DESIGN OP EXPERIMENTS

A Realistic Approach
 STATISTICS

Textbooks and Monographs

A SERIES EDITED BY

D. B. O W E N , Coordinating Editor
Department of Statistics
Southern Methodist University
Dallas , Texas

PETER LEWIS P A U L D. M I N T O N
Naval Postgraduate School Virginia Commonwealth University
Monterey ,California Richmond ,Virginia

J O H N W. P R A T T
Harvard University
Boston, Massachusetts

V o l u m e 1: The Generalized Jackknife Statistic

H. L. Gray and W. R. Schucany
Vo lu m e 2: Multivariate Analysis
Anant M. Kshirsagar
Vo lu m e 3: Statistics and Society
Walter T. Federer
Vo lu m e 4 : Multivariate Analysis:
A Selected and Abstracted Bibliography, 1957 -1972
Kocherlakota Subrahmaniam and Kathleen Subrahmaniam
(out of print)

V o l u m e 5: Design of Experiments: A Realistic Approach

Virgil L. Anderson and Robert A. McLean

O t h e r Vo l u m e s in P r e p a r a t io n
DESIGN OF EXPERIMENTS
A Realistic Approach

IT * __*1 V A 1 _ 1 n L I I II V
^
Department of Statistics Department of Statistics
Purdue University University of Tennessee
Lafayette, Indiana Knoxville, Tennessee

MARCEL

WL M a r c e l D e k k e r, In c. N e w Y o rk • B a s e l • H o n g K o n g
COPYRIGHT © 1974 by MARCEL DEKKER, INC. ALL RIGHTS RESERVED.

Neither this book nor any part may be reproduced or transmitted

in any form or by any means, electronic or mechanical, including
photocopying, microfilming, and recording, or by any information
storage and retrieval system, without permission in writing from
the publisher.

MARCEL DEKKER, INC.

270 Madison Avenue, New York, New York 10016

LIBRARY OF CONGRESS CATALOG CARD NUMBER: 73-90767

ISBN 978-0-8247-6131-8

Current printing (last digit):

20 19 18 17 16

PRINTED IN THE UNITED STATES OF AMERICA

To Avis and Marjorie
For when fear is overcome, curiosity and constructiveness are free,
and man passes by natural impulse towards the understanding and
embellishment of life.

Will Durant: Our Oriental Heritage

 CONTENTS

PROBLEMS xi

PREFACE xiii

1. REVIEW OF SOME BASIC STATISTICAL CONCEPTS 1

1.1 Testing Hypotheses and Sample Size 1
1.2 One Way Classification, Assumptions Met 6
1.2.1 ANOVA Notation and Rationale 12
1.3 Unequal Variances and Transformations inANOVA 16
1.3.1 Bartlett Test (Equal Subclass Numbers) 20
1.3.2 Bartlett and Kendall log s2 ANOVA 21
1.3.3 Burr-Foster Q-Test of Homogeneity 22
1.3.4 Transformation of y 23
1.3.5 Other Transformations 24
1.3.6 Test for Normality 25
1.4 Curve Fitting in One Way Classification 27
1.4.1 Orthogonal Polynomials 28
1.4.2 Lack of Fit Principle 34
1.5 References 38

2. SOME INTERMEDIATE DATA ANALYSIS CONCEPTS 40

2.1 Two Factor Experiments with One Observation Per Cell 41
2.1.1 Both Factors Qualitative 45
2.1.2 One Factor Qualitative and the Other Quantitative 47
2.1.3 Both Factors Quantitative 48
2.1.4 Summary of Section 2.1 51
2.2 Two Factor Experiments with More Than One
Observation Per Cell 51
2.2.1 Expected Mean Square Algorithm 52
Example. 2.1: Fixed Model 53
Example. 2.2: Random Model 56
Example. 2.3: Mixed Model 56
Example. 2.4: Fixed Model 59
Example. 2.5: Fixed Model (More than one
observation per cell) 59
Example. 2.6: Random Model 60
Example. 2.7: Mixed Model (M fixed, T random) 60
2.2.2 ANOVA of Two Way Classification with More Than
One Observation Per Cell 62
Example. 2.8: Both Qualitative Factors 66

v
 CONTENTS

Example 2.9: One Factor Qualitative and

One Quantitative 71
Example 2.10: Both Factors Quantitative 73
2.3 References 78

A SCIENTIFIC APPROACH TO EXPERIMENTATION 79

3.1 Recognition That a Problem Exists 81
3.2 Formulation of the Problem 82
3.3 Agreeing on Factors and Levels to Be Used
in the Experiment 83
3.4 Specifying the Variables to Be Measured 84
3.5 Definition of the Inference Space for the Problem 84
3.6 Random Selection of the Experimental Units 85
3.7 Assignment of Treatments to the Experimental Units 86
3.8 Outline of the Analysis Corresponding to the
Design before the Data Are Taken 87
3.9 Collection of the Data 89
3.10 Analysis of the Data 90
3.11 Conclusions 90
3.12 Implementation 91
3.13 Summary 91
3.14 References 93

COMPLETELY RANDOMIZED DESIGN (CRD) 94

4.1 One Factor 94
4. 2 More Than One Factor 98
4.2.1 All Fixed Factors 98
Example. 4.1: Engineering 105
4.2.2 Some Factors Fixed and Some Random 111
Example. 4.2: Engineering 113
4.2.2 All Random Factors 116
4.2.4 Spacing Factor Levels 119
4.3 References 121

RANDOMIZED COMPLETE BLOCK DESIGN (RCBD) 123

5.1 Blocks Fixed 125
5.1.1 A Medical-Engineering Example 131
5.1.2 RCBD More Than One Observation Per Cell (Example) 132
Example. 5.1: Medical 135
5.1.3 RCBD When Interaction Is Present and One
Observation Per Cell 137
5.2 Blocks Random 138
Example 5.2: Genetics 142
5.3 Allocation of Experimental Effort in RCBD 143
5.4 Relative Efficiency of Designs 151
5.5 References 153
CONTENTS vii

6. NESTED (HIERARCHICAL) AND NESTEDFACTORIAL DESIGNS 154

6.1 Nested (Hierarchical) 154
6.1.1 All Factors Random 154
Example. 6.1: Economics 156
Example. 6.2: Genetics 159
6.1.2 Fixed and Random Factors 160
6.2 Nested Factorial 162
6.2.1 Cardiac Valve Experiment 163
Example. 6.3: Medical 167
6.2.2 Social Science Application 171
6.2.3 Nutrition Application 172
6.2.4 An Application in Ammunition Manufacturing 175
6.3 References 180

7. SPLIT PLOT TYPE DESIGNS 181

7.1 Split Plot Designs 181
Example. 7.1: Cardiac Valve 189
Example. 7.2: Metallurgy 193
7.2 Split-Split Plot Designs 199
7.3 Regression Analyses fromSplit Plot Data 206
7.4 References 208

8. LATIN SQUARE TYPE DESIGNS 210

8.1 Latin Square 210
Example. 8.1: Engineering 216
8.2 Associated Designs 218
Example 8.2: Pharmacy 219
8.3 References 223

9. 2n FACTORIAL EXPERIMENTS (COMPLETE AND INCOMPLETE BLOCKS) 225

9.1 Complete Blocks 226
9.1.1 The 22 Factorial 227
9.1.2 The 23 Factorial 229
9.1.3 The 2n Factorial 231
Example 9.1: Metallurgy 232
Example 9.2: Metallurgy 237
9.2 Incomplete Block 241
9.2.1 The 22 Factorial 241
9.2.2 The 23 Factorial 243
9.2.3 General Approach 245
Example 9.3: Animal Science 248
9.3 References 251
viii CONTENTS

10. FRACTIONAL FACTORIAL EXPERIMENTS FOR TWO-LEVELED FACTORS 252

10.1 The 1/2 Replication 253
Example 10.1: Highway Engineering 256
10.2 A More Complicated Design (1/4 Replication) 260
Example 10.2: General 261
10.3 Blocking within the Fractional Replication 268
10.4 Some Other Versions and Developments in
Fractional Factorials 270
10.4.1 Parallel Fractional Replicates 270
10.4.2 The 2n”P Fractional Designs 272
10.4.3 Symmetrical and Asymmetrical Fractional
Factorial Plans 272
10.4.4 The 3/4 Fractional Factorial 273
10.4.5 Case When Some Three Factor Interactions Cannot
Be Assumed Negligible 275
10.5 General 2n Problems 277
10.6 References 278

THREE-LEVEL FACTORIAL EXPERIMENTS 280

11.1 Confounding in a 3n System 281
11.1.1 The 32 Factorial 282
11.1.2 Other 3-Leveled Factorial Experiments 284
11.1.3 Method of Confounding 285
11.2 System of Confounding 289
2
11.2.1 A 3 Factorial Experiment in Blocks of 3 289
3
11.2.2 Two 3 Factorial Experiments 291
4
11.2.3 Two 3 Factorial Experiments 291
11.3 Fractional Replications 293
11.3.1 A 1/3 Replicate of 35 Factorial 293
11.3.2 A General Approach 294
11.4 A Special Use of Fractional Factorials 295
11.5 Special Latin Square Designs as Fractional
Factorials 298
11.6 Extension of 3n to pn, Where p Is a Prime Number 300
11.7 References 301

