority of the cities in tht"
sample show an increase ill size from 1no to 1930 of the order of 20% . From the sample
data we have
Y =L >" :c 6262, x =1: x, = 5054 .
Consequently the ratio e timate of the 1930 total for all 196 citr~s is
./', y 6262
TR =;X = 5054(22,919) =28,397

The corresponding estimate based on tbe sample mean per city is

y ... Ny = (196)(6262) - 25,048

49 ,I'
The oorrect total in 1930 is 29,351. , III I ,.
 152
SAMPLING TP.(l{N]OUEs

TABLE 6,1
Srzn OF 49 LARoE UNIT£O STATES Cmu (in 1000's) IN 1920 (~J ANO 1930 (311)
~I Yi ~.
'!II :/;1 !II
76 80 2
138 50 243 291
143 507 634
67 67 87 105
179 260 30
29 SO 121 111
113 71 79
381 464 SO 6-4 256 288
23 4S 44
37 58 43 61
bJ 77
120 89 25 57
lIS 64
61 63 94 85
69 64 77
387 459 43 50
56 142 298
93 104 40
317
172 60 36 46
183 40 64
78 106 161 232
38 52 74
66 86 136 93
60
139 45 S3
57 116 130
46 65 36 54
46 53 50 58
48 75

'. 70
r-
60

50
r- - .
200 rollo .. lIml ttl

.g40
:::I
W!~
...~30 r - -
r -
20

10
1'- h )( cienolft p<>puillion tolal

0 t~ ,-I'
J..
~r--------- __________________________--,
g20r--- __
~

110 r-----.=-a
0 ,'::8-A11!~-
1

"" 6.1 Experimental comparison of tbe ratio estimate witb the estimate based on the sample mean.
 RATIO ESTIMATORS 153
Figure 6.1 shows the ratio estimate and the estimate based on the sample mean per city
for each of 200 simple random samples of size 49 drawn from thls population. A substantial
improvement in precision from the ratio method is apparent.

6.3 APPROXIMATE VARIANCE OF THE RATIO

ESTIMATE
The distribution of the ratio estimate has proved annoyingly intractable
because both y and x vary from sample to sample. The known theoretical results
fall short of what we would like to know for practical applications. The principal
results are stated first without proof.
The ratio estimate is consistent (this is obvious). It is biased, except for some
special types of population, although the bias is neglible in large samples. The
limiting distribution of the ratio estimate, as n becomes very large, is normal,
subject to some mild restrictions of the type of population from which we are
sampling. In samples of moderate size the distribution shows a tendency to
positive skewnes in the kinds of populations for which the method is most often
used. We have an exact formula foJ' the bias but for the sampling variance of the
e. timate only an approximation valid in large samples.
These results amount to saying that there is no difficulty if the sample is wac
enough so that (a) the ratio is nearly normaJJy distributed and (b) the large-sample
formula for its variance is valid. As working rule. the large-sample results may hre
used if the sample size exceeds 30 and is also large enough so that the ooefticien15
of variation of i and j are both less than 10%.
Theorem 6.1. The ratio estimates of the population total Y, the population
mean. Yo and the population ratio Y/X are, respectivel).

(6.2)

(6.3)

ff (Y,-Rx,)21
V(R).:.!i(L- N-l J
1
(6.4)

where f =n/N is the sampling fraction. The method used in theorem 2.5 shows
 154 SAMPLING TECHNIQUES

that (6.2), (6.3), and (6.4) are also approximations to the mean square errors of the
e timator in these formulas.
The !lrgu"!.e~t l~ading lJ the approximate result (6.4) was given in theorem 2.5 .
Since YR = XR, YR = NXA., the other two results follow immediately.

Coronary l. There are various alternative form of the result. Since Y = RX,
we may write
A N 2 (1- f) N - - 2
V(YR )= n(N-l) i~1 [(Y;- Y)-R(x;- X))
2
N (1- f) [ - 2 ' 2 - 2
=n(N-l) r(Yi- Y) +R r(xl- X)
-2R r (Yi - Y)(.t; - X)]
The correlation coefficient p between Yi and X; in the finite population is defined
by the equation
N _ _
E(Yi - y)(Xi - X) r (y; - y)(Xi - X)
P =JE(YI - 9)2E(x; - *>2 (N-l)S~..

This leads to the result

VCYR ) -= N (1- f)(Sy 2+ R 2S ..2 -2RpS~,,)

2
'r (6.5)
n

An equivalent form is
2
y2(S/ s.. 2Sy",)
V(YR ) = (l-,n-::- ~+~- .........
A
(6.6)
n Y X YX
where S YJC = pS,sx is the covariance between Y. and X,. This relation may also be
written as

(6.7)

where Cyy,C.... are the squares of the coefficients of variation (cv) of Yi and XI
respectively, and CyX is the relative covariance.

Y
Corollary 2. Since YR , R , and R differ only by known multipliers, the
coefficient of variation (i.e., the standard error divided by the quantity being
estimated) is the same for aU three estimates. From (6,7) the square of this cv is
V(YR ) I- f (Cyy+C
(cv) 2 =~=- . .... - 2C'yO' ) (68)
•
r Y n I ,
 RAno ESnMATORS 155
The quantity (cv)' has been called the relative variance by Hansen et al. (1953). Its
use avoids repetition of variance formulas for related quantities like the estimated
population total and mean.

6.4 ESTIMATION OF 1HE VARIANCE FROM A SAMPLE

From equation (6.2).

As already mentioned in section 2.11. we take

I (Yi- RxY
(n -1)

as a sample estimate of the population variance. Thi ' estimate bas a bias of
order lin.
For the estimated variance, vCYR ), this gives
~ N2(1-f) " ~
v(YR) = 1I(n -1) i~1 (YI - Rx/)' (6.9)

This result may be expressed in several different ways. For example,

~ f)(~ 1. J:2 ~ 1. ~
2
v( YR ) =N1I(n(1-
-1) t.. Yi + i... XI -
I(
A
21( t.. y,x,) (6.10)

= N (1- f)(s/+R1.s/-2Rsy~)
2
(6.11)
n
where Sy~ = E(Yt - y)(x/ - x)/(n -1)is the sample covariance between Y, and X"
There are two alternative formulas for the sample estimate of the variance.
Since YR = NXR. one form for R is

(6.12)

Since, however, R = Pix. the quantity X need not be known and is sometimes
not known when estimating R. This suggests the alternative form

(6.13)

This form could also be used for v('YR ), taking v£9'~) =X2tJ2(R).
This raises the question: If X is known, is VI preferable W "2? The n;w r is Qot
v,
at present clear. P. S. R. S. Raoand J. N. K. Rao (1971) studi d .he bi es :11 and ,
 156 SAMPLING TECHNIQUES

V2 analytically in an infinite population with the linear regression model

y, = a + (lx, +e,
with
B'(eilxl) = 0, V(el lx, ) = 8x/ , E(e,e,lx;x,) = 0,

were Os t s 2 and XI has a g1mrna distribution ax"- Ie- x. The range Os t s 2 was
'studied because in applications the residual variance of YI is thought to increase
with X , at different rates in different populations. They found V 2 Jess biased for
Os t s 3/2, but also less stable for I = 0 or I = 1.

6.S CONFIDENCE LIMITS

If the sample is large enough 0 that the normal approxi ation appUes,
confidence limits (or Y and R may be obtained.

Y: 9'R ±zJ;;(Y,J (6.14)

R : R ±z.rvrJi) (6.15)

where z is tell rma] deviate corre po nding to the chosen confidence probability.
In secor'n 6.3 it was suggested that the normal approximation holds reasonably
well if the . ample size is at least 30 and is large enough so that the cv's of 9 and 1
are botH k 'lt than 0.1. When these conditions do not apply, the formula for v(R)
tends to gi\'-t vaJues that are too low and the positive skewness in the distribution
of R may become noticeable.
An alternative method of computing confidence limits, which takes some
account or the skewness of the distribution of R. has been used in biological assay
(Heller, 1932 ; Paulson, 1942). The method requires that 9 and i follow a
bivariate normal distribution, so tbat (9 - Ri) is norma11y distributed. It follows
that in simple random samples tbe quantity

y-Ri
(6.16)
J[(N - n)/ Nn}/ s/ + R 2s/ - 2Rs y%

is approximately normally distributed with mean zero and unit standard devia-
tion.
The value of R is unknown, but any contemplated value of R which makes this
normal deviate large enough may be regarded as rejected by the sample data.
Consequently, confidence limits for R are found by setting (6.16) equal to ±z and
solving tbe resulting quadratic equation for R. The confidence limits are approxi-
mate since the two roots of the quadratic are imaginary with some samples. Such
cases become rare if the cv's of j and i are Jess than 0.3.
 KA flU ~"l1MATORS 157
After some manipulation, the two roots may be expressed as

R A(l-z 2c,.)±zJ(c,,+cu- 2c91) - z2(C9jlCu - c,l>

= l - z 2cu (6.17)

where
2
N-ns
C99=--~
Nn y
is the square of the estimated cv of y, with analogous definitions of c9' and Cu. If
Z2 C,,, Z2 CU• and Z2 C,. are aU small relative to 1, the limits reduce to

R =A :t:zJc,,+cu -zCp.
This expression is the same as the normal approximation (6.15).
Even with bivariate normality, the FieUer limits have been criticized as not
conservative enough. James, Wilkinson, and Venables (1975) explain the nature
of the difficulty and present an alternative method.

6.6 COMPARISON OF 11IE RAnO ESTIMATE wrm

11IE MEAN PER UNIT
The type of estimate of Y that was studied in preceding chapters is Ny, where y
is the mean per unit for the sample (in simple random sampUng) or a weighted
mean per unit (in stratified random sampling). Estimates of this kind are called
estimates based on the mean per unit or estimates 0 tained by simple txpatlSion.
Theorem 6.2 In large samples, with simple random sampling, the ratio
estimate YR has a smaller variance than the estimate Y =Ny obtained by simple
expansion, jf

> y ~)/(~) "'" coefficient of variation of XI

p 2\X Y 2(coeflicient ofvariation of YI)
Proof. For Y we have

For the ratio C".stimate we have from (6.S)

V(Y.R) =_N'l(I- f) (Sy 7. + R 2S" 7. - 2R.pSyS,,)

n
Hence the ratio estimate bas the smaller variance if
S/ +R 2Sx2 - lRpSySx < $,"
 158 SAMPLING TECHNIQUES

If R = PI X is positive, this condition becomes

(6.18)

6.7 CONDmONS UNDER WHICH THE RATIO ESTIMATOR

IS A BEST LINEAR UNBIASED ESTIMATOR
A well-known result in regression theory indicates the type of population under
which the ratio estimate may be called the best among a wide class of estimates.
The result was first proved for infinite populations. Brewer (1963b) and Royall
(1970a) extended the result to finite populations. The result holds if two condi-
tions are satisfied.

1. The relation between y, .a nd X I is a straight line through the origin.

2. The variance of YI about this line is proportional to X"

A "best linear unbiased estimator" is defined as follows. Consider all estimators

Yof Y that are linear functions of the sample values y" tha! is, that are of the form

flYI + f2 Y2 + . .. + fnYn
where the I's do not depend on the Yj, although they may be functions of the Xj. The
choice of /'s is restricted to those that give unbiased estimation of Y. The
estimator with the smallest variance is called the best linear unbiased estimator
(BLUE).
Formally, Brewer and Royall assume that the N population values (YI. x,) are a
random sample from a superpopulation in which

(6. 19)

where the £ j are independent of the X, and x, > O. ln arrays in which X , is fixed, £ ,
has mean 0 and variance AXj. The x,(i = 1, 2, ... , N) are known .
In the randomization theory used thus far in this book, the finite population
total Y has been regarded as a fixed quantity. Under model (6.19), on the other
N
hand, Y =pX +I £ 1 is a random variable. In defining an unbiased estimator under
this model, Brewer and Royall use a concept of unbiased ness which differs from
that in randomization theory. They regard an estimator Yas unbiased if B(Y) =
B( y) in repeated selections of the finite population and sample under the model.
Such an estimator might be called model-unbiased.
Theorem 6.3. Under model (6.19) the ratio estimator YR = Xy/i is a best
linear unbiased estimator for any sample, random or not, selected solely according
to the values of the X"
 RATIO ESTIMATORS 159
Proof. Since E(sdx,) =0 in 'repeated sampling, it follow from (6.19) that
N
Y==~X+LSi : E(Y)=={3X (6.20)

Furthermore, with the model (6.19) any linear estimator Y is of the form
(6.21)

If we keep the n sample values Xi fixed in repeated sampling under the model
(6.19),

(6.22)

A n
From (6.20) and (6.22), Y is clearly model-unbiased if L l;Xi -x. Minimizing
V( n under this condition by a Lagrange multiplier gives
21;x, = ex,: I, = constant == XJ ni (6.23)
The constant must have the value X/ni in order to satisfy the model-unbiased
f
condition / XI = X. Hence the BLUE estimator Y is nyX/ ni = Xy/ i == YR , the
usual ratio estimator. This completes the proof.
Furthermore, from (6.20) and (6.21), with 1= X/ni,
"" It .N "N
YR - Y == L ';E{ - I Sj :;: (X/ ni)(L S;) - L E; (6.24)

(6.25)
N- "
where I denotes tbe sum over the (N - n) population values that are not in the
sample. Hence

V( YR ) = A(X - n!~2(ni) + A(X _ nil = A(X - ~)X (6.26)

(nx) nx
A model-unbia ed estimator of A from this 'Sample is easily shown to b:
.. "1 ,.
A =I -(YI- Rx I )2/(n -1) (6.27)
XI

where R == f/i, as usual. This value may be substitu'ted in (6.26) to give a

model-unbiased sample estimate of V(YR ). • I •

. The practical relevance of these results is that they suggest the conditions under
which the ratio estimator is superior not only to y but is the best ot a,whole class of
estimators. When we are tzying to decid~ what kind of estimate to lI:se, a graph in
160 SAMPLING TECHNIOUES

which the sample values of Yj are plotted against those of XI is helpful. If this graph
shows a straight line relation passing through the origin and if the variance of the
points YI about the line seems roughly proportional to Xit the ratio estimator will be
hard to beat.
Sometimes the variance of the YI in arrays in which XI is fixed is not proportional
to XI' If this residual variance is of the form AV(XI), where V(XI) is known, Brewer
and Royall showed that the BLUE estimator becomes
"
y=#WiY;X; (6.28)
L" WiX/
where Wi = l/v(xl)' In a population sample of Greece, Jessen et al. (1947) judged
that the residual variance increased roughly as x? This suggests a weighted
regression with WI = l/xI2, which gives

(6.29)

For a given population and given n, V(YR ) in (6.26) is clearly minimized, given
every XI > 0, when the sample consists of the n largest XI in the population. 10 16
small natural populations of the type to which ratio estimates have been applied,
RO'Yall (1970) found for samples having n = 2 to 12 that selection of the n largest
XI usually increased the accuracy of YR. .
In summary, the Brewer-Royall results show that the assumption of a certain
~ of modelleads to an unbiased ratio estimator and formulas for V(YR ) and
v(YR ) that are simple and exact for any n > 1. The results might be used in
practice in cases where examination of the Y, X pairs from the available data
suggests that the model is reasonably correct. The variance formulas (6.26) and
(6.27) appear to be sensitive to inaccuracy in the model, although this issue needs
further study.
Further work by Royall and Herson (1973) discusses the type of sample
distribution needed with respect to the XI in order that YR remains unbiased when
there is a polynomial regression of YI on XI.

6.8 BIAS OF THE RATIO ESTIMATE

In general, the ratio estimate has a bias of order 1/ n. Since the s.e. of the
estimate is of order I/Jn, the quantity (bias/s.e.) is also of order 1/Jn anq
becomes negligible as n becomes large. In practice, this quantity is usually
unimportant in samples of moderate size. Its value in small samples is of interest,
 RA TIO ESTIMATORS 161
however, in stratified sampling with many strata, where we may wish to compute
and examine ratio estimates in individual strata with small samples in the strata.
Two useful results about the bias are presented.
The first gives the leading term in the bias when it is expanded in a Taylor's
series.

Write
1 1 1( i - .t\ - I 1 ( i - X) (6.30)
i =X + (i - X) = X 1+ X- J == X 1- X
Hence,

" y - Ri( i-X\ (6.31)

R-R == - x l - -XJ

Now
E (y- Ri) = Y - RX =O
so that the leading term in the bias comes from the second term inside the
brackets. Furthermore,
- - - 1- 1
E y(i - X) = E (y - Y) (i - X) = -pSyS" (6.32)
n

by theorem 2.3 (p. 25) and the definition of p. Also,

E X-(x- - X)
- =E( x- - X)- 2 =I --IS 2
n "
Hence the leading term in the bias is

E (l? - R) == ~(RS/- PSySx ) (6.33)

== 1- f(Cu - Cyx)R (6.34)

n

For a rigorous justification of (6.34), see David and Sukhatme (1974).

Now the leading term in VCR) is

V(R) = (~{)(S/ +R 2S/- 2RPSyS.. ) (6.35)

from (6.S), substituting R = YR / NX

162 SAMPLING TE HNIQlJES

From (6)3) and (6.~5) the leading term in the quantity (bias/s.e.), which is the
same for R. Y R , and Y R , may be expressed as
(bias) _ (RS" - pSy)
~ = cv (x)(R 2S"2_ 2RpSyS,, + Sy2) 1/2 (6.36)

where cv (i) = ./1 - IS,,/ JnX. By substituting sample estimates of the terms in
(6.36), Kish, Namboodiri, and Pilla; (1962) computed the (bias/s.e.) values for
numerous item in various national and more localized studies. In the national
studie nearly all the (bias/s.e.) values were < 0.03 and almost the only values
> 0. 10 in their studies occurred for a single stratum with nh = 6 small hospitals.
The second result, due to Hartley and Ross (1954). gives an exact result for the
bias and an upper bound to the ·ratio of the bias to the standard error. Consider the
covariance, in simple random samples of size n, of the quantities Rand i. We have

cov (R, i) = E(f i) - E(R) E(i) (6.3 7)

= Y-XE(R) (6.38)
Hence
~ Y 1 ~ 1 ~
E(R) = X - X cOV (R, i) = R - X cov (R, i) (6.39)

Thus the bias in R is - cov (R. £)/ X. Unlike the Taylor approximation (6.33) to
the bias. this expression is exact.
Furthermore,

' 10
lblas . R~I = IPR ..fO"RO"xl
-
X
O"AO"i
S X
since Rand i cannot have a correlation> 1. Hence
Ibias in RI 0"1 f _
' - - - - - ' < -= = cv 0 x (6.40)
O"R X
The same bound applies, of course, to the bias in Y and R YR'
Thus, if the cv of i is
less than 0.1, the bias may safely be regarded as negligible in relation to the s.e.

6.9 ACCURACY OF THE FORMULAS FOR THE

VARIANCE AND ESTIMATED VARIANCE
e
With small samples, say n < 30 anq xx large.!. it has long been suspected that the
large-sample formula given for. VCR) and v(R) are undere timates. By a Taylor
 RATIO ESTIMATORS 163
series expansion, Sukhatme (1954) expressed the error in VCR) in terms of the
bivariate moments of y and x. Unfortunately, the result is too complicated to lead
to a useful guide for practical applications.
If y and x follow a bivariate normal distribution, Sukhatme's result simplifies
considerably. Let
VI = (Cyy + Cxx -2Cyx)/n
denote the first approximation to the relative variance of R, ignoring the fpc . To
terms of order 1/ n 2,

E(R
R
_R)2== VI(1 + 3Cn xx + 6CXX _p 2Cyy + Co -2Cyx)
n Cyy + Cxx - 2 CyX
(6 .41)

Since the right-hand term inside the parentheses is less than 6Cu /n , this gives
E(R ;Ry < V(1+9: 1
xx
) (6.42)

to terms of order 1/ n 2 . Now Cxx / n is the square of the coefficient of variation of i .

Thus, if n is large enough so that the cv of i is less, than 0.1, use of VI hould not
underestimate by more than 9% . In practice, the multiplier 9 in (6.42) appears to
be unduly high as compared with (6.41) . For instance, if Cn = Cyy, (6.41) reduces
to

(R-R)2 [C
E ~ == V I 1+ ;X (6 - 3p ) ] (6.43)

Since p is almost always positive in applications of the ratio method, a multiplier

between 3 and 6 is more representative. However, the effects of nonnormality in y
and x al 0 enter into the term of order l/n 2 .
From a Monte Carlo study by Rao (1968) of small natural populations, some
illustrative results on the biases of the large-sample formulas for VCR) and vI(R)
will be quoted for simple random samples from eight populations each with
N > 30. The populations are described in Rao (1969). The formulas for V(R) and
vI(R) , given in (6.4) and (6.12), will be appraised as estimators of the true
MSE{lh
For a given population the quantities 100[MSE(R) - V(R)]/MSE(R) and
100[MSE(R)-Evl(R))/MSE(R) are the percent underestimates of the true
MSE(R). The averages of the e percents in the eight populations were as shown in
Table 6.2 .
In these data the pen:ent underestimation in VCR) scarcely declines at all with
increasing n. This is explained in part by the circumstance that in one population
VCR) gave an overestimate which declined with n. The indication!; from these
results are that the biases in vl(R) are much more serious in small samples than
those in R itself, and are unsatisfactory at least up to n = 12. For n =4, Koop
 164 SAMPLING TECHNIQUES

TABLE 6.2
Average percent
underestimation of MSE(R)
n
Estimator 4 6 8 12

V(R) in (6.4) 14 14 14 12
v,(R) in (6.12) 31 23 21 18

(1968) found underestimations in vI(R) averaging 25% in three populations with

N== 20.
An alternative estimator to VI (R) that looks promising as regards bias reduction
is presented in section 6.17.

6.10 RATIO ESTIMATES IN STRATIFIED RANDOM

SAMPLING
There are two ways in which a ratio estimate of the population total Y can be
made. One is to make a separate ratio estimate of the total of each stratum and add
these totals. If Yh, Xh are the sample totals in the h th stratum and X h is the stratum
total of the Xh" this estimate YRs (s for separate) is

" ./', ,Yh ,Yh

rRs == L.. -Xh == L.. -;- X h (6.44)
h Xh h Xh
No assumption is made that the true ratio remains constant from stratum to
stratum. The estimate requires a knowledge of the separate totals X h .
Theorem 6.4. If an independent simple random sample is drawn in each
stratum and sample sizes are large in all strata,
A ,N/(l-fh) 2 22
V(YRs)==L.. (Syh +Rh SXh -2RhPhSyhSxh) (6.45)
h nh

where Rh = Yh/ X h is the true ratio in stratum h, and Ph is defined as before in each
stratum.
Proof. Apply formu la (6.2), section 6.3, for a simple random sample to give in
stratutn h,

(6.46)

Since YRs == L YRh and sampling is independent in each stratum, V(YRs )=

h
L V(YRh ) and the result (6.45) follows.
 RATIO ESTIMATORS 165
This formula is valid only if the sample in each stratum is large enough so that
the approximate variance formula applies to each stratum. This limitation should
be noted in practical applications.
Moreover, when the nh are small and the number of strata L is large, the bias in
YR. may not be negligible in relation to its standard error, as the following crude
argum~nt suggests.
In a single stratum we have seen (section 6.8) that
Ibiasin
A
YRhl :S cv 0 f-Xh
0"( Y Rh )
If the bias has the same sign in all strata, as may happen, the bias in Y10 will be
roughly L times that in YR h . But the standard error of YRs is only of the order of
JL times that of YRh • Hence the ratio
Ibiasin YRJI
O"(VR .)
is of order

JL(cv of i h )
For example, with 50 strata and the cv of Xh about 0.1 in each stratum, the bias
in YR. might be as large as 0.7 times its standard error. The contribution of the
bias to the mean square error of Y10 would then be about one third.
Although in practice the bias is usually much smaller than it upper bound, the
d~nger of bias with the eparate ratio estimate should be kept in mind if
.JL(cv of i h ) elfCeeds, say, 0.3.

6.11 TIlE COMBINED RATIO ESTIMATE

An alternative estimate is derived from a single combined ratio (Hansen,
Hurwitz, and Gurney, 1946). From the sample data we compute

(6.47)

These are the standard estimates of the population totals Y and X. respectively,
made from a stratified sample. The combined ratio estimate, YRc (c for combined)
IS

YRc
A _ Y"x -- --::-
- -,.-
Y"X (6.48)
X. , x"
where y., = Viti N, i ll = X.,IN are the estimated population means from a
stratified sample.
166 S"AMPUNG TECHNIOUES

The estimate YRo does not require a knowledge of the X It , but only of X.
The combined estimate is much less subject to the risk of bias than the separate
estimate. Using the approach of Hartley and Ross in section 6.8, we have, writing
It = y,,/x'h
cov eRe, X.,) = E{ ~., .X
.,) - E(1t)E(x,,)
\i.,
= Y-XE(lt) (6.49)
Hence

and

(6 .50)

Thus the biases in Re , YRc are negligible relative to their standard errors, provided
only that the cv of X" is less than 0.1 .
Theorem 6.S. If the total sample size n is large,
2
A
V( YRc) ="N
L..
It {I - flt)
(Sylt
2
+ R 2 S,,1t2 - 2RphSv1tS,,1t ) (6.51)
It nit

Proof. This follows the same argument as theorem 2.5. In the present case the
key equation is

(6.52)

Now consider the variate UIt ; = Yiti - Rx hi • The right side of (6.52) is Na." where a.,
is the weighted mean of the variate UIt; in a stratified sample. Furthermore, the
population mean 0 of Uh; is zero, since R = Y/ x.
Hence we may apply to all theorem 5.3 for the variance of the estimated mean
from a stratified random sample. This gives

(6.53)

where
 RATIO ESTIMATORS 167
From equation (6.45) and (6.5)) it is interesting to note that the approximate
variances of YIb and YR assume the same general form, the difference being that
the population ratio R" in the individual strata in (6.45) are all replaced by R in
(6.51) .

6.12 COMPARISON OF THE COMBINED AND

SEP ARA TE ESTIMATES
We may write
V( YR c ) - V( YRJ

-- '-
'" N/(l- fh ) [ (R 2 - Rh 2) Sxl, 2 - 2 ( R - Rh )PhSvl,Sxh ]
I, nh

In si tuations in which the ratio estimate is appropriate the last term on the right
is usuall y mall. (It vanishes if within each tratum the relation between YIII and X/t i
is a straight line through the origin.) Thus, unless Rh is constant from stratum to
stratum, the use of a eparate ratio estimate in each stratum is likely to be more
preci e if the s a~mple in each stratum is large enough so that the approximate
formula for V( YR. ) is valid , and the cumulative bias that can affect Y Ib (section
6.10) is negligible . With only a small sample in each stratum, the combined
estimate is to be recommended unless there is good empirical evidence to the
contrary.
For ample estimates of the e variances we substitute sample estimates of R"
and R in the appropriate places. The sample mean squares Sy/ and sx/ are
substituted for the corresponding variances and the sample covariance for the
term PI,Sv"Sxh' The ample mean square and covariance must be calculated
separa te ly for each stratum.

Example. The data come from a census of all farms in leffer on County, Iowa. In this
example }I~, represents acres in corn and x'" acres in the farm . The population is divided
into two strata, the first stratum containing farms of as many as 160 acres. We assume a
sample of 100 farms . When tratified samplil1g is used, we will suppose that 70 farms are
taken from stratum 1 and 30 from stratum 2, this being roughly the optimum allocation .
The data are given in Table 6.3. The last three quantities, Q~, V h ', and Vh ", are auxiliary
quantities to be used in the computations, the last two being defined later.
We con 'ider five methods of estimating the population mean corn acres per farm . The fpc
are ignored .
1. Simple rand m ample: mean per farm estimate.
S.2 620
VI =-=-=6.20
n 100
168 SAMPLING TECHNIQUES

TABLE 6.3
DATA fROM JEFFERSON CoUNTY, IOWA
Size
Strata (farm acres) Nil. Sy,,' SyZl\ Su,'

1 (}'-16O 1580 312 494 2055 0.2350

2 More than 160 430 922 858 7357 0.2109
For complete pop. 2010 620 1453 7619 0.2242

Strata Y" X" n" Q" = W"I/ n" V,,' V,,'

19.40 82.56 70 0.008828 193 194

2 51.63 244.85 .30 0.001525 887 907
For complete pop. 26.30 117.28 100

2. Simple random sample: ratio estimate.

V2 - !,
-,S12 + R 2 S. 2 - 2RS,.)
n

= 1~0[620+ (0.2242)2(7619) - 2(0.2242)(1453)]

=3.51
3 . Stratified random sample: mean per farm estimate.
~2
~t V3=r-" S'l,/ = !:Q,.sylt':: 4.16
n~

4. Stratified random sample: ratio estimate usmg a separate ratio in each stratum.
V. = L O~(S,~2 +R~2S.~2_2R"S1M) =L Qh Vh ' =3.06
5. Stratified random sampling: ratio estimate using a combined ratio.
V, = LO,,(S,/+R 2S./-2RS,M) =L Q" V"ff", 3.10
The relative precisions of the various methods can be summarized as follows,

Method of Relative
Sampling method Estimation Precision

1. Simple random Mean per farm 100

~. Simple random Ratio 177
3. Stratified random Mean per farm 149
4. Stratified random Separate ratio 203
5. Stratified random Combined ratio 200
 RATIO ESTIMATORS 169
The results bring out an interesting point of wide application. Stratification by size of
farm accomplishes the same general purpose as a ratio estimate in which the denominator is
farm size. Both devices diminish the effect of variations in farm s.ize on the sampling error of
the estimated mean corn acres per farm . For instance, the gain in precision from a ratio
estimate is 77% when simple random sampling is used, but it is only 36% (203 against 149)
when stratified sampling is used .
In the design of surveys there may be a choice between introducing a factor into the
stratification or utilizing it in the method of estimation. The best decision depends on the
circumstances. Relevant points are : (a) some factors, for example, geographical location,
are more easily introduced into the stratification than into the method of estimation; (b) the
issue depend on the nature of the relation between y, and X,. All simple methods of
estimation work most effectively with a linear relation. With a complex or discontinuous
relation, stratification may be more effective. since. if1here are enough strata, stratification
will eliminate the effects of almost any kind of relation between YI and X,. (c) If some
important variates are roughly proportional to X" but others are roughly proportional to
another variate Zi' it is better to use Xi and z, as denominators in ratio estimates than to
stratify by one of them.

6.13 SHORT·Cur COMPUTATION OF

TIlE ESTIMATED VARIANCE
If nil := 2 in all strata, Keyfitz (1957) has given simple methods for computing the
approximations to the estimated variances of VI! or Rc;_ or, more generally, of
functions of one or more variables of the form Y II . For R c we have
A Nil
Y. L YII L 2 (YIII + Yill.)
R =~=~= _" _____- (6.54)
c XII L XII ... Nil ( + )
II ~ 2 XIII Xu

The Keyfitz method uses the identity that for nil =2,
2-
2Syll 2:= 2 L (YIII - 911)2 a! (Yhl- Yh2? = (dYh)2 (6.55)
leI

where dYh = (YIII- YII2). Hence

v( YI )= (~h) 2 2(1_ /h)Sy/ = (1- !h)(Yhl' - Yh2,)2:= (1- /h)(dYIl ') 2 (6.56)

where y~1 = NhYh/2. Similarly, for the sample estimate of the covariance,

(6.57)
Now

R - R =~- R == YII-RX,,= Y(Y,,_ X,,) (6.58)

c X., X XYX
170 SAMPLING TECHNIQUES

Since sampling is independent in different strata,

(6.59)

(6.60)

Keyfitz (1957) has extended this method to cover poststtatified estimators and
multistage sampling, and to give variances of differences of estimates from
successive surveys in periodic samples. Woodruff (1971) gives a general approach
that handles nonlinear estimators, unequal probabilities of selection, and samples
of size nh in the strata. As an illustration of Woodruff's approach, consider a
function feY> where Yrepresents the vector or set of m variables Y, = L Y'h' With
h
simple random sampling in stratum h, the flh are of the form (Nh/ nh) LYihh the
sum extending over the nil sample units in stratum h. By Taylor's approximation,
iJ
fch-f(Y)==I
, a
! (fI- Y,)=LL !}B! (V,h -
~I ' h U .lj
Yih) (6.61)

The trick is to reverse the order of summation, writing

(6.62)

where
"
(6.63)

By evaluating the (af/BY;) at the estimates fI,

the UIII can be calculated from (6.63)
for each sample unit in stratum h. With simple random sampling in the strata, the
usual formula for the estimated variance of the sum UII apj?lies. Hence, from
(6.62), an approximate sample estimate of the variance of fey) is
n.
L(Uhl- Uh)2
v[f(Y>J == L (Nil - nll)nll _;_i_ __
(6.64)
II Nil (nk - 1)
The advantage of this approach is that the covariances of the 'fJII need not be
calcuJated.
To illustrate the Uk l for the estimator k
= V/g, it helps to write Rc= Vtf V2• so
that
YJ af 1 af YJ
(6.65)
f(Y)=Y2 : aY2 - .
- Y ' aY2 ---~
2
 RATIO ESTIMATORS 171
Hence, from (6.63), with nh = 2,
. Nh(YlhJ YI ) N h (Ylh, - R cY2hl)
Uhi = - -"--~Y 2 h' =: - (6.66}
nh Yz Y2 nh Yz
In terms of the original Yh" Xhi,

(6.67)

It estimated variance is approximated as in (6.64).

The Keyfitz method is used, for instance, in the Health lnterview Survey of the
Natronal Center for Health Statistics. The sampling unit is a cluster unit-a county
or group of contiguous counties. Each unit is divided into segments of about 9
households, an average of 13 segments being chosen from a sampled unit. The
variables Yh .. Yh 2 might be the numbers of persons with a specific illness and Xh II
Xh 2 the total numbers of persons in the samples from the two units in stratum h.
In addition to the initial geographic stratification of the counties, the persons in
the sample are poststratified by age, sex, and color. Thus, instead of the estimator
YR ( = XR, of the total number ill, the estimator is
Y pS = 1... XaRa
A '" A
=:
Y"
r Xa"-1r" (6.68)
a a Xa
where a represents an age-sex-color class and Xa is the known total population in
this class. Here we have a function of two set of random variables - Ya and K".
Furthermore, for many reasons, Yah, and Ya ' /. i for two different classes a and a' in a
cluster unit may be correlated ; for example with an infectious disease the number
of cases may be high for all classes in the unit. Applying (6.63) and ignoring the
fpc's, we have with nh = 2,

V(YPS) =:~[~X,,(d~( Yf;h')f (6.69)

== r [r x,,(.t)
h" X"
(~
Y
_ ~)]
Xo a
2 (6.70)

where dYoh' = Nh (Yah I - Y"hl)/2.

In this application the Keyfitz method also handles further complexities of the
survey-selection of primary units with unequal probabilities. adju tments for
nonresponse, and use of the method of collapsed strata. Furthermore, since
computer time even with this simple method permits variance calculations for
only a limited number of items, charts of the relation between s.d. (Y)/ Y and Y
are given for ditItrent types of item to help in predicting s.d. for items for(n
which the s.d. has not been computed. Bean (1970) gives a clear presentation of
these methods and results.
172 SAMPLING TEOiNlOUES

6.14 OP11MUM ALLOCATION WITH A RATIO ESTIMATE

The optimum allocation of the nil may be different with a ratio estimate than
with a mean per unit. Consider first the variate YItr. From theorem 6.4 its '.'ariance
is

_ ~Nh(NI - nh)Sdll,
- .(..
2 .
WIt
hS
dll
2_ 1
- --.(..
~ dIII2 (6.71)
II nil N II - l l _ 1
where d"i = Yhl - R"xlll is the deviation of Y"I from Rllx"I' By the methods given in
Chapter 5 for finding optimum allocation, it follows that (6.71) is minimized
subject to a total cost of the form L Clln", when

With a mean per unit it will be recalled that for minimum variance nil is chosen
proportional to N"Sy,,/J0..
In the planning of a sample, the allocation with a ratio estimate may appear
a little perplexing, because it seems difficult to speculate about likely values of
Sd/t. Two rules are helpful. With a population in which the ratio estimate is a best
linear unbiased estimate, Sdk will be roughly proportional to JX;., (by theorem
6.3). In this case the nil should be proportional to NII JJ[,,;J0.. Sometimes the
variance of dill may be more nearly proportional to XII 2. This leads to the
allocation of nh proportional to N"xll/..rc;,. that is, to the stratum total of XIIi,
divided by the square root of the cost per unit. An example of this type is discussed
by Hansen, Hurwitz, and Gurney (1946) for a sample designed to estimate sales of
retail stores.
If the estimate YRc is to be used. the same general argument applies.

Example. The different methods of allocation can be compared from data collected in a
complete enumeration of 257 commercial peach orchards in North Carolina in June 1946
(Finkner, 1950). The purpose was to determine the most efficient sampling procedure for
estimating commercial peach production in this area. Information was obtained on the
number of peach trees and the estimated total peach production in each orchard. The high
correlation between these two variables suggested the use of a ratio estimate: One very
large orchard was omitted.
For thi illustration, the area is divided geographically into three strata. The number of
peach trees in an orchard is denoted by XIII and the estimated production in bushels of
peaches by YIII' Only the first ratio estimate Y116 (based on a separate ratio in each stratum)
will be considered, since the principle is the same for both types of stratiiJed ratio estimate.
Four methods of allocation are compared;__(_a) nil proportional 10 Nil' (b) nil pro-
portional to N"s.II, (c) nil proportional to NIIJX,., and (d) nil proportional to N"xll =Xh •
The sample size is 100. The data for these comparisons are summarized in Table 6.4.
 RA no ESTlMATORS 173
TABLE 6.4
DATA FROM THE NORTH CAROLINA PEACH SURVEY

Strata S",," S... s••• S.. S•• X. y. R. s..'

I 5186 6462 8699 72.01 93.27 53.80 69.48 1.29133 658
2 2367 3100 4614 48.65 67.93 31.07 43.64 1.40475 573
3 4817 4817 7311 69.83 85.51 56.97 66.39 1.16547 2706

Pop. 3898 4434 6409 62.43 80.06 44.45 56.47 1.27053 1433

Strata N. (a) N.S•• (b) vX. N.VX. (c) N.X. (d)

1 47 18 4384 22 7.33 344.5 20 2529 22

2 118 46 8016 40 5.57 657.3 39 3666 32
3 91 36 1781 38 7.55 687.1 41 5184 46

Pop. 256 100 20181 100 20.45 1688.9 100 11379 100

The upper part of the table shows the basic data. The method employed to calculate the
four variances was first to find the nk for each type of allocation. These values are shown in
the columns headed (a) through (d) in the lower part of the table. Thus, with allocation (a),
n~ = nN./ N , so that in the first stratum
_ (100)(47) 8
n, - 256 1
When the n~ have been obtained, the corresponding V( YEU ) is found by substituting in
the formula

where
Sdh 2 = S •• 2 + R h 2 S •• 2 - 2Rk S y••
The quantities Sd~ 2 are given on the extreme right of the top half of Table 6.4.

TABLE 6.5
CoMPARlSON OF FOUR METHODS OF ALLOCATION
Variance
Method of
Allocation: n. Strata
Proportional Relative
to 2 3 Total Precision

1. N" 49,824 105,833 376,215 531,872 100

2. N"S." 35,144 131,847 343,446 510,437 104
3. N"vX" 41,750 136,964 300,312 479,026 111
4. N"g" 35,144 181,710 240,888 457,742 116
174 SAMPLING TECHNIQUES

The variances and relative precisions are shown in Table 6.5.

There is not much to choose among the different allocations, as would be expected, since
the n" do not differ greatly in the four methods. Method 4, in which allocation is
proportional to the total number of peach trees in the stratum, appears a trifle superior to
the others.

6.15 UNBIASED RATlO·TYPE ESTIMATES

As we have noted, estimates of the ratio type that are unbiased or subject to a
smaller bias than R or YR may be useful in surveys with many strata and small
samples in each stratum if the separate ratio estimate seems appropriate. Three
methods that give unbiased estimates and three methods that remove the term of
order lin in the bias [see (6.33)) will be discussed brieRy.
In comparing these methods, relevant questions are : (a) Does the MSE of the
method compare favorably with that of the ordinary ratio estimate ? (b) Does the
method provide a satisfactory sample estimate of variance? This is a difficulty with
R, as we have seen. We first describe the methods . The unbiased methods require
knowledge of X.

Unbiased Methods
One estimate, due to Hartley and Ross (1954), can be derived by starting with
the mean 1 of the ratios yJ Xi and correcting it for bias .
.,
1=-n1"Lr, =-L~
1 " Y
n X,
Now
1 N _ 1 N Yi (1 N ) _
- L ri(xi-X)= - L _' X, - - L rl X
N ,_ I N I _ I X, N I- I

= Y -XE(ri)= X[R-E(r,)] (6.72)

But in simple random sampling E(1) =E(rd . Hence
1 N _
bias in f = E(f) - R = -x-N L ri(xI-X)
1
_
1
(6 .73)

By theorem 2.3, an unbiased sample estimate of

1 N _
- - L r/(xi -
N - l ,_I
X)

is

1 ~ (
- - t... r, XI-X_) = -n - (y-rx
_ __)
n-l i_ 1 n-l
 RAno ESTIMATORS 175
On substituting into (6.73), the estimate f, corrected for bias, becomes
A _ n (N - 1) _ __
RHR = r + (n -l)NX(Y - rx) (6.74)

The corresponding unbiased estimate of the population total Y is

~
J(HR
X- _X+ n(N -l)(_Y -rx
- r
--) (6.75)
n -1
By similar arguments, another unbiased estimate (Mickey, 1959) is derived
from the n ratios RI obtained by removing each unit in turn from the sample, so
r
that Rj = y/r x over the remaining (n -1) members. If it denotes the mean of
tHe Rio Mickey's estimate is

(6.76)

As a third method, Lahiri (1951) showed that the ordinary ratio estimate R is
r
unbiased if the sample is drawn with probability proportional to XI' Perhaps the
simplest method of doing this (Midzuno, 1951) is to draw the first member of the
sample with probability proportional to XI. The remaining (n -1) members of the
sample are drawn with equal probability. It is easy to prove (exercise 6.10) that
with this method the probability that a specific sample is drawn is proportional to
r Xj, and that R = t yJ t XI is unbiased for this method of sample selection.
Methods with bias of order 1/n 2
These methods consist of an adjustment to R. The first, due to Quenouille
(1956), is applicable to a broad class of statistical problems in which the proposed
estimate has a bias of order lin. It has been given the name of the jackknife
method, to denote a tool with many uses. The utility of this method for ratio
estimates was pointed out by Durbin (1959).
Ignoring the fpc for the moment, the bias of estimates Like R may be expanded
in a series of the form
A hI b2
B(R)=R+-+-+" . (6.77)
n n2
If n = mg, let the sample be divided at random into g groups of size m. From (6.77)
B(gR) = gR +~+~+
2 ... (6.78)
m gm
r
Now let RI be the ordinary ratio Y/ L x, computed from the sample after
omitting the jth group. Since RI is obtained from a simple random sample of size
 176 SAMPLING TECHNIQUES

m(g - I), we have

(6.79)

Hence
A] bl' b2
E [(g-l)R j = (g-I)R + m + (g-l)m2+· .. (6.80)

Subtraction from (6.78) gives, to order n - 2 ,

E[gR-(g-l)R.]=R b2 R _ b2 - g -
J g(g - 1)m 2 n 2(g - 1)
The bias is now of order l/n 2 • We can construct g estimates of this type, one for
each group. Quenouille's estimator (the jackknife) is the average of these g
estimates, that is,
Ro = gR - (g - I)R _ (6.8 1)
where R_ is the average of the g quantities Rj • As Queoouille showed, ·the
A

variance of Ro differs from that of R by terms of order 1/ n 2. Any increase in

variance due to this adjustment for bias should therefore be negligible in moder-
ately large samples. The choice m = 1, g = n seems best with the jackknife in small
samples.
If the fpc cannot be ignored, the leading term in the bias of R, as in (6.33), is of
the f6rm b l (1- /)/n . It can be shown (exercise 6.10) that in order to remove both
the terms in 1/" and 1/ N, we need
Ro = w}~ -( w - 1)JL (6.82)
where w = g[I-(n - m)/N], or with m = 1, g = n, w = n[I-(n - 1)/N].
Beale's (1962) estimator is
R = y +[(1 - /)/n] (Syx/.i) _ k 1 +[(1- /)/n]cyX
(6.83)
B .i+[(I - /)/n](s/ /.i) l+[(l-/)/n]c....
while Tin's (1965) is the closely related quantity
2
I(~ r= RA[ I -(1-/)(Sx
-- ~[(1
Syx)] =1(
.i2- yi I - --- /)(c.... -c yx )] (6.84)
n n

where SyX= I" (YI- y)(x, - .i)/(n - 1), s/ = I" (XI - .i) 2 /(n - 1), so that cyx , c;.. are
the sample relative covariance and relative variance of x.
The structure of RT may be seen by noting that from (6.34) the leading term in
the expected value of R may be written

R [ 1 +-(I n--/)(Cu - Cyx ) ]

 RATIO ESTIMATORS 177
Evidently, Rr is adjusting R by a sample estimate of the needed bias correction.
Rs and Rr are identical to terms of order 1/ n and in general perform very
similarly.

6.16 COMPARISON OF THE METHODS

Example. The artificial population in Table 6.6 contains three strata with Nh = 4, nh =
2 in each stratum. The population was deliberately const{lJcted so that (a) R h varies
markedly from stratum to stratum, thus favoring a separate 'l'.atio estimate YR " and (b) R h
overestimates in each stratum, with the threat of a serious cumulative bias in YR ,. The

TABLE 6.6
A SMALL ARTIFICIAL POPULATION
Stratum

II III

Y x y x y x
2 2 2 3 I
3 4 5 4 7 3
4 6 9 8 9 4
11 20 24 23 25 12

Totals 20 32 40 36 44 20
Rh 0.625 1.111 2.200

following methods of estimating the population total Y were compared .

Simple expansion: L NhYh
Combined ratio : (y"ji,,)X
Separate ratio : L (Yhli.) X h =L R~h
The remaining estimate, the separate Hartley-Ross, Lahiri, and QuenouilleJ Beale ... and
Tin methods, have the same form as the eparate ratio estimate. except that R (HR)h, R (Ll h,
and 0 on. replace Rh . (For nh =2. the Hartley-Ross and Mickey methods are identical.)
There are 6 3 = 216 possible samples. Results are exact apart from rounding error . For help
in some computations I am indebted to Dr. 10 eph Sedransk.
The results in Tl.1ble 6.7 show some interesting features . For the combined ratio estimate,
the contribution of the (bias)' to the mean square error is trivial, de pite the extreme
conditions. but this estimate does poorly as regards variance because of the wide variation
in the R h • As judged by the MSE, the separate ratio estimate i much more accurate than
the combined estimate, but it is badly biased. Of the unbiased methods. Hartley-Ro
shows relatively high variance, as it has been found to do for fl" = 2 in some studies on
natural populations. The Lahiri method does particularly well This population suits the
Lahiri method because one unit in each stratum has unu ually high values of both Yt and x"
 178 SAMPl..ING TECHNIQUES

TABLE 6.7
RESULTS FOR DIFFJ;RENT ESTIMATES OF Y
Method Variance (Bias)' MSE

Simple expansion 820.3 0.0 820.3

Combined ratio 262 .8 6.5 269.3
Separate ratio 35.9 Rl 60.0
Separate Hartley-Ross ]53 .6 0.0 153.6
Separate Lahiri ]9.6 0.0 19.6
Separate Ouenouille 42.9 1.1 44 .0
Separate Beale 28.9 8.0 3(..9
Separate Tin 28.6 5.7 34.3

with the consequ~nce ~ that this un:: has a high probability of being drawn and that sample
containing this unit give gnod estimates of R h •
The Ouenouille. Beale. and Tin methods all produced substantial decreases in bias as
compared with the separate ratio estimate, and all had smalle r MSE's, so that in this
ex 'mple they achieved their principal purposes.
The . tudy by J. N. K. Rao (1969) of natural populations cited in section 6.9
compared the Quenouille, Beale, and Tin methods for n = 2, 4, 6, 8, on 15 such
populations. For n = 2, the most severe test, the medians and the upper quartiles
of the quantities I biaslNMSE were as follows : Ro, 3% , 7% ; RB , 8% , 12%; Rr,
8%,19%; as against 15% , 20% for k The more complex methods appear to help
ma.terially as regards bias in these tiny amples.
The same study compared the MSE's of five of the estimates in this ection with
that of R. (Lahiri 's method was omitted, since the study Wl'.s confined to simple
random samples.) For each method the ratio 100 M SE(P ) )/MSE(R), and so
forth, was calculated for each population.
For n = 4, Quenouille' and Mickey's estimates weIr slig ltly inferior to R in
these populations but, for n ~ 6, all methods had av rage ~ ISE's very close to
those of R. For a very mall sample from a single pOpul:Hion, this study suggests
that these more comj'k l( methods have no material advantage in accuracy ever k
But the fact that they reduce bias with little or no increase in MSE in a single
stratum should give them an advantage in a separate ratio estimate with numerous
strata having small samples.
Under a linear regression model, comparisons of the MSE's of these methods
for small n by P.S.RS. Rao and J. N. K. Rao (1971), Hutchinson (1971), and J. N.
K. Rao and Kuzik (1974) gave results in general agreeing with those from "the
natural populations.

6.17 IMPROVED ESTIMATION OF VARIANCE

One estimator worth consideration for moderate or small samples was
suggested by Tukey (1958) for thl! jackknife (QuenouilJe's) estimate Ro. With g
 RATIO ESTIMATORS 179
groups, n == mg, and f negligible, Ro is the average of the g quantities
R,== gR - (g - 1)R,
where RI = I y/I x, computed after omitting group j . If the R; could be regarded
as g independent estimates of R then, with simple random sampling, an unbiased
estimate of V(Ro) would be
, ... t ... 2
vCR )",,(1-f)I(R i -Ro ) (6.85)
o g (g . . . 1)
Since R; - Ro = -(8 - 1)(Ri - R_), (6.85) is more easily calculated as

v(Ro) (1- f)(g - 1)

g
f (R, - R - f (6.86)

where Ji- is tFe mean of the g quantities RI'

The R; or R ; for different j are, of course, not independent, and formula (6.86)
is an approximation. So far, the analytical properties of v(Ro) have been
established only for large amples. Arvesen (1969) showed that for a broad
class of estimates inducting Ro that are symmetrical in the elements of the sample
[estimates known as Hoeffding's U-Statistics (1948)], the formula v(Ro)
becomes unbiased either for fixed g or for g = n as n becomes large.
From Rao 's (1969) study of eight natural populations, small-sample average
percent underestimations in the standard vl(k, as an estimate of the true
MSE(R) were reported in section 6.9 for n == 4. 6, 8, 12. The corresponding
average percent biases in v(Ro) are shown for comparison in Table 6.8; these are
the averages of the eight number ; 100[v(Rol - MSE(Ro)]/MSE(R o )'
Table 6.8 also gives the averages of the quantities 100 [v(Ro)-
MSE<fi)]/MSE(R). These avera$es are ,of interest to the inves!igator who
uses R but i willing to replace v,(R) by v(!?o} as an estimator of V(~) if it seem
less biased. In view of the biase in Rand Ro. the comparisons of V(Ro) are made
with the MS£'s as more appropriate.

TABLE 6.8
A VERAOE PERCENT BIAS IN ESTIMATORS OF V ARJANCE.
n=
Average of 4 6 8 12

1OO(v(Ro) - MSE(Rol]/MSE(R o ) +11% +10% +6% +1%

100(v(Rpl - MSE(~)]/MSE(~) + 11 % + 10% +6% + 1%
I OO{v,(R) - MSE(R)J/MSE(R) -3 1% -23% -21 % - 18%

In these populations v(Ro) is a slight overestimate of botb MSE(R o ) and

MSE(rt). while vI(R) has substantialllf'lgative biases in these small sample~.
 180 SAMPLING TECHNIQUES

The stability of v(.Ro) relative to that of vlR), as judged by the squares of the
coefficients of variation of these variance estimates, was poor in these samples. In
studies by Rao and Beegle (1968) of v(Ro) and VI (R) under a linear re~ression
model of y on x in an infinite population with x normal, however, v(Ro) and
vl(R) appeared about equaUy stable for n =4 to n = 12.
With a separate ratio estimate and numerous strata these results suggest that
L Xh2V(R oh ) is superior to L X,/VI(l~h) as an estimator of VO\., ). The former is
likely to be freer from bias and both should have adequate stability. But with only
a few strata the issue is questionable until further comparisons appear.

6.18 COMPARISON OF lWO RATIOS

In analytical surveys it is frequently necessary to estimate the difference R - R'
between two ratios and to compute the standard error of R - k. The formulas
given here are for the estimated variance of R - R', si nee these are the ones most
commonly required. The fpc terms are omitted for reasons presented in section
2.14.
Simple random sampling is assumed at first. Three cases can be distinguished.

The Two Ratios Are Independent

This occurs when the units are classified into two distinct classes and we wish to
compare ratios estimated separately in the two classes. For instance, in a study uf
house'hold expenditures, a simple random sample of households might be sub-
divided into owned and rented houses in order to compare the proportions of
income spent on upkeep of the house in the two classes. If tbe estimated ratios are
denoted by R = 9/f, k = f/i', then .

6.87)

The Two Ratios Have the Same Denominator

When the unit is a cluster of families, we might wish to compare the proportion
of adult males who use electric shavers with the proportion who use razors. In any
unit, y = number of adult males using electric shavers, y' == number of adult males
using razors, and x = total number of adult males.
A A,
- -,
y-y
R - R =--
i

If d j = Yj - y/, the estimated variance of R - R' may be computed as

(6.88)
 RATIO ESTIMATORS 181
The Two Ratios Have Different Denominators But May Be Correlated ..
An example is the comparison of the proportion of men who smoke with the
proportion of women who smoke, in a survey in which the unit is a cluster of
house . Mathematically, this is the most general case.
v(R - R') = v(R)+v(R')-2 COy (RR') (6.89)
y
The oni unfamiliar term is cov (RR') . Writing, in the usual way,
A y- Rx R' - R" Y'- R'x'
R - R=---
X X'
we have
COy (RR') = k, COY (YI- Rx;)(y/ - R'x,')

A sample estimate may be computed as follows:

... ,.. 1 " A ,.. "" ,..

cov(RR')' ( l ) __ ,L(Yly/-Ry/x;-R'y,x/+RR'x,x/) (6.90)

nn- xx

Example. The 1954 field trial of the Salk polio vaccine wa conducted among children
in the first three grades in all schools in a number of counties. The counties were not
randomly selected, since those with a history of previous polio anacks were favored, but fOI
this illustration, it will be assumed that they are a random sample from some population.
Children whose parents did not give permission to participate in the trial were called the
" not inoculated" group and, of course, received no shots. Half of the children who received
permission were given three shots of an inert liquid anc! were called the "placebo" group.
From the data in Table 6.9, compare the frequencies R, R' of paralytic polio in the "not
inoculated" and " placebo" groups. To reduce the amount of data, the comparison is
.restricted to 34 counties, each having more than 4000 children in the two groups combined.
In these data any variation in the polio attack rate from county to county would produce a
positive correlation between.R. and .R.'.
The following quantities are derived from the totals.

Placebo '. .R. = ~ '"' 0 525687 ,

167.4'
i= 167.4 =4.9235
34

Not inoculated: R' "" 2::6 "" 0.347857, x' = 284.6 = 8.3706
34
For v(Jh v(R') and coy (M'), all uncorrected sums of squares and products among the
four variates are required.

v(R)= ( 1 ) 2(l:y2_2Al:YX+ R 2l:x2)

n n- 1i
1
=- (34)(33)(4.9235)2 [(564) - (1.05137)(822.2) + (0.27635)(1661.92)J

= 0.00584
 182 SAMPLING TECHNIQUES

TABLE 6.9
NUMBER. OF CKILOREN (~, X') AND OF PAIlAl.YTIC CASES (y, y') PEIl CoUNTY
X· ~' yt y' ~ x' y y'

4.1 2.4 0 0 13.8 25.6 3 3

3.5 8.0 1 6 10.5 8.1 2 0
4.1 6.1 7 2 21.6 25 .9 10 7
2.6 4.6 2 3.5 6.7 2 2
2.4 1.5 2 1 6.8 7.3 3 8
2.2 1.9 0 0 2.3 3.7 0 I
1.1 4.0 I 2.6 2.9 2 0
1.6 4.0 2 6.0 11.1 3 1
5.7 7.8 1 4 11.0 14.8 7 11
3.3 11.0 3 7 19.4 42.5 II 14
1.0 3.8 0 I 6.8 13.7 6 2
2.0 5.2 0 1.2 4.0 3 I
8.3 19.0 4 4 5.4 9.3 II 6
1.0 3.7 I 5 1.7 2.6 0 2
1.1 4.2 0 I 2.1 2.3 0 0
2.3 6.8 I 2 I.S 2.6 0 0
1.9 3.5 0 2 3.0 4.0 0 2

Totals 167.4 284.6 88 99

" • ~.~' "" numbers of "placebo" and "not inoculated" children (in 1000's)
t Y. y' = numbers of paralytic polio cases in the placebo and not inoculated
groups

Similarly. we find v(R') = 0.00240.

1 ~
COy (RR') ---_-_,(Lyy'- l< Ly'x-R'Lyx'+RR'LXX')
A AA

n(n - l)xx
(497) - (O.52569)(844.6) -{O.34786)(1397 .4)
+ (0.52569)(0.34786)(2690.8)
(34)(33)(4.9235)(8.3706)
= 0.00127
Hence

s.e.(R - R') == JO.00584 + 0.00240 - 0.00254 = 0.0754

Since R - R' = 0.1778, the difference approaches significance at tite 5% level (thedistribu-
tion of R - R' may be somewhat skew for this size of sample). A possible explanation is that
the not-inoculated children may have had more natural protection against polio than the
placebo children.
 RATIO ESTIMATORS 183
The same problem may arise in stratified samples in which the domains of study
cut acro s strata. If Rh , Rn' appear to vary from stratum to stratum, the
comparison will probably be based on an examination of the values of Rn - Rk ' in
individual strata. By finding the standard errors of Rh - RI.' it is possible to
determine whether these differences vary from stratum to stratum and, if not, to
compute an efficient over-all difference.
If the Rn, RIr ' exhibit no real variation from stratum to stratum , it may be
sufficient to compare the combined estimates Rc and Rc'. As before,
vCRc- Re') = v(Rc)+vCRe') -2 cov CR~/) C6 .91 )
where, putting d h, = (YhI- Yh) - Re(Xhi -.ih),
..
v(R c) =;
1
-=2 r. (N/ r.
X$I h nh nh -
2
1) d h•
,
(6.92)

(6.93)

A more thorough discussion of the comparison of ratios, including shortcut

computing formulas when the sample permits them, has been given by Kish and
Hess (l959b).

6.19 RATIO OF TWO RATIOS

In some applications we want to estimate the ratio RI R' of two ratios . Thus, in
the preceding example (section 6.18), we might be interested in the ratio R I R ' of
the paralytic polio case rates for " placebo" and " not inoculated" children, or the
ratio of the proportions of males and females in the labor force from a cluster
sample. If data on (y,x ) are available in the same sample for two time periods, the
quantity might be the ratio of the weekly expenditures on food per household at
the two times.
With a simple random sample (e.g., of clusters) the sample estimate of RI R ' is
RI k = (YI.i)/Y'I.i') , sometimes called a double ratio estimate. As with the single
ratio the leading term in the bias of RI k is of order lin, but is more complex than
for R or k . We may write

RI k = (RI R ')(1 +89)(1 +8.i')/(1 +8.i)(1 +8.Y') (6.94)

where 8y denotes (9 - 9)1 Y, and so forth. When this expression is expanded,

there are six quadratic terms of order 1/ n that enter into the bias of RI k . Rao
and Pereira (1968) give an exact expression for the bias.
Formula (6.8) for the relative variance V(R)/ R 2= CAR of a ratio can be
written
 184 SAMPLING TECHNIQUES

From this, the leading term in V(R/ k) is

" IB) 2(CAA +CA'.V:- 2CItA· )

V(R/R')=\jf (6.95)

where
.C RR · = Cyy'+ Cu,- Cv1,-Cn , (6.96)
, For the corresponding sample estimate v(R/ k ) we substitute sample estimates of
the terms in (6.95).
blimp/e. For the ratio of placebo and not inoculated case rates, R/ R =
0.52569/0.34786 = 1.511. Estimate the s.e. of this ratio. The computations in the preced-
ing example give
A 0.00584 c _ 0.00240 _ .
CAR = (0.5257)2 0.0211 ; A'A' - (0.3479)2 0.0198,

C 0.00127 00069
AA' = (0.5257)(0.3479) = . 4
veRI R') =(1.511)2(0.0211 + 0.0198) -0.0139) = 0.0617
s.e. (R/ R') = 0.248
The double ratio estimate has occasionally been used in place of YR = Ax "to estimate a
population total Y, as suggested by Keyfitz (Yates, 1960). Suppose that R '=(y'/x') is
known for the same sample from a previous period and that R' = Y'/ X' is also known. If
R'/ R~has been found, say, to be slightly > 1, we might argue intuitively that R is also likely
to give an overestimate of R that should be adjusted downward by dividing it by the ratio
R'/R'. This leads to the double ratio estimate YOR •
R' R
YOR =If(Rx)=lf(R 'X) (6.97)

Since the relative variance of YR is C AA , while that of YOR is

CAA +CA ·A· - 2CRA· (6.98)
the dou~le ratio will give a more precise estimate in large samples if the correlation between
Rand R' is nigh enough.

6.20 MULTIV ARIATE RATIO ESTIMATES

Olkin (1958) has extended the ratio estimate to the situation in which p
auxiliary x-variables (x~. X2, •.. x p ) are available. For the population total , the
proposed estimate, say Y MR for multivariate ratio, is
 RATIO ESTIMATORS 185
where the Wi are weights to be determined to maximize the precision of YMR.
subject to L WI = 1. This type of estimate appears appropriate when the regres-
sion of y on x" Xl • ••. Xp is linear and passes through the origin. The population
totals X, mu t be known .
The method is described for two x-variate. ince ' this should be the most
frequent application. We have
YMR - Y= W,CYR , - Y)+ W2(YR1 - Y)
Hence. assuming negligible bias.
.... .,.... ... A 2 A

V(YMR ) = W, - V(YR ,)+2 W, W 2 cov (Y R , Y R , ) + W 2 V(YR 2 )

= W, l V" +2 W, W z V 12 + W 22 V 12 (6.99)
where VII :: V(YR ,) , etc. The values of WI. W 2 • that minimize the variance.
subject to W, + W 2 = I, are found to be
V n - V'\2
W, =--=--..:=-.-
VI' + V n - 2 V' 2'
and the minimum variance is

(6.100)

With p variates, it is necessary to compute the inverse V" of the matrix V.i' Then
the optimum W, = L.lI , where L, is the sum of the elements in the ith column of
V" and L is the sum of all the p 2elements of V,i. The minimum variance is IlL.
1n practice. the weights are determined from estimated variances and
covariances VI/" From (6.7) in ection 6.3,
(1 - I) y2
';"_~- (cyv +clI - 2c yl)
n

where Cyy "" s/ I9z • etc. The covariance can be expressed as

(I-I) y2
VJ Z= (Cy y +Cll - Cyl-Cy Z)

. n
A convenient method of computation is first to obtain the matrix

If VI/ = 'Iv,'/(I- I) yl, the matrix v,/ is easily obtained by taking diagonal contrasts
186 SAMPLING TECHNIQUES

in C, that is,
VII' = C yy + C 11 - C y 1 - C y I

Vl2' =C yy +C12 -CyJ -C y 2 etc.

The factor (1- f) y 2/ n ·is not needed when computing the Wi, but it must be
inserted when computing the minimum variance. Thus
", )_(1_f)y2 (VJ/V22'-VI2,2)
V m ,"
( r MR - (" 2 ') (6.101)
n VII +V22 - Vl 2

In view of the amount of computation involved, this estimate will probably be

restricted to smaller surveys of specialized scope. The method is capable of giving
a marked increase in precision over YR , or YR , alone.

6.21 PRODUcr ESTIMATORS

If an auxiliary variate x has a negative correlation with y where x and yare
variates that take only positive values, a natural analogue of the ratio estimator is
the product estimator, for which

(6 .102)

By the usual Taylor series expansion, the analogue of (6.8) for the product
estimator in a large simple random sample is
·f
(6.103)

where (CV)2 is the square of the coefficient of variation of either of Yp. P.S.R.S. Yp
Rao and Mudholkar (1967) have extended Olkin's multivariate ratio estimator to
a weighted combination of ratio estimators (for Xi positively correlated with y)
and product estimators (for XI negatively correlated with y) .

EXERCISES
6.1 A pilot survey of 21 households gave the following data for numbers of members
(x). children (y,). cars (Y2), and TV sets (Yl) .

x YI Y2 Ya x Yl Y2 Ya x YI Y2 Ya

5 3 1 3 2 0 0 1 6 3 2 0
2 0 1 I 3 1 1 1 4 2 1 I
4 1 2 0 2 0 2 0 4 2 I 1
4 2 I 1 6 4 2 1 3 1 0 1
6 4 1 1 3 1 0 0 2 0 2 I
3 1 1 2 4 2 1 J 4 2 I I
5 3 I 1 5 3 1 I 3 I J J
 RATIO ESTIMATORS 187
Assuming that the total population X is known, would you recommend t-kat ratio
estimates be used instead of simple 'expansions for estimating total numbers of children,
cars, and 1V sets?
6.2 l~ a field of barley the grain, y" and the grain plus straw, x" were weighted for each
of a large number of sampling units located at random over the field . The total produce
(grain plus straw) of the whole field was also weighed. The following data were obtained :
t:" = 1.13, cY' = 0.78, c... = 1.11. Compute the gain in precision obtained by estimating the
grain yield of the field from the ratio of grain to total produce instead of from the mean yield
of grain per unit.
It requires 20 min to cut, thresh , and weigh the grain on each unit, 2 min to weigh the
straw on each unit, and 2 hr to collect and weigh the total produce of the field . How many
units must be taken per field in order that the ratio estimate may be more economical than
the mean per unit?
6.3 For the data in Table 6.1, YR =28,367 and C09 = 0.0142068, Co. = 0.0 146541 ,
en = 0.0156830. Compute the 95% quadratic confidence limits for Y and compare them
with the limits found by the normal approximation.
6.4 The values of y and x are measured for each unit in a simple random sample from a
population . If X, the population mean of x, is known, which of the following procedures do
you recommend for estimating Y/X"I (a) Always use y/X. (b) Sometimes use y/X and
sometimes y/i. (e) Always use i ii . Give reasons for your answer.
6.5 The following data are for a small artificial population with N = 8 and two strata of
equal size .
Stratum I Stratum 2

X II 111 ( x 2,t Y2 i
2 0 10 7
5 3 18 15
9 7 21 10
IS 10 25 16

For a stratified random sample in whicb"1 = "2= 2. compare the MSE's of YR. and YR c by
working out the results for all possible samples. To what extent is the difference in MSE's
due to biases in the estimates?
6.6. In exercise 6.5 compute the variance given by using Lahiri ' method of sample
election within each stratum and a eparate ratio estimate.
6.7 Forty-five states of the United States (excluding the five largest) were arranged in
nine strata with five states each, states in the same stratum having roughly the same ratio of
"h
1950 to 1940 population . A stratified random sample with = 2 gave the following results
for 1960 population (y) and 1950 population (x), in millions.

Stratum
2 3 4 5 6 7 8 9

YAI 0.23 0.63 0.97 2.54 4.67 4.32 4.56 1.79 2.18
XAl 0.13 O.SO 0.91 2.01 3.93 3.96 4.06 1.91 1.90
VAl 4.95 2.85 0.61 6.07 3.96 1.41 3.57 1.86 1.75
Xu 2.78 2.38 0.53 4.84 3.44 l.3a 3.29 2.01 1.32
188 SAMPLING TECHNIQUES

Give that the 1950 population total X is 97.94, estimate the 1960 population by the
combined ratio estimate. Find the standard error of your estimate by Keyfitz' short-cut
method (section 6.13). The correct 1960 total was 114.99. Does your estimate agree with
this figure within sampling errors?
6.8 In the example of a bivariate ratio estimate given by Olki~, a sample of 50 cities was
drawn from a population of 200 large cities. The variates y, X" Xl are the numbers of
i~abitants ~er city in 1950,1940, and 1930, respectively. For the population, Y== 1699,
X, == 1482,X2 == 1420(in 100's) and, for the sample, y = 1896, x, = 1693, Xl = 1643 . The C
matrix as defined in section 6.20 is
y X, Xl
1.213 1.241 1.256
1.241 • 1.302 1.335
1.256 1.335 1.381
Estimate Y by (a) the sample mean, (b) the ratio of 1950 to 1940 numbers of inhabitants,
and (c) the bivariate ratio estimate. Compute the estimated standard error of each
es ·mate.
6.9 Prove that with Midzuno's method of sample selection (section 6.15) the probabil-
ity that any specific sample will be drawn is
(n - l)!(N - n)! I" (X,)
(N-l)! X
6.10 In small populations the leading term in the bias of R in simple random samples of
size n is of the form

E(R - R) = b,(I - f) =!!.!_!!.!