MIXED FACTORIAL EXPERIMENTS AND OTHER INCOMPLETE

BLOCK DESIGNS 302
12.1 Designs for Factorial Experiments with the
Number of Levels the Same Prime Power 302
12.1.1 One Factor (To Demonstrate Pseudofactor) 303
12.1.2 Two Factors 304
12.2 Designs in Which the Number of Levels Are
Different Prime Numbers 307
12.2.1 Design with 9 Blocks of 4 for the 2 x 2 x 3 x 3 309
12.2.2 Design with 6 Blocks of 6 for the 2 x 2 x 3 x 3 309
CONTENTS ix

12.2.3 Design with 4 Blocks of 9 for the 2 x 2 x 3 x 3 309

12.2.4 Design with 3 Blocks of 12 for the 2 x 2 x 3 x 3 309
12.2.5 Design with 2 Blocks of 18 for the 2 x 2 x 3 x 3 310
12.2.6 A Partially Balanced Incomplete Block Design
(Four Replications of 6 Blocks of 6) 310
12.3 Designs That Have the Number of Levelsas
Products of Prime Numbers 310
Example. 12.1: Industrial Engineering 311
12.4 Designs in Which the Pseudofactors Have
Different Powers on the Prime Number of Levels 316
12.5 Designs in Which the Number of Levels Are
Products of Powers of Prime Numbers 317
12.6 Fractional Mixed Factorials 2m3n 317
12.7 Lattice Designs (Pseudofactorial) 319
12.7.1 One-Restrictional Designs 320
12.7.2 Two-Restrictional Designs 323
12.7.3 Designs with More Than Two Restrictions 324
12.8 Additional Incomplete Block Designs 324
12.9 References 325

13. RESPONSE SURFACE EXPLORATION 326

13.1 Fixed Designs 327
13.1.1 Random Balance 327
13.1.2 Systematic Supersaturated 330
13.1.3 Random Method 333
13.1.4 Mixture Designs 335
Example. 13.1: Chemical Flare 343
13.1.5 Rotatable Designs 348
13.1.6 Fractional Factorials (Orthogonal) 350
13.1.7 Complete Factorials (Orthogonal) 352
13.1.8 Composite Design 353
13.2 Sequential Designs 361
13.2.1 Univariate Method or One-Factor-at-A-Time 361
13.2.2 Simplex (Sequential) 362
13.2.3 Computer Aided Designs 367
13.2.4 Steepest Ascent Method 370
13.2.5 Canonical Analysis 374
13.2.6 Evolutionary Operation (EVOP) 379
Example, 13.2: Steel Manufacturing 380
13.3 References 386

APPENDICES 388

AUTHOR INDEX 413

SUBJECT INDEX 415

 PROBLEMS

Problem
Number Applied Area Page
1.1.1 Civil Engineering 6
1.3.1 Pharmacy 21
1.3.2 Pharmacy 24
1.3.3 Pharmacy 27
1.4.1 Engineering 37
1.4.2 General 37
1.4.3 General 37

2.1.1 Aircraft (Engineering) 51

2.2.1 Steel (Engineering) 71
2.2.2 Steel (Engineering) 71
2.2.3 Steel (Engineering) 73
2.2.4 Animal Nutrition 77

4.1.1 Engineering 97
4.2.1 Engineering 98
4.2.2 Animal Nutrition 110
4.2.3 General 110
4.2.4 Agronomy 110
4.2.5 General 111
4.2.6 General 111
4.2.7 Engineering 116
4.2.8 Engineering 116
4.2.9 Engineering 119
4.2.10 Genetics 120

5.4.1 Photo (Engineering) 152

5.4.2 Agronomy 153

6.1.1 Genetics 160

6.1.2 Genetics 160
6.1.3 Aluminum (Engineering) 161
6.1.4 Agronomy 162
6.2.1 Medical 170
6.2.2 Poultry Nutrition 170
6.2.3 Engineering 174
6.2.4 Psychology 179
6.2.5 Advertising 179

7.1.1 Agronomy and Animal Nutrition 198

xi
7.1.2 Medical 198
7.2.1 Alloy (Engineering) 202
7.2.2 Battery (Engineering) 203
7.2.3 Steel (Engineering) 204
7.3.1 Highway (Engineering) 206

8.2.1 Plastics (Engineering) 222

8.2.2 Pharmacy 222

9.1.1 General 229

9.1.2 Metallurgy 232
9.1.3 Metallurgy 239
9.1.4 Steel (Engineering) 239
9.2.1 General 242
9.2.2 Animal Nutrition 250
9.2.3 Engineering 250

10.1.1 Highway (Engineering) 258

10.1.2 Highway (Engineering) 259
10.2.1 Metallurgy 261
10.2.2 General 266
10.2.3 General 267
10.3.1 Plastic (Engineering) 270
10.5.1 General 277
10.5.2 Engineering 277
10.5.3 General 277
10.5.4 General 277
10.5.5 General 278

11.1.1 General 285

11.2.1 General 290
11.2.2 General 291
11.2.3 Sociology 292
11.2.4 Sociology 293
11.6.1 Engineering 300

12.3.1 Industrial Engineering 316

12.3.2 Industrial Engineering 316
12.3.3 Industrial Engineering 316
12.3.4 Cardiac (Farm) 316
12.5.1 General 317
12.5.2 Plastics (Engineering) 317

13.1.1 General 339

13.1.2 Alloy (Engineering) 347
13.1.3 General 360
13.1.4 Engineering 360
13.1.5 Engineering 360
13.2.1 Electrical Engineering 367
 PREFACE

In this book we have tried to express rather complex ideas on

how and why scientific investigators should design experiments. It
is assumed that the reader knows only high school algebra and basic
2
statistical concepts such as normal, t, x > and F distributions;
sampling distributions of the mean and variance; and has been
exposed to simple linear regression and one way analysis of variance.
Many of the ideas have been tried out on students at Purdue
University, the University of Tennessee and in numerous industrial
type classrooms.

The authors have both taught these concepts at this

comparatively low mathematical level to an extremely diversified
audience. The students have ranged in mathematical competence from
Ph.D. candidates in mathematics to students who had only high
school algebra. The experimental backgrounds of the students have
ranged from undergraduates with no experience to a professor with
a Ph.D. (who took the course for credit). The latter had many years
of experience but no formal training in design of experiments. In
all cases, if the student had no fear of statistics, wanted to learn
and possessed a certain amount of mathematical maturity, the varied
backgrounds did not detract from the success of the course. In fact,
many times this diversity made students recognize similarities among
problems from different fields of applications. This gave them a
broader scientific approach to experimental problems than a similar
course taught only with familiar examples. This book is meant to
be read by people from any field who work, or will possibly work,
on problems dealing with scientific experiments. Some of the areas
from which students have come are Agriculture (all departments),
Biology, Bionucleonics, Chemistry, Computer Sciences, Education,

xiii
xiv PREFACE

Engineering (most of the schools), Genetics, Home Economics,

Industrial Management, Mathematics, Medicine, Pharmacy, Social
Sciences, Statistics and Veterinary Science. The varied background
of the students makes it impossible to present examples for
everyone. The mature students, however, appear to have little
difficulty in visualizing similar examples in their own fields.

This book is unusual inregard to the emphasis on the

usefulness of the results ofthe experiment. We want the reader to
appreciate the inferences hecan make once he has the results. This
means he must understand how the samples were taken, or in design of
experiments terminology he must know what restrictions on
randomization he has imposed on his observations and how these
restrictions affect his inferences. The usual concepts of blocking
and efficiency of designs are given, but we wish to emphasize the
real dangers the experimenter nans when blocking and sometimes how
little information can be obtained when he inadvertently or
purposely makes restrictions on randomization. We use a restriction
error in a pedagogical sense to make the reader grasp certain
deficiencies that happen so often in experimentation. It is not
uncommon for a so called "expert" in design of experiments to ruin
an experiment because the emphasis in so many design of experiments
books has been the straight application of the designs set up for
agricultural experiments. These people do not understand the
assumptions made in this particular subject matter area and apply
the techniques without careful thought. One of our main purposes,
then, in this book is to try to give a broad, general base for
designing experiments in any subject matter area and have the
reader thoroughly understand what and why he is doing certain
things at the design stage.

The book is arranged so that the reader may go from the simple
to the complex designs and grasp the appropriate analyses from the
resulting data. In Chapter 1 we review analyses for one way
classified data and in Chapter 2 we introduce two way analyses.
Random sampling concepts are used in these two chapters, assuming
PREFACE xv

the design of each experiment is completely randomized although the

design is not fully defined until Chapter 4. The method of review
that is utilized is to present some non-standard, as well as
standard, material. The most controversial subject that is
presented in this review is that of the use of multiple comparison
tests. The reason for this controversy is that there is no one
correct answer as to whether one should always strive to have small
Type I or Type II errors for either the individual comparisons
among treatment means or for the entire experiment. Obviously, the
selection of the most appropriate multiple comparison test is
dependent upon the individual situation and usually the experimenter
does not have sufficient information available to make a good
selection. Consequently, since we have experienced and observed
larger economic losses whenever Type I errors are made, a modified
version of the Student-Newman-Keuls test, which has excellent
control over the Type I error rate, is used throughout this text.
The modification arises by requiring a significant F test in the
analysis before the multiple comparison test is utilized. This
appears to be a very realistic approach when one is involved with
experiments containing several factors and associated interactions.

Chapter 3 is the real beginning of this design book. In this

chapter we wish to show the reader where design of experiments fits
in the scientific method and how most experiments come about.

Chapters 4 through 7 cover our concepts of design, the thorough

usage of the restriction error concept and the unusual analyses that
result. This is the body of the book, the basic designs that set
the stage for a most realistic look at the problems encountered by
an experimenter when he decides to design an experiment.

Chapter 8 expresses a view on Latin square type designs and

Chapters 9 and 10 attempt to provide a practitioner's approach to
the 2n completely randomized factorials. Chapters 11 and 12 have
useful designs in them but if time were a factor in using this book
for a quarter or a semester, it would do little harm to go to
xvi PREFACE

Chapter 13 directly from Chapter 10. We believe it would be bad

for a student not to be exposed to the designs related to response
surface exploration. Hence, we recommend that, even though Chapter
13 is not as thorough as Myers, R. H., Response Surface Metholology,
Allyn and Bacon, Inc. (1971), it should be covered in a design of
experiments course.

The authors chose to leave out the use of concomitant variables

(including covariance) in designing experiments in this book because
of the possible background of the readers and the experience of the
authors in teaching courses with and without this additional
information. It has been our experience that students do appreciate
the difference between concomitant information and blocking, or
additional terms in models expressing the effect of concomitantly
measured variables; however, to cover this material in the desired
depth to make it understandable would require more time in the
analysis phase (Chapters 1 and 2) than we believe is warranted for
this text. For those readers who would like to know something
about these concepts we recommend they read Chapter 4 of Planning
Experiments, Cox, D. R., Wiley and Sons, Inc., 1958.

We are indebted to Oliver and Boyd Publishers for allowing us

to reprint most of pages 609 and 610 and Section 11.4 from their
book edited by Davies, 0. L., entitled Design and Analysis of
Industrial Experiments (1963). These appear in Appendix 4 and
Section 13.2.4, respectively. We also wish to thank S. S. Shapiro
and M. B. Wilk for allowing us to use their tables in Appendices
9 and 10, and H. Leon Harter for allowing us to reproduce in
Appendix 6 part of the tables in Order Statistics and their Use in
Testing and Estimation, Volume 1: Tests Based on Range and
Studentized Range of Samples from a Normal Population (Gov't.
Printing Office, Washington, D. C. , 1970). We are indebted to the
Biometrika Trustees for giving us permission to use Tables 8, 12,
and 18 from Biometrika Tables for Statisticians, Volume I, third
edition (1966). These tables appear in Appendices 7, 3, and 5 in
this book.
PREFACE xvii

We gratefully acknowledge the help and criticism from students

and colleagues who gave freely of their time through the many
revisions of this book.

Most of all we wish to thank Mrs. Eddie Hicks of the Statistics

Department, Purdue University, for her patience through the endless
changes and expert typing of the final manuscript.

Robert A. McLean Virgil L. Anderson

Knoxville, Tennessee West Lafayette, Indiana
DESIGN OF EXPERIMENTS
A Realistic Approach
 Chapter 1

REVIEW OF SOME BASIC STATISTICAL CONCEPTS

Before the ideas of designing experiments can be fully

appreciated certain standard statistical techniques must be
presented. The ideas developed here are not intended to encompass
all of the elementary procedures, but merely to display enough
examples to introduce the terminology required to structure the
material on designing experiments.

Our methodology throughout the book will be to introduce and

explain ideas through examples and provide references for details.

In the first two chapters the design of the experiment is

assumed to be completely randomized. That is, the structure of the
experiment is assumed to be such that the treatments were placed on
the experimental units completely at random.

Details for constructing a completely randomized design are

given in Chapter 4. Suffice it to say here that for the first two
chapters the structure of the experimental arrangement of the
material used in the investigations is such that no restriction on
randomization of the treatments onto the experimental units is
assumed.

1.1 TESTING HYPOTHESES AND SAMPLE SIZE

In a Highway Research Organization one aspect of a most

complicated experiment had to do with comparing the effect of
percent cement and curing time on the strength of cement-stabilized
base material. The dependent variable, y, was pounds per square

1
2 1. BASIC STATISTICAL CONCEPTS

inch (psi) measured by an indirect tensile strength method. For

this part of the study, previous investigations showed that the
measured variable responded well enough for the usual t-test
assumptions (normal distribution and equal variances) to hold.
Hence the measured variable, psi, could be used as the variable to
be analyzed.

In order to get an indication of the number of observations

that would be necessary to show a significant increase in psi using
8% cement and 7 days of curing time over using 4% cement and 21 days
of curing time if there truly was an increase, the experimenter
decided to run 10 cylindrical specimens, each of which may be
considered as an experimental unit, measuring 4 inches in diameter
and 2 inches in height for each treatment. This would allow an
estimate of the variance from the 20 specimens and then one could
estimate the additional number of specimens needed to show
significance if it really existed. Stein (1945) shows a
theoretically sound procedure for handling this type of sequential
sampling; however, an approximation of this technique will be
utilized here.

For this particular experiment the investigator was willing to

run a 5% risk of saying there was a significant difference when
there really was not.

As pointed out in this study, most experimenters know quite a

bit about their product and reflect this knowledge by using a one-
tail test. In this case the investigator is only interested in
obtaining a less expensive product (4% cement) which will perform
the same task as the more expensive product (8% cement). Most
investigators ignore the fact that they have two factors, i.e.,
percent cement and curing time, confounded and merely compare
"Method 1" to "Method 2.” This is a satisfactory procedure in this
case, but, as we will see later, when one wants to write a
mathematical model which relates percent cement and curing time to
tensile strength one will not be able to use an experiment which
1.1 TESTING HYPOTHESES 3

confounds the two factors. ’’Confounds’1 merely means that the effects
of the two factors are not separable.

Assume that the experiment was carried out appropriately at the

mixing and curing stages and the specimens were obtained by a
completely randomized procedure. (By completely randomized, we mean
that a table of random numbers, Appendix 1, or some procedure, was
used to decode the order of obtaining the 20 specimens.) The data
obtained are given in Table 1.1.

TABLE 1.1

4% cement 8% cement
and and
21 days curing time 7 days curing time
(psi) (psi)

70 89
81 79
82 85
68 76
74 74
76 81
65 70
74 79
60 65
70 72
10 10
I y ,• = 720 = 770
i = i 11
yx = 72 = 77
^2
10 - 9 10 9
I (yu - y p - 422 = 460
l& n - y2J
i=l 11 1

The inference from this experiment is that the results apply to

stabilized base material in which all other factors are held constant.
Such factors are water at 3%, wrapping specimen during curing,
4 1. BASIC STATISTICAL CONCEPTS

limestone soil, curing temperature at 40°F, high compactive effort,

and gyratory shear compaction.

The analysis of the data follows:

Hq: = u2 where = population mean psi
of 4% cement at 21 days
curing

**1: ^1 < ^2 ^0ne w2 = P0?11^ ^ 011 mean psi

of 8% cement at 7 days
curing

In this case the experimenter is only interested in finding out

whether or not there is a significant decrease in the mean psi when
he uses the 4% cement cured for 21 days as compared to the 8% cement
cured for 7 days. Under the assumption that the psi observations are
normally distributed, the test statistic is

_ yi ' X2 _ 72 - 77 I 5
tdf = S - - -,v 3.13 =
yl ' y 2 /49(°-2)

where t ^ may be compared to t in Appendix 3 (df = n^ + n^ - 2 = 18),

y^ and y n are sample means estimating and respectively,

10 9 10
I (y li - yP + ^
i=l 1=1
n^ + n^ - 2

Since n^ = n^ = 10 and a = 0.05 the critical value for t is

-1.734. The value a is the probability of making a Type I error,
i.e., rejecting the null hypothesis (H^) when in fact it should have
been accepted. The selection of the magnitude of a is done by the
experimenter and is possibly based on the economics of the problem
or in many cases merely selected at a value, say 0.05, which is
consistent with previous studies of this type. Note that
i.l TESTING HYPOTHESES 5

10 9 10 9
I Cyu - vx) + I (y2i -
i=l 11 1 i=l _ 882 _
n, + n - 2 18 ” 4y

is a pooled estimate of the within-treatment variability which will

2
be denoted as Sp00^ecj = 49. This calculation is the one that requires
the assumption of equal variances. The method of testing for
homogeneous variances is reviewed in Section 1.3. This pooled
estimate of the within variance is the best estimate of the true
2
variance, a , and will be used to determine the approximate sample
size required.