., n n N
where b, does <oot depenq on n, N,. If n = mg and. the sample is divided at random
into g groups of size m, let R{ == ~ y/'i x taken over the remaining (n - m) sample members
when group i is omitted from the sample. Show that in the bias of the estimate
wR -(w - l)R{
both terms in b, vanish if w = gel - en - m)/N).
CHAPTER 7

Regression Estimators

7.1 THE LINEAR REGRESSION ES11MATE

Like the ratio estimate, the linear regression estimate is designed to increase
precision by the use of an auxiliary variate Xi that is correlated with Yi' When the
relation between YI and XI is examined, it may be found.that although the relation
is approximately linear, the line does not go through the origin. This suggests an
estimate based on the linear regression of Yi on Xi rather than on the ratio of the
two variables.
We suppose that Y, and Xi are each obtained for every unit in the sample and
that the population mean X of the X, is known. The linear regression estimate of
Y, the population mean of the YI, is
Ylr = Y + b(X - x) (7.1)
where the subscript /r denotes linear regression and b is an estimate of the change
in Y when x is increased by unity. The rationale of this estimate is that if x is below
average we should expect y also to be below average by an amount b(X - x)
because of the regression of Yi on Xi' For an estimate of the population total Y, we
take fir = NYI,.
Watson (1937) used a regression of leaf area on leaf weight to estimate the
average area of the leaves on a plant. The procedure was to weigh all the leaves on
the plant. For a small sample of leaves, the area and the weight of each leaf were
determined. The sample mean leaf area was then adjusted by means of the
regression on leaf weight. lhe point of the application is, of course, that the weight
of a leaf can be found quickly but determination of its area is more time
consuming.
This example illustrates a general situation in which regression estimates are
helpful. Suppose that we can make a rapid estimate XI of some characteristic for
every unit and can also, by some more costly method, determine the correct value
Yi of the characteristic for a simple random sample of the units. A rat expert might
make a quick eye estimate of the number of rats in each block in a city area and
then determine, by trapping, the actual number of rats in each of a simple random
sample of the blocks. In another application described by Yates (1960), an eye
189
190 SAMPLING TECHNIQUES

estimate of the volume of timber was made on each of a population of ie-acre

plots, and the actual timber volume was measured for a sample of the plots. The
regression e timate
y+b(X-i)

adjusts the sample mean of the actual measurements by the' regression of the
act'lal measurements on the rapid estimates. The rapid estimates need not be free
from bias. If Xi - Yi = D , so that the rapid estimate is perfect except for a constant
bias D, then with b = 1 the regression estimate becomes
y + (X - i) = X + (y - i)
= (pop. mean Of rapid estimate)+(adjustment for bias)

If no linear regression model is assumed, our knowledge of the properties of the

regression estimate is of the same scope as our knowledge for the ratio estimate.
The regression estimate is consistent, in the trivial sense that when the sample
comprises the whole popu'lation, i <= X, and the regression estimate reduces to Y.
As will be shown, the estimate is in general bia ed, but the ratio of the bias to the
standard error becomes small when the sample is large . We possess a large-sample
formula for the variance of the estimate, but more information is needed about the
distribution of the estimate in small samples and about the value of n required for
the practical usc of large-sample results.
By a suitable choice of b, the regression estimJte includes as particular cases
both,he mean per unit and the ratio estimate. Obviously if b is taken as zero, YI,
reduces to y. If b = y/ x,
9- y-:.
YI, == Y+-:(X - xl = -:X = Y R (7.2)
X X

7.1 REGRESSION ESTIMATES WITH PREASSIGNED b

AI~ho1Jgh , in most applications, b is estimated from the results of the sample, it
is sometimes reasonable to choos the value of b in advance . In repeated surveys,
previous calculations may have been shown that the sample values of b remain
fairly constallt ; or, jf x is the value of Y at a recent census, general knowledge of
the po ulation may suggest that b is not far from unity, so that b = 1 i~ I ho ,en .
Since the sampling theory vf regre sion estimates when b is preassigned IS both
simple and informative, this case i considered first.

Theorem 7.1. In simple random sampling, in which bo is a preassigI\ed

constant, the linear regression estimate

91' = 9 + boex - x)
 REGRESSION ESTIMATORS 191
is unbiased, with variance
N - - 2
_ 1- f i~1 [(y, - )') - bo(x,- X)]
V(YI,) = -;- N -1 (7.3)

= 1 - f (5/ - 2b 0 5 yx + bo25/) (7.4)

n
Note that no assumption is required about the relation between Y and x in the
finite population.
Proof. Since bo is constant in repeated sampling,
£(YI,) = E(y) + boE(i - X) = y (7.5)
by theorem 2.1. Furthermore, YI, is the sample mean of the quantities YI-
bo(x,- X), whose population mean is Y. Hence, by theorem 2.2,
N - - 2
_ 1- j~l [(Yj - Y) - bo(xj - X)]
f
V(YI,) = ---;- . N- 1 (7.6)

= I-f(5/-2b05yx+b025/) (7.7)
n
Corollary. An unbiased sample estimate of V(YI,) is

_ 1- f t
[(YI- y) - bo(xj - i)]2
V(YI,) = -n- '--'----n---1--- (7.8)

_ I-I( 2 ")b
- - - Sy - ... oSyx + b0 2 Sx 2) (7 .9)
n
This follows at once by applying theorem 2.4 to the variate YI':' bo(xi - X).
I

A natur':tl question at this point is: What is the best value of bo? The answer is
given in theorem 7.2.
Theorem 7.2. The value of b o that minimizes V(YI,) is
N _ _

5 L (y, - y)(XI-X)
b
o
= B =:;A = '-i-_I_ _ _ _ __ (7.10)
5x f (XI -X)2

which may b .;: called the linear regression coefficient of y on X 10 the finite
population. Note that B does not depend on the properties of any sample that is
192 SAMPLING TECHNIQUES

drawn, and therefore could theoretically be preassigned. The resulting minimum

variance is

(7.11 )

where p is the population correlation coefficient between y and x.

Proof. In expression (7.4), for VC.YI,), put

(7.12)

This gives

V(yI, )=1- I [S2_2S (SYX+ d)+S 2(SY/+2~YX+d2)]

n Y yx S/ x S/ S/
= l~f[ (S/- ~:22)+d2S/] (7.13)

Clearly, this is minimized when d = O. Since p 2 = Sy/ / S/ S/,

- I - f S. 2 (1 -
V ...," (YI,) = -
2
p ) (7 .14)
n
The same analysis may be used to show how far b o can depart from B without
incurring a substantial loss of precision . From (7 .13) and (7. 14),

(7. 15 )

(b - B)2S 2]
= Vrni .. (YI,) [1 + S~2( 1 _ P 2) (7. 16)

Since BSx = pSy, this may be written

V«YI, ) = V"'i" (YI,) [1 + (~ -1r (1 ~:2)] (7.17)

Thus, if the proportional increase in variance is to be less than a , we must have

(7.18)

For example, if p = 0.7, the increase in variance is less than 10%, (a == 0.1),
provided that

I~ -11 < J(0.1)(O.51)/(0.49) = 0.32

 REGRESSION ESTIMATORS 193
EXpression (7.18) makes it clear that in order to ensure a small proportional
increase in variance bol B must be close to I if p i very high but can depart
substantially from 1 if p i only moderate.

7.3 REGRESSION ESTtMATES WHEN h IS COMPUTED FROM

THE SAMPLE
Theorem 7.2 suggests that if h must be computed from the sample an effective
estimate is likely to be the familiar least squares estimate of B, that is,

L" (y, - y)(x, - x)

b= '-'-....:1_ _ _ _ __
(7.19)

The theory of linear regression plays a prominent part in statistical methodol-

ogy. The standard results of this theory are not entirely suitable for sample surveys
because they require the assumptions that the population regression of y on x is
linear, that the residual variance of y about the regression line is constant, and that
the population is infinite. If the first two assumptions are violently wrong, a linear
regression estimate will probably not be used. However, in surveys in which the
regression of yon x is thought to be approximate ly linear, it is helpful to be able to
use Ylr without having to assume exact linearity or constant residual variance.
C:onsequently we present an approach that makes no assumption of any specific
relation between y, and x,. As in the analogous theory for the ratio estimate, only
large-sample results are obtained.
With b as in (7 .19), the linear regression estimator of Y in simple random
samples is
Ylr =Y+ b(X - x) =Y- b(x - X) (7.20)

The estimator Yin like YR, will be shown in section 7.7 to have a bias of order II n.
In finding the sampling error of YI" replace tht< sample b in (7.20) by the
population regression coefficient B in (7.10). In Theorem 7.3 the error committed
if" this approximation will be shown to be of order 1/-rn
relative to the terms
retained. We first examine the relation between band B.
Introduce the variate t!j defined by the relation
(7.21)
N
Two properties of the ej are that L e, = 0 and

N N N
L el(x, -X}= I (YI- Y)(x, - X}-B L (Xi _ X)2=0 (7.22)
194 SAMPLING TECHNIQUES

by definition of B. Now

b= f Yl(Xi - x)/f (Xi - X)2

=f [Y+B(Xi -X}+el ](Xi -X)/ L(XI - X?
(7.23)

A result needed in Theorem 7.3 is that (b - B) is of order l/Fn. By theorem 2.3,

n N
L el(xl - i)/(n - 1) is an unbiased estimate of
n
L e, (x, - X}/(N - 1) which, by
(7.22), is zero. Thus L ei(x, - x)/(n - 1) is distributed about a zero mean in
repeated samples. Since the standard error of a sample covariance is known to be
f
of order l/Fn, e,(x, - i)/(n -1) iii of order l /Fn. But L (Xi - x // (n - 1) == s/ is
of order unity. Hence (b - B ), which from (7.23) is the ratio of the e two
quantities, is of order I/Fn.

Theortm 7.3. If b is the least squares estimate of Band

Ylr =Y+ b(X - i) (7.24)
then in simple random samples of size n, with n large,

V(Ylr) == (1- f) S.2( I - p 2) (7.25)

n
where p = Sv,J SySX is the population correlation between y and x.
Proof. The sampling error of .)ilr arises from the quantity
y" - y =Y- Y +b(X - x) (7.26)
As an approximation, replace YI, by
(7.27)
where B is the population linear regression coefficient of y on x. The error
committed in this approximation is (B - b )(X - x). This quantity is of order 1/ n in
a simple random sample of size n, since (b - Bland (i - X) are both of order l/Fn.
But the sampling error in Ylr is of order I/Jn, since it is the error in the sample
mean of the variate (Yi - Bxd . Hence the leading term in B(Ylr - Y)2is V(Ylr) ' By
(7.11), in large samples,

-
B(YJr- YI, =(I-f)S
y)- 2=. V(-) - - ) '2(1 -p 2) (7.28)
11
 REGRESSION ESTIMATORS 195
7.4 SAMPLE ESTIMATE OF VARIANCE
As a sample estimate. of V(YI,), valid in large samples, we may use

(7.29)

(7.30)

the latter being the usual short-cut computing formula. The derivation is as
fonows.
In theorem 7.3, equation (7.28), we had, since S/O- p2) = S/,

From theorem 2.4, an unbiased estimate of S/ is

2 1 ~ 2
S. = -- t.. (el-i)
n - 1i- 1
Now, from equation (7.21), it follows that

el-e = (YI- Y)- B(x,-.i) = [(YI- y) -b(x; - .i»+(b - B)(x,-.i) (7.31)

The second term on the right, of order 1/Fn, may be neglected in relation to the
first term, which is of order unity. Hence in large samples we may ·..tSe

(7.32)

as an estimat~ of S/. The divi or (n - 2) instead of (n -1) i suggested in

(7.29) and (7.30) because it is used in standard regression theory and is known to
give an unbiased estimate of S/ if the T'opulation is infinite and the regression
is linear.

7.5 LARGE-SAMPLE COMPARISON WITH 'mE RATIO

ESTIMATE AND THE MEAN PER UNIT
For these comparisons the sample size n must be large enough so that the
approximate formulas for the variances of the ratio and regression estimates are
valid. The three comparable variances for the estimated population mean Yare as
follows.
196 SAMPLING TF.C'HNh)UES

(regression)

(ratio)

_ N-n ,
V()'):::: Nn S, - (mean per unit)

It is apparent that the variance of the regre!>sion estimate is ~maller than thai of
the mean per unit unle!.s p = 0, in which case the two variances are equal.
The variance of the regression e~timate is less than that of the ratio e~timate if
(7.33)

This is equivalent to the inequalities

(pS, - RS, )2 "> 0 or (7.34)

Thus the regression estimate i more precise than the ratio estimate unless
X, i~ a straight line through
B :::: R . This occurs when the relation between Y, and
th .... origin.

Example. The precision of the regression. ratio. and mean per unit estimates from a
simple random sample can be compared hy using data collected in the complete enumera-
tIOn of peach orchards described on p. I 7~. In thb example. )', is the estimated peach
pr9,duction in an orchard and x, the number of peach tree, In the orchard. We will compare
the' estimates of the total production oC the 256 orchard~. made from a sample of 100
orchards. It is doubtful whether the ample is large enough to make the variance formula
fully valid, since tbe cv's of y and i are both omewhat higher than 10%. but the example
.will serve to illustrate the computations. The basic data are a~ follows .
~
S, 2 =6409 S,. = 4434 S, 2 = 3898
R = 1.270 p =0.887 fI = 100 N = 256

v('Y,:) = N(N - nl S / (l _ p 2)
n

= (256)(156) (6409)( 1 - 0 787) = 545 000

100 . .

• _ N(N-n) 2 2'
V(YR ) - (S. + R S, - 2PS.. )
n

= (25~~~56) [6409+ (1.613)(389R) -2(1.270)(4434)]

= 573,000

V(n =N(N -11) S.2= 2.559,000

n
 REGRESSION ESTIMATORS ]97
There is little to choose between the regression and the ratio estimates, as might be
expected from the nature of the variables. Both techniques are greally superior to the mean
per unit .

The preceding result on the superiority of the linear regression estimate is

strictly a large-sa mple result. In small samples on natural populations the regres-
sion estimate appears disappointing in performance. In eiE,ht natural populations
of the type in whi ch the ratio estimate has been used . Rao (1969) found in a Monte
Carlo study that the average of the rati~ MSE(Yi,)/MSE(YR ) w~s 1.15 for
n = 12, 1.36 for n = 8, and 1.51 for n = 6. 'ftese lower efficie ncies of } ' I, were not
due to greater biases in the regression estimates, the corresponding variance ratios
being almost the same.

i.6 ACCURACY OF THE LARGE·SAMPLE FORMlTLAS FOR

V(JtI,) AND v(Jt,,)

No general ana lytical results are available on the accuracy of the apflrDximate
formulas Ci .25) for V(YI, ) and (7 .29) fer V(YI,) in mode rate or small samples. The
approximate estimator in (7 .25) and (7.29) are

- = -(1 - -{J Sv 2( 1 - p)
V( YI,) 2 (7.25)
n

(7.29)

Suppose that the y, for i == 1, 2, .. . N are a random sample from an infinite

population under the model
y, == a +f3x, +e, (7.35)

where for fixed x" the e, are independently distributed with mea n 0, variance
2
CT, = q / (1 - p 2). With this model, Cochran (1942) gave the result that to terms of
2
order 1/n •

(7.36)

where G] = k 3x/ q/ is Fisher's measure of relative skewness of the distribution of

x. Since S/(l_p2) in (7.25) is an unbiased estimate of O}(1_ p 2) under this
model, (7.36) suggests that with x symmetrically distributed the percent under-
estimation by V(y/,) is 100/(n - 2) with this model.
From Monte Carlo studies on eight small natural populations, (Rao, 1968),
Table 7 .1 shows for n = 6, 8, 12, the a\ erage percent underestimation of the
variance of Y/, by the approximations V(YI,) and v(y/,).
198 SAMPLlNG TECHNIOUES

TABLE 7.1
A VERAGE PERCENT UNDERESTIMA nON OF THE VARIANCE OF 9"
n
Estimator 6 8 12

V(y,,) in (7 .25) 38 34 28
v(y,,) in (7 .29) ' 48 42 33

For x, symmetrical, formula (7.36) suggests percent underestimations by V(YI,)

of 25, 17, and 10% for n = 6,8, 12. The percents for V(YI,) in Table 7.1 are
substantially higher. by amounts judged unlikely to be accounted for by skewness
in x in these populations-more likely by deficiencies in the linear model. The
underestimations are still greater for the sample estimates f variance V(YI,)'
Furthermore, comparison with Table 6.2, page 164. which applies to the ratio
estimate in the same eight populations, shows that the percent underestimations
in Vand V for YI, are at least twice those for YR in samples of the same size.

7.7 BIAS OF THE LINEAR REGRESSION ESTIMATE

The estimator YI, has a bias of order 1/ n in simple random sampling. We have
(7.37)
" -
Thus one expres ion for the bias is - Eb(i - X) = - cov (b, i) . The leading term in
the bias turns out to be
-(1 - f) Eei(X, - X)2
2 (7.38)
n Sx
This term represents a contribution from the quadratic component of the regres-
sion of y on i . Thus, if a sample plot of Yi against x, appears approximately linear,
there should be little risk of major bias in fl,'
To show (7 .38) requires some algebraic development. By (7.23), page 194,
n
te,ex/- i)
b = B +---- (7.39)
f(x , -i)2

Replace f (x, - i)2 by its leading term, nS 2

x • Also write

(7.40)
 REGRESSION ESTIMATORS 199
Hence the leading term in the bias - Eb(i - X) of Y'r is the average of

(7.41)

Let UI = ej(xj - X). By (7.22) its population mean 0 = O. The average value of the
first term in (7.41) may therefore be written
- E(u - O)(i -;R) (1- f) E(u j - O)(XI- X) (7.42)
52 = - -n- S2
x x

by theorem (2.3) (p. 25) for the average value of a sample covariance in simple
random sampling. This in turn equals (7.38), namely
_ (1 - f) Ee;(x; - X)2
n S/
In the second term in (7.41), e is O(1/,J;,.) and (i-;R)2is O(l/n), so that this term
is of smaller order than (7.38). Thus (7.38) is the leading term in the bias of YI,'

7.8 TIlE LINEAR REGRESSION ESTIMATOR

UNDER A LINEAR REGRESSION MODEL
Suppose that the finite population value y, (i = 1, 2, ... N) are randomly
drawn from an infinite superpopulation in which
y = a +(3x +£ (7 .43)
2
where the £ are independent, with means 0 and variance 0-. for fixed x. By direct
substitution from the model we find that
n n
LYj(xt - i) L £j(x; -i)
b= = {3+--- - (7.44)
£ (x, - i)2 £
(XI - i)2
n
_ _ L £1(Xj -i)
YI,- Y =(En -EN) + ( X - i ) - - - - (7.45)
(Xt- i )2 £
where En and EN are means over the sample> and the finite population. It follows
from (7.45) that under this model, E(YI, - y) = 0, so that y" is model-unbiased for
any size of sample.
As regards variance, it follows from (7.45) that for a given set of x's,
- 2 , [( 1 1 ) (X - i)2 ] (7.46)
V(y,,)=E(YI,-Y) =0-, - - N +n .
n L (xl-i)2
 200 SAMPLING TECHNIQUE

This result holds for any n > 1 and any sample selected solely by the values of x.
This approach and its generalization to the case of unequal residual variances
were given by Royall (1970). Under this model a pu.rposive sampling plan that
succeeded in making i = X would minimize V()'I,) for given n.
Also. for any ample selected solely according to the values of the x,. the usual
lea. t square. estimator

~, ' = £
[()'I - n- b (x, - i)f /( 11 - 2) (7.47)

is a modd- unblased estimator of a, 2 (or 11 • 2.

Thu~. in problems in which this model applies, simple exact results ahout the
mean and variance of 91, can bc c\tahli~hed . valid for any sample SI7.e . ~ anti
requiring only sample selection according to the values of the x,. thr random
element oeing supplied gratis by the distrihution of the E'S assuPled in the model.

7.9 REGRESSION ESTIMATES IN STRATIFIED SAMPLING

A!. with the ratio cstimate, two types of regression estimate can be made in
stratified random sampling. In the first estimate 91" (s for separate). a separate
regression estimate is computed for each stratum mean, that is,
(7.4H)

Then, with WII = N II / N,

'f

91r,' = I W"y"" (7.49)

"
This estimate is appropriate when it is thought that the true regression coefficients
Eh vary from stratum to stratum.
The second regression estimate. Ylre (c for combined), is appropriate when the
Eh are presumed to he the same in all trata. To compute Ylre, wt:' first find

y" = L WIIYII i,,=I Whih

h h

Then
(7.50)
The two estimates will be considered first in the case in which the bll and bare
chosen in advallce, since their properties are unusually simple in this situation.
From section 7 2, fir" is an unbiased estimate of Yin so that }it,. is an unbiased
estimate of Y. ')ince sampling is independent in different strata, it follows from
theorem 7.1 tr \t

- ~ W,,2(1-fh) 2 b 2S 2) (7.51)
V((YI,..) = '" (Syh - 2b"Sy!IC/o + h x/o
h nil
 REGRESSION ESTIMATORS 201
Theorem 7.2 shows that V(Ylr., ) is minimized when bll =BI., the true regression
coefficient in stratum h. The minimum value of the variance may be written

_ ) ~ W/(l- Ih) ( 2 S~"h)

V"'h,(YI" =/... S,'h - S ~ (7.52)
h nh .h

Turning to the combined estimate with preassigned b, (7.50) shows that YIn: is
also an unbiased estimate of Y in this case. Since Ylr< is the usual e, timate from a
stratified sample for the variate Ylu + b(X - Xh,), we may apply theorem 5.3 to this
variate, giving the result
. - )_~ W h (l-lh)(s 2 ?
2
V(YI,.. - /... vh - _bSyxh + b 2 Sxh 2 ) (7.53)
h nh

The value of b that minimizes this variance is

2
Be = L W/( 1 - III )S,xl'/L W/(l- fh )Sxh (7.54)
I, nl, h nl,
The quantity Be i!. a weighted mean of the stratum regression coefficients
Bh = S"xhl S,/. If we write

then Be = 1: ahBh/!. all'

From (7.52) and (7.53), with Be in place of b, we find
\ImIn ()'/,..) - Von,n (,91,,) = L ahB/ - (L all) v/
= L ah(Bh - Be)2 (7.55)
This result show that with the optimum choices the separate estimate has a
smaller variance than the combined f"stimate unless Bh is the same in all strata.
These optimum choices would, of course, require advance knowledge of the Svrh
and Sxh 2 values.

7.10 REGRESSION COEFFICENTS ESTIMATED

FROM THE SAMPLE

The preceding analysis is helpful in indicating the type of sample estimates bit
and b that may be efficient when used in regression estimates. With the separate
estimate, the analysis suggests that we take
L (Yhl - 91, )(Xhl - Xh)
bh =i (7.56)
L (XIIi -XII)2
i

the within-stratum least ,squares' estimate of Bh •

 202 SAMPLING TECHNIQUES

Applying theorem 7.3 to each stratum, we have

V( -
Ylrs
) = i...
~W/( l-fh)S~h 2(1_ Ph 2) (7.57)
h nh

provided that the sample size nh is large in aiJ strata. To obtain a sample estimate
of variance, substitute

S;' Xh=
nh'
~2[L(Yh'-Yh)2-bh2 L(Xlli-Xh)2]
,
(7.58)

in place of Sy/(I-p/) in (7.57).

The estimate Ylrs suffers from the same difficdty as the corresponding ratio
estimate, in that the ratio of the bias to the standard error may become appreci-
able. It follows from section 7.7 that the regression estimates Ylrll in the individual
strata may have biases of order 1/ nh and the biases may be of the same sign in all
strata, so that the over-all bias in Yin may also be of order 1/ nh. Since the leading
term in the bias comes from the quadratic regression of Yh, on XIIi, as shown in
section 7.7, this danger is most acute when the relation between the variates
approximates the quadratic rather than the linear type.
With the combined estimate, we saw that the variance is minimized when
b = Be, as defined in (7.54). This suggests that we take
W h2(1-fh) _ _ / W/(l - fh) _ 2
be = L ( 1) L (Yh, - Yh)(Xhi - Xh) L ( 1) L (Xlti - Xh)
It nh nh - , h nil nh - I
"
as a sample estimate of Be. If the stratification is proportional and if we may
replace the (nh -1) in be by nit, be reduces to the familiar pooled least squares
estimate

In certain circumstances other estimates may be preierable to be or be'. For

instance, if the true regression coefficients Bh are the same in all strata but the
residual variances about the regression line differ substantially from one stratum
to another, a different weighted mean of the bl" weighting inversely as the
estimated variance, may be more precise. However, the gain in precision as it
affects Ylrc is likely to be small.
Since
Ylrc - Y = y,,- Y +bc(X -ill)
=[y" - Y + Bc(X - x,,)]+(bc - Bc)CX - ill) (7 .59)
it follows that if sampling errors of be are negligible
- ) ~ Wh2(1- fh)(S 2 B' co B 2S 2)
V (Ylre =i...
h
yll - 2 ~yJCh + c
nil
xII (7.60)
 REGRESSION ESTIMATORS 203
As an estimate of V(Ylrr), we may take

V (Ylrr
'" W/(l-
- ) = £., (
fh) -) - bc (Xhc - Xh
'" [(Yhi - Yh
1) £., - )]2 (7.61)
h nh nh - ;

7.11 COMPARISON OF THE lWO lYPES OF

REGRESSION ESTIMATE
Hard and fast rules cannot be given to decide whether the separate or the
combined estimate is better in any specific situation. The defects of the separate
estimate are that it is more liable to bias when samples are small within the
individual strata and that its variance has a larger contribution from sampling
errors in the regression coefficients. The defect of the combined estimate is that its
variance is inflated if the population regression coefficients differ from stratum to
stratum.
If we are confident that the -l"egressions are linear and if Bh appears to be
roughly the same in all strata, the combined estimate is to be preferred. If the
regressions appear linear (so that the danger of bias seems small) but Bh seems to
vary markedly from stratum to stratum, the separate estimate is advisable. If there
is some curvilinearity in the regressions when a linear regression estimate is used,
the combined estimate is probably safer unless the samples are large in all strata.
Estimlstors of the regression type that are unbiased have been developed by
Mickey (1959) and Williams (1963), but have not yet been extensively tried. Rao
(1969) found Mickey's estimator usually inferior to the standard regression and
ratio estimators in natural populations. A jackknifed version can also be con-
structed. With n == mg, let YClr)j be the standard regression estimate, computed
from the ample with group j omitted. (j = 1.2, . .. , g). Then the jackknifed form
is
(7.62)

EXERCISES
7.1 An experienced fanner makes an eye estimate of the weight of peaches XI on each
tree in an orchard of N = 200 trees. He finds a total weight of X = 11,600 lb. The peaches
are picked and weighed on a simple random sample oC 10 trees, with the following results.

Tree Number

2 ) 4 5 6 7 8 9 10 Total

Actual wt. YI 61 42 50 58 67 45 39 57 71 5) 54)

Est. wt. x, 59 47 52 60 67 48 44 58 76 58 569
 204 SAMPLING TECHNIQUES

As an estimate of the total actual weight y, we take

Y =N[X +(9 -xl)
Compute the estimate and find its standard error.
7.:! Does it appear that the linear regression estimate. with the sample least . quares b,
would give a more precise estimate in 7.1?
7.3 From the sample data in Table 6.1 (p. I :;2) compute the regression estimate of the
1930 total number of inhabitants in the 196 large cities. Find the approximate standard
erroc of this estimate and compare its precision with that of the catio estimate.
7.4 In exercise 7.3 find the estimated total number of inhabitants and its standard eFfor
if b is taken as 1.
7.S In the following population with N = 5, verify (a) that the regression of }' on :x is
linear and (b) that the linear regression estimate is unhiased in simple random sa mples with
n = 3. The h. x) pairs are (3 ,0). (5 , 0). (8.2), (8 . 3). (12.3).
7.6 A rough measurement x, made on each unit. is related to the true measurement I'
un the unit by the equation
x = },+e+d·
where d is a constant bias and e is an error of measurement, uncorrelated with y, which has
mean zero and variance S/ in the population, as~umed infinite. In simple random ~ample.'i
of size n compare the variances of (a) the " difference " estimate [9 + (X - .r )) of the mean )/
and (b ) the linear regression estimate. using the value of b thaI gives minimum variance .
tThe variances may depend on S/ .)
7.7 By working out all po~sible cases, compare the MSE's of the separate and
combined regreSSion estimates of the total Y of the following population. when simple
random samples of size 2 are drawn from each stratum. For each estimate, how much doe~
its bilis co ntribute to the MSE?

Stratum I Stratum 2
.1"11 '!I 1 , x2 , ,1/2,

4 0 5 7
6 3 6 12
7 5 8 13

Use the ordinary least squares estimates of the B's , b. and b, on pp . 201-2.
7.8 In the popula~ion of exercise 7.7. show that if the optimum preassigned B could be
used in each case, V(Y,,,) = 4.39, VCY,~) = 4 .43. both estima!es bei,l&, of cour e, unbiased .
7.9 By the same method , compare the MSE's of the separate and combined ratio
estimates in the population in exercise 7.7. Since the ratio Y/ X is 8/17 =0.47 in stratum 1
=
and 32/19 1.68 in stratum 2, large sample theory would suggest that Y to would be
superior to Ync' You will find, however, that in these, tiny saTples, Yn , has the smaller
MSE. Its superiority is not due to smaller bias, neither YR. nor Y nc being ~naterialll' biased.
As another disagreement with large-sample theory, yOll will find that Y R.t and Y Rc have
smaller MSE's than the corresponding regression estimates.
 HAPTER 8

Systematic Sampling

8.1 DESCRIPTION
Thi s method of sa mpling is at first sight quite different from simple random
sampling. Suppose th at the N units in the population are numbered I to N in some
orde r. To se lect a sa mpl e of n units, we,take a unit at random from the first k units
a nd e very kth unit the reafte r. For instance. if k is 15 and if the first unit drawn i ~
numbe r 13, the subsequent units are numbers 28, 43, 58, and so on. The se lecti on
of the first unit dete rmines the whole sample . This ~ " pe is called an every kth
systematic sample .
The apparent advantages of this method over ,im.lle randO' n :;dmpling are a
follows.

1. It is easie r ~o draw a sample and of' ,;!.1 easier to execute without mistai .:s.
Thjs is a particular advantage Whf" .l th r. Jrawing is done in the field . Even when
drawing is do n in an office there may be a substantial saving in time. For instance,
it the units are described on cards that re all of the same size and lie in a file
~ a card can ~ drawn out every inch along the file as mea ured by a ruler .
This ~ration is speedy, whereas simple random sampljng would be slow. Of
course. this method departs slightly from the strict " every kth" rule . ...
2. Intuitively, systematic sampling seems likely to be more precise than s~le
random sampling. In effect, it stratifies the population into n strata, which coftSist
of the first k unit , the second k units, and so on. We might therefore expect1f,e
systematic sample to be about as precise as the corresponding stratified random
sample with one unit per tratum . The differe.1ce is that with the systematic sample
the unit occur at the same relative position in the stratum, whereas with the
stratified random sample the position in the stratl1m i jeterr.lined separately by
randomization within each stratum (see Fig. 8.1). The systematic sample is spread
more evenly over the population, and this fact has sOlletimes made systematic
sampling considerably more precise than stratified random sampling.
One variant of the systematic sample is to choose each unit at o7;'ear the center
of the stratum; that is, in tead of star ing the seque Ice by a random number
chosen between 1 and k, we take the starting number as (k + ])/2 if k is odd and
205
 206 SAMPLING TECHNIQUES

SYSTEMATIC SAMPLING

x =systematIc sample 0 =stratifaed random sample

Ll(----o-Lxo----l-x~.~)(__o__b)(--j
It 2/t 3/t 4/t 5/t 6/t
UnIt number

Fig. 8.1 Systematic and stratified random sampling.

either k/2 or (k + 2)/2 if k is even (Madow, t 953). This procedure carries the idea
of systematic sampling to its logical conclusion. If Yi can be considered a
continuous function of a continuous variable i, there are grounds for expecting
that this centrally located sample will be more precise than one randomly located.
Limited investigation on some natural populations supports this opinion,
although centrally located samples tend to behave erratically. Attention here will
be confined to samples with some random element.
TABLE 8.1
THE POSSIBLE SYSTEMAllC SAMPLES FOR N = 23, k = 5
Systematic sample number

II III IV V

2 3 4 5
6 7 8 9 10
., II 12 13 14 IS
16 17 18 19 20
21 22 23

Since N is not in general an integral multiple of k, different systematic samples

from the same finite f')pulation may vary by one unit in size. Thus, with N = 23,
k = 5, the numbers of the units in the five systematic samples are shown in Table
8.1. The first three samples have n = 5 and the last two have" = 4. Thi fact
introduces a disturbance into the theory of systematic sampling. The disturbance
is probably negligible if n exceeds 50 and will be ignored, for simplicity, in the
presentation of theory. It is unlikely to be large even when n is small.
Another method, suggested by Laltiri in 1952 (see Murthy, 1967, p. 139)
provides both a constant sample size and an unbiased sample mean. Regard the N
units as arranged round a circle and let k now be the integer nearest to N/ n. Select
a random number between 1 and N and take every kth unit thereafter, going
round the circle until the desired n units have been chosen. Suppose we want n = 5
with N = 23. Then k = 5.H the random number is 19 we take units 19, 1,6,11,16.
It is easily verified that every unit has an equal probability of election with this
method. If n = 4 units are wanted with N = 23 we take k = 6.
 SYSTEMATIC SAMPLING 207
8.2 RELATION TO CLUSTER SAMPLING
There is another way of looking at systematic sampling. With N = nk, the k
po sible systematic samples are shown in the columns of Table 8.2. It is evident
from this table that the population has been divided into k large sampling units,
each of which contains n of the original units. The operation of choosing a
randomly located systematic sample is just the operation of choosing one of these
large sampling units at random . Thus systematic sampling amounts to the selec-
tion of a single complex sampling unit that constitutes the whole sample . A
systematic sample is a simple random sample of one cluster unit from a population
of k cluster units.

TABLE 8.2
COMPOSITION OF THE k SYSTEMATIC SAMPLES

Sample number
2 k

Yl !h !I, !t.
Yk 01 .114. ~ ilk. , !/Z •

Y(" _ l) k. ) .'1" , 11 .4 • ~ "/(11 1 11. - I ?I,d.

Means Yl ii2 .17 t ih

8.3 VARIANCE OF THE ESTIMATED MEAN

Several formulas have been developed for the variance of y,~, the mean of a
ystematic ample . The three given below apply to any kind of clu ter sampling in
which the clusters contain n elements and the sample consi. ts of one cluster. In
these formulas we a sume N = nk.
If N = nk, it is ea y to verify that y,y is an unbiased estimate of Y for a randomly
located systematic sample.
In the following analysi the symbol Y.} denotes th jth member of the ith
systematic sample, a that j = 1, 2, ... , fl , i = 1,2, . .. , k . The :nean of the ith
sample is denoted by y,.
Theorem 8.1. The variance of the mean of a ystematic sample is

V( - )=N - I S1_ k (n - 1)S2

Y'Y N N ""Y
(8.1)

where
20 SAMPLING TECHNIQUES

is the variance among units that lie within the same systematic sample. The
denominator of this variance, k (n - 1), is constructed by the usual rules in the
:lOalysis of variance: each of the k samples contributes (n -1) degrees of freedom
to the sum of squares in the numerator .
Proof. By the usual identity of the analysis of variance

, J

= n L (Yi - yf + L L (Yij - y,f

I , j

But the variance of )i,>" is by definition

_ )k __ ?
V(ySY) = -k I (y, - Y)
,=1
Hence
(N - 1)5~ = nk V(YsI' ) + k (n - I )5~,sy (8.2)
The result follows.
Corollary. The mean of a systematic sample is more precise than the mean of a
simple random sample if and on ly if
5:"'l'> 5 2
ProJ[. If )i is the mean of a simple random sample of size n,

"
N - n 52
V(y) = - - -
N n
From (8 . 1) , V(Y.fY) < V(y) if and only if
1\f - 1 2 k(n-I) , N - n 52
--5 - S- < - - l....- (8.3)
NNw" Nn
that is, if

k(n - l)5 2wsy > ( N - l -N-n)

-- 5 2 = k(n-l)5 2 (8.4)
n
This important result, which applies to cluster sampling in general, states that
systematic sampling is more precise than simple random sampling jf the variance
within the systematic samples is larger than the population variance as a whole .
Systematic sampling is precise when units withjn the same sample are heterogene-
ous and is imprecise when they are homogeneous . The result is obvious intuitively.
If there is little variation within a systematic sample relative to that in the
population, the succe sive units in the sample are repeating more or less the same
information.
Another form for the variance i given in theorem B.2.
 SYSTEMATIC SA MPLING 209
Theorem 8.2.

(8.5)

where P... is the correlation coefficient between pairs of units that are in the same
systematic sample. It is defined a
_ E(y,/ - n(y,u - n
(8.6)
P... - E(Y'j - 9)2
where the numerator is averaged over all kn(n - 1)/2 distinct pairs, and the
denominator over all N values of Yor Since the ?enominator is (N - 1)5 2 / N, this
gives
2 I __

P... = (n - 1)( N _} )5~ ,LJ j~" (y,/ - Y)(y,.. - Y) (8 .7)

Proof.
k
n 2 kV(Yry) = n 1 L: (y, - Y)2
1= I

k
= L [(y, , - Y) +(y, ~ - f" )+ "' + (Ytn-Yll
;_ 1

The squared terms amount to the total sum of . quares of deviations from Y, that
is, to (N - 1)5 2 . This gives
n 2 kV(Ysy) = (N - l)5 2 +2L:
t
L
j<u
(y,/- Y(Y,u- n (8.8)

(8.9)
Hence
2
5 (N
V(Y,y) ;: -; - l) [1 + (n - l)p ... ]
-;::; (B .10)

This result shows that positive correlation between units in the arne sample
inflates the variance of th ample mean. Even a small positive correlation may
have a large effect, becau e of the multiplier (n -1).
The two preceding theorems express V(YJY) in terms of 52, hence relate it to the
variance for a simple random sample. There j an analogue of theorem 8.2 that
expresse V(Y.y) in term of the variance for a stratified random sample in which
the strata are composed of the first k units, the second k units, and so on. In our
notation the subscriptj in Ylj denotes the stratum . The tratum mean is written f j.
Theorem 8.3.

V (Y. y) 5~,,(N
=--;;- ~- ")l 1 +\n - l )p.....,] (~ . ll )
210 SAMPLING TECHNIOUI!.S

where

(8.12)

This is the variance among units that lie in the same stratum. The divi or n(k - 1)
is used because each of the It strata contributes (k - 1) dewees bf lreedom.
Furthermore,
_ E(YIl- Y.,)(Yiu - 9.u) (8.13)
PWII - E(YI/ _ 9./)2
This quantity is the correlation between the deviation from the stratum means of
pairs of items hat are in the same systematic sample.
- 2 fL (Ylj -
PW'I - n(n -l)(k - l); _ lj<u
9 )(YI~ - 9..J
S!"
(8. 14)

The proof is similar to that of theorem 8.2.

Coronary. A systematic sample has the same precisioh a the corre-
sponding stratified random sample, with one unit pet stratum, if PWSl = O. This
follows beca use for thjs type of stratified random sample V(Ysr ) is (theorem 5.3,
corollary 3)

V( - ) =
y"
(NN- n) S!$/
,. (8. 15)
'r

Other formulas (or V(.y,y), appropriate to an autocorrelated population, have

been given by W. G . and L. H. Madow (1944). who made the first theoretical study
of the precision of systematic sampling.
Example. The data in Table 8.3 are for a small artificial population that exhibits a fairly
steady rising trend. We have N = 40, k = 10, It = 4. Each column represents a systematic
sample, and the rows are the strata. The example illustrates the situation in which the
" within-stratum" correlation is positive. For instance, in the first sample each of the four
numbers 0, 6, 18, and 26 lies below the mean of the stratum to which it belongs. Thi is
consistently true, with a few exceptions, in the first five systematic samples. In the last five
samples, deviations from the strata means are mostly positive. Thus the cross-product
terms in p.." are predominantly positive. From theorem 8.3 we expect systematic sampling
to be less precise than stratified random sampling with one unit per stratum.
The variance V (y.,.) is found directly from the systematic sample totals as
_ _ _ 1 k _ _ 2_ I k _ - 2
V (y.,) - VI, - -k L (Yr. - Y) - 2""k L ("Yr. - " Y)
1_ 1 n 1-1

=_1_ [(50)2 + (58)2 + ... + (8W _ (727)2] = II 63

160 10 .
For random and stratified random sampling, we need an analysis of variance of the
population into "between rows" and " within rows." This is presented in Table 8.4. hence
 SYSTEMATIC SAMPLING 211
TABLE 8.3
DATA FOR 10 SYSTEMATIC SAMPLES WITH n - 4, N = kn = 40

Systematic sample nllmbers Strata

Strata 2 3 4 5 6 7 8 9 10 means

0 I I 2 5 4 7 7 8 6 4. 1
II 6 II 9 10 13 12 15 16 16 17 12.2
III 18 19 20 20 24 23 25 28 29 27 23 .3
IV 26 30 )1 31 33 32 35 37 38 38 33.1

Totals 50 58 61 63 75 71 82 88 91 88 72.7

TABLE 8.4
ANALYSIS OF VARIANCE

df ss ms

Between rows (strata) 3 4828 .3

Wilhin strata 36 485 .5 13.49 = S!"
Totals 39 5313 .8 136.25 .., SJ

the variances of the estimated means from simple random and stratified random samples
are as follows.
N-
V... = ( N
")S2 9 136.25
-;=10' - 4 - = 30.66

V.. =(NN-")S!" .. =.2..10 . 13.49

4
=3.04

Both stratified random sampling and systematic sampling are much more
eff ctive than simple random ampling but. as anticipated. systematic sampling is
less precise than stratjfied random sampling.
Table 8.5 shows the same data. with the order of the observations reversed in
the second and fourth strata. This has the effect of making pwJt negative, because it
makes the majority of the cross products between deviations from the strata
means negative for pairs of observations that lie in the same systematic sample. In
the first systematic sample, for instance, the deviations from the strata means are
now - 4.1, +4.8, -5.3, +4.9. Of the six products of pairs of deviations, four are
negative. Roughly the same situation applies in every systematic sample.
This change does not affect V",n and v.,.
With systematic sampling, it brings
about a dramatic increase in precision, as i seen when the systematic sample totals
in Table 8.5 are compared with th se in Table 8.3. We now have

(73) +(74) + .. '+(65) ----w- = 0.46

1 [ 2 2 (727)2]
V.,. = 160 2
 212 SAMPUNO lCCHNIQUES

TABLE 8.S
DATA IN TABU 8.3, WITH THE ORDER REVERSED IN STRATA II AND IV

Systematic sample numbers Strata

Strata 2 3 4 5 6 7 8 9 10 means

0 I I 2 5 4 7 7 8 6 4.1
II 17 16 16 15 12 13 10 9 8 6 12.2
III
IV !l 19 20 20 24 23 25 28 29 27
38 37 35 32 33 31 31 30 26
23.3
33.1

Totals 73 74 74 72 73 73 73 75 75 65 72.7

It is sometimes possible to exploit [his result by numbering the units to create

negative correlations within strata. Accurate knowledge of the trends within the
population is required. However, as will be'seen later, the situation in Table 8.S is
one in which it is difficult to obtain from the sample a good estimate of the
standard error of f. y •

8.4 COMPARISON OF SVSTEMAnc wrm STRATIFIED

RANDOM SAMPLING
The performance of systematic sampling in relation to that of stratified or
simple random sampling is greatly dependent on the properties of the population.
There are populations for which systematic ~ampling is extremely precise and
others for which it it is less precise than simple random sampling. For some
populations and some values of n, V(Y•.,) may even increase when a larger sample
is taken-a startling departure from good behavior. Thus it is difficult to give
general advice about the situations in which systematic sampling is to be recom-
mended. A knowledge of the structure of the population is necessary for its most
effective use.
Two lines of research on this problem have been followed . One is to compare
the different types of sampling on artificial populations in which YI is some simple
function of i. The other is to make the comparisons for natural populations. Some
of the principal results are presented in the succeeding sections.

8.5 POPULA nONS IN "RANDOM" ORDER

Systematic sampling is sometimes used, for its convenience, in populations in
which the numbering of the units IS effectively random. This is so in sampling from
a file arranged alphabetically by surnames, if the item that is being measured has
no relation to the surname of the individual. There will then be no trend or
 SYSTEMATI SAMPLING 213
stratification in YI as we proceed along the fil e and no correlation between
neighboring values .
In this situation we would expect systematic sampling to be essentially equival-
ent to simple random sampling and to have the same variance . For any single finite
population, with given values of nand k, this is not exactly true, because \/'Y'
which is based on only k degrees of freedom. is rather errati c when k is small and
may turn out to be either greater or smaller than V",", There (Ire two results which
show that on the average the twO variances are equal .

Theorem 8.4. Consider all N! finite populations that are formed by the N!
permutations of any se t of numoer Yl> Y2, ... ,YN' Then, on the:: average over
these finite populations.

(8.16)

Note that V'O" is the same for all permutations

This result. proved by W. G. and L. H. Madow (1944), show that if the order of
the items in a specific finite population can be regarded a drawn at random from
the N! permutations, systematic sampli ng is on the average equivale nt to simple
random sampling.
The second approach is to regard the finite population as drawn at random from
an infinite superpopulation which ha certain properties. The result that is proved
does not apply to any si ngle finite population (i .e .• to any specific et of values
Yh Y2• ... , YN) but to the average of all finite populations that can be drawn from
the infinite popUlation .
The symbol ~ denotes averages over all finite populations that can be drawn
from this superpopulation.

Theorem 8.S. )f the variates y, (i =" 1, 2 •...• N) are drawn at random from a
superpopulation in which

g'(y, - /-lY = a/
Then

The crucial conditio ns are that all YI have the same mean J,L, that i • there is no
tre nd, and that no linear correlation exi ts between the values y, and y, at two
different points. The variance u/ may change from point to point in the series.
Proof. For any specific finite population.
214 SAMPLING TECHNIQUES

Now

Since YI and Yj are uncorrelated (i ;t j) ,

- 2 1 ~ 2
~(Y - JL) = N 2 1:-IU1 (8.17)

Hence

(B .18)

This gives

(8.19)

Turning to V. y , let y" denote the mean of the uth systematic sample, For any
specific finite population,
1 k _ - 2
V'Y =-
k L (y" -
,, - I
y) (B .20)
"
=.!.[k "i:- I(y,, - p.)2-k(Y - JL)2] (B .21)

By the theorem for the variance of the mean of an uncorrelated sample from an
infinite population ,

(8.22)

(8.23)

8.6 PO PULAnONS wrm LINEAR TREND

If the population consists solely of a linear trend, as illustrated in Fig. B.2, it is
fairly easy to guess the nature of the results. From Fig. 8.2, it looks as if V" and V"
(with one unit per stratum) will both be smaller than Vran , Furthermore, v., will be
larger than V,,, for if the systematic sample is too low in one stratum it is too low in
 SYSTEMATIC SAMPLrNG 215
y,

)( • systematic 5ampif
o - strat,f*' random u mple

~.__..,.1
~.~.
~x
"
Fla. 8.2 Systematic sampling In a population with linear trend.

all strata, whereas stratified random sampling gives an opportunity for within-
stratum errors to cancel.
To examine the effects mathematically, we may assume that YI "" i. We have
f i = N(N+ I) f j2= N(N + 1)(2N+ 1)
,_) 2' I-I 6
The population variance S2 is given by

S 2= N~ I(IY/-NY2)

=_l_[N(N .-1)(2N + 1) N(N+ })2] = N(N+ 1) (8.24)

N-l 6 4 12
Hence the variance of the mean of a simple random sample is
V =N-n.~~==n(k-l).N(N+l) (k-l)(N+l)
(8.25)
raft N n N 12n 12
To find the variance' within strata, S•.}, we need only replace N by k in (8.24).
This gives
V = N - n . S./==n(k- l).k(~ +1):;(k2 - 1)
(8.26)
" N n "k 1211 12n
For systematic sampling, the mean of the second. ample exceeds that of the first
by 1; the mean of the third exceeds that of the ccond by 1, and so on. Thus the
means j" may be replaced by the number I, 2•... ,k. Hence. by a further
application of (8.24),
 216 SAMPLING TECHNIQUES

This give
1 _ - z e- 1
Vsy ="k L (Yu - Y) =1:2 ( .27)

From the formulas (8.25), (8 .26), and (8.27) we deduce, as anticipated,

z
V =e- l < v =k -l<v =( k-l)(N+1) (8.28)
SI 12n - sy 12 - ran 12

Equality occurs only when n = 1. Thus, for removing the effect of a linear trend ,
suspected or unsuspected. the systematic sample is much more effective than the
simple random sample but les effective than the stratified random sample.

8.7 METHODS FOR POPULA110NS WfI1f LINEAR TRENDS

The performance of systematic sampling in the presence of a linear trend can be
improved in several ways . One is to use a centrally located sample. Another is to
change the estimate from an unweighted to a weighted mean in which all internal
members of the sample have weight unity (before division by n ) but different
weights are given to the first and last members. If the random number drawn
between 1 and k is i, these weights are
n(2i - k -l)
1±~:---~ (8.29)
2( n - I)k

die + sign being used for the first member, the - sign for the last. For any i, the two
weights obviously add to 2. The reader may verify that if the population consists of
a linear trend and N = nk the weighted sample mean gives the correct population
mean. The performance of these end corrections has been examined by Yates
(1948), to whom the) are due.
Bellhouse and Rao (1975) have extended the Yates correction to the case
N.e nk when the systematic sample is drawn by Lahiri's circular method (section
8.1), which guarantees constant n. As before, the weights different from 1 are
applied to the first and last sample numbers in the original serial order of the
population. For ~xamp le. if the starting random number in drawing the sample is
19 with N = 23, n = 5, units 19, 1, 6, 11, 16 constituting the sample, the first and
last members are Yl and Y,q . Two cases ari e .
ase 1. Small i for which i + (n - 1)k:s N, so that the n units are obtained
• without passing over YN' The weight for the first (+) and last (-) members are
1±n[2i+(n- l)k -(N+l)] ,
(8.30)
2(n -I)k
Case 2. i + (n -1)k > N. Let nz be the number of sample units obtained after
=
passing over YN' Thus, with i 19, nz = 4. The weights forthe first (+) and last (-)
 SYSTEMATIC SAMPLING 217
members are

1 ±2(:-k)[2i +(n - 1)k - (N+ 1)-2n2;] (8.31)

In both cases the internal sample members receive weight 1 in the sample total.
With N = 23, n = k = 5, i = 19, n2 = 4, the first and last weights are 1 ± (-7/18).
Hence Yt receives a weight 11/18, while Yt9 receives 25/18.
Two alternative methods attempt to change the method of sample selection so
that the sample mean is unaffected by a linear trend . With N = nk and n even, a
method suggested by Sethi (1965) divides the population into n/2 strata of size 2k,
choosing two units equidistant from the end of each stratum. With starting
random number i, the n/ 2 pairs of units are those numbered
[i+2jk, 2(j+l)k -i+l ], j=0 , 1, 2, ... !n-1 (8.32)
This selection removes the effect of a linear trend in any stratum of 2k units,
even if the linear slope varies from stratum to stratum. Murthy (1967) has called
the method balanced systematic sampling.
The modified method of Singh et al. (1968) chooses pairs of units equidistant
from the ends of the population. With n even, the n/2 c:quidistant pairs that start
with unit i (i = 1, 2, . .. , k ) are
[i +jk, (N - jk )-i+ 1], j = 0 , 1,2, .. . ! n -1 (8. 33)
With n odd in these methods, j goes up to !{n - 1)-1 in (8.32) and (8.33). The
baJal1ced method (8.32) adds the remaining sample member near the end at
[i + (n - 1)k]; the modified method near the middle at [i +!(n - 1)k). The effect of
a linear trend is not completely eliminated in y for n odd.
Comparisons of the performances of these two methods with Yates' corrections
and with ordinary systematic sampling have been made on superpopulation
models representing linear and parabolic trends, periodic and autocorrelated
variation (Bellhouse and Rao, 1975), and on a few small natural populations by
these authors and by Singh (personal communication). In general the three
methods (Yates, balanced, modified) performed similarly. being superior to
ordinary systematic sampling in the presence of a linear or parabolic trend.
The population in Table 8.3, p. 211, for example, is one on which these methods
hould perform very well. Ordinary systematic sampling gave V,)' = 11.63. Com-
parable variances for the other methods (n = 4, k = 10) are: Yates, 1.29; Sethi
(balanced), 0.46; Singh (modified), 0.34. The balanced method happens to be that
obtained in Table 8.5 by reversal of strata II and IV in Table 8.3.

8.8 POPULATIONS WITII PERIODIC VARIATION

If the population consists of a periodic trend, for example, a simple sine curve,
the effectivenes of the systematic sampl depends on the value of k. This may be
218 SAMPLING TECHNIQUES

seen pictorially in Fig. 8.3. In this representation the height of the curve is the
observation YI' The sample points A represent the case least favorable to the
systematic sample. This case holds whenever k is equal to the period of the sine
curve or is an integrai multiple of the period. Every observation within the
systematic sample is exactly the same, so that the sample is no more precise than a
single observation taken at random from the population .

Fil. 8.3 Periodic variation.

The most favorable case (sample B) occurs when k is an odd mUltiple of the
half-period. Every systematic sample has a mean exactly qual to the true
population mean, since successive deviations above and below the middle line
cancel. The sampling variance of the mean is therefore zero. Between these two
cases the sample has various degrees of effectiveness, depending on the relation
between k and the wavelength.
Populations that exhibit an exact sine curve are not likely to be encountered in
practice. Populations with a more or less definite periodic trend are, however, not
uncommon. Examples are the flow of road traffic past a point on a road over 24
hours of the day and store sales over seven days of the week. For estimatiQg an
aver~ge over a time period, a systematic sample daily at 4 p.m. or every Tuellday
would obviously be unwise . Instead, the strategy is to stagger the sample over the
periodic curve, for example, by seeing that every weekday is equally represented
in the case of store sales.
Some populations have a kind of periodic effect that is less obvious. A eril's of
weekly payrolls in a small sector of a factory may always list the workers in the
same order and may contain between 19 and 23 names every week , A systemlltic
sample of 1 in 20 names over a period of weeks might consist mainly of th~ reWeds
of one worker or of the records of two or three workers. Similarly, a systematic
sample of names from a city directory might contain too many heads of house-
holds, or too many children. If there is time to study the periodic structure, a
systematic sample can usually be designed to capitalize on it. Failing this, a simple
or stratified random sample is preferable when a periodic effect is suspected but
not well known .
In some natural populations quasiperiodic variation may be present that would
be difficult to anticipate. L. H. Madow (1946) found evidence pointing this way in
a bed of hardwood seedling stock in a rather small population (N = 420). Finney
(1950) discussed a similar phenomenon in timber volume per strip in the Dehra
Dun forest, although in a reexamination of the data Milne (1959) suggested that
 SYSTEMATIC SAMPLING 219
the apparent periodicity might have been produced by the process of measure-
ment. The effect of quasiperiodicity is that systematic sampling performs poorly at
some values of n and particularly well for others. Whether this effect occurs
frequently is not known. Matern (1960) cites examples in which natural forces
(e.g., tides) might produce a spatial periodic variation, but he is of the opinion that
no clear case has been found in forest surveys.

8.9 AUfOCORRELATED POPULATIONS

With many natural populations, there is reason to expect that two observations
y" Yl will be more nearly alike when i and j are close together in the series than
when they are distant. This happens whenever natural forces induce a slow change
as we proceed along the eries. In a mathematical model for this effect we may
suppose that Yi and Yl are positively correlated, the correlation between them
being a function solely of their distance apart, i - j , and diminishing as this distance
increases. Although this model is oversimplified, it may represent one of the
salient features of many natural populations.
In order to investigate whether this model does apply to a population, we can
calculate the set of correlations p" for pairs of items that are u units apart and plot
this correlation against u. This curve, or the function it represents, is called a
correlogram. Even if the model is valid, the correlogram will not be a smooth
function for any finite population because irregularities are introduced by the
finite nature of the population. In a comparison of systematic with stratified
random sampling for this model these irregularities make it difficult to derive
results for any single finite population. The comparison can be made over the
average of a whole series of finite populations, which are drawn at random from an
infinite superpopulation to which the model applies. This technique has already
been applied in theorem 8.5 and in sections 6.7,7.8.
Thus we assume that the observations y, (i = 1, 2, ... , N) are drawn from a
superpopulation in which
~(y, ) =#" ~ (y, - #,)2 =0- 2, ~(y, _ #' )(YI+" _ #' ) = p,,0-2 (8.34)
where
p" 2: Pv 2: 0, whenever u < v
The drawing of one set of y, from this superpopulation creates a single finite
population of size N.
The average variance for systematic sampling is denoted by
~v,)' = ~E(y~)'
-
- -
Y)
2

For this class of populations it is easy to show that stratified random sampling is
superior to simple random sampling, but no general result can be established
about systematic sampling. Within the cla.~ there are superpopulations in which
220 SAMPLING TECHNIQUES

systematic sampling is superior to stratified random sampling, but there are also
superpopulations in which systematic sampling is inferior to simple random
sampling for certain values of k.
A general theorem can be obtained if it is further assumed that the correlogram
is concave upwards.
Theorem 8.6. If, in additic-n to conditions (8.34), we have
5} = Pu +I +Pu - I-2pu;';:'0 (u = 2,3, ... , (kn -2)] (8 .35)
then
Z' VJy :s; Z'VJ,:s; g'Vron (8.36)
for any size of sample. Furthermore, unless 5.. 2 :: 0, U = 2, 3, ... , (kn - 2),
Z'VJY < ~VJ' (8.37)
A proof has been given by Cochran (1946).
A sketch 0; the argument for n = 2 illustrates the role played by the "concave
upwards " condition. In the syst~matic sample the members of the pair are always
k units aput. Hence
~V(YJY):: ~(O'2 + 0'2+ 2PkO'2) = i O' 2(1 + Pic) (8.38)
With the stratified sample, there are k possible positions for the unit drawn from
e
each stratum, making combinations of positions. The numbers of combinations
1, 2, ... (2k -1) units apart are as foUows .
"
Distance 2 . . . (k -1) k (k + 1) ... (2k -1) Total

Number 2 ... (k -1) k (k - 1) ...

Hence the average value of V(YJ,) , taken over the e combinations, may be
written

~V(YJ') = ~:z.2 C~: u(2 + Pu + P2k-u) + k(l + pdJ (8.39)

Similarly, Z'V(Ys~) may be expressed as

~V(YIY) ::~C~:U(2+2PIJ+k(1 +Pt)] (8.40)

hence
(8.41)

But jf
Pu +1 +Pu - I ;,;:.2pu (u = 2, 3, ... )
 SYSTEMATIC SAMPLING 221
it is easy to show that every term inside tbe brackets is positive. This completes
the proof. In short, the average distance apart is k for both the systematic
and the stratified sample, but because of the concavity the stratified sample
loses more in precision when the distance is less than k than it gains when the
distance exceeds k .
QuenouiUe (1949) has shown that the inequalities in theorem 8.6 remain valid
when two of the conditions are relaxed so that

'I,'(y, - ,.,.Y = a/ (8.42)

In this event each of the three average variances is increased by the same amount.
As far as practical applications are concerned, correlograms that are concave
upward have been proposed by several writers as models for specific natural
populatIons. The function Pu = tanh (u - 3/ 5 ) was suggested by Fisher and Macken-
zie (1922) for th correlation between the weekly rainfall at two weather st.ations
that are a distance u apart; the function Pu = e - Au by Osborne (1942) and Matern
(1 947) for forestry and land use surveys ; and the function Pu = (/- u)/I by Wold
(1938) for certain types of economic time series.

8.10 NATURAL POPULATIONS

'nvestigations have been made on a variety of natural populations. The data are
described in Table 8 .6. The first three studies were made from maps. In the first
study the finite population consists of 288 altitudes at successive distances of 0.1
mile in indulating country. In the next two the data are the fractions of the lengths
of lines drawn on a cover-type map that lie in a certain type of cover (e.g., grass).
These examp es mi ght be considered the closest to continuous variation in the
mathematical sense.
The next three studies are based on temperatures for 192 consecutive days: (a)
12 in. under the soil, (b) 4 in. under the soil. (c) in air. This trio represents a
gradation in the direction of greater influence of erratic day-to-day changes in the
weather compared with slow seasonal influences.
The remaining studies deal with plant or tree yield in sequences that lie along a
line. In the study on potatoes, which is typical of the group the finite population
consists of the total yields of 96 rows in a field . Since 00 exhaustive search of the
literature ha been made, further data may be available .
1'1 some of the studies v,y is compared with tile variance V,12 for a stratified
random sample with strata of size 2k and two units per stratum. Thi comparisu .
is of interest because an unbiased estimate of V,,2 can be obtained from the sample
data. This cannot be done for v,tl (with strata of si .e k and 1 unit peT stratum) or
for V,y. Other writers report comparisons of V,v with both /"1 and V.. .l . Th e
majority of the sources do not present comparisons with V"", in readily u ·able
form, but it ap~ars that in general V,,2 gave gain in preci<:ioll over V N',
222 SAMPLING TECHNIQUES

TABLE 8.6
NATURAL POPULATIONS USED IN STUDIES OF SYSTEMATIC SAMPLINa
Refere"nce N Type of Data

Yates (1948), 288 Altitudes read at intervals of 0.1 mile from

table 13 ordnance survey map.
Osborne (1942) • Per cent of area in (0) cultivated land, (b) shrub,
(c ) grass, (cI) woodland on parallel lines drawn
on a cover-type map.
Osborne (1942) • Per cent of area in Douglas fir on parallel lines
drawn on a cover-type map.
Yates (1948) 192 Soil temperature (12 in. under grass) for 192 con-
secutive days.
Yates ( 1948) J92 Soil temperature (4 in . under bare soil) for 192 days.
Yates (1948) 192 Air temperature for 192 days.
Vates (1948) 96 Yields of 96 rows of potatoes.
Finney (1948) 160 Volume of saleable timber per strip, 3 chains
wide and of varying length (MI. Stuart forest).
Finne) (1948) 288 Volume of virgin timber per strip, 2.~ chains wide,
80 cha.i ns long (Black's Mountain forest).
Finney (1950) 292 Volume of timber per strip, 2 chains wide and of
varying length (Dehra Ot\n forest).
Joh ns('n (1943) 400t Number of seedlings per J-ft-bcd-width in 4 bed~
of hardwood seedbc!d stock.
'. Joh n..;on (1943) 400t Number of seedlings per I-ft-bed-width in 3 beds
of coniferous seedbed stock.
}ohn!.On (1943) 400t Number of seedlings per J-ft-bed-width in 6 beds
of r.oniferous transplant stock.

• Theoretically, N is infinite, if lines that are infinitely thin can be envisaged.

t A.pproximately . The number varied from bed to bed.

In the pap..:rs by Yates and Finney comparison are given for a range of values of
nand k within each finite population. III these cases the data in Table 8.7 are the
geometric means of the variance ratios for the individual values of k. The other
writers make computations for only one value of k per population but may give
datu for different items or for several populations of the same natural type. Here,
again, geometric means of the variance ratios were taken.
Although the data are limited in extent, the results are impressive. In the studies
'that permit comparison with \1,'1 systematic sampling shows a consistent gain in
precision which, although modest, is worth having. The median of the ratios
\1,11/ V,)' is 1.4. The gains in comparison with \1,'2 are substantial. the median ratio
being 1.9.
 SYS'reMA TIC SAMPLING 223
TABLE 8.7
RELATIVE PRECISION OF SYSTEMATIC AND STRATIFIED RANDoM SAMPLlNG
Relative Predsion of
Systematic to Stratified
Range
Data of k V,tI / v. y ~M2/V"

Altitudes 2-20 2.99 5.68

Per cent area (4 cover types) 4.42
Per cent area (Douglas fir) 1.83

Soil temperature (12 in.) 2-24 2.42 4.23

Soil temperature (4 in.) 4-24 1.45 2.07
Air temperature 4-24 1.26 1.65

Potatoes 3-16 1.37 1.90

Timber volume (Mt. Stuart) 2-32 1.07 1.35
Timber volume (Black 's Mt.) 2- 24 1.19 1.44
Timber volume (Dehra Dun) 2-32 1.39 1.89
Hardwood seedling 14 1.89
Coniferous seedlings 14-24 2.22
Coniferous transplant 12-22 0.93

The internal trend of the results agrees with expectations, although not too
much should be made of this in view of the small number of studies. The gains are
largest for the types of data in which we would guess that variation would be
nearest to continuous. The decline in Vllt! VI)' from soil to air temperatures would
also be anticipated from this viewpoint. In the last three items (forest nursery
data) , the only one showing no gain is coniferous transplant stock, which is older
and more uniform than seedling stock.

8.11 ESTIMATION OF 1HE VARIANCE FROM A

SINGLE SAMPLE
From the results of a simple random sample with n > 1, we can calculate an
unbiased estimate of the variance of tbe sample mean, the estimate being
unbiased whatever the form of tM population. Since a systematic sample can be
regarded as a simple random sample with n = 1, this useful property does not hold
for the systematic ample. As an illu tration, con ider the "sine curve" example.
Let
. 'IT;
YI = m+aslO-Z
224 SAMPLING TECHNIQUES

where k = 4 and i = 1, 2, . . . , 4n. The successive ob ervations in the population

are
(m +a ), m, (m - a ), m, (m +a ), m, (m - a ), m, ...

If i = 1 is chosen as the first member, all member of the systematic sample have
the value (m +a). For the other three pos ible choice of i, all member have the
values m, (m - a) , or m, respectively. Thus from a single sample we have no mean
of estimating the value of a. But the true sampling variance of the mean of the
2
ystematic sample is a / 2. The illustration show that it is impo sible to con truct
an estimated variance that is unbiased if periodic var iation i present.
These results do not mean that nothing can be done . Excluding the case of
periodic variation, we might know enough about the structure of the population to
be able to develop a mathematical model that adequately represe nts the type of
variation present. We might then be:: able to manufacture a formula for the
estimated variance that is approximately unbiased for this model, although it may
be badly biased for other models. The decision to use one of these models must
rest on the judgment of the sampler .
Some simple models with their corresponding e timated variances are pre-
sented below. No proofs are given .
The simplest mudels apply to populations in which y, is composed of a trend
plus a " random " component. Thus
Y; = IL; +e;
where ILl is some function of i. For the random componen t, we assume that there is
a superpopuJation in which

~(e, ) = 0, g'(e/) = 0/ , (i ¢. j)

A proposed formula s,/ for the estimated vari ance is called unbiased if
'lE (s,y 2) = ~ \I,y
that is, if it is unbiased over all finite populations that can be drawn from thf'
superpopulation.

PopuJation in "Random" Order

ILj = constant (i = 1, 2, . . . , N )
1 _ N - n L(YI- Y. y)2
S .y l - Nn n- 1 (R .43)

TItis case applies when we are confident that the order is essentially random with
respect to the items being measured . The variance formu la is the same as that {or a
simple random sample and is unbiaserl if the model is correct.
 SYSTEMATI SAMPLING 225
Stratification Effects Only
lLi constant (rk+1 S is rk+k)

(8.44)

In thi case the mean is constant within each stratum of k units. The e timate S ;y2'
which is based on the mean square successive difference, is not unbiased. It
contains an unwanted contribution from the difference between IL'S in neighbor-
ing strata, and the first and last strata carry too little weight in estimating the
random component of the variance. With a reasonably large sample, this estimate
would in general be too high , assuming that the model is correct.

Linear Trend
ILl :; IL + f3i
2 N - n n' r(Yi- 2Yi+k +Y, +2k?
(1 s i s n - 2) (8.45)
S'r 3 =N n2 6(n-2)
The estimate is based on successive quadratic terms in the equence Yi' The sum of
squares contains (n - 2) terms. With a linear trend we have seen (section 8.7) that
the trend can be eliminated by the use of end corrections. The term n'/n 2 is the
sum of squares of the weights in YW'Y- Unless n is small, n'/n 2 can be replaced by
the usual factor 1/ n. Because the strata at the ends receive too little weight, the
estimate is biased unless a} is constant, but it should be satisfactory if n is large
and lhe model is correct.
If continuous variation of a more complex type is present, the preceding
formulas may gi e poor results. In Table 8.8 the second and third formulas are
applied to six forest nursery beds (John on, 1943). The quadratic formula i
slightly better than that based on successive differences, but both give consistently
serious overestimate .
, TABLE B.B
VARIANCES OF SAMPLE MEAN NUMBERS OF SEEDLINGS (JOHNSO 's DATA)
Actual
0
Bed V SY S.; ' 2 S';'3

Silver maple I 0.91 2.8 2.5

2 0.74 3.6 2.9
American elm 4 .8 28 .4 12.6
2 15.5 22.6 18.6
White spruce I 5.5 17.2 11.2
2 2.0 11.6 6.4
White pine 8.2 21.0 21.9
226 SAMPLING TECHNIQUES

Formulas developed from simple assumptions about the nature of the

correlogram have been given by Osborne (1942), Cochran (1946), and Matern
(1947).lf the successive observations in the systematic sample are denoted by YI',
Y/, and so forth, Yates (1949) suggested an estimator based on differences d u of
which the first is
(8.46)
The next difference, d 2, may start with Y9', and soon. For the estimated variance of
YIY we take
v(Y.y)
- ---;;-
_N- n d/K
L - 57g (8.47)
I'In II - I .

The factor 7.5 is the sum of squares of the coefficients in any duo and g is the
number of differences that the sample provides (g == n/9) . In natural populations
that Yates examined. a formula of this type was superior to S ;y2 based on
successive differences but still overestimated V(Y. y)'
In summary, there is no dearth of formulas for the estimated variance. but all
appear to have limited applicability.
With N = nk, suppose that n is divisible by an integer m (say to). The following
method uses systematic sampling in part and provides an unbia"ed sample
estimate of V(Ys y ) based on (m -1) degrees of freedom . Draw a simple random
sample of size m from the units numbered 1 to mk. For every unit in this sample,
take also every (mk)th thereafter. In effect, this method divides the population
intb mk clusters each of size N/mk = n/m, and chooses m clusters at random , so
that we have a simple random sample of m clusters. For instance, suppose that we
want a 20% sample from a population with N :; 2400, so that n =- 480, k :; 5. Take
a simple random sample of size m = 10 from units numbered 1 to mk = 50, and
take every 50th unit thereafter. We then have 10 cluster samples each of size
2400/50 = 48.
Gautschi (1957) has examined. the accuracy of this method under the popula-
tion structures considered in this chapter. As might be anticipated, the accuracy
lies between those of simple random sampling and of sy tematic sampling with
m =l.

S.U STRATIFIED SYSTEMATIC SAMPLING

We have seen that if the units are ordered appropriately, systematic sampling
provides a kind of stratification with equal sampling fractions . If we stratify by
some other criterion, we may draw a separate systematic sample within each.
stratum with starting points independently determined. This is suitable if separate
estimates are wanted for each stratum or if unequal sampling fractions are to be
used. This method will, of course, be more precise than stratified random sampling
 SYSTEMATIC SAMPLING 227
if systematic sampling within strata is more precise than simple random sampling
within strata.
If 9,ylo. is the mean of the systematic sample in stratum h, the estimate of the
population mean Y and its variance are
2
9.ISY =I Wh 9,yh , V(Y.ISY) =I Wh V(Y.yh)
With only a few strata, the problem of finding a sample estimate of this quantity
amounts to that already discussed of finding a satisfactory sample estimate of
V(Ysyh) in each stratum .
When the strata are more numerous, an estimate based on the method of
" collapsed strata" (section SA.12) may be preferable. From the results in that
section, it follows that the estimate
v(Y,ISY) = I' Wh2(Y,yh - Y,yj)2 (8.48)
where the sum extend over the pairs of strata, is on the average an overestimate,
even if periodic variation is present within the strata. ...
An unbiased estimate of the error variance can·be obtained if two systematic
samples, with a different random start and an interval 2k, are drawn ~ithin each
stratum, one df being provided by each stratum. There will be some loss in
precision if ystematic sampling is effective. If there are many strata, one systema-
tic sample can be used in most of them, drawing two in a random subs ample ·of
strata for the purposes of estimating the error.

8.13 SYSTEMATIC SAMPLING IN 1WO DIMENSIONS

III sampling an area, the simplest extension of the one-dimensional systematic
sample is the "square grid" pattern shown in Fig. BAa. The sample is completely
determined by the choice of a pair of random numbers to fix the coordinates of the
upper left unit. The performance of the square grid has been studied both on
theoretical and natural populations. Matern (1960) has investigated the best type
of sample when the correlation between any two points in the area is a monotone
decreasing concave upward function of their distance apart d. For correlograms
like e - Ad the square grid does well, being superior to simple or stratified random
sampling with one unit per stratum, although Matern give reasons for expecting
that the best pattern for this situation is a triangular network in which the points lie
at the vertices of equilateral triangle .
In 14 agricultural uniformity trials, Haynes (1948) found that the square grid
had about the same precision as simple random sampling in two dimensions.
Mjlne (1959) examined the central square grid, in which the point lies at the center
of the square, in 50 uniformity trials. 1t performed better than simple random
sampling and perhaps ·slightly better than stratified random sampling, although
this difference was not statistically Significant. These results suggest that, at least
for data of this type, autocorrelation effects are weak. For estimating the are.a
 228 AMPLlNG T £CHNIOUES

covered by forest or by water on a map, Matern found the square grid superior to
the random method in two examples.
Figure 8.4b shows an alternative systematic sample, called an unaligned
sample. The coordinates of the upper left unit are selected first by a pair of random
numbers. Two additional random numbers determine the horizontal coordinates

x :
I )( : x
1
II II x
I I I I
I x
I
----'1'----..,-----
I. I
J
----r----,-----
)( I I
I

x I x I x I I )(
I I I I
1 I 1 x I
----~----+---- -
I J
-----r----4-----
I
l( I
I I
" I "
I
I x
I I I
)(

1 I I )( 1
(a) Aligned or " square gnd" (b) Unaligned sample
sample

Fig.8.4 Two types of Iwo-di mensional sys letnatic~ampl e.

of the remaining two units in the first column of strata. Another two are needed to
fix the vertical coordinates of the remaining units in the first row of strata. The
constant interval k (equal to the sides of the squares) then fil('es the locations of all
·points. lnvestigations by Quenouille (1949) and Das (1950) for si mple two-
dimensional correlograms indicate that the unaligned pattern will often be
superior both to the square grid and ~o stratified random sampling.
urther evidence of the superiority of an unali gned sample is obtained from
experience in exp~rimental design , in which the latin square has been found a
precise method for arranging treatment in a rectangular field . The 5 x 5 latin
square in Fig. 8.5a may be regarded as a division of the field into five systematic
. samples, one for each letter. There is some evidence that this particular quare.
which is called the " knight's move " latin square, i lightly more precise thtln 3
randomly chosen 5 x 5 square , probably b.~cause alignment is absent it1 the
diagonals as well as in rows and columns.
The principle of the latin square has been used by Homeyer and Black (1946) in
sampling rectangular fields of oats. Each field contained 21 plots. The three
possible systematic samples are denoted by the letters A, B, and C, respectively, in
Fig. 8.5b. This arrangement, with one of the letters chosen at random in each field .
gave an increase in precision of around 25% over stratified random sampling with
rows as strata. The arrangement does not quite satisfy the latin square propert)
because each letter appears three times in one column and twice in the other
columns, but it approaches this property a nearly as possible.
Yates (1960), who terms arrangements of this type lattice sampling, discusse
their use in two- and three-dimensional sampling. In three dimensions ead: roW.
 SYSTEMATI SAMPLING 229
A lJ C D E A B C
D E A B C B C A
B C D E A C A B'
E A B C D A B C
C D E A B B' C A
c: A ' B
A B C
(a), " Knight's move" lat in square (b) Systematic design for a 3 x 7 rec-
tangular field
Fi~. 8.S Two systemat ic designs based on the latin square.

column, and vertical level can be represented in the sample by choosingp units out
of the p 3 in the population . With p2 units in the sa Ie, each of the p2
combinations of levels of rows and columns, of rows and ical heights, and of
columns and vertical heights can be represented . Patter on (1954) has investi-
gated the arrangements that provide an unbiased estimate of error.

8.14 SUMM ARY

Systematic samples are convenient to draw and to execute. In most of the
studies reported in this chapter, both on artificial and'On natural populations, they
compared favorably in precision with stratified random samples. Their disadvan-
tages are that they may give poor precision when unsuspected periodicity is
pre ent and that no trustwo rthy me thod for estimating V('Y,fY) from the sample
data is known .
In the light of these results syste matic sampling can safely be recommended in
the following situations. '

1. Where the ordering of the population is essentially random or contains at

most a mild stratification . Here systematic sampling i used for convenience, with
little expectation of a gain in precision . Sample estimates of error that are
reasonably unbiased are available (section 8.1 )).
2. Where a stratification with numerous strata is emp10yed and an independent
systematic sample is drawn frolll each stratum. The effects of hidden periodicitie
tend to cancel out in this situation, and an estimate of error that is known to be an
overe timate can be obtained (section 8 .12). Alternatively, we can use half the
number of strata and draw two systematic samples, with independent random
starts, from each tratum . This method gives an unbiased estimate of error .
3, For sub ampling cluster units (Chapter 10). In this case an unbiased or
almost unbia ed estimate of the sampling error can be obtained in mo t practical
situations. This is a common use of systematic sampling.
4. For sampling populations with variation of a continuous type, provided that
an estimate of the sampling error is not regularly required . If a series of surveys of
8 O~~~N~~.N~~ ~OOOO~O~M
~