Based on the numerical results above some practitioners would

say the inference must be made that there is no difference in using
8% cement for 7 days curing and 4% cement for 21 days curing. It is
true that the means are not statistically significant at the 0.05
probability level, but what are the consequences of accepting H^ in
this case?

An error of accepting Hq when in fact it should have been

rejected is referred to as a Type II error. The probability of
making an error of this kind and consequently accepting a false
hypothesis is referred to as 8. Naturally one would like to have 8
as small as possible because a decision error of this type would
result in an economic loss.

The next step in this experimental procedure is to decide on the

number of additional specimens to take to reduce 8 to a reasonable
size, say, 0.10. This decision on the size of 8 and a, for that
matter, must be made before the experiment is run or the sequential
procedure recommended here or by Stein (1945) must be followed. In
either case the Type I and II errors must be understood by the
experimenter and be decided upon.

If for example, there was an actual difference of 5 psi between

the two populations, i.e., 6 = ~ ~ ^ anc* us^-n8 sp00ie<i = 7 as
6 1. BASIC STATISTICAL CONCEPTS

the a in Appendix 4, 3 can be estimated. In this case D = 6 /a = 0.71

and for a one-tailed test with a = 0.05 and n^ = = 10 we find
that 3 > 0.5. This is extremely bad from a practitioner’s view
because he wants to make the right decision as frequently as he can.
Note that we have utilized the information obtained in our
experiment as if these were true values, rather than just estimates
of these values. One will resort to this type of reasoning quite
frequently and hence conduct the experiment in a sequential fashion.
The practitioner must recognize that the 3 values in Appendix 4 are
approximate, as a result of estimating a, but they do provide
excellent guidelines for obtaining sample sizes.

Continuing our example, let us assume that a = 0.05 (one tail),

6 = 5, a =7 and 3 = 0.10, what should the sample size be for each
treatment? Using D = 6 /a = 5/7 = 0.7 in Appendix 4 we find n^ = ^
= n = 36. Hence we should take 26 more specimens of each (percent
of cement, curing time) treatment combination to force 3 to be small
and allow a more clear-cut decision.

Problem 1.1.1. If the economics of the problem caused the

experimenter to let 6 * 7 and he could take only 13 more
observations per treatment after the first 10, what combination
of a and 3 would you recommend? Explain your reasoning.

1.2 ONE WAY CLASSIFICATION, ASSUMPTIONS MET

In Section 1.1 general ideas on testing hypotheses and

determining sample size in experimentation were discussed. The
example in that section had only two treatments and utilized the
t-test to bring out these general ideas.

In this section the generalization is to set up a testing

procedure to handle any number of treatments, namely, the analysis
of variance (ANOVA). The example will have only four treatments
but it should be understood that the technique is easily extended
to any number of treatments. To allow the reader to review the
1.2 ONE WAY CLASSIFICATION 7

ANOVA method without other disturbances, we will say that all

assumptions necessary to make the F-test have been met. In the
following sections discussions will cover cases in which some
assumptions have not been met.

A research worker in pharmacy was interested in finding out

whether or not there were differences in disintegration times of a
narrow group of four types of tablets; A, B, C, and D. He was
interested in only these four types which make the factor fixed.
Hence the mean time (in seconds) for disintegration should be
analyzed. If he had been interested in the population of all tablets
from which these four were drawn and he had drawn these four at
random from the possible types, he could have called the factor
random and analyzed the variance of the means of disintegration time.
The inference space, that space within which the investigator may
apply the results of his experiment, is much broader and has wider
implication for the random model than for the fixed model. This is
a general concept for any number of factors in a large experiment.
Two references on fixed and random factors are Eisenhart (1947) and
Dixon and Massey (1969, Chapter 10).

This experiment on tablet manufacturing was set up so that the

manufacturing and testing procedures were carried out in a completely
randomized manner. In addition, previous studies of tablets of these
general types indicated that if there were differences among types,
at a = 0.05 level, that the number of tablets per type had to be at
least six in order for the experiment to detect this difference.
References on the subject of determining sample size in ANOVA
problems include Kastenbaum, Hoel and Bowman (1970a), (1970b) and
Bratcher, Moran and Zimmer (1970).

This experiment had six tablets selected at random from each

type and the disintegration time (in seconds) was measured. It is
assumed that the response variable in this case is normally and
independently distributed, that the terms in the model are linear
and that the variance among the tablets within types are all equal.
8 1. BASIC STATISTICAL CONCEPTS

The model used as the basis for the analysis is

y. . = v + T. + e,... X.= 2' "" t = 5 (1.2.1)

'il i (i)l 1= 1. 2, . . n = 6
where
th
y.. = the disintegration time of the j tablet in
1*J .th
l type

y = overall mean

T^ = the effect of the i**1 type of tablet (fixed)

th th
= the random error of the j tablet in the i
2
type, NID (0, a ) or normally and independently
2
distributed with zero mean and variance o .

The data are presented in Table 1.2.1.

TABLE 1.2.1

Disintegration Times (seconds) for Types of Tablets

Type of Tablets (i) Observations (j) Totals Means

A 6, 2, 5, 4, 6, 7 30 5.0
B 10, 8, 11, 7, 7, 9 52 8.7
C 3, 7, 6, 4, 8, 6 34 5.7
D 10, 4, 6, 6, 7, 8 41 6.8

Sums of squares (SS):

Given that E l y. . = 157

i] J
and I E y?. = 1141
• -13
i 3 J
(a) Correction term to allow for deviation around mean (CT)
1.2. ONE WAY CLASSIFICATION 9

(c) SS total = I I y 2.. - CT = 1141.0 - 1027.0 = 114.0

i j
(d) SS Within Types (error) = SS Total - SS Types
= 114.0 - 46.5 = 67.5

From these results we can set up the ANOVA in Table 1.2.2. An

abbreviated form of notation for the above material and the intuitive
reasoning which supports the F-test shown in Table 1.2.2 is given in
Section 1.2.1. The student who is truly interested in understanding
the philosophy behind the ANOVA should study this section in detail.

TABLE 1.2.2

ANOVA for Tablet Manufacturing Data in Table 1.2.1a

(0.05)
Source df SS MS F
crit

Types of tablets 3 46.5 15.5 4.6 3.1

Within types
(error) 20 67.5 3.4

Total 23 114.0

adf = degrees of freedom; SS = sums of squares,

deviations around means; MS = SS/df = mean square; F = MS
types/MS error calculated from data; = the
critical value of F for a = 0.05 and df are 3 and 20 found'
in Appendix 5.

Conclusions:

Since the calculated F from the data is greater than the

critical value of F at the 0.05 level, the experimenter concludes
that there is a significant difference among the mean disintegration
times of the four types of tablets. This does not allow him to say
anything about which type has the shorter or longer disintegration
time, however. Since there is an overall significant effect of types,
the investigator would ordinarily examine the individual means to
10 1. BASIC STATISTICAL CONCEPTS

find this out. One test that allows investigation of all possible
pairs of means in a sequential manner, has very good power (1-3)
and keeps the a level constant for investigation of all pairs of
means is the Newman-Keuls test. The two basic references for this
test (See the preface for a philosophical discussion of multiple
comparison tests) are Newman (1939) and Keuls (1952). In addition
to the Newman-Keuls test the researcher is apt to encounter other
multiple comparison procedures such as: the least significant
difference, Scheffe's method, Tukeyfs honestly significant difference,
and Duncan's multiple range test. Two excellent references which
compare these various methods are Bancroft (1968, Chapter 8), and
Carmer and Swanson (1973).

To run the Newman-Keuls test, arrange the means in rank position

1 2 3 4
B D C A
8.7 6.8 5.7 5.0

Next prepare a table of differences of means from the largest

difference to the smallest in all possible pairs forming a triangular
arrangement as follows:

Number of Means Spanned

k k-1 k-2
1.2 ONE WAY CLASSIFICATION 11

where column headings indicate the number of means spanned for each
test on that diagonal.

Prepare a list of the least significant ranges as follows:

Rfc = <l0(k*df) s-; where qa(k,df) means the entry

in Appendix 6 for the desired a and k = 4,
number of means compared in the experiment,
and s- = the standard error of the mean

’ -y ' ■ / ¥ ■

where n = number of observations in that type.