--
~o
~
-M - ----
•• o
N

~OO~~.M~O~.~~~O~~.OOOOO
f'""') MM---N-NN--N -

00
~NOOOMO~N~~-~~OO-~~~-~ ~
1 00 -- --N-- -NN----NN~N M
V

o ,..,
•I ~
~~~~-O~~.~O-~.M~~~~O
-~- --NN--- N_ t"") ,..,
o
N

00
N
~

8
I~
00

o~ V-~~N~~~~~~VVV~~V~~~
~
I . ~NNN~.~.NN~--N---MMN
~
~

0-
~

III

",'3
- 0
fI-
lii
230
 SYSTEMATIC SAMPLING 231
this type is being made, an occasional check on the sampling errors may be
sufficient. Yates (1948) has shown how this may be done by taking supplementary
observations.

EXERCISES
8. J The data in the table (p.230) are the numbers of seedlings for each foot of bed in a
bed 200 ft long.
Find the variance of the mean of a systematic sample consisting of every twentieth foot .
Compare this with the variances for (a ) a simple random sample, (b) a !>tratified random
sample with two units per stratum, ic) a stratified random sample with one unit per stratum.
All samples have n '" 1O. [L (y, - Y) 2", 23,601 .]
8.2 A population of 360 households (numbered 1 to 360) in Baltimore is arranged
alphabeticall y in a file by the surname of the head of the household. Households in which
the head is nonwhite occur at the following numbers: 28, 31 -3 3, 36-41 . 44, 45, 47 , 55 , 56,
58, 68 , 69,82,83, 85, 86, 89- 94,98, 99 , 101 , 107- 110, 114, 154, 156, 178, 223 ,224, 296,
298- 300,302- 304, 306- 323, 325-331 , 333, 335- 339, 341 , 342. (The nonwhite hou e-
holds show some " clumping" because of an association between surname and coloL)
Compare the precision of a 1-in-8 systematic sample with a simple random sample of the
same size for estimating the proportion of households in which the head is nonwhite.
8 .3 A ne ighborhood contains three compact communities, conSisting, respectively, of
people of Anglo-Saxon, Polish, and Italian descent. There is an up-to-date directory. In it
the persons in a house are listed in the following order: husband, wife, children (by age),
others. Houses are listed in order along streets. The average number of persons per house is
five .
The choice is between a systematic sample of every fifth person in the directory and a
20% simple random sample . For which of the following variables do you expect the
systematic sample to be more precise? (a ) Proportion of people of Poli h descent, (b)
proportion of males, (c ) proportion of children . Give reasons.
8 .4 In a directory of 13 houses on a street the persons are listed as follows: M = male
adult, F =female adult, m =male child, f = fem ale child .
Household
I 2 3 4 5 6 7 8 9 10 11 12 13
M M M M M M M M M M M M M
F F F F F F F F F F F F F
f
m
.rm m m
m
f
m
f m m m f
m
f
f f f f
f f f m

Compare the variances given by a systematic sample of one in five persons and a 20%
simple random sample for estimating (a) the proportion of males, (b) the proportion of
children, (c) the proportion of persons living in professional households (households 1,2,3,
12, and J3 are described as professional). Do the results support your answers to exercise
8 .3? For the systematic sample, number down each column, then go to the top of the next
column.
8.5 In exercise 8.1 we might estimate V(y,,) y (a) regarding each systematic sample as
a simple random sample, (b) pretending that each 1-in-20 systematic sample is composed
232 SAMPLING TECHNIQUES

of two l-in-40 systematic samples with a separate random start. For each method , compare
the average of the estimated variances with the actual variance of Y'Y'
8.6 In a population consisting of a linear trend (section 8.6) show that a systematic
sample is less precise than a stratified random sample with strata of size 2k and two units
per stratum if n > (4k + 2)/(k + 1).
8.7 A two-dimensional population with a linear trend may be represented by the
relation
YII = i+j (i,j = 1, 2, , . ..• nk)
where y" is the value in the ith row and jth column. The population contains N 1 = n 2k 1
units.
A systematic square grid sample is selected by drawing at random twO independent
starting coordinates io, jo, each between 1 and k . The sample. of size n 2 , contains all units
whose coordinates are of the form
io + yk, i o+5k
where y. 5 are any two integers between 0 and (n -1), inclusive.
Show that the mean of this sample has the same precision as the mean of a simple random
sample of size n 2 .
8.8 Jf the comparison in exercise 8.7 were made for a three-dimen ional population
with linear trend , what result would you expect?
8.9 In a population with y, = i 2 (i = 1. 2 . ... , 16), compare the values of E( Y- 9)1
given by every Jah systematic sampjing and by the Yates, Sethi, and Singh et al. methods for
n =4, k = 4, N = 16.
CHAPTER 9

Single-Stage Cluster Sampling:

Clusters of Equal Sizes

9.1 REASONS FOR CLUSTER SAMPLING

Several references have been made in preceding chapters to surveys in which
he sampling unit consists of a group or cluster of smaller units that we have called
elements or subunits. There are two main reasons for the widespread application
of cluster sampling. Although the first intention may be to use the elements as
'am pling units, it is found in many surveys that no reliable Ii t of the elements in
the population is avai lable and that it would be prohibitively expensive to
construct such a list . Jn many countries there are no complete and up-to-date lists
of the people, the houses, or the farms in any large geographic region. From maps
of the region , however, it can be divided into areal units such as block in the cities
and egments of land with readily identifiable boundaries in the rural parts. In the
United States these clusters are often cho en because they solve the problem of
constructing a list of sampling units .
Even when a Ii t of individual hou e. is available , economic con iderations may
point to the choice of a larger cluster unit. For a given size of ample, a small unit
usually gives more precise results than a large unit. For example, a simple random
'ample of 600 houses covers a town more evenly than 20 city blocks containing ar
average of 30 houses apiece . But greater field cost are incurred in locating 600
houses and in travel between them than in locating 20 blocks and visiting all the
houses in these blocks. When cost is balanced against precision. the larger unit
may prove superior.
A rational choice between two types or size of unit may be made by the familiar
principle of electing the unit that give the smalJer variance for a given cost or the
smalJer cost for a prescribed variance . As in many practical decisions, there may
be imponderable factors : one type of unit may have ome special convenience or
disadvantage that is difficult to include in a calculation of costs. In ampling a
growing crop, some experiences suggest that a smalJ unit may give biased
e timates because of uncertainty about the exact boundaries of the unit. Homeyer
and Black (1946) found that units 2 x 2 ft gave yields of oat about 8% higher than
23
234 SAMPLING TECHNIQUES

units 3 x 3 ft, possibly because samplers tend to place boundary plants inside the
unit when there is doubt . Sukhatme (1947) cities similar results for wheat and rice .
This Chapter deals with the case in which every cluster unit~contains the same
number of elements or subunits.

9.2 A SIMPLE RULE

When the problem is to compare a few specific sizes or type of unit, the
following result is hetpful. •
Theorem 9.1. This applies to simple random sampling in which the fpc is
negligible. The quantity to be estimated is the population total. For the uth type of
unit, let
M" = relative size of .unit
S" 2 = variance among the unit totals
C" =relative cost of measuring one unit
Then relative cost for specified variance or relative variance for specified cost
OCC"S}/M}.
Proof. Suppose V is the specified variance of the estimated population total.
For the uth type of unit the estimate is N,S" with variance
N 2S 2
n =-
" _"_ (9.1 )
" V
.,
2 2
The cost of taking n" units is C"n" = CuN" S" / V. Since N"M" = constant for
different types of unit, the cost i proportional to C"S" 2 / M" 2 . On the other hand, if
the cost C is specified, n" = C/C" and, in (9.1), VOCC"S}/M} . This completes
the proof.
CoroUary 1. If we define the relative net precision of a unit as inversely
proportional to the variance obtained for fixed cost, theorem 9.1 may be stated as
. . . M,,2
re Iatlve net preCIsIOn oc C S 2 (9.2)
""
.coronary 2. In the analysis of variance, the variances for units of dl~erent
sizes are often computed on what is called a common basis-usually that applic-
able to the smallest unit. To put the variances on a common basis, the variance S" 2
among the totals of units of size M" is divided by M". Let

S.. '2 = ~2 = variance among unit totals (on a common basis)

"
C.. 1 oc ~" = relative cost of taking a given bulk of sample
"
 SINGLE-STAGE CLUSTER SAMPLING 235
Then theorem 9.1 and corollary 1 may be stated as follows .
2
· cost f ' . oc CuS u
re Iatlve l preCISion
or equa M 2 ex:: C'S,2
u u (9.3)
u

I I . .. 1 (9.4)
re atIVe net precIsion oc Cu'Su ,2

If differences in the costs of taking the sample are ignored (i .e. , if Cu ' is
constant), the relative ne t precision with the uth unit oc 1/ Su'2. Kish's de/! factors
for the differen t units (section 4.11) are therefore proportional to the Su,2=
2
su / Mu.
Example. Johnson 's data (1941) for a bed of white pine seedlings provide a simple
example. The bed conta ined six rows, each 434 ft lo ng. There are many ways in which the
bed can be divided into sampling units. Data for four types of unit are shown in Table 9.1.
Sint;C the bed was completely counted , the data are correct population values .

TABLE 9.1
OAT ... FOR FOUR TYPES Of S... MPLtNG UNtT
Type of Unit

I-ft 2-ft I-ft 2-ft

Preliminary Data row row bed bed

M• ... relative size of unit I 2 6 12

N. = number of units in pop. 2604 1302 434 217
S .2 = pop. variance per un it 2.537 6.746 23 .094 68.558
Number of feet of row that can
be counted in IS min . 44 62 78 108

The units were

1 ft of a single row
2 ft of a single row
1 ft of the width of the bed
2 ft of the width of the bed
With the first two units it was assumed that sampling would be stratified by rows, so tbat
the S. 2 represent variances within rows. Simple random sampling was assumed for the last
two units.
Since the principal cost is that of locating and counting the units, costs were estimated by
a time study (last row of Table 9.1). With the larger units, a greater bulk of sample can be
counted in 15 min, less time being spent in moving from one unit to anotber.
The quantity to be estimated is the total number of seedlings in the bed. In the notation of
theorem 9.1., Table 9.1 gives the value of M. and S}. The relative values of C.' expressed as
the time required to count one unit, are as f lIows.
236 SAMPLING TECHNIQUES

I -ft 2-ft loft 2-ft

row row bed bed

c. (in IS-min times) ..

I 6
'iii

By theorem 9. 1, corollary 1, the relative net precisions are worked out in Table <.) .2.
The last line of Table 9.2 gives the relative precisions when that of the smallest unit is
taken as 100. The I-ft bed appears to be the best unit .

TABLE 9.2
RELATIVE NET PRECISIONS OF THE FOUR UNITS

I-ft row 2-ft row I-ft bed 2-ft bed

M.' 44 (4)(62) = 18.38 (36)(78) (144)(108)

- - = 17.34
Coo S.' 2.537 (2)(6.746) (6)(23 .094) = 20.27 ( 12)(6R.558) = 18.90

100 106 J 17 109

I

The variances among units, expressed on a common basis, are also worth looking at. The
values of S.'2= S. 2/ M. , applicable to a single foot of row, are, respectively, 2. 537, 3.373,
3.849.5.713 . Note that these variances increase steadily with increasing size of unit. This
result is commonly found (although exceptions may occur) . Since the relative net precision
ert/C.'S. '2, the cost of taking a given bulk of sample must decrea e with the larger units if
they are to prove economical.
'f

Theorem 9.1 and its corollaries remain valid for stratified sampling with
proportional allocation if all strata are of the same size and if S} , Su'2 represent
average variances within strata. This is so, under the conditions stated, because
the variance of the estimated population total, ignoring the fpc, is N 2 Su 2 / n, and
therefore assumes the same form as in simple random sampling. Theorem 9.1
does not hold for more complex types of sampling.
The preceding results are intended merely as an illustration of the general
procedure. Comparisons among units should always be made for the kind of
sampling that is to be used in practice or, if this has not been decided, for the kinds
that are under consideration. Changes in the method of sampling or of estimation
will alter the relative net precisions of the different units. Even with a fixed method
of sampling and estimation, relative net precisions vary with size of sample if the
cost is not a linear function of size or if the size is large enough so that the fpc must
be taken into account.
There is usually more than one item to consider. One approach is to fix the total
cost and work out the relative net precisions for each type of unit and each item.
Unless one type is uniformly superior, some compromise decision is made, giving
principal weight to the most important items.
 SINGLE-STAGE CLUSTER SAMPLIN G 237
TABLE 9.3
ESTIM ATED STANDARD ERRORS ( ~;.; ) fO R fOU R SIZES Of UNIT. W ITH SIMPL.E
RAN DOM SAMPLING
Best
Items S/4 S/2 S 2S Uni t

Number of swi ne 5.0 4.9 5.3 6.2 S(2

Number of horses 3.4 3.3 3.6 4.2 S/2
Number of sheep 17.4 15 .7 14.9 14.3 2S
Number of c hicken s 3.0 3.0 3.3 3.8 S /4. S/2
Number of eggs yesterday 5.7 5.2 4.9 4 .7 2S

Number of cattle 4.7 4.6 4.8 -S.5 S /2

Number of cows milked 3.7 3.6 3.8 4 .4 S/2
Number of gallons of milk 4.4 4.2 4.4 4.9 S/2
Dairy products receipts 5.5 5.2 5.4 6.0 S/2
Number of farm acres 2.9 2.8 3.0 3.5 S/2

Number of corn acres 3.7 3.5 3.8 4.4 S/2

Number of oat acres 4.6 4.8 5.6 7.0 S/4
Corn Yield 1.6 1.7 2.0 2.5 S/4
Oat Yield 1.6 1.5 1.6 1.8 S/2
Commercial feed el(penditures 12.6 13.6 16.7 2 1.8 S/4

Total cl(pendltures. operator 7.8 8. 1 9.6 12.0 S/4

Total receipts. operator 6.2 6 .5 7.7 9.8 S[4
Net cash income, opera tor 6.8 6.9 7.8 9.5 S/4

In view of the numerous factor that influence the results, a study of optimum
ize of unit in an extensive survey is a large ta k. A good example for farm
sampling is de cribed by Jes en (1942). An excerpt from his results is given in
Table 9.3. This compares four sizes of unit-a quarter-section, a half- ection , a
section, and a block consisting of two contiguous ections. The section is an area 1
mile square, containing on the average slightly under four farms. In this compari-
son the total field cost ($1000), the length of questionnaire (60 min to complete),
and the travel cost (5 cents per mile) are all specified, becau e relative net
precisions change if any of the e variables is altered. Costs are at a 1939 level.
The data in the table are the relative tandard errors (in per cent) of the
estimated means per iarm for 18 items. No unit is best for all items. The
half-section and the quarter-section are, however. superior to the larger units for
all except two items, with little to choose between the half- and quarter-sections.
The half-section would probably be preferred , because the problem of Identifying
the boundaries accurately is easier.
238 SAMPLING TECHNIOUES

9'.3 COMPARI ONS OF PRECISION MADE FROM

SURVEY DATA
In the nursery seedling example the variances for the different types of unit
were obtained from a complete count of the population . Except with mall
populations, however, it i seldom feasible to conduct a survey solely for the
purpose of comparison . Information about the optimum unit is more usually
procured as an ingenious by-product of a survey whose main purpose is to make
estimates .
Suppose that in a survey each unit can be divided into M smaller units. Instead
of recording only the totals for each "large" unit in the sample, we record data
separately for each of the M small units. A comparison can then be made of the
precision of the large and small units . A simple random sample of size n will be
assumed at first .
The analysis of variance in Table 9.4 can be computed from the sample .

TABLE 9.4
ANALYSI S OF VARIANCE OF THE SAMPLE DATA (ON A SMALL-UNIT BASIS)

df ms
Between large units (n - I)
Between small units within large n(M - I)

&tween small units in sample nM - 1

The estimated variance of a large unit (on a small -unit basis) is Sb 2 . It might be
thOUght that an appropriate estimate of the variance of a small unit would be the
mean square between all small units in the sample; that is,
2 (n-l)s/+n(M - l)s,}
s = nM - l (9.5)

This estimate, although often satisfactory, is Slightly biased because the sample is
not a simple random sample of small units, since these are sampled in contiguous
groups of M units.
An unbiased estimate is obtained from the sample by constructing an analysis of
variance, as in Table 9.5. for the whole population, which contains N large units
and NM small units.
By its definition, the population variance among small units is given by the last
tine of the table, that is.
2
S 2 = (N - 1)Sb2+N(M - 1)Sw
(9.6)
NM - l
 •
SINGLE-STAGE CLUSTER SAMPLING 239
TABLE 9.5 1
ANALYSIS OF VARIANCE fOR THE WHOLE POPULATION (ON A SMALL-UNIT BASI.)
df

Between large units N - I

Between small units within
large units N(M - I)

Between small units in the (N - I)Sb2 + N(M - 1)5,.,2

NM - I S2 = NM _ 1
population

With simple random sampling, s/ in Table 9.4 is an unbiased estimate of S/ (this

follows from section 2.3). It may be shown easily tha t Sw 2 is an unbiased estimate of
S,} . Hence an unbiased estimate of the variance S2 among all small units in the
population is
2
52 = eN - 1)s/ + N(M - J) sw
(9.7)
NM - l
Clearly. this expression is almost the same as the simpler expression
52== Sb 2 + eM - I )S~?
(9.8)
M
If 11 > 50, (9.5) for S 2 also reduces to (9 .8), so that S 2 is a satisfactory approxima--
tion to S 2 for n > 50.
The two estimates Sb 2 (for the large unit) and 52(for the small unit) are on a
common basis and may be substituted in theorem 9.1, corollary 2.
If the sa mple is large, the small units may be measured for a random subsample
of the large units (say 100 out of 600). Alternatively, two small unit , chosen at
random from each large unit, might be measured . More th an one size of sma1l unit
may be investigated simultaneously, provided that we take data that give an
unbiased estimate of Sw 2 for each small unit.
With stratified sampling, the variances for the large and small units can be
estimated by these methods separately in each stratum and then substituted in the
appropriate formula for the variance of the estimate from a stratified sample.
~ ~

Example. The data come from a farm sample taken in North Carolina in 1942 in order
to estimate (arm employment (Finkner, Morgan, and Monroe, 1943). The method of
drawing the sample was to locate points at random on the map and to choos as sampling
units the three farms that were nearest to each point. This method is not recommended
because a large farm has a greater chance o( inclusion in the sample than a small farm, and
an isolated farm has a greater chl\ncc than another in a densely farmed area. Any effects of
this bias will be ignored .
240 SAMPLING TECHNIQUES

TABLE 9.6
SAMPLE ANALYSIS OF VARIANCE (NUMBER OF PAID WORKERS)
(SINGLE-FARM BASIS)

df ms

Between units within strata 825 6.2 18

Between fa rm within units 2768 2.9 18

Between farms within strata 3593 3.676

From the sa mpl e data for individual farm s, the group of three farms can be compared
with the individu a l farm as a sampling unit. The ite m chosen is the numbe r of paid workers.
The sample was stratified, the stra tum bei ng a group of townships similar in de nsi ty of farm
population and in ratio of cropland to farmland . Since the sampling fract io n was 1.9%, th e
fpc can be ignored.

The correct procedure is to compute N h2 S h21nh separa te ly within each stratum for the two
types o f unit , using an a na lysis of variance and expression (9 .8). We will use a simpler
procedure as an approximation .
The strata containe d in general between 300 a nd 450 farms, and ei ther two or thrl!e
3- farm units we re ta ke n in each stratum to make th e sampling a pproxima tl! ly proportional.
Assuming proportiona lit y, th at i. , nhl Nh = nl N, we may write
2
• N 2 • N - 2
" V(Y,,)=-I NhSh =-Sh
n n
if we assume further that t~e S/ do not vary grea tl y among strata, so th at they may be
replaced by their average, Sh 2•
Estimates o f 5h 2 are obtained from the analysis of varian ce in Table 9.6, which is on a
single-farm basis.
For the group of three farms , th e mean square ShJ' = 6.218 serve~ as the estima te o f 5,,2
o n a si ngle-farm basis , For the indivIdual farm , u ing (9.S), we have

6.218+2(2.918) =4 .018
3
By theorem 9.1, co ro llary 2, the two figures. 6.218 for the group of three farm s and 4.018
for the individual farm , indicate the re lative va riances obtained for a fixed total size of
sample. Th e group of farm s gives a bout two thirds tb e precision of the single farm .
Con. idera tion of costs would pres umably make tJle..res~1t more favorable to the three-farm
unit.

9.4 VARIANCE IN TERMS OF INTRACLUSTER CORRELATION

Variance formulas are sometimes expressed in terms of the correlation coeffi-
cientp between elements in the same cluster. This approach has already been used
for systematic sampling (section 8.3) . .
 SINGLE·STAGE CLUSTER SAMPLING 241
Let YII be the observed value for the jth element within the ith Wlit, and let Yi be
the unit total. In cluster sampling we need to distinguish between two kinds of
average: the mean per unit Y = L Y;/N and the mean per element Y = L Y/ NM =
YI M. The variance among elements is
L (Yij - 9)2
5 2 =-,,-1.1_ __
NM - J
The intracluster correlation coefficient p was defined (section 8.3) as

_ -l'A( _
_ E( Yij 1) Yik
l'/'\
1 )
2L
1
I (y" - })(Ylk - })
,<k
p- E(Yi'- Y)2 --'-="(M
.!.....;.::..--1,-)(NM----l-=-)S:-:;'2- (9.9)

The number of terms (cross products) in the nwnerator E is NM(M -1)/2, and in
the denominator E is (NM - 1)5 21NM.
Theorem 9.2. A simple random sample of n clusters, each containing M
elements, is drawn from the N clusters in the population. Then the sample mean
per element y is an unbiased estimate of Y with variance
• 1- / NM - l 2
V(y)=-;;-' M 2 (N-l/ [l+(M -l)p]

== I - f S2 [l+(M - l)p] (9. 10)

nM
where p is the intracluster correlation coefficient. n

Proof. Let Yi denote the total for the ith cluster and y =I yJ n. By theorems
2.1 and 2.2, j is an unbiased estimate of Y with variance
V( -) = ( 1- f) L (Yi - '9)2
Y n N- l
But y = My and Y = Mr. Hence y is an unbiased estimate of Ywith variance
V(y) = 1- fI(YI- y)2 (9.11)
nM2 N- l
But
(YI- Y) = (YIl- 9)+(YI2- 9)+ ... +(YIM- Y)
Square and sum over aU N clusters,

=(NM - 1)52 + (M - 1)(NM - 1)p52

=(NM - l)5 2[t +(M - l)p] (9.12)
242 SAMPLING TECHNIQUES

using the definition of pin (9.9). Sub titute ill (9.11) for V(9) . This gives

V (y) = I - f. ~J-S2[ 1 +(M - l)p]

n M -(N-l)

This completes the proof.

If a simple random sample of nM elements is taken, the formula for V(y) is the
same as (9.10) except for the term in brackcl~. The factor

1 +(M -- l)p
shows by how much the variance is changed by the use of a cluster instead of an
element as sampling unit. This factor is therefore Kish's (deff) for clusters of size M
(section 4.11). If p > 0, the cluster is less precise for a given bulk of sample . If
p <0, as sometimes happens, the cluster is more precilte. Theorem 9.2 is a simple
extension of theorem &.2, p. 209.
2
An alternative expre ion can be given for p. Let Sb denote the variance among
cluster totals, on a single unit basis. Then

1: (y, - 9)2 = (N - 1)MS/

Equation 9.12 can be written as

(N - 1)MSb2 = (NM - 1)S2[1 + (M - l)p] •

so l'hat
(N-l)MS/-{NM-l)S2 S/-S2
(9.13)
P= (NM - l )(M - 1)S2 =- eM _ 1)S2
when terms in 1/N are negligible.
The value of the within-c1uste-r mean square
NM
S,} = 1:1: (Yjj - fM 2/N(M - 1) (9.14)
I j

is worth noting. By the one-way analyltis of variance,

= (N~- 1) S2[1 +(M -1)p] + N(M - 1)S,/

by (9.12). Hence,

(9.15)
 SINGLE·STAGE CLUSTER SAMPLING 243
A good discu ion of the numerical values of p for different items and different
sizes of cluster is given by Han en, Hurwitz, and Madow (1953), who regard p as a
"measure of homogeneity" of the cluster.

9.S VARIANCE FUNcnONS

In some types of urveys, for example, soil sampling, crop cutting, and surveys
of farming that utihze an areal sampling unit. the size of the cluster unit may be
capable of almost continuous variation. In the earch for the best 4nlt the problem
is not that of choosing between two or three specific sizes that have been tried but
of finding the optimum v~due of M regarded as a continuous variable. Thi!:
problem requires a method of predicting the variance Sb 2 betwe rl units in the
population as a function of M By the analysis of variance, Sb 2 can be found if we
know (a) the variance S2 between all elements. in the population and (b) the
variance S,. 2 between ch::ment· that lie ltl the same unit. Our approach i to predict
S,., ~ and S 2 and to find Sb~ bv the analysis of variance .
The sample data produce estimates of S2 and S..2 for the size of unit actually
used . Since S2js the var iance among elements, it is not affected by the size of the
unit . However, S,.,2 will he affected . It might be expected to increa e a the size
of the large unit increases. If the large units to be examined differ hule in size from
the unit actually used, a first approximation is to regard S.} as constant, using the
estimate given by the sample data. An investigation by McVay (lQ47) suggests
that this approximation may often be atisfactory .
As a better approximation, attempts ha ve been made (Jessen, 1942 ;
Mahalanobis, 1944; Hendricks, J 944) to develop a general law to predict how S", 2
changes with the size of unit. In everaI agricultural surveys, S,} appeared to be
related to M by the empirical formula
S,.,2=-AMII (g>O) (9.16)
where A and g are constants that do no depend on M. In this fo rmula S", 2
in' reases steadily a M increases. Usually g is ~ma ll. A curve of this type might be
expected when there are forces that exert a !timilar influence on elements close
together . Climate, soil type . topography, and access to maIkets tend to give
neighboring farms similar feature .
Theoretically, the formula is open to objection, sinoe it makes S,.," incr.:ase
without bound as M increases. If we as Ulllr;, a seem rea \mable. that there is no
correlation between elemen!~ that are far apart, a formula in which S}
approaches an upper bound ' ith large J I would be more appropriate. However, a
formula will ')uffice if il giv~s a goo I fit over the range of M thut is under
investigation.
If this form la fits, log S .. 2 shoW.l pll't as a straight lin against Jog M. Values ot
S ..2for at leaST two values of M aT > n eded in order to estimate the constants log A
and g. At least three value!. of Mare t"IC"ce sary {or a ny appraisal of the linearity of
the fit.
 244 SAMPLING TECHNIQUES

From the analysis of variance in Table 9.5 (p. 239) we find

2 (NM - 1)S2- N(M - 1)S.}
Sb = N- 1

(NM -1)S2- N(M ":" l)AMK

= N- 1
== MS Z - (M - 1)AMIf (9.17)
Hendricks (1944) has pointed out that the complete population might be
regarded as a single large sampling unit containing NM elements. If (9.16) holds,
then S 2 = A (NM)g. The advantage of this device is that the values of A and g can
now be estimated from the data for a survey in which only one value of M was
used. The two equations that lead to the estimates are
log S.} =log A + g log M (9.18)
log S 2= log A + g log (NM) (9.19)
The formula for S/ becomes [from (9.17)]
S/ == AMK[MN' -(M - 1)] (9.20)
This method furnishes no check on the correctness of (9. 16), which might hold
well enough for small values of M but fail for a value as large as NM.
Formula 9.16 is presented as an example of the methodology rather than as a
general law. The reader who faces a similar problem should construct and test
wnatever type of formula seems most appropriate to his material. In some cases
log S/ might be a linear function of M, as Fairfield Smith (1938) suggested for
yield data.

9.6 A COST FUNCllON

In an extensive survey the nature of the field costs plays a large part in
determining the optimum unit. As an illustration of the role of cost factors, we will
describe a cost function developed by Jessen (1942) for farm surveys in which the
large units are clusters of neighboring farms .
Two components of field cost are distinguished. The component clMn com-
prise costs that vary directly with the total number of elements (farms). Thus C1
contains the cost of the interview and the cost of travel from farm to farm within
the cluster.
The second component, cz.[;" measures the cost of travel between the clusters.
Tests on a map showed that this cost, for a fixed population, varies approximately
as the square root of the number of clusters. Total field cost is therefore

(9.21)
 SINOLE·STAGE a.USTER SAMPUNO 245
Assuming simple random sampling and ignoring the fpc, the variance of the mean
per element; is S/lnM. From (9.17), this equals
• S2-(M-I)AM,- 1
V(y) = (9.22)
n
To determine the optimum size of unit, we find M, and incidentaJJy n, to
minimize V for fixed C. The general solution is complicated, although its
application in a numerical problem presents no great difficulty.
By some manipUlation we can obtain the equation that gives the optimum M.
First solve the cost equation (9.21) as a quadratic in../n. This gives

(9.23)

The equation to be minimized is

C+AV=c 1Mn+C2J;;+AV
Differentiating, and noting that aVIan = - Vln , we obtain the equations
I - 1/2 A aV AV
n: CI M +ic 2 n =---=- (9.24)
an n
>.av
M: cln=--- (9.25)
aM
Divide (9.25) by (9.24) to eliminate A. This leads to
n av cln
-=- clM+~c2n - 1J2
V -aM
or

(9.26)

If we substitute for J;; from (9.23), we obtain, after some simplification,

M av ( 4Ccl~ - t/2
--::; 1+-- - 1 (9.27)
VaM c/
By writing out the left side of this equation in full and changing signs on both
sides, we find
AMB- I[gM-(g-l)] 1-{1+ 4Cc t M\ -t/2 (9.28)
s2 - (M - l)AM' 1 ~J
This equation gives the optimum M. The left side does not involve any of the cost
 246 SAMPLING TECHNIQUES

factors , being dependent only on the shape of the variance function. Botb sides
can be seen to be increasing functions of M , for g > 0, M 2: 1, within the region of
interest. Suppose that the solution has been found for specified values of C, CI' and
C2, and we wish to examine the effect of an increase in cIon this solution. The left
side does not depend on Cit but the right side increases as Cl increases. However,
the optimum M is found to decrea e because of the term clM on the right. A
decrease in C2 produces a similar effect.
Now C 1 increa es if the length of interview increases, whereas C 2 decreases if
travel becomes cheaper or jf the farms in a given area become denser. These facts
lead to the conclusion that tbe optimum size of unit become smaller when
length of interview increases
travel becomes cheaper
the elements (farms) become more dense
total amount of money used (C) increases
This conclusion is a consequence of the type of cost function and would require
reexamination with a different function. It iJIustrates the fact that the optimum
unit i not a fixed characteristic of the population, but depends also on the type of
survey and on the levels of prices and wages.
Hansen, Hurwitz, and Madow (1953) give an exceUent discussion of the
construction of cost functions for surveys involving cluster sampling.

9.7 CLUSTER SAMPLING FOR PROPORTIONS

"
The same technique apply to cluster sampling for proportions. Suppose that
the M elements in any cluster can be classified into two classes and that Pi = aJ M
is the proportion in class C in the ith cluster. A simple random sample of n clusters
is taken, and the average P of the observed Pi in the sample i used as the estimate
of the population proportion P.
It will be recalled (section 3.12) that we cannot use binomial theory to find V(p)
but must apply the formula for continuous variates to the Pi' This gives
N 2
N-n 1~1 (Pi - P ) . N-n
V(p)= Nn N-l = N7n 'f,(p,_p)2 (9.29)

Alternatively, if we take a simple random sample of nM elements, the variance of

P is obtained by binomial theory (theorem 3.2) as

v. . ( )= (NM-nM) PO == N-n PO (9.30)

b,n P NM - l nM N nM

jf N is large. Consequently the factor. the deff,

V(p) . M'f,(PI-P)2
(Nlarge) (9.31)
Vb1n(P) = NPQ
 SINGL.E ·STAOE CLUSlER SAMPLING 247
show the relative change in the variance caused by the use of clusters. Numerical
values of this factor are helpful in making prelimjnary estimates of sample size
with cluliter sampling. The required sample size is fir t estimated by the binomial
formula and then multiplied by the factor to indicate the size that will be necessary
with cluster sampling. For an illustration, see Cornfield (1951).
Ii the cluster sizes M; are variable, tht> estimate P = L aJL M, is a ratio estimate.
Its variance is given approximately by the formula (section 3. 12)

(9.32)

where M =L MJ N is the average size of cluster.

If this sample is compared with a simple random sample of nM elements, we
find , as a generalization of (9.3 1),
V(p) . L M ?( Pi - p )2
(9.33)
Vbl,.(p) = NMPO
As with continuous variates, the relationship of size of cluster to between-
cluster variance can be investigated. either by expressing the factor in (9 . 31) and
(9.33) as a function of M, or by seeking a relation between the within -cluster
variance and M. If we assign the value 1 to any unit that falls in class C and 0 to any
other unit, the fundamental analysis of variance equation for fixed M is
NMP(1 - P) ; ML (PI- P)2+ ML P((1- PI)
(9.34)
total ss = ss between clusters + ss within clusters
From this relation the mean square within clusters can be computed and plotted as
a function of M. McVay (1947) describes how this analysi can be used to
investigate optimum cluster size.

EXERCISES
9.1 For the data in Table 9.1 compare the relative net precisions of the four types of
unit when the object is to estimate the total number of seedlings in the bed with a standard
error of 200 seedlings. (Note that the fpc is involved.)
9.2 For the data in Table 3.5 (pAi7) estimate the relative precision of the household to
the individual for estimating the sex ratio and the proportion of people who had seen a
doctor in the past 12 months. assuming simple random samphng.
9.3 A population consisting of 2500 elements is divided into 10 strata. each containing
50 large units composed of five elements. The analysis of variance of the population for an
item is as follows, on an element basis.
df ms

Between strata 9 30.6

Between large units within strata 490 3.0
Between elements within large units 2000 1.6
 248 SAMPLING TECHNIQUES

Ignoring the fpc, is the relative precision of the large to the small unit greater with simple
random sampling than with stratified random sampling (proportional allocation)?
9.4 A population containing LNM elements is divided into L strata, each baving IV
large units, each of which contains M small units. The following quantities come from the
analysis of variance of the population, on an element basis.
5,2 = mean square between strata
5 2 2 = mean square between large units within strata
5]2 = mean square between elements within strata
If N is large and the fpc is ignored. show that the relative precision of the large to the small
unit (element) is improved by stratification if
(M - I) M 1
~<5/ - 5/
9.5 In a rural survey in which the sampling unit is a cluster of M farms, the cost of taking
a sample of n units is

C = 4tMn + 6Wn
where t is the time in hours spent getting the answers from a single farmer . If $2000 is spent
on the survey. the value of n for M = 1. 5, 10; t =i. 2, work out as follows.
M

5 10

t == i hr 400 131 74
t == 2 hr 156 40 21

Verify two of these values to ensure that you understand the use of the formula.
The variance of the sample mean (ignoring the fpc) is
52
Mn [1 +(M- I)p]

=
If p = 0.1 for all M between 1 and 10, which size of unit is most precise for (a) t i hr, (b)
t = 2 hr? How do you explain the difference in results?
9.6 If $5000 were available for the survey, would you expect the optimum size of unit
to decrease or increase (relative to that for $20oo)? Give reasons. You may, if you wish,
find the optimum size in order to check your argument.
 CHAP TER 9A

Single-Stage Cluster Sampling:

Clusters of Unequal Sizes

9A.l CLUSTER UNITS OF UNEQUAL SIZES

In most applications the cluster units (e.g., counties, cities, city blocks) contain
different numbers of elements or subunits (areal units, households, persons). This
chapter deals with some of the numerous methods of sample selection and
estimation that have been produced for cluster units of unequal sizes. Let M/ be
the number of elements in the ith unit. For the estimation of the population total Y
of the Y/j, two methods are aJready familiar to us.

Simple random sample of dusters: unbiased estimate

As before, let

denote the item total for the ith cluster unit. Given a simple random sample of n of
the N population units, an unbiased estimate of Yis (by theorem 2.1, coroUary)
~ N"
y =- L y, (9A.l)
n 1- 1
By theorem 2.2, its variance is
N
A

V(Y) =
N 2( l - f) L (Y/ - Y)2
;_1-..:.,I_ __
(9A.2)
11 N- 1
where Y= Y/ N is the population mean per cluster unit.
The estimate Yis often found to be of poor precision. This occurs when the YI
(means per elemel1t) vary little from unit to unit and the M j vary greatly. In this
event the YI = M/YI also vary greatly from unit to unit and the variance (9A.2) is
large.
249
 250 SAMPUNO TECHNIQUES

Simple Random Sample of austers: Ratio-to·Size Estimate

Let
N
Mo == L M; = total number of elements in the population.
1M )

If the M, and hence Mo are al1 known, an alternative is a ratio estimate in which
M; is taken as the auxiliary variate X"

L" Y,
YR = Mo..!.:::!._.
n
= Mo (sample mean per element)
L M,
I- I

In the notation of the ratio estimate the population ratio R = Y/ X = Y/ Mo= 9',
the population mean per element. By theorem 6.1, assuming that the number of
clusters in the sample is large,

(9A.3)

N
== N 2 (1- f) L M/(y; - 9')2
(9A.4)
n N- J
"

As (9A.4) shows, the variance of YR depends on the variability among the means
per element and is often found to be much smaller than V( n.
Note that YR requires a knowledge of the total Mo of a1l the M;, while Y does
not. The reverse is true when we are estimating the population mean per element.
In this case the corresponding estimates are

.)_y = - Y = -N- L Y"

n -'-
YR = -
YR = -nf- Y, == sample mean per element
Mo nMo Mo LM;

Thus, -r
R requires knowledge of only tbe M; that fall in the selected sample.

,
9A.2 SAMPLING WITII PROBABILITY PROPORTIONAL TO SIZE
If all the M, are known, another technique, developed by Hansen and Hurwitz
(1943), is to select the units with probabilities proportional to their sizes M,. One
 SINGLE-STAGE CLUSTER SAMPLING 251
method of selecting a single unit is illustrated in the followin g smaJI popula tion of
N = 7 uni ts.

Size Assigned
Unit M, 'f.M, Range

1 3 3 1-3
2 I 4 4
3 II 15 5- 15
4 6 21 16-21
5 4 25 22-25
6 2 27 26-27
7 3 30 28-30

The cumulative . ums of the M j are formed. To select a unit, draw a random
number between I and Mo = 30. Suppose that this is 19. In the sum, number 19
falls in unit 4, which cover the range from numbers 16 to 21 . mclu",ive. With this
method of drawing. the probahiljty that any unit another is "elected is propor-
tional to its size.
This method of selecting a unit is convenient when N is only moderate, or in
stratified sampling when the Nh are moderate or small, but the cumulation of th
M I can be time-consuming with N large (e.g., N = 20,000). For this case, Lahiri
(1951) has given an alternative method that avoids the cumulation. Let Mmu
be the large t of the MI' Draw a random number between t and N ;
suppose this is i. Now draw another random number m between I and MffUi1(' If m
is less than or equal to M;, the ith UOlt is selected. If not, try another pair of random
numbers. Naturally, thi method involves the fewest rejections whcn the M j do
not differ too much in size.
Now consider n > 1. Assume at present that sampling is with "'placement. To
select a second unit by the cumulative method, draw a new random number
between 1 and 30. However, unlike sampling withou t replacement, we do not
forbid the selection of unit 4 a second time. Wtth this rule, the probabilities of
selection remain proportional to the sizes at each draw. An advant ge of selection
with replacement is that the formulas for the true and estimated variances of the
estimates are simple.
In sampling without replacement, on the other hand (section 9 A.6), keeping the
selection probabilities proport ional to the chosen sizes is more difficult and sooner
or later becomes impossible as n increa es. This may be seen in the extreme
(although impractical) ca en = 7 in the preceding example. If selection were made
without replacement, every unit would be certain to be chosen. irrespective of the
original sizes M I' However, for ~tr8 tified sampling in which the Nil are small, milch
research has heen done (section 9A.6 If) to deve lop practical methods of sampling
with unequal probabilities without replacement.
 252 SAMPLlNG TECRNIQUES

9A.3 SELECI10N WITI:I UNEQUAL PROBABILmES

WITH REPLACEMENT
Let the ith unit be selected with probability MJ Mo and with replacement,
where M o = '- M I' We will show that an unbiased estimate of the population lotal
Yis

= M o (mean of the unit means per element) (9A.5)

For com pari on with subsequent methods, this estimate will be denoted by Y,.."...
Furthermore,
b Mo ~ --ih2
(9A.6)
V(r,.."..) = - '- M, (YI - I}
n I- I
so that the variance of Ypps , like that of YR , depends on the variability of the unit
means per element.
In some applications the sizes M I are known only approximately. In others the
"size" is not the number of elements in the unit but a measure of its bigness that is
though! to be highly correlated with the unit total YI' For instance, the "size" of a
hospital might be measured by the total number of beds or by the average number
of occupied beds over some time period. Similarly, various measures of the "size"
of a'restaurant, a bank, or an agricultural district can be devised. Consequently,
we will consider a measure of size M;' and a corresponding probability of selection
N
ZI = M/ / M o', where Mo' = L M/ . As far as the theoretical results are concerned,
the ZI can be any set of positive numbers that add to lover the population. It will
be shown that '

Y = _!_
ppz
n
f ~
I- I Z,
(9A. 7)

is an unbiased estimate of Y with variance

V(Yppz ) = -1 L
A N (YI
ZI - - Y)2 (9A.8)
n I_ I ZI

The proofs utilize a method introduced in section 2.10. Let II be the number of
times that the ith unit appears in a specific sample of size, n, where II may have any
of the values 0,1,2, ... , n. Consider the joint frequency distribution of the II for
all N units in the population.
The method of drawing the sample is equivalent to the standard probability
problem in which n balls are thrown into N boxes, the probability that a ball goes
 SINGLE-STAGE CLUSTER SAMPLING 253
into the ith box being z, at every throw. Consequently the joint distribution of the
Ii is the multinomial expression

For the multinomial, the following properties of the distribution of the I, are well
known.
V(!,):: nZi(I - z,) , (9A.9)

Theorem 9A.l. If a sample of n units is drawn with probabilities Zi and with

re placement, then

(9A.I0)

is an unbiased estimate of Y with variance

VCYppz )= -I rN Z,(y-l._ 'Y~2

J (9A.ll )
n i- I Zi

Proof. We may write

Y.
ppz
='!'(t
n
YJ+ r/2+ .. '+ 1 YN)=.!. ~ I X!.
I ZI Z2 n NZN ':'1 'ZI

where the sum extends over all units in the population. In repeated sampling the
t's are the random v riables, whereas the Y, and the Xi are a set of IIlced numbers .
Hence, since E(t, ) =nx, by (9A.9).
1 N Y, N
E(YppJ =- L (nz,)-= L Yi = Y
n '- I Z, i-I

so that Y""z is unbiased. Also,

V(Y""r)=~[f (l!)2 V(t;)+2 I-If ;>1f l!_li.CoV(tllj)]
n I - I Z, Z, ZI
(9A.12)

1[ N
=- L (1!.)2zl(1-z,) -2 L L :..!.!J.
V.Y ZIZ,]
N N
(9A.13)
n I- I Z, Z,
, - I j>I'<:,

= -1 (N
L..L_ 2
Y y 2 =-1 ) -- YJ2
LN z, (YI (9A.ll)
n I - I ZI n 1- 1 Z,

since L z, = 1. This completes the proof.

Takingzi =M;/Mo in theorem 9A.l give tbecorrespon<lin~ r su N lnl ~.ampl
ing with probability proportional to size.
 254 SAMPLING TECHNIQUES

An alternative expression for V(')'>ppz) can be given . From (9A.13),

~ . N Y12(1 - Zi) N N
nV( Y ppz ) :: I 2 I I YIY, (9A.14)
, "1 Z, I- I J> i

Since ( I - z,) equals the sum of all other z 's in the population, the coefficient of
y," / Zi in (9A.14) contains a term zi for any j¢. i. Similarly. the coefficient of z, y/ /
(;ontains a term ::,. Hence,
V ( }' ) ""'.~ ,,'!, ~ ('Y/ZI+Y/ ZI _ 2YY)
~. n i -- I /-> 1 Z, Z/ ' f

=- -
1 N N
I I zlz, (Y...!. _ ily ) 2 (9A.15)
n ' - '/ > 1 Zj Zf

11Ieorem 9A.1. If it sample of n unit!> is drawn with probability proportional

to z, with replacement. an unbiased sample estimate of V ( Ypp. ) is, for any n > I,

LI ( Ypp~)= ,~,
n (yi- Yppz )2/ n (n - l) (9A.16)

Proof. By the usual algebraic identity,

f (l1._Yppz) = f (ll._y\J2- n('Ypp: _ y)2

j~ 1 ZI
2

,- I ZI
(9A.17)

Henc<': ~!.~'m (9A. 16)

E[n(n - l)v(Ypp;)] = E I
n (y-'--Y)2 - nV(Y ppz ) (9A.18)
I_ I Z,

by efinllll)O of '(Yppz)' Introducing the variables tit we get

lI(n - l)E[v(tppz)] = E IN tl (Y-'- - YJ\2 - nV(Yppz )

I_ I Zi

=n I" Z,(Y-'- - Y)2 -nV(Yppz )

I- I Zi

=n(n - 1) V( ~:.".) (9A.19)

u iog ([JA.Il) in theorem 9A.1. This completes the proof.
SlOCt; f ;".. is the mean of the n values y,/Z" formula (9A.I6) has a very simpl .
form.
When selection is strictly proportional to size (i.e., ZI = M,/M o), theorems 9 A.I,
9A. 2 are expressed as follows.
1beorem 9A.3. If a sample of n units is drawn with probabilities proportional
tu size, Z, = MJ Mo and, with replacement,
./',
TpJn = -
Mo"
n i _I
I (y,)
-M, = -Mo ~ _
I.. (YI) = Moj
n I _I
(9A.20)
 SINGLE-STAGE a..USTER SAMPLING 255
where 9 is the unweighted mean of the unit means, is an unbiased e timate of Y
with variance

(9A.2l)

These resul ts follow from theorem 9A.l , since YI = YiM; and 9" = YjMo.
Theorem 9A.4. Under the conditions of theorem 9A.3, an unbiased sample
estimate of V ( Y"". ) is

v(Ypps) = Mo2 £ (y;-y)2/ n (n - I)

i- I
(9A.22)

'The result follows by substituting ZI = MJ Mo in (9A. 16), since Y, = yd M/ and

Y"", =Moy.
9A.4 1HE OPTIMUM MEASURE OF SIZE
In cases in which the measure of size M/ is some estimate.of the bigness of the
I

unit, a question of interest is: what measure of size minimizes the variance of Y""z ?
Now,

1"" (Y;
V (Y""z)=-LZ; -- Y )2=-1 (Ny/ - Y 2)
L-
n ZI n Zi

This expression become~ zero if Z, oc YI: that is, Z, = yJ Y. If the YI are all positive,
this set oC ZI is an acceptable set of probabilities. Consequently, the best measures
of size are numbers proportional to the item totals y/ for the units.
This result is not of direct practical application; if the YI were known in advance
for the whole population the sample would be unnecessary. The result suggests
that if the YI are relatively stable through time, the most recently available
previous values of the YI may be the best measures of size for this item. In practice,
of course, a single measure of size must be used for all items in selecting the
sample. If there is a choice between different measures of size, the measure most
nearly proportional to the unit totals of the principal items is likely to be best.

9A.S RELATIVE ACCURACIES OF THREE TECHNIQUES

In this section we compare the accuracies of three preceding techniques for
estimating the population total with cluster units of unequal sizes (assuming that
the M; are known if the technique requires them).

1. Selection: equal probabilities, Estimate; 9"".

2. Selection : equal probabilities. E tim te; YR'
3. Selection : probabilityocsize. Estimate; 9""".,.
 256 SAMPLING TECHNIQUES

There i no simple rule for deciding which is most accurate. The issue depends
on the relation between Yi and Mi and on the variance of Yt as a function of M, . The
situation favorable to the ratio and pps estimates is that in which Yt is unrelated to
Mi. The situation favorable to Y.. is that in which the unit total y, is unrelated to M,.
Some guidance can be obtained by expressing the variance of the three
estimate in a comparable form. We assume that (N - 1) == N and write E(Yi -
N
Y/ = I: (Yi - YlIN. We also a sume that the bias of YR is negligible .
For Y.. we have, from (9A.2),
n V( Y.. )= N 2 (1 - f)E(y, - Y)2::: (1 - f)E(NYi - Y)2 (9A.23)
For YR. from (9A.4),
n V( YR )::: N 2 (l - f)EM/ (YI - 9')l = (1- [JE{~) \Moy; - Y)2
(9A.24)
where M = I:M,IN= MolN.
From (9A.2l). for YpJM,

nV( Yp"., )= NMoEM,(Y; - Y)2= Mo2E(~)(YI - 9')2

=E(~)(MOYI - Y)2 (9A.25 )

• From (9A.23), (9A.24), (9A.25), we see that V( Y.. )depends on the accuracy of
the quantitie. Ny, =NM,y; as estimates of Y, while V(YR ) and V( YpJM ) depend on
the accuracy of the quantities MoY; = MoyJ M; as estimates of Y. If Yi is unrelated
to M" we expect the MoYI to be more accurate than the NM,y" and the reverse if y,
is unreldted to M,.
As regards YR and Ypl''' note from (9A.24) ~and (9A.25) that V(YR ) giv~s
relatively greater weight to large units than V(YpJM )' Note also that Y.. and YR
benefit from the fpc term, which can become substantial in small strata (e.g., with
nh = 2, Nh = 10). This point stimulated the development of unequal probability
selection without replacement. Formula (9A.24) for YR , of course, holds only in
large samples.
Further comparisons among the methods have been made from an infinite
population model by Cochran, 2nd ed. , Des Raj (1954, 1958), Yates (1960).
Zarcovic (1 960)~ and Foreman and Brewer (1971). Most writers assume that the
finite population is a random sample from an infinite superpopulation in which
19A.26)
which il> hoped to approximate the relation that holds in many surveys. Some
assumption must also be made abou t the variance of t, in clusters of given size.
 SINGLE·STAGE CLUSTER SAMPLING 257
From (9.12) in Section 9.4, we get on dividing by (N - I) == N,
V(el ) = V(YIIMt)
== MtS2[MIP +(1- p))
As an apptoximation this suggest'i V(el) = cMJ' where 1 < g < 2 in most applica-
tions. From the model (9A.26) we get
I'r a eN
T = -=+{3+-= (9A.27)
M M
We assume here that eN is negligible ; this amounts to ignoring the fpc.
It follows that,
nV(YJ
N2 E(YI - Y/ = E[{3 (M, - M) + etl 2

={3 2V(M,) + cE(MtB ) (9A.28)

nV~fR) == EM/ (YI- 9')2::: E(y; - Mt9')2

2
- 2 a V(M;) 6
= E[a(l - M;/M)+e;] = M2 + cE(Mt) (9A.29)

nV~pprJ =MEMI(y, - 9')2 =MEMt[ a(~; - k) +;;J

2...J -
= a cL (M~: ] + VME(Mt'l -
2
I
) == a ;;~I) +CME(MtB - 1)
2
(9A.30)

Hence, approximately comparable variances under this model are

~" = {3 2V(M;) + CE(M,B) (9A.31)

nVR -. . :. . a 2 ----=-or-
V(Mi ) + cE(M.B) (9A.32)
N2 M< I

nVr::: a2~(~i) +cME(Mt, - I) (9A.33)

N M
Consider first a = 0, the case in which YI is unrelated to M,. Clearly,
VR < V u
If {3 (which in this case becomes 9') is large, the superiority of the ratio estimate
may be great. With a = 0,
Vpp.r < V" for g 2: I
This holds because the covariance of M; and M;,- I is positive if g> 1 and zero if
258 SAMPLING TF;rHNIOl/ES

g = 1, so that with g ?:. 1,

E(M/l = E[(M,)(M/ I)) ?:. ME(M/ I)
If f3 = 0 and a ye n, so that the unit lOcal y, i unrelated to M b Y" alway~ beats YR
and beats f'pp, except p sibly in the unlikely case g = 2 with f3 =: O.
When neither a nor {3 vanishes, the relative performances of Yu and )'tR' }'p,,,
depend on the relative sizes of 10'1 and 1131. or instance. YR heats Y" if
13 2 > a·2 / M~ . as noted hy Foreman and Brewer ( 1971 ).
As regard. VR versu), V,>" " the oefficients of (l' ~ are approximately the . arne in
Vn and VII'''' From the terms in c, V,'PI Vn in the ca~e g > I expected to apply to
most applications that have 13 'I- O. while '. '",,, .' V R if g < 1.
Th e re su lt~ from thi~ model agree with the co nclu~ions suggested earlier in thi~
section. If y, shows no trend or only a modera te trend as M, increases. the ra tio
method with equal probabilities and the pps method are more preci~e than
unbiased estimation with equal probabilitie~ and may be much more precise . Yu i~
superior if the unit total ', is unrelated to M, . There is Ie s to choose between the
ratio and the pps estimates. Since g b expected to lie between I and 2, the pps
estimate is usuall y mme precise and it~ results are not restricted to large sample~ .
The estimates Yu and }'R are he lped by the fpc term when this is appreciahle .
To anticipate later work (section lJA.I2l. the evidence ~ug,gests that the best pps
methods without replacement benefit from about the same size of fpc term as in
equal -probability,ampling.

!)A.6 SAMPLING WITH UNEQUAL PROBABILITIES WITHOUT

REPLACEMENT
Much of this work wa produced for extensive ~ urveys in which the cluster units
had first heen stratified by some otht:r principle (e.g., geographic location) into a
. ubstantial number of re latively ~mall strata, only a small number of cluster units
heing drawn from each stratum . The ca<;c:: ~h = 2. which provides one degree of
freedom from each ~tratum for estimating !>ampling errors, is of particular
interest.
Suppo~e that two units are to be drawn from a stratum. The first unit is drawn
with probabilities zl' proportional to some measure of size. Let the ith unit be
~elected . If we follow the most natural method , at the second draw one of the
remaining units is se lected with assigned probabilities z/( I - z,) . Henee the total
probability 11', that the ith unit will he selected at either the first or the second draw

.. 11', : Z,
N ~z ( N Z
+ 2: ~ == Z, I + I ~
, .. ,(1-z,)
)

", l -z,
(9A.34)

: ; z,( I + A __ z,_)
1 - z,
(9 A .35)

where A =I z/(1- z,) taken over all N units.

 SINGLE-STAGE CLUSTER SAMPLING 259
Suppo e that 1r, = 2z j , the relative probabilities of selection of the units remain-
ing proportional to the measure of size Z/. The simple estimator introduced by
Horvitz and Thompson (1952),

'YUT =f, l!..

1r,
=! f Xl
2 i Z,
will then have zero variance jf the z, are proportional to the y" since Zj = yJ Yand
every selected unit will give the correct estimate yJz, = Y. However, in (9A.35),
the quantities z,' ::: 1rJ2 are always closer to equality than the original Zj because
of the second factor in (9A.35) . ln the example given by Yates and Grundy (1953),
with N = 4, n = 2, Zj = 0.1 , 0.2, 0.3, and 0.4, the z/ are found to be 0.1173,
0.2206,0.3042, and 0.3579.
Three approaches have been used to cope with this distortion of the intended
probabilities of selection. One is to retain the preceding natural method of sample
selection, but make appropriate changes in the method of estimating the popula-
tion total. One is to u e a different method of sample selection that keeps 1rj :::;; nZ j
subject to certain restrictions on the values of the nz,. One is to accept some
di tortion of the probabilities if this has compensating advantages (e.g. in simplic-
ity or generality). Examples of the three methods will be given in section 9A.B and
following ections.
First, we present the best-known general. estimate of the population total for
unequal probability ampling without replacement.

9A.7 THE HORVITZ-THOMPSON ES11MATOR

A sample of n units is selected, without replacement, by some method . Let
probability that the ith unit is in the sample
1r/ :::;;

1r'J = probability that the ith and jth units are both in the sample
The following relations hold :
N N N
I 1r/ ::: n, I 1r/J = (n -1 )1rI, I, ,I> ; 1r'J= i n (n - l) (9A.36)
lqil,'

To establish the econd relation, let pes) denote the probability of a sample
consisting of n specified units . Then 1r'J :::;; IP(s) over all samples containing the ith
and jth unit , and 1r/ = IP(s) over all samples containing the ith unit. When we
take I1rjj for j ¢ i, every P(s) for a sample containing the ith unit is counted (n -1)
times in the sum, since there are (n - 1) other values of j in the sample . This prove
the econd relation . The third relation follows from the second.
The Horvitz- Thompson (1952) estimator of the population total is
A
YH1' =
I" -y, (9A.37)
i 1T',

where y/ is the measurement for the ith unit.

260 SAMPLING TECHNIQUES

Theorem 9 A.S. If 'TT; > 0, (i = 1, 2, ... , N),

yHT=rl.!.
, TT,
is an unbiased estimator of Y, with variance

VCYHT ) = f
,_ I
(1-TT')y/+2
1T;
ff
i- I j> i
('TTlj -77"I77"j)YIYj
1T't'TTj
(9A.38)

where TT;j is the probability that units i and j both are in the sample.
Proof. Let (; (i = 1, 2, . ..• N) be a random variable that takes the value 1 if the
ith unit i drawn and zero otherwise. Then (; follows the binomial distribution for a
sample of size 1, with probability 'TT, . Thus
E(l,) = 77";, (9A.39)
The value of Cov (t;(,) is also required . Since ('(f is 1 only if both units appear in the
sample,
Cov (t;t, ) = E(/;t,) - E(t;}E(/,) = TT" - 77"I'TTf (9A.40)
Hence, regarding the Y, as fixed and the t, as random variables.
~ (~(,y;)
E( YHT') = E '- -
f a 1 1T;
= '-~
j_ 1
Yf = Y

"

(9A.41)

This proves the theorem .

This variance may be expressed in another form by using the first two of the
relations in (9A.36). These give
L (77"ij -
,,,, 77";77"f) = (n - 1)77", -TT;('. - 'TTl) = - TT;(l - 'TT;)

Substituting for (1-77";) in the first term in (9A.41),

Hence,

(9A.42)
 SINGLE·STAGE ClUSTER SAMPliNG 261
Cor~Uary. From (9A.41), using the tl method, an unbiased sample estimator
of V(YHT ) is seen to be
~ (l - +2 ~ ~
( ./',HT ) -_ I...
vir
1T1)
2 YI
2
I... I...
(1Tlj - 1T;1Tj)
YIYI (9A.43)
1 1T1 1 />1 1T11T/1TI/

provided that none of the 1Tlj in the population vanishes.

A different sample estimator has been given by Yates and Grundy (1953) and
by Sen (1953). From (9A.42), this estimator is

V2( 9'HT) =f f (1T,1Tj - 1T;) (Y; _ l1.) 2 (9A.44)

I 1> 1 1Tij 1T1 1Tj

with the same restriction on the 1T1/'

Since the terms (1TI1T; -1TI/) often vary widely, being sometimes negative, VI and
V 2 tend to be unstable quantities. Both estimators can assume negative values for
some sample selection methods. Rao and Singh (1973) compared the coefficients
of variation of VI and V2 in samples of n :: 2 from 34 small natural populations
found in books and papers on sample surveys, using Brewer's sample selection
method (section 9A.8). for which 1T; = 2z; as desired. The estimator V 2 was
considerably more stable as well as being always ~O for this method, while V I
frequently took negative values.
We tum now to methods of sample selection. The bibliography by Brewer and
Hanif (l969) mentions over 30 methods, of which 4 methods will be described.
Most methods become steadily more complex as " increases beyond 2, a few
extend fairly easily. The presentation will give most attention first to n = 2, both
for simplicity and because this is a common situation in nationwide samples using
many small strata with "II == 2.

9A.8 BREWER'S METHOD

For n = 2 this method of sample selection keeps 1T; = 2z1 and uses the Horvitz-
Thompson estimator

YHT == l!.+1L=!(!!+l'.L)
1T1 1Tj 2 ZI z/

Using different approaches, methods produced by Brewer (1963), Rao (1965).

and Durbin (1967) all gave the arne 11"1 and 1Tli values. We assume every z, < 1/2.
Brewer draws the first unit with wha he calls revised probabilities proportional
to zl(1- zl)/(1- 2zl ), and the second unit with probabilities z;/(1- z,), where j is
the unit drawn first. The divisor needed to convert the zl(1 - zl )/(1 - 2z l ) into
actual probabilities is their sum

D= r zl(1 -zl):::: ! r z,(2- 2Z )_!(1+r_3_) I (9A.45)

'-I 1- 2zl 2'_1 1-2z, 2 1 1-2z,
262 SAMPLING TECHNIQUE

With IT = 2 the probability that the ith unit is drawn is the sum of the
probabilities that it was drawn first and drawn second. Thus

(9A.46)

noting (9A.45, for D. Similarly,

7T = ~(_1_ +_1_) =2z, z, (1 -2, - z)

(9A.47)
" D 1- 2z, 1- 2z, D (1-2 z,)(1 - 2z,)

Since this method use the HTestimate , theorem 9A.5 and its corollary provIde
formulas for the variance and estimated variance of YH .
The method has two desirable properties. Brewer (1963) has shown that its
variance is always less than that of the e timate YpPl in . ampling with replacement.
Second, som~ algebra shows (Rao, 19(5) that (7T,7T,-- 7T"J > 0 for all i :j: j, so that
the Yates-Grundy estimate V2 of the variance is alway!> positive .
Durbin's ( 1967) approach draws the first unit (i) with probability z,. If unit i was
drawn first . the probability that unit j is drawn second is made proportional to

Z.[ I + 1 ] (9AAH )
' (1 - 2z,) (1 -2z, )
In this case the divisor of the proportion~ is
N [I I ] ( I - z,) N z N Z
L zi +- - - = - -+ L = 1+ L ____::L_
I '" ~ ( 1- 2z,) (1 - 2z, ) 1-2z, " ,(1 - 2z, ) ,: 1 ( - 2z,
(9AAf.J)

Thus the divisor is equal to 2D in Brewer's (9AA5) . The probability that the ith
unit was drawn first and the jth unit second is, therefore.

P(i)P(j Jil = i5 [(I _12z ,) + 0 _12z, J (9A.50)

By symmetry, this equals P(j)P(iJj) , so that Durbin's 7T" is the same as Brewer's in
(9A.47) . .
Sampford (1967) has extended this method to samples of size n, provided
nZI < 1 for all units in the population. With his method of sample selection, the
probability that the sample consi ts, for example. of units I. 2, ... , n is a natmal
extension of (9AA7), being proportional to

( 1- f z,) fI
;- 1 i- I
Z'/ fI
, .. ,
(1 - nz,) (91' .51)
 SINGLE·STAGE CLUSTER SAMPLING 263
For this method it can be shown that 1fj = nZi. A formula for 1fij is given, with
advice on its calculation by computer. The HT estimator of Y is used, so that
formulas are available for its variance and estimated variance. The Yates-Grundy
estimator V2 in (9A.44) is always positive. Several methods of actually drawing the
sample so as to satisfy (9A.51) are suggested by Sampford. One is to draw the first
unit with probabilities z, and all subsequent units with probabilities proportional
to z/(l- nZj) with replacement. If a sample with n distinct units is obtained, this is
accepted. An attempt at a sample is rejected as soon as a unit appears twice . This
method can be seen to lead to (9A.51). As a guide to its speed, a formula is given
for the expected number of attempts required to obtain a sample.
For n = 2 Durbin's (1967) method of drawing the sample, unlike the Brewer
method, has the property that the unconditional probability of drawing unit i is z,
at both the first and the second draws. ]n multistage sampling in surveys repeated
at regular intervals, Fellegi (1963) pointed out earlier that it is necessary or
advisable to drop unit~ and replace them from time to time on some regular
pattern called a rotation scheme, because of the undesirability of long-continued
questioning of the same persons. He produced a method of selection of successive
units that also has the Durbin property. His method, based on iterative calcula-
tions, is similar to the Brewer- Durbin method , but has slightly different 1fir

9A.9 MURTHY'S METHOD

This method uses the first selection technique suggested (section 9A.6). the
successive units being drawn with probabilities z" z/(l- Z,). Zk/(l- Z, - z). and
so on . Murthy's estimator (1957) follows earlier work by Des Raj (1956a).
who produced ingenious unbiased estimates based on the specific order in which
the n units in the sample were drawn . Murthy showed that corresponding to any
ordered estimatc of this class we can construct an unordered estimate that is also
unbiased and has smaller variance.
His propo ed e<;limator is
"
L P (sl i)y,
YM '" I P(s) (9A.52)

where
p (sli) = conditional probability of getting the set of unit that was drawn , given
that the ith unit was drawn first

pes) = unconditional probability of getting the set of units that was drawn

We now prove that the eStimate YM is unbiased . For any unit i in the
population, L P(sli) = 1, taken over all s mples having unit i drawn first. To show
264 SAMPLING TECHNIQUES

this for n =2, where j is the other unit in the . ample ,

N

Z·
L zf
P(sli) =-=:L..; LP(sli)=~= 1
I -zi l -zi
For n = 3, with j and k the second and third unit,
N N N
L p(sli) = L L Z/Zk/(l - zl)(l- z, _. zi) == L z/(l - ZI) = 1
,"I k "'J /.,,1
and so on for n > 3. Hence, when we sum L pes ) YM over all samples of size " , the
coefficient of YI in the sum is 1, so that

(9A.53)

General expressions for VCYM) and l' ( f'M) for any n have been given by
Murthy (1957).
When n = :' , the sample consisting of units i and j ,

P(s li) =-.!L; P(sjj)=_Z_, (9A.54)

I -zi 1-zi
p es) = 1TiJ = zIP(sli) + z~(slj} = ZIZP- z, - zt>I(1- zi)( l - zi) (9A.55)
The estimate is therefore, using (9A.52), (9A54), (9A.55),

(9A.56)

(9A57)

After substituting for YM"from (9A56) and for y l and rearranging, we find

(9A.58)

By compari on with V(Yppz ) for sampling with replacement, formula (9Al5),

clearly V(YM ) < V(Yppz ) ,
Since for any proposed v( YM) its average is L pes )vCYM), an unbiased sample
estimate of wriance for n = 2 is

(9A.59)

which is always positive.

 SINGLE-STAGE CLUSTER SAMPLING 265
The corresponding forplUla for samples of size n is

v( YM ) = [P(~)f {~ It [P(s)P(sli, j) - P(Sli)P(slj)]ZIZi(~-~r} (9A.60)

where P(slij) is the conditional probability of getting the sample given that units i
and j were selected (in either order) in the first two draws. A computer program
for calculating the P(sJi , and P(slij) has been given by Bayless (1968).

9A.IO ME11IODS RELATED TO SYSTEMATIC SAMPLING

A method suggested by Madow (1949), which Murthy (1967) notes has been
used in Indian surveys, is an extension of systematic sampling. Its advantages are
that the sample is easy to draw for any n, the method keeps 1T'1 = nzt. and it
provides an unbiased estimate of Y. Furthermore, the sampler may sometimes
have enough advance knowledge to arrange the units at least roughly in an order
under which this systematic sampling performs well. A drawback of the systematic
method is, as usual , the absence of an unbiased v ( Y,ys}'
A sample of size n may be drawn using either the z, or the measures of size M I '
Ir.
from which the ZI = M;' M/ = M/1Mo' were derived. If the M/ are used, form a
column of the cumulative totals 'ft of the quantities nM/. From this column,
within the interval 1 to nMo' assign a range of size nM,' to unit i. For sample
selection, draw a random number r between 1 and Mo' and select the n units for
which the assigned range includes the numbers r, , + M o', . .. , r + (n -1)Mo'.
This method is illustrated by the data used in Section 9 A.2 with N = 7. We
suppose n = 3. Having formed the T, and the assigned ranges, draw a random
number between 1 and Mo' = 30, say' = 17. The units selected are those whose
ranges assigned include the numbers 17, 4 7, 77, that is, units 3, 4, 6.
Size Assigned Units
Unit M/ T, =3 IM,' Range Selected

1 3 9 1- 9
2 1 12 10-12
3 1] 45 13-45 17 (unit 3)
4 6 63 46-63 47 (unit 4)
5 4 75 64- 75
6 2 81 76-81 77 (unit 6)
7 3 90 82- 90

M'
0 -- 30

If nZj < 1 (i.e., nM;' So Mo') for all i, any unit has a probability nZI of being
selected, and no unit is selected more than once. For instance, unit 5 is selected if
4s r So 15, giving probability 12/30 = 3z s. If nZI> 1 for one or more units, such
 SAMPLI NG TECHN IQUES

units may be selected more than once in the sa mple, but the average frequency of
<;e lection is 7r, = nz,. This happens in the example for unit 3, since 3M )' = 33 >
30 = Mo'. Unit 3 is chosen once if 1 _ r s 12 or 16 S r S 30 (in all 27 choices), and
rwic I;' if 13 _ , S 15 (3 choices) . The average frequency of , election i (I x 27 + 2 x
3)/ 30 "" 33/30 = 3z 3 •
It ,follows th at
y'Y:. = ~l!. = _!_~~
~ l- (9A.61)
I 7r, n I Z,
I~un unbiased e timate of Y.
Hartlcy and Rao ( I (62) examined this method wi th the units first arranged in
random order . With the restriction nz, < I. (all i) . they obtained approximate
expressions for V( Y,yJ and v( Y,yJ .

9A.1I mE RAO, HARTLEY. COCHRAN METIIOD

For a ~ample of size n, this method first form~ n random group of units. olle unit
to be drawn from each group . The numb er~ of unib N 1• N 2 , • •• , n in the
rc~pective group, may be chosen in advance : we will ,ee that it is advantageous to
make the N~ as equal as possible . if ZR is the tot.ti mca~ ure of ize for group g, the
ith unit in the group is given probability of ~e l ectinn zJ ZR' The estimate of the
populal ion total. Rao , Hartley. and Cochran ( IlJ( 2 ). i ~

(t) A .o2)

"
where YK , ZM refer tn the unit drawn from group g.
Since the Zx will not be equa l, this method does not keep the probabilities of
-..:lcction pr oportional to the sizes, and there i~ some evidence th at its e~timator
~u lfcr' CI ~ l i!!ht l o ~s in precision. Its advantages are its ~impl i.:i1 y and generality.
In dcvcll)ping V('YR H C ) we average over two stages. Stage I b the randomiza-
Imn mtn group. . stage 2 the selection of a un it wit hin each group. For any specIfic
'plit into gr~) ups, Ys in (9A.62) is an unbiased estimator of the group total Y~. and
hence f:',( YHHC ) = Y. A well -k nown formula f\ r finding a var,iance over two
\!ages of sampling, proved in Chapt er 1O. i ~
V( Y'lH(" ) = El[V~(YRIi )] + Vl[E2(YR llr l] (10 .2)
S1l1ce 2( YRNC ) = Y and is a constant, it h a~ lero variance and the second term in
( 10.2) di~appears in this application .
For V~ ( YRllci we can use . within a gro up. the variance formula for sampling
with replacement, si nce only one unit i5 se lected from each group. Within a gr up
the pf("lbabi lities of selection are z,/ ZR' By (9A. 15) in ection 9 A.3 we get, for any
,peclfic spli1 (with nR = 1),
(9A .63)
 INGLE-STAOE CL STER SAMPLING 267
taken over the pair of units in group g. Over the set of random subdivisions in to
~roup • the probability that any pair of units falls in group g is Ng(Ns -1)/
N(N - I) . Hence the average val ue of V 2 ( YIr ) over the randomization is

l::,V~(y~)==NK(NM~ l) E f z,z/~-lLr = N (N _-l)(f: Y/_y g B 2

) (9A.64)
N(N I) , }> I zJ N(N I)
Z, i - I Z,

!rom (9A .63). Since YnHc = ... Y"

(r =I
l V 2( Y~ RH )] = ~ N(N
N/-N)
(""t -z-
N
y, Y2)
2 n(I r-IN/-: N) ~ )
(9A.65)
EI - l) . = N (N -l) V(Y ppz

Thu V( YRHC ) i<; si mpl y a mUltiple of the variance of the estimate Ypl" in sampling
with replacement. If N/n is integral. the choice Nr. = N/n minimizes the multi -
plier. 111 th i~ case
• (n-1).
, '( ) RHI ) - 1- -=l '" () '1'1" ) (9A.66)

It N = nR + k, with R integral and k ..-- n, the b !>t t:holC(' is to make k gwups of

"t:7.e (R + 1I, the remaining (n - k I of size R. Thi<; gives
. 1 k [n -
k )] •
V( YJ<H( I = I - N _ 1 + N(N _ 1) V( Yp/,: )
(n- (YA.67)

If N/ n I" intcgral.(QA.661 glve~ V( } ' RHcl/ V( )',,,,z);: (N - n lieN - J). the same
ratio a~ obtained in . implc random ~a mpling .
An unoi..1"t!d "allance estimator can be show n In be

• (._£, N/ - N) n ( v ~ -
)'

v( YnHcl == ( n ) ~ Zs ,L&_ Y RHC (9A.6R)

N2 L N/ ~ I z~
~- I

With N = fiR +k. and k groups of size (R + 1). formula (9A.68) becomes
A N + k(n-k) - Nn n
_
2

V(YR Hc)-N2( n _ l) _ k(n _ k )f ZIr

(It ZR-YRHC
~ )2 9A.69)

9A.U NUMERI AL COMPARISON

In the Iiterutu re, comparisons of tht· perfo rmances ot some of the methods bay
been made , parti ularly by Rao and Bayless (1969, J 97t J, in three l\ua tion~ . (a I
on small artificial populations [e.g., the population!. with ,"i = 4, n =:; on tructcd
by Yate and Grundy (1953)], (b) under the linear rcgre 'sion mex. '1 u cd Jtl
section 9A.5. and (c) on 20 natural populations.
 268 SAMPLING TECHNIQUES

Seven methods are compared here on three artificial populations with N =5,
n = 2. The relative sizes ZI of the units are the same in all three populations
(A, B, C) (Table 9 A. I).' In A the mean per element, which is proportional to yJ Zit
is uncorrelated with Z I' In B the mean per element rises as the sizes increase. In C
the unit totals have little relation to the sizes.

TABLE 9A.l
THREE SMALL ARTIFICIAL POPULATIONS
Relative sizes ll ;) 0.1 0.1 0.2 0.3 0.3

Yi 0.3 0.5 0.8 0.9 1.5

Population A
yJz. 3 5 4 3 5
y, 0.3 0.3 0.8 1.5 1.5
Population B
y,/z, 3 3 4 5 5
Y, 0.7 0.6 0.4 0.9 0.6
Population C
y,/7., 7 0 2 3 2

The plans compared are as follows .

1. Simple random sampling of the units (equal probability). Estimate: 9'SRS'=

Ny.
2. Simple random sampling of the units. timate : Ratio to size, YR =
0, Lyt/'f.. ZI'
3. Sampling with probability proportional to size with replacement. Estimate:
Y,,,u =(l/n) L (y;/ z, ).
4. Brewer's method, sampling without replacement. Estimate : 'YB =
(l/n) L (Y;/Zf)'
5 . Murthy's method. Estimate : YM = [(1- z,) y, +(1 - z,).!:'.t.]/(2
ZI z, - Z, - z,).
6. The RHC method with one group of three units, one of two units. Estimate:
YRHC =L Z,(y,/ z,) .
7. Madow's systematic method, with units arranged in increasing order of yJ ZI'
Estimate: 9'.yo =(1/ n) L (Ydzl ).

To obtain a rough idea of what to expect in the relative performances of

methods like B, M, and RHC versus simple random sampling with equal
probabilities, note that the relative variances of the newer methods to V( 9'SRS )
are approximately the relative coefficients of variation of the quantities Y;/ ZI and
YI' The ratios CV(yi/ZI)/CV(YI) are 0.19 in population A, 0.11 in B, and 6.7 in C.
Thus the unequal probability methods should have a relative efficiency of around
5 in A, 9 in B, but only 0.15 in C. This comparison is somewhat unfair to the
unequal-probability methods, which weight the squared errors in y;/ z, by the
 SINGLE·STAGE CLUSTER SAMPLING 269
measure of size. The MSE 's of YR should be fairly close to the variances of the
unequal-probability methods.
In a choice of a method in practice, major considerations are the ease with
which the sample can be drawn , the simplicity of the estimate, the accuracy of the
est imate, and the availability of an estimator of the variance of the estimate. With
n = 2, all methods are simple. As regards accuracy, the systematic method is
shown here only in its most favorable case, which the sampler rarely has the
knowledge to use. All methods except Y,Y5 and YR provide unbiased sample
estimates of error variance. Table 9A.2 presents the variances (with the MSE
for YR ).

TABLE 9A.2
VARIANCES OF THE ESTIMATED POPULATION
TOTALS
Population

Estimate A B C

t~RS 1.575 2.715 0.248

YR(MSE) 0.344 0.351 1.42 1
Y ppz 0.400 0.320 1.480
YB 0.246 0 .248 1.251
YM 0.267 0.237 1.130
Y RHC 0.320 0.256 1.184
y... 0.150 0.140 0.760

In equal-probability selection, the ratio of V( Y) in sampling without replace-

ment to V(y) in sampling with replace!Dent is (~ - n) /(N - 12=O.75 for
these populations. For YRH , the ratio V( YRHC )/ V( Yp " . ) , where Yppz denotes
sampling with replacement is 0.8 in Table 9A.2 . For the other unequal-
probability methods, the ratio varies from population to population. The average
of the three ratios in A, B , and C is 0.74 for Brewer's method and 0.73 for
Murthy's method, about the same ratio as for equal-probability methods. Results
for n = 3, 4 on small natural populations by Bayless and Rao (1970) agree in
general with (N - n)/(N - 1).
In populations A and B, all other methods are much more accurate than YSRS
with simple random sampling. The ratio-to-size estimate with SRS performs
roughly similarly to the unequal probability methods, although not quite as well.
In the latter, there is little to choose among the Brewer Murthy, or RHC
methods. The systematic method at its best performs very well.
In population C, in which the unit totals bear little relation to the sizes, simple
random sampling with equal probabilities is much the best. As noted at the end of
 270 AMPLING TECHNIQUES

section 9A.13, this superiority is probably due to the estimator Ny, not the SRS
feature.
Rao anti Bayless (1969) compared 10 unequal-probability methods in 20
natural populations found in books and papers on sampling, with N ranging from
9 to 35. They confined themselves to methods (a) known to have smaller variances
than Yppz and (b) providing a positive unbiased variance e timator. Among the
method pres~nted here, they compared the efficiencies of YM, YRH and Ypp:
with that of Y8' For n = 2, there was little to choose among the three " without
replacement" methods, with YM slightly ahead, beating YRHC whenever the two
methods differed in precision . Also, we have noted (section 9A.7) that th
variance estimator may be unstable for methods using the Horvitz- Thompson
estimate . The Rao-Bayless results compare the coefficients of variation of v( YM) ,
VCYRHc), and vCYB ) in the Yates-Grundy form with that of v(Yppz ) , as measures
of the stabilities of the variance estimators. Relative to v( YB ) as 100%, the
median efficiencies of the other variance estimators \\ ere : V(YRH )=109%,
v(YM ) = 104%, v( Yppz) = 97%, the three methods all showing a few large indi-
vidual gains.
Bayless and Rao (1970) give similar comparisons for n = 3 (14 populations) and
n = 4 (10 populations). Sampford's extension of the Brewer method was used . For
n much beyond 2, both Sampford's and Murthy's methods require computer aid
in calculating the needed probabilities. The variances of YM and Ys agreed closely
in nearly all populations, with YRHC slightly behind, its median efficiency relative
to Ys dropping to 92% for n = 4. In stability of the variance estimators, on
the other hand, the superiority of YRHC and YM increased, the median relative
-'efficiencies being 118% (n = 3) and 129% (n = 4) for v( YRHc) and 110% (n = 3)
and 120% (n = 4) for v(YM ) relative to 100% for v(Ys ).
Rao and Bayless also compared the efficiencies of Yand v( y) for some of the
estimators under the linear regression model of ection 9A.S with a = O. While
comparative results depended on the power g, the general trend was similar to that
in the natural populations.

9A.13 STRATIFIED AND RATIO ESTIMATES

The preceding formulas have been presented as they would apply to a single
stratum, although the concentration on small n implied previous stratification by
another principle (e.g., geographic location, urban-rural), The extenSion of the
formulas to this stratification is as usual. For any method, if nl" Zhl, 7Th l, Yhl. and so
forth, refer to stratum h,

(9A.70)

Ratio estimates enter either when the variable of interest is a ratio (e.g.,
unemployed females/females eligible for work), or when they are used to increase
 SINGLE·STAGE CLUSTER SAMPLING 271
precision , In unequal-probability sampling a single choice of the z, or Zit, must be
made for all variables for which estimate are required. For instance, the Zj or Ziti
may be proportional to the total recent sales of a type of business, where the
survey has to estimate current sales of individual classes of items as well as current
total sales. For some classes, sales may not be closely proportional to the Z" In
such cases, use of the familiar "ratio to the same variable last time" estimate may
bring substantial increases in precision.
With unequal-probability sampling the change in formulas to those for ratio
estimators is easily made. In an unstratified population, replace Yfor any method
by XYI X. For instance, with the HT estimator, we use X(f
yJ'TT'i) (x,/Trl) for If
r
the ratio form instead of (yJ 11',) . For the standard approximations to the MSE
and estimated MSE of a ratio estimate, replace y, by d, = (y, - Rx,) in V( Y) and by
d;' = (Yi = Rx, ) in v(Y) .
For instance, with the ratio form of the Horvitz-Thompson estimator, the
approltimation to V2, the estimated variance in the Yates-Grundy form, is from
(9A.44),p. 261 , .

V2[ YIfl"(R») == rr
I j '> i
(11',11'1 - 11"1)
1Tij
(d;' -~
"tT,
,\2
7TJ~J
(9A. 71)

In a stratified population it is likely with smaJ) nit that the combined ratio
estimate (section 6.11) will be used [i.e., Y = X(r Y,,)/(r X,,)). For approximate
variance formulas, replace YIt, by dll , = YIII - RXIII in V and by d"i' = Ylol - Rxlol in v.
When the y, for an item are not related to the ZI and no suitable ratio estimate is
feasible for such an item, Rao (1966) has investigated alternative estimators.
These are produced from any of the unequal-probability estimators (m: M ,
RHC, etc.) by replacing yslz, by Ny" wherever y, appears in the estimator. Thus,
for the Horvitz-Thomp on estimator, the alternative form is

(9A.72)

while for the RHC method

"
YlH = NrZ,y,
, (9A.73)

The estimators are biased but intuition suggests that if the YI have no relation to
the z" the biases should be relatively small. By the same method as used in finding
VCYIfl") in (9A.42), we get
N2 N N
v[Y1n-]=-rr
n
r (11'111'1- 11'jj)(Y, - Yj) 2
I j >1
(9A.74)

Thus V[ Y*Ifl") depends on the amount of variation in the unit totals, as with
V( YSR ) . In population C, Table 9A.2, this method gave 0.266 and 0.280 for the
272 SAMPLING T CHNIQUES

MSE's of the alternative forms of YHT and YM , these forms doing almost as well a
YSRS' For both methods the (Bias) 2 term was about 4% of the MSE.
EXERCISES
9A.l Horvitz and Thompson (!9~7.) give the foil wing data for eye estimates M, of the
numbers of households and for the actual numbers 1', in 20 city blocks in Ames, Iowa. To
assist in the calculations, value of Yi and 9// M, are also given. A sample of II = 1 block is
chosen . Compute the variances of the total number of households Y, as obtained by (a) the
unbia ed estimate in sampling with equal probabilities. (b) the ratio estimate in sampling
with equal probabilities, (c) sampling with probability proportional to M,. (For the ratio
estimate, compute the true mean square error, not the approximate formula.)

Mi Yi iii yNMt M,· Yi ilt y?/Mi

9 9 1.0000 9.000 19 19 1.0000 19.0;00

9 13 1.4444 18.778 21 25 1.1905 29.762
12 12 1.0000 12.000 23 27 1.1739 1.696
12 12 1.0000 12.000 24 21 0.8750 18.375
12 14 1.1667 16.333 24 35 1.4583 51 .04i
14 17 1.2143 20.643 25 22 0.8800 19.360
14 15 1.0714 16.071 26 25 0.9615 24.038
17 20 1.1765 23 .529 27 27 1.0000 27.000
18 19 1.0556 20.056 30 47 1.5667 73.633
18 18 1.0000 18.000 40 37 0.9250 34.225

Do 'the results agree with the dist:ussion in section 9A.5?

9A.2 A questionnaire is to be sent to a sample of high schools to find oul which schools
provide certain facilities, for example, a course in Russian or a swimming pool. II M, is the
number of students in the ith school, the quantity to be estimated for any given facility is the
proportion P of high-school students who are in schoolS having the facility, that is,
'i,M,
p = -"'--
N
'i, M ,

where'i, is a sum over those schools with the facility.

ow
A sample of II schools is drawn with probability proportional to M, with replacement.
For one facility, a schools out of II are found to possess it. (a) Show that P=a/II i an
unbiased estimate of P and that its true variance is P(l - P)/ II. (Hint. In the corollary to
theorem 9.4 let y, = M, if the school has the facility and 0 otherwise .) (b) Show that an
unbiased estimate of YeP) is v(P) =P(l - P)/(II - ]) .
9A.3 The large units in a population arrange themselves int a finite number of ~ize
classes: all units in class h contain M~ small units. (a) Under what condition does ampling
with pps give, on the average, the same distribution of the size classes in the sample as
stratification by size of unit, with optimum allocation for fixed sample size? (b) If the
variance among large units in cla~s h is kM~, where k is a constant for all cia ses, what
 SINGLE-STAGE CLUSTER SAMPLING 273
system of probabilitie, of selection of the units gives a sample in which the sizes have
approximately the sa me distribution as a stratified random sa mpl e with optimum allocation
for fixed sa mpl e size?
9Ao4 For a po pulation with N = 3, Z, =~,!, ~ and y, = 7,5,2, two units are drawn
without replacement, the first with probability proportional to z,' the seco nd with probabil·
ity proportional to the remaining sizes . (a) Verify that 1T', = ~i,. 1T'2 ""~, 1T'J = Mand that
11
1T" 2 = lo, 17' :0'
= 1;8. 17'~, = (b) For this method of sample selection, compare the variance~ of
Yin and YM and also compare them with the va ri ance of YRHC using its-method of ~ample
,election. You may either COns tru ct .all three yossible estimates or use the variance
fo rmulas . (c) Show that the ratio of V( Y M) to V( Y pp,) in sampling with replacement is clo~e
to the value! that applies for equal probability sampling.
9A.S For the population in exercise 9Ao4. a second va riable had values y ~, = H. 5. 9, !:lot
at all ~Iosely related to the z" so that with the sampli ng method used 10 exercise 9A.4. Y m
and Y would be expected to perform poorly for this variab le . For Rao\ estimator
Y1rr= I .S(Y2' + Yo,), compare its M E wi th the variance of Y~"" "" I .Sly" + Y2,)
111 equal-probabi lity samplinj!. How much does bias contribute to the MSE ?
9A.6 For Brewer 's method with n = 2, section 9A.!l showed that

'IT = 4Z'ZI(I-Z' -ZI )/(I+~_Z_'_)

'I (I - 2z,)(1-22I ) i 1 -22,
(a) Show that if every z, <!,
0 < ."." < 4z,zl (eve ry i 'F j)
(b) Show that this result makes the Yates-Grundy estimator of variance always positive
for this method .
Hint. For (a) it is sufficient to show th:lt

(I - z. - z)= (I-2z )(1 - 2z )[] +_i_,-+~J

' I ' , 1 - 2z, I - 2z,
9A.7 (a) For Durbin's method with n = 2, verify directly that the probability that the
jth unit is drawn second is'zl as stated on p. 263 .
(b) With N = 4, Z, = O. J, 0.2, 0 .3, 004, calculate the probability that with Brewer's
method unit J is drawn first and the probability that it is drawn second . Verify that the two
probabilities add to 0 .2 = 2z,.
9A.8 In Madow's systematic method, a unit may be chosen more than once in the
sample if nZI > 1 (i.e., nMl ' > Mol Show (as stated on p. 266) that for such units the average
freq uency of selection is nZI, so that the Horvitz-Thompson estimator of Y remains
unbiased for nZI > 1.
 CHAPTER 10

Subsampling with Units

of Equal Size

10.1 TWO-STAGE SAMPLING

Suppo e that each unit in the population can be divided into a number of
mailer units, or subunits. A sample of n units has heen selected . If subunits within
a selected unit give similar results, it seems uneconomical to measure them all. A
common practice is to select and measure a sample of the subunits in any chosen
unit. This technique is called subsampJing, since the unit i not measured com -
pletely but is itself sampled . Another name , due to Mahalanobis, is two -stage
sampling, because the sample is taken in two steps. The first is to select a sample of
units, often called the primary units. and the second is to select a sample of
second-stage units or subunits from each cho en primary unit .
'f Subsampling has a great variety of applications, which go far beyond the

immediate scope of sample surveys. Whenever any process involves chemical,

physical, or biological tests that can be performed on a smaiJ amount of
material, it is likely to be drawn as a subsample from a larger amount that is itself
a sample.
In this chapter we consider the simplest case in which every unit contains the
same number M of subunits, of which m are chosen when any unit is subsampled.
A schematic representation of a two-stage sample, in which M = 9 and m == 2, is
shown in Fig. 10.1.
The principal advantage of two-stage sampling is that it is more flexible than
one-stage sampling. It reduces to one-stage sampling when m == M but, unless this
is the best choice for m, we have the opportunity of taking some smaller value that
appears more efficient. As u ual, the issue reduces to a balance between statistical
preci ion and cost. When subunits in the ame unit agree very closely, considera-
tions of precision suggest a small value of m. On the other hand, it is sometimes
almost as cheap to measure the whole of a unit as to subsample it, for example,
when the unit is a household and a si ngle respondent can give accurate data about
all members of the household.
274
 lJl:ISAMPLlNG WITH UN ITS OF EQUAL SI ZE 275

I:8J denotes an element ,n the sample

FiR. 10.1 Sche mati c representa ti on of two-stage sa mpling (IV; 81. n ; 5. M ; 9. m '" 2).

10.2 FINDING MEAN AND VARIANCES IN

TWO-STAGE SAMPLING
In two-stage sam pling the ampling plan gives first a method for selecting n
units. Then, for each selected unit , a method is given for selecting the specified
number of subunits from it . In finding the mean and variance of an estimate,
averages must be taken over all samples that can be generated by this two-stage
process. One way of calculating this average is first to average the estimat over all
second-stage selections that can be drawn from a fixed set of " units that the plan
l>elects. Then we average over all possible elections of n units by the plan . For an
estimate 6. this method can be expressed as
£(8) = E,[E2(8)] (10.1)
where E denotes expected or average value over all samples, £2 denotes
averaging over all possible second-stage selections from a fixed et of units, and £ I
denote averaging over all first-stage selections.
276 SAMPLING T CHNIQUES

For V(O) this method give the following easily remembered result.
\/(0)= V I [E 2 (8)]+E I [V2(0)J (10.2)
where V 2 (0) is the variance over all possible subsample selections for a given set of
units . To show this, let 0 = E(O) (where 0 is not necessarily the quantity that 8 is
designed to estimate, since 8 may be bia ed) . By definition,
V(O)=E(O - O)1 = E I E 2 (0 - 0)2 (10.3)
But
E 2 ( 0- 0)2 = E 2 ( O~) - 20E2 ( 0) + 0 2 ( LO.4)
= [E2 (8)J + V 2 (0) - 20E2 (0) + (} 2
2
(l0.5)
Average now over first-stage select ions. Since E) E 2(0) = 8,
YeO) =: E I [E 2(0)]2- (} 2 + E I[ v2 ee)] (10.6)
= V I [£2(0)) + E)[ V 2(0)] (10.7)
Formula (10.7) extends naturally to three 0: more stages. For three-stage
sampling,
(10.7')

.. 10.3 VARIANCE OF THE ESTIMATED MEAN IN

lWO-STAGE SAMPLING
The following notation is used.
YiJ = value obtained for the jth subunit in the ith primary unit

Yr = f
i- I
bL:: sample mean per subunit in the ith primary unit
m

y= I
i- I
~ = over-all sample mean per subunit
n
N
I (Y, - 9)2
2 _I ~ ) . . .
5I - N _} :: variance among pnmary uOlt means

N M - 2
I I (Yi/- Y i )
5 22 = i- I ~~M _ 1) variance among subunits within primary units

Note that Yj denotes the total over-all subunits in the ith unit (denoted by Yi in
Chapters 9 and 9A).
 SUBSAMPLING WITH UNITS OF EQUAL SIZE 277
Theorem 10.1. If the n units and the m subunits from each chosen unit are
selected by simple random sampling, y is an unbiased estimate of Y with
variance

(N -n) --+
V( Y-) == - -
N
S/n (M-m)
---
M
s/
mn
(10.8)

Proof. With simple random sampling at both stages,

E(9) == E 1[E 2 (9)] == E1(~ f ¥i) = (~f ¥;) = y (l0.9)

For V(y), we use formul a ( 10.2).

(10.10)

"
Since E 2 (y) = I Y, / n, the first term on the right is the variance of the mean per
subunit for a one-stage simple random sample of n units. Hence, by
Theorem 2.2,

(10. 11)
n
Furthermore. with y =: I y,/n and simple random sampling used at the second
stage,

-) _ (M - m)
V 2 (Y - 2
f S2/ (10.12)
Mn m
M
where S2/= I(Y;j- Y,) 2/(M-I) is the variance among subunits for the ith
I "
primary unit. When we average over the first-stage samples, I 5 2 / / n averages to
N i

I 5 2/ / N = 5 22 •
i
Hence

(10.13)

The theorem follows from formula (10.10), on adding (10.11) and (10.13).
If II = n/ Nand i2 = m/ M are the sampling fractions in the first and second
stages, an alternative form of the result is

(10.14)
 278 SAMPLING TECHNIQUES

10.4 SAMPLE ESTIMATION OF THE VARIANCE

Theorem 10.2. Under the conditions of theorem 10.1, an unbiased e timate
of V(y) i

(10.15)

whe re I, = 11/ N, 12= m/ M , and

, f (y, - 9) 2 L L (YiJ - jiJ 2

5, • = =-=n'---l~ 5/ =..;.'...J.j_ _~ ( 10. 10)
n(m - 1)

Proof.

(10.17)

Hence

where f'n = f YJ n. The last term on the right holds because subsampling is
independent in different units and 9= L" y,/ n. Thus,
'f

( 10.19)

Multiplying by (1 - I,) / n(n - 1) and averaging over the first stage of simple
random sampling,

(10.20)

By comparison with (10.14) for V(y), note that the term in 5 2 2 is too small by the
amount II (1-12)522/ mn. Since £, E 2(5/ ) = 5/, an unbiased estimate of V(y) is
therefore
_) _ 1- 1, 2+1,(1 - 12)5 2 2
v (Y - --s, (10.21)
n mn

Coronary. A result used later is that from (10.20).

2
£(S, 2) =5/+ (1 - 12)8/ = 5/ +5/ _ 5M2 (10.22)
m m
 SUBSAMPLING WITH UNITS OF EO AL SIZE 279
2
[I follows that an unbiased e tim ate of SI is
~
2
'\·2 ( I- h)
$1 -
m
NOles on Theorem 10.2. [f m = M, that i , I~ = I , formula (I 0. 15) b\.:come~ that
appropriate to simple random sampling of the units . If n == N. the formula is that
for proportional stratified random sampling. since primary unib may then be
regarded as strata, all of which are sampled. In this connection. tWll-stagt:
~a mpling is a kind of incomplete . tratification , with the units as strata .
When II = n/ N is negligible. we obtain the simple result ,

2 f (y, _ 9)2
v(y) = ~ = -' ~_I-,---_- ( 111 . 2 I)
n n (n-I)

Thus the estimated variance can be computed from a knowledge of the unit me an,
only. Thi result is helpful when ubsampling is systematic, becau~e in thi~ CH'n!
we cannot compute an unbiased estimate of S: 2 • But ( 10.:13) still applies. providcd
that n/ N is small. If II/ N is not small. (10.23) overc ·timates hy the amount
[IS//n. a~ seen from (10.20) and ( IO.I ..n

10.5 THE ESTIMATION OF PROPORTION

If the ubunits are cia 'ified into two classes and we estimate the rroportiol1 that
falls in the first class, the preceding formulas can be appli ed by the usual dC\'ice of
defining y" as I if the corresponding subunit falls into this class allLi a~ ze ro
otherwise. Let p, = aJ m be the proportion falling in the fir. ! cia. ~ in the ~uhsamplc
from the ith unit. The two estimated variances s12 and s/ required for theore m
10.1 work out as follows:

f (p, _ p)2
s/ = ,-' ...:,1_ _-
n- 1
In n

s/= n(m - I) i~1 Plql

where p = "i.pJn. Consequently, by theorem 10.2,
_ 1- I, ~ ( -)2 II (J - 12) ~
00.24)
vCP)= n (n - 1)t...I PI-P + n 2( m - 1) t...I PIl;

Example. In a study of plant disease th plants were grown in 160 small plots
containing nine plants each. A random sample of 40 plots wa chosen and three random
plants in each ampled plot were examined for the pre ence of disease. It wa found that 22
 280 SAMPLING TECHNIQUES

plots had no diseased plants (ou t of three), 11 had one, 4 had two, and 3 had three . Estimate
the proportion of diseased plants and its s.e. The symbol <p de notes the frequencies 22, t 1.
4,3.
We have N = 160, M = 9. n '" 40, m = 3. In finding s , 2 and S2 2 , it is convenie nt to work at
first with the numbers of diseased plants (3p,) and the numbers of healthy plants (3qJ The
calculations are et out as follow .
Frequency
3p ; </> 9p,q, 9</>p ,q, 3</>1' , 9.pp,t

0 22 0 0 0 0
II 2 22 II II
2 4 2 8 8 16
3 3 0 0 9 27

40 30 28 54

_ 3 L <pp, 28
p = 3 L <p = 120 = 0.233

_ 2 1 ( (28)2)
L<P(P, - P) = (9) 54 - = 3.822
40
30
L<PP4, "'"9= 3.333

lienee, from the formula immediately before this example,

__ (3)(3.822) + (2)(3.333)
v(fi) - (4)(40)(39) (4)(3)(1600)(2) 0.00201
The proportion diseased is 0.233 with s.e. 0.045. The approximate formula s,/-In, from
(10.23), gives 0.049, a reasonably good estimate considering that 1, = i.

10.6 OPTIMUM SAMPLING AND SUBSAMPLING FRAcnONS

These depend on the type of cost function . If travel costs between units are
unimportant, one form that has proved useful is
C=c l n+c2nm
The first component of cost, Cln, is proportional to the number of primary
units in the sample; the second, c2nm, is proportjonal to the total number of
second-stage units or elements. From theorem 10.1, V(f ) may be written
_ 1( 2
V(y)=; SI - M
S22)+ mn1 S2 - N1 S ) 2 2
(10.25
)

The last term on the right doe 'Dot depend on the choice of nand m. Minimizing V
 svnSAMPLlNG WITH UNITS OF EQUAL SIZE 281
for fixed C, or C for fixed V, is equivalent to minimizing the product

( V+ ~S/)C ~ [( S1 2- ~) + ~2JcCI +C2m)

By the Cauchy-Schwa~z inequality ,

_ S2 FrTr"..
mop. - J 2 2 CI/C 2 (10.26)
S, - S2 / M
provided S ,2 > S2" / M. Round mop. to the nearest integer. If m is an integer such
that m < mop. < m + 1, a slightly better rule for mop. small is (Cameron, J951):
round up if m ~p. > m(m + 1), otherwise round down . If mop. > M orif S/ < S/ /M,
take m = M , using one-stage sampling. (The product (V + Sl 2 / N)C is a strictly
decreasing function of m when S / < S / / M .)
The value of 11 is found by solving either the cost equation or the variance
equation , depending on which has been preassigned .
In most practical situations the optimum is relatively flat. An error of a few units
in the choice of m produces only a small loss of precision. as the following example
illustrates. Write

S 2= S
u I
2_ S/
M (1O.27)

Example. LeI

then

mop< = D.JiO = 4. I
We will regard total cost as fixed and see how the variance of y changes witb m. N is
assumed large . From (10.14).
2
.) S. l S2
V( Y =- + -
n 11m

= (S}+ ~2)C. +Cmc2

eliminating n by mean of the cost equation. This gives

- , ' 1+S/
V(y) =s}r.( -- )(c,
-+ m)=S}c
-- 2( 1.69)
J+ - (IO+m)
( mS} C2 C In

Omitting the constant factor. the.: relative variance can be calculated for different values of
m. Table 10. I shows these vlJrianres and the relative precisions (with the maximum
precision for m = 4 taken as the standard) . .
For any value of m between 2 and 9, Ih loss of precision relative to the optimum is less
than 12%.
284 SAMPLING TECHNIQUES

deals with the choice of n' and m' . If s I 2 is the variance hetween unit means and S 2 2
is the var.iance between subunits within units, as defined in section 10.4, (10.22)
gives

•
E(s
I
2) = (s s/)
M I
+ S/ = s + S2
2_
m' " m'
2
2
(10.31 )

Foc the simple cost function cln +c2nm, we had

S2 ~
m"p, = J 2 2 CI/CZ
51 - 52 /M
As an estimate of mop, from the pilot survey, (10.31) suggests that we take

S2 .;;;;;
mup, = J 2 2 J cl / r, = 2 2 JcJcz (10 .32 )
Sl - S 2 /m' J(m'si / S2 ) - I
The estimate m"p, is subject to a sampling error that depends on the sampling error
2
of the ratio SI /S /. From the analysis of variance it is known that m's I2/ s/ is
distributed a

F
t::'( l+m ,S}
S/ )
where F has (n ' - I ) and n '(m ' - ') degrees of freedom , provided that the y" are
normally distributed . This result leads to the sampling di stribution of "lop, for
given values of n' and m ', that is,

Jm'c, / co
JF( + ;~))
A

(10 .33 )
mop, = 1 m - 1

Example. For the example in section I 0.6, in which

52 = 1.35.,
consider how well m o", is estimated from a pilot sample with n ' = 10 and m ' = 4 . From
(10.33),
6.324 6.324
=-r============
A

m
0 ", • JFlI +(4/1.69)] - 1 JJ .367F - l
where F has 9 and 30 df. To find the Iimit~ within whLch mop, will lie 80% of the time, we
have, from the] 0% one-tailed significance levels of F,
Flo (9, 30) = 1.8490, F.90 (9, 30) = 1/ F IO (30, 9) = 1/2.2547 = 0.4435
Substitution of these values of F gives
lower limit, m.,.. = 2.8; upper limit, m.", = 9.0
 SUBSAMPLING WITH UNITS OF E QUA L SIZE 285
As shown in Table 10.1. any m in this range gives a degree of precision close to the
optimum . Thu~, with n' = 1n, m ' '" 4, the chances are 8 in 10 that the loss of precision is
small with normal data .
The 80 and 95°1., limits for n' "" 5. 10. 20 and m' = 4 appear in Table 10.3. With n' = 20,
we are almost certain to estimate map' with precision close to the optimum . This is not so
with n' = 5.

TABLE 10.3
LOWER AND UPPER LIMITS FOR ,nUll I
n' 80·;'; 90~~

5 2.5. oc I.B, 0-

10 2.R.9.0 2.3,
20 3.1.6.4 2.7,9.1

If the ratio C I / c ~ is the same in the pilot survey as in the main ~urvcy . the cost of
the pilot survey will be proportional to C In' + C2n' m '. Brooks (1955) gives a table
of th values of (n '. m ' ) in the most economical pilot survey that provide. an
expected relative precisicll1 of 9() % in the estimation of mOT'" Table J0.4 show part
of thi:- table .

TABLE 10.4
PILOT SAMPLE DbSIGNS HAVING AN EXI'FCTED RE~ATIVE PRH: IStO'-i Of 90. 0

cl /c~ ~ 1 2 4 B 16 32 64

S2 2 / S ,, 2 n Ill ' n' /II ' ,,' III ' n' nl ' n /11 ' n' 1/1 n' /II

1 7 3 6 4 6 5 5 6 5 7 4 10 4 12
2 Ii 5 '7 7 6 9 6 9 5 13 5 14 4 20
4 9 9 8 II 8 12 7 14 7 15 5 25 5 27
8 10 14 10 15 9 17 9 IB B 22 6 32 5 44
16 10 :5 10 27 10 27 10 28 II 37 7 46 6 60
32 10 46 10 47 10 48 10 49 9 58 8 69 6 102
64 10 92 10 93 10 96 10 100 10 104 8 137 7 169

The computations assume that Nand M are large : the designs are con. ervative
if fpc terms are taken into account. Note that no more than J 0 primary units are
required and that the designs are relatively insensitive to the ratio CI/C 2'

10.8 THREE·STAGE SAMPLING

The process of subsampling can be carried to a third tag by sampling the
subunits instead of enumerating th m completely. For instance. in urveys to
286 SAMPLING T CH NIQ U S

estimate crop production in India (Sukhatme, 1947). the village is a conve nient
ampling unit. Within a village, only some of the fields growing the crop in
question are selected , so that the field is a subunit. When a field is sel cted, only
certain parts of it are cut for the determination of yield pe r acre; thus, the subunit
itse lf is sampled . If phy ical or chemical analyses of the crop arc involved. an
additional subsa mpling may be used. si nce these determinations are often made
on a part of the sample cut from a field .
The result s are a straightforward extension of those for two-stage sampling and
are given briefly. The population contains N first -stage units. each with M
, econd-stage units. each of which has K third-stage unit s. The corresponding
numbe rs for the sa mple arc n, m, and k, respective ly. Let Y'J" be the value obtai ned
for the uth third-stage unit in the jth second -stage unit drawn from the ith primary
unit . The relevant population mea ns per third -stage unit are as follows .
K MK NMK

LY'J" LY"" L 1:1: )",,,

9:, :: -'-"-- ~= ' 1 (.4
Y" ==7 ' MK' NMK
The following population varia n ce~ are required.
N •
L (,Y, - Y) "
S 1=...:.'_ _ __
I N- J
/'1M
'f L L (Y,) -: Y,)2
<; 2 _ ..;.
' -',_ _ __
2 - N(M - I)

Theorem 10.3. If si mple random sa mpling is used at all three stage, thc
sa mple me an y per third-stage unit is an unbiased estimate of~ with variance
V(~) = .!._=.uS 2 +~ - ! 2S 2+ 1- ! 1S 2 (10.34)
. n I nm 2 nmk ~

where!1 = n/ N,f2 == m/ M.h = k/ K are ~he sampling fraction s at the three stage .
Proof. Only the principal steps are indicated. Write
y- y= (y- Y,,,")+(Y,,,,, - Y")+('Y,, - 9) (10.35)
where Y"", is the population mean of the"rn second-stage units that were selected
and Y" is the popu lation mean of the" primary units that were selected. When we
square and tak e the average , the cross-product terms vanish . The contributions of
 UBSAMPLlNG WITH UNITS OF Oli AI. SIZE 2 7
the squared terms turn out to be a~ follows .

When these three terms are added . the theorem is obtained .

Theorem 10.4. An unbiased estimate of V(y ) from the ~amplc is

") 1- / , , Id 1- [1) , 1,12(1 - /,) ,
v (y = --s ,- + s,- + S 1- ( 10.36)
n nm - nmk .

where s/ , s/, S ,2 are the sample analogues of ,2,5/ ,5/. respectively.

Proof. Thi!> may be proved hy the methods in section 10 .4 Of alternatively by
showing that

(10.37)

and E(s \~ ) '" 5 / To ohtain the first result, let y,,, denote the meC'ln over the m
second-stage unit~ in the ith primary unit. given that all K clement!>. were
enumerated at the third stage . Let y"
he the mean of the n v a lue~ Y,k' Then . from
(I() 22) for two-stage sampling, it follow!. that

( IO.3R)

ow, if 9, is the sample mean for the ith primary unit , write

(IO .3Y)

By first averaging over samples in which the first -stage and . econd - ~tage Ul1lt~
ar fixed, it lTIay he shown th at

IE'" [ . . , 11 J~ (1 - I ,) /
-( - I)
'- (y, - y,,,l - (y - y,,) = k ( In.-HI)
n- /11 '

a nd that the cross-product t!!fnlS from (J n.3l) contrihute nothing. Thii> el>tabli~he~
 288 SAMPLING TECHNIQUES

the result for £(SI2). That for £(s/) is found similarly. Hence

E(vCm= 1 - II(SI2+ 1-[2 S22 + 1-13S/)

n m mk

As with two-stage sampling. it is clear from (10.36) that if II negligible

v(y)reduces to

('10.42)

This estimate is conservative if II is not negligible ,

With a cost function of the form
(10.43)
the optimum valucs of k and marc

Therextcnsion of the results in this section to additional stages of sampling should

.. be clear from the structure of the formulas.

10.9 STRATIFlED SAMPLING OF THE UNITS

Subsampling may be combined with any type of sampling of the primary units.
The subsampling itself may employ stratification or systematicsampling. Variance
formulas for these modification~ can be built up from the formulas for the simpler
methods .
Results are given for stratified sampling of the primary units in a two-stage
sample. The primary unit sizes are assumed constant for a given stratum but may
vary from stratum to stratum. This situation occurs when primary units are
stratified by size so that sizes within a stratum become constant or nearly so.
The h th stratum contains Nh primary units, each with MI• . econd-stage units;
the corresponding sample numbers are nh and mh o The estimated population
mean per second-stage unit is

(10.45)
 SUBSAMPLING WITH UNITS OF EQUAL SIZE 289
where Wh = NhMh/'i.NhMh is the relative size of the stratum in terms of second-
stage units and 91t is the ample mean in the stratum. By applying theorem to.1
within each stratum, we have

(10.46)

where flit = nit/ N h, flit = mh/ M h.

From theorem] 0.2, an unbiased sample estimate is

• ) _ '" IV
v ( y" - t... ""It
2[1 -flh
--Slh
2 +[lh(1- [2h)
S 2h
2] (10.47)
h nh nhmh
Corresponding variances for the estimated population total are obtained by
multiplying formulas (10.46) and (l 0.47) by ('i.NhMh )2.

10.10 OPTIMUM ALLOCATION WITH STRATIFIED SAMPLING

Thi deals with the best choice of the nit and the mhoIf travel co ts between units
are not a major factor , the cost may be represented adequately by the formula

C=1: Clhnh +1: c2l,n hmh (10.48)

h h

From (10.46), the variance may be rewritten as

The quantity

where A is 11 Lagrange multiplier, is a function of the variables nh and (nhmh)'

Hence, to minimize V for fixed C. or vice versa, we have

(10.49)

(10.50)

The e give

(10.5])
290 SA MP LING TECH NIQUES

The formul a for optimum mlr is ex actl y the sa me a!- in unstra tifie d sampling
[(10.26 ) in se ' tjon IO.n).
From ( I (JAY ). ~ in ce WI, NIrMIr

N hM IrS,,1r
( I () S2 J
~
Sin ce se lf-wei ghting estim a tes are conve nie nt. we consider unde r wh a t cir-
cumsta nces th e o ptimum a llocation lea ds to a !>e lf-weighting estim ate . From
( 10.45), it foll o ws tha I ji" is se lf-we ighting if III,mf, / N h M h =- i,, =c() n ~ l a nl since . in
this e ve nt.

The conditio n is, as mi ght be expecte d . tha I the proha hil ity f" of sckl:t ing a ' lih ulllt
be the sa me in a ll stra ta.
From ( 10,50), the o ptimum a llocation g i ve~

f l O q)
'f

Fre que ntl y ( ~ Ir ' the cost pe r se cond -stage unit. will he app rox im atdy I he ~ , tlnL'
in la rge a nd sma ll prima ry unit ~: but S :' h may he gre ater in large lln ih th all I I I
sma ll. Howeve r. since the optimum i ~ fl a t. a ,e lf -weighting sa mple wi ll o it ,' l1 h, '
a lmost as preci:-.e as the optimum . Note th il t this re, ult ho ld!> eve n if th t.: np llll' 11m
sampling of prima ry units is fa r from propo rtion a l.

EXERCISES
10. 1 A ~e t o f 20,O()() rec() rd~ a re More d in 4()() fil e dr a we r ~. eac h CO l1la inin g 5() rl!CI l i l k
In JI two-stage ~a mrl e. fi ve records a rc drawn at ra ndo m from each o f X() ra ndoml y , .:Ic.:c lc'd
dra we rs. For on e ite m. the e~tim a t es o f va ri a nce we re S l~ = 3h::! . o5/:c H0 5. a, d efi ned in
sect io n lOA . (a J o mpute th e sta nda rd e rro r o f th e mea n pcr record fm nl tIm ,a mple . (h i
Co mpare thi s with th e sta nd a rd e rro r give n hy th e approx im a tc fo rmul a I 10.2 3) in ' cet'on
10.4 .
10.2 From th e results of a pilot two -stage ~a mpl e. 111 whi ch ", ' \uhunit \ wcre dlo'el1
v)
fro m each of n ' unit s, it is useful to bc abl e to e~ tim a t e the va lue o f V( th " l w(lu ld he g i~ e n
by a subse que nt , a mpl e ha ving m ~ ub u nit s from each of n ul1it " S ho w th a t a n lInhia \ cd
estima te of Vf.Yl is

V'( y_) = (N
---n) ~ Sll (1- -m+-mn- - -mil)
s ,2+ - -
N n mn m' m 'N MN
 SUBSAMPLlNG WITH UNITS OF OUAL SIZE 291
2
where 5,2 and 52 are computed from the preliminary sample. Hint. Use theorem 10. I and
the result (10 .22 1:
, , S, 2 S,2
. ,' ) = S ' ._ .....!-+--=-
E(~
M In
JO.3 In sa mpling wheat fields in Kansas, with the field as a primary unit . King and
McCarty (1941) report the following mean squares for yield in bushels per acre: 5,' = 165.
s/ = 6f1 . Two ~utJsamples were taken per field . For a sample of n fields . compare the
ar ian ce~ of th e ~a mpl e mean as given by (a) the sample as actually taken . (b) four
, ubsamples per field fr.-,m n field~ , (e) co mpletely harvesting n fields .
Nand M ma y be as~ umed large and constant . In (e) assume that complete harvesting is
equiva lent tn single -stage sa mpling (i.e .. to having m = M).
10.4 In the ,arne ,urvey. with two subsa mples per field. the. mean squares for the
percentage of prott:in were S,2 = 7.73. s/ =' 1.43 . How ma ny fields are required to estimate
the mean Yield 1(1 within ± I bushel and the mean protein percentage to within ± 1. apart
frnm a l -in-10 chan ce in each case') Perform the calculations (a) assuming that two
'uh~ample~ [lc:r ficld arc taken in the mCtin survey. (b 1 assuming complete harvesting of a
field in the main survey .
10.5 For the wheat-yield data in exercise 10.3. what is the value of c,lc ] in a linear cost
function if th c estimated op timum m i\ 1?
10,6 If ml {\,f and 11/ N are hath small and the cost function i~ linear. show that n; = 1.
give, a smaller value of V(,vl than m = I if
e l 2,S',2
-> -
(2 S1~
I fl . 7 A large department s tore handle~ about 20,000 account~ receivable per month . A
2% ~arnplc (111 - 4(0) wa~ verified each month over a 2-yea r period (n = 24) . The numbers
of account, found to be in error per month «l ut of 400) were (i n order of magnitude) O. 0 , l.
I. 2.4.4, 5. S. S. 5. 6, 6. 6. 7, 7. R. 9 . 9, 10. In, D . 14. 17. the time pattern heing erratic.
!-rom the results In sect ion 1O ..'i. co mput e s, ~ and s]~. Hence compute the st<lOdard error of
p. as an estimate of the percentage of accounts that are in error over a period of a year. that
would At! obtained from vt!r ifying (a) 1200 accounts from a slIlgle month . chosen at
random, (b) 30n accounts from eac h of four random months, (c) 100 accounts each month .
HiM. Either lI ~e the formula in exercise 10 ,2 with m ' = 400 or obtain unbiased estimate~ of
S, 2 and S/ and usc theorem 10.1,
10,1{ In planning a two-~tage survey it was expected that c,/c l would be about 4 and
that S/ / S/ would lie between 5 and SO. (a) What value of m would you choo 'e from Table
10,2'1 (b) Sup[lose that after the survey was completed it was found that c ,I c! was close to 8
and 5/ I S.. ~ wa~ about 25 Compute the relative preci ion given by your m to that given by '
the optimum m. (c) Make the same' computation for c ,/ c~ = 4. S/ I S} '" 100,
10.9 If p i~ the correlation coefficient between second-stage unit. in the same primary
unit , prove that

(This e tablishes a result used in section 10.6 .)

10. 10 how that if S} > O. in the notation of section 10,6. a simple random sample of n
primary units, with I element chose n per unit , is more precise than a imple random sample
of /I elemen ts (n > I . M > J). Show that the plecision of the two methods is equal if nl N is
negligible , Would you expect this intuitively?
CHAPTER 11

Subsampling with Units of

Unequal Sizes

11.1 INTRODUcnON
In sampling extensive populations, primary units that vary 10 size are encoun -
tered frequently . Moreover, considerations of cost often dictate the use of
multistage sampling, so that the problems discussed in this chapter are of common
occurrence. If the sizes do not vary greatly, one method is to stratify by size of
primary unit, so that the units within a stratum become equal, or nearly so. The
formulas in section 10.9 may then be an adequate approximation . Often, how-
ever, substantial differences in size remain within some strata, and sometimes it is
advisable to base the stratification on other variables. In a review of the British
Social Surveys, which are nationwide samples with districts as primary units, Gray
and Corlett (1950) point out that size was at first included as one of the variables
for stratification but that another factor was found more desirable when the
characteristics of the population became better known.
Some concentrated effort is required in order to obtain a good working
knowledge of multistage sampling when the units vary in size, because the
technique is flexible . The units may be chosen either with equal probabilities or
with probabilities proportional to size or to some estimate of size. Various rule
can be devised to determine the sampling and subsampling fractions , and various
methods of estimation are available. The advantages of the different methods
depend on the nature of the population, on the field costs, and on the supplemen-
tary data that are at our disposal.
The first part of this chapter is devoted to a description of the principal methods
that are in use. We will begin with a population that consists of a single stratum.
The extension to stratified sampling can be made, as in preceding chapters, by
summing the appropriate variance formulas over the strata. For simplicity, we
assume at first that only a single primary unit is chosen, that is, that n = 1. This case
is not so impractical as it might appear at fir t sight, because when there is a large
number of strata we may achieve satisfactory precision in estimation even though
nh = 1. In the monthly surveys taken by the U .S. Census Bureau to estimate
292
 SUBSAMPLING WITH UNITS OF UNEOUAL SIZES 293
numbers of employed people, the primary unit is a county or a group of
neighboring counties. This is a large unit, but it has administrative advantages that
decrease costs. Since counties are far from uniform in their characteristi,cs,
stratification is extended to the point at which only one is selected from each
stratum. Con equently, the theory to be discussed is applicable to a single stratum
in this sampling plan.
As in preceding chapters, the quantities to be estimated rna}' be the population
total Y, the papulation mean (usually the mean per ubunit y), or a ratio of two
variates.
Notatidn. The observation for the jth subunit within the itb unit is denoted by
Ylj. The following symbols refer to the ith unit.

Population Sample

Number of subunits M, m,
Mean per subunit Y, 9,
Total Y,= M,Y, y, =m,9,

The following symbols refer to the whole population or sample.

Population Sample

N
Number of subunit Mo= ~M,
"
~m,
N
"
Total Y = ~ 1'; ~y;

Mean per subunit 9 = Y/Mo Y"' ~ y,/~ m,

Mean per primary unit Y = Y/ N 9 -= ~ yJn

11.2 SAMPLING METIIODS WHEN 11 =1

Suppo e that the ith unit is selected and that it contains M j subunits, of which mj
are sampled at random . We consider three methods of estimating 9', the mean per
ubunit, or second-stage unit. as it is often called.
I. Units Chosen with Equal Probability
Estimate = YI = Yi·
The estimate is the sample mean per subunit. It is biased, for in repeated sampling
from the same unit the average of YI is y" and, since every unit has an equal chance
of being selected, the average of Y, is
1 N _
N i~1 Y, = f" (say)
294 \ A \1I' LI~G T E'H ~ I OUE .

Bu t the popu lation mea l) is

N
2: M,Y,
Y' =;,..
, _;I_ - where M = 2: M,
M , 1

I k nce Ihe bi a~ equ ab (Yo- Y)' Si nce the met hod is bi ase d. we will comput the
mean square error (MS E) about Y. Write
9, - Y = (.v, - Y. )+ ( Y, - Y,,) ( Y" - fJ
Square and take the expectation ove r all possible sample . All contributi ons from
cross- product term va nish. The cx pec t a tion ~ of the squared terms foll ow easily
by the methods given in Chapt er 10. We fi nd

MS EiY d = _!_ 2:N (M ,- m, l 5~+..!..

'
~ (y,- Ya )~ +(y,, + y)2 ( 11.1 )
N , I M, m, N"
where

i!> the va ri ance amo ng s ubunit ~ in the ith unit.

The MSE of 91 contains three compone nts: one arisi ng from variation within
units. onc from va ri atio n betwee n the true mea ns of the un its. and one from the
bia ..
The v a lu e~ of the m, have not bee n !>peci fied . The most common choice is either
1<lt tak e all m, equal or to ta ke m, proportinn al to M:. that is, to \ ubsa mple a fixe d
proporti on of whatever un it is selected. The choice of the f1l, affect only the fi rst
of the three compo nents of the vari ance- the componen t that ari ses from
vari ation within units.
II . Units Chosen with Equal Probability
.,.
. - NM,y,
E stlmate = YII ==--
M il
This estimate is unbiased. Since y, is an unbi ased estimate of Y" the product
MS, is an unbiased estimate of the unit total Y, . Hence NM,y; i an unbiased
estimate of the populatio n total Y. Dividing by M o. lhe total number of subunits in
the popUl ation, we obtain an unbiased estimate of 9.
To tind V(YII) which. of course. equals its MSE. we have
_ ~, NM;'y; ""
YII - T = - -- T
Mo
NM,
=- - (y,
_ - y-)
,+ (NMI
- - y- ,- .,.'
T
)
Mo M..
 U BSAMPLING WITH UNITS OF U NEQU A L SI ZES 295
Now M; V, = Y" the total for the unit, and Y = NV/ M o, where V is the population
mean per unit. This gives
.[', NM;
Y - r = -(Y - Y- )+-N( Y - -
Y)
II Mil " Mo '
Hence

N ~ S2/ N ~ - 2
V(Y II ) = M 2 t... M ,(M , - m, ) - + - 2 t... ( Y, - Y) ( 11.2)
II ,= 1 m, M o , - 1
The be twee n-units compo ne nt of this variance (second te rm o n the right)
re prese nts the variation among the u.nit totals Y,. This ompo ne nt is affected bo th
by va ri ati o ns in the M, from unit to unit by variatio ns in the mean s Y, pe r e le me nt.
If the units vary conside rabl y in size. this compo ne nt is la rge , even tho ugh the
mea ns pe r e le me nt 'F, are almo ·t constant from unit to unit. Freque ntl y this
compone nt is so large tha t YII has a much highe r MSE th an the biased estimate )II '
T hu. ne ither me thod I no r me thod II is full y satisfactory.

III. Units Chosen with ProbabiJity Proportional to Size

stimate = 9111 = Y, =sa mple mea n
Thi ~ technique i. due to Hanse n and Hurwitz ( 1943) . It gives a sample me an
th at is unbiased and i no t subject to the infl atio n of the va ri ance in me thod II .
In re peated sampling. the ith unit a ppears with re la tive frequency M,/ M".
He nce
_ ~ M, - •
E (Y II.) = t... - Y, =Y
,wi Mu
Furthe rmo re,
YIII - Y = (Y III - V,) + (Y, - f')
Ave rage first over samples in which the ith unit is . e lected .

E(YIII - f') 2= ( M 'M-

, I
In,) S2/
m,
+ ( Y, - f') 2

Now average ove r all possible selection of the unit. Since th e ith unit is selected
with re lative frequ ency M;/ M o,
_
V(YIII) = M
1 [ LN (M, - m, )-+
S2,2 - 'fJ 2J
LNM, ( Y{- ( 11.3)
o i- I m, ,- I

Note that , as in method I, the between- unit compo ne nt a rises fro m differe nces
amo'ng the means per subunit Vi in the successive units. If these means per subunit
are nearly equal , thi component is small .
296 SAMPLl~G TE HNtQUES

TABL I 1.1
AR11I1C"1 ,\L POPULATION WI rH UNITS Of UN~OUI\L SIZES
Unn !l lI 1\1 , )' . S2.
2
9, 9, - y

0, I 2 0.500 0.5 -2.25

~ I. 2. 2. 3 4 l< 0.667 2.0 -0.75
3 3. 3. 4. 4, 5. 5 6 24 0.800 4.0 + 1.25

Totals 12 D

Example. Lei u, apply thc\c rc~ult~ to a small popu lati on . artificially constructed. The
d"I,1 ilrc pre ~entcd in Tahle II . I. There arc three unit ~, with 2, 4. and 6 e lements.
Icspertively. The leader may verify the figur<:\ given fsn Y" Sz.', a nd V,. The population
mean ' i~ ;~. or 2 .75 . The unweighted mean () f thc Y, is 2.167 = fl., so that the bias in
method I i~ - O.:'iH 3. It ~ ~quare . thl' contribution t,,1he MSE. i~ () 140.
One unit is to he selected and two suhunits sampled from it. We consider four methods.
two of which arc variant, l1f method I.
Method la .
Selection ; unit with eq ual probability. m, = 2.
Estimate ; 9, (hiased) .
Method lb.
Selecrion ; unit wllh equal probability. m, =~ M, .
E~timate ; 9, (hiased) .
Method II.
Selection ; unit with equal probability. m, = 2 .
'. Estimate : NM,y'! Mil (unbiased) .
Method Ill.
Selection: unit with probabi lit y MJ Mo. m, = 2 .
Estimate : 9, (unbiased) .
Method Ib (proportional subsampling) doe . not guarantee a sample size gf 2 (it may be I ,
2 . or 3), but the average ample size is 2 .
13y application of the sampling error formulas (11 . 1), (11.2), and (11 .3), we obtain the
rc~ults in Tahle 11 .2.

TABLE 11.2
MSE 's OF SAMPLE ESTIMATES OF Y
Contribution to MSE from Total
Method Within Units Between Units Bias MSE

la 0. 144 2.056 0.340 2.540

In O.IIG 2.056 0.340 2.579
II n.256 5.792 0.000 6.048
III O. IH9 1.813 (l.000 2.002

Althoug:h the example is artificial, the results are typical of those found in
l'ompuisons made on many populations. Method III gives the smallest MSE
 SUBSAMPLlNO WITII UNITS OF UNEQUAL SIZES 297
because it has the smallest contribution from variation between units. Method n,
although unbiased, is V~I;'y inferior. Method la (equal size of subsample) is slightly
better than method Ib (proportional subsampling).
Some comparisons of these methods have also been made on actual popula-
tions. For six items (total workers, total agricultural workers, total nonagricultural
workers, estimated separately for males and females) , Hansen and Hurwitz
(1943) found that method III produced large reductions in the contribution from
variation between units as compared with the unbiased method U, and reductions
that averaged 30% as compared with method 1. (They assumed the contribution
from variation within units to be negligible.) In estimating typical farm items for
the state of North Carolina, Jebe (1952) reported reductions in the total variance
of the order of 15% as compared ith methods of type I. In both studie the
primary unit was a county.

11.3 SAMPLING WITH PROBABILITY PROPORTIONAL

• TO ESTIMATED SIZE
As mentioned in Chapter 9, the sizes M/ of the units are sometimes known only
approximately from previous data, and in other surveys several possible measures
of the size of a unit may be available. Let Z; be the probability or relative size
as igned to the ith unit, where the Z; are any set of positive numbers that add to
unity . We still assume n = 1.
Method IV. An unbiased estimate of }' is

(11.4)

This follows because, in repeated sampling, the ith unit appears with relative
frequency Zj, so that

The variance of YIV is obtained in the usual way. Write

[by (11.4)]

=1 [M; _ -
- -(y;-Y;)+ (Mi - - MoT"")]
- Y;
Mo Z; Z;

In the variance , each square receives a weight Z; . Hence

V(Ylv) = ~[ f: M,(M; - m; ) 5 2/ + f Zi(M/Y; _Mot) 2] (11.5)

Mo i- I ZI m; ;- 1 Z;
 298 SAMPLI NG T CHN1QUES

If Zi = M./ Mo, ( 1 L.5) reduces to ( 11.3) (or VcYlII) ' If Zi = 1/ N (initial prob-
abi lities equal), ( 1 1.5) reduces to ( I 1.2) for the variance of the unbiased estim ate
when probabilities are equal.
Unless Zi = M,/ Mo. the betwee n- units compone nt in (11 .5) is affected to some
extent by variations in the sizes M, as well as by variations in the means pe r
element Y,.

TABLE
, 11.3
COMPUTATION OF V(Yrd
M,(M , - III ,) Y, Y,
Uni t M, M ,/Mo z, III, S2,2 Y, Y
,
1J ''',
Z, -,
2 0. 17 0.2 2 0 0.500 I 5 - 28
2 4 0.33 0.4 2 10 0.667 8 20 - 13
3 6 0.50 0.4 2 30 0.800 24 60 +27

Example. Table 11 .3 shows th e computat ions for fi ndin g V(Y,v) in the artificial
pop ulation in Table 11 . 1. The ::, have been take n as 0. 2, 0.4, and 0.4 , and th e nI, = 2.
From (11.5). the varia nce comes o ut (lS f()lJow~ :
'" .. \"' M,(M, - m,)Sz/ M 2 0 ~ .,
wnhln-ufllts contrlhutlon = ~ / 0 = . L 1.,
l ,m,

.,
between-units contribution = L Z,( Z,Y; - y)2( M,/ = 3.583
Comparison with Table I 1.2 shows that me thod IV has a lower varia nce than
the unbiased method I I in which th e primary unit is cho!.en with equal prob-
a bilities. but method IV is decidedly inferior to method I or me thod Ill. In this
exampl e me thod IV pays too high a price in order to obtain a n unbiased estimate .
Con equently. it is natural to consider ~hether the sa mple mean (as in method
I) would be be tte r tha n the estimate adopted in method IV.

V. Unit~ Chosen with Prob bility Proportional To Estimated Size

stima te = Yv ;: y, = sample mean
The estima te is biased since, for example .

E(M = L. ZI Yi = f'z
If the ZI are good estimates, f'z is close to the correct mean f'::: L M; YJ M o and
the bias is small.
If we write
Yv - f' = (y, - Yi) + ( Y; - 9,) + (9'. - f')
 SUBSAMPLING WITH UNITS OF UNEQUAL SIZES 299
the three components of the MSE work out as f.ollows.

MSE(yv) = ~ zl(MI-ml) S2/ + ~ z,(9;- 9',)2+(9z - fJ2 (11.6)

I- I M, m, 1- 1
Example. If the values of z, and m, are chosen as in Table 11.3 , the reader may verify
that the conlponents of the variance of Yv are as shown in Table I 1.4 .

TABL 11.4
CONTRIBUTIONS TO THE MSE IN METHOD v
Within Between Bias Total
Units Units MSE

0. 173 1.800 0.062 2.035

This i superior to all method s except method In (pps ) and is almost as good as
method Ill .

11.4 SUMMARY OF METHODS FOR n =1

The five methods of estimatrllg the mean per element Y and their MS 's in the
numerical example arc summarized in Table 11 .5.
For estimating the population total Y = Mo Y, the preceding estimates .are
multiplied by Mn and their MSE's in Table 11.5 by M I/ = 144. Their relative
performances remain the same . •

TABLE J 1.5 ,
TWO'_Sl AGE SAMPLING METHOlJS (n = I)
Probabilities in Estimate Bias MSE
Method Selecting Units of }' Status in Example

Equal ii, Biased la : 2.541

Ib: 2.579
NM,fj,
II Equal Unbiased 6.048
M
MI
III - 0:. size ii, Unbiased 2.002
Mo
M lii i
IV z, 0:. estimated size Unbiased 3.796
:iMU
V :, a: estimated size y, Biased 2.035
300 SAMPLING TECHNIQUES

11.S SAMPLING METiiODS WHEN n > 1

For n > 1 the principal sampling methods are natural extensions both of the
preceding methods I to IV and of the methods discussed in Chapter 9A for
one-stage sampling with cluster units of l'nequal sizes. The quantities most
commonly estimated are population totals, population means per subunit, and
means or proportions that have the structure of ratio estimates.
In many application the quantity Mo , the total number of subunits in the
populatiun, is not known , but only the MI for !he sample primary units that are
drawn, since these Mi can be counted by listing. It is worth noting, therefore, that
methods II and IV and their extensions to n > 1 do not require knowledge of Mo
for estimation of the population total. Method I and its extensions do not require
knowledge of Mo for estimating the mean per subunit. Method III , probability
proportional to size, requires knowledge of Mo .
Estimates are called self-weighting when they are simply a multiple of the total
over all subunits in the sample. In view of the practical convenience of self-
weighting estimates in large-scale surveys, the condition under which each
estimate becomes self-weighting will be noted. For unbiased estimates the
condition is, as will be seen, that every second-stage unit in the population or,
more generally, every unit of the final stage of sampling, has an equal chance of
being drawn .

11.6 TWO USEFUL RESULTS

, In finding variances and sample estimates, two generaJ results are useful. They
sbow how to extend one-stage sampling results to two-stage or more generally
multistage sampling. They were first proved by Durbin (1953) for estimators that
are sums of estimators for the sample primary units, extended by Des Raj (1966)
to linear functions of the primary unit estimators, and extended to still more
complex cases by Rao (1975b). We will follow the approach used by Des Raj .
Primary units are chosen without replacement, with equaJ or unequal prob-
abilities. In the ith primary unit, let Y, be an unbiased estimator of the unit total
Y; , with second-stage variance 0"2/_ Subsampling is independent in different
primary units . Consider an unbiased estimator of the population total Y of the
form
(11. 7)

where the weight WIJ is known for every sample s, and may depend on other
primary units that are in the sample as well as on unit i.
We will adopt the device used in section 2.9 of letting Wi.,' be a random variable
that equals WI. if unit i appears in the sample and equals zero otherwise. This gives

(11.8)
 SUBSAMPUNG WITH UNITS OF UNEQUAL SIZES 301
Now
(11.9)

jf and only jf E I (Wi.') = 1 for every i.

'Theorem 11.1.

Ven = v(.t,. 1Wi.r Y;) = v(i,- 1Wj,J Y;) +I,- IE 1(w;.,,2)oo2/ (11.10)

Proof. By formula (1 O.2~,

V(n = V 1[E 2(n]+E 1[V2(n]

= v(,t WbY;)+EI[ f Wb,2V 2(Y;)]

,- I i- I
(11.11)

since the second-stage covariance between Y; and fj (i # j) is zero, because

subsampling is independent. Hence

V(n = v(JI Wi. Yi) + J/E t (W;;2)oo2/] (11.12)

Thi completes the proof.

The first term in (11.12) is the variance of the corresponding estimator when
every primary sample unit is completely enumerated, and it is obtained from
results in single-stage sampling,
"Example. For the two-stage analogue of the Horvitz- Thompson estimator, YfIT =
L Y./ 'TT" the weight w,,' = 1/ 'TTl if unit i is in the sample and zero otherwise. Hence,
E,( W,,'2) = 'TTJ'TT/ = 1/ 'TTh where 'TT, is the probability that primary unit i is drawn. Further-
more, if m, subunit are drawn from the M, by simple random sampling whenever unit i is
drawn,
2_
U 2, -
V ( y' ) _ M, (M, -
'2 J -
mJ S 21 2 (11.13)
m,
Hence, by theorem J 1.1, using formula (9A.42) for the single-stage analogue of V( YHT),
we have

(11.14)

Theorem 11.2. Suppose that we have an unbiased estimate U2i2 of the

second-stage variance 002/ of Yt. and an unbiased sample e timate of Va: Wi. Y;) =
N
V(r Wb'Y;) from one-stage sampling. The latter will be a quadratic of the form

(11.15)
302 SAMPLING TF.CIINIOUES

Then an unbiased ample estimator of V(I W;s Y,) is

U " . )+L" Y.1t.fU2'~

(L" w,.,Y, )= ("La,sY/+2I"Lb;jsY,Y/ (11. 16)
I I J -'

Thus the rule for constructing the sample estimator of V(f Wi' Y,) is: In an
unbiased estimator of V([ W; s Y,) from one-stage sampling, insert Y, wherever Y,
appears. To this add the te rm I (W"U1/). where f w,., Y, = Y. and rT2/ is an
unbiased e's timator of V 2 ( ) ' ,) .
Proof.

( I I. J 7)

Introduce the rantlom variabie. a where a,..' = 0" if unit i i)' in the sample
L, ' .

and a,..' = 0 otherwise. Similarly, let b :/., = b,l' if units i and j are in the sample and
b;il = 0 otherwise. From ( II . J 5) for o ne-stage sampling.
N ) N N N
U( LW,s'Y, =L O,, ' Y,1+ 2L Lb;I'Y''Y; ( I I. I H)
I I J .,

If this is to be unbi ased , comparison of ( 11 .1 R) and (11 . 17) shows th at we must

have E I (a,,') = V( W" 'I . Now for our variance es tima tor ( J 1. 16)
'r

EIEz(I
,
a,: f'/+2I .Ib;jj Yf~) +~I 2(I W,..'U2,,1 )
I I '"

( 11.1 9)

In ( 11.19) we have used the results that £(a;..') = V( w,,') and that for ;] ' i,
EI(w,,')= I = E I 1 (w,,'). Continuing.

E[ u(f w,., Y)] = v(f y,) + f, E Wi" 1( W;.: 2)U2/ = v(£, w" Y,) ( 11 .20)

by (I 1. 12). Thi)' comple tes th e proof.

We consider now some speci fic estimators of the population total Y. In
extensive surveys, selection of primary units 'with unequal probabilities has
become the most commonly used technique. Sections 11.9 (sampling with
replacement) and 11.10 ( ampling with ut replacement) deal with these methods;
section II . 13 prese ntc; the ratio estimator in ppz sa mpling. Other sections (11.7
and 11 .8) give the corresponding methods when primary units are chosen with
equal probabilities.
 SUBSAMPLINO WTTH UNITS OF UNEQUAL SIZES 303
11.7 UNITS SELECTED WITH EQUAL PROBABILmES:
UNBIASED ESTIMATOR
Except where noted otherwise, the mi ample subuni ts in the ith unit are chosen
by simple random sampling. The unbiased estimator of the population total is

(11.21)

To apply theorem 11 .1, note that

Wi.
N
=; ; N !:!
E(wIS ') == !!_ n = 1 '.

Hence, by theorem 11 .1 and earlier results ,

2
V('Yu) = N (1 - I,)I (Y, - y)2 +!:! f M,20 - /-t,)S2/ (11.22)
n (N- J) n m,

where 12. == m,/ M, . The estimator becomes self-weighting if h is con. tant, (= 12,
say). We then have
,. N "I,...
nf 2 II y"
Y" =- 01.23)
i ,

The qua ntity nhl N is, of course, the probability that any second-stage unit is
drawn .
For an unbiased sample estimate of variance , theorem 11.2 gives, from (11.16),

(l 1.24)

11.8 UNITS SELECfED WITH EQUAL PROBABILmES:

RATIO TO SIZE ESTIMATE
This e timator of the population total Y is

( I 1.25) .

This is a typical ratio estimator. since both numerator and denominator vary
from sample to sample. It is used mainly for estimating mean per subunit, for
which knowledge of Mo is not required, and for which it is an extension of me thod
I with n ::;:: 1.
304 SAMPLING TECHNIQUES

To find its approximate MSE, write

( 11.26)

where Y = Y/ Mo.
~ " ~
Since Yu = (N/ n) L M,Yh we can obtain the approximate MSE( YR) from for-
mula (11 .22) for V( YJ by substituting M,(y; - Y) for M,y, or, more generally,
(Ylj - Y) for Ylj' Now (11 .22) is

N
In the substitution, 1"; =M, Y; becomes M, (Y; - Y) while Y =L y,/ N is
replaced by O. Also S2;2 = L (YiJ - Y,)
- 2
/(M; - 1) remains unchanged when (Y'J - ~
C J
j
replaces Y;j . This gives the result
2 N 2 - ~2 2 2
MSE( YR)='= ~(1 - II)LM, (1"; - r J +!:!.. r.. M, (1 - 12, )S2, (11.27)
n N- 1 n '";

As with Yu , this estimator becomes self-weighting if

"

121='"' rfi = -Nrfi

- = constant = 12 = -=
MI M Mo

In this event the within-units contribution may be expressed more simply by

putting ml = NrfiM,/ Mo. giving
N

(1 - 1 )L M/ (Y, - Y) 2+ Mo2(1 - 12) ~

2
MSE(:Y; ) ='= N (M;)s .2
R n I (N _ 1) nm t... Mo 2,

(11.28)
The resemblance to the corresponding formula when primary units are of equal
size~ may be noted. From (10.14) section 10.3, multiplying by Mo 2 since Y = MoY,
we have

(11.29)

The difference is that in (11 .28) the contributions to the MSE from the primary
units are weighted , larger units receiving greater weight.
 SUBSAMPLINO WITH UNIT OF UNEQUAL SIZES 305
By theorem 11.2 an approximate sample estimate of MSE(YR ) in (11.27) is
given by

(11. 30)

..." "
where 9'R =I M;MI ~ :;: YR/M o·
When this estimator is used to estimate the po pulation mean per subunit, we
have fR =I Y,;£M; and Vd'R) =V(YH )/M o2 . When Mo is not known , we
substitute M o = N(IM'/ n) for Mo in calculating v cfrR)'
When primary units are selected with equal probabiliriel>, an alternative esti-
mate of the population mean per subun it (another extension of method I) is

This estimate is self-weighting if m; = con tant, as in the following example. When

m, and 'P, are uncorrelated, this estimate may be satisfactory , but it is liable to a
bias that does not vanish even with n large.

Example. From the volume American Mer. of Science. 20 pages were selected at
random . On each page the ages of two scienti t , from two biographies also selected at
random, were recorded. The total number of biographies per page varies in general from
about 14 to 21 . Estimate the average age and its standard error from the data in Table 11 .6,
using the ratio estimate.
F.rom the extreme right column,
f = L.M1YI= 17,121.5 47.7 years
R L. MI 359
Since 11/ N is negligible, we have from (lJ . 30), dividing by Mo".

(f )== f M~(Yi - fR) 2

V R nM2 (n -1)
The numerator is computed as

r. (M/y/? - 2~ )M/ + -f-R 2 L. M/

R '[.( }l.f,y,

== 15,375,020 - (95 .3844)(309,747.5)+(2274.55)(6481) == 571,300

Stnce M. == 359/2() 15 estimated from the ample , this gives

v
(f R
) = (20)(571,300) = 4.67
(19)(3 W
ScrK) = 2.16 years
 306 SAMPLING TECHNIQUES

TABLE 11.6
AOES OF 40 S CIENTISTS IN American Men of Science (n - 20, m - 2)
Ages
Unit Total
No. M; Yn Y;. Yi Miiii

1 15 47 30 77 S77.S
2 19 38 SI 89 84S.S
3 19 43 45 88 836.0
4 16 S5 41 96 768.0
S 16 S9 4S 104 832.0
6 19 39 38 77 731.5
7 18 43 43 86 774.0
8 18 49 SI 100 900.0
9 18 4S 3S 80 720.0
10 18 46 S9 lOS 94S.0
11 20 71 64 13S 1,350.0
12 18 35 46 81 729.0
13 19 61 S4 115 1,092.5
14 19 4S 87 132 1,254.0
IS 18 31 38 69 621.0
16 16 64 39 103 824.0
17 16 63 47 110 880.0
18 19 36 33 69 655.S
" 19 19 61 39 100 950.0
20 19 54 34 88 836.0

Totals 3S9 1,904 17,121.S

11.9 UNITS SELECI'ED WI11I UNEQUAL PROBABILITIES WI11I

REPLACEMENT: UNBIASED ESTIMATOR
Primary units are selected with probabilities proportional to ZI with replace-
ment. Results for Zj =Md Mo (probability proportional to size) follow as a special
case.
The subsampJe of mj subunits from the ith unjt is assumed to be randomly
drawn without replacement. If the ith unit is selected more than once, we suppose
that on each selection the whole subsarnple is replaced, a new independent
drawing of ml units being made without replacement from the complete unit.
An unbiased estimate of the population total is

(11.31)
 SUBSAMPLING WITH UNITS OF UNEQUAL SIZES 307
For n =1, it was shown in section 11.3 that this estimator, MoYIV= YIv, is
unbiased. Its variance is obtained from formula (11.5) on mUltiplying by Mo 2 as

V( Yrv) = f ZI( YI - y)2 + f MI(MI - mj )S2? (11.32)

1- 1 Zj I_I Zlml

With this method of sampling, the estimator Yppz is the mean of n independent
estimates of the form Yrv. Consequently, from classical sampling theory 'Yppz is
unbiased and

V('Yppz ) =.! V(Y )=.! f ZI(l'i_ Y)\.!

IV
f M/(I-f21)S2/ (11.33)
n n I- I Zi n 1-1 mlZ,

Furthermore, given n independent estimates ¥IV = l'il ZI, an unbiased sample

estimator of V(YIV ) is, of course,

"(9; ./'. )2
I --r.
V
(¥ )_' - 1 Z,
IV - (n-l)
ppz
(11.34)

Consequently, an unbiased sample estimator of V(¥ppz) is the very simple

expression

v(Y.
ppz
f(Y'-y.
)= 1-1 Z,
n(n -1)
r
ppz
(11.35)

These results hold also for multistage sampling, provided that 9; is an unbiased
estimator of l'i and that subsampling is independent whenever a primary unit is
drawn.
To discover when ¥ppz becomes self-weighting in two-stage sampling, write

(11.36)

The condition is therefore

nZIm,
- - =constant =fo (11.37)
~
This expression is the probability that any specific second~stage unit is drawn.
From (11.37), m;/MI = fol nZI. If the probability fo is cbosen in advance, the field
worker can be told in advance what second-stage fraction mtf~ to take from any
primary unit that has been chosen. For instance, suppose fo= 1/50=0.02 and
n =60 primary units have been chosen If %1 ::< 0.0026 for one selected unit, we
must have m;/~ =0.02/(60)(0.0026), or 1 in 7.8.
 308 SAMPLING TECHNIQUES

With a self-weighting sample , the variance estimat Jr (11.35) takes the simpler
form

(l1.38)

where Yi = L Yij is the ample total in the ith unit. The simplicity of the estimated
j
variances (11 .35) and (l1.3R) is an attractive feature of with-replacement
sampling.
With this method there are other ways in which the subsamples may be drawn.
If the ith unit is selected I, times, one variant is to draw a single subsample of size
lim, without replacement provided that Mi > m/I, . Sukhatme (1954) has shown
~ N
that with this method , V( Y ppz ) is reduced by (n - ]) L Mi52,21 n. Another method
is to draw a single subsample of size m, no matter how many times the ith unit is
selected. The estimate M,y./ z, from this unit receives a eight I, (the number of
times that the unit has been drawn) with either method . The effect i to increase
VCYppz ) by
(n -)) f M/(l - 12.)5 2/

n mi
For the same co t, the differences in precision among these methods are seldom
likely to be substantial.
If z, =Mil Mo the unbiased estimate (J 1.31) reduces to
.,
A Mo" _
Ypp.. =-Ly, (11.39)
n i

Clearly this estimate becomes self-weighting when m, = m, so that YPI" = Mop·

An unbiased sample estimator is
Mo2 "(_ ~)2
v(Ypp.t) - n(n - 1)~ Yi- Mo
A _
(11.40)

11.10 UNITS SELECTED WITHOUT REPLACEMENT

For any of the " without-replacement" methods studied in Chapter 9A, the
formulas for the variance and estimated variance in two-stage sampling are
obtained from theorems 11.1. 11.2. With the Brewer or Durbin methods.

(11.41)

with variance
Vrf'B) = ff
I j >i
('IT/?I'j - 'lTlj)( Yi_l1) + f M?O - h.i )S2l
'lT1 'IT]
2

I ml?l'j
l
(11.42)
 SUBSAMPLING WITH UN1TS OF UNEOUAL SIZES 309
For the corresponding estimator Ypp , in sampling with replacement, we have,
from (1) .33), with 17'; = nz"

V(Yppz) = 2. f z/(Y; - Y)\ f M/(l- f2i)S2j2 (11.43)

n l, ml17'j

Since the "within-units" contributions to the variance are the same in (11. 33)
and (1) .43), any relative gain in precision from sele<..'ting units without replace-
ment is watered down in two-stage samples by the within-units contribution to the
variance. For instance. in a stratified multistage sample of n = 441?rovinces opt of
N= 147 provinces, Des Raj (1964) found that the ratio V{YWD.l~)/V(YWR)
averaged about 0.79 over seven items for the between-units component, but the
average ratio over both components was 0.92.
With n = 2 an unbiased sample estimate of V(Ys ) is

V ( .c.iJ )_(
I - 17' ] 17'217'12
-]- l)(M]YI
-
M2Y2
_., _ -
~ Mj2( 1 - f2j)s~:2
) 2+ L.. (11. 4 4)
17'1 17'2 ,_ I tnl7T,

where suffixes 1. 2 denote the chosen units .

The extension to the Sampford estimator for n > 2 is straightforward . The
alternative RHe estimator

Y~ RHC --~L.. ZgMSY8 (11.45)

8 Zg

where Zg = I z/ over the gth group and Mg. Y,l' and z, refer to the ,!nit drawn from
the group. For a sample of n units an unbia~ed estimator of V( YRHC ) is

V(YRHcJ= '-r
CN/- N) fz,l( - 2 MgY LYRH
)2 +fzgM_g2(1-f 2,)S2B2
( '" . '- N g 18 z~ S Z8 m~
(11.46)
Formulas (l J .44) alJd (11.46) rt..q uire separate calculation of the between-units
and the withjn-units contributions to the estimated variance. In extensive surveys
with many strata and numerous items, the complexity of such formulas makes the
estimation of variances u task requrring more computer time than can be devoted
to it in practice. recall the much simpler form of the variance formula (11.35) when
units are drawn with replacement; [hat is,

1 "
V(YPP')= n{n _nL
(9;;-Y" )2
ppz

where y;" =Mjy,/ ZI. This result make with-replacement sampling appealing if it
doe not involve too much loss of precision.
Among without-replacement methods, Platek and Singh (1972), in planning
the redesign of the Canadian Labor Force Survey for strata in which n~ == 6 or a
312 AMPUNG TECHNIQlJES

This follows by the method used in Theorem 2.5. For

A Y XA A A A A

YR - Y=X - Y=x(Y-RX)== (Y-RX)= D (11.49)

x
where d ij = Yij - RXij' Since £(0) = 0 with any sampling plan for which Y and f<
are unbiased , we find E( YR - Y)2 by taking the formula for V( fJ for thi plan and
substituting d ij = YiJ - RXij in place of Y'j'
For example, consider Yu = (N/n) L M1y, = (N/n) L Yi in sampling with equal
... " " A

probabilities. From formula (11.22) in section 11.7,

2
V( Y ) = N (1 - fl)L (Yi - Yi +!!.. ~ M /(l- f ld S2/ (11.50)
u
n N- ] n m,
Hence, with equal-probability sampling and YR = xt/ XU,
V( YR ) == N
2
(l-fl)f (Y, - RXI) 2+!!.. ~ M /( l - h)S~u (I 1.51)
n N- l n mj

where

The conditions under which YR reduces to a multiple of the ratio of the sample
totals 'i..'i..Yij/I.I.X// are always those under which the corresponding Y becomes
self-weighting; in this case hi = mJ M, = constant.
For the estimated variance, substitution of d j / = Y// - &// for Y;j in (11.24) for
v(YJ gives
2
f
v( Y ) == N (1-fl) L (Y;- RX/)2 +!!.. M / (l- f2j)S~'2j
R (11.52)
n n- 1 n mj

Similarly, with ppz selection with replacement we had for

1n
Yppz =- LMt9Jzj
n
(from (11 .33) in section 11.9)

V(Yppz ) =.!. ~ z,(lL y)2 +..!. ~ M/(1-/2,)S2/ (11.53)

n Z, n z,m/

Hence, for YR = xY/f< in sampling with replacement,

n
f z,
V(YR ) ==..!. .!.(y; - RXj)2+.!.
n
f M/(l-z,m,/2/)S~2' (11.54)
 UBSAMPLING WITH UNITS OF U"!EQUAl. SIZES 313
From ( 11 .35) a sample estimate v( YR ) that is slightly biased is

v(Y ) =
R
. 1
n(n - I)
t(Y.-z,RX,)2 (11.55)

With Urewer"~ mdhod flf ~ampl in g without replacement. YRB = XYo/ XB •

when! £
'f·B :;:. M;'j;/ 71", . For n "" 2, formula!. (11.42) and (11.44) gi\ie . for the ratio
estimate,

(11.56)

where D, = Y, - RX, and units j and j arc the two selected.

(
,( yO
R8
) : ,_
\11,71", 71"'1
l_l)(!2~_!Z;_)2+ ~_ M,2(1-h)s~ l.: ( 11.57)
71"j 71", " m 71",

11.13 CHOICE OF SAMPLING AND SliD. AMPUNG

FRACfIONS. EQUAL PROS \ RIUT'ES
hi!. prohlem l~ ..ii ,cussed fif~{ tor t he ratio- tn-size ~ . malt. when units arc ch ·:1~cn
with lequal proh<thlir H_!>. The subsampling fraction 12= /,1.1 M j is assumed con-
'>tanl. ~('I tP ;·, !he ('strmate i~ the sample mean per subun .. .
The ~lCnrk,t ,;ost function c(1ntain:- three term!> .

Cu = n. ed co'! per pnmary uni t

! = co't per whumt
(, = co'>t of listing per subunit in a ,e lected unit

The third term is included becau~e the sampler mu.st u<;lll1lJy Ii!" the elements in
,illY se lected unit and verify their numbel in order :,) draw a SI IU\ampie. Henc~
" n
cost = cun + C2 L m, + c, L ,'vI,
This formula is not usable as it stancis, since the cost ri ;:pends ('In (he parti .. ular et
of units that is chosen . Instead, consider the average co!.t over n unit!>, which
equals

where (I nolY I"d u les the average cost of listing ~ unil.

We determllll.' fI <.IIH..! ni = {,M so a. to minimil l' \'f VI for "ell Cf)SI \ r "k'
verSQ. From ( I J.2 ) in section 11.8, dividing by Mf)~ = (M, :')2,
314 SAMPUNG TECHNIQUES

Write
N 2 - -6-\ 2
5 2 _ LM, (Y,-SJ
b - M 2(N - l)

This is a weighted variance among unit means rer element. It is analogous to the
variance 5/ in section 10.3 and reduces to 5, if all Ms are equal. We may also
write

52 2_~
-i., -
Mi 5 212
Mo
This is a weighted mean of the within-unit variances. It reduces to the 5 2 2 of
section 10.3 if all M, are equal.
t2
In this notation, since = ril/M,
2
MSE(Y) 1 ('5b 2 --=-
.. =- 52 ) 1 52 2 - -
+-_ 1 Sb 2 (11.59)
n\ M nm N
Applying the Cauchy-Schwarz inequality as usual to 01.58) and (11 .59) we get

- . 52 ~ (1 60)
op1
m = JS/ - 5//MV~ 1.
The methods given in section 10.8 for utilizing knowledge about the ratios S2/ Sb
and C lie. to guide the selection of mop' are applicable here. The unbiased estimate
when units are drawn with equal probabilities can be handled similarly.
The next section presents a more general analysis of this problem.

11.14 OmMUM SELECflON PROBABlLmES AND

SAMPLING AND SUBSAMPLING RATES
An important early analysis by Hansen and Hurwitz (1949) determines at the
same time the optimum Z, as functions of the Ms and the optimum sampling and
subsampling fractions. Selection of units is with replacement. The analysis is
presented for YR in the self-weighting form, so that m, =foM,/nz, =toMb" ,
As in Section 11.13, the cost function is
" Ms
" m, + c, L
C = c"n + C2 L
Although the Ms are known, a listing cost is included for those units that are in the
sample, since listing may be needed to provide a frame for subsampling.
The average cost of sampling n units is
N
E(C)=c"n+CJoMo+c, L'TrIMs (11.61)
 SUBSAMPUNG WJ'Jlf UNITS OF UNEQUAL SIZES 315
By (11.54) in section 11.12,

V(9"R) =.!.[f
n
!..CY, - RX;)2+ M,(Mt - m,) S~21]
Z, Z,ml
(11.62)

Since d,j= Y'j - Rx,j> we may write (1', - RX, ) = MtD,. Noting that 17', = nz"
M,jnz,m, = l/fo, and combining the first and third terms in (11.62), we get

(11.63)

The problem is to choose n. fo. and the 17', = nz, to minimize V subject to fixed
average cost and to the restriction
N
LZ, = 1,
Take A and p. as Lagrangian multipliers and minimize

V+A[ elln +cdoMo+ C, f 1T,M,- B(C)] +p.( n - f 17'1) (11 .64)

Differentiation gives
n: (l1.65)

(11.66)

Relations (11.65) and (11.66) lead to

Z, ;" : ' OCM,(D? - 2~'r/'l(c . +c,M,)1/2 (l1.67)

Since the individuaJ values of (D? '- S~2,j~) will not be known, we consider
how the average vaJue may depend on the size of unit Mt. using the following
rough argument. Suppose a population were divided into units of size M Since
B(D,) = 0, formula (9.10) on p. 241 gives
-2
B(D,)~ V(D1)
-
=s/
M[l +(M-l)PM]

where sl is the population variance of the d jj and PM is the intraunit correlation

for units of size M. Also, from (9.15), p. 242,
E(S2<121 )=S'/{1- PM)
hence

-2 SJ2I). Sl[
E ( DI _~2
- - = - l+(M - ' I)PM-(I-PM))=PllfUd
M, M
316 SAMPLING TECHNIQUES

From (11.67) thi gives as an approximation

Mj.r;;:;.
z, ex: 1"''''==== (11.6R)
-Jc,,+c,Mj
With p positive. PM, may be expected to decrease as M, increa es, since subunits
far apart are less subject to common influences. but this decrease may be only
~Iight for JP:;. Deductions from (11.68) are as follows.

1. If the cost of Ii ting c,M1 is unimportant, Z, ex: M j (i.e., pps selection) is best if
..;p;; changes little over the range of sizes in the population . If {ii;, decreases
noticeably, optimum probabilities li es between ZI M, and Zj ex:.JM,.
2. If listing cost predominates, Zj should lie between JM, and constant (equal
probabilitie~) .
3. If listing costs and fixed costs are of the same order of magnitude. : , '" ,'\,1,
may be a good compromise.
Differentiation of (I J .64) with respect to the overall sampling fraction f () gives

(I 1.69)

The value' of ;. .~ f(lund in term~ of the known 7T'S by adding (11.66) over all units.
This step leflds to the result

(11.70)

Comparison with 01.60) will show thatfo has the same structure as in sampling
with equal probabilities, remembering that'in (11.60), fr. = ntflopl/ Mo and C I =
c,,+c,M
The opwnum n is found from the average co t equation (11.01).

11.15 STRA TIFfED SAMPLING. UNBIASED ESTIMATORS

For the unbia ed methods (Y", Y",m YR , etc.) the extension to stratified sam-
pling is straightforward. The subscript h denotes the stratum.
The estirna!ed populat illn total is Y., = L YIo , with

VCY,,) = t
h
V( Y,,). (1 1.71)

These variances are obtainable from the formulas already given.

 SUBSAMPLING WITH UNITS OF UNEQUAL SIZES 317
We may note the conditions under which the estimates become self-weighting.
For Yppz ( ection 11.9) or VB (section 11.10),

(11. 72)

where y,,/ is the total over the m", subunits from the ith unit in stratum h. These
estimates are seen to be self-weighting within strata if the probability fo;, =
n"z,,/mhJ M h, of selecting any subunit in the stratum is constant within the stratum.
In this event, the estimate becomes
A L l n.
Y.,=L,LYh, (11.73)
"JO" I

which is completely self-weighting if fo;, is the same in all strata, as might be

anticipated intuitively.
If units are of the same size within a given stratum (i.e., M,,/ = M,,), it was shown
(section 10.10) that sample allocation leading to a completely elf-weighting
estimate is close to optimum, provided that S2"/JZ;;. is reasonably constant.
Whel_! the M h , vary within strata, a corresponding result for estimators like Yppz
and YB is as follows. Suppose that the cost function is linear as in ections 11.13
and 11.14, and that the estimator is self-weighting within strata (i.e., tO h=
nhz"/m,,.! M h,). Then the optimum fOh for given expected cost may be shown to be
of the form

(11.74)

where MOh = I.Mhi • Thus. choice of tOil == fo, which makes these estimators com-
pletely self-weighting, will be near-optimal as regards precision unless either the
silll or the C2h vary widely from stratum to stratum. especially since this choice
affects only the second-stage contribution to the variance.

11.16 STRATIFIED SAMPLING. RAno ESTIMATES

The formulas for the separate estimate YR. follow from those in section . j .12
fOla single stratum, assuming a large ample in each stratum and independent
sampling in different strata.
For the combined estimate, YRc = XY,,/ X."

Hence, we substitute dldl =Yhij - Rx"11 for Yhij to obtain the approximate formulru
318 SAMPLING TECHNIQUES

for V('YRc) from those for VCY,, ) by the sampling plan used. FOT V('9'RC) we
substitute diuj = Y'd, - ReX"lj in v( '9'.. ).
For example, with unequal probabilities with replacement (section 11 .9),
formula (11.33) leads to

(11.75)

where
D"I = YII; - RXIo;
M.. - X
2
Sd2101= (M"I-I)
1
7«Yhlj - RX"lj)-(YIo1 - R "I)]
2

For the estimated variance, formula (11.35) gives, after the substitution,

(11.76)

where
"..
I:DIoI '
D"I ,= Mhld"I', 6",=_1__,
Z" I nIt
"

11.17 NONLINEAR ESTIMATORS IN COMPLEX SURVEYS

In addition to the estimation of totals, means, ratios and differences among
them, analyses of survey data may involve estimates of more complex mathemati-
cal structure (e,g. simple and partial correlation coefficients, medians, and other
percentiles). The objectives of the analyses may include jittaching a confidence
interval to the quantity estimated or performing a test of significance.
Given a random sample from an infinite popuJation, statistical theory has
produced a variety of methods for meeting these obj~ctives-exact small-sample
methods based on assumptions about the nature of the distribution followed by
the observations, and approximate large-Sflmple methods reql.Jirlng fewcr
assumptions, With the more complex types of sample stpdied i~ this ~k, on the
other band, we have been able to give methods for computing unbiased estimators
of the variances of unbiased linear estimators like Y,,, Y", .. or f'rtlT' Assuming the
sample large enough so that these estima tors have distri butions close to normality,
the normal tables supply confidence limits and tests of significance. There remain
many problems with nonlinear estimators in complex surveys.
 SUBSAMPLINO WITH UNITS OF UNEQUAL SIZES 319
Three approximate methods for estimating the standard errors of nonlinear
estimators have been produced. They will be described for stratified random
samples with nlo = 2 in all strata, the case for which they have received the most
study. All methods give the usual unbiased variance estimates when the estimator
is linear. The sample can be a multistage sample, primary units being drawn either
with equal probabilities or with unequal probabilities with replacement.

11.18 TAYLOR SERIES EXPANSION

This is the method that produced the approximate formulas for the estimated
variances of R, and R." R, in stratified sampling. The function to be estimated,
I( Y" Y 2 , ..• , Y,,) = I{y) is expressed as a function of the population totals of
certain variables, the estimate I(Y> being the corresponding function of unbiased
sample estimates of Y. For variable j with nh = 2. when units are chosen with ppz
with replacement, the population total and its linear estimator are
A L2}J1o L
lJ=LL
10
'=L(yjhl+yjd
2Z,ti 10
i
(11.77)

where Yilo,:::: }J1ti/2Z hi and Yjh; is an unbiased sample estimate of Yjlti from the
subsample in this unit.
From (11.35) an unbiased estimate of V(}J) works out as
2
v(}J) = LL 2(yjlti- ji,,)2;!!i L (yj" 1- yj"2)2 (11. 78)
" I h

Hence, b~ Woodruff's (1971) extension in section (6.13) of the Keyfitz .short-cut

method for nil = 2, a Taylor series approximation to the variance of ICY) is

v[/CY)) =
iiI) (Yilo 1- Yilo~
. t [Ie~ (ay; I ,] 2.
(11.79)

Expression of a nonlinear function of the measured variable in tho form I(Y)

may require some work and care and is not possible with SOOle functions of
interest. Consider a simple example-simple random single-stage sampling with
primary units of equal sizes, where the correlation between the unit totals of two
variables VI and V 2 is to be estimated. The population value I(y) is
320 SAMPUNG TECHNIQlIES

In terms of our variables Yjll" this is a function of five .such variables.

11.19 BALANCED REPEATED REPLICATIONS

We consider this method first for a linear function of a single variable. Let
I( Y) = Y,with its estimate ICY) = Y =r(Y1I1' + YII2'), where YIII' = Y1II/2zlIi· An
unbiased variance estimate v( Y) is, frod; (11. 78),
(11.80)

Select a half-sample H by choosing one unit from eacb stratum. The estimate of
r
I(y) = Y from this half-sample is 1(11) = 2 YhI' , where hi is the unit chosen in
II .
stratum h. Hence, if I(S) denotes the estimate of I( Y) made from the whole
sample,
1(1I)-/(S) = 2 LYhI' - L (YIII ' + Y1I2') = L±(Yhl'- Y1I 2') (11.81)
II II II

the signs depending on whether unit 1 or unit 2 was selected in stratum h.

Comparison of (11.81) and (11.80) shows that for any of the 2L possible
selection, (f(H) - /(S )Y contains the correct squared terms (YIII' - YII 2,)2 in v( Y)
in every stratum. However, for any specific balf-sam ple every pair of strata
contributes a cross-product term with a ± sign to [/(H) _/(S)]2. McCarthy (1966)
noted that a set of balanced half-samples can be found in which, for every pair of
strata, half the cross-product signs in the set are + and half - . If this set contains g
different half-samples, the unwanted cross-product terms cancel when we sum
over the set giving
11. L A

- L (f(H,)- / (s)f - L (YII1'- Yh /)2 = v(Y) (11.82)

gill
If Ci is tbe complementary half-sample to H" second and tbird forms are
v(Y)e! t (f(Ci) - /(S)]2_-41 t [/(H,) - I(Ci)f
g i g i
(11.83)

A fourth estimator is the average of the first two above.

Balanced sets of this type were given earlier by Plackett and Burman (1 946) for
use in experimental design for g any multiple of 4. For L strata the smallest
balanced set has g equal to the smallest multiple of 4 that is > L. With L = 5 strata,
g = 8. Table 11 .7 shows a balanced set for L = 5, + denoting unit 1, - denoting
unit 2 in a stratum.
 SUBSAMPLING WITH morTS OF UNEOUAL SIZES 321
TABLE 11.7
BALANCED HALF- SAMPLES FOR L =5 STRATA
HALF- SAMPLE
Stratum 2 3 4 5 6 7 8

1 + + + +
2 + + + +
3 + + + +
4 + + + +
5 + + + +

When this balanced repeated replication (BRR ) method is applied to a non-

linear function fey ) of several variables. the four estimators

and the average of the first two, all differ to orne extent. Also, I (S) = feY) is