*k_i - qa0c-Mf) s-

R2 = %(2,df) s-

and for our example using a = 0.05,

= R4= q0#()5 (4,20) s- = (3.96) (0.75) = 2.97

Rk-1 = R3= q0.05 (3,20) Sy = (3.58)(0.75) = 2.68

Rk-2 = R2= q0.05 (2'20) sy = (2.95)(0.75) = 2.21

Referring back tothe triangular table, we look at each diagonal

element and compare in order from k to 2

B-A = 3.7 vs 2.97 significant

B-C = 3.0 vs 2.68 significant
D-A = 1.8 vs 2.68 .'. not significant
B-D = 1.9 vs 2.21 not significant
D-C = 1.1 vs 2.21 /• not significant
C-A = 0.7 vs 2.21 not significant

Hence the final conclusions for comparing individual means are:

(a) Type B has a significantly longer disintegration time
than types A and C.
(b) All other types are not distinguishable.
12 1. BASIC STATISTICAL CONCEPTS

In general, when making individual comparisons of this type one

has one additional rule to follow. That is, one continues across the
first row until the first nonsignificant difference is encountered.
At this point one must stop and declare all remaining differences in
the first row nonsignificant.

In some instances it will be possible to continue across the row

and find another value which is greater than its respective least
significant range. This value would still be declared not significant
in order to avoid any logical contradictions. In addition, all
differences below any nonsignificant difference are declared
nonsignificant. Next, one repeats this procedure for row 2, and
continues until finished. Note that our example satisfies this
particular rule.

To complete this investigation the research worker may want the

longest disintegration type for some reason and in this case he should
choose B. Next he would almost undoubtedly like to know the 95%
confidence interval on the mean disintegration time. To obtain this
confidence interval he merely needs to set up

(y - ^CO-025) y < y < (y + ^(0*025) ,

lyB 20 df V B 20 df yJ

and use Appendix 3 to obtain

(0.025) _
20 df “ 2*086

Hence he has the interval

[8.7 - (2.086)(0.75) < y < 8.7 + (2.086)(0.75)]

or
(7.1 < y < 10.3)

as the 95% confidence interval on the mean disintegration time (in

seconds) of the type B tablet.

1.2.1 ANOVA Notation and Rationale

The type of notation that is used in most statistical books on

1.2 ONE WAY CLASSIFICATION 13

the ANOVA is the so called l!dot" notation. To introduce this notation

let us consider the general form of the model which is given in
Eq. (1.2.1). The symbols representing the data are displayed in an
array for n observations for each of t treatments as shown in the
following tabulation.

Observation

Treatment 1 2 3 ... j ... n Total Mean

The symbol T^# represents the total of all observations for the ith
treatment and the "dot" indicates a summation over the subscript that
has been replaced by the dot, that is T2# = y^y In addition,

y\# = T^#/n, the sample mean of the i ^ treatments. Note that the
grand total of all observations is T#> and the grand mean is
y
• • • = T • •/tn.

The method of calculation of the sums of squares for types of

tablets and the total sum of squares was indicated in the previous
section. The error sum of squares shown in Table 1.2.2 was then
obtained by subtracting SS types from SS total. In order to under­
stand the reasoning behind this calculation, observe that the basic
definition for the total sum of squares is
14 1. BASIC STATISTICAL CONCEPTS

1 n 7
SS total =I I (y. . - y . J
i=l j=l 13

2
By expanding ][ J (y.^ - y ##) this expression can be seen to be equal

to the formula for calculation, namely,

t n ?
SS total *I I yf. - CT
i=l j=l 13

In addition, by writing

(y^ - y . . ) 2 ■ [Cy±3 - yi#) + (yi# - y. . ) ] 2

it can be seen that

SS total = I 5 (y - y )2+ { ? (y - y..)2

i=l j=l 13 1 i=l j=l 1

t n
+ I I2(y.. -yi#)(y. - y.J
i=l j=l ij i i

Note that in the last term of the above equation that 2 (y\# - y ##) is
a constant with respect to the inner summation over j and that

l (y ^ - yj.) = o
j= i 13 1
consequently

SS total = 1 1 (y.. - y i#)2 + 1 1 ( ? i . - y..)2

i=l j=l 13 1 i=l j=l 1

Observing that

l l fti. - y..)2 = n I tii. - y..)2 (i.2.2)

i=l j=l i=l 1
1.2 ONE WAY CLASSIFICATION 15

= SS treatments

we see that

t n 2
SS total = I I ( Y ii - Yim) + SS treatments
i=l j=l 1J 1

so that

t n 2
SS error = I I (y.. - y. )
i=l j=l 1J 1

which is intuitively correct since (y^ - y^.)2 represents the

sum of squares within treatment i and if divided by its degrees of

2
freedom, n - 1, would yield an estimate of a .

This fact should be intuitive since all the variation within

treatment i is the result of the n e.. random variables each of which
2
has been assumed to have variance a . Thus the outer summation over
i produces a pooled sum of squares which divided by the total degrees
2
of freedom t(n - 1) yields the best estimate of a .

It should be noted that the above data array and the algebraic
relation among the SS total, SS treatments, and SS error was presented
assuming an equal number of observations per cell. This assumption
was only made for ease of notation and is not a mathematical
requirement. This statement remains valid for the following
discussion on the F-test utilized to test the hypothesis of no
significant differences among treatment means.

The F-test is well known for its use in testing for equality of
two variances. This test is carried out by computing the ratio of
two independent variance estimates. If this ratio is greater than
16 1. •BASIC STATISTICAL CONCEPTS

some tabular value it is concluded that the population variance

estimated by the numerator variance estimate is greater than the
other population variance. Suffice it to say that it must be
demonstrated that under the null hypothesis of no treatment effects
that MS treatments and MS error are estimating the same population
variance and hence (MS treatments)/(MS error) is distributed as F.
The fact that these estimates are independent will be assumed.

It was shown above that (SS error)/t(n - 1) » MS error is an

2
estimate of a . Thus it remains to show that MS treatments is also
2
an estimate of a under the hypothesis of no differences among
treatment means. Under the assumption of the null hypothesis
(Hq : T^ = 0), Eq. (1.2.1) can be rewritten as

yij " v * e(i)j {j = 1, n)

Thus, each of the treatment means, y\#, i = 1, ..., t, is merely a

random observation from the distribution of means which, by assumption,
2
must have mean u and variance a /n. It follows from Eq. (1.2.2) that

, t 2 1
rr r I (Xi. - y.J = f7ft~ T T css treatments)
i=l v

= i (MS treatments)

2
and hence must be an estimate of a /n so that MS treatments is an
2
estimate of a under the null hypothesis. Intuitively, it has been
demonstrated that the ratio

MS treatments
MS error

is distributed as F with t - 1 and t(n - 1) degrees of freedom.

1.3 UNEQUAL VARIANCES AND TRANSFORMATIONS IN ANOVA

In most ANOVA problems the assumption which was historically

1.3 UNEQUAL VARIANCES 17

thought to have been most critical was homogeneous variances. Box

(1954), however, demonstrated that the F-test in the ANOVA is most
*
robust for a while working with a fixed model with equal sample
sizes. Box (1954) showed that for relatively large (one variance was
up to 9 times larger than another) departures from homogeneity that
the a level may only change from 0.05 to about 0.06 which is not
considered to be of any practical importance. (It should be pointed
out that the only time the a level increased drastically in Box!s
study was when the sample size was negatively correlated with the
size of the variance.) Whenever larger departures from homogeneity
occur it is felt that the data should be transformed in order for the
ANOVA to produce meaningful results. The intent of this section is
to discuss tests for homogeneity of variances, and the various
transformations that may be utilized in the case of ’’too much”
heterogeneity.

No one to our knowledge, has come up with an a-level on

homogeneity tests which will indicate when the experimenter should
become concerned about making a transformation. A set of working
rules which seem to be effective for the practitioner are as follows:

i. If the homogeneity test is accepted at a = 0.01 level,

do not transform.

ii. If the homogeneity test is rejected at a = 0.001 level,

transform.

iii. If the result of the homogeneity test is somewhere

between a = 0.01 and 0.001, try very hard to find
out from the investigator the theoretical distribution
of the data. If there is a practical reason to transform
go ahead and transform; otherwise we recommend not to
transform. With the availability of transgeneration
options in statistical programs for digital computers

*
Robust for a in this situation means that even if the variances
are heterogeneous the a-level for the F-test on treatment means is not
changed materially.
18 1. BASIC STATISTICAL CONCEPTS

there is a tendency to try various transformations, make

the homogeneity test in each case and select the
transformation which yields the test statistic which is
most favorably located in the acceptance region. While
this practice is not all bad, there is danger in losing
"touch" with the meaning of the variable being
investigated. Hence use as much theory about the
variable as possible and select the transformation which
makes sense from a physical point of view. An example
of this type of selection is presented in Section 1.3.4.

An example in manufacturing tablets demonstrates tests of

homogeneity of variances and the uses of transformations. Four types
of magnesium trisilicate tablets:

(A) magnesium stearate 16mesh granule size

(B) talc powder 16 mesh granule size
(C) liquid petrolatum 16mesh granule size
(D) magnesium stearate 20mesh granule size

were completely randomized in manufacturing and testing.