biased, and the variance estimates do not include the correct bias contribution to
MSE[f(Y)]. However. note that if we drew m independent samples Sj, each ', ith a
complementary H;, C. the quantity

(11.R4)

would be an unbiased estimate of V(f(H)]. Thus. roughly speaking, the BRR

method amount to assuming, for I(S), that (a) its bias is negligible, (b)
v(/(S )] == (1/2) vr/(H», and (c) calculation of v[f(S)] from BRR for a single
stratified S agrees well enough with its calculation from independent samples Si to
be useful.
The repeated replication method has been extensively used both for sample
selection and variance estimation by the U.S. Census Bureau in the 19505 and by
Deming (1956, J960). The idea goes back to the Mahalanobi technique of
interpenetrating sub amples (Mahala nobis, 1944). McCarthy (1969) reviews the
properties of the BRR method, with further di cussion by Kish and Frankel
(1974).

11.20 THE JACKNIFE METHOD

For R in simple random samples this method was de cribed in section 6.17. It
was propo ed a a means of estimating V(Ro), but could also be tried for
e timating VCR) . Omit unit j from the sample and calculate Ri , the ratio estimate
from the rest of the sample. Ignoring the fpc, an approximate estimator of V(R)
can be obtained from (6.86) with g = n. If we a ume R=R_, the mean of the Ri ,
322 SAMPLING TElI-l!\,J 'tI<:S

this es~im .. to, becomes

v(R ):-- (n - 1)
n
f (R
j
~.:.._:.!2.(
- - .8.)2 or:
i n )
'(5
I
) - 1'(S)]2 (11.85)

the expre 'sion on th extreme right illdicati h w the method might be applied to
other nonhnear estimation.
For a linear estimaror like j it is easil ' ,,('rified th t formula (11.85) reduces to
the usual v(y) = L ~yj - 9)2/ n 1/1 - 1).
In extending thi method to stratified sampl,. .. with nh =-1, Frankel (1971)
suggests omitting one unit at random from stratum II (or h ;:, 1,2, .. . L in turn ,
calculatinfl; f(Sn) from the remaining !)C'!mpie uf size l2l - 1). One (orm of the
lacknife e<;timate of V[f( S)) is then
1..
v[l(5») == L [f(5h ) - f(5)Y (11.86)
h

For a Iin<;;ar estimator f i Y), thiS variance e':tirnatc;r reduce, to the usual unbiased
estimator in ppz sampling with repla~ment.
As with BRR, there are four analogolls versiOn!> of the laeknife estimator of
v[/(5»).

U .2 l COMP ARlSON OF TIlE mREE APPROACHES

The Taylor. HRR, and J methods have had very limited rigorous theoretical
justification. Appraisal of their performance with different estimators and types of
surveys has relt'd thus far primarily on Monte Carlo studies. One study is
describetl by Frankel (1971) and Kish and Frankel (1974) ,
The sample was a ingle- tage sample of c1ust~r units of slightly differing sizes, a
unit having on the average J 4. I households. The porml ' tion of N = 3240 un its or
45,737 households was dividt:d in turn into (1, 12, and 30 equa l ~sized strata, with
"h = 2 in each stratum. Simple random sampling witrun strata was used. The data
were taken from the Cuo ent Population. u:-v 1,' (U .S. Census Bureau). The eight
original variables for a Guseh ld were the number of reT ons, number under 18,
number in labor force, inc,ome, Rge, sex, and years of Sc oli ng of bousehold bead,
and total income. Various means (ratios), differences of means, and simple, partial
and multiple regression or correlation cot>fficients (three independent variables)
were computed. In numbers of primary units the samrles were quite small, 12,24,
and 60, being about 170, 340, and 850 in numbers of households.
In computing variance ec;timators, all {our forms of the BRR and the J methods
were compared. The Taylor method was not used for partial and multiple
regressions, ,because of difficulty in expressing the partial derivative af;a~ in
manageable form. Since the sampling method was proportional stratification
without replacement, the fpc (1- f) was included in all variance estimates, with
f = n/ N, although scarcely needed with these small samples.
Except for the m ultiple correlation coefficients, the estimators f(Y) had rela~
tively unimportant biases for all three sample sizes, the ratios /Bias/ls.e. being
 SUBSAMPLING WITH UNIts OF UNEQUAL SIZES 323
<0.1 for ratios, differences of ratios, and simple regression coefficients, and
around 0.2 for simple and partial correlation coefficients. When the approximate
v(i) are regarded as estimators of MSE(i), all three methods-Taylor, BRR, and
J-performed well for ratios and differences of ratios, the average IBias in
v(V)I!MSE(Y) being under 5%. For simple correlation coefficients, the BRR
variance estimators had substantially smaller biases than tbe Tor J estimators.
The reverse was true for simple regression coefficients With both BRR and J the
best of the four v (Y) methods was the average of the H aDd C estimates (i.e. with
BRR)

ig{~ (f(H,) - /(S )]2+ ~ [fCG) - /(S)Y} (11.87)

In confidence interval tatement:; and tests of si~ificance, the important issue is

how well the tail probabilities of the variate [f(y)-/(y)]/s.e.(f(Y)] agree with
those of Student'S t with L d./. For selected t-values Frankel (1971) gives the
Monte Carlo estimates of the tail frequencies for the related variate (f(Y)-
EfcY)]/s.e.[f(Y)]. Som~ results for 6 and 30 strata are shown in 1 able 11 .8forthe
t-values that Frankel presents nearest to the 5 and 10% two-tailed levels. Version
(11.87> of the BRR and J methods is the one shown.

TABLE 11 .8
A VERAGE TAlL PROBABIUTIES OF (f cY)-
Ef (y»)/s.e.(f (Y))
CoMPARED WITH THOSE OF STU()ENTS't
6 strata
P(/) "" .042. 1='2.576 P(/) = .098, 1= 1. 960
BRR ) Taylor BRR J Taylor

Ra!iO'l .044 .049 052 .096 .106 .112

b'. .034 .048 .05 .085 .11 7 .127
,'s· .052 .069 .084 .114 .137 .163
Partial,'s .043 .063 .092 .132
Multiple R's .065 .088 .105 .160

30 str8ta
P(t ) ... .059, t '" 1.960 P(t) = .110, 1=1.645
BR R J Taylor BRR J Taylor

Ratios .056 .057 .057 .109 .111 .112

b's
, 's
..062
. )89
.01\7
.frln
.068
.102
.110
.138
.116
.153
.116
.164
Panial ,'s .103 .\ 1 .1 6 .181
Multiple R 's .175 .207 .265 .297

• b = simple regression coefficient.

•,= simple oorreiation coefficient.
 324 SAMPLING TECHNIQUES

To summarize, BRR consistently performs best in Table 11.8. Except for

multiple R's, it can be regarded as adequate for practical use if one has the view
that in data analysis a tabular 5% tail value represent an actual tail value
somewhere between 3 and 8 %. J does slightly better than Taylor. Except for BRR
with ratios and simple regressions, all method give actual tail frequencies higher
than the I-tables, so that confidence probabilities are overstated. A puzzling
feature is that for correlation coefficients the increase in ampJe size from 12 to 60
has not brought a corresponding improvement in the clo eness of the actual to the
t tail frequencies .
This study opens up a wide area for investigation of the methods with different
survey plans and different types of estimator I(y) .
In a Monte Carlo study of a larger, more complex sample (two-stage pps
sampling with replacement, including both stratification and poststratification)
Bean (1975) compared the Taylor and BRR methods for estimators of the ratio
type. Both methods gave satisfactory variance estimates and adequate two-sided
confidence probabilities calculated from the normal distribution . Sufficient skew-
ness remained, however, so that one-sided confidence intervals could not be
trusted.

EXERCISES
11.1 By working out the estimates for all possible samples tbat can be drawn from the
artificial population in Table 11.1, by methods la, lb, II, and III, verify tbe total MSE's
~iven in Table 11.2.
11 .2 For methods II (eQ'!al probabilities, unbiased estimate) and III{ pps selection),
recompute the variance of f' for the example in Table 11 .1 when m, = 1. Show that the
=
precision of method 1lI in relation to method II is lower for m, 1 than for m, 2. What=
general result does this illustrate?
11 .3 For the population in Table 11 .1, if the estimated izes Z, are 0. 1, 0.3 and 0.6, with
m, "" 2, show that the unbiased estimate (method IV) gives a smaller variance than pps
sampling. What is the explanation of this result?
11.4 The elements in a population with three primary units are classified into two
classes. The unit sizes M, and the proportions P, of elements that belong to the tirst class are
as follows.
MI = 100, Ml = 200, M 3= 300, PI = 0 .40, Pz = 0.45, P3 =0.35
For a sample consisting of 50 elements from one primary unit, compare the MSE's of
methods la, II, and III (or estimatlhg the proportion of elements in the first class in the
population. (In the variance formulas in 5ection 11.2, s,z is approximately P,O,.)
11 .5 A sample of n primary units is selected with equal probabilities. From each chose.n
unit, a constant fraction 12 of the subunits is taken. U a, out of the m, subunits in the ith unit
fall in class C, show that the rati<rto-size estimate (section 11.8) of the population
proportion in class C is p = Ia,/Im,. From formula (1l.36);..ihow that an estimate of
MSE(;;) i.

where P, =a,1 mi.

 UBSAMPUNG WITH UNITS OF UNEQUAL SIZES 325
11.6 A firm witb 36 factories decides to check the condition of some equipment of
which Mo= 25,012 pieces are in use. A random sample of 12 factories is taken, a 10%
subsample being checked in each selected factory. The numbers of pieces checked (mj ) and
the numbers found with signs of deterioration (aJ are as follows .

at at
Factory mi aj Pj""- Factory m, aj Pt=-
mj mt

65 8 0.123 7 85 18 0.212
2 82 21 0.256 8 73 )1 0.15)
3 52 4 0.077 9 50 7 0.140
4 91 12 0.132 10 76 9 0.118
5 62 I 0.016 II 64 20 0.312
6 69 3 0.043 12 50 2 0.040

Estimate the percentage and the total number of defective pieces in use and give
estimates of their standard errors.
Note. Since M,/ M == mil m, the between-units component of v(fi ) may be computed as
1- /
-2( 1 1)U:: a/- 2,; r a,m, +pl r m/ )
nm n -
and, since the m, are fairly large, the within-units componellt as
11(1- / 2 ) r
(nm) 2 a,q,

11 .7 If primary units are selected with equal probabilities and /2is constant, show that
in the notation of exercise 11.5 the unbiased estimate of a population proportion is
p = NIall nMo/2 and that, if terms in 1I ml are negligible, its variance may be computed as
_ 1- / ) " 2 /) (1- / 2)"
v(P )- n (n - 1)-
m 2r (a.- a)+ (nm-)2 ra,ql
Calculate p and its standard error for the data in exercise 11.6.
11.8 A sample of n primary units i chosen with probabilities proportional to estimated
sizes z, (with replacement) and with a constant expected over-all sampling fraction /0. Show.
that the unbiased and the ratio-to-size estimates of the population total are, respectively,
Tl lo and ™./r" m , where T is the ample total. (It follows that if Mo is not known the
l
unbiased estimate can be used, but not the ratio to size. For e timating tile population mean
per ubunit, the situation is rever ed.)
11.9 In a study of overcrowding in a large city one stratum contained 100 blocks of
which 10 were chosen with probabilities proportional to estimated size (with replacement) .
An expected over-all ampljng fraction I. ;: 2% was used. Estimate the total number of
persons and the average persoh per room and their s.e .'s from the data below.
Block 1 2 3 4 5 6 7 8 9 10
Rooms 60 52 58 56 62 51 72 48 71 58
Persons 115 80 82 93 105 109 130 93 109 95
11 .10 For Durbin's method (section 11.10) f simplif)'ing variance estimation in ppz
sampling without replacement, a simple method of sample selection due essentially to Kish
326 SAMPLING TECHNIQUES

(1965), is as follows. The subscript h to denote the stratum will be omitted and the numl:>er
of primary units is assumed to be even.
Arrange the units in order of increasing z, and mark them off in pail's. The method is
exact only if z, = z, for members of the same pair; this will be assumed here. Select two units
ppz with replacement. If two different units are drawn, accept both. If the same unit is drawn
twice, let the sample consist of the two members of the pair to which this unit belongs. Show
that for this method: (a) '71", = 2z" (b) for units not in the same pair, '71"" = 2z,z, = '71",'71"/2, so
that '71",'71",'71",,-1 - 1 = 1, and (c) for units in the same pair, '71"/1 := 4z,z, = '71",'71"" so that
.."..,'71",'71",,- 1-1 = O.
11.11 In section 11 .9, formula (11.33) for V( Y"...) in sampling with replacement was
proved under the plan that whenever the ith unit was selected, an independent simple
random subsample of size m, was drawn from the whole of the unit. Prove the following
results for two alternative plans.
(a) When the ith unit is selected ti times, a simple random subsample of size m,t, is
drawn from it (assume m,t,sM,). Under this plan, V(Y"...) in (11.41) is reduced by
N
(n -1) L MiSZ/ I/I (Sukhatme, J 954).
(b) When the ith unit is selected t, times, a simple random subsample of size m, is
drawn . Then V(Y"..,) in (J 1.41) is increased by
(n -1) N 2 2
--LM,
n
(1-/2')S21 1m,

N
In both (a) and (b), Ypp% = L f,M,Y/ nz" the ith unit receiving weight t,.
CHAPTER12

Double Sampling

12.1 DESCRIPTION OF TIlE TECHNIQUE

A we have seen, a number of sampling techniques depend on the possession of

advance information about an auxiliary variate X" Ratio and regression estimates
require a knowledge of the population mean X. If it is desired to stratify the
population according to the values of the Xi, their frequency distribution must be
known.
When such information is lacking, it is sometimes relatively cheap to take a
large preliminary sample in which Xi alone is measured. The purpose of this
sample is to furnish a good estimate of X or of the frequency distribution of Xi' In a
survey whose function is to make estimates for some other variate Yi, it may pay to
devote part of the resources to this preliminary sample, although this means that
the size of the sample in the main survey on Yi must be decreased. This technique is
known as double sampling or two-phase sampling. As the discussion implies, the
technique is profitable only if the gain in precision from ratio or regression
estimates or stratification more than offsets the loss in precision due to the
reduction in the size of the main sample.
Double sampling may be appropriate when the information about Xi is on file
cards that have not been tabulated. For instance, in surveys of the German civilian
population in 1945, the sample from any town was usually drawn from rationing
registration lists. In addition to geographic stratification within the town, for
which data were usually already available, stratification by sex and age was
proposed. Since the sample had to be drawn in :1 hurry and since the lists were in
constant use, tabulation of the complete age and sex distribution was not feasible.
A moderately large systematic ample could, however, be drawn quickly. Each
person drawn was classified into tne appropriate age-sex class. From these data
the much smaller list of persons to be interviewed was selected.

12.2 DOUBLE SAMPLING FOR STRA1mCAll0N

The theory was first given by Neyman (193 ). The population is to be stratified
into L classes (strata). The first sample i a simple random sample of size n '.
327
 328 SAMPLING TECHNIOUES

Let
Wh = N h / N = proportion of popula,tion falling in stratum It
W" ;: n hi n' ;: proportion of fir t sample falling in stratum h
Then W" i an unbiased estimate of Who
The second sample is a stratified random sample of size n in which the Jill are
mea ured: nh units are drawn from stratum h. Usually the second sample in
stratum It is a random subsample from the nlr' in the stratum . The objective of the
first sample is to estimate the strata weights ; that of the second sample is to
estimate the strata means Vh .
The population mean V == L Wh Vh . As an estimate we use
L
y..,=hL- l w"y" (12 .1)

The problem is to choose n' and the nh to minimize V(ji.,) for given cost.
We must then verify whether the minimum variance is smaller than can be
attained by a single simple random ample in which y, alone is measured. In
presenting the theory, we assume that the n h are a random subsample of the nh '.
Thus, nit = l1"nh' , where 0 < 111t :S 1 and the I1h are chosen in advance . Repeated
sampling implies a fresh drawing of both the first and the second samples, so that
the Wh , nil, and jill are all random variables. The problem is therefore one of
stratification in which the trata sizes are not known exactly (section SA.2).
' r Two approximations will be made for simpliCIty. The first sample size n' is
assumed large enough so that every Wh > O. Second, when we come to discuss
optimum strategy, every optimum I1h as found by the formula is assumed :S 1.
Theorem 12.1. The estimate ji., is unbiased .
Proof. Average first over samples in which the Wh are fixed. Since jih is the
mean of a simple random sample from the stratum, E(jih) = Vir' Furthermore,
when we average over different selections of the first sample, E(w/t) ::; Wh , since
the first sample is itself a simple random sample. Hence
(12 .2)
Theorem 12.2. If Ute first sample is random and of size n', the second sample
i a random subsample of the first , of size nil = "hnil', where 0 < "h :S 1, and the IIh
are fixed ,
2

N + t -n-'-
1 1 ) L WhSh 1 )
V(y.,) == s:1 -,- (
I1h - 1 (12.3)

where S 2 is the ;>opulation variance.

Proof The t>roo( is easily obtained by the following device . Suppa e that the
Yhi were measu red on all n,,' first-sample units in stratum h, not just on the random
 DOUBLE SAMPUNG 329
subsample of nIl. Then, since WI> = n,,'/n',

is the mean of a simple random sample of ize ,,' from the popula,t ion . Hence,
averaging over repeated selections of the sample of size .n',

(12.4)
But
L L L
YII = L w"y" = L W"YIo'+L WIr(Y~ -YII') (12.5)
" " "
Let the subscript 2 refer to an average over aU random subsamples of nil units that
can be drawn from a given n~' units. Oearly, E 2(YII) = YII'. Results that follow
immediately are (see exercise 2.16):
cov (YII', (Y~ -y,,')] = 0:
(12.6)
cov (YII', y,,) = V(YII ') : V(Yh - Yit ') = V(Yh) - V(Yh')
Hence, for fixed w",
V 2 [1: WII(YII - YII ')] =]=1: W,,2Sh2(_!__~)
nIl nIl
=L W,,~,,2
n
(_!_-1)
JI"
(12.7)

since nil = JIlin" ' = "11 w"n' .

Averaging over the di tribution of the w" obtained by repeated selections of the
first sample, we have, from (12.4), (12.5) and (12.7),

2(
V(Y., )= S -;-- L W"S/ ( 1
1 1) +1:-,- --1 ) (12.8)
n N " n ""
Corollary L The result can be expressed in a number of different forms. By
the analysis of variance,
(N-l)S2=1:(Nh-l)S/+1:N~( i\-Y)2 (12.9)
Hence, if g' =(N-n ')/(N-l), multiplying by g'/n'N gives
l)S/+!:1: W,,(YI. - Y)2
(12.10)
From (12.3) this gives
L WIr S,,2( 1 ) g' L I g' - 2
V(y.,) =1:-,- --1 +-;1: (WII-N"' )S", '1--; 1:L W,,(Y,,-y)
-

" n v" n " n "

(12.11)
3 2 SAMPLING TEOINIQUES

where cp = S2/ SK,2 is the relative efficIency of proportional stratification to simple

random sampling, Thus, if we guessed that stratification would reduce V(y) in half
.so that cp = 2, then IIh := (C' / C) I/ 2

Emmplt. This example did not arise from a double sampling problem, but illustrates
orne feature of the solution . We use the Jefferson data from p . 168. In estimating corn
acres, assume that we could either take a simple random sample of farms or devote some
resources to classifying farms into two strata by farm size. Relevant population data are for
corn acres.

Strata W. •
2
5. y.
1 0 ,786 312 17.7 19.404
2 0 ,2 14 922 30.4 51 .626

Population 620 '!6.300

Suppose thaI C* = 100, c. = I = c. and that S ] / .''' is negligible. This implie that if double
sampling is not used, we can afford to take a sample of " '" 100 fal ms, giving V (ji) 6.20. =
Let c' be the cost per farm of classifying farms into stratum 1 ( :S 160 acres) and stratum 2
( > 160 acre) . onsider the questions :
1. For what values of r'/ e does double sampling bring an increase in precision ?
2. What is the opumum double sampling plan if c' = c/ I 00, and what is th e resulting
V ( ji., )?
" 3. In problem 2, bow do the plan and the value of V (ji,, ) change if the IIh are guessed as
twice the optimum fracti o ns'!

1. from the population data, I. W.S. = 20.4 . ( 2_ I. W.5. 2 ) '" 177 , Hence from (1 2.22)
V",," (ji,,) = 0 .01 (20.4+ 13 .3Jc')2
If this is to be less than 6.20 lor simple random sampling, c' < 0, 11 ; that is, c'/ c < 1/ 9.
2. If c'/c :: 1/ 100, the n with the optimum plan, (1 2.22) gives
V",,"(y,, ) = 0,01 (20.4 + I.W = 4.71
Note th at if classificati on by farm size cost nothing, we would have
V", .. (ji,,) =0,01 (20.4)2 = 4. 16
As regards the details o[ the plan with c'/ c = 1/100, from (12.21 ), we get
II, >: 0.133 ; V 2= 0.229
Since I.W"II" = 0 ,]535, we find , from (12 .18), that n' = 612 . In return thi give
expec,ted vaJues of 64, 30 for "It
n 2 • Thus nearly all the money is spent on
measurement: only 6% on classification.
3. If we guess II, = 0.266, 112 == 0.458. then I. W.II. = 0.307 and from (12. J 8), "' = 315,
=
leading to an expected" I 66, n2 == 31. From (12.3), V(j.,) for this plan will be found
to be 4.85 , only a 3% increase over the optimum 4.71 in problem 2.
 DOUBLE SAMPLING 333
12.4 ESTIMA TED VARIANCE IN DOUBLE SAMPLING FOR
STRA11FICAnON
If lin ' and 1/."1 are both negligible with respect to 1 (e.g., <0.02), an almost
unbiased sample e timate of V(y.,) in (12.14) is imply the sample copy of this
formula.
(12.24)

(12.24')

where g' = (N - n')/(N -1). This formula will suffice for almost all applications.
With I,' Nand 1/ n' not negligible a m re complex algebraicexpr ~ion is needed .

Theorem 12.3. An unbiased sample estimate of V(y., ) in doubl.: sampling is

_ n'(N - l)[ l( 1 1)
g' l(W~ g' _ _ ~]
v(y.,) = ( , 1)N LW"S" - ' - - N +~LSh N---;-- +~LW,,(y,,-ys.)
1)
n - " n "h n " n "It n It
(12 .25)
Proof. From (11.13) the general form of the variance LO be estimated i

By averaging first for fixed n' and WIt and then over, ariations in the WI.. the
lIverage of ~·"s" 2 in ( 12.25) is WI,S" 2 • while that of Sit 2 is Sit ~. These result~ will he
used after equation (12.31).
In the last term in (12.25)
~ ~
L., Wh (- -)2
Yit - Y., = L., - 2
Wh)'" -
-2
y" (12.27)
A veraging first for fixed W"'

~
Ev...
_ 2
WltYh ) =
~
L., WII
- ~ ~ ~( l
Yh + L., wS" - - ' - -
N
1) (12.2 )
I'hw"n Wh

Furthermore,

(12 2 ,

Also.
(12.. (I)
334 SAMPLING TECHNI lJE~

Subtracting (12.30) from (I'"' 29) and multiplying by g'l n' gives,

!:E L Wh(Yh - Ys, )2= !:[L Wh (Y,, - Y)2+L S/t2C:n'-~) - V(Y6')]

(12.31)
Substitute (12 .3J) in finding (n' · 1 )NEv(Y.,)/n'(N- l) from (12 .25). We get

(n' - l)N 'C _ ( g') _

(n' - l)N _
n'tN -l)L~o(y,,)= 1- ;;; V (Y6/)= n '(N _ l)V(Ysl)

This proves the r suit . Notl! that the two middle terms in (12.25) are of order
l/n'N and l / n ,JJlh and are negligible relatIve to terms retained if liN and lin'
are negligible. Thi supports the <;impler form (12.24).
Rao (197.1) has given the result 0 2.2 5) in terms of the "h
and nit' as follows .

,- ,=1:"'- J "
V\,t " , N 'h
(n,, ' ·· 1_~C.!)W"s,.z + (N - n ') "w (- _ - )2
n'- 1 N- ] "h N tn' - I); h Yh y"
(12.32)

Corollary. To use (12 .24) in the estimation of a proportion, put Ph for Yh and
nlcP"q;, !(II , 1) fO! 5,/ .
"
&amplt , In II ~i mr>le random sample of 374 households from a large district, 292 were
occupied hy white ramilies and 82 by nonwhite families . A sample of about one in four
households gftve th following data on ownership.

Owned Rented Total

-----------------------------------
Whil e 31 43 74
N."lnwhite 4 14 ]8

E.sumate the prop,mion of rented households in the area rrom which the sample was drawr
an its standard error.
If the first stratum consists of the white-occupied households.

292
W, == 374 == 0 .78,

43
PI = 74 == 0.58,

PAl =' W ,p, + WlP l '"' 0.624

n' ==374,
 DOUBLE SAMPLING 335
It is readily found that only the first term in (12.24) is of importance. Hence
v(p ) =L ~ w,.p.q~ = L w~ 2Phqh = (0.78)'(0.2436) + (0.22)2(0.1716)
" h (nh -1) IIhll ' h (n h - 1) 73 17
= 0.00252
s(P,,) = 0 .050
The eSlimated proportion of rented households is 0.62±O.050.

12.5 DOUBLE SAMPLING FOR ANAL YDCAL

COMPARISONS
Section 5A.13 dealt with the determination of sample sizes in subgroups of the
population for certain analytical comparisons among L subgroup means. If the
oi.jective is to have the va.iances of the differences between the means of every .
pair of subgroups all equal to V, it was noted that a simple compromise allocation,
often adequate. is to specify or minimize only the average variance of the
L (L - 1)/2 differences.

(12.33)

In another application, Booth and Sedransk (l969), the sUbgroups formed a

2 x 2 fac.lorial classification, the problem being to estimate row and column effects
with equal precision. In this case the variance to be specified or minimized took
the form
1 2 2 S 2
- rL«(h 2 +8,/).:::!L = v (12.34)
2 I / n'j
where the 8;. and 8./ are weights chosen by the investigator, the symbols (ij)
denoting the 4 subgroups.
In the discussion in section 5A.l3, subgroup could be sampled directly. The
problem becomes one of double samplin~ when members of the subgroups cannot
be identified in advance but units can be classified relatively cheaply into the
subgroups from an initial simple random sample of size n·. If n/ of the initial
"I
sample fall in subgroup i, the for tbe final sample that is measured to provide the
desired comparisons are ,drawn from the nj·. The simplest cost function is
c = ,'n'+c L n, = c'n'+cn (12.35)
where we assume c'« c.
For fixed nl. both relations (12.33) and (12 .34) are of the form

(12.36)

where we assume that the S/, SI/ and hence the al2 are known in advance.
336 SAMPLING TECHNIQUES

When the objective is to minimize V for given C, Sedran k (1965) and Booth
and Sedransk ( J 969) try different values of "' in turn . For given n " the value of " is
"j
then known from (12.35). For given n, the that minimize Vare = nat/CI aj). "j
Difficulty may arise in usinr these "i, however, becau e the "I' provided by the
initial sample are random variables. In ome subgroups we may find "j' <
natiCI ai) so that the minimizing nl exceeds the available n;'. Allocation rules to
handle this situation will be illustrated for L = 3 groups, from Sedransk (1965).
Number the subgroups in increasing order of n/CI a,)/nal, and let CI a,) denote
the sum over all classes except the first . Take I

if

, nal
if nl < CIal )

, (n - "l )a2
if n2 2:: CI al) (12.37)

if

where L a, =
I
r al - a l'
These rules are not .::omplete, but will cover most n' likely to be near optimal
The principle is to keep clo e to the nl ex: a, allocation.. See Booth and Sedran 'k
(1969 ) for more detail.
Ex,,,,,,le_ An example in which double sampliftg should perform welJ is a follows :
c= 2,000, c· = 1, C= 10. The three subgroups are of relative size WI == 0.05 , W2 = 0.25,
W, = 0.70, a,2== S,2 = 10 (i = 1,2, 3). Consider first single sampling. Since it costs 11
monetary units to select, classify, and measure a sampling unit, we can afford" = 182 with
single sampling.
Optimum allocation would require equal "" but on the average the values of from "I
singJesampling with n == 182 are 9. J, 45 .5, 127.4. A sumingE(fln;)= 1I E (n, ), the average
V from single sampling is approximately

E(V) = 10 ( - 1+- 1 + --1 ) == 1.40

9.1 45 .5 127.4
With double sampling, calculations for different n' show that the optimum n' is close to
620. giving n = 13 . With n = 138 the optimal n, is 46 in each class. However, n' "" 620
pTovide ~ only an expected 31 in class 1. Hence, on the average, the allocation rule (12.37)
 DOUBLE SAMPLING 337
lUI double sampling gives n I"" £(n l ') = (0 OS)(620) = 31, " 2 = n ) = 53.5. leading to

E( V):!: I o(..!_+...2_)
31
....,11.70
53 .~

a 50% reduction from E( V) with single sampling.

Rao (I973) handles these problems by the method used with stratified sam-
pling: this method specifies the fraction v, of the n;' in the ith subgroup that are to
be measured . An advantage is that the optimum n' can be determined analytically.
With C = c' n' + en as before, the expected co:,! is •

where W; is the relative size of ubgroup i. Assuming as before that

E(l/ni)=l/E(ni), the average Vis
L 2
E( V) = r -{!i-
,11Willi
(12.39)

By the Cauchy-Schwarz inequality, the optimal "I for fixed n ' is given by
n'\+:", ==a:a i } (C*- n'c') (12.40)
Q, e

provided all ", 5. 1, upstirution of ,,'W,JI; from (12.40) into (12.39) gives. for the
minimal E( VI ,

E(V) ::;: cCL aY (12.41)

C*-n'c'
Since H( V) in (12 ,1 1)i.ietTCases as n' decreases, the optimum n ' will deere se at
least un til it reliches the value ml, !oay, at which one of the II,. ay hecomes 1. At "I.
this point (12.40) and (12.41) no longer appl '. From (1 2.401. "I becomes 1 when

(12.42)

Solving (12.42) for mI . the value of n' at which I I"" I, we get

C*
t1l1 = ' (12.43)
re'+cWla: o,.lIaJ3
To examine values of It' smaller than "'1. set "1 =0 1 and use the Cauchy-
Schwarz inequ:uity to 0 tnm the remaining "I' W" nnd, for i> 1.

n'\\'11 .: ~ C* ~C'+C ~ I) (12.44)

i I <La,)
338 SAMPLING TECHNIQUES

where L 0, = L 0, - OJ . The resulting minimal B( V) is now

I

a1
2 ca:: 0,)2
I
(12.45)
£(V):: n'WI + C* - n'(c' + cWI)
From this E(V) the derivative dE(v)ldn' vanishes at

a:: a)
n' = C* I{ (C'+CWI)+~[CWl(C'+CWI)]I/2} (i2.46)

The value m2 at which III and 112 are both 1, so that (12.45) cea es to hold, is

cW a:: a/)]
2

m2= C* I[Cc'+ c W 1)+ a; (12.47)

Thus expressions (12.45) and (12.46) apply only over the range mt 2: n' ~ m 2 ' If
d£(v)ldn ' does not vanish for n'~ m2, we need to set Ill> 112, and so forth , = 1 in
tum until the turning point of E( V) is found . In many situations for which double
sampling is economical, however, the turning value of ECV) occurs for m I ~ n' 2:
m 2·
ExIIrnple. For the worked example in this section, C* = 2000, c' = 1, c = 10, a/ = S/ =
10, W. =0.05, 0.25,0.70. We have, from (12.43) and (12 .47) .
'f
2000
,
m = [1 +(10)(0.05)(3) =800
2000
m 2 = [1.5 + (10)(.25)2) - 308

Furthermore. dE(V)/dn ' in formula (12.46) vanishes at

, 2000 '
n = {U'+(2)[(0.5)(1.5))' /2} = 619

Since this value lies between 800 and 308, it gives the required minimum. Formula (12.44)
gives 112 = .346, " 3 = .124. Numerically this solution is essentially the same as that found by
~~ransk:'s method, which had n ' "" 620 and similar values of '" in the three subgroups.
Both methods extend easily to the ca e of differential costs of measurement in
different subgroups. Suggestions are also given (Rao, 1973) for the case where
EO I nj) is substantially larger than II E(n,) .

12.6 REGRESSION ESTIMATORS

In some applications of double sampling the auxiliary variate x/ has been used
to make a regression estimate of Y. In the first (large) sample of size n', we
 DOUBLE SAMPUNG 339
measure only XI; in the second, a random subsample of size n = rm' :: n'/ k where
the fraction JI is chosen in advance, we measure both XI and YI' The estimate of Y is

y" = y+b(i' -i) (12.48)

where i', i are the means of the XI in the first and second samples a~d b is the least
squares regression coefficient of YI on X" computed from the second sample.
If no assumption is made about the presence of a linear regression in the
population, y" will be biased, just as in one-stage sampling (Chapter 7). An
approximation to V(y,,) can be given, assuming random sampling and lin and
1/n' negligible with respect to 1.

V( -,,) . S/O - p2) +p S/ _~

2
(12.49)
Y n n' N
Proof. In finding the sampling error of YI, in simple random sampling, we
showed that if b in YI, is replaced by the finite population regression coefficient
B = 5",1 S.2, the error in the approximation is of order 1/.J;, relative to that in j".
The same device applies here. We therefore examine the variance of the approxi-
mation
y" =.y + B(i' - i)
The subscripts 1, 2 will denote variations over the first and second phases of
sampling. Let U j = Yj - BXj. In the second phase, regard the large sample as a finite
population. Then, since the small sample was drawn at random from the large
sample,

where S",2 is the variance of u within the large sample. It follows that

V(y,,) =V(y,,) = Vl(j')+Et(~- ~,)S . '2 (12.50)

(1 1)
n' N)'
(1 1)
= - - - S 2 + --- S 2 (l-p Z)
n n' Y
(12.51)

(12.49)

This completes the proof.

As in section 7.8, suppose we assume, following Royall (1970), that the finite
population is itself a random sample from an infinite superpopulation in which a
linear regression model holds. Then y". becomes model-unbiased and exact
340 SAMPLING TECHNIQUES

small-sample results for its variance can be obtained. Let the regression model in
the superpopulation be
Y =0: + f3X+E (12 .5_2)
where, for given x's, the E'S are independent with means 0 and variance tT/(l-
p2), where u y and pare n_ow parameters of the superpopulation.
On substituting for y, Y, and b from (12.52), straightforward algebra gives
n

- y- _ - - + a( -' X)_ +(x'-x)L EI(Xj - x) (12.53)

YI, - - En - EN fJ X - n

L (Xi-X?
By averaging over the distribution of the E'S, it follows from (12 .53) that YI, is
model-unbiased for fixed x's in the finite population and the two samples.
Furthermore, from (12.53),
_
[ -y) X- 21 ] 2 2· ( 1 1 ) 2( -, X)- 2 2( 2) (x' - i)2
E(y,, =Uy (1- P ) ;;_ - N +{3 x - +Uy 1- p L(XI- i)2

(12.54)
The last term on the right in (12.54) arises from the sampling error of b and is of
order 1/n relative to the first two terms on th(: right. Averaging the first two terms
on the right over the distribution of the X 's created by repeated random selections
of the finite population and the two samples, we get

" EV(y,,)==tTy 2( I - p 2 ) (1 l(
;;_ - N1 ) + p 2tTy -;;;
1 - N1 ) (12.55)

2( 2) 2 2 2
= tTy I - p +~_£L (12.56)
n n' N
This expression has the same form as (12.49) except that in (12.56) and p u/
refer to the superpopulation.
Double sampling with regression has been extended by Khan and Tripathi
(1967) to the case where p auxiliary x variables are measured in the second
sample, Y being estimated by the mUltiple linear regression of y on these
variables. With the second sample a random subsample of the first and with
multivariate normality assumed for y and the x 's, the extension of (12.54) for
p > 1 gives for the average variance

V(91,) = S/(l - R2)[1+(n'~n)

2
p ]+R S/ _ §.i_ (12.57)
n n (n - p-2) n' N

where R is the multiple correlation coefficient between y and the x 's.

The case where the second sample is drawn independently of .the first was
considered by Chameli Bose (1943).
 DOUBLE SAMPLING 341
12.7 OP11MUM ALLOCATION AND COMPARISON WITH
SINGLE SAMPLING
From the variance formula (12.49). which assumes lin negligible, double
sampling with a regression estimate can be compared with a single simple random
sample with no regression adjustment. We have
S 2 S 2(1_p2) p 2S}
V+::.L .= Y +...._._: C=cn+c'n ' (12.58)
N n n'
By the Cauchy- Schwarz inequality, the product VC is minimized when
cn 2 c'n ,2
so that (12.59)
S/ O-p 2) p 2 S/
Substitution in VC gives
CS 2
( VC)mi" :: S/(Jc(1- p2) +JC'p2)2 - ~ (12.60)

Thus, for a specified cost C,

(12.61)

If all resources are devoted instead to a single sample with no regression

adjustment , this sample has size CI c and the variance of its mean is

V(y) = cS/ -~ (12.62)

C N
Hence, optimum use of double sampling gives a smaller variance if

C > (Jc(1-p2)+~)2 (12.63)

This inequality may be expressed in two ways . .

(12.64)

or
2 4(c/c1
(12.65)
p >(1 +e/e,)2

Equations (12.64) and (12.05) give lhe critical ranges of cle' for given p and of p
for given c/e' that make double sampling profitable.
Figure 12.1 plots the values of the r tio cl e ' (on a log scale) against p. Curve I is
the relationship when double and single sampling are equally precise; curve II
342 SAMPUNG TECHNIOUES

holds when V..", =0.8 V(j), that is, when double sampling gives a 25% increase
in precision ; and curve III refers to a 50% increase in precision. For example,
when p = 0.8, double sampling equals single sampling in precision if cl c' is 4,
gives a 25% increase in precision if clc' is about 7'i., and a 50% increase if clc' is
about 13.

100
!
E m\
::I
'E
I;::
50
.5 40
""c 30
\ \
n\ \
::I

&
"§20
2
tE
~~ "i'\.. \ \
10
::I
... 8 ...... '\.

~ 6 ........
" "-
...
.5 5 ............
........,_
" ,,~
C
"
::I

& 3
~ ~,\

"" ,'.~
'(

§ 2

~
"0
:8
•
a::
1
0.4 0.5 0.6 0.7 0.8 0.9 1.0
p• correlation ~ Yi .00 "i

.... 12.1 ttelation between c/ c' and p for three fixed values of the relative precision of double and
single sampling.
Curve I: double and single sampling equally precise.
Curve II : double sampling gives 25 per cent increase in precision.
Curve III: double sampling gives 50 per cent increase in precision.

For practical use, the curves overestimate the gains to be achieved from double
sampling, because the best values of nand n' must either be estimated from
previous data or be guessed. Some allowance for errors in these estimations
should be made before deciding to adopt double sampling.
For any p, there is an upper limit to the gain in I,recision from double sampling.
This occurs when information on i' is obtained free (c' =0). The upper limit to the
relative precision is 1/(1-p2).
 DOUBLE SAMPLING 343
12.8 ESTIMATED VARIANCE IN DOUBLE SAMPLING
FOR REGRESSION
If terms in 1/ n are negligible, VC9,,) is given by (12.49):
V(- ) . S/(l_ p 2) + p 2S/_~
y" n n, N
With a linear regression model, the quantity

s;.x = n ~ 2[ f 1-1
(Yi - 9)2 - b
2
f
1- 1
(Xi - x?] (12.66)

is an unbiased estimate of S/(I _p2). Since

S
2
=
r (YI- ji)2
y n-1
IS an unbiased estimate of S/, it follows that
2 2
Sy -Sy.x

is an unbiased estimate of p 2 S/.

Thus a sample estimate of V(YI,) is
$2 S 2 S2 S 2
y ,y.. ::.L
v(y,,) =::l!:+ (12.67)
n n N
If the second sample is small and terms in 1/ n are not negligible relative to 1, an
estimate of variance suggested for simple random samples from (12.54) is
_)_
v ( y"
2
-Sy.x
[.!. +r(i'-i)2]
( -)2 +
s/-s;.x
,
~
N (12.68)
n X,-X n
This is a hybrid of the conditional variance and the average variance.

12.9 RAnO ES11MATORS

If the first sample i used to obtain i' as an estimate of X in a ratio estimate of Yo
the estimator of Y is
- ji-, (12.69)
YR =-:X
x
To find the approximate variance, write

.9R-Y=~i'- Y
X

=
.9 - -) +:i(i'-X)
(:iX-Y ji -

-: x - X)-
X(_Y - Ri)+Y(-'
=-::-
x X
344 SAMPLlNG TECHNIQUES

The first component is the error of the ordinary ratio estimate (section 2.1 1). In
obtaining the appropriate error variance in ection 2. ] ] . we replaced the factor
Xli by unity in this term. To the same order of approximation. we replace the
factor y/i in the second component by the population ratio R = Y/X. Thus
YR - Y =(y - Ri)+R(i' -X) (12.70)
If the second sample is a random subsampJe of the first,

- . (1 1)
V 2(.YR - Y). ;;- n' Sd
,2
(12.71)

where Sd, 2 is the variance within the second sample of the variate d, = Y, - Rx"
Averaging now over repeated random selection of the first ample.
V(YR) = \/1 £2+£1 V 2

(1 1) 2
n N SY + ---
== -;-- rt
(1 1)
II'
(S Y2 -2RSY.. +R 2 S.)
2 (12.72)

since S/2 is an unbiased estimate of S,/ = S/ - 2RSy~ + R 2S/.

Separating ,he terms in 1/ nand 1/ n' we get
2
2RSyx - R S/ -~
2
V(- ). S/,-2RSy., + R S/
YR +, N (12.72')
n n

• 12.10 REPEATED SAMPLING OF THE SAME PO PULAnON

The practice of relying on samples for the collection of important series of data
that are published at regular intervals has become common. In part, this is due to a
realization that with a dynamic population a census at infreq uent intervals is of
limited use. Highly precise information about the characteristics of a population in
July ] 960 and July 1970 may not help much in planning that demands a
knowledge of the population in 1976. A series of small samples at annual or even
. horter intervals may be more serviceable.
When the same population (apart from the changes that the passage of lime
introduces) is sampled repeatedly, the sampler is in an ideal po ition to make
realistic estimates both of costs and of variances and to apply the techniques that
lead to optimum efficiency of sampling. One important question is how frequently
and in what manner the sample should be changed as time progresses. Many
considerations affect the decision. People may be unwilling to give the same type
of information time after time. The respondents may be influenced by information
which they receive at the interviews, and this may make them progressively less
representative as time proceeds. Sometimes, however, cooperation is better in a
second interview than in the first, and when the information is technical or
confidential the second visit may produce more accurate data than the first.
 DOUBLE SAMPLING 345
The remainder of this chapter considers the question of replacement of the
sample and the related question of making estimates from the series of repeated
samples. The topic is appropriate fo the present chapter because double sampling
techniques can be utilized.
Given the data from a series of samples, there are three kinds of quantity for
which we may wish estimates.
1. The change in Y from one occasion to the next.
2. The average value of Y over all occasions.
3. The average value of Y for the most recent occasion.
In most surveys, interest centers on the current average (3), particularly if the
characteristics of the population are likely to change rapidly with time. With a
population in which time change are low, On the other hand, an annual average
(2) taken over 12 monthly samples or four quarterly samples may be adequate for
the major uses. This.would be the situation in a study of the prevalence of chronic
diseases of long duration. With a disease whose prevalence shows marked
seasonal variation, the current data are of major interest, but annual averages are
also useful for comparisons between different regions and different years. Esti-
mates of change (1) are wanted mainly in attempts to study the effects of forces
that are known to have acted on the population. For instance, if a bill is passed that
is upposed to stimulate the building of houses. it is interesting to know whether
the building rate of new houses ha increased in the succeeding year (with a
reali7..ation that an increase may not be entirely due to the bill).
Suppose that we are free to alter or retain the composition of the sample and
that the total size of sample is to be the same on all occasions. If we wish to
maximize precision, the following statements can be made about replacement
policy:
1. For estimating change, it is best to retain the same sample throughout all
occasions.
2. For each estimating the average aver all occasions, it is best to draw a new
ample on each occasion.
3. For current estimates, equal precision is obtained either by keeping the same
sample or by changing it on every occasion. Replacement of part of the sample on
each occasion may be better than these alternatives.
Statements 1 and 2 hold because there is nearly always a positive correlation
between the me.tsurements on the same unit on two sllccessive occasions. The
estimated change on a unit has variance 5 12 + 52 2 - 2p515l , where the subscripts
refer to the occasions. If the change is estimated from two different units, the
variance is 5 12 +522 . In estimating the over-all mean for the two occasions, the
variance is (5 12 +52 2 + 2pSI52)/4 if the same unit i retained and (5 12 + 52 2 )/4 if a
new unit is cho en.
Statement 3, which is less obvious, is investigated in succeeding sections.
346 SAMPUNG TECHNIQUES

12.11 SAMPLING ON 1WO OCCASIONS

Suppose that the samples are of the same size n on both occasions and that the
current estimates are of primary interest. Replacement policy has been examined
by Jessen (1942). For simplicity, we assume that simple random sampling is used.
The mean of the first sample has varience S/ In, there being no previous
information to utilize. In selecting the second sample, m of the units in the first
sample are retained (m for matched). The remaining u units (u for unmatched)
are discarded and replaced by a new selection from the units not previously
selected.
Notation.
Yk .. =mean of unmatched portion on occasion h
Ykm = mean of matched portion on occasion h
Yk = mean of whole sample on occasion h
The unmatched and matched portions of the second sample provide indepen-
dent estimates Y2 .. ', Y2," ' of Y2, as shown in Table 12.1 . In the matched portion we
use a double-sampling regression estimate, wnere the "large" sample is the first
sample and the auxiliary variate XI is the value of Yi on the first occasion. The
variance of Y2m' comes from (12.49), p. 339: note that our m and n correspond to
nand n', respectively, in (12.49). The fpc is ignored.

TABLE 12.1
ESTIMATES FROM THE UNMATCHED AND MATCHF.D PORTIONS
'r
Estimate Variance

Unmatched :

Matched :

The best combined estimate of Y2 is found by weighting the two independent

estimates inversely as their variances. If W2 ,,' W2m are the inverse variances, this
estimate is
Y2' = q,2YZ.' + (1 - q,Z)Y2m' . (12.73)
where
.I. _ W 2 ..
'1'2 -
W 2 .. +W2m
By least squares theory, the variance of Y2' is

V(Y2') = w:
2..
: w2m
 DOUBLE SAMrLlNG 347
From Table 12.1, thi works out.after simplification as
V{ - ') _ S/(n - up2)
Y2 - n2_u2p2 (12.74)

Note that if u = 0 (complete matching) or if u = n (no matching) this variance

has the same value, 5/In.
The optimum value of u is found by minimizing (12.74) with respect to variation
in u. This gives
u 1
(12.75)
;;=1+~'
When the optimum u is substituted in (12.74), the minimum variance works
out as
5/
V..".(Y2') = - (1 +Jl-p2) (12.76)
2n

Table 12.2 shows for a series of values of p tbe optimum percent that should be
matched and the relative gain in precision compared with no matching. The best

TABLE 12.2
OPTIMUM % MATCHED
% gain with
p
Optimum
% matched
% gain in
precision
I - --
m
n
I
3
m
-n -- I
4

O.S 46 7 7 6
0.6 44 11 11 9
0 .7 42 11 17 IS
0.8 38 2S 2S 23
0.9 30 39 39 39
0.95 24 52 SO 52
1.0 0 100 67 75

percentage to match never exceeds 50% and decreases steadily as p increases.

When p =: 1, the formula suggests m = 0, which lies outside the range of our
assumptions, since m has been assumed reasonably large. The correct procedure
in this case is to take m =: 2. The two matched units are sufficient ~ determine tbe
regression line exactl .
The greatest attainable gain in precision is 100% when p "" 1. UnIfls p is high,
the gains are modest.
 348 AMPUNG TECHNIQUES

Although the optimum percentage to match varie with p, only a single

percentage can be u ed in practice frail Items in a urvey. The right-hand
columns of Table 12.2 show the percent gains in precision when one third and one
fourth of the units are matched. Both are good compromi e" except for items in
which p exceeds 0.95.
Kulldorff (1963) gives an e tcnsive discu sion of this model. He considers the
case where on the econd oc(:asio n the c.ost of measuring a matched unit (i.e. , one
previously measured) may differ from that of measuring a new unmatched unit
and does not assume equal ample ~izes on two occasions. Thus, apart from fixed
costs, his cost on the second ccasion is

C2 = me,,, + LIe., : (12 .77)

where ~ :: em/ru' If sample sizes are the same on the two occasions so that
m + u = n, the optimum unma~ch ed proportion on the second occasion is found by
minimizing
ve2 (n - up 2) (1 - JLP2)
~:7 = [M + u(l - 8)] (n 2_ U2p 2) [~+ #L(1 -~)J(I _ ~ 2p2) (12.78)

where ~ = u/ n and V comes tram (12.74). If ~ < 1, matching being cheaper, the
optimum t,ruportion matched is, of course, greater than the values in Table 12.2.
He also ·Ieal with the case where the costs are to be the same on the two
' l()ccasion ~"
In some ap;,li.:atlOns the data for occasion 1 provide several auxiliary variables
correiait'd with Yl, one of which will, of course, usually be YI. For example, in
estImating th kill Y7 of waterfowl per hunter in Ontario from 1968 to 1969, Sen
(l973a) found that the kill per hunter and the number of days hunted in 1967 and
1968 were bul.h correlated with Y2' In this paper he extended the preceding
analysis to the cas where 92," is adjusted by its multiple linear regression on the
3uxiliary variables and where the samples on the two occasions are of unequal
sizes. With large samples of equal size the only change in (12.76) for V"",(jiz') is to
replace p '2 by .he .'quare R 2 of the multiple correIa n coefficient between Y2 and
the auxiljpry "'ariables (as uming multivariate normality). The corresponding
theory for the case where Y2m is adjusted by the muJtivariate ratio estimate is
gIven in Sen (1972) for equal sample sizes and in Sen (1973b) (or unequal sample
sizes.

U.ll SAMPLING ON MORE THAN NO OCCASIONS

The general problem of replacement has been studied by Yates (1960) and
Patterson (1950), with respect to both current estimates and estimates of change.
When there are more than two occasions, the opportunities (or a flexible u e of the
data are increased. On occasion h we may have parts of the sample that are
 DOUDL SAMPLING 349
matched with occasion h - I, parts that are matched with both occasions h - I and
h - 2, and ·so on. In attempting to improve the current estimate, we might try a
multiple regression involving all marchinj.!s to previous occasions. It is also
possible to revise the current estimate for occa ion h - ] after the data for occasion
h are known. In the revi ed estimate the regression of occasion h - 1 on both
occasion h - 2 and occasion h could be utilized, as uming 'that suitably matched
portions of the sample were available.

TABLE 12.3
EsTIMATES OF y~ ON THE hth Oc ASION

Estimate Variance

Unmatched :

Matched :

The present section contains an introduction to the subject. Attention will be

re trieted to current estimates in which only the regre sion on the sample
immediately preceding is used . This results in some los of precision but, since the
correlation p usually decreases as the time interval between the occasions is
increased, the loss of precision will seldom be great. The variance S2 and the
correlation coefficient p between the item values on the same unit on two
succes ive occasions are assumed constant throughout.
On the h th occasion let mh and u" be the numbers of units that are matched and
unmatched, respectively, with the (h -l)th occasion. The two estimates of Yh that
can be made are given in Table 12.3. The only change in procedure from the
second occasion (Table 12.1) is that in the regression adjustment of the estimate
from the matched portion we use the improved estimate Y~ _ I instead of the
sample mean Yit-T'
The variance of the matched estimate Y"m in Table 12.3 is derived from (12.49)
I

on page 339. Note that (a) our m corresponds to the n in (12.49) and (b) the term
p 2 S// n 'in (12.49), which equals 8 2 V(i ' ), is replaced by p2V(Y~ _ I)' since B =p
when S is constant on successive occasions and Y~ _ I corresponds to i ' in the
earlier analysis.
We now examine the precision obtained if the optimum mIl and Uk and the
optimum weights are used on every occasion. It will be found that the optimum
mh/n increase steadily on ucces ive occasions, rapidly approaching a limiting
. value of i.
350 SAMPUNG T£CHNIQUES

Weighting inversely as the variance, the best estimate of y" is

y,,' = </>"y",,' +(1- </>")y,,,"' (12.79)
where </>" = W",,/(W,," + W" ... ). This gives
1 g,,5 2
V(y,,') = w.lou + w.'"n n
-
where gil denotes the ratio of the variance on occasion h to that on the first
occasion. Substituting for Wlow, W"," from Table 12.3, we have
5 2 n 2 1
V(-,)=-=5(W",,+Wlom ) = UII+(1 2) 2 (12.80)
y" gil __:::j!_+P g" - I
mil n
We now choose mlo and u" to maximize this quantity and therefore to minimize
V(j,,'). Writing u,,::: n -mIl and differentiating the right side of (12.80) with
respect to mIl we obtain
1 _ ~2 = (1 _ p2+p2 g "_1)2
mil mIt n
This gives, on solving for the optimum mIl, say,

(12.81)
'f

When this value is substituted in (12.80), the relation becomes; after some
algebraic manipulation,

1 -~
-1 = 1 +--__;-=== (12.82)
glo g"-1(1 +Jl- p 2)
This relation may be written
r" = 1 +br"_1
where r" = 1/KIt and rl = 1/gl ::: 1. Repeated use ofthis recurrence relation gives
1
- = rIo = 1 + b + b 2 +, .. + b 11 - 1 =I-b"
--
~ I-b

where, from (12.82), b = (1-~)/(1+~). Since O<b<l, the limiting

variance factor goo is

goo= I-b =
2Jl-1l
r=---? (12.83)
l+vl-p-
 DOUBLE SAMPUNO 351
Hence the variance of jilt' tends to

V(Yoo') = S2 ( 2~ ) {12.84}
n 1 +Jl - p 2
Finally, the limiting value of mit is obtained from (12.81) as

moo ~ 1
-;-=goo(1 +J1 _ p2) 2
irrespective of the value of p .
Table 12.4 shows the optimum percentage to matCh-100m";", as found from
(12.81)-and the resulting variances for p = 0.7, 0.8, 0.9 and 0.95 and for a series
of values of h.
TABLE 12.4
OPTIMUM % MATCHED AND VAiUANCES

% matched loom,,/n g" os n V(j" ,)/$1

p- p'"
h 0.7 0.8 0.9 0.95 0.7 0.8 0.9 0.95

2 42 38 30 24 0.857 0.800 0.718 0.656

3 49 47 42 36 0.837 0.762 0.646 0.556
4 50 49 47 43 0.834 0.753 0.622 0.515
5 SO 50 49 46 0.833 0.751 0.613 0.495
00 SO 50 SO 50 0.833 0.750 0.607 0.476

By the fourth occasion, the optimum percent matched is close to 50 for all the
values of p shown, although a smaller amount of matching is indicated for the
second and third occasions. The reductions in variance, (1 - &), are modest if pis
less than 0.8.

12.13 SIMPLInCADONS AND FURDIER DEVELOPMENTS

In practical application the preceding analysis may need modification. We
assumed that all replacement policies cost the same and are equally feasible. With
human populations, field costs are likely to be lower if the ame units are retained
for a number of occasions. If estimates of the change in the population total or
mean are of interest, thi factor also points toward matching more than half the
units from one occasion to the next.
It is convenient also to keep the weights and the proportion matched constant,
instead of changing them on every occasion. O>nsequently, we will investigate the
 352 SAMPLING TECHNIQUES

variances of ji,, ' and of the estimated change (ji" '- YI. _I) when m, U, and 4> are held
constant. We continue t() write V(y" ,) = g"S2/ n, although the actual value of gIl
will be different {rom that in the preceding section.
The estimate is now
YIi' :::: 4>Y" .. ' + (1 - (/J }Y"m'
Substituting the expressions for the two variances (from Table 12.3), we have
S2
V(y,,'} == ~ ;' 4> 2V(YIt .. '} + (1 - 4»2 V(Y"m')
n
S2[4>2 + (1-4>}2(1 _ p2)] + S2p2(1_(/J)2gh _1
u m n
Hence
_ [4> 2+(I_ 4» 2(1_p2)]+ 2(1 A..)2
g,, - -; A P -", g" - I (12.85)

where foL = u/ n, A = m/ n. Write this relation as

gil = a +bg" _1
By repeated application. we have, ~ince gl = I,
a (1- bit - I) b" I
g,, = I- b +
Since b :::: p 2(1 - 4»2 is Icss tttan I , the limiting value is
a A4> 2+ foL(1 - 4> )2(1 - p2)
g ,= I - b = AfoL[1 - p2(] - 4» 2] (12.86)

The value of the weight 4> that minimizes the limiting variance may be found by
differentiating (12.86). This leads to a quadratic equation whose appropriate root
is
Ji7(Jl- p2+ 4 AfoLP ':! -~J
4>oPI:::: 2Ap-

In practice, the value of p will not be known exactly and will differ from item to
item. A compromise value can usually be chosen. Oearly, 4>orx will be less than
foL = u/n, since the matched part of the sample gives higher precision per unit than
the unmatched part. For example, with foL =0.25, that is, ! of the sample
unmatched, q,opr turns out to be 0.216, 0.198, and 0.164 for p = 0.7 ,0.8,0.9. The
choice of 4> :::: 0.2 would be adequate for this range of p.
For the estimate of change, we have
V(y" '- Yl.-l} = V(y,,') + V(Yi. - J) - 2 cov (Y,,'YI. _,) (12.87)
 DOUBLE SAMPLING 353
To find the covariance term, note that if YIII, Yh-I,I are the values for the ith unit
in the matched set on occasions hand (h -1), our model is
YIII = Yk +P(Yh - l,l- Yh - 1)+ehl
where the ehl are independent of the y 's. From thjs model it is found by
sub titution that
9h ... ' = Yhm + P(ji;' - I - jih'- I,,,,) = Yh + P(jii. - I - Y" -I) + els m
Hence the covariance of jill ... ' and jil. _1 is p V(jil. _I)' But
cov (ji" 'y I. - I) = COV {[I/>jihu + (1- I/> )Yhm ']jil. - I} = p(1- 1/» V(y;' _t)
since jillu is independent of ji ~ _ I' From (12.87), thjs gives

V(jih ' - Y;' _I) = 52{gh + gh - l[l- 2p(1- I/>)J} (12.88)

n
From (12.85) and (12.87), the variances of jih' and (jih' - YI._I) may be com-
puted for any values of m, 1/>, and p. Table 12.5 shows the resulting percent gains in
efficiency for these estimates relative to the estimates obtained from independent
TABLE 12.5
PERCENT GAINS IN EFFICIENCY FOR THE CuRRENT ESTIMATE Yh' AND THE
ESTIMATE OF CHANGE (Yh ' - Y~ - l)' PROPORTIONS MATCHED : ! AND,

Current estimate: percent gain = 100(1- gh)/gh

P = 0.7 p = 0 .8 p =0.9 P = 0.95

I ~ .1 2 I
h ! • i .; ! • !

2 14 10 22 14 33 19 41 22
3 16 14 30 20 52 32 67 39
4 17 15 32 24 59 40 79 52
co 17 15 33 26 62 50 89 74

RO G
15 10 27 18 56 40

h Estimate of change: percent gain = 100(2 - gh ')/ gh'

2 106 153 156 233 245 399 326 565

3 113 160 170 245 277 415 365 588
4 115 160 174 251 285 424 388 603
co lIS 163 178 251 292 440 397 624

RO 101 163 166 269 381 605

-Described on p. 355.
 354 SAMPLING TECHNIQUES

samples on each occasion. The proportions matched are A = m/ it = ~ and ~. The

weight ¢ was taken as 0.35 for A = ~ and 0.2 for A =~.
Not surprisingly, the major features of Table 12.5 are the large gains in
efficiency for the estimates of change when p is at lea to. 7. Moreover, increase in
the proportion matched from ~ to ~ produces sub tantial gains in efficiency for the
estimates of change at the expense of smaller losses in efficiency for the current
estimates. The results uggest that retention of t ~ or ~ from one occasion to the
next may be a good practical policy if current estimates and estimates of change
are both important.
Comparison of the gains in efficiency for the current estimate YII in Table 12.5
I

with the optimum gains from Table 12.4 suggests that after the econd occasion
little precision is lost by using a constant weight and a fixed proportion matched,
unlessp 2: 0.95 .
If p exceeds 0.8, the regression coefficient h = p may be replaced by I with only
a small additional loss of precision . This gives an estimat YII" of the form
( l2 .8Y)
1n the important Current Population Survey taken monthly by the U.S. Bureau
of the Census, one quarter of the second-stage units are replaced each month, so
that an individual household remains in the sample during four consecutive
months. The household is omitted for the eight ucceeding months but is then
brought back for another four months, thus increasing lightly the precision of
year-to-year comparisons.
" The composite estimate used in this survey is of a form related to (12.89) but
slightly different.
Yh" = (1 - K)YII + K(Yh - 1+ Ylom - Yh - I.m) (12 .90)
where K is a constant weighting factor. The difference is that Yh, the current
estimate for the whole sample, takes the place of the Yhu in (12.89). The quantities
Yhl1l, 9h - I.,., YII in (12 .90) are ratio estimates of a fairly complex type .
The variance of Yh" (due to Bcrshad) is given in Hansen , Hurwitz, and Madow
(1953); see also the Appendix in Hansen et al. (1955). The estimator of the
month-to-month change is

Since the primary units remain unchanged, only the within-units components of
V(Yh") and V(dh ) are affected by the sample rotation policy.
Rao and Graham (1964) have examined the performance of composite
estimators in rotation policies of this type, in which a respondent remains in the
sample for' months and then drops out for m months. They used as models an
exponential correlogram and a linear correJogram in time, descending to zero. A
more complex correlogram, needed if there is a high correlation between months
hand (h - 12), has been studi d by Graham (1973).
 DOUBLE SAMPLING 3'55
Their gains in efficiency for r = 2, 4 and m = 00 correspond to the results for
A = i. ~ in Table 12.5. For an exponential correlogram in which the correlation
between result on the same unit on occasions h, h - i is pi, the lines labeled RG in
Table 12.5 show for p = 0.7,0.8,0.9, their percent gains in efficiency. For each p
they use in (12.90) the optimum K for the current estimates as h -+ 00 . As Table
12.5 shows, they also find that the gains in efficiency from the composite
estimators are much greater in estimating change than current level.
In a more g~neral framework, Scott and Smith (1974) have discussed the role of
time series methods in making estimates in repeated surveys of various types.
In another rotation policy a new sample is drawn on each occasion, with no
matching. With weekly or monthly sampling, this plan i appropriate when annual
estimates, and to a lesser extent semiannual or quarterly estimates, are of primary
importance, for example, in an illness survey with emphasis on chronic diseases . If
the questionnaire obtains for any unit the results fOT the preceding month as well
as for the current month, we can consider composite estimates of the form
(12.91)
where Yh = e timate made from current data in the current sample
Yh - I.h = estimate made from previous month's data in the current sample
Y~ _ I = composite estimate for the previous month
The theory is discussed by Hansen, Hurwitz, and Madow (1953) and Woodruff
(1959), who apply it to a survey of retail sales, and by Eckler (1955). In the Retail
Trade Survey the composite estimate involves a ratio estimate, being of the form

Yh" = (1 - W)Yh+ W (_Yh )Y~- l

Yh - I.h
where W is a weighting factor . Since month-to-month correlations are very high ,
averaging around 0.98, the gains in precision are substantial. One month later a
revised composite estimate for month h is computed, using the results for month h
from the new sample taken in month (h + 1).
With this method, it'is essential that the data obtained for the preceding month
from the current sample be accurate. This may not be so when the data depend on
the unrecorded memory of the respondent, although the method may work
succes fully if the data are of a type that the respondent .records carefully as a
routine matter.
EXERCISES
12. 1 $3000 is allocated for a survey to estimate a proportion. The main survey will cost
$10 pe.r sampling \lOl t . Information is available in files, at a cost of $0.25 per sampling unit,
that enables the UOlts to be classified into two strata of about equal sizes. If the true
proportion is 0.2 in stratum 1 and 0.8 in tratum 2, estimate the optimum n, n ', and the
resulting value of V(p.,) . Does doubl sampling produce a gain in precision over single
all'.pling? (The ratios ,,'IN, II~/N~ may be ignored.) ,
356 SAMPLING TECHNIOUES

12.2 For the Wh , Ph in exercise 12.1 , find the cost ratios c.1 c_' for which double
sampling is more economical than single sampling.
12.3 A population contains L strata of equal size. If V",. denotes the variance of the
meim of a simple random sample and V," VIIs are the corresponding variances for stratified
random sampling with proportional allocation and for double sampling with stratification.
show that, approximately,

t (Yh- Y)l
nV..... =Sh2+_h____
L

t(Yh - YV
- 2 n k
nV,u = Sk + - ; - - - -
n L
where Sh 2 is the average variance within strata. (N and n ' may both be assumed large
relative to L, and the nh in double sampling may be assumed equal to n/ L.)
Hence, if (RP)" denotes the relative precision of the stratified sample to the s.imple
random sample. with a corresponding definition for (RP)d.o' show that
(RP)"
(RP)d. = 1 + (nl n ')[(RP)" - 1]

For (RP).. = 2. plot (RP)", against nln'. How small must this ratio be in order that
·(RP)... = 1.9?
12.4 If p '" 0.8 in double sampling for regression , how large must n' be relative to n, if
the loss in precision due to sampling errors in the mean of the large sample is to be less than
10%.
12.5 In an application of dt>uble sampling for regression, the small sample was of size
87 and the large sample of size 300. The following computations apply to the small sample.
L (y,- y)2 = 17,283, L (Yi - y)(x,- i) = 5114, L (x, - ,f)2 = 3248

Compute the standard error of the regression estimate of Y.

12.6 For p =0.95. verify the data given in Table 12.4 for the optimum percentage that
should be matched and for the gain in precision relative to no matching. Compute the
corresponding percent gains in precision if one third of the units are retained from the first
to the second occasion and one half of the units are retained on each subsequent occasion.
12.7 In simple random sampling on two occasions, suppose that the estimate on the
second occasion is. in the notation of section 12.11 ,
y/ = (J - cjl)(y, + Y2m - y,,,,) +cjlY2u
(a) Ignoring the fpc, show that

V(Y/1 = S2{(1 _cjl)2 (1 +",(1 - 2p)J+ cjll}

n A ~

where A = ml n, ~ = ul n. (b) For given p, A. ~, find the value of cjl that minimizes V(jiz").
Show that if p exceeds! the best weight <p lies between ~ and 11-1(1 +~).
 DOUBLE SAMPLING 357
*,
12.8 For IJ. = IJ. =!, P = 0.8, and p = 0.9, compare V (yz") in the preceding exercise
with the variance of the optimum composite regression estimate Y2', as given by equation
t n
12.74. (In 92" take tP = 0.2 when IJ. = and tP = 0.4 when IJ. = Verify that for these values
of p the estimate ;12" is almost as precise as 92' for both IJ. =~ and IJ. =!.
12.9 An independent sample of size n is drawn each month . From the sample taken in
any month, data are obtained for the current and the preceding month. A composite
estimate 9~' is made as in (12.91), section 12.13 .
9~' = 9k +4>.(YI.-I- YA - I.h)
The model is
YAI = Y h +p(Y. 1.1 - Y. - l)+e~1
where eAI is independent of the y's and has variance (l_p 2). Show that (a)
9h'- Y. = i. +4>.(9;'- , - Y" - I)+(P - 4>.)(Y,, - l.. - Y" - I)
(b) If V(y. ') = 8~S 2 / n, where S2 is constant on all occaslons,
s. = (I _ p2) + 4>.28. _ + (p -tP.)2
1

(c) The optimum 4>k "'" p/(l + 8h - l) and the resulting optimum 8. is
p2
8h = 1- ---
1 +8,, -1
(d) The limiting 8. is 800 = JI - p'. These results were given by Eckler (1955).
12 .10 If E" = V(y)/V(y,,) and E,._ = V(y)/V(y,._) are the relative efficiencies of the
linear regression and ratio estimates of the sample mean of a simple random sample, show
that for both Y" and y" the corresponding relative efficiency to 9 in double 'ampling with
optimum choice of nln' is (ignore liN),

EtU =E/( + .jfJE-IY

1

Hence note that with either of these estimators, double sampling will not be highly effective
unless e'/e is small (e.g., <1/10) . For example, with c'/ e = 1/10, E =6 gives EtU = 2.1.
12.11 In sampling on two occasions, suppose that SI == 52 = S and that the samples are
large, so that the regression coefficients of Yll on yHand of y" on Y21 in the matched part of
the samples on the two occasions are both effecti ;rely equal to p. The estimate Y2' in section
12.10 is constructed and an analogous estimate Yl ' u ing the regression of YlI on Y ~I. Show
that
(i)

(ii)

(One way of doing this is to express (92':t Yl') as linear functions of (ji2," ± 91 ... ) and
(92.:t 91.), which are uncorrelated).
Note that, as intuition suggests, (i) IS minimized when u = 0, while (ii) is minimized when
U " n.
12.12 The most favorable case for the application of the method in exercise 12.1 occurs
when the true proportion is 0 in stratum 1. In estimating the total number of units Y, in the
358 SAMPLING TECHNIQUES

population that possess an attribute costly to measure, this happens when there is a second
attribute, cheap to measure, such that only units having the second attribute can posse s the
first attribute. In a simple random sample of size n', count the number m' who have
attribute 2. Draw a subsample ohize JIm' = m' / k from these, and count the number r who
have attribute 1.
(0) From theorems 12.1 and 12.2, show tbat 'P"J = Nkr/n is an unbiased estimate of Y J
with variance

V('P") =
J
fY2P
n'
I
[Q (1-~)
I N
+ P2(k - 1)]
P +P2
J

where PI> (P, + P 1 ) are the population proportions having attributes 1,2. Assume 1/N (P , +
PJ negligible.
(b) With c = 10, c' = 1, the investigator guesses that P J = 0.25, P2 = 0.15 . Is double
sampling profitable in this case?
CHAPTER 13

Sources of Error in Surveys

13.1 INfRODUcnON
The theory pre ented in preceding chapters assumes throughout that orne kind
of probability sampling is used and that the observation y, on the ith unit is the
correct value for that unit. The err r of estimate arises solely from the random
sampling variation that is present when n of the un its are measur d in~tead of the
complete population of N units.
These assumptions hold reasonably well in the simpler types of surveys in which
the measuring devices are accurate and the quality of work is high . In complex
surveys, particularly when difficult problems of mt: asurement are involved, the
assumptions may be far from true. Three additional source of error that may be
present are as follows.

1. Failure to measure some of the units in the chosen sample. This may occur by
oversight or, with human populations, because of failure to locate some indi idu-
als or their refusal to answer the questions when located .
2. Errors of measurement on a unit. The measuring device may be biased or
imprecise. With human populations th respondents may not possess accurate
information or they may give biased answers.
3. Errors introduced in editing, coding and tabulating the re ults.

These source of error necessitate a modification of the tandard theory of

sampling. The principal aims of such a modification are to provide guidance about
the allocation of resources between the reduction of random sampling errors and
the reduction of the other error and to develop methods for computing standard
errors and confidence limits that remain valid when tbe other errors are pre ent.

13.2 EFFECTS OF NONRESPONSE

We will use the term non response to refer to the failure to measure some of the
units in the selected sample. In the study of nonrespon e it is convienicnt to think
of the population as divided into two "strata", the first consi ti g of all unit!. for
359
360 SAMPLI NG TECHNIQUES

which measurements would be obtained if the units happened to fall in the sample,
the second of the units for which no measurements would be obtained. The
composltl()n f the two strata depe nd Intimately on the methods used to find the
units and obtain the da ta . A survey in which at least three calls are made. if
necessary. on every hou e and in which a supervisor wi th exceptional powers of
persuasion calls o n all persons who refuse to give data will have a much smalle r
"nonresponse" stratum than one in which only a single attempt is made for every
hou e .

TABLE 13. 1
RESPONSIOS TO THREE R EQU~STS IN A MA ILED INQ UIRY

Average Number
Number of ~~of of Fruit Trees
G ru wers Populallo n per Grower

Response to first mailing , ~!)

10 456
Response to second maIli ng 543 17 382
Response :0 th ird mail ing 434 14 340
Nonrespondenls after 3 mai lings 1839 59 290

Total popu lallon 3116 100 329

"
This division into two distinct strata is, of co urse, (10 oversimplification . hance
plays a part in determining whether a Itn!t IS found and mea ured in a given
numbe r of atte mpts. In a more complet(.; ~pecification of the problem we would
attach to each unit :l probability repre~e nting the chance that it would be
measured by a given field method if it fe ll in the sample.
The sa mple provides no information about the nonresponse stratum 2. This
wo uld not matter if it could be ass umed that the characteristics of stratum 2 are the::
same as those of stratum 1. Where check s have been made, however, it has often
been found that units in the " nonresponse " stratum differ from units that are
measurable . An illustration appears in Table 13. 1. The data come from an
experimental sampling of fruit orchards in North Carolina in 1946. Three
successive mailings of the same questionnaire were sent to growers. For one of the
questions-number of fruit trees-<:ompJete data were available for the popula -
tion (Finkner, J 950).
The steady decline in the number of fruit trees per grower in the successive
responses is evident , these numbers being 456 for respondents to the first mailing,
382 in the second mailing, 340 in the third , and 290 for the refusals to all three
letters. The total response was poor, more than half the population failing to give
data even after three attempts.
 SOURCES OF ERROR IN SURVEYS 361
We now consider the effects of nonresponse on the sample estimate. Let N I , N2
be the numbers of units in the two strata and let WI = Nil N, W 2 = N21 N, so that
W 2 is the proportion of nonresponse in the population. Assume that a simple
random sample is drawn from the population. When the field work is completed,
we have data for a simple random sample from stratum 1 but no data from stratum
2. Hence the amount of bias in the sample mean is
£(91) - ¥ = ¥I- ¥ = ¥I -( W) 5\ + W 2 ¥2)
=W2( ¥J - Y2) (13.1)
The amount of bias is the product of the proportion of nonresponse and the
difference between the means in the two strata. Since the sample provides no
information about Y2 , the size of the bias is unknown unless bounds can be placed
on Y2 from some source other than the sample data. With a continuous variate,
Ihe only bounds that can be assigned with certainty are often 0 wide as to be
useless.
Consequently, with continuou data, any sizable proportion of nonresponse
usually makes it impossible to assign useful confidence limits to Y from the sample
results. We are left in the position of relying on some guess about the size of the
bias, without data to substantiate the guess.
In sampling for proportions the situation is a little easier, since the unknown
proportion P2 in stratum 2 must lie between 0 and 1. If W 2 is known, these bounds
for P2 enable us to construct confidence limits for the population proportion P.
Suppose that a simple random sample of It units is drawn and that measurements
are obtained for It J of the units In the sample. Assuming It I large enough, 95%
confidence limi t!. for PI are given by

PI ±2.JPIQll lt l
where PI is the sample proportion and the fpc is ignored.
When we try to derive a confidence statement about P, we are on safe ground if
we assume P2 = 0 when finding A
and P2 =< 1 when finding Pu . Thus we might
take, for 95% limits,

A = W I(PI -2.JPIQ dlt,)+ W2(0) (13.2)

Pu = WI (PI +2.JPIQ t!n l ) + W 2(1) (13.3)

It is easy to verify that the e limits are conservative, that is, that
Pr(PL s P s fi u »O.95
The limits can be narrowed a little by a more careful argument (Cochran,
Mosteller, and Tukey, 1954, p. 280), since P2 cannot be 0 and 1 simultaneously, as
assumed above.
 362 SAMPLING rr:. 'HN1QUES

The limits are distressingly wide unless W2 is very small. Table J3.2 hows the
average limits for a sample size" = I 000 and a series of values of W2 and PI ' Since
the limit in (13.2) and (13.3) depend on the value of n I tnumber of respondents in
thesampJe), we have taken nl =- n WI, its average value, in computing Tahle 13.2.

TABL 13.2
95 % CONFIDENCE LIMITS FOR P ( • 0) WHt.N" "" 1000
~o
Nonresponse. Sample Percentage. J OOPI
IOOW2 5 10 20 50

0 (3 .6. 6.4) (8.1. 11.9) (17.5. 22 .5) (46.7, 53.2)

5 (3 .4, 11.1) \7.6, 16.3) (16.5, 26.5) (44.4, 55 .6)
10 (3.2, 15.8) (7.2.20.8) (15 .6, 30.4) (42.0, 58.0)
15 (3.0, 20.5) (6.8,25 .2) (14.7,34.3 ) (39 .6, 60.4)
20 (2.8, 25.2) (6.3, 29.7) (13 .7, 3 .3) (37.2. 62.H)

The rapid increase in the width of the confidence interval with increasing W~ i~
evident. It is of interest to examine what values of n would he needed to give the
ame widths of confidence interval if W 2 were zero. This is easily done when PI i~
50% . For W 2 =5%, Table 13 .2 shows that the half-width of the eonfidcm.l
interval is 5.6. The equivalent sample , izc n. , as uming no nonre~ponse, is found
from the equation

"
5.6 =- 2J(SO)(50)/ n,
n. = 320
For W 2 =- In, 1S, and 20%. the values of II, are 155 ,90. and 60, respectively. It
is evidently worthwhile to devote a ubstantial proportion of the resources to the
reduction of nome ponse.
If the population nonrespon e rate W2 is not known , as will usually be the case,
conservatIve confidence limits can be calculated from the sample data by a method
suggested by a student. In calculating the lower limit, assume that all ample
nonrespondents would have given a negative response . In calculating the upper
limit, assume that sample no ore pondents would have given a positive response .
For example, suppose n == 1000, n J = 800, and PI = 10%, so that 80 sample
members give a positive response and the sample non response rate is 20%. Then,
in percents,
PL ::= 8 - 2J(8)(92)/ 1000 = 6.3%

Pu = 28+ 2J (28)(72)/1000 = 30.8%

The limits arC a little wider than those for PI == 10%, W1 = 20% in Table 13.2.
 SOURCES OF ERROR IN SURVEYS 363
If W 2 is known from previous experience in the particular type of survey,
Birnbaum and Sirken (1950a, 1950b) give a method of finding the sample size n
that guarantees with ri k a an absolute error in the sample proportion less than a
specified amount d. No advance knowledge of P 1,P2 , or P is assumed. If there
were no nonresponse, we would take (by section 4.4),
2
n = t,}PO/d (13.4)
where ta is the normal deviat<... ...orresponding to the risk that the error exceeds d.
With no advance information about P, the least favorable case is P = 0.5, giving
t 2
n=-"-2 (13.5)
4d
By taking the least favorable combination of the bias W2 (P 1 - P2 ) and the value
of Ph Birnbaum and Sirken show that a value of n that still guarantees an error
less than d, with risk a, is approximately
t 2
II == 0 1 (13.6)
4d (d - W2)W1
Notc that no value of n suffices if W 2 > d. If W 2 = O. this equation reduces to (13 .5)
apart from the term - 1. which comes from an approximation in the analysis. Some
values of n given by Birnbaum and Sirken' method are shown in Table 13.3.
This table tells the same sad story as Table 13.2. If we are content with a crude
estimate (d ..;; 20). amounts of nonresponse up to 10% can be handled by doubling
the sample ize. However. any sizable percentage of nonresponse makes it
impo sible or very costly to attain a highly guaranteed preciSIon by increasing the
sample size among the respondents.

TABLE 13.3

SMALLEST VALUE OF n fOR GIVEN LIMIT OF ERROR d, WITH RISK IX = O.OS

i'~
Nonresponse. d(%>
lOOW 20 15 10 5

0 24 43 96 384
2 27 SO 122 653
4 3J 60 J66 2000
6 36 75 255
8 43 99 521
10 53 142
15 112
 364 SAMPLING CHNIQUES

13.3 TYPES OF NONRESPONSE

Some methods for handling the nonresponse problem are described in succeed-
ing sections. A rough classification of the types of nonresponse is as follows .
I. Noncoverage . This is failure to locate or to visit some units in the sample.
This is a problem with areal sampling units. in which the interviewer must find
and list all dwellings (according to some definition) in a city block . It ari 'es also