Ten tablets were measured for disintegration time in seconds for

each of the four types. The data are presented in Table 1.3.1.

TABLE 1.3.1

Time (seconds) of Disintegration of Tablets for Four Types of Tablets

Types of tablets

20 42 8 12 50 124 151 178

28 25 10 24 67 72 125 151
36 24 12 10 90 78 180 152
16 31 16 19 103 70 149 161
25 33 9 10 90 76 175 118

Mean 28 13 82 154
Variance 59.55 26.22 430.89 436.22
1.3 UNEQUAL VARIANCES 19

On looking at the data one can see that there tends to be a

definite positive relationship between the means and the variance.
Ordinarily when one finds a relationship of this type between the
means and variances one also finds a lack of normality in the data.
It is our experience that transformations which improve the
heterogeneity situation also improve the lack of normality problem.
The relationship between the assumption of homogeneity of variances
and that of normality is discussed by Cochran (1947).

This visual relationship provides a warning that tests of the

homogeneity of variances should be considered. Three of these tests
follow in the subsections. The Bartlett (1937) test is an old,
well-established test that has been programmed for most computers,
but it is a bit sensitive to nonnormality, especially if the tails
of the distribution are too long it tends to show significance too
often. Using our working rules if the investigator has no
theoretical knowledge of his variable, we would suggest using
a = 0.001 when the distribution of y seems to have excessively long
tails. This will tend to prevent too many unnecessary transformations.

Another test, suggested by Bartlett and Kendall (1946), utilizes

2 2
the log s as a variable. They suggest running ANOVA on the log s ,
2
assuming the variances of log s are homogeneous, in order to find
out which treatments cause this heterogeneity. The basis for using
2
log s as the variable is that it is approximately normally distributed
if n 5. This test is not used as much as we think it should be
because too many practitioners do not know about it. One drawback,
2
of course, is that to obtain a minimum of two observations of log s
per cell there should be at least 10 observations of y in each cell.
The group of 10 observations may be broken randomly into two groups
2
of 5 observations each and log s calculated for each group. Not
only does this test furnish the experimenter with another means of
investigating the assumption of homogeneity of variance but also
provides a procedure for determining the causal effects that
influence the variability of a process. Ordinarily the ANOVA is
used to study the effects of the factors on the response variable
20 1. BASIC STATISTICAL CONCEPTS

2
but through the log s technique one can also study the relationship
between the factors and variability of the process.

The only other test of homogeneity we will describe is the

Burr-Foster Q test, Burr and Foster (1972). This test has not been
published and we believe it should be available to the practitioner
because the test statistic is easy to calculate and does not have
the sensitivity of normality departures encountered with the Bartlett
test.

1.3.1 Bartlett Test (Equal Subclass Numbers)

Table 1.3.2 shows the procedure to test for homogeneity of the

variances using Bartlett’s test for equal subclass numbers on data in
Table 1.3.1. Snedecor and Cochran (1967, p. 296 and 297), give a
procedure for unequal subclass numbers.

TABLE 1.3.2

Bartlett’s Test of Homogeneity

2 2
k Type s log s

1 A 59.55 1.77488
2 B 26.22 1.41863
3 C 430.89 2.63437
4 D 436.22 2.63971
Total 952.88 8.46759

T Zs2 952.88
s = 238.22
k 4

log s2 = 2.37698

M = 2.3026 (df) [k log s2 - I log s2]

where: df = degrees of freedom per variance = 9
k = number of treatments or samples (type) = 4
M = 2.3026 (9) [4(2.37698) - 8.46759] = 21.$6
1.3 UNEQUAL VARIANCES 21

2 M 21.56 *
*3df = C “ 0 4 6 " 20-6
and
xcrit 3df(°-001) = 16-3

Reject homogeneity of variances.

1.3.2 BARTLETT and KENDALL log s2 ANOVA

The data in Table 1.3.1 were randomly assigned to two groups of

5 for each type of tablet. Table 1.3.3 shows the ANOVA of the log s^

TABLE 1.3.3
2
Log s Analysis of Homogeneity

Type of tablet
A B C D

2 . s2 2 . 2 2 2 2 2
s log s log s s log s s log s
59 1.7709 10 1.0000 449.5 2.6527 493.0 2.6928
52.5 1.7202 39 1.5911 510.0 2.7076 478.5 2.6799

2
ANOVA of log s

Source df MS F (0.01)
^calc crit(3,4)
Types of tablet 3 0.96876 21.8 16.7
Within error _4_ 0.04439
Total 7

Since Fca^c (calculated from data) > Fcr^t we reject the

hypothesis that the variances are homogeneous at a = 0.01. A further
refinement of the analysis would be to run a Newman-Keuls test on the
2
type of tablets means of log s and show which ones are different.

Problem 1.3.1. Run the Newman-Keuls test on means of the

(log s^Vs of Table 1.3.3 and explain the results.
22 1. •BASIC STATISTICAL CONCEPTS

1.3.3 BURR-FOSTER Q-Test of Homogeneity

The Q-test for equality of variances is based on a statistic

which is a monotone function of the coefficient of variation of the
sample variances. As such it offers promise as a preliminary test
for the assumption of homogeneity of population variances which is
needed in the ANOVA technique. Although the Q-test is not a so-called
quick test, the test statistic is sufficiently simple to permit
calculation of its value on a desk calculator. A sample variance
taking the value zero does not disrupt this test (as it does to
Bartlett*s test).

For equal sample sizes, n, from each of p parent populations,

2 th
let s^ (for i = 1, ..., p) denote the i sample variance. Denoting
the value of the test statistic by q, we have:

q = (s* + ... + sp )/(s2 + s2)2

2
For unequal sample sizes we specify that each sample variance, s^,
be calculated by dividing by the degrees of freedom, v^, rather than
by the sample size, n^ (where = n^ - 1, for i = 1, ..., p ) . Let
v denote the arithmetic average of the degrees of freedom. In this
case we have:

q = v (V J + ... + vpsJ)/(V 2 + ... + vps2)2 .

Large values of q lead to rejection of the hypothesis of equal

population variances. The critical values are given in Appendix 8
for various numbers of parent populations, p, and various possibilities
for equal degrees of freedom v (where v = n - 1). This table can be
used directly for equal degrees of freedom. For unequal degrees of
freedom, q is calculated as indicated above, but v is to be substituted
for v, provided that V and the harmonic mean of the v*s do not differ
greatly. For large sample sizes (v > 60), we note that p v(pq - l)/2
is asymptotically chi-square with (p - 1) degrees of freedom.

For the tablet example, the Burr-Foster test is carried out as

follows:
1.3 UNEQUAL VARIANCES 23

2 4
Types_______ s___________ s
A 59.55 3,546.20
B 26.22 687.49
C 430.89 185,666.19
D 436.22 190,287.89
Total 952.88 380,187.77

= 380,187,77 „ Q>41fl
q =
■ (952.88)

where: v = 9; p = 4 and using

linear interpolation one gets

(°;0D a 0 398
ncrit

q <°:?01) - 0 . 4 M
ncrit

Since q(0.01) < calculated q < q(0.001) and theory indicates

that disintegration time is not normally distributed, we reject the
hypothesis and transform the variable (note paragraph iii of Section
1.3).

1.3.4 Transformation of y

The earlier results of this section indicated that if Bartlett*s

test were used in analyzing the data in Table 1.3.1 the hypothesis of
homogeneity of variances would be rejected at the 0.001 level. For
2
both the log s and Q-test the hypothesis was rejected at the 0.01
level. Based on this evidence, along with information from the
experimenter that this type of data usually appeared to have the
sample mean proportional to the sample variance (Dixon and Massey,
1969 p. 324) it was decided that the analysis of these data should
proceed by utilizing the square-root transformation. The results of
24 1. BASIC STATISTICAL CONCEPTS

this transformation on the data of Table 1.3.1 are shown in Table

1.3.4. In addition the results of performing Bartlettfs test on the
transformed data, Sy, is also included in this table.

TABLE 1.3.4

Analysis of Homogeneity of the Variances Using / y

Type log [ s ^ ]

A 0.538 -0.2692
B 0.441 -0.3566
C 1.279 +0.1069
D 0.734 -0.1343
Total 2.992 -0.6522

M = 2.3026(9) [4 (-0.1261) + 0.6522]

= 3,063
C = 1.046

= 2.93 : X^fCO.Ol) = 11.34, X^CO.OS) = 7.81

Accept the hypothesis of the homogeneity of the

variances using Jy (Using Bartlett's test)

It is seen here that the hypothesis of homogeneity of variances is

now accepted even at the 0.05 level. This result is also substantiated
by the Q-test. At this point one is now ready to carry out the ANOVA
on Jy as described in Section 1.2.

Problem 1.3.2. (a) Carry out the Q-test on the data furnished
in Table 1.3.4. (b) Run the ANOVA on Jy in Table 1.3.1, p. 18
and explain your results, (c) Determine the 95% confidence
interval for the mean of the Type D population.