from the use of incomplete list ', ometimes weather or poor transportation
facilities make it impossible to reach certain units during the period of the survey.
2. Not-at-homes. This group contains persons who reside at home but are
temporarily away from the house. Families in which both parents work and
families without children are harder to reach than families with very young
children or with old people confined to, the hou e.
3. Unable to answer. The respondent may not have the information wanted in
ertain question or may be unwilling to give it. Skillful w rding and pretesting of
the questionnaire are a safeguard.
4. The "hard core." Perso ns who adamantly refuse to be interviewed, who
are incapacitated. or who are far from home during the whole time available for
field work con. titute this sector. It represents a ource of bias that per 'i ,t no
matter how much effort is put into completeness of returns.

The detection and measurement of noncoverage are difficult. With areal

sampling, one method is to revisit the primary units, making a careful listing that
erves as a check. Comparisons of counts of numbers of people or dwellings with
"those in another survey sometimes give a warning that Some have been missed.
When the principal frame is a directory of street addresses, it may be sup-
plemented by an art~a sample, the purpose of which is to sample parts of the town
(new building) not adequately covered by the directory and , in parts in which the
directory seems accurate, to search for addresses missing from the directory .
Surveys using these methods, with a discussion of the problem of noncoverage, arc
described by Kish and Hess (1958) and Woolsey (1956).
In regard to the not-at-homes, the problem is easier in surveys in which an y
adult in the home is capable of answering the questions than in those in which a
single adult, chosen at random, is to be interviewed. A single adult is usually
preferred if the survey is one of individuals in which one person cannot report
accurately for another or if intrahousehold correlations are expected to be high, 0
that the measurement of more than one per on per family is uneconomical. In this
connection, a useful method of selecting a ingle person from a household was
developed by Kish (1949). The original plan was intended for hou eholds that
may contain as many as six eligible persons, but the procedure can be adapted for
smaller or larger households.
The interviewer lists on the schedule the Ligible per ons in the household and
then numbers them : males first in order of decreasing age, then females in order of
 \(lURCES OF F.R RO R IN SU RVEYS 365
o<.:crc<t\ing agc . ach ~c h cdu l e has printed on it one of the se ts of instructions in
Ta hle 13.4 .
Each e li ~i hl e person in a house hold of a give n size has an equal chance 01 bei ng
,dected. except that ad ult<; :1 and 5 in households of size 5 are slightly overrep-
H·,en ted. Since male respondents are concentrated in Tables A, 8 , and C, the
Ill ll'rvil:wcr ca n devote evening. calls to house hold ~ so designated .
TABL 13.4
I NSTRUCTIONS fOR SELtcTING A SINGa R ESPONDENT

If Number of Adults in
Hou ehold is
Relative
Table
Frequency
of Use
Number I
I2 I3 I4 I5 I6
Select Adult Numbered
--
1/6 A I J J J 1 1
JIl2 81 I I I I 2 2
1/ 12 B2 I I I 2 2 2
1/6 I I 2 2 3 3
1/6 D I 2 2 3 4 4
1/ 12 EI I 2 3 3 3 5
1/ 12 E2 1 2 3 4 5 5
1/6 F 1 2 3 4 5 6

13.4 CALL·BACK
A "wndard technique is to specify the number of t:all -backs. or a minimum
numhcr. that must be made on any unit before abandoning it as " unable to
(,'III :\(.:t." · Steph'an and McCarthy (195Rl give data fr m a number of surveys on
thl' P\!rn~ l1taAe !)f the total sample obtained at each call. Average results are
,h,\\l.n In Table 135.

TABLE 13.5
NUMBER OF CALLS REQUIR ED FOR COMPLETED I NTERVIEWS
% of Sample contacted on
First Second Third or Later Per Cent
Respondent Call Call Call No nresponse Total

An)' adult" 70 17 8 5 100

Random adult 37 32 23 8 100

"TwlI . urveys in which the respondent Wit a housewife and a farm operator, respectively.
have been included in the "any adult" group.
366 SAMPLING TECHNIQUES

In surveys in which any adult in the house could answer the que tions, the first
call obtained about 70% of the sample and the first two calls, 87%. The increased
cost of sampling when a randomly chosen adult is to be interviewed is evident, the
first call producing only 37% of the required interviews. The marked success of
the second call reflects the work of the interviewer in finding out in advance when
the desired respondent would be at home and available.
Little has been published on the reJative costs of later calls to the first call. Later
calls would be expected to be more expensive per completed interview, since the
houses are more sparsely located in the area assigned to the interviewer and since
the occupants are presumably people who spend more than an average amount of
time away from home. From British experience, Durbin (1954) suggests that later
calls may be less expensive than would be anticipated. The following figures
(Table 13.6) show estimated relative costs per completed interview (i.e., money
spent on jth calls divided by number of new interviews obtained) for each caJl up
to the fifth in a special study reported by Durbin and Stuart (I 954).

TABLE 13.6
RELATIVE COSTS PER NEW CoMP1.ETED INTERVIEW AT THE iTH CA1.1.
Call 2 3 4 S
Relative cost 100 112 127 lSI 2S0

The estimation of these costs requires care. If the desired respondent is not at
home at the first call, the interviewer may spend time inquiring when this person
will be at home and making a tentative appointment. In the costing such time
should be assigned to the second call instead of to an unsuccessful first call.
A more useful measure is the average cost per completed interview over all
interviews obtained up to the ith call. These figures give the relative costs of
obtaining n completed interviews when we insist on i calls before the interviewer
gives up. In order to compute these figures , we must know bow many interviews
are obtained at each call. In Table 13 .7 these calculations are made under two sets
of assumptions. The first simulates surveys in which any adult can answer the
questions, the second those demanding a random adult. The data on numbers of
interviews obtained were -taken from Table 13.5.
The details of the calculation are shown only for the first assumption, the
method being exactly the same for the second. The symbol no denotes the original
sample size.
Insistence on up to three calls costs only 4 % more per completed interview than
single calls jf any adult is a satisfactory respondent, and only 10% more if a
random adult must be interviewed. How typical these results are is not known, but
the method provides realistic estimates of the cost of inSisting on call-backs jf the
necessary cost and sample size data have been collected. There i al 0 the time
factor: call-backs delay the final results.
 SOURCES OF ERROR IN SURVEYS 367
TABLE 13.7
RELATIVE CosTS PER COMPLETED INTERVIEW UP TO THE ith CALI.

Respondent = Any Adult

"Random"
At ith Call Up to ith Call Adult

No. Cost Total Cost No. Cost

Call Relative of of No. of Total per of per
Cost Ints. • lnts. lots. Cost Int. Ints. lnt.

I 100 O.70n. 70n. 0.70n. 70n. 100 O.37n. 100

2 112 0. 17n. 19.04n. 0.87n. 89.04n. 102 0.32n. 106
3 127 0.07n. 8.89n. 0.94n. 97.93n. 104 0.J6n. 110
4 151 O.04n. 6.0411. 0.98n. 103 .97n. 106 0.09n. J 14
5 250 0 .0211. 5.00". 1.00n. 108 .97n. 109 0.06n. 122

• Interviews

13.5 A MATHEMATICAL MODEL OF TIlE EFFECTS

OF CALL-BACKS
Deming (1953) developed a useful and flexible mathematical model for exa-
mining in more detail the consequences of different call-back policies. The
population is divided into r classes, according to the probability that the respon-
dent will be found at home . Let
w" = probability that a respondent in the jth c1a~s will be reached on or before the
ith call
Pj = proportion of the population falling in the jth class
IL, = item mean for the jth class
0', ~ = item variance for the jth class

or si mplicity we assume w" 0 for all classes, although the method is

easily adapted to include persons impossible to reach. If 9" is the mean for th se
in class j who were reached on or before the ith call, it is al 0 a sumed that
E (9,,) = ILl'
The true population mean for the item i

(13.7)

Consider the composition of the ample after i calls. The persons in the sample can
be cia sified into (r + ) cia ses as follows : in the first c1as and interviewed ; in the
second class and interviewed; and so on. The (r + ))th class consists of all tho e not
yet interviewed after i calls. If the fpc is ignored, the numbers falling in tbe e
 368 SAMPLING TECHNIQUES

(r+ 1) das es are distributed according to the multinomial

[waPI + WI2P2+ ... + WI,p, + (1- L wiJP,))"·
where no is the initial si.ze of the sample
It follows that the number nl who have been interviewed in the course of i cans
is binomially distributed with number of trials "" no and probability of success
L Wi,P,· Hence
,
E(n;) = expected number of interviews, in; calls = no L wi/Pi
, (13.8)

For fixed n h the numbers of interviews nl, obtained (j = 1, 2, ... , r) follow a

multinomial with probabilities WiIP/L Wi/PI' It follows that
- n,wl/p,
E( nil In,) - ,
t- wliPi
Hence, if Yj is the sample mean obtained after i calls,
E (Vi Inj) = dL njjYij) := L nlWijPjIL, = L WijPjIL, = iii (13.9)
C\ ni nl L WijPj L WljP,
Since this result does not depend on ni, the unconditional mean of 9, is also ii,. The
bias in the estimate Y is therefore (ii, - ii )·
The conditional variance of YI for given n, is found similarly to be
'f

fwi/pj[a/ + (IL/- iiYJ

V(ydnl) :=..!.../- - - , - - - - (l3.10)
nl1: wllPI
/
The unconditional variance, ignoring terms of order 1/ n/, is given approximately
by replacing nl in (13.10) by its expected value form (13.8).
Finally, the mean square error of the estimate obtained after i calls is
(13.11)
The cost of making i calls must also be considered. The expected number of new
interviews obtained in the kth call is L (Wkl- wlc - IJ)PI' Hence, if Cle is the cost per
completed interview at the kth call, the expected total cost of making i calls is
noC(i), where
C(i) = CI L WI/P/ +C2 L (W2/- Wlj)Pj + ... + CI L (wll- WI- tJ)P/

Eu.ple. A population with three classes is shown in Table 13.8. The p/ and 1111/ are
intended to represent surveys in which a random adult is interviewed. At 'the lint call the
probabilities 1111/ of obtaining an interview are taken as 0.6, 0.3, and 0.1 in the three classes.
At the second and subsequent calls, the conditional probabilities of interviewing a per on
 SOURCES OF ERROR IN SURVEYS 369
missed previously are 0.9, 0.5, and 0.2. These figures were made higher than the
correspondi ng probabilities at the first call in order to represent the effect of intelligent
inquiry by the interviewer.

TABLE 13.B
CHARACTERISTICS Of THE THRE E CLASSES

Class

I 2 3

p, 0.45 0.50 0.05

0.6 + (0.4)[1 - (0.1)1-1) 0.3 + (0.7)(1 - (0.5)' lJ O. I + (0.9)(1 - (0.8)1-1)
"'",
II' S5 50 45
111', 60 50 40

TABLE 13 .9
NUMBER OF INTERVIEWS. COS1S PER INTERVIEW AND BIASES
Number of Average
Number of Calls Interviews Cost per 11
Required Obtained Interview Bias Bias

I 0.425n. 100 + 1.118 +2.235

2 O.77ln o 105 +0.711 + 1.421
3 0.882n. 108 +0.421 +0.842
4 0.933". 110 +0.266 +0.532
5 0.960n. 114 +0.180 +0.360

The item being estimated is a binomial percentage close to 50%. Two sets of ILl are
considered (I, 11). For si mplicity, the within-class variances 0/ = 1L;(100 - IL,) were all take n
as 2500. The relative costs per completed interview at successive calls were those given in
Ta ble 13 .6.
Table 13.9 shows (a) expected total number of interviews obtained for a total of i calls,
(b) the average cost of these calls per interview. and (c) the bias (ji, - ji) in the estimate y
under as umptions 1 and II about the ILl" .
I n II , for example. the true population mean ji is 54 %. The mean ji 1 obtained from first
calls is 56.235%, giving the bia~ of + 2.235% shown in the table. A policy that requires
three caUs reduces this bias to +0.842% .
The values of MSE(y) obtained hom a given expenditure of money were compared for
the different call-back policies. In the first comparisons the amount of money is sufficient to
take n. "" 500 if only one call is made. From Table) 3.9 the e~pected number of interviews
obtained in the first call is E(n,) = (500)(0.425) = 212.5 . 1f twoealls are made, this expected
number must be reduced to £(n2) = 212.5/1.05 = 202.4. to maintain the same cost, and
similarly (or 3, 4, and 5 call-backs. These value of E("I) were substituted in equation
(13.10) to give V(y) and hence MSE(y).
370 SAMPLLNG TECHNIQU ES

TABLE 13 .10
VALUES OF MSE(y) FOR DIFFERENT CALL-BACK POLICIES
COSTING THE SAME AMOUNT

no = 500
(for first calls only) no = 1000 no = 2000
Number of
Calls No
Required Bias I· II· II II

11.8 13.0 16.9 7. 1 10.9 4 .2 8.0

2 12.4 12.9 14.6 6.7 8.3 3.6 S.2
3 12.7 12.9 13.6 6.S 7. 1 3.4 3.9
4 13.0 13.1 13.4 6.6 6.9 3.3 3.6
5 13.5 13.5 13.8 6.8 6.9 3.4 3.S

• These represent populations with smaller (I) and greater (II) amounts of
bias, as defined in Table 13.8.

Table 13 .10 presents the re ulting MSE's for three amounts of expenditure,
corresponding to " . = 500. 1000, 2000 for a single call. When n. = 500, the values of
MSE(y) are also given for the " no bias" situation in which every IL} = SO. Thi column
shows the effect of call- backs when they are unnece sary, since no bias results from
confining the survey to a single call.
The policies giving the lowe t MSE 's are shown in boldface type . Consider first the
smallest sample size, " . = 500. If call-backs are unnecessary , a policy demanding as many
'as four call-backs results in only a modest increase in the MSE. In I, involving the smaller
amount of bias, the different policies produce about the same accuracy, although three is
the optimum . In II , three to five call-backs are satisfactory, a single call giving a MSE about
25% above the minimum .
For the larger sample sizes the optimum number of call-backs increases to four or five,
and the use of a single call results in more substantial losses of accuracy.

This is, of course, only an illustration. The importance of the method is that as
information accumulates about costs and relative biases an economical policy can
be worked out for any specific type of survey.

13.6 OYIlMUM SAMPLING FRACDON AMONG 11IE

NONRESPONDENfS
After the first attempt to reach the persons in the sample has been made,
another approach, due to Hansen and Hurwitz (1946), is to take a random
subsarnpJe of the persons who have not been reached and make a major effort to
interview everyone in the subsample. This technique was first developed for
surveys in which the initial attempt was made by mail, a subsample of per ons who
did not return the completed questionnaire being approacbed by the more
expensive method of a personal interview.
 SOURCES OF ERROR IN SURVEYS 37 I

This method can be regarded as an application of the techniquc of doubl

sampling for stratification presented in section 12 .2 We first take a simple random
sample of n ' units. Let n I' be the number of units in the sample that provide th
data sought and 11 2' the number in the nonresponse group . By intensive efforts, the
data are later obtained from a random subsample 11 2 = " 211 2 ' out of the 11 2" Hansel
and Hurwitz use the notation 1/2 = 1/ k, so that 11 2 = 11 2 '/ k.
In the framework of section 12. 2 there are two strata. Stratum I consists ot
those who would respond to a first attempt, with a measured samplc o f size,
n 1= n I', so that "I = I . Stratum 2 consists of those who would respond to the '
second attempt, with 11 2 = 11 2'/ k . ,
The cost of taking the sample is l

' +c n ,+ C211 2' (13 . 12)

con I I k

where the c 's are the costs per unit: Co is the cost of making the first attempt. c I i~
the co t of processing the results from the first attempt. and C2 is the cost of getting
and processing the data in the second tratum . If WI ' W 2 are the population,
proportion in the two strata. the expected cost is

c. = con ' + CI W In ' + -

C2 W 2" '
-k- ( 1.1 . 13 )

As an estimate of Y. we take
(1.1 . 1 ~

where YI> Y2 are the means of the amples of si zes TIl = II I' and 11 2 = n 2' / k . By
theorem 12 .1, the estimate ji' will be unbiased if responses are obtained fwm all
the elected random subsample of size " 2 = 11 2'jk .
By formula (l2.3), the variance of ji' is
, (I 1) ,-+W,S2
V(y) = --,.-- -,- -1- 1)
-
2
(
11 N 11 1/ 2

= (!_!)S 2+ (k - ) W2 S/ (HI5)
11 ' N 11'

The quantities 11' and k are then chosen to minimize the product C( V + 05 2 / N) .
From (13.15) and (13.13) we have

V+S 2/N = (S 2_ W 2 S / ) + kW2 S/ ( 11. If»

n' ,, '

( 1J . 17)
 SAMPUNG TECHNIQUES

By Cauchy's inequality, the optimum k is

(13 .18)

The initial sample size n' may be chosen either to minimize C for pecified Vor
V for specified C by solving fol' n' from (13.16) or (13.17). If V is specified,
, N[5 2 +(k-1)W2 5/ ]
(13 .19)
n op1 = (NV + 52)

where V is the value specified for the variance of the estimated population mean .
The solutions require a knowledge of W 2 : this can often be estimated from
previous experience. In addition to 52, whosc value must be estimated in advance
in any "sample size" problem, the solutions also involve 5 22 , the variance in the
nonre ponse stratum. The value of 5 2 2 may be harder to predict ; it will probably
not be the same as 52. For i:1stance, in surveys made by mail of most kinds of
economic enterprise, the respondents tend to be larger operators, with larger
between-unit variances than the non respondents.
If W2 is not well known, a satisfactory approximation is to work out the value of
n ~PI from a provisional (13.18) and (13.19) for a range of assumed values of W2
between 0 and a safe upper limit. The maximum n~Plin this series is adopted as the
initial sample size n'. When the replies to the mail survey have been received, the
value of n2' is known . In seeking the value of k to be used with this method, we use
'f the variance vrCji') conditional on the known values of n2 and n'. This can be

,btained from (12.4) and 12.7) as

Vc('Y') = 52C, - ~) + (k - ~,~ 2'522 (13.20)

Equation (J 3.20) is solved to find the k that gives the desired conditional
variance. The cost for this method is usually only slightly higher than the optimum
cost for known W 2 •

Example. This example is condensed from the paper by Hansen and Hurwitz (1946).
The first sample is taken by mail and the response rate WI is expected to be 50%. The
precision desired is that which would be given by a simple random sample of size 1000 if
there were no nonresponse. The cost of mailing a questionnaire is 10 cents, and the cost of
processing the completed questionnaire is 40 cents. To carry out a personal interview costs
S4.10.
How many questionnaires should be sent out and what percentage of the nonrespon-
dents should be interviewed?
In terms of the cost function (13.12) the unit costs in dollars are as follows.
Co = cost of first attempt = 0.1
C I =cost of processing data for a respondent =0.4
C2 = cost of obtaining and processing data
for a nonrespondent = 4.5
 SOURCES OF ERROR IN SURVEYS

The optimum n' and k can be found from (l J .19) and (J 3. 18). If the variances 52 and 5/
are assumed equal and N is assumed to be large. then

/C 2(1- W2) / (4 .5)(0.5) r;;-;

k.",= V co+c, W, V.0.1 +(0.4)(0.5) = '0/7.5=2.739
5 2 [1 +(k -I)W J
n;"" = V 2.= IOOOil +(l.739)(0.5))

= 1870
Note that we have put 5 / V = 1000. or V "" 5 2 /1000, since this is the variance that the
2

sample mean would have if a sample of 1000 were taken and complete response were
obtained .
Consequentl). 1870 questionnaires l>hould be mailed . Of the 935 that are not returned.
we interview a random subsample of 935/2.739. or 341. The cost is S2095.

As Durbin (1954) ha~ pointed out, subsampling is unlikely to show a marked

profit unless C2 is large in relation to (co+c, WI) ' The two quantities are compara-
ble, since (co+c, WI) is the expected cost per unit of making the first attempt and
processing the results and C2 has the same meaning for the second attempt. From
the equations it can be shown that the ratio of the cost of obtaining a prescribed V
with k = 1 (no subsampling) to the minimum cost for optimum k is

S2(CO+C, W , +C2W 2) co+c, W, +C2 W7

[.J(S2- W 2 S 2 2 )(CO+ C, W ,)+v'c; W 2 S 2 ]2 [,lw, (co +c , Wt)+v'c; W2 ]2
if S2 and S/ are approximately equal. If r is the ratio uf C2 to (co + c 1 Wd, the cost
ratio becomes

For instance, the cost ratio is 1.029 for WI = 0.5 if r '" 4, 1.146 if r =
10, and 1.228 if r = 16. If, however, S 2 is substantially greater than S/" there is
more to be gained from subsampling.
With stratified sampling, the optimum values of the nh I and the kh in the
individual strata are rather compJex. A good approximation is to estimate first, by
the methods in sections 5.5 and~ . 9, the sample sizes nolt that would be required in
the strata if there were no nO)1response. Now, from (13.19), if W2 = 0, we have
NS 2
(13.21)
n" == NV+S2

Hence (13.19) call be written as

. _ [1 (k - 1) w2S/] (1 3.22)
""pt - n" + 52
 ·374 SAMPLING TECHNIQUES

This equation. applied separately to each stratum, gives an approximation to

the optimum nh '. The values of kh are found by applying (13.18) in each stratum.
These techniques can be used with ratio or regression estimates. With the ratio
estimate, the quantities S2 and S22 are replaced by S/ and S2/, where d l =
YI - RXj. With a regression estimate, S2 becomes S2(1 - p 2) and S22 becomes
S/ (l_p 2).

13.7 ADJUSTMENfS FOR BIAS WITHOUT CALL-BACKS

An ingenious method of diminishing the biases present in the results of the first
call was suggested by HartJey (1946) and developed by Politz and Simmons (1949.
1950) and Simmons (1954). Suppose that 311 calls are made during the evening on
the six week-nights. The respondent is asked whether he was at home, at the time
of the interview, on each of the five preceding weeknights. U the respondent states
that he was at home t nights out of five, the ratio (t + 1)/6 is taken as an estimate of
the frequency TT with which he is at home during interviewing hours .
The resuJts from the first call are sorted into six groups according to the v31ue of
t, (0, 1, 2,3,4, 5). 1n the tth group let n, be the number of interviews obtained and
y" the item mean. The Politz-Simmons estimate of the population mean II- is
s
_ L 6n,y,/(t + 1)
,-0
YPS = '---'-5- - - - (13.23)
·f
L 6n,/(t+ 1)
,- 0
This approach recognizes that the first calJ results are unduly weighted with
persons who are at home most of the time. Since a person who is at home, on the
average, a proportion TT of the time has a relative chance TT of appearing in the
sample, his response should receive a weight 1/ TT. The quantity 6/ (I + 1) is used as
an estimate of l/TT. Thus Yps is less biased than the sample mean y from the first
call, but its variance is greater, because an unweighted mean is replaced by a
weighted mean with estimated weights.
In presenting the mean and variance of YPS, we use the notation of section 13.5.
The population is divided into classes, people in the jth class being at home a
fraction TTl of the time. Note that the tth group (i.e., persons at home t nights out of
the preceding five) will contain persons {rom various classes. Let n/" Yj, be the
number and the item mean for those in class j and group t. Then YPS may be
written as follows.
_ _ LL 6njtYj,/(t+ 1) N ( )
(13.24)
YPS - L L 6nj,/(I+ 1) D say

This is a ratio type of estimate. In large samples its mean is approximately

E(N)/E(D) .
 SOURCES OF ERROR IN SURVEYS 375
If no is the initial size of sample (responses plus not-at-homes) and nj is the
number from class j who are interviewed, the following assumptions are made.

(i) !!.J.. is a binomial estimate of P/1Tj

no

(ii) B( nj' InJ)= nj 1!(55!_ t)! '7Tj '(I - '7Tj

)5- '

(iii) B(Yj') = J.Lj, for any j and 1

Assumption (ii) is open to question. Without giving a detailed discussion, it

assumes that people report correctly whether they were at home.
For given j, using assumption (ii),

E,-
f. (6)
,- I nj' --1
t+
;0
f. (
,- I -
nj '-
t+
-1
6) t.1(55!_ t . '(
)11T"j } - '7TJ)
5- , (13.25)

(13.26)

Hence

(13.27)

uSing assumption (i). Furthermore, since B(y},) = J.L} for any j and I, this gives the
result

(13.28)

j- l
t pj[l - (1- '7Tj )6]

Since the true mean ji = L PjJ.LJ> some bias remains in yps. In a certain sense, this
estimate bas the same bias as Y6, the sample mean given by the call-back method
with a requirement that as many as six calls be made if necessary. In ection 13.5,
equation (13.9), it was shown that the call-back method, with a total of j caUs,
gives an unbiased estimate of jijl= :E WljPjJ.LJi:E WljPj, where WIJ is the probability
that a person in class j who falls in.the sample will be interviewed. Now Wl j:=; '7TJ' If
at subsequent calls the probability of finding at home a person not previously
reached remains at '7Tj, then
Wlj = [1 - (1- '7Tj)I]

so that ilps = ji6' However, with the call-back method the probability of an
interview at a later call may be greater than 'TTj as a result of information obtained
376 SAMPLING TECHNIQUES

by the interviewer at the first or earlier calls. In this event the call-back method has
less bias after six calls.
The variance of YPS is rather complicated. With the usual approximation for a
ratio estimate, it may be expressed, following D ming (1953), as

V(yPs)= IUa: ,",pjBj[u/+(l1-j _~ps)2]

no
+ (no - 1) L (7T;pi(Bj - A/)(I1-; - tLps)2)
where

Bf = ,t [(1 ~t)rt!(55~t)! 7T/(l-7Tj)S-'

Although this expression is difficult to appraise without applying it to specific

populations, two comments can be made. If the 11-/ do not differ greatly, that is, if
the bias from first calls is moderate, the dominating term is the first.

This expression tends to be 25 to 35% higher than the variance of the unweighted
mean of the first calls. Also, V(yps) contains a term that does not decrease as no
increases and becomes important in very large samples.
To summarize, comparisons made on simulated populations by Deming (1953),
Durbin (1954), and myself suggest that this method shows to best advantage, in
relation to call-backs, when the biases from early calls are substantial and the
sample is large. The reductions in MSE for the same outlay are small, however,
unless call-backs cost substantially more than postulated here. The Politz-
Simmons technique has the advantage of sliving time. Errors and incompleteness
in the values of t, not considered in the analysis, are a disadvantage. The method
may also be applied, as suggested by Simmons (1954), in conjunction with several
call-backs.
Several other methods for mitigating the "not-at-home" bias have been
proposed. Bartholomew's (1961) applies to a survey with two calls. He supposes
that, for those not at home on the first call, the interviewer, by careful inquiry, can
make the probability of finding them on the second call approximately equal. If
this is so, the n2 persons interviewed at the second call are a random subsample of
the (no - nl) persons missed at the first call. Hence [n 191 + (no - n I)Y2]/ no is an
unbiased estimate of the mean of the initial target sample .. The method worked
weU on some British surveys to which Bartholomew applied it. In repeated
 SOURCES OF ERROR IN SURVEYS 377
surveys Kish and Hess (1959a) suggest that nonresponses from recent surveys
may serve as a repla~ment for nonresponses in a current survey. Wherever the
bias from early calls shows a systematic pattern, as in Table 13.1, Hendricks
(1949) has outlined extrapolation methods to estimate the average results that
would be given by nonrespondents.

13.8 A MATHEMAllCAL MODEL FOR ERRORS OF

MEASUREMENT
Conceptually, we can imagine that a large number of independent repetitions
of the measurement on the ith unit are possible. Let Ylu be the value obtained in
the ath repetition. Then
(13.29)
where ILl = correct value
elo ::: error of measurement.
The idea of a " correct value" requires a little discussion. With some items the
concept is simple and concrete. For instance, in an inventory taken by sampling,
the correct value may be the number of fan belts lying on a shelf at 12 noon on a
specified day. In some cases the correct value can be defined operationally. A
person's correct diastolic blood pressure at a specified time might be defined as the
value obtained when it is measured by a certain standard instrument under
carefully prescribed conditions. We may realize, however, that our standard
instrument is itself subject to errors of measurement, and we may expect that in
course of time a more precise instrument will be developed. With other items, for
instance, some aspect of an employee's attitude toward his employer or of a
person's feelings of ability to cope with his day-to-day problems, nobody may
claim to have a satisfactory method of measuring the "correct value." Neverthe-
less, the concept is useful even in such cases.
Under repeated measurements of the same unit, the errors flu will follow a
freq uency distribution. For the ith unit, let elu have mean PI and variance o/. The
term PI represents a bias in the measurements. The magnitudes of P, and u/ will,
of course, depend on the nature of the item being measured aDd on the measuring
instrument. They may depend also on numerous other factors. With human
populations the prevailing economic and political climate and the amount and
type of advance publicity received by the survey may influence the responses to
the questionnaire.
The next step is to consider how the errors of measurement change when we
move from one unit to another. Various complications can arise.
For the bias component P" there may be a constant bias, say, E(/3I) = p, that
affects all units in the population. There will also be a component (PI - P) that
follows a frequency distribution over the population. This component may be
 378 SAMPLING TECHNIQUES

correlated with the correct value 1'-1 ; for instance, the measuring device may
consistently underestimate high values of 1'-; and overestimate low values.
There may be a correlation between the values of tit> on different units in the
same sample. The simplest example is the "interviewer bias." Dramatic differ-
ences are sometimes found in the mean values of Ylo obtained by different
interviewers who are sampling comparable parts of the same population (see
Lienau, 1941 , Mahalanobis, 1946, and Barr, 1957).
A similar effect has appeared when samples of a growing crop are cut by
different teams and when chemical or biological analyses are done in different
laboratories. The human factor is not the only cause for correlations among units
that are measured at about the same time. Many measuring processes are affected
by the weather; some use raw materials whose quality varies from batch to batch.
In estimating the current sale price of homes built some years ago, Hansen,
Hurwitz, and Bershad (1961) point out that if some houses in the sample have
been sold recently their prices establish a level that guides the interviewer and the
household~r in assigning values to houses that have oot been sold for many years.
In fact, the average price recorded for the sample may depend on the order in
which the recently sold houses appear in the sample.
In orde.r to handle these intrasample correlations in their most general terms, a
more complex model than that presented here is required. In particular, the
notation for elo and PI would have to indicate that their values may depend on the
other units present in the sample However, the types of correlation that are
believed to be most common in practice can be represented by the present model
br by simple extensions of it.
The components of the error of measurement are summarized in Table 13.11.
We have noted further that values of Pi and dlo on different units in the same
sample may be correlated with one another, where dlo - t;t> - PI'

TABLE 13.11
CoMPONENTS OF THE ERROR OF MEASUREMENT ON THE ith UNIT

Symbol Nature of Component

Constant bias over all units

Variable component of bias, which follows some frequency
distribution with mean zero as i varies and may be
correlatl:d with the correct value I-' {
Fluctuating component of error, which follows some
frequency distribution with mean zero and variance
ul as tlt varies for fixed i

Models in general similar to the preceding have been developed by Hansen

et at (1951), Su.lchatme and Seth (-1952), Hansen, Hurwitz, and Bershad (1961),
 SOURCES OF ERROR IN SURVEYS 379
and Hansen, Hurwitz. and Pritzker (1965) . FeIJegi (1964) has given a more
extensive model tbat includes numerol}s cro s correlations that may exist between
different sources of measurement error in practical situations.
In their 1961 and 1965 papers, Hansen et al. expressed our model in slightly
different terminology, including adding a subscript G to all variables as a reminder
that the errors and their sizes may depend on the general (G) conditions of the
survey. They express results in terms of the du. and of
p..;' =E(Yi",!i) = p..i +{3j (13.30)
The quantity p.., ' (which they denote by 1";, or PI with a proportion) is
conceptually the average obtained from many repetitions of the measuring
process on unit i. Hence
(13.31)
They call d'a the response deviation on unit i, whereas we have called it the
fluctuating component of the measurement error. Thus
YIQ - f.~ = du. +(p..;' - p.. ') + (p..' - p..) ( 13.32)
where p.. ' is the population mean of the J.L;'.
Averaging over the sample, we have
Yu - p.. = dn+ (ii' - p.. ') + (J.L ' - p.. ) (13.33)
This gives the formula
MS (Yo) = V (do )+ V(ii')+(p..'- p.. /+2 cov (do, ii' ) (13.34)
For the sample mean Ya, the terms on the right are called, respectively, the
respon e variance, the sampling variance, the quare of the overall bias, and twice
the covariance between the sample's average response deviation and sampling
error. Since E(d;a Ii) = 0 under our model, this covariance vanishcl: under repeti-
tions of the measuring process on the same set of sample units in the same order. It
might not vanish under repetitions over different samples or different orders if the
dlo are affected by the other units in the sample. Although this term will be ignored
here for simplicity, Fellegi (1964) has shown in a Canadian study that it may
materially bias some methods of estimating components of the response variance
V(do ). .
Using Cornfield's approach, as described ;n section 2.9, Koch (1973) has given a
general decomposition of tbe MSE of the estimator in multivariate sample
surveys, with applications to subclass means.

13.9 EFFECI'S OF CONSTANT BIAS

Suppose that the measurements YI on aU units are subject to h constant bias f3
whose magnitude is unknown . Then the mean y f a simple r udom sample is also
 380 SAMPLING TECHNIQUES

subject to bias {3. In the estimated error variance, which we attach to the sample
mean, the bias cancels out, since this estimate is derived from a sum of squares of
tt:rms (y/- Y)2 . Consequently, the usual computation of confidence limits for Y
from the sample data takes no account of the bias. The same results hold in
stratified random sampling.
The situation is essentially the same with regression and ratio estimates.
Consider the regression estimate
5;" = y+b(X-i)
where both the Yi and the Xi may be subject to constant bia es {3y and {3~ ,
rt:spectively. Since the least squares estimate b remains unchanged and since the
bias {3x cancels out of the term (.Y' - i) , it follows that YI, is subject to a bias (3y . It is
easy to verify tpat the sample c"timate of V(YI,) contains no contribution due to
the biases
With the ratio e timate
Y-
YR =-::: X
x
the bias is <llso [3)" to a first approximation , since in large samples E(Xji) is
approximaH.ly 1 even if tht Xj are subject to a constant bias. In large samples the
sample esti'nate of varian.;e
(-) (N- n) 1: (Yj-RX/) 2 (13.35)
., V YR =-~ n -1

will be aJJ;1OSt free ff'om bias as an estimate of

E(YR - Y)2
that is, as a'; (;o;timate of the variance about the biased mean Y.
To summari /.e. a constant bias passes undetected by the sample data. As we
have seen (sc 'tion 1.8), the 95% confidence probabilities are almost unaffected if
the ratio of (3y to the standard error of the estimated mean is less than 0.1 but, as
the ratio increases beyond this value, the computation of confidence limits
bec.:ome misleaumg. Estimates of change from one time period to another, or
from one stratum to another, remain unbiased, provided that the bias is constant
throughout.

J3.10 EFFECfS OF ERRORS TIlAT ARE

UNCORRELATED WI11IIN TIlE SAMPLE
In this section a few results about V(Ya ) and v(Ya) are given for simple random
sampling in the simplest situation in which errors of measurement are uncorre-
lated within the sample. This situation may apply in surveys taken from records, in
 OURCES OF ERROR IN SURVEYS 381
self-filled questionnaires (as in mail surveys) in which members of the same
sample do not consult one another, and in some surveys of inanimate populations
in which the measurement is objective. The assumption of no correlation cannot
be made lightly: intrasample correlation may enter, for example, in interviewing,
editing, coding, and transferring the data to the computer if the same person
handles a number of elements in the sample.
From (13.34) we get, for the variance,
(13.36)
In finding variances we average first over repetitions of the measuring process
on a specific sample and then over different simple random samples. Now
V(dlo ) = 0/ and errors are uncorrelated on different units within the sample.
Hence, for simple random sampling, where S denote a specific sample,

E(d" 21S) = -\
n
f 0/
i
(13 .37)

When we average subsequently over all simple random samples, we have

N
1 N I _t L (jJ,, '- J.l..)2
V(Y")= -nNLO/+-
, n I N- 1 (13.38)

(13.39)

where (Td 2 denotes the population average of the variances of the errors of
measurement. In the Hansen et al. terminology, (Td 2 / /I is the response variance of
the sample estimate y".
With uncorrelated errors the same model can be applied in estimating a
population proportion P (Hansen , Hurwitz, and Bershad, 196] ). For any unit, let
the correct value J.l..i be 1 if the unit is in class C and zero otherwise. If errors of
measurement occur, this implies that units are sometimes incorrcctly classified.
For the ith unit, the recorded value YI" is sometimes I , sometimes O. Let p. denote
the proportion of measurements on the ith unit for which Y,a = 1. 111en, for given i,
YI" is a binomial variate in repeated measurement, with mean J.l..I ' = P; while the
variance of dlo is PiQ,. Hence, if Po = Yell is the sample estimate, (13.38) becomes
N
V(P ) = V( - ) = _1 ~ p.Q + (1 -
" Yo nNL..' I n
n L N(Pi- - 1p )2 (13.40)

1 (N N N )
<n(N-1) LP;-LP/+LP/-Nr

N PO (13.41)
n(N-1 )
 382 SAMPLING TECHNIQUES

where
N
p='IP,
N
As (13.41) shows, the sum of the response variance and the sampling variance
has an upper limit NPO/n(N - l). This upper limit is also the variance of the
sample mean of the correct measurements on the units if the fpc is ignored and
there is no overall bias. For the correct measurement ILi on any unit is then a
binomial variate with mean P, so that
S 2 N
~ = n(N _1/0 2:: V(p.. ) (13.42)

This rather puzzling result holds because (i) the sampling variance entering into
(13.39) and (13 .40) is that of IL,' = (ILl + (3,) and (ii) in the e timation of a
proportion, ILl and {3i are always negatively correlated. When IL, = I , {3, :s; 0, since
Pi = (ILl + {3/) :S; 1, and similarly when ILi = 0, f3i 2:: 0. Thus the term
S
=e_
,2

n
in (13 .39) is always less than
s;
::L
n
by just about the response variance of Ya'
With uncorrelated errors, a useful result is that the usual formula for v (y", ) in
simple random sampling remains unbiased if the fpc is negligible. From theorem
2.4 corollary, this formula, developed by assuming no errors of measurement, is

v(Yo) = 1- / 5 2 = I - If (Yia - y"l (13 .43)

n n (n - 1)
Now
(13.44)
Squaring and averaging first over repeated measurements and then over
repeated sample selections, we obtain

E(
-)
V Yo =
1-/
-
2 1- / 2
-Ud +- -S.. , (13.45)
n n
while, from (13.39)

V( y_.. )_1
- -Ud
2+(1
--
- /)S ,2 (13.46)
n n"
Thus Ev(yo) = V(y.. ) if f is negligible.
 SOURCES OF E.RROR IN SURVEYS 383
in the same way the formulas in preceding chapters for the sample estimates of
sampling error variances can be shown to remain valid in stratified and multistage
sampling, as do the large-sample formulas applicable to ratio and regression
estimates, provided that errors of measurement in Yu. and Xu. are uncorrelated
within the sample and that the fpc's are negligible.

13.11 EFFECI'S OF INTRASAMPLE CORRELAnON

BElWEEN ERRORS OF MEASUREMENT
Suppose that some or all of the values of du. for units in the same sample are
correlated. The term d,/ can be written

(13.47)

Hence, averaging over repeated measurements and simple random samples,

(13.48)

where the products are averaged over all pairs of units in the same sample. By
analogy with cluster sampling, the average intrasample correlation coefficient p...,
may be defined by the equation
E (du. djo)= Pwu./ (13 .49)
This gives, from (13.48),
2
- Ud
V(d., ) = -[1 +(n - 1)pw] (13.50)
n
Hansen, Hurwitz, and Bershad (1 961) have calJed V(da ) the total response
variance as it affects the sample mean. Its component u/ / n is called the simple
response variance, while the term (n - 1)P..u/ / n is the co"elated component of
the total response variance.
From (13.34) we get, assuming cov (da • ji.') =0,
_
V(y.,) = -
1- 1 ·2
u/[ 1+(n - 1)p", ] +-S,. (13.51)
n n
The average value of the usual v(Ya) in (13.43) is found in the same way to be

(13.52)

Since P.. is likely to be positive for many types of measurement error, the standard
formula for v(Y.. ) is usually an underestimate in this case and makes the sample
estimate appear more precise than it is. This is true even when the fpc is negligible.
384 SAMPLING TECHNIQUES

Perhaps the most frequent example of intrasample correlation between errors

of measurement is the intrainterviewer correlation previously mentioned, par-
ticularly on questions involving opinions and judgment. Suppose that n = mk, and
that each of the k interviewers obtains data from m respondents. U we can assume
that there is no correlation between the errors of measurement of different
interviewers,
- u/
Veda) = -(1 +(m - l)pw] (13.53)
n
where p", is the average intrainterviewer correlation coefficient.
Even a small Pw may make a major contribution to V(Ya) , since it is multiplied by
roughly the size m of the interviewer's assignment. There may also be some
correlation between errors of measurement for different interviewers (e.g., if they
have been trained or directed by the same supervisor).
This model represents only the simplest type of intrasample correlation. With
stratified sampling, for instance, a coder may process results from several strata
and through a misunderstanding of instructions may introduce correlated errors
that extend over the strata. The mathematical model can be adapted to apply to
situations of this type.

13.12 SUMMARY OF THE EFFECI'S OF ERRORS

OF MEASUREMENT
In terms of the model, the mean Y of a simple random sample would be
unbiased, with varianOi. ::-", ~ In (ignoring the fpc). if all measurements were fully
accurate. As a result of the types of errors of measurement discussed here, the
mean may be subject to a bias of amount (3, and its mean square error is
_ 1
MSE(yo) = _ {S", ,2 + u/[l + (n - t)p",]}+ {3 2 (13.54)
n
where JL/ :::: ILl + (3/.
Formula 13.54 contains two terms, S", ,2I nand u/ (1 - Pw)/ n, that decrease as
lin. The remaining two terms, p",u/ and {3 2, appear at first sight to be indepen-
dent of n. This is probably an oversimplification. Any material change in the size
of sample may require a change in the field methods of measurement, and this may
affect Pw and {3 2. However, these two terms should change relatively slowly, if at
all, with n. Thus. in large samples, the MSE is likely to be dominated by these two
term , the ordinary sampling variance becoming unimportant and misleading as a
guide to the real accuracy of the results.

13.13 THE STUDY OF ERRORS OF MEASUREMENT

In recent years much of the research on sampling practice has been devoted to
the study of errors of measurement. The objectives are to discover the compo-
 SOURCES OF ERROR IN SURVEYS 385
nents that make a large contribution to the MSE and to find ways of decreasing
these contributions. Some of the principal methods are described in this and the
following sections. It is already clear that progress will be slow and expensive. One
reason is that, as already mentioned, the measurement errors depend intimately
botb on the items and on the measuring process. Results about measurement
errors found in one survey can seldom be assumed to apply to other surveys.
Ideally, the best method of studying errors of measurement is to obtain the
correct values #J.I' In practice. this approach is limited to items for which a feasible
method of finding #J.I exists and by problems of expense and execution. Examples
are given by Belloc (1954), who compared data on hospitalization as reported in
household interviews with the hospital records for the individual, and by Gray
(1955), who compared employees' statements of sick leave with the personnel
office records. Checks of this type~ometimes called "record-checks"-are
possible with items such as age, occupation, number of years of schooling, and
price paid for car. One difficulty is that sometimes the records contain no exact
match of the person interviewed.
Failing a method of determining the correct value, an alternative is to remeas-
ure by an independent method hat is considered more accurate. Kish and Lansing
(1954) engaged professional appraisers to estimate the selling prices of homes that
had already been reported by the home owners. In surveys of illness respondents'
replies hive been compared either with doctors' records on the respondents or
with the re ults of a complete medical examination [Sagen, Dunham, and Sim-
mons, 1959, Trussell and Elinson, 1959). The results of such comparisons are not
easy to interpret in terms of the model, since the superior instrument is itself
subject to measurement errors, but the comparisons will at least indicate the items
for which the routine instrument agrees well with the superior instrument and
those for which It does not.
In household interview surveys, a method with a similar purpose is to reinter-
view a subsample of the respondents by more expert inttrviewers, the question-
naire being .more detailed and probing. After the reinterview. the expert discusses
with the respondent any discrepancies between the original and the repeat
answers, the objectives being to determine the most accurate answer and the
reason for the discrepancy. Much useful information, may be gained. In pre-
nting some devices for the measurement of response errors, Madow (1965) has
discussed the use of double sampling, with a difference estimator of the form
[ji' - (ji -,I)] as a means of reducing response bias. · Here, ,I is the mean of
unbiased or less biased measurements made on the sub ample, while ji' and ji are
means of the original measurements on the sample and subsample.
Occasionally, overall comparisons between the results of two different surveys
are feasible. For a number of items, the results of the U.S . . Census can be
compared with those given by the Current Population Survey taken at the same
time. Since the Survey is considered more accurate, particuJarly for items difficult
to measure, rough estimates of the measurement bias fJ in the Census data can be
386 SAMPUNG TECHNIQUES

made (Hansen, Hurwitz, and Bershati, 1961). A number of comparisons between

the re ults of quota samples and probability samples are discussed by Stephan and
McCarthy (1958).
In the following ections we consider some methods designed to produce
quantitative estimates of the components of the total variance (response variance
plus sampling variance) of a sample estimate.

13.14 REPEATED MEASUREMENT OF SUBSAMPLES

In recent studies, interest has centered on estimating: (a) the total response
variance and the relative sizes of the simple response variance and the correlated
component as contributors to it, and (b) the relative sizes of the total response
variance and the sampling variance. For (a) a common method is to select a
subsample of the measuring agents (interviewers, coders etc.) and remeasure
their assignments by second agents presumably of the same skill. Suppose mk is
the size of the subsample, m being the size of an agent's assignment and k the
number of agents chosen from the original sample.
Let Yn, YI2 be the two measurements on the ith unit of an agent's assignment.
If (13.31) holds,
(13.55)

Hence, averaging over the assignment,

'f

(13.56)

This expre sion estimates Ud 12, the simple response variance for agent 1 in the
survey, if two conditions (A) hold: no correlation between response errors dill dl2
on the same unit; and Ud/=Ud2 2 • Equation (13.56) applies to a single pair of
agents and is averaged over the subs ample of k pairs of agents ,
Conditions A may hold when the mea urement is coding, the agents being
coders of similar skill trained by different supervisors, neither coder eeing t~
otber's work. With interviewing, a positive cov (dll d12 ) is to be expected because
some respondents repeat a first incorrect response from memory. In this event
I (Y11 - Yi2f 12m underestimates Ud 12. It also underestimates Ud,2 if the second
agent is more skilled than the first, as shown for a (0,1) measurement by Hansen,
Hurwitz, and Pritzker (1965). Moreover, L (YI1 - Y12)2/2m has been found to
decljne if the second interviewer is given the response obtained by the first
interviewer, even if told not to look at them until the repeat interview is completed
(Koon , 1973). Thes compLexitie illustrate why the realistic study of errors f
measurement is difficult.
 SOURCES Of EIlROIl IN SURVEYS 387
For the total response variance the relevant estimate from (13.55) for a single
pair of agents is <Y.I- .9.2)2/2, where Y.I, 9.2 are the means of the m first and second
measurements. Under conditions (A).

E(Y.I;y'z)2 :;:\1+(m-1)pwl (13.57)

where Pw is the correlation between re ponse errors on different units by the same
agent. Equation (13.57) provide only a s!ngle degree of freedom, but is averaged
over the k pairs of agents. Havi ng an estimate of O'd 2 from (13.56), we can estimate
the relative size of the imple respon. e variance and the correlated component.
In interview survey the correlated component is usually found much larger than
the simple re pon e variance except for basic items such as age, sex, and marital
status (Fellegi, 1964). Partly for this reason the 1970 U.S. Census used self-
enumeration by mail extensively (Hansen and Waksberg. 1970).
If (13.55) holds, we can also study the ratio of the simple response variance to
the sampling variance that applies to the sample in an agent's assignment. For
under conditions A,
z '"
E 2:I (y,,, - j,,,,)2/2(m - 1) = u/(l-p",)+S,/ (13.58)
" i

Pritzker and Hanson (J 962) have called the ratio

u/ (13.58)'
1= 2 2
(Ud + S.. · )
the index of inCOflSistency. It is analogous to the quantity (1- c/» , where c/> is the
coefficient of reliability used in studying errors of measurement in p ychology. If
Pw is negligible, I can be estimated from the ratio of (13.56) to (13.58).
With a (0, 1) variate, equations (13.40) and 13.4]) in eetion 13.10, putting
n = 1, show that for a single measurement of a single unit, the sum of the simple
response variance and the sampling variance is V(Yla)= PQ, where P = I P;/N,
assuming! negligible. The set of joint values (YIl, Yll) obtained by the two agents
can be summarized in the following 2 x 2 frequency table .

Nurnberofre pon e
Second agent
1 0 Total

First 1 a b a +b
agent 0 c d c+d

a +c b+d m

Thus, b is the number of units on which the first agent record 1, the second
 388 SAMPUNG TEOINIQUES

agent O. Under conditions A

'" 2
L (YII- Y,2)
,;'/ = I =: (b +c)/2m (13.59:
2m
As an estimate of I from the sub ample, one choice is
.. ,;/ m(b+c)
(13.60;
I = PQ = (a + b)(c +d)+(a +c)(b +d)
.which averages the PO estimates from the two sets"Of measurements. Estimates 01
the index have been published for both Census and sample survey items b )
Pritzker and Hanson (1962), FeUegi (1964), and Koons (1973). These estimat !
are useful in comparing the relative unreliabilities of measurement for differeD1
items, in successive censuses, or as between different methods of measurement,
thus providing some appraisal of new methods of mea ·urement.
The interpretation of these comparisons is, of course, often clouded by doubts
as to whether conditions A apply. More complex estimates based on more
realistic assumptions are given by FeUegi (1964).

13.1S INTERPENETRA11NG SUBSAMPLES

This technique, particularly useful for the study of correlated errors, was
proposed by Mahalanobis (1946). To present is in the simplest terms, a random
sample of n units is divided at random into Ie ubsamples, each subsample
containing m = n/ k units. The field WOK and processing of the sample are
planned so that there is no correlation between the errors of measurement of any
two units in different subsamples. For instance, suppo e that the correlation with
which we have to deal arises solely from biases of the interviewers. If each of k
interviewers is assigned to a different subsample and if there is no correlation
between errors of measurement for different interviewers, we have an example of
the technique.
With the same mathematical model, it is convenient to label the units by double
subscripts. Let
Yip = p.'q+d,p (13 .61 )
where i denotes the subsample (interviewer) and i the member within the
subsample. The fpc is ignored.
Since the ith subsample is a random subsample, it is itself a simple random
sample of size m. Hence, by (13.51), the variance of its mean is

V(Yu.) =..!.{S,/ + O'd2[1 + (m - 1)p.. ]) (13.62)

m
where P.. is the correlation between the diP obtained by the same interviewer'l
 SOURCES Of ERROR IN SURVEYS 389
Since errors are independent in the different subsarnples,

V(Ya)=~ V(j,a) =;{S,/+U/[1 +(m -1)p.. ]) (13.63)

From the sample results we can compute an analysis of variance of the km

observations into the components "Between interviewers (sub amples)" with
(k -1) degrees of freedom and "Within interviewers" with Ic(m -1) degrees of
freedom. It is easy to verify that the expected values of the mean squares,
averaged over selections of the: interviewers and of the random samples and
subsamples, work out as in Table 13.12.

TARLE 13.12
ExpECTAnONS Of THE MEAN SQUARES (ON A SINGLE-UNrr BASIS)
df ms E(ms)

Between
interviewers
(subsampJes)
k -l S,/+ (1'<lZCl +(m -l)p.. J

Within
interviewers
k(m-1) of 2 L2:(Ylloo-Y"')~ S,/ + (1'/(1 - P.. )
.. k(m-l)

Table 13.12 contains two importantresuJts. By comparison with (13.63) we see

that s//n is an unbiased estimate of V(Ya), ignoring the fpc. Thus interpenetrat-
ing subsamples provide an estimate of V(Ya) that takes proper account of both the
simiJle response variance and the correlated component.
The analysis also enables us to estimate the correlated component, since

(13.64)

Consequently, comparison of ( '" - 1) (Sb' - s,.2)/ m with $b2 estimate the rela-
tive amount which the correlat ··1 oompon nt of the respon e variance contributes
to the total variance of Y•. Witl. measurements in which the correlated com~nent
is much larger than the simple el.ponse 'ariance, the ratio (m -1){S,,2- sw )/ms/
has been used alternatively as a me sure of the relative contributiou of the total
response variance to the total variance of o. Tepping and Boland (1972) present
estimates of this ratio for items in the Current Population Survey.
When the interpenetration method is a plied in a multistage sample covering a
wide geographic area, the most common practice is to have pairs of interviewers
390 SAMPLING TECHNIQUES

measure interpenetrating subsamples drawn from the smallest clusters among the
succes ive stages. In this way the number of ultimate units in an interviewer's
assignment is kept at its customary level in the survey, although the interviewer
has to travel over twice the usual area. For a single cluster, Table 13.12 provides
1 df between interviewers and 2(m -1) df within interviewers: the carre ponding
mean square are averaged over the c cluster cho en for the study. The sampling
variance that is measured i , of course, only that within the last stage of clustering.
As a reminder of this fact , the sampling variance term in E(ms) is sometimes
written 5,/( 1 - P.) instead of 5,}, where P. is the intracluster correlation among
the sampling errors for different subunits.
The interpenetration method was used in this form in Response Variance Study
I by the U.S. Census Bureau (1968), designed to estimate the correlated compo-
nents of the total response variances of items in the 1960 Census. The areas in this
study were compact clusters of households, the clu ters being scattered all over
the U.S.A. In any sample duster, two interpenetrating subsamples were formed ,
each subsampJe being assigned to a different interviewer.
In halt the cl~ the two interviewer had different crew leaders. In this half it
wa a umed, as seem reasonable,.that the response errors of the two interview-
ers were uncorrelated. Thus (s/-s..,2)/m estimates Pwu/, where s,,2, is now the
average mean square between interviewers in the same cluster. In the other half of
the clusters the two interviewers had the same crew leader. The objective here
was to measure the extent to which "crew leader effect" induced a covariance
between d.l and d.l for the two interviewers in a cluster. If 0, Sb 2 in Table 13.12
now estimates
2 1 - -
5,., (l-P.)+O"II [1 +(m - l)p..,] - mE cov (d.I d.2 )
and (Sb l_ s,/)/ m estimates
(13.65)
Comparison of the two sets of values of (s,.2- s..,2)/m revtals the presence of a
"crew leader effect." Since differences between estimates of variance like Sb 2 nad
2
sw are unstable, large numbers of clusters in the two halves of this type of study
are necessary to measure "crew leader effect" with any precision.
The interpenetration technique extends to stratified and multistage sampling. If
the primary intere t is in an unbiased estimate of V(ji.,) that takes proper account
of the effects of errors of measurement, aU that is necessary is that the sample
consist of a number of subsamples of the same structure in which we are sure that
errors of measurement are independent in different subsamples. Strictly, tbi
requires that different interviewing teams, supervisors, and data processor be
used in different subsamples. If ji,.. is the mean of the ith subsample, the quantity
l:(YIa - y.}2/k(k -1} is an unbiased estimate of V(y.. ), with (k -1) df. This result
holds because the subsample can be regarded as a ingle complex sampling unit,
the sample being in effect II. s~ple random sample of these complex units, with
 SOURCES O · ERROR tN SURVEYS 391
uncorrelated errors of measurement between different complex units. Conse-
quently, the results in section 13.10 apply.
Numerou applications of thi method, sometimes called replicated sampling,
are described by Deming (1960), who has u ed the method extensively. For other
discussions of its advantages, see Jones (1955) and Koop (1960). Travel costs of
interviewers are increa ed by interpenetration, but this can be mitigated if the
ample i~ stratified into compact areas. For instance, each stratum might contain
two random samples, assigned to a different interviewer. Each interviewer is
required to travel over the whole stratum in tead of over only half the stratum.
Every stratum provides 1 df for the t:stimate of V(Yo) .

13.16 COMBINATION OF INTERPENETRATION

AND REPEATED MEASUREMENT
As we lJ:lv£' seen (Seclion 13.14), repeated measurement of the units in an
agent's a signment by another agent of similar quality provides estimate of the
simple response variance and the total response variance if cundjtions A apply,
although it may underestimate in tr:terview surveys if the respondent's errors on
the two occasions are positively correlated. The interpenetration scheme (section
13.15) prov.des estimates of the correlated component of the response variance
lind its contribution to the total variance, re spon ~e variance plus sampling
variance .
Much more may be learned from an IDgeniou combination of interpenetration
and repetition, as used by ellegi (J 964) in a tudy of response errors in the 1961
Canadian Census of Population. The study wa conducted in 134 Enumeration
Areas (E.A. 's), each containing about J50 hou eholds, the size of an interviewer's
assignment. Contiguous E. A.' were grouped into 67 pairs. Two interviewers
were assigned to each pair, each interviewing a random half of the household in
the pair. Thus each enumerator had the regular work load, but spread over twice
the area. Then the assignments of the two interviewers w re switched, giving the
desired combi nation of interpenetration and repeated measurement.
If S"S2 'denote the two interpenetrating subsamples and II> 12 , the two
interviewers, comparison of (ItSI) with (12S1) or (/IS2) with (12S2) gives the
"repeated mea urement" analysis, while comparison of (J IS.) with (1 2 S2 ) or (lIS2)
with (/'},SI) gives the "interpenetration" analysi . These comparison lead to
estimates of the simple response variunce, the correlated component, the total
response variance, and the index of incon tsteDey. The sampling variance involved
is that between' households within pair of E.A. 'so More extensive analysis by
Fellegi also estimates the covariance between sampling.and respon e deviation
for the same interviewer. This term has been neglected in the model presented
here, but FeUegi hows that it may create sizable biases in the estimates of {JUJ ~
A good exposition of th strengths and weaknesses of different variants of the
repeated measurement and the in rpenetration approaches ha en given by
392 S MPUNG TECHNIQUES

BaUar and Daieruus (1969). Hansen and Waksberg (1970) review the re~arch
work of the U.S. Census Bureau on measurement errors as they affect the Census
and some of the most important sample survey takell by the Census Btl eau.
Resulting changes in the 1970 Census included more wide pread use of self-
enumeration (omitting interviewer biases) 10 a Census by mail, further use of
sampling as distinct from a complete Census, and advunce computer election of
the sample to avoid some biases that had been delt'l."ted in selection of the
final-stage sample by the interviewer. Disturbing errors found in data on occupa-
tion, industry, and housing quality as well a recall problems in expenditure data
are under continuing study.

13.17 SENSI11VE QUESTIONS: RANDOMIZED RESPONSES

A situation likely to lead either to refusals to answer or to eva.<;ive answers
occurs when a question in a survey is sensitive or highly person41 (e.g. does the
respondent regularly engage in shoplifting or use drugs?). Consider first the
estimation of a binomial proportion-the proportion 11'" of respondents who
belong to a certain class A or have committeu a certain act . By ingenious use of a
randomizing device, Warner (1965) showed that it i possible to estimate this
proportion without the respondent revealing His or her personal status with
respect to this question. The objective is to encourage truthful answers while fully
preserving confidentiality.
The randomizing device, such as a spinning arrow or a box with red and white
b,alIs, selects one of two statements or questions, each requiring a Hyes" or "no"
response, to be presented to the respondent. The mterviewer does nol know
which Question any respondent has answered, but does know the relative prob-
abilities P and (1- P) with which the two statements tilt: presented. The success of
the method depends, of course, on the respondent's being convinced that by
participating he or she will not be revealing personal status with regard to the
sensitive issue.
In Warner's original proposal the two statemefJts are:
"I am a member of class A." (presented WIth probability P)
"I am not a member of class A."
With a random sample of n respondents the interviewer records a binomial
estimate j, = mIn of the proportion t/J of "yes" an wers. If the questions are
answered truthfully, the relation between t/J and 11'" in the population is
t/J = P11'A + (1- P)(1-11'A) = (2P - 1)11'" + (1 - P) (13.66)
With known P, this relation suggests the estimate

A (j,-(l - P)] (P;6 -21) (13.67)

11'AW == (2P - l)
 SOURCES OF ERROR IN SURVEYS 393
This estimate turns out to be the maximum likelihood estimate of 1TA' (The suffix
W denotes " Warner" .) The estimate is unbiased, with variance

(13.68)

Writing (1- cp ) in the form

(1- cp) = (2P -1)(1 - 1TA)+(1 - P) (13.69)
we find ea i1y

(13.70)

The first term in V (7TAW) is the variance that V(7TA) would have if al1 n
respondents answered truthfully a direct question about class A membership.
Except for 1TA near ~ and P > 0.85, the second term is greater than the first, often
much greater. The method is thus quite imprecise in general. This might be
expected, since the interviewer does not know whether a " yes" answer implies
membership in class A or the opposite. As Warner showed, however, his method
may give a smaller MSE than a direct sensitive question would, if -the latter
produced numerous refusals or false answers.

13.18 THE UNRELATED SECOND QUESTION

As an alternative to the Warner method, Simmons suggested (Horvitz, Shah,
and Simmons, 1967) that respondent cooperation might improve if the second
statement was not in any way sensitive, belDS unrelated to the first . For example,

"I was born in the month of May."

The first statement remains unchanged . If all respond truthfully, the population
proportion of "yes" answers is now
(13 .71)
where 1TU is the proportion in the sampled population who were born in May. If
1Tu is known, the obvious (and maximum likelihood) estimate of 1TA is

.. [~ - (l- P)17"u] (13 .72)

1TAU = p

with variance

(1.3-: 73)
"
394 SAMPLING TECHNIQUES

Morton (Greenberg et ai., 1969, p . 532) has suggested bow tbe case, 1TU known
can always be acbieved. A box contains red, white, and blue baJls in known
proportions P h P2 , P3 . Drawing a red ball produces the sensitive statement.
Drawing a white or a olue ball produces tbe statement: "The color of this ball is
white." Thus. 1TU = P2/(P2 + P3).
Dowling and Shachtman (1975) have shown that V(?T,;,U) < V(?TAw) for all
1TA. 1TU, provided that P exceeds about l. (The variance of ?TAW i symmetrical
about P =~ but that of ?TAU is not, a small P proviCJing few responses on the
sensitive question with this method.)
If it is necessary to estimate both 1TA and 1TU we can have two random samples
of sizes n h n2. with different proportions Pit P 2 for the sensitive question. With
cfJh cfJ2 denoting the proportions of "Yes" answers in the populations defined by
the choices PI and P 2 •
cfJl = P 1 1TA + (1- P ,)7Tu (13.74)
cfJ2 = P 21TA + (1 - P 2)1TU (13 .75)
These relations suggest the estimate
A [~I(I - P2) -~(1-PJ)]
1TA U= (P 1 -P2 ) (13.76)

with
V(1T,;,U) =
(P I - P2 ) nl n2
?]
1 2[cfJl(1 - cfJl )(1-P2)2+ cfJ 2(J- cfJ2)(1- P I (13.77)

If PI >1· Greenberg et ai. (1969 showed that this variance is min'mized when
P2 = 0, that is, when an in the second (n2) sample are a!,ked the unrelated 1T(J
question. Moors (1971) has recommended ~hi procedure, but Greenberg et aI.
(1969) suggest PI + P 2 = 1 as a working rule, in case the choice P2 = 0 might
weaken cooperation by respondents. When PI =0.8, for example, 80% of ample
1 and 20% of sample 2 would be asked the sensitive question on the average.
With the optimum n I, n2 for given n = nl + n2, the Cauchy-Schwarz inequality
shows that the resulting minimum variance of ?TAU is

Vmj,,(1TAU) = n(P
I
~ P 2)z[(I- P2)JcfJ1(1- ~1)+(1 - P )JcfJ2(1 -cfJ2)]2
I (13.78)

The minimizing n t! n l ratio is

(13.79)

This choice requires advance estimates of 1TA and 1TU, but the optimum is fairly
. Greenberg et aI. (1969) give recommendations aboutthe choices of Ph P'}" n I'
 SOURCE.<; OF ERROR fN SURVEYS 95
and n2 ' For preservation of confidentiality it helps to have 'lTu approximately
equal to 17"A-
Numerous variants of these methods have been studied (e.g., use of two
unrelated questions), as well as the biases produced in '77-A and '77-AU if a fraction
of the respondents answer the question falsely. Greenberg et al. (1971) have
applied the two-sample technique to estimate the mean ILA for a sensitive discrete
or continuous variable by methods analogous to those leading to equations
(1 3.74) and (13.75). The unrelated question estimates the mean lLu {or a
nonsensitive variable. Random subgroups of"; ubjects (i = 1,2) receive the
sensitive question with probability P" the nonsensitive with probability (1 - PI)'
The recorded variahle Z, for a subject therefore follows a mixture of two
distributions in proportions P" (1- Pi), one distribution with mean ILA ' variance
u ,/. the other with mean J-Lu. variance uu2 • Hence
£(I,) = P,lLA;' (1 - PJlLu (13 .RO)
V(z,) = P,<TA2.+ (I - Pi)UU2.+ Pi (1- P, (jLA - p,(J)l (13.81)
Analogous to (13.76), the estimate of ILA is
~ ((l- P1)ZI-(1- J'I )Z2]
ILA U: (PI - P2 ) (13.82)

For maxunlJm efficiency the condition lLu =ILA' uu2 =0 are required , while for
preservation of anonymity, lL u = ILA, u u 2 = UA 2 are best.
Warner (1971) has given a theoretical framework for a broad cia s of ran-
domized response models. As (j 3.74) and (]3.75) suggest, the trick is to estimate
certain linear functions of the sensitive and unt ' lat d ?T'S or IL 'S with as many
equations as there are parameters to be estimated.
The method has been applied to obtain estimate:> of the proportion
illegitimate births, of induced abortions, of users of ht'roin, of persons having
contact with organized crime, and of mean income and number of abortions as
continuous-discrete applications. Since the method bas attracted widespread
attention, further applications are likely to appear. An excellent review is given by
Horvitz, Greenberg, and Abernathy (1975) .
There have recently been discussions of ule degree of privacy that the resron-
dents have in different versions of randomized interviewing. In some versions the
interviewer may be able to guess the statu of some respondents with regard to the
sensitive issue with a fairly high probability of being correct-an undesirable
feature for this method.

13.19 SUMMARY
In regard to tI cir effects on the formulas given in preccJing rha)1lers, non
piing error may be classified as foJ)ows.
396 SAMPLING TECHNIQUES

1. With noncoverage and nonrespon e, the most important consequence is that

estimates may become biasc::d, because the part of the population that is not
reached may differ from the part that is sampled. There is now ample evidence
that these biases vary considerably from item to item and from survey to survey,
beang sometimes negligible and sometimes large. A second consequence is, of
course, that the variances of estimates are increased because the sample 'actuaBy
obtained is smaller than the target ample. This factor can be allowed for, at least
approximately, in electing the size of the target sample.
2. Errors of measurement that are independent from unit to unit within the
sample and average to zero over the whole population are properly taken into
a count in the usual formulas for computing the standard errors of the estimates,
provided that fpc terms are negligible. Such errors decrease the precision of the
estimates, and it is worthwhile to find out whether this decrease is serious.
3. If errors of me ltsuremenl on difterent unjts in the sample are correlated, the
usual formulas for the standard errdrs are biased. The standard errors are likely to
be too small, ,incc the correlations are mostly positive in practice. This type of
disturbance IS as ·j) overlooked and may often have passed unnoticed.
4. A constant bias that affects all units alike is hardest of all to detect. No
manipulatiom, of U\e sample data will reveal this bias.
As this chapter ha~ indicated, the ~tudy of theSe problems is slow and difficult.
Neverthele~ . a goud beginning has been made. Much ingenuity has been shown in
devislOg . :chn,que:o for the assessment and control of nonsampling errors.
Althougb I h,e; i!t a field in which broad generalizations are hard to attain,
informatJotl is accumulating about the nature and magnitudes of errors of
m asuremt'nt in difIerent types of survey. More is being learned also about what
can be accomplished with good training and supervision of the interviewers, with
prt:testing, with nl_chanisms to control the quality of the field work, and with a
post!>urvey aopraisal of the successes and weaknesses in the operation.
Uuder certain assumptions, a second measurement of a subsample of the units
by another interviewer of similar skill provides estimates of the simple response
variance and I s ratio to the total response variance, as well a an approximate
e 'timat of its ratio to the sampling variance to whici thi . ubsample is subject. A
~('"nd device- interpenetrating subsamples-iumishes estimates of the total
varittnce sampling variance plus response variance) and an estimate of the
correlatt:d component of the response variance. Combination of interpenetration
and repeat measurement is particularly fruitful.

EXERCISES
13.1 Suppose that, by field methods of different intensities, it is possible to make the
"response" stratum consist of 60, 80, 90, or 95% of the whole population. For a percel'ltage
that is to be estimated, the true "response" stratum means are: 60% stratum, 40.7; SO%
stratum, 43.5; 90% stratum, 44.S; 95% stratum, 45.4; last 5%, 59.0. (0) For a method that
 SOURCES OF ERROR IN SURVEYS 97
samples only the 60% stratum, show that the root mean square en:or of the estimated
percentage for the whole population is

J(2414/n)+28 .94
where n is the number of completed questionnaires obtained. (b) Show that a root mean
square error of 5% cannot be achieved by a method with 60% response but can be obtained
with slightly over 100 completed questionnaires for the methods that have a response of
80% or better. (c) If a root mean square error of 2% is prescribed, what methods can
achieve it and what sample sizes are needed?
13.2 In 13.1 (c) suppose that it costs $5 per completed questionnaire for the field
method that has a 90% response . To obtain a completed questionnaire from the nexl 5% of
the population costs $20. For a root mean square error of 2% , is it cheaper to use the
method with 90% response rale or that with 95% response rate?
13.3 A population consists of two strata of equal sizes. The probability of finding the
respondent at home and willing to be interviewed at any call i 0.9 for persons in stratum 1
and 0.4 for persons in stratum 2. (a) In the notation of section 13.5 show that
WI! = 1- (0.1)', W'2 =1 - (0.6)'
~b) If the original sample size is no, compute the total expected number of interviews
obtained for I, 2, 3, 4 and 5 calls. (c) If the relative costs per completed intervjew at the ith
call are tOO, 120, 150. 200. and 300 for j :: ),2,3, 4, 5, respectively, compute the average
cost per interview for all intervi~ws obtained up to the jth call. (d ) The money available for
the survey is enough to pay for 300 completed first calls. If the policy is 10 insist on i calls,
what are the expected total numbers of completed interviews that can be obtained for the
same amount of money when j:o 1, 2, 3, 4, 5?
13.4 In exercise 13.3 persons in stratum) have a mean of 40% for some binomial
percentage that is being estimated and persons in stratum 2 have a mean of 60%. (a )
Compute tbe bias in the sample mean for j = 1, 2, 3, 4, 5 calls. (b) Colppute tbe variances of
the sample means for the cost situation in part (d ) of exercise 13.3. (To save computing, the
variance may be taken as 2600/11" where II, i the expected total number of interview
obtained.) (c ) Which policy gives the lowest MSE~
13.5 In section 13.6 (subsampling of the noorespondents) verify the formula (p. 373)
for the ratio of the expected cost of obtaining a specifi d V with no subsampling to the
minimum expected cost.
. F(co+c ,W.+c, W 1 )
Rabo = --;.==========-.....:..c~;=­
[J(F - W,)(co+c ., W,)+ W,Jc:]l
where F = 5' /5/. Let Co =c, = 1, Cl = 16. (a) If F :; 1, show that the cost ratio has a
maximum 1.25 when W, = 0.2 or 0.25. (b) If F = 1.5, show that the maximum is 1.4) for
W1 .,. 0.3 or 0.35.
13.6 In a survey on poultry and pigs kept in gardens and certain small holdings (Gray,
1957) a postal inquiry with several reminders was followed by interviews of a subsample of
nonrespondents. By advance judgment, k = 2 was chosen (i.e., a 50% su~mple) . The
following data were available after the survey for one important item, il' the notation of
exercise 13.5'.

~.O.15, ~.9.5, W. = O.8,

Co Co
398 SAMPLING TECHNIQUES

By finding VC for k = 2 and for the optimum k , determine whether k = 2 was a good
choice.
13.7 In a survey by the Politz-Simmons method 390 respondents in an initial sample of
660 were found at home on the first call . The number who stated that they were at home on
O. 1, ... , 5 of the five previous nights and the number answering yes to a question in the
survey were as follows .
0/ 5 1/ 5 2/ 5 3/5 4/5 5/ 5

Number 14 35 55 74 94 118
Yes answers 4 )3 20 30 42 156

Compute the Politz-Simmons estimate of the proportion of " yes" answers in the
population and compare it with the simple binomial estimate.
13.8 A population with N = 6 contains three units for which the correct answer to a
question is yes and three for which it is no. Owing to errors of measurement, the probability
of obtaining a "yes" response on a yes unit is 0.9. and the probabili ty of obtaining a " no"
response on a no unit is also 0.9. (a) By working out the distribution of all possible
responses for samples of size 2, show that the probabilities are 0.218 , 0.564, and 0.2 18 that
the sample gives 0 , 1, 2 "yes" responses. (b) Show that th variance of the estimated
proportion of "yes" respon e is 0.1090. Verify results (13.40) and (1 3.41) in section
13 .10. (c) What would be the variance of the estimated proportion of " yes" responses if
there were no errors of measurement?
13.9 In part of the 1942 BengaJ Labour Enquiry (Mahalanobi , 1946) a random
sample of about 175 families was taken in each of three strata . The sample in each stratum
was divided into five random subsamples, each assigned to a different interviewer. The five
interviewers worked in aU three strata. For expenditure n food , the relevant part of the
aQjllysis of variance (on a Single-family basis) is as follows .

df ms E(ms)

Between interviewers 4 22. 3 u,/+ u/+3 5u/i + 105u/

rnterviews x strata 8 9.6 u ,. / + u/ + 35u/s1
Within subsampJes 510 9.9 u,/ +u/

If 8.. W h, represent biases of interviewer i, the model for a single family is

Yhljo = iih + gl + Wlo' + (lA-' III! - ii~) + dllljo
Variances:
Verify the expressions given for E(ms ) and estimate the proportion of the total variance
of the mean that may be ascribed to enumerator biases.
13.10 Consider an illegal act that 10% of the population have commined ('ITA = O.I) . If
all respondents an wer trutbfully, compare tbe V( rrA ) for n = 500 given by (a) a direct
sensitive question, (b) the Warner method with P = 0.8, (c) the unrelated question method
with 'IT. =0.2 known, (d) the two-sample unrelated question methoo with 'IT. actually 0.2
but unknown, when PI = 0.8, P2 =0 as recommended by Moors, (e) the same method when
PI = 0.8, P1 = 1- PI ' Assume that you can use the ptimal nl/n 2 in (d) nnd (e) . Some
decimals are avoided by calculating V( lOO7TA ) = 10' V(7TA) for the methods.
 SOURCES OF ERROR eN SURVEYS 399
13.11 In exercise 13.10, uppo e that all respondents answer truthfully with any of the
randomized response methods (b). (c). (d), or (e), but thaI under a single direct sensitive
question, method (0) , some respondents who have committed the act deny this. For
n = 500, which of the randomized response methods give a smaller MSE( 'Ii-A) than method
(0 ) if (i) 15% , (ij) 20%, (iii) 25% of those who have committed the act deny it under
method (0 )1
 References

Annitage. P. (1947). A comparison of stratllied with unrestricted random sampling from a finite
population. Biometrika. 34, 273-280.
Arvcscn. J. N. (1969). Jackknifing U-Statistics. AIIII. Math. Slal., ..o; 2076-2100.
Avadhani. M. S. and Sukhatme. B. V. (1973). Controlled sampling with equal probabilities and
without replacement. lilt. Stat. Rev., 41, 175-183.

Bailar, B. A . (1975). The effects of rotation group bias on estimates from panel surveys. Jour. Amer.
Stat. Assoc., 70, 23- 30.
Bailar. B. A., and Dalenius, T. (1969). Estimating response variance components of the U.S. Bureau
of the Census Survey Model. SmsIchya. B31, 341- 360.
Barr, A. (1957). Differences between experienced interviewen. App. Stat , 6, 180-188.
Bartholomew. D . J. (1961). A method of allowing for " not-al-home" biu in sample surveys. App.
Stat.• 10,52-59.
Bartlett. M. S. (1949). Fitting a straight line when both variablet are lub;ect to enor. Biometric.t. 5,
., 207- 212.
Basu. D. (1958). On sampling with and without replacement. Smtkhya, lO, 287-294.
Bayless. D. L. (1968). Variance estimation in sampling from finite populations. Ph.D. Thesis. Texas A
& M Univenity.
Bayless, D. L., and Rao, J. N. K. (1970). An empirical study of stabilities of estimators and variance
estimators in unequal probability sampling (n - 3 or 4). Jour. Amer. Stat. Assoc .• 65, 1645-1667.
Beale, E. M. L. (I 962). Some uses of computers in operational research. Industrklk Organisation,ll,
51-2.
Bean, J. A. (1970) . Estimation and sampling variance in the health interview survey. National Center
for Health Statistics, Washington, D.C., Series 2. 38.
Bean, J. A. (1975). Distribution and properties of variance estimaton for complex multistage
probability samples. National Center for Health Statistics. WashinJton. D.C .. Series 2, 65.
Beardwood, J. , Halton, J. H.. and Hammersley, J . M . (1959). The .hortest path through many points,
Proc. Cambridge PhIl. Soc., 55, 299-327.
BeJlhouse, D. R. and Rao, J. N. K. (1975). Systematicsarnplingin the presence of. trend. Biometrilul,
61, 694-697.
Belloc, N. B. (1954). Validation of morbidity survey data by comparison with hospital records. Jour.
Amer. SIal. Assoc., 49, 832-846.
Birnbaum, Z. W., and Sirken, M. O. (1950a). Bias due to nonavailability in sampling surveys. JoW'.
Amer. Stat. Assoc .• 45, 98-111.
Birnbaum, Z. W., and Sirken, M. O . (1950b). On the total error due to noninterview and to random
sampling. btl. JoW'. Opinion and Attitude Res., 4, 179-191 .
Blythe, R. H. (1945). The economics of sample size applied to the scaling of saw-logs. Biom. Bull., 1,
67- 70.
400
 REFERENCES 401
Booth, G .• and Sedransk. J. (1969). Planning some two-factor comparative surveys. Jour. Amer. SIal.
Assoc .• 64, 560-573.
Bose, Charneli (1943). Note on the sampling error in the method of double sampling. Sankhya, 6, 330.
Brewer, K. W. R. (196.3a) . A model of systematic sampl ing with unequal probabilities. Ausrralian
Jour. Sial., 5, 5-13.
Brewer, K. W. R. (1963b) . Ratio e~ti mati on in finite populations; Some results deducible from the
assumption of an underlying stochastic process. Ausrralian Jour. Sial .. 5, 93- 105.
Brewer, K. W. R .. and Hanif, M. (1969). Sampling without replacement and probability of indusion
proportional to size . I Methods using the Horvitz and Thompson estimator. II Methods uSing
special estimators. Unpublished manuscript.
Brewer, K. W. R .. and Hanif. M. (1970). Durbin's new multistage variance estimator. Jour. Roy. Stal.
Soc .. 832, 302-311.
Brooks, S. (1955). The cstim tion of an optimum subsampling number: Jour. Amer. tat. Assoc .. SO,
398-415.
Bryant, E. C , Harrley, H. 0 ., and Jessen, R. J . ( 1960). De ign and estimation in two-way stratifica tion.
Jour. Amu. Slat. Assoc., 55, 10S-124,
Buckland, W. R. (195 1). A review of the literature of systematic sampling. Jour. Roy. Slat. Soc., 813,
20tl- 215
Burstein. H, (1975). Finite population correction for binomial confidence limits. Jour. A mer, SlOt.
Assoc" 70, 67--69,

Cameron. J, M. (1951) , Use of variance components in preparingscheduJes for the sampling of baled
wool. Bioml!rics, 7. 11.\ - 90,
Chatterjee, S, ( 19M) . A programming algorithm and its Slat: tical al'phcatioo~, O.N,R. Tech, Re.pt. I.
Departme nt of Sta\l~lIcs . Harvard Univer. ity. Cambndge.
Chatterjee. S, (1 (67) . A note on ,)ptimum tratification , Skand. AkI.. 50, 40--44.
Chatterjee, S, ( 1968). Multivariate stratified ~ urvey~. Jour. Amer. Suu. Assoc .. 63, 530-534.
Chatterjee, S. (I (72). A tudy of optimum allocation in multivariate stratified surveys. Ska,ui. Akl.,
SS, 73- 1l0
Chung, J . H .. and DeLury, D. S , (1950) . Conjidmce Limits for Ihe Hypergeometric DistribullOn .
University of Toronto Pre~~, Toronto, Canada.
Cochran, W G . (1942) . ampling theory when the sampling units are of unequal sizes, Jour, Aml'r,
SIal, Assoc" 37, 199-2 I 2.
Cochran, W. G . (194(,i, Relative accuracy of systematic and stratified random samples for a certain
c1assof populatlnns. Ann . Malh , 5101 .• 17,164- 177 .
('..ochran. W. G. (196 I ) Comparison of methods for determining stratum boundaries. Bull, Int. SIal,
irul., 38, 2, 345-358,
Cochran, W, G .. Mosteller. F .. and Tukey, J. W, ( 1954). Slalistical Problems of lhe Kinsey R eporl.
American Slalistical Asrocialion, Washington, D.C.. p, 280,
Coleman, J. S. (1%6), Equality of Educalional OpPDrlunity, u.s. Government Printing Office,
Washington, D.C.
Cornell, F. G . (1947) . A stratified random sample of a small finite population. Jour, mer. SIal. Assoc.,
41, 523-532.
Cornfield. J. (1944) . n sample from finite populations, Jour, Amer. SIal. Assn::., 39, 23&-239,
Cornfield. J, (1951). The determination of sample size. Amer. Jour. Pub. Heallh, 41. 654-'661 ,
Cox, D. R, (1952). Estimation by double sampling, Biomerrika, 39, 217-227,

Dalenius, T . (1957), Sampling ill ",eden. Contributions to the methods and theories of sample survey
practice. Almqvi t and Wicksell, Stockholm .
Dalcnius, T " and Gurney, M. (1951) . The proll" m of optimum stratification. II. Skand. Akr.. 34,
133- 148.
 402 SAMPLING TECHNf l iES

Dale ni us, T .. and Hodges, J. L., .1r. (1959). Mirumum V:;I .!let: sttatiflcation. Jour. A mer. Sial. Assoc ..
54,88-101.
Das, A. C (1950). Two-djmensional yslematic samrltng and the associated stratified and random
sampl inA. Sankhya, 10. 95-1 08.
David, F. N .. a nd Neyman. J. (t~38). Ex tension of t he Markoff theorem of leaS! square .. l UI Res.
Mem., I , 105.
David, 1. P., and Sukhlllme, B. V. (1974). On the bias. a nd mean '\<j uare error of the ratio esllmlltor .
Jour. A mer. SUII. A (!c •• 69, 464-466.
Deming, W. c. (1953) . Or. a probability mechanism 1(1 alta in an economi balance be tween the
resultant error of non·rc~pon:>t: and the bias of non rc .ponse. Jour. A mer. Sial. AlSIlC., 41.
743-772.
Deming, W. E . (1956) . On sillllJls fications of sampling design through repl ication with equal plCb-
abilities and without stagt; • . Juur. A mer. Stat. Assoc .. S1 . 24-53.
Deming, W. E . (1960). Sample Desigl1'" Business Research. John Wi ley and Sons, Ne w York .
Deming, W. E., lind Simmoll ~, W. R. (1946) . On the desi~ nf 3 ~ample for dealer inven tories. Jour.
A mer. Stat. Assoc .• 41. 10-33 .
Des Raj (1954 ). On samphng with probabilities proportion"1 tf) size . Gallila, 5, 175- 1112.
Des Raj 11950a). Some esti m~llul' ill sampling wilh varying proba bilities without replace me nt. JOllr.
Am.:r. :>Ial. Assoc .. 51, 26v- 2M4.
Des Raj (1956b). A nOle Oil the determinatiun of optimum probabilities in sampling without
replacement . Scm khya, 17, 197- 200.
Des Raj ( J 1J5lS). On th~ relative accuracy of ome sampling techniques. JOllr. Amtr. SIal. Assoc., 53,
9B-I01.
Des Raj (lY64). 1 he use of systematic sampling with probability proportional to size In a lar~e-scale
survey. Jour. A mer. Sial. Assoc .. 59, 25 1-255.
Des Raj (1966). Some rcmarK~ on a simple proced ure of sampling without replacement. Jour. A mer.
Stal. Assoc .. 61 ,39 1- 396 .
• Des Raj, and Khamis, S. H. (195lS) . Some remarks on sampling WIth replaceme nt. A nn . Malh . SUlI.,
, 29,550-557.
Dowling, T . A. , and Sllat'l ltrlJan, R. H. (1975) . On til: relative efficie llCY of randomized response
models. Jour. Amer. Scar Assoc., 70, 84- R7.
Ourbin, 1. (1953) . Some r~sl.!~ in sampling th('ory when thl: uni ts are selected with unequal
prObabilities. Jour. Roy. tal. Soc., B15, 26~- ~b9 .
Durbin. 1. (1954). Non-response and call - back~ in ~ u n·lj r. . Bull. lilt. Stal. Insl., 34, 72- 86.
Durbin. J. (1958). Sampling theory for cstlm l~ based on few r individuah than the number selected.
Bull. Inl. Sial. Insl., 36, 3, 113- 11 9.
Durbin, 1. (1959). A note on the applka tion of Oue nouille 's method..,f bias reduction to the e~timation
of ratios . Biometrika, 46, 477--480.
Durbin. 1. (1967). Design of multi-stage surveys for the estimalJ(ll1..,f sampling errors A pp. Sial .. 16,
152-1 64 .
Durbin. J., and Stuart, A. (1954). Callbacks and clustering in sample surveys : an experimental study.
Jour. Roy. Stat. Soc., A1l7, 387-428.

Eckler, A. R. (1955). Rotation sampling. Ann . Math . Stal., l6, 664-685.

Ekman, G. (l959). An approximation useful in Univariate stratification. Ann. Math . Sial., 30,
219-229.
ErdOs, P., and Renyi, A. (1959) . On the central limit theorem for amples from a finite population.
Pub. Malh . [nsl. Hungarian Acad. Sci., 4, 49- 57.
Ericson, W. A. (1969) . Subjective Bayesian models in sampling finite populations. Jour. Roy. Slat.
Soc., B31, 195- 233.
Evans, W. D. (195 1). On stratification and optimum allocations. Jour. A mer. Stal. Assoc., 46,
95- 104.
 REFERENCES 403
Fellegi, I. (1963). Sampling with varying probabilities without replacement: rotating and non-rotating
samples. Jour. Amer. Slal. Assoc .• 58, 183- 201.
Fellegi, I. (19 4). Response variance and its «tilnllllon. Jour. Amer. Slat. A.uoc .. 59, 1016-1041.
Feller, W. (1957). An Introduction 10 Probability The(Jry and lIS Applicmil/l1ii, John Wiley and Sons,
New York . second edition.
Fieller, E . C. ( 1932). The distribution of the imlex in a normal bivariate population. Biormtrika, 24,
428-440.
Finkner, A. L. (J950). Method of sampling for' estimating commercial peach production in North
Carolina. North Carolina Agr. Exp. Stat. Tech . Bull., 91.
Finkner, A. L.. Morgan. J. J., and Monroe, R. J. (1943). Methods of estimating farm employment from
sample data in North Carolina. N. C Agr. Exp. Sta. Tech . Bull., 75.
Finney, D. J. (1948). Random and ;ystem3tic sampling in timber surveys. Forestry, 22, 1- 36.
Finney, D. J . (1949). On a method of estimatmg freq uencies. Biometrika, 36,233- 234
Finney, D. J. (1950). An example (If periodic variation in forest sampling. For~str)', 23, 96-111.
Fisher, R. A. (1958). Statistical Methods tnr H.esearch Workers . Oliver and B(lyd. Edmburgh,
thincenth edition, section 2 1. founh ed (1 932).
Fi~heT, R. A., and Mackenzie. W. A. (1922). rh correlation of weekly rainfall. Quart. Jour. Roy. Met.
StJL ., 48, 234- 245.
Fisher, R A ., dnd Yates, F. (1957\. Slali.'itical Tables for Biological. Aencu1tw-al and Medical
Researcll. Oliver and BOjd , Edinburgh, fifth edition.
Foreman. E. K., and Brewer, K. W. R (1971). The cfficic:nt u~e of supplementary information in
standard ~mphng procedures. Jour. Roy. Stal. Soc. B33, 391-400.
Frankel, M. R . (1971). Inference trom Mlrvey samples. Institute for Social Research, /\nn Arbor,
Mich .
Fuller, W. A. (1906). Estimation employing post strata. Jour. Amer. SIal. Assoc., 61.1172- 1183.
Fuller, W. A. (1970). Sampling with random strutum boundarie~ . Jour. Roy. SIal. Soc .. BJ2, 209-226.
Fuller, W. A., and Burmeister, L. F. (1972). Estimators for samples selected from two o verlapping
frames Proc. Soc. SIal. Stcl. Amtr Stal. A.uoc., 245-249 .

tJallup, G . (1972). OpinIOn polling in a democracy. Slutisn'cs, a GUIM Illlhc Unknown . J. M. Tanur et
af. (ed~ .), HOlden -Day, Inc., San Francisco, 146-152.
Gautschi, W. (1957). Some remarks on ystematic saml'ling Alltl . Math . Stat., 28, 385-394.
GodamlJe , V. P. (1955) . A UIIified theory of sampling from finite populations. Jour. Roy. Slat. Soc.,
BI7,269- 278.
Goodman, L. A., lind Hartley, H. O. (1958). The pr (,I';on of unbiased ratio-type estimators. Jour.
Amer. SIal. Assoc., 53, 491-508 .
Goodman, R., and Kish, L. ('1950). Controlled sel ction-3 technique in probability sampling. Jour.
Amer. Stal. A.uoc., 45, 350-372.
Graham, J. E. (1973). Composite estimation in two cycle r tation sampling designs. Comm. in Sral.. l,
419-431.
Gray, P. G . (1955). The m mary factor in social surveys. Jour. Armr. Stal. Assoc., SO, 344-363.
Gray, P. G. (1957). A sample survey with both a postal and an interview stage. App. SIal., 6, 139-153.
Gray, P. G., and Corlett, T . (1950). Sampling for the social survey. Jour. Roy. SIal. Soc., .A1l3,
150-206.
Greenberg, B. G ., et al . (1969). The unrelated q ue 'uon Tandomi:ted re ponse model: Theoretical
framework . Jour. Amer. SIal. Assoc., 64, 5:!U-539.
Greenberg, B. G . e t al. (1\171). Application of the randomized response technique in obtaining
quantitative data. Jour. Anurr. Stal. Assoc., 6(;, 243-250.
Grundy. P. M., Healy, M. J. R., and Rees. P. H. (1 54). DeCIsion betwee n two altemati"es- how
many e~perimenlS? Biormtrics, 10, 317~ J'
Grundy, P. M., Healy, M. J. R .. lind Recs, O. H. (1956). Eoonomic chOice of the amount of
expenmenla!lon. Jour. Roy. Stal. 5«., B18, 32-55.
404 SAMPLING TECHNIQUES

Hagood, M. J ., and Bernert, E. H . (1945). Component indexes as a basis for stratification. Jour. A~r.
SUlI. Assoc., 40, 330-341.
Hajek, J. (1958). Some contributions to the theory of probability sampling. Bull. Inl. Slat.lnst., 36. 3,
127- 134.
Hajek, J . (1960). Limiting distributions in simple random sampl.ing (rom a finite population. Pub.
Ma/h . lns/. Hungarian Acad. Sci., 5, 361 - 374 .•
Haldane: J. B. S. (1945). On a method of e timating frequencies. Bjo~trika , 33, 222-225.
Hansen , M. H. , et al. (1951). Response errors in surveys. Jour. Amer. Stat. Assoc., 46, 147-190.
Hansen , M. H., and Hurwitz, W. N. (1942). Relative efficiencies of various sampling units in
population inquiries. Jour. Amer. Stat. Assoc., 37, 89-94.
Hansen. M. H .. and Hurwitz. W. N. (1943). On the theory of sampling from finite populations. Ann.
Math . Stal .. 14, 333-362.
Hansen, M. H .. and Hurwitz, W. N. (1946). The problem of nonrespon 'e in sample surveys. Jour.
A mer. Slat. Assoc.. 41, 517- 529.
Hansen, M. H .. and Hurwitz, W. N. (1949). On the determination of the optimum probabilities in
ampling. Ann. Math . SIal., 20, 426-432.
Hansen, M. H .. Hurwitz, W. N., and Bershad, M. (1961). Measuremen errors in censuses and surveys.
Bull. Int. Slat. Illst., 38, 2, 359-374.
Hansen, M. H., Hurwitz, W. N., and Gurney, M. (1946). Problems and method~ of the sample survey
of busines . Jour. Amer. Stat. Assoc., 41, 173- 189.
Hansen, M. H ., Hurwitz, W. N., and Jabine, T. B. (1963). The use of imperfect lists for probability
sampling at the U .S. Bureau of the Census. Bull. Int. Stat. Inst., 40, I, 497- 517.
Hansen. M . H., Hurwitz, W. N., and Madow. W. G . (1953). Sampk SUJ'Vey Methods and Theory. John
Wiley and Sons. New York , Vols. I and II .
Hansen , M. H ., Hurwitz, W. N., Nisselson, H ., and Steinberg, J . (1955). The redesign of the censu
current population.survey. Jour. Amer. Stat. Assoc., 50, 701 - 719.
Hansen, M. H., Hurwitz, W. N., and Pritzker, L. (1965). The estimation and interpretation of gross
differences and the simple response variance. Contributions W SUlJistics presented W Professor P. C.
Mahalanobi.! . Pergamon Press, Oxford, and Statistical Publishing Society, Calcutta, 111 - 136.
Hansen, M. H., and Wak berg, J. (1970). Research 00 non-sampling errors in censuses and surveys.
Rev. Int. Stat. Inst .• 38, 318- 332.
Hanson, R. H., and Marks, E. S. (1958). lnfiuence of tbe interviewer on the accuracy of survey results.
Jour. Amer. Stat. Assoc., 53, 63~55.
Hartley. H. O. (1946). Discussion of paper by F. Yates. Jour. Roy. Stat. Sac. , 109, 37 .
Hartley. H. O. (1959). Analytic Studies of Survey DaUl. Istituto di Statistica. Rome, volume in honor of
Corrado Gini.
Hartley, H. O . (1962). Multiple frame surveys. Proc. Sac. Stat. Sect. Amer. SUIt. Assoc., 203-206.
Hartley, H . O . (1974). Multiple frame methodology and selected applications. Sankhya, CJ6,99-1 18.
Hartley, H. 0., and Hocking, R. (1963). Convexing programming by tangential approximation.
Management Science, 9, 6()()-{\12.
Hartley, H. 0 .• and Rao, J. N. K . (1962). Sampling with unequal probabilities and without replace-
ment. Ann. Math. Stat., 33, 350-374.
Hartley, H . 0., and Rao, J. N. K. (1968). A new estimation theory [or sample surveys. Biometrika, 55,
547-557.
Hartley, H. 0 ., and Rao, J. N. K. (1969). A new estimation theory for sample surveys, lL In New
Developments in Survey Sampling, N. L. Johnson and H. Smith (eds.), Wiley-Interscience, New
York , 147-169.
Hartley, H. 0., Rao, J. N. K. , and Kiefe r, O. (1969). Variance estimation with one unit per stratum.
Jour. Amer. Stat. Assoc., 64, 841-851.
Hartley, H. 0 ., and Ross, A. (1954). Unbiased ratio estimates. Nature, 174, 270-27] . .
Harvard Computation Laboratory (1955). Tabks of the CUmulative Binomial Probability Distribution.
Harvard University Press, Cambridge, Mat~s.
 REFERENCES 405
Haynes. J. D. (194R) . An empirical investigation of amphng methods for an area. M. S. thesis.
University of North Carolina.
Hendricks. W. A. (1944). The relatt¥e efficiencies of groups of farms as sampling units. Jow. Amer.
SUlI. Assoc., 39, 367-376.
Hendricks, W. A. (1949). Adjustment for bias by non-response in mailed surveys. Agr. &otI. ReI.• 1.
52-56.
Hendricks, W. A. (1956). The MalMmatical Theory of Sampling. Scarecrow Press, New Brunswick.
N.J .
Hess. I., Riedel, D. C .• and Fitzpatrick, T. B. (1976). Probability Sampling of HospiUl/s and Patients.
University of Michigan, Ann Arbor, Mich .• second edition.
Hess. I., Sethi. V. K.• and Balakrishnan, T. R. (1966). Stratification: A practical investigation. Jour.
Amer. Stal. Auoc., 61, 74-90.
Hoetfding. W. (1948). A class of statistics with asymptotically normal distribution . Ann. Malh . SIal .•
19.293-325.
Homeyer, P. G., and Black, C. A. (19<46). Sampling replicated field experiments on oats for yield
determinations. Proc. Sail Sci. Soc. America, 11,341-344.
Horvitz. D. G. (1952). Sampling and field procedures of the PitL~burgh morbidity survey. Pub. Heallh .
Re~.67, 1003-1012.
Horvitz, D. Y., Greenberg, B. G ., and Abernathy, J. R. (1975). Recent developments in randomized
response designs. A slll'Vey of statistic41 duign and /inearmode/s. J. N. Srivastava (ed.), American
Elsevier Publishing Co., New York, 271 - 285.
Horvitz, D. G ., Shah, B. V., and Simmons. W. R. (1967). The unrelated randomized response model.
Proc. Soc. Stat. Sect. Amer. Stal. Assoc., 65- 72 .
Horvitz, D. G., and Thompson, D. J. (1952). A generalization of sampling without replacement from a
finite universe. Jour. Amer. Stat. Assoc., 47, 66~5 .
Huddleston. H. F., Qaypool, P. L., and Hocking, R. R. (1970). Optimum sample allocation to stralA
using convex programming. App. Stal. • 9, 273-278.
Hutchinsol\. M. C. (1971). A Monte Carlo comparison of some ratio estimators. Biometrika, sa.
313-321.
Hyman, H. H. (t 954). llllertHewing in Soci41 Research, University I)f O1icago Press, Chicago, III.

James, A. T .• Wilkinson. G. N.• and Venables. W. N. (1975). Interval e~timates for a ratio of means.
Sankhya (in press). .
Jebe, E. H. (1952). Estimation for sub-sampling designs employing the county as a prillUlty sampling
unit. Jour. Amer. SIal. Assoc .• 47, 49-70.
n.
Jensen, A. (1926). Report on the representative method in statistics. Bull. 1111. Stat. Insl., 359-377.
Jessen, R. J. (1942) . Statistical investigation of a sample survey for obtaining farm facts. Iowa Agr.
fup. Sta. Res. Bull., 304 .
Jessen. R. J. (1955). Determining the fruit count on a tree by randomized branch sampling. Biometrics,
U,99-109.
Jessen, R. J.. et aI. (1947). On a population ,ample for Greece. Jour. Amer. SIal. Assoc., 41, 357- 184_
Jessen, R. J., and Houseman. E. E . (1944). Statistical investigations 01 farm sample surveys taken in
Iowa, florida and California. Iowa Agr. fup. Sta. Res. Bull.• 329 .
Johnson, F. A. (1941). A statistical study of sampling methods for tree nursery inventories. M. S.
thesi , Iowa State College.
JohNOn, F. A. (1943). A statistical ~tudy of sampling methods for tree nursery inventories. Jow.
F_slry, 41. 67~89 .
Jones, H. W. (1955). Jnvestigating the properties of a sample mean by employing random subsamplc
means. Jow. A mer. Stat. Assoc., 51. 54- 83.

Kempthorne, O . (1969). Some remarks on inference in finite sampling. New ~velopmenll in Survey
Sampling, N. L. Johnson aAd H. Smith. Jr. (eds.), John Wiley &. Sons. New York, 671-tiCJS.
406 SAMPLING TECHNIQUES

Kendall. M. G., and Smith, B. B. (1938). Randomness and random sampling numbers. Jour. Roy. Sit
Soc., 101, 147- 166.
Keyfitz, N. (1957). Estimates of sampling variance where two units are selected from each stratulT
Jour. Amer. Stal. Assoc., 52, 503-510.
Khan . ... and Tripathl, T . P. (J 967). The use of multIvariate auxiliary information in double-samplinJ
J . Ind. Stat. Assoc .. 5, 42-411.
King, A. J., and McCarty. D. E. (194 I) . Application of sampling to agricultural statistics with empha51
on stratified <amples. Jour. Marktting, April , 462-474.
Kis~r. C. V .. and Whelpton. P. K. (1953). Resume of the Indianapolis study of social and psychologic;
factors affecting fertility. Population Studies, 7, 95-110.
KISh, L. ( 1949). A procedure for objective respondent selection within the household . Jour. Ame.
5101. Assoc., 44, 380-187 .
Kish. L. (1957). Confidence limit> for clustered samples. Amer. Soc. Rev. , 22,154-165.
Kish , L. ( 1'165). Survey Sampling. John Wiley & Sons. New York .
Klsh. L.. and Frankel. M. R . (1974) . Inference from complex samples. Jour. Roy. Stat. Soc., 836,1-37
Kish , L.. and Hess, 1. (1958) . On non coverage of sample dwellings . Jour. A mer. Stat. Assoc.. 53
509-524.
Kish. L.. and Hess. I. (1959a) . A "replacement" procedure for reducing the bias of nonre~ponse
Amer. Statistician, 13, 4, 17- ISI .
Kish. L.. and Hess. I. (J 959b). On vanances of ratios and thelT differences in multistage samples. Jou.
Amer. Stal. A.ttor .• 54, 416-44t.
Kish. L.. and Lansing. J. B. (J 954). Response errors in estimating the value of homes. Jour. Amer. Sta
Assoc .• 49, 520-5311.
Kish , L., Namboodiri. N. K., and Pillal, R. K. (1962) . The ratio bias in surveys. Jour. A mer. 510
Ass()( .. 57, Rti3-S7t.
Koch . G. \ I 'J73) . An alternative approach to multivariate response error models for sample surve
data with application~ to esti mators involving subelass means. Jour. A mer. Stat. Ass., 611
906-913 .
Kokan . A. R. (1963) . Optimum allocation in multivariate surveys. Jour. Roy. Slat. Soc., AIU
557-565 .
Koons . D. A. (197:1) . Quality control and measurement of nonsampling error in the Health lnterviel
Survey . Nal. (:t nter for Health Stat .• 5eries 2.54 .
Koop. J . ('. ( 196()) . On theoretica l questions underlYing the technique of replicated or interpenetratin
,amples. Proc. Soc. SIal. Secl. Amer. Sial. Assoc.. 19ti-205.
Koop. J. C i: aIiR). An exercise in ratio estimation. Amer. Statisticiafl, 22, 1.29-30.
J..:ulldorff. G. (1<)(>3) . Some problems of optimum allocation for sampling on two occasions. Rev. In
Sial. losl.. 31 , 24-57.
Lahiri . D. 8 . (1951) . A method for sample ~election providing unbiased ratio estimates. Bull. lnl. Sta,
Inst .. 33, 2. 133- 14(J.
Lieberman. C . J .. and Owen. D. B. (1961). Tables of tlte Hyptrgeometric Probability Distribution
Slanford UnIversity Press. Stanford. Calif.
Lienau. C. C. ( 1941). Selection. training and performance of the National Health Survey field staff
Amer. lour. HVllime.34, 110- 132.
Lund. R. . (I'Iolli . EstImators in multiple frame surveys. Proc. Soc. Sci . Sect. Amer. Stat. Assoc.
282-2RIl .

McCartQy. P. J. (19/)6) . Replication : An approach to the an8lysi~ of data from complex survey
National Center for Health Statistia;. Washington. D C., Series, 2, 14.
McCarthy, P. J. ( 1969). Pse udo-replication : Half-samples. Rev. (nl. Slat. [nst., 37, 239-2&,.
McVay. F. E . (1947). Sampling methods applied to estimating numbers of commercial orchards in
commercial peach area. lour. Amer. Stal. Assoc.• 42, 533-540.
Madow. L. H. (1946). Systematic sampling and its relation to other sampling designs. Jour. Amer. 51,
Assoc.. 41, 207-214.
 REFERENCES 407
Madow, L. H . (1950). 00 the use of the county as a primary samphng unit (or ~tate estimates. Jour.
Amer. Stat. Assoc. , 45, 30-47.
Madow. W. G. (1948 ). On the limiting distributions of estimates based on samples from finite
universes. Ann. Math . Stal., 19, 535-545 .
Madow. W. G. (1949). On the theory of ystematic sampling,n. Ann Math. Stat .• lO. 333- 354.
Madow. W. G . (\953). On the theory of systematic sampling. Ill. Ann. Math . Stat .. 24.101 - 106.
Madow. W. G ., and Madow, L. H. (1944). On the theory of sY~lematicsampling. Ann. Math. Stal .. 1S,
1-24.
Madow. W. G . (1965). On some 8 peets 01 response error measurement. Proc. Soc. Stat. Soc. Amer.
Stal. Assoc., 1112-192 .
Mahalanobis. P. . (1944). On large-scale sample surveys. Phil. Trans. Roy. Soc. London. B231,
329-451.
Mahalanobis. P. C. (1946). Recent experiment' in statisllcal ~ampling in the Indian Statistical
Institute. Jour. Roy. Stal. Soc .. 109,3 25-370.
Matern. B. (1947). Methods of esumating the accuracy of line and sample plot surveys. Medd. ir.
Sialens Skogsiorsknings institul.• 36, 1- 138.
Matern, B. (1960). Spatial variation . Medd. ir. Sla/ens Skogsiorsknings Institut .• 49, 5. 1-144.
Mickey. M. R. (1959). Some finite population unbiased ratio and regression estimators. Jour. Amer.
Stal. Assoc .• 54, 594-612.
Midzuno, H. (1951). On the sampling system with probability proportionate to sum of sizes Ann. /t1S1.
Stal. Math .. l, 99-108 .
Milne. A. (1959). The centric systematic area sample treated as a random sample. Bio~trics, 15,
:!70-297.
Moor~ , 1. J. A . (1971). Optimization of the unrelated Quesuon randomized response model. Jour.
Amer. SIal. Assoc .. 66, 627-629.
Murthy. M. N. (1957). Ordered and unordered estimators in sampling without replacement. Sankhya.
18,379- 390.
Murthy, M. N. (1967). Sampling 'J'Mory and Melhods. Statistical Publishing Society. Calcutta. India.

Naroin. R . D . (1951). On sampling without n'pll1ccment with varying probabilities. Jour. Ind. Soc.
Agric. Sial .• 3, 169- 174.
National Bureau of Standard (1950). Tables vi I~ Binomial Prohability Distribution . U.S. Govern-
ment Printing Office. Washington. D .C.
Neter, J. (1972). How accountants save monel' by sampling. Slati.rtics. A Guide /0 Iht Unknown, J. M.
Tanur et 81. (eds.), Holden-Day. Inc .• San Francisco. 203-211 .
Neyman. J. (1934) . On the twO different aspects of the repre entative method: The method 01 stratified
sampling and the method of purposive selection. Jour. Roy. SIal. Soc .• 97, 558-606.
Neyman. J. (1938). Contribution 10 Ihe theory of ampling human populations. Jour. Amer. Stat.
Assoc.• 33,101 - 116.
Nordbotten , S. (1956). Allocation in tratified ~ampling by means of linear programming. Skand. Akl.
1idskr., 39, 1-6.
Nordin. J. A. (IQ44). Determining sample ize. Jour. Amer. StaL Assoc .• 39,497-506.

Ikin. 1. (\ 9~8). Multivariate ratio E'stimation for finite populations. Biometrika. 45, IS4-1b:i .
Osborne, J. G. (1942). Sampling errors of yslematic and random surveys of cover-type areas . Juur.
, A mer. lat. Assoc .. 37, 256-264.

Patterson, H. D. (\ 950). Sampling on ucce sive occasions with partial replacement 01 units. Jour. Roy.
Sial. Soc., 8 12, 241-25~ .
Patterson, H . D. (1954). The errors of I ttice ampling. Jour. Roy. Stal. Soc .• BI6. 14(~ 149.
Pau lson, E. (1942). A note on th estimation of me mean values for 8 bivariate distribution . Ann.
Math . Stat.• 13, 44()-444 .
 408 SA MPLING TECHNIOUES

Payne . . L. ( 1951). The Arl of Askinll Ouestions. Princeton University Press. Princeton, N.J .
Placke tt . R. L .. and Burman. J. P. ( 1946). The design of optimum multifactorial experiments.
Biometrika , 33, 305- 325.
Plat ek. R .. and Singh. M. P. (1972). ome aspects o f redesign f the Canadian Labor Force Survey.
Proc. Soc. Stal. Secl. Amer. tal. Assoc .• 397-4()2.
, Politz. A. N .. and immon . . W. R. ( 194'1, 19501 An a ttempt to get the " no t a t homes" into the sample
without callback,. Jour. Amer.. tal. Assoc .. 44, 9- :1 1, and 45,136- 137.
Pritzker. L. . and Ha nson, R. ( 19(2). Meas urt:ment e rrors in the 1960 Ce nsus of Population . Proc. Soc.
Slat. Secl. A mer. SIal. ASSQ(·.. 80-89 .

Ouenouille. M. H. (1949). Problems in plane sa mpling. Ann. Malh . SIal., 20, 355-375.
Ouenouille. M. H . ( 1956). Notes on bia, in e tims tion . Biomelrika. 43, 353-360.

Raiffa . H .. and Schla ifer. R. (196 1). Applied Statistical Decision Theory. Harvard Business School.
ambridge . Mas .
Rand Corporation (1955). A Million Random Digits. Free Press. Glencoe, III.
Rao. C. R. (1971) . $9me aspects of statistica l inference in problems of sampling from finite
populations. Foundations of Slalisticallnference. V. P. Godambe and D . A . Sprott. (eds.). Holt.
Rinehart. and Winston. Toronto. Canada 177 202.
Rao. J. N. K. (1962) . On the estimation of the relative efficiency of sampling procedu res. A nn . Insl.
5 101. Malh .. 14, 143- J 50.
Rao. J . N. K. (I 965J. On two simple chemes of unequal probability sa mpling without replacement.
Jour. Ind. 5101. Assoc .. 3, 173-180.
Rao, J . N. K. (196(,) . Alternative estimators in pps sampling for mUltiple characteristics. Sankhya ,
A23, 47-60.
Rao. J . N. K. ( 1968). Some sma ll sample results in ratio and regression estimation. Jour. Ind. 5101.
Assoc" 6, 160- 168 .
Rao, J . N . K. (1969) . Ratio and regression estimators. New Developments ill Survey Sampling, N. L .
., Johnson and H. Smith. Jr. (eds.), John Wiley & Sons. New York. 2 n - 234.
Rao. J . N. K. ( 1973) . On double sampling for stratification and analytical surveys. Biometrika. 60,
125- 133 .
Rao. J . N . K . (1975a). On the foundations of survey sampling. In A Survey of latlstical Design
and Lillear Modt!ls. J. N. Srivastava (ed.). American Elsevier Publi shi ng Co. New York,
489- 505 .
Rao. J. N. K. (1 975b). Unbiased variance estimation for multistage designs. Sankhya (in press).
Rao. J. N. K .• and Bayless. D. L. ( 1969). An empirical study of the stabilities of e timators and variance
es timator~ in unequal probability sampling of two units per stratum . Jour. Amer. Stal. Assoc., 64,
540-559.
Rao. J. N. K., and Beegle. L. D . ( 1967). A Monte Carlo study of some ratio estimators. Sankhya, 829,
47-56 .
Rao. J. N. K.• and Graham . J . E . ( 1964). Rotation designs for sampling on repeated occasions. Jour.
Amer. . 101. Assoc., 59, 492- 509.
R o. J . N. K., a nd Kuzik , R . A. (1974) . Sampling errors in ratio estimation . .Indian Jour. Stal. 36, C,
43- 58.
Rao. J. N. K.. Hartley. H . 0 .. and Cochran, W. G . (1962). A simple procedure of unequal probability
sa mpling without replacement. Jour. Roy. SIal. Soc. 824, 482-491. I

Rao, J . N. K .• and Pereira. N. P. (1968). On double ratio estimators. Sankhya , AJO, 83-90.
Rao, J . N. K.. and Singh, M. P. ( 1973). On the choice of estimator in survey sampling. Australian Jour.
Slat .. )5, 95- 104 .
Rao, P. S. R . 5 ., and Mudholkar. G . . (1967). Generalized multivariate estimations for th mean of
finite populations. Jour. Amer. Stat. AfSOC., 62, 1oo8- JOI2.
 REFERENCES 409
Rao, P. S. R . S., and Rao, J. N. K. (1971). Small sample results for ratio estimators. Biomt!lrika, !8,
625-630.
Robson, D. S. (1952). Multiple sampling of attributes. Jour. Amer. Stal. Assoc., 47, 203-215 .
Robson, D. S. (1957). Applications of multivariate polykaYi to the theory oC unbiased ratio type
estimation. Jour. Amer. Stat. Assoc., 52, 51]-522.
Robson, D. S., and King, A. J. (1953). Double sampling and the Curtis impact survey. Cornell Univ.
Agr. Exp. Sta . Mem., 231.
Romig, H. G. (l952). 50-tOO Binomial Tables. John Wiley & Sons, New York.
Ror, J., and Cha.lcravarti, 1. M. (1960). Estimating the mean of a finite population. Ann. Math. Stat.,
31,392- 398.
Royall, R . M . (1968). An old approach to finite population sampling theory. lour. Amer. Stal. Assoc.,
63,1269-1279.
Royall. R . M. (I 970a). On finite population sampling theory under certain linear regression models.
BkJmetrika, 57, 377-387
Royall, R. M . (1970b). inite population sampling-<>n labels in estimation., Ann. MIlIh . Star., 41,
1774-1779.
Royall, R. M . (1971). Linear regression model in finite population sampling tbeory. Fou'llialions of
Statistical Inference, V. P. Godambe , and D. A. Sprott (eds.), Holt, Rinehart, & Winston,
Toronto, C'lnada, 259-279.
Royall . R. M .. and Herson. J. (1973). Robust estimation in finite populations, I. lour. Amt!f. Stat.
Assoc., 68, 880-889.

Sagen, O . K., Dunham, R. E ., and Simmons, W. R . (1959). Health statistics from record sour,ces and
household interviews compared . Proc. Soc. Sial. SeCI. Amer. Sial. Assoc. , 6-15.
Sampford, M. R. (1967 ). On sampling without replacement with unequal probabilities of selection.
Biometrika, 54, 499- 513 .
Sandelius, M . (1951 ). Truncated inverse binomial sampling. Ska'lliinavisk Aktuarietidskrift, 34.
41-44.
Siirndal, C. E . ( I Q72). Sample survey theory vs. general slatisticaltheory: Estimation of the population
mean, Rev. Inl. Stat. Insl., 40, )- 12.
Satterthwaite, F. E. (1946). An approximate distribution of estimates of variance components.
Biomlltrics, 2, I 10-114.
Scau , A. J., and Smith. T . M . F. (1974). Analysis of repeated surveys using time series methods. Jour.
A mer. SIal. Assoc.. 69, 674-{)78.
Sedransk, J. (1965 ). A double sampling scheme for analytical surveys. lour. Amt!r. Stat. Assoc., 60,
985- 1004.
Sedransk, J . (1967). Designing some multi-factor analytical studies. Jour. Amer. Stal. Assoc., 62,
1121- 1139.
Sen, A. R. (J 953). On the estimate of variance in sampling with varying probabilities. lour. Ind. Soc.
Agne. Slat., 5,119- 127.
Se n, A . R . (1972) . Successive sampling with p (p ~ 1) auxiliary variables. Ann. Math . Suu. , 43,
2031 - 2034 .
Sen, A. R. (1973a). Theory and application of sampling on repeated occasion with several auxiliary
variables. Biometrics, 29, 383- 385.
Sen, A . R . (1973b). Some theory of sampling on sucoe ive oa:asion . Australian Jour. Stal., 15,
105-110.
Seth, G . R ., and Rao, J. N. K. (1964) . On th comparison between simple random sampliog with and
without replacement. Sankhya, AU, 85- 86.
Sethi, V. K. (J 963). A note on optimum tratification tor e timating the population means. Australian
Jour. Stat., 5, 20-33 .
Sethi, V. K. (1965). On optimum pairing of units. Sankhya, 827, 315- 320.
410 SAMPLING TECHNIQUES

Simmons, W. R. (1954). A plan to account for "not-at-homes" by combining weighting and callbacks.
Jour. of Marketing, 11,42-53.
Singh, D., Jindal, K. K., and Garg, J. N., (1968). On modified systematic sampling. Biometrika, 55,
541-546.
Sittig, J. (1951). The economic choice of sampling system in acceptance sampling. Bull. Inl. Sial. IrUI.,
33, V, 51-84.
Sionim, M. J. (1960). Sampling in a Nutshell. Simon & Schuster, New York.
Smith, H. F. (1938). An empirical law describing heterogeneity in the yields of agricultural crops. Jour.
Agric. Sci. , 28, 1- 23.
Smith, T. M. F. (1976). The foundations of survey sampling: A Review. Jour. Roy. SIal. Soc., Al39.
183- 204 .
Snedecor, G . W., and Cochran, W. G . (1967). Statistical Melhods. lowa Stale University Press, Ames,
Iowa, sixth edition.
Srinath, K. P. (1970. Multiphase sampling in nonresponse problems. Jour. Amer. Slat. Assoc., ]6,
583-586.
Stein, C. (1945). A two-sample test for a linear hypothesis whose power is independent of the variance .
Ann. Math . Slat. , 16, 243- 258.
Stephan, F. F. (1941). Stratification in representative sampling. Jour. Marketing, 6, 38-46.
Stephan, F. F. (1945). The expected value and variance of the reciprocal and other negative powen; of
a positive Bernoulli variate. Ann. Math . Slat .. ]6, 50-61.
Stephan, F., and McCarthy, P. J . (1958). Sampling Opinions. John Wiley and Sons. New York , p. 243 .
Stuart, A. (1954). A simple presentation of optimum sampling results , Jour, Roy. Slat. Soc. , 816,
239- 241 .
Sukhatme, P. V, (1935). Contribution to the theory of the representative method. Supp. Jour. Roy.
Slat. Soc .. 2, 253- 268 .
Sukhatme, P. V. (1947). The problem of plot size in large-scale yield surveys. Jour. Amer. Stat. Assoc.,
42, 297-310.
Sukhatme, P. V. (1954). Sampling Theory of Surveys, With Applications. Iowa State College Press,
Ames, Iowa .
• ukhatme, P. V.. and Seth, G . R. (1952) , Non -sampling errors in surveys. Jour. Ind. Soc. Agr. Stat .• 4,
5-41 .
Sukhatme, P. V.. and Sukhatme, B. V. (1970). Sampling Theory of Surveys With Applications. Food
and Agriculture Organization, Rome, second edition.

Tepping, B. J., and Boland, K. L. (1972). Response variance in the Current Population Survey. U .S,
Bureau of the Census Working Paper No. 36, U. S. Government Printing Office, Washington.
D.C.
Tin. M. (1965). Comparison of some ratio estimators. Jour. Arner. Slat, Assoc. , 60, 294-307.
Trueblood, R. M., and Cyert, R . M. (1957). Sampling Techniques in Accounting, Prentice-Hall.
Englewood Cliffs, N.J.
Trussell, R. E. , and Elinson, J. (1959). Orronic Illness in a lArge City, Harvard University Press,
Cambridge. Mass .• pp. 339-370.
Tschuprow, A. A . (1923) , On the mathematical expectation of the moments of frequency distributions
in the case of correlated observations. Metron, 2, 461-493, 646-683.
Tu ey, J. W. (1950). Some sampling simplified. Jour. Amer. Slat. Assoc .• 45, 50 1- 5 19.
Tukey. J. W, (1958). Bias and confidence in not-quite larse samples. Ann. Math. Stat., 29,614.

U. N. Statistical Office (J 950). The preparation of sample survey reports. Stat. Papers Series C. No. I.
U. N. Statistical Office (1960). Sample Surveys of Cu"entlnUrest. Eighth Report.
U. S. Bureau of the Census. (1968). Evaluation and Research Program of the U.S. Qnsus of Population
and HOUSing, 1960: Effects of Interviews and Oew Leaders. Series ER 60. No.7. Washington ,
D.C.
 REFERENCES 411
Warner, S. L. (1965). Randomized response: A survey technique for eliminating e\lasive answer bias.
Jour. Amer. SUl/. Assoc. , 60, 63-69.
Warner, S. L. (197 1). The linear randomized respolh'ie model. Jour. Amer. SUl/. As.roc.,66, 884-,I!88.
Watson, D. J. (1937). The estimation of leaf areas. Jour. Agr. Sci., 27, 474.
West, O . M . (1951). Th<! Results of Applying a Simple Random Sampling Process to Farm Management
DaUl. Agncultural Experiment Station. ('.om ell University.
Williams, W. H. (1963). The precision of some unbiased regression estimators. 8iometrilro, 17.
267-274.
Wishart, J. (1952). Moment-coefficients of the k -statistics in samples from a finite population.
Biometrika, 39, 1- 13.
Wold, H . O . A. (1954). A Study in lhe Analysis of Stationary TIme Series. Almqvist and Wicksell,
StockhOlm, second edition .
Woodruff, R. S. (1959). The use of ro tating samples in the Census Bureau's Monthly Surveys. Proc.
Soc. Stat. Sect. Amer. Slat. Assoc., 130-138.
Woodruff, R. S. (1971). A Simple method for approximating the variance of a complicated estimate.
Jour. Amer. Slat. Assoc .• 66, 41 1-41 4.
Woolsey, T. D . (1956). Sampling methods for a small household survey. PIlb. Health Moru>graphs, No.
40.

Yates, F. (l948). Systematic sampling. Phil. Trans. Roy. Soc. London, AUI . 345-377 .
Yates, F. (1960). Sampling Methods for Censuses and Surveys. Otarles Griffin and Co., London, third
edition.
Yates, F., and Grundy, P. M. (1953). Selection without replacementfrom within strata with probability
proportio nal 10 siu. Jour. R oy. Sial. Soc., 815, 25J-26 J.

Zarkovic, S. S. (1960). On the effici e ncy of sampling with variou~ probabilities and the selection of
units with replacement. Metrika, 3, 53-60.
Zukllovitsky. S. I., and Avdeyeva. L. r. (1966). Liluar and Convex Programming. W. B. Saunders.
Philadelphia.
Answers to Exercises
1.1 (a) Examples of problems of definition are decisions whether to count words in a preface.' or
index and how mathematical symbols are treated as "words." In a book such as this one with many
mathematical symbols, however. it seems unlikely that a cou nt of words, either including or omitting
symbols, would be wanted. (b) (1) The pages constitute a convenient trame. A disadvantage of the
page as a sampling unit in which we count all words on any sample page is that with numerous
illuslrations the number of words per page may be quite variable because of the incomplete pages. It
may be worthwhile first to list all the incomplete pages, formi ng two 5ubpopulations or strata, one of
incomplete pages and one of complete pages. that are sampled separately by the method of stratified
sa mpling described in Chapter 5. (2) A problem with the line is that obtai ning a listing of lines so that
lines can be sampled directly is time-consuming. Also. there are incomplete lines at the end of
paragraphs. Since words per line should be fairly stable, however, the solution may be to use two-stage
sampling (Chapter 11), first drawing a sample of pages and then counting the number of lines on each
selected page and drawing a subsample of lines on these pages.
1.1 This question supposes that we first draw a sample of cards with equal probability. (a) If
sampJe names not in the target population are discarded. the only problem is that the size of the sample
of names from the target population will generally be less th an the number of cards and will be a
random variable, depending on tbe cards that bappen to be chosen. (b) The problem is that names
appearing on several cards have higher probabilities of selection. One way of handling tbis is to count
tbe number of cards on which a selected name appears and usc this number in making the e limate by
Ij)ethods appropriate to se lection with unequal probabilities (Chapter 9A). Another way that gives
each name an equal chance but may involve many rejections is to retain a card only if it is the first of the
set on which this name appears. (c) As in (b), names are being selected witb unequal probabilities. I
know of no easy method of giVing each name an equal chance. If the number of cards on which each
name appears has been recorded somewhere, an unequal-probability method can be used . as in (b) .
1.3 Suggestions are : (a ) a recent directory of department and luggage stores . (b ) the repositories
for lost article~ maintained by the subway and bus companies, and (c) hospitals and private physicians
in the geographical area in which snake bites occur, plus any public health organization to which
reporting of bites is compulsory. Weaknesses in all three frames are likely to be incompleteness. plus
bigh costs in (c) if snake bites are rare and not centrally reported. (d ) A list of households is often used
as a frame for selecti ng a sample of families . Although there will be some incompletene s (families who
cannot be reached). [he major problem may be errors in measurement.
1.4 A problem is incompleteness because of new construction. In a sample of addresses, new
dwellings can usually be bandied by the interviewer, who checks, for any sample addres , whether
there are new dwellings between this address and the next address in the directory and . if so, includes
these new dwellings in the sample. Whole areas of new construction may not be mentioned in the
directory and require development of a separate frame. Drawing a list of addresses is preferable to
drawing a list of persons, since addresses are more permanent. For reasons of travel expense, however.
the sampling unit may be a city block from which a subsample of dwellings is elected.
l.! S80,39O and S82,970.
1.6 The confidence probability is about 0.054 (found from t '" - 1.67 with 25 tlegrees
freedom). This assumes tbat future receipts follow the same frequency distribution as the.la.mple of 26
0'
receipts, and that this distribution is normal.

412
 ANSWERS TO EXERCISES 4 13

1.7 When the MSE i~ due entirely to bias. the estimate is always wrong by IJMSE. The
probability of an error ~JJMSE 15 therefore unity and th e probability of an error 2: 1. 96../MSE or
2: 2.576JMSE is zero.

2.4 Y- 51.473 . Probabi lity about 0.9.

1.5 Yes.0'( Y) is '18.4.

1.6 Y ~ 20,238•.11 Y) .. 849.

1.7 ( a ) Public: R = 15.46. Private: R "12.75. (b) Public: s( R )" 0.761. Private : s(RI-0.727 .
For the fpc we take f - 100/ 468. (c) 14 .2< R < 16.7.
2.8 Diff ./s.e'di~~ 2.71 / 1 186 - 2.28. P about 0.023. Note that the fpc IS not used 10 computing
s.e·dl"·
1.9 (a ) 9408. s.c. - 780: (b) 9472 . s.e. v 1104 .
2.10 S.c. (in 1000'~) - (aI1 4.800: (b) 31100: (e) 3140.
2.11 Q 2. (a) 2.7; (b) 2 .4 .
2.12 (01 II ~ 60 . with 30 from each domain : (b ) 11 - 110 wif! do if the number of owners in the
'ample he~ Hllywh e,~ l,,,lweeo 20 nod ',0. W"h n .. 80, Ihe proh:.bility that this happens is only ~hout
0.54 (from the binomIal table'). With n ~ 100. any samp le number of owner' between 19 and 8J will
, uihoe. the probability that this happens being about 0.94..
1.14 (0) 420. (b) 490; (c) bot h lire unbiased ; (d) estimate (h)

3.2 1066 1334 a\ gl\ ~r. by the normal .. pproximat,on. equation 319.
3.3 NClirly conclusive.
3.6 ta, 16.2 oj: 3.6%. (b l J 738± ZIW famihe
3.7 171j(l:l 268 famih e .
3.8 A~ an exact result.
VlAoJ N ,lnO ,
~~ ~ 11 ;(j': 1T)(O t+ Pl 1T)
" w N, = N(1 - 1Tl. and ,n large samples II, "'. ,, (1 - 1T) . These <ub~titulions give the <tated res ult . In
order Ihal V(A , ) V{A , 'I he small. we mu't hav€' ?T() - 0, )/0 , 18rg< This means that 0 , musl br
smd ll: in o th er words. rhe proportion of domain 1 rhal 1ie~ in cia ~ C musr be large. For !!lveo 0 ,. 1r
shou ld be large .
3.9 All give A u = 13 . By th e hypergeometric, the prohability of no UOlI< ,n C in the sample is
0.060 1 {or A -12 and 0.0434 for A,, -13. B the binomial. P p -04507 and.j'J-fPu = 0 .4114.
giving A u'" 12.3. Page 59, Ex . 3 gives 0.061 (or Au "" 12 and 0.044 for All '" 13 .
3.U E limal c (b) see ms more preci~ .

3.U nIt' hiihest va lue is POIII as compared with PO/ "," by Ih(' binomIa l formu la. Thi~ occurs
when evny cluster consists entirely of I ' orenrirely of 0'5. The lowest value can be zero if every cluster
gives the arne proportion P. (This is possible only for certa,n va lues of P and m).
3.13 VarianI' ,. 0 .0(1 184 b} the ratio method and 0.00160 h) !'be t>inofllial formula .
3.14 Average size of sample - ml P

4.J 735 houses. This sample size is needed for Iwo-car household il P ~ I n%
4.2 About 260 sheets.
4.3 (a ) 2475; (b) 4950 .
4.4 ,, - 2 1 (laking t - 2).
4.5 11 - 484 . For number of unemployed. tbe cv uld be abou t 1~% .
414 SAMPLING TECHNIQUes

4.6 62 more.
4.7 (a) n - 278; (b) n - 2315; (c)" - 3046.
4.8 If a rectangular distribution i~ anumed within each class, we talce 52 = 0.083h 2 or 5 .. 0.29h.
This gives estimates of 230, 580, 2030. and 11,600 in the four classes. If the right-triangular
di~tribution is used in the fourth claloS, we take 5 - 0.24h, giving 9600 for this class.

UNS~2/3
4.11 nup, ~ ( --=
2 c J2
4.12 (a) n '" 1250; (b) n ~ 679. In this part, we can give a dummy variate y, the value +100 for a
(YtlS, No) answer. -100 for a (No, Y ) a nswer, and 0 otherwise. Thcn Y=P 1 - P2 in perccntages.
With an advance sample of nl '" 200, formula (4.7) in Section 4.7 can be used, givmg n - 679.
4.13 (a) The probability that a family of four persons blS 1,2. and 3 females is approximately 1/4,
1/1.. and 1/4. respectively. For estim.~tmg the proportion P of females, a simple random sample of n
families gives VIP) = 0.03125/n. PS agllin't 0.0625/n for a sample of 4n penons. The dtff factor is
about 1/2. In the corresponding CYJI.llplc In Table 3.5 with 30 households of unequal SIZes, the
eSllmated del! factor Wb 0.475 . (bl fne dt:ff factor would be slightly raised by famWes witb identical
twins, sInce the prop rtions of families With I and 3 females would be sllahtly increased.
5.1 (a) Neyman allocation give. /II = 0.87. n2 = 3. 13. (b) There are three possible estimates
und", optimllm 11I1..-':lIlIon and nine under proportional allocation. V_(ii,,) - i .. 0.167 : V"",p(ii.. ) -
11 " 0.583 . (.I) ForlJl uht 5.27 gives V..,(9 .. ) a 0.159.
5.2 (a)" , ~3 -15 . /1 2= 625 ; (b)" 1 &750. n ~- 250 .
5.3 RP - 181.,u fOl propo{tiollllJ allocatiol' lind 214% for optimum allocation.
5.5 Whc;:n W , - W . the relative incceu"s equal 0 .029 for Cl/CI " 2 and 0 .111 for C2/C} - 4 .
.5.6 t
(a ) II;!" .. t /I . / n - (b l n '" 264, nl " 88, n2 - 176; (c) 11936.
5.7 (a ) S2lf,!! agai nst S 1936. tb l No . The minimum field OO1It to reduce V to 1 is 52230.
5.8 ( 0 I n , " 384. 11 2 '" 192; (b) "1 '" 400, " . - 1600; (c) n,- 1200, n2 - 2400 .
" .5.9 Frll florlal ,"crea~e ~ ~.
5.10 n1 '" 's41, IIl - 313, n3 - 146.
5.12 In population 1. v_ -0.143/n; V.... -0.134/n. In population 2, V_ - 0.0491/n, V.... -
0 .0423/0. Tt t I·cdu.:tion in variance from optimum allocation is about 6% in population 1 as against
14% in popUlation 2.
5.14 (a) 11 we guess P, - 4S%, P."'2S%, P,-7 .S% 8S a compromise, this give. nl -268,
n2 - 116, n, ~ :\ 6 : (b ) s.c. '" 0.0225; (c) s.e. '" 0.0241.
5.15 As II uoproaches N. 8 stage is reached in which the standard formula n~ OCN~5~ for Neyman
optimum allOCll tivn is no longer applicable. since it would require n~ > N" in at least one stratum. As
noted in SC(;tion .8, formula (5 .27) then ceases to hold. The student isin error if he claims that (S.27)1s
always wrone; the formu la has a limited range which , however. covers nearly all applications.
SA.4 No. In each of the wont cases [I(w" - Wit) Y,,)2 is (0.105)2 ~ 0.0110. Thus, with stratifica-
tion, M~E(.y..,), as given by formula (5A.6), is 0.0108+0.0110-0,0218. With simple random
samplina, V(Y) >= O.Ol77.
SA.6 (a) n ... 1024. The optimum allocation for the teOOnd variate (average amount invested)
lalistie£ both requirements.
SA. 7 WI - 0.728, W1 - 0.272, S, '" 1.806, S2 - 4.698 (in the coded scale). (a) The optimum
sample sizes are ", - 0.50711, 111- 0.49311. (b) V(y ) .. 31.9S/n, V ...,(YoU) " 6.72/n.
SA.8 (b) r../1fYJ
o
dy - r· J2(1- y)d,- 2J'2[1-(1- a)3/2]/3. Heneewewaot[l-(1 -a)'/2]-i.
Jo
SA.9 In Exercise SA.7" WI - 0.728, as In the DaleniUII-HodgC$ rule, comes closest to •• tisfying
the Ekman rule. In Exercise SA.8. the Ekman rule gives a - (3 -..fS)/2 - 0.38.
 ANSWBRS TO EXERCISES 4J5

!lA.10 The optima are L .. 7 for p - 0.95, L '" 5 for p = 0.9, and L - 4 for p e 0.8. Ei ther LaS or
L - 6 i, a good compromise.
SA.ll (a) Gain in precision is about 110%. (b) Gain from proportional stratification oversimpl~
random sampling is about 90%
SA.l1 Increase "I as the hint suggests, leave " ," 400. We require nl . )40. giving n '"' 540.

6.1 For the ratio estimate VCYR ) ~ N'(1- nS/ In and for simple expansion VO?') c
2
N ( I ._ nS, 2/". where d : ()I - RIC). For the sample of 21 households the eMimates of S/ and S, 2 are as
follows . Number of children, si
0 .49,K s/
1.61 ; number of cars, . / a 0.41,
8 s/
= 0.39; number of
TV se ts, s/ - O.S 1, s/ -
0.45 . The ratio estimate appears superior for children.
6.2 Gain - 66% . At leal 11 units b)llhc ral;o melhod.
6.3 Quadratic limits (27.100, 29,870); normal limits (27,030, 29.700).
6.4 Apply theorem 6.3 to the estimation of R "" YIX. With large .s amples, use YI X if r s (cv of
IC)/ 2(cv of )I), and use iii otherwise, where r and the cv's are sample estimates.
6.5 The MSE's are 46.5 for the separate ratio estimate and 40.6 for tbe combined ratio estimate.
In both cases the contribution of bias to the MSE is negligible.
6.6 For Lahiri's method, vel>Ib) ~ 40.1.
6.7 Estimated population total z 116.21 millions. The relative variance is 0.00111 , so that the
s.e. is (0.0333)(116.21) - 3.87 millions. The esti mate is within I s.e. of the true total.
6.8 The estimates are (a) 1896, (b) 1660, (c ) 1689. In (c ) we find WI'" 2.38. - 1.38. w, B

Estimated s.e.'s are (a ) 256, (b) 36.9, (c) 18.6. For the s.e. in (b) I used tbe formula S.e. '"
~RJ(l - n(e.. +CIl - 2e. '>ln, where YR is tbe ratio estimate of Y. that ;s, 1660. For the S.e . in (c ) I
used Y"MR - 1689.

7.1 Estimate" 11.080; s.e. - ) 52 (including tbe fpc) .

7.1 No. since b is very close to 1.
7.3 Y" ....28,)'7 7 ± 570. The relative precision is· 113% .
7." 27,751 * 694.
7.6 For the difference estimate, V(y) - S/ I .. , for the linear regression estimale, V(y,, ) ~
s.2 S121.. (S/ + S/) . The regression estimate has the smaUer variance, but its superiority i» uaWaportant
if S. I S/is small.
7.7 MSECY".) - 34.5, Bias'm 9 .7; MSE(Y,rc) - 11.9, Bias' -1.2.
7.9 MSE(YIb ) - 8.9. MSE('YRr )-6.7.

8.1 Variances are 8.19 (systematic), 11.27 (simple random), 8.25 (stratified, 2), 7.46 (strati-
fied,I) .
8.1 V, •• - 0 .00141 , V ••• - 0.00340.
8.3 The systematic sample should be superior for the proportion of people of Polish de cent,
since this variable exhibits geographical stratification. It is likely to be inferior for proportion of
children because the sampling interval, 1 in 5, coincides with the average size of a household . The same
is true. though to a smaller extent, for proportion of males.
8.4 The variances are as follows. Males, V... - 0 .0204, V... - 0.0216; children , V... - 0.0204 ,
v. •. - 0.0776; professional, v... -
0.0192, V", - 0.00 J 6.
8.5 Actual variance . 8.19. Method (a) gives 11.29. For method (b) the estimated variance from
a single sample is (1 - n(ji'l - ji,,)' / 4, where YIl' )i'2 are the means of the two halves. The average is
3.24. The serious underestimation is unexpec d.
8.7 80th variances are (k' - 1)/6.
 416 SAMPLING TECHNIQUES

8.8 Simple random samplin g is bener unless II - 1 or k '" 1.

8.9 Every kth unit, V( h -n
= 362.2; Yates, MSE( = 7.3 ; Sethi, V( y) = 21.0 ; SinSh et al. .
V( Y) = 8 LO. The last two estimators are unbiased in this eumple.

9.1 Relative costs of using the four types of unit are a 100. 90.1. 79.7. and 77.8 (takinll, the first
unit as a standard) .
9.2 Relative precision of the household is 211 % for the sex ratio and 38% for the proportion who
had seen a doctor.
9.1 Relative precision of the large unit is 0.566 with simple random sampling and 0 .625 with
stratified random sampling.
9.5 ( a)M~5:(b ) M"1.
9.6 The optimum M should decrease because travel cost. which varies as,r,., becomes relatively
less important when II increases.

9A.l (a) 34.242 : (b) 5534 : (d 6493 .

9.4..3 (a) If the s.d. among large units in class II oc M~ . (b l If probability oc.JM;,.
9.4..4 (b ) V ( Y/-IT)= 1.75 , V ( Y.. l '" 0. 27, V ( YRHCl= 0.33 . fn tltis problem y./z, va ies httle , but
Ytn performs relativel poorly because with the method of sample selection ...., '" 2z , Murthy 's
estimator is devised for this method of sample selection having V( YoW )" 0 if y./ z, is constant, as
(9A.60) shows , (c l V(YM)/V(Y"", ) ~ 0.54, while (N - II)/(N - I) " ! ,
9A.!; MSE( y/~, ,, ,) '" 7.06. (Bias)~ / MSE = 0.065 . V( YSRS ) '" 6.5.

10.1 (a ) 2.00 ; (b ) 2. D .
10.3 (a ) 165 / n ; ( 1) ) 148.5/ n ; (c) 132 / n.
10.4 (a) II '"' 660 fields ; (b ) ,, = 530 field~ . Protein requires fewer fields than yield,
10.5 C, / C2- 8.
" 11>.7 (a) 0,93 %; (b l 0.5 1% : (c) 0.36%.
10.8 ( a ) E,ther m il = 7 o r mn = 8 is suitable ; (b ) 89% for rn o = 7 and 93 % for m o ~ 8; (c) 86% for
,
mo = 7 and 89% for m " = 8.

11.2 The relat.ive precision of ill to II drops from 3,02 to 2.75 . If two sampling plan differ
primarily in their between-units contribution to the variance. the relative precision of the superior plan
will in general decrease as the ratio of the within-units variance to the total variance increases,
11.3 The explanation is. roughly speaking. that with these data the Y.! z, are more stable than the
Yi / M ,. If we look t, - fl. 13.
and H.
the between-units contribution in method IV would vanish
11.4 Total variance: 0.00504 (10), 0.02358 (11). 0,00554 (Ill).
11.6 Estimated percentage J4 .2±2, J6. Estimated number 3540±540,
n.7 Estimated percentage 13 .9±2.49,
11.9 Total rooms, 2Q,400, total persons, 50,550, persons per room, 1.72; s,e.'s: lotal persons.
2,440, pe~ns per room 0,066 ,

12.1 II'" 267 , II ' = 1320 or II - 268. n' - 1280. V(PII) with optimum allocation i 6.67 wh.en PIt is
in % 's. With single sampling. V ( p.) - 8,33.
12.:Z c./c.' > 9.
12.3 "/11' = 1/19.
12.4 ,,' >1611 ,
11.5 By formula (12.67). s.e. = 1.25, ignoring 1/ N.
12.6 Per cent gain {rom the second to the sixth occasion are 50, 75,91. 100, and 105, respectively.
 ANSWERS TO EXERCISES 417

11.' The values 0111 V(Y2")/5 ' and n V(~,')/ 5' are a~ lollows: I' -:t () -
0.8 : 0.885. 0.875; I' -~ ;
p - 0.9: 0.843. 0.840; I' - i.
p. 0.8: 0.824, 0.810; I' '' !. p m 0.9: 0.752. 0.746.
n.n (a) LeI y, - I for any unit that has the tirst attnbutc and y, - 0 otherwise and let stratum 1 be
the stratum in which every unit has the second attribute. In the notation 01 theorems 12.1. 12.2. with
I/N neglisible. 5' - PIQI and 5 1' - PIP,/(P) + P1)" Results in (a) follow from the theorems ; (b) if
c> I, the expected cost. double samplina with "I - 1/2. k - 2 (the optimum) sives V( Y,.) ..
N 1 (0 .844)/C>. while a simple random sample lives ven ..N ' (l .875)/C· . over twice as large.

13.1 (c) 90% response with 1047 completed questionnaires or 95% respOnse with 701 completed
Questionnaires.
13.2 The method with 90% response costs $5235 . ThaI with 95% re ponse costs $5.7895 per
completed questionnaire. or 54058 total cost.
13.3 (b) 0.6Sno. 0.815no. 0.8915no. 0.9351no. 0.9611no; (c) 100, 104, 108. 112. 117; (d ) 300.
288.277.267.256.
13.4 (a) Bias (in %). - 3.85. - 2.15, - 1.21. - 0.69. - 0.40; (b) varianceSarc 8.67. 9.03 . 9.39. 9.74,
10.16; (c) four calls.
13.6 Yes. VC for k .. 2 Is only about 2% over the minimum Vc.
13.7 Politz-Simmons estimate. 39.7·%; binomial. 42.3% .
13.' (c) Variance - 0.1.
13.9 II each enumerator's error of measurement were independent from family to family , the
variance of the sample mean would be (tT• .' + tT/ + 1711 , + (711)/525 instcad of (17,/ + o} ... 35171/ +
\()Su/}/SlS . En\lmerator biasc\ contrib\lte abo\lt 55% of the tota\ "aritonce .
13.10 10·V( ..... )- V(100"... ) - (a) 1.80 (Direct); (b ) 10.69 (Warner ); (c ) 3.30 ( 7T" krown); (d )
5.12 (Moo"); (t ) 6.30 (P, " 1- PI) '
13.11 (a ) 10·MSE(..... ) - 3.81 ; .fr"u is superior if 7T u is known . (b ) 10·MSE( ..... ) '" 5.47 ; the
two-question ..... u 15 also superior if P, - O. (c ) 10·MSE(,.... ) . 7,64. All methods arc superior except
Warner's orisinal method.
 Author Index

Abernathy , J.R. , 395 Co~. D. R., 79

Arvesen. J. N.. 179 Cyen. R.M.. 12
Avadhani , M.S .. 48 . 127
Avdeyeva , L.l., 122 Da~~;lu~. T .. 122. 127. 129. DO. 131. 134.J~~.

Bailar. B.A., 392 Das. A.C., 228

Balakrishnan, T.R. , 130, 134 DaVid, J.P .. 161
Barr, A . , 378 DeLury . D.B.. 57. 59
Banholomew , D.J ., 376 Deming. W. E. , 4. 81. 148 . 321. 367 . .176 ..WI
Basu, 0 . ,30 Des RaJ. 30. 256. 26J . JOO. 309
Bayles5, D.L. , 265. 267, 269, 270 Dowling, T. A.. 94
Beale, E. M.L., 176, 178 Dunham , " . E.. 385
bean, J.A.. 135, 171.324 Durbin. 1.. 142. 144. 175.261.262,263 . 300.
8eardwood , J .. 96 310. 325. 366, 373. 376
Beegle. L.D . , 180
BeJlhouse, D.R .. 2 16. 217 Eclclcr. A R .355
Bellae , N.B., 385 Ekman. G . • 130
Bemen. E. H., 131 Elinson 1., 385
Bershad . M.. 354. 378. 381, 383. 386 ErdOs. P.. 39
Birnbaum, Z.W.. 363 Evans. W D., 11 7
Blac~ . C. A.. 228 233
Bly.Ihe . R.H., 84, 88 Fellegi. I.. 263. 279. 387. 388 . 391
n oland, K.L .. 389 Feller, W.. 39
Booth . G., 122, 123.335.336 Fieller. E.C .. 156
Bose. Chameli. 340 Finkner, A.L., 172.239,360
Brewer. K.W.R. , 158 , 256.261.310 Finney, D.J .. 7 1, 222
Brooks. 5 .. 282. 285 Fisher, R.A.. 42 . 59. 221
Bryant , E.C . • 121\ Fitzpatnck. T.B .. 126. 127
llunnan. J .P. , 320 Foreman, E.K.. 256
Burmeister , L.F. , 146 Frankel. M.R . 321.322
Burstein. H .. 59 Fuller, W. A.. 140. 146. 149

Cameron, I .M. , 28 1 Gallup . G .. 4

Chakravani . I.M . , 48 Garg, 1 .N .. 217
Chatterjee. S., 12 1. 122, 147 Gautschi. W. , 226
Chung, J.H. , 57, 59 Godambe. V.P., 45 . 48
Cochnin , W.G" 19. 130. 131,132, 197 .220, Goodman. R.. 126
226. 256 , 266 , 361 Graham . 1. E.• 354
Colerr.an, 1.5 ., 4 Gray . P.G ...185
Cornell. F.G. , 106 Greenberg. B. G., 394 . 395
'Cornfield. J , 28, 80, 247, 37<1 Grund y. P M.. 84
Corlcll, T., 292 Gurney. M. , IJO, 165. 172,26 1
419
 AUT HOR IND EX

H"good. M.. 131 Koop . J!'C .. 163 .39 1

Hajc~ . J ..39 Kulldorf . G .. 348
Haldane. J.B.S.. 77 Kuzik . R. A.• 178
Hullon . J. h .. 400
HammcI'olc}. J. M.. 400 Lahiri . D.B.. 175.206,251
Hunif. M.. 26 1. 310 Ldnsing. J.B.• 383
Hun,en. M. H .. 22. 68. 139. 145. 155. 165. 172. Lieberman . G') .. 57
24-'.246.250 . 283.297.314.354.355.370. Lienau. C.C .. 378
~72. 378 . 381. 383. 386. 392 Lund . R.E. . 146
Han"lO . R.. 387 . ~88
Hartley . H.O .. 45 . 122 . 124 . 139. 142. 145. 146. McCarthy. P,) .. 136.320. 32 1.365.386
162. 174 . 266.374
McCarty. D. E.. 131
Hayne • . J. D .. 227 McVay. F. E .. 243. 247
H~ndnck •. W. E.. 243. 244. 377
Mackenzie . W.A.. 221
Her,"'n. J.. 160 Madow. L.H .. 210.213 . 21M
He". I .. 126. In 134. 183.364 .377 Madow . W.G .. 39.139. 145.206.210.2 13.
H,lCki ng. R.R .. 122 243. 246. 265. 21n. 354. 355. 385
Hodge •. J.L .. 129 MahaJanobi s. P. C.. 243 . 274. 321. 378. 388
Hocffding. W.. 179 Matern, B., 219. 221. 226. 227
Homeyer. P.G . 228. 233 Mi c ~ e)'. M.R .. 175.203
Horvitz. D.G .. 44. 259 . 272. 393. 395 Midzuno. H.. 175
Houseman . E. E.. 38. 102 Milne. A.• 218. 227
Huddleston . H.F .. 122 Monroe . RJ ., 239
Hurwitz . W. N.. 22. 68 . 139.145.165 . 172.243 . Moors. J,) . A.. 394
246.250.28:1.297.314 .354.355. 370.372, Morgun. J.J .. 239
378 . 381. 383. 386 Morton . R.. 394
Hutchin'on. M.C .. 17R Mosteller. F.. 361
Hr:nan . H.H.. 6 Mudholkar, G.S.• 186
Murthy, M. N.. 22. 30.130.206.217.263.265
Jubinl·. T. B. . 145
Jume,. A. T .. 157
Namboodiri . N.K . 162
Jcbc. E H.. 297
Neter. J .. 4
Jen,en. A .. II
Ney man . J.• 99. 101 .3 27
k"en . RJ .. 6.38.102.119 . 124.1 60.237.243 .
NordboHen . S .. 149
2oW. 346
Nordin, J. A.. 84
Jindal. K.K.. 217
Johnson. F. A.. 222. 225. 235
Jone, . H.W.. 39 1 Olkin. l. , 184. 186
Osborne, J.G .. 22 1. 222. 226
Kcmpthorne. 0 .. 45 Owen. D. B.• S7
Kendall. M.G .• 19
Keylitz. N.. 169. 170. 184 Patterson. H.D .. 348
Khan" ~ . S.H .. 30 Paulson. E.• 156
Khan . S .. 340 Payne. S. L.• 6
Kiefer . G .• 139 Pereira. P., 183
King . A.J .• 131.330 Pillal. R.K .• 162
KI>er. C. V.. 4 Plackett, R.L. . 320
KI~h . L.. 68 . 126. 162. 183.242. 32 1. 322,325. Platek . R .• 309
3M . 377. 38:l Poli ll. A. N.. 374
Kl~h . G .. 379 Pritzker, L .. 379. 386. 3R7 . 388
Kokan . A. R.. 122 . 149
K,,<IO'. D.A.. 386. 388 Quenouille, M.H., 175, 178.22 1. 228
 AUTHOR IND EX 421

Raiffa . H .. 85 Snedecor. G. W.. 19

Rao . C. R., 45 Srinath. K.P .. 330
Rao. J. N.K.. 45. 136 . 139. 155. 163. 178 . 179. Stein. C .. 78
180 .183. 197.203.216.2 17.261.262.266. Stephan. F .. 118. 131. 135 , 136.365.386
267.269.270.300.311.330.337.338.354 Stuan. A .. 97. 366
R~o . P.S. R.S . 155 . 178 . 186 Sukhatme. B.V ., 48, 117, 127 , 161
Renyi. A., 39 Sukhatme. P.V .. 117. 163.308.326.378
Riedel. D.C .• 126 . 127
Robson, D.S . • 330 Tepping. BJ .• 389
Romig . H.G . • 55. 59 Thompson . D.J .. 44 , 259. 272
Ross, A .. 162. 174 Tin. M .• 176. 178
Roy. J .. 48 Tripathi. T.P . • 340'
Royall. R. M .• 45. 48.15 . 160.200.339 Trueblood . R .M., 12
Trussell. R. E. • 385
Sagen. O.K.. 385 Tschuprow. A.A. , 99
S~mpford. M.R .. 262. 263 Tukey . J.W. , 29. 178 .36 1
Sande1ius. M .. 71
Sarndal. C. E.. 49 Venables, W. N .. 157
Sunenhwaite. F .E.. 96
Sch1alfer. R .. 85 Waksberg. J .• 387. 392
Scott. A.J .• 355 Warner . S.L.. 392. 393, 395
Sedransk. J . 122. 123 . 335. 336 Watson. OJ .. 189
Sen. A. R .. 261. 348 West. Q.M . , 44
cth. G R.. 30. 37R Whelpton. P.K .• 4
Sethi. V.K.. 130. 134. 217 Wilkinson. G .• 157
Shuchtman. R.H . . 394 Williams. W. A .• 203
Shah. B. V .. 393 Wishan. J. , 29
S,mm m•. W. R .. 148. 374. 385. 393 Wold. H .O .A. , 221
Singh. D .. 217 Woodruff, R.S .• 170. 319. 355
Singh. M.P . 261. 309 Woolsey . T. D. , 364
Sirken. M.G .. 363
Sittig. J .. 85 Yates, F . . 59. 84. 121. 122. 142 . 184. lR9. 216.
Slonim . M.J .. 4 217.222.226.228.256,261.348
Smith. B.B .. 19
Smith. " .F .. 244 Zarcovic. S.S .. 256
Smith . T .M. F .. 45 . 355 ZukhovilSkay. S.L., 122
Subject Index
Acceptance sampling, 3 variance of mean per element, 240
Advantages of sampling, 1- 2 clusters of unequal sizes, 249
Aligned systematic sample, 228 comparison of ~ u, ~ R, ~ ""', 255-258
Analytical surveys, 4. Su also Domains of study equal-probability selection, unbi sed
Area sample, 89 estimator Yu, 249
Attributes, su Proportions ratio estimator YR, 250
1('
Autocorrelated populations, 219 Horvitz-Thompson estimator, 2.59-261
optimum selection probabilities, 25.5
Balanced repeated replications, 320 ratio estimators , 27 1
Best linear unbia ed estimate, 1.58 selection:
Bias: methods related to syStemallc sampling, "
definition, 13 265
deliberate use of biased procedure , 12, 15 without replacement (WOR' , 258-259 '
.1
due to errors in stratum weights, 117 selection WOR, Brewer's method, 261
due to nonre ponse, 359-361 comparison of methods , 270
of interviewer, 378 Murthy ' s method, 263
of ratio estimator, 160 Rao, Hartley , Cochran method , 266
of regression estimator, 198 in str..tifJed sampling , 270
in small areal units, 233 unequal -probabil ity selection with replace·
Binomial distribution, 59 ment (WR), ~ pp'., 251- 255
confidence limi ts, 59 Cl uster sampl ing (with subsamplingl, su
• erroneous use in cl uster sampling , 64-68 Two-stage sampling
references to tables, 5.5 Coefficient of variation(cv l , .54
for sample proportions, 55 Collapsed strata method, 139, 227
Boundaries of strata , rule for determining , Combined ratio estimate, 165
127- 130 compared with separate ratio estimate, 161
estimated variance, 167
Call-backs. 365 in estimating domain means , 143
effects on response ratio, 365 optimum allocation for, 112
mathematical model , 367 in stratified two-stage sampling. 318
optimum policy, 369-370 upper bound to relative bias. 166
relative cost, 366-367 variance, 166
Cauchy-Schwan inequality, 97 Combined regression estimate, 202
Census, use of sampling, 2 compared with separ-..te regression e timate. 203
Centrally-located systematic samples, 205 estimated variance, 203
Change, estimates of, 345, 3.52--353 variance, 202
Cluster samplina (single-stale): Composite estimate. in repeated sampling, 3.54
CIUSlCn c:I. equal sizes, 233 Conditional distribution, of proportions, 61
compared with SRS of elements, 242 Confidence limits. 12
estimation of proportions, 64-65, 246 conditional, 60
optimum size and shape of cluster, 234-240, definition for attributes. 57
246 effect of bias on, 14
variance as function of size of cluster, 243 effect of nonresponse on, 361- 363
422
 SUBJECT INDEX 423

Pop 'f means in simple random sampling. 27 Double sampling. 327

sa r proponions or percentages. 57-(J) ratio estimates. 343-344
(ur ratio estimates. 156 Double sampling with regression estimates, 335
in stratified random sampling. 95 comparison with simple random ampling. 141
validity of normal approxi mation . 39-44 estimated variance . 343
Consistency: optimum sample sizes. 341
definition. 2 I variance. 339-340
of mean of simple random sample. 21 Double sampling with slrdtification. 327
of ratio estimate ; 153 comp~rison with simple random sam-)ling. 332
of regressIOn estimate. 190 estimated variance. 333
Cor rolled selection. 124 estimation of pr ·ons. 330
Co. eetion for continuity. 58. 64 optimum ample sizes. 331
"e. Teet value." 377 variance. 328
(orrelation coefficient:
in finite population~. 154 E. average over all possible samples. II
intracluster . 209 E , . El. averages over first and second stage. 275
within a systematic sample, 209 Elements. 31
Correlogram. 219 End corrections. 216
('''st function : Errors in surveys. types of. 359
'or number of sirdla . 134 Erron; of measurement. 377
in determining optimum probabilities of effects of constant bialo. 379
selecting primary UOlts . 3 11- 318 effect> of correlation between errors. 383
in determining optimum sampling fraction effect, of errors uncOrTelsted within sample.
for nonrespondents. 371 380 •
in determining optimum size of cluster unit. mathematical model . 377- 379
244 summary of effect" 384
in determining optimum subsampling techniqu~ for studying. 384-392
fraction . 280. 313-316 u 'e of interpenetrating subsamples , 389. 39[;
in determining sample size. 84 Estimates of population variaJlces: •
in double sampling: effects of errors in Sh 00 tncision of sb'atifi
for analytical comparisons. 335 sampling. 115
for regression estimates. 341 for detennining sample size. 78
with stratification . 331 Expansion faclor. 21
in stratified random sampli ng, 96 Eye estimates. 189
in two-frame sampli ng. 145
variance of sample means. 2~ Finite population correction (fpcl, 24
'Jmulative f rule. 129 rule for ignoring. 25
in stratified random sampling. 92
llegrees of freedom. effective number in stratified in two-stage sampling , 277
sampli ng . 96 Frdme.6
Descriptive surveys, 4 estimates when frame has units nOl in target
sign effect (Deffl. 21 population . 37-38
Domains of study (subpopulations l • 34 sampling from two frames. 144-146
, mpari~ons between nleans. 39
comparisons between propnnions nnd total s. 64 Geographic strJ!ification. 102
domains CUlling acros~ strata , 142- '44 construction of strata, 13 I
eslimate~ of mean~ and totals (continuous gains In preci ion. 102- 103
variate l • 34-35
eslimates of proportions and tOlJlls (0-1 variate). Hard core in n-onresponse. 364
63 Horvitz-Thompson estimator.",,59-261
ample size needed. 82-83 Hypergeometric distribution. ~5
tl'lua as domain. 140 a conditional distribution. 61
424 SUBJECT INDEX

confidence limits . 57 Politz and Simmons' method, 374-377

reference to charts and tables, 57 reason... for. 364
Normal distribution , 8
Index of inconsistency, 387 as approximation to hypergeometric. 57
Inflation factor, 21 as limiting distribution of sampl e mean . 39
Interpenetrating subsamples, 388-392 use in surveys, 11
Interviewer bias, 378, 384 validity with means from continuous dala .
mathematical models, 371-379 39-44
Intracluster correlation, 209 Notation:
Inverse sample (Haldane's method). 77 for errors of measurement, 377-379
Item, definition. 20 for proponions, SO, 63
for rotio estImateS, 151
Jacknife method . 174 for simple random sampl ing . 20
for estimating variance of a ratio, 178 for stratified sampling, 90
for variance of nonlinear function. 321 for two-stage sampling , 276
for variances of estimates, 27
Keyfitz , short-cut method of variance estimation.
169-171
Optimum allocation in stratiffed ~umpling:
Knight' s move latin square, 229
comparison with proponional allocation, 99,
Labels, 44 109
Lattice sampling. 228 comparison with simple random sampling.. 99.
Linear regression estimator. see Regression 109
estimator wi th double sampling. 330
Listing of primary units , effect of cost on effect of deviations from the optimum. 115
optimum selection probability, 316 effect of errors:
Lo 5 function, 83. 121 in Sh. 115-117
in stratum size,. 11 7
Mail surveys, 145, 360, 37U estimati on from previous data. 102. 131
Malching in repeated ampling of same for fixed sample ~ize . 9 99
p pulalion. 364-365 for fixed total cost. 96-98
Mean of sa mple. optimum prupenies, 44-45 with more than (Jne item. I 19- 123
Mean square error. definition, 15 with rat io estImate .... 172
justification for use of. 15 requiring more thun IOOt;f sampling.. 104
relation to variance and bias. 15 in &ampling for proponions . 108
Measure of homogeneity. 248 Optimum allocution in stratified two-,tage
Model·unbiased . 158 sa mpling. 280
Multivariate ralio e.~tima\e, 185 Optimum per cent mutched. in repeated ,umpling. .
. 347.3SO
Neyman allocation , 99 Optimum ,he of subsample:
best Slrdtum boundaries for. 127 primary unns of equal size . 2110
Noncoverage , 364 primary Unit, of unequal size. 313-3 16
Nonnormality:
effect of stratifi cation on. 44 Percentages. estimation of. su Proponions.
effect on confidence limits . 41 estimation of
frequently encountered in sampling practice. 40 Periodic variation, effect on systemutic sampling.
Nonrespollse. 359 2 17
bias produced by. 359-36 1 Pilot louevey:
effect of call·backs. 365-370 use in elotimuting optimum ,amphn)! and
effect on confidence limits, 361-363 ~ur.su mpling If'~ct"",, . 283-285
effect on variance in stratified sa mpling. 144 Uloe in est;mating sample loize. 78
optimum sa mpJi~ fraclion among Politz and ~immons ' method for hundling
nonrespondents, 370-374 nonresponloc. 374-377
 SUBJECT INDEX 425

'Opulation, 5 Random numbers, 19

sampled,S Randomized rc 'ponse method, 392
target, 5 other variants, 395
oststratification, l34 unrelated second question. 393
compared with proportional allcx:allon, 134 Warner', original method. 392
'recision contrasted wIth accuracy , 16 Random sampling, see Simple random <;ampling
r(lative, 99 Rare items, Haldane' s method. 77
'retest , 7 Ratio estimator, 30, ISO
'rimary sampling units (primary unit~l , 274 accuracy of approximate \'ariance, 162
'robabiliry proportional LO size, pps, (one-stage adjustments to decrease bias, 174
sampling). 250 bias , 160-162
clmpared with equal probabilities. 25 258 in cluster sampling, 31, 66
method of drawing sample. 250 compared with mean per unit, 157
S!lcction without replacement, 258 compared WIth regre,sion estimate. 195
'\)1 {two-stage sampling', 295, 308 compared with sttatified sampling, 169
ompared with equal probabilities, 299. 3 10 condition, under which unbiased, 158
election without replacement, 308 confidence limit~. 156
'bability ampling. definition and properties. 9 consistency, 153
Itematives to . 10 in double sampling, 343
oouct estimator, 186 estimated variance. 32. 155, 178
ional allocation in stratified sampling, 9 I Hanl ey- R os~ estim..te, 174
dvice on use. 103. 110 jackknife estimate of variance. 178
omparison with optimum allocation, 99, 103. Lahiri ' ~ method. 175
109 multivariate, 185
mpari son with p05tstmtification , 134 optimum allocation in stratified sampli ng. 172
mpanson with simple random sampling. 99, optimum conditions for, 160
109 as spl!ciai case of regression estimate. 190
I timating proponions. 109 standard errOr for comparison of two ratios, 180
s lple self-weighting, 91 in stratified random sampling. 164
ance formulas, 92- 95 in stratified two"stage ~a mpling . 317
Pnpnions, estimation of, 50 unbiased ratio-type estimates, 174
lc1ustcr ~ampling. 64-68, 246 upper bound to relative bias, 162
double sampling , 330 variance in large samples , 30, 153
feet of nonresponse, 361 Set! also Combined ra,io estimate; Separate ratio
'ect of population P on precision . 53 estimate
th more than two classes, 60. 61 Ratio of two ratio estimators. 183
simple mndom sampling. 50-68 Record checks, 385
e of sample for, 74 Regression coofticient in finite populations, 191
stl"oItified random ampling, 107- 111 Regre~sion estimator, 189
wo-stage sampling (units of equal ize l .279 accuracy of-large-sample varian e, 197
two-stage ampling {units of unequal silel. bias. 198
31 I compared with mean per unit. 195
Ive selection, 10 compared with ratio estimate, 195
conditions under which unbiru ed, 199
rJUc confidence limits for I"dtiO e, timate. 156 in double sampling, 38
'tative characteristics, 50, See also effect of error in slope. 192
Proponions, estimation of estimated variance, 195
u sampling, 135 with preassigned slope. 190
in repeated sampling of same population
ing factor. 21 346-355
om member of household, method of in stratified random sampling, 201, 202
selecting, 365 uses, 189
428 SUBJECT INDEX
Unit (sampling unit l •
definition . 6 in opinion po. " 4
United States Census. use of sampling. 2-3
Uses of sample surveys. 2-4 VIriaDoe, definitioa 0( SI and CT I , 22. 25
by business and industry. 3 Variance function. 243
in decennial censuses. 2-3 Variance 0( population. advance estimates. 78-8 I
in market research. 3 Variance of sample ertimalCs. su Standard error
by members of United Nations. 2
 L'BRARy1.
CENTRAL INSTITUTE "(
OF FISHERIES EOUCAllON BOMBA
( ICAR) .

C NTRAL INST UTI Of PI, liS

. ATiON
VERSOVA ANOHeR. BOMBAY-400 081.
{I CAR)
Boob may be retained from 8 pertod not
excHding 16 days.
Book. may b. r.newed on request at the
discretion of the Ubrarian.
3. Dog-earing the pages of a book. marklno
or writing ther.in with ink or pencil. tearing I

I..
or taking out ita pagel or oturwlse demag-
Ing it. will constitute an 'injury to I book.
Any luch injury to a book .eriou.
oftenee; Un.... II borrower pointS out the
Injury et the ti .f borrowing the beok.
he thall be required to replace thl book or
pav ita price,
I, to ICMp thf.