1.3.5 Other Transformations

As mentioned previously, transformations are utilized to obtain

a variate which when analyzed via ANOVA techniques will conform to
the basic assumptions of the analysis of variance. The most familiar
1.3 UNEQUAL VARIANCES 25

paper on this subject is that by Bartlett (1947). This paper

contains a listing of several often used transformations along with
the relationship which should exist between the mean and a function
of the variance of population.

The most commonly used transformation for "growth"-type data,

for example, changes in weight of animals caused by diets, is log y.
Log y is also used if the mean is proportioned to the standard
deviation.

In percentage data or proportions (p) dealing with the binomial

(0 or 1 response), the variance is proportional to [mean (1 - mean)]
and the correct transformation is arc sin Sp. Usually one does not
carry out this transformation if the data are all between p = 0.30
to 0.70 because the variance is fairly well stabilized in this region.

If the standard deviation is proportional to the square of the

mean the reciprocal (1/y) is quite often used.

Ostle (1963, p. 340), provides a summary table of the four

transformations given here.

In addition, recently we have encountered data in which the mean

is negatively correlated with the variance. In this situation the
transformation, vlf - vfe - y, where B is the upper limit of all y's
in the data, has worked quite well.

1.3.6 Test for Normality

There are numerous tests for normality, however, we recommend

the W test developed by Shapiro and Wilk (1965). This test is such
that it is appropriate for a test of the composite hypothesis of
normality, i.e., one does not have to include the mean and variance
as part of the hypothesis as is common in some other tests for
normality such as the Kolmogorov-Smirnov and chi-squared tests.
These tests as well as others were compared by Shapiro, Wilk, and
Chen (1968) and they demonstrated that the W test was generally
superior in detecting nonnormality when evaluated on various
26 1. • BASIC STATISTICAL CONCEPTS

symmetric, asymmetric, short- and long-tailed alternatives over

sample sizes ranging from 10 to 50.

Tables of coefficients required to calculate W are given in

Appendix 9 and the percentiles of W appear in Appendix 10. When
using these tables one must keep in mind that small values of W
indicate nonnormality.

As an example to illustrate the use of this test consider the

data of tablet Type B which appears in Table 1.3.1. The following
steps must be carried out:

1. Order the n observations as .1 .1 ••• 1. Xn•

2. Compute E(yi - y) .

3. If n is even, n = 2k, compute

k
b = V i + i (V i +i - y p

where the values of aR appear in Appendix 9. If n is

odd, n = 2k + 1 then one omits the sample median, y^+^> and
calculates

k
b = Vi+i (V i +i -

4. Compute W = b2/E(y^ - y)2.

5. Compare W to the percentage points given in Appendix 10.

Again, small values of W indicate nonnormality.

In our example we have the following for each step:

1. Ordered observations: 8, 9, 10, 10, 10, 12, 12, 16, 19, 24.

2. S(yi - y)2 = 236.

3. b = (0.5739)(24-8) + (0.3291)(19-9) + ... + (0.0399)(12-10)
= 14.0826.

4. W = (14.08)2/236 = 0.840 (calculated).

1.4 CURVE FITTING 27

5. W(0.05, 10) = 0.842, thus we would reject the hypothesis of

normality at the a = 0.05 level (from Appendix 10).

The square root transformation on these data will tend to normalize

them. A reference on transformations is Bartlett (1947).

Problem 1.3.5. Use the W test to test the data utilized in the
above example for normality after carrying out the square root
transformation.

1.4 CURVE FITTING IN ONE WAY CLASSIFICATION

In most design of experiments problems involving quantitative

factors one usually becomes interested in describing the relationship
between the dependent variable (response or yield) and the independent
factors by means of a mathematical equation. In this section we will
describe the procedure of using orthogonal polynomials to obtain an
equation which estimates the response surface for the case of one
quantitative factor. The next chapter will generalize this procedure
for the case of more than one quantitative factor.

The procedure of determining estimates of coefficients in the

mathematical model and determining their relative importance in
explaining the total variation (sum of squares) in the response
variable is normally referred to as regression analysis. For
detailed procedures on this subject see Draper and Smith (1966),
Anderson and Bancroft (1952), Ostle (1963), and Myers (1971).
Regression analysis as described in these references is based on the
method of least squares. The use of orthogonal polynomials for
estimating the response surface associated with a "well"-designed
experiment is equivalent to using the method of least squares. In
this context, a we11-designed experiment implies that the same number
of observations are taken at each level of the independent variable
(X) and that the levels of X are equally spaced. It is quite
feasible that one would be interested in designing an experiment
where some function of X was equally spaced rather than the values
of X. For example, the effect of the logarithm of time on the
28 1. BASIC STATISTICAL CONCEPTS

electrical characteristics of electronic components is frequently

studied. In this case all calculations are carried out using the
logarithm of time instead of time itself. For the case of unequally
spaced X values one is referred to Anderson and Bancroft (1952) and
Robson (1959).

It should be pointed out to the practitioner that the use of

orthogonal polynomials merely estimates the response surface with a
certain degree polynomial. In most cases this polynomial will be
an approximation of some other mathematical function. If the true
functional relationship between the dependent and independent
variable is known it should be used in the analysis of the problem.
The use of the true function may possibly involve the use of nonlinear
models (see Draper and Smith, 1966, Chapter 10).

1.4.1 Orthogonal Polynomials

To introduce the topic of orthogonal polynomials we first

consider the case with one observation per treatment combination
(X level). We are also only considering one independent variable
where the responses are observed at equally spaced values of X. An
example might be hardness readings (y) taken at equal intervals across
a block of cured synthetic material. In this example the distance
from one edge would be the X value. These X values will be assumed
to be the values 1, 2, 3, ..., n where n is the number ofX levels.
If this is not the case and we have a factorX* which has levels
2, 9, 16, and 23, then the levels of X will become 1, 2, 3, and 4 if
we code the Xf values with the formula

X = [Xf + (the spacing) - (the smallest X* value)]/(the spacing)

The practitioner immediately becomes interested in finding the

smallest degree of the polynomial in X which best describes his data.
The general form of the model is
1.4 CURVE FITTING 29

which can be written as an orthogonal polynomial model

k k
where the Zfs are functions of X and the afs are the corresponding
coefficients.

The Z(X) expressions with the various subscripts are the

orthogonal polynomials which implies that Z = 0 ^or r t s* The
formulas for the first three orthogonal polynomials are given below
where the Afs are scale factors and make the Z values become integers
for easy calculations. References for this material are Anderson and
Bancroft (1952) p. 210, Ostle (1963) p. 194, and Myers (1971) p. 35.

Z = Z
1 linear
2
= X2 [(X - X)2 -

In order to obtain the estimate of the response surface one must also
determine estimates of the coefficients of the polynomials. For
example,

a.
1

The numerator of the and are sometimes called contrasts since

Z Z^ and Z 1 ^ in each case is zero. In general, if T^ represents
a total of r observations and the c.'s are constants where Ic. = 0,
l i
then C « is a contrast and
30 1. BASIC STATISTICAL CONCEPTS

When one recalls that we are going to assume that the X values
take on the values 1, 2, ..., n, it is readily seen that one is able
to create a table of Z values in order to simplify calculations.
Appendix 11 gives an abbreviated table of Z's along with the
2
corresponding Xfs and E(Z) . The case where n = 6 is reproduced in
Table 1.4.1 for illustrative purposes.

TABLE 1.4.1

Orthogonal Polynomials for n = 6

Notice for example that these values are orthogonal, e.g., EZ^ Z^^
® (-5)(-5) + (-3)(7) + ... + (5)(5) = 0. In analyzing a set of
data for trends of various degrees it becomes necessary to be able
to compute the sums of squares which is explained by each of the
terms included in the model of Eq. (1.4.2). For example, the
formula required for the sum of squares (SS) for the quadratic term
is

SSquadratic = = “l l ^ l P

The same procedure is used for any degree terms given in Eq. (1.4.2).
The following example concerns the observation of hardness as
measured every 2 centimeters across a block of synthetic material.
The data and the method of calculating the afs are shown in
Table 1.4.2.
1.4 CURVE FITTING 31

TABLE 1.4.2

Calculation Method for Determining the a fs

x» y X
Z1 ^1 Z11 yZll Zlll ^Zlll

2 16 1 -5 -80 5 80 -5 -80
4 14 2 -3 -42 -1 -14 7 98

o
6 10 3 -1 -10 -4 4 40

00
i
8 12 4 1 12 -4 -48 -4
10 13 5 3 39 -1 -13 -7 -91
12 15 6 5 75 5 75 5 75
80 -6 40 -6
O
olvo
O1

13.33 a -6
ii

ao = -0.0857
1 70 ~
volo

40
1l

0.4762 -0.0333
II

ii
|rH

®ir 84 ~ alll
oo

The resulting linear equation for the response surface would be

y = aQ+ ax Zx = 13.33 -0.0857 [A^X-X)]

= 13.33 - 0.0857 [2(X-3.5)]

= 13.93 - 0.1714 X

= 13.93 - 0.1714

= 13.93 - 0.0857 X'

The analysis of variance for this part of the calculations is given

in Table 1.4.3. The sum of squares due to linear is basically that
portion of the total sum of squares in y which can be explained by
the first term in the polynomial. In regression analysis terms this
is often referred to as the sum of squares due to regression. The
error sum of squares is thought of as the sum of squares about the
regression line. The three sums of squares in this ANOVA were
computed as follows
32 1. BASIC STATISTICAL CONCEPTS

SSlinear = = V * y Zl3 = (-0.0857) (-6) = 0.514

sstotai = z<y - y)2 = 2 y2 - - 1 0 9 0 -^ = 23.333

SS = Z(y- (13.93 - 0.0857 X'))2 « SS. . - SS... = 22.819

error w v total linear

TABLE 1.4.3

ANOVA for the Linear Model Using Data of Table 1.4.2

Source df SS MS R2

,514)
Linear regression 1 0.514 0.514 0.022 = | 2 ^;
,333)
Error 4 22.819 5.705
Total 5 23.333

The proportion of the total sum of squares in y which is explained

2
by regression is given the name R .

In general the degrees of freedom for regression is oneless

than the number of coefficients estimated which would be k when one
uses the models given in Eq. (1.4.1) and (1.4.2). As in theANOVA
section the degrees of freedom for total is the number of observations
2
minus one. The coefficient, R , has a major weakness in that as the
2
degrees of freedom for error becomes small, R may be quitelarge
(near one) but the prediction value may not begood. One may test
2 2
the hypothesis that p , the parameter estimated by R , is zero by
the test

p - MS due to regression_ 0.514 _ ^

k,n-k-l ~ MS error - 5.705

which is quite non-significant. Thus we would conclude that this

model does not contribute significantly towards explaining the
variation in y.

If a plot of the data given in Table 1.4.2 were made it would

easily be seen that a linear model would not be expected to explain
1.4 CURVE FITTING 33

much of the variation in y. However, we continue this exercise as

this method will carry over to the case of several independent
variables where data plots are not as easily interpreted. Thus, on
2
the basis of the small R and the appearance of a data plot, one
would want to include the quadratic polynomial in the response surface
model. In ordinary regression analysis a completely new set of
calculations would have to be performed. With orthogonal polynomials
this is not necessary as one is permitted to merely add on the
quadratic term in the model and add together the two sums of squares,
i.e., 1inear and quadratic.

The best fitting quadratic model for this example becomes

y = y + «i z i + ° n zn
2
= 13.33 - 0.0857 [X (X-Y)] + 0.4762 X2[(X-X)2 -

20.60 - 5.1715 X + 0.7143 X2

= 20.60 - 2.5857 X' + 0.1786 X'2

The sum of squares contributed by the quadratic term is

an(Z y Zn ) = 0.4762 x 40 = 19.048

and the sum of squares due to regression becomes

SS . = 0.514 + 19.048 = 19.562

regression

The ANOVA for this model is given in Table 1.4.4.

TABLE 1.4.4

ANOVA for the Quadratic Model Using Data of Table 1.4.2

Source df SS MS F R2

Linear 1 0.514
Quadratic _1 19.048

Regression 2 19.562 9.781 7.78

°-»38 ■ £ 1“
Error 3 3.771 1.257
Total 5 23.333
34 1. BASIC STATISTICAL CONCEPTS

2
The figures in Table 1.4.4 still do not indicate a significant R
2
when one uses the F-test; however, the relatively large jump in R
from 0.022 for the linear model to 0.838 for the quadratic model would
probably lead one to consider the quadratic model as the appropriate
model. If one continued this example and calculated the cubic sum of
squares one would not find much of an increase in the regression sum
of squares. In order to obtain any further information on the fit of
this model we need additional observations at the same X values which
is the topic of the next section.

1.4.2 Lack of Fit Principle

When one has repeated observations at each X value (when using

orthogonal polynomials it is desirable to have an equal number of
observations at each X value) the degrees of freedom for regression
remains at k, assuming the model given in Eq. (1.4.1). Consequently,
the degrees of freedom for error is increased by the number of
additional observations. At this point we can consider partitioning
the error sum of squares and the degrees of freedom into two parts.
One part is that associated with '’pure error" or the within cell sum
of squares which is pooled across all X values. To calculate the
pooled error one would calculate the sum of squares about the mean
response at each X level and sum over all X, thus

m.
■i ( I Xj)2
y v 2 j=l
SSpooled error “ x yj m7 ^

where m. = number of observations at the i ^ X value and the summation

over j takes place only for the y values associated with the i X
value. The degrees of freedom for pure error is - 1) where
there are n X values. The "pure error" mean square thus becomes an
estimate of the variance of y. The second part of the initial error
sum of squares is now attributed to "lack of fit" and the degrees of
freedom are calculated by subtraction. Under the null hypothesis
1.4 CURVE FITTING 35

that the selected model is adequate to account for all possible

variation in y we see that the initial error mean square is also an
estimate of the variance of y. Thus when we partition the error sum
of squares the second part (the lack of fit sum of squares) divided
by its degrees of freedom will also be an estimate of the variance
of y. Thus the ratio of the lack of fit mean square to the pure
error mean square under the null hypothesis will be distributed as
F with n-k-1 and (nu - 1) degrees of freedom. The reader should
be able to picture the situation that occurs when the null hypothesis
is false. The initial error sum of squares is large due to the poor
fit. The pure error sum of squares will remain the same thus forcing
the lack of fit sum of squares to be large. Hence, the lack of fit
mean square becomes larger than it should be if the null hypothesis
were true.

The above procedure is illustrated by means of an example. This

example is merely an extension of the example given in the previous
section. Consider that two sets of hardness readings were taken
completely at random over the same set of X values. Thus we have
two observations at each X value. The calculations are essentially
the same except that one uses the mean of the y !s for each X value
rather than the individual y's which were used in Section 1.4.1,
where there was only one yfor each X value. For example,

n
SS due to the quadratic term = a.,(m J y. Z.J
i=l 1 L

where m is the number of observations at each X value and y. is the

th
mean of the y fs for the i X value. The data and the calculations
for this example through the quadratic term appear in Table 1.4.5.

The ANOVA for this example appears in Table 1.4.6. Note that
the lack of fit is carried out twice and in general it would be
carried out on successive terms until the hypothesis of an adequate
model is accepted or there are no degrees of freedom remaining for
lack of fit. One must recognize that this testing procedure involves
36 1. BASIC STATISTICAL CONCEPTS

TABLE 1.4.5

Calculation Method Using Two Observations per Level

X* y X y *Z1 *Z11 ^Zlll

2 16,16 1 16.0 -80.0 80.0 -80.0

4 14,13 2 13.5 -40.5 -13.5 94.5
6 10,12 3 11.0 -11.0 -44.0 44.0
8 12,11 4 11.5 11.5 -46.0 -46.0
10 13,12 5 12.5 37.5 -12.5 -87.5
12 15,14 6 14.5 72.5 72.5 72.5
79.0 -10.0 36.5 -2.5

79
13.17
ao 55 6 ~
"4h-*
11
o|o

=
1!

-0.14
al

= 36.5 = 0.43
all 84 *

2 X (-10)2
SSCap SSL = 2.86
70

2 X (36.5)2
SS(on ) SSQ = = 31.72
84

contingent tests (Bozivich et al., 1956) and the a level will

increase. Hence the 3-error willdecrease and ifthehypothesis is
accepted one may have more terms in themodelthanrequired, but
there should be no more terms required in the model than that
which is determined. This is a strong reason to use this method
to evolve a model.

The appropriate quadratic models for this example in terms of

the coded and uncoded variable are as follows:

y = 20.25 - 4.85X + 0.65X2

= 20.25 - 2.42X' + 0.16X'2
1.4 CURVE FITTING 37

TABLE 1.4.6

ANOVA for the Data of Table 1.4.5

Source df SS MS F R2

Linear 1 2.86 0.072

Lack of Fit 32.81 x 8.20 12.30
Quadratic ) 31.721 ) 0.872
Lack of Fit 3J 1.09J 0.36 0.54
Error j6 4.00 0.67
Total 11 39.67

Problem 1.4.1. Verify all calculations and results of the trend

analysis example above.

Problem 1.4.2. An engineer conducted an experiment on a jet

aircraft engine in which he was interested in finding out the
polynomial that best described the temperature conditions
around the engine. Once he obtained the equation he could keep
the temperature as even as possible by designing the appropriate
cooling system.
At each of seven equally spaced positions around the engine,
the investigator took five readings at random over a specified
time. The results were as follows:

y = °F
Positions

1 2 3 4 5 6 7
1750 1700 1730 1710 1630 1665 1775
1775 1750 1700 1665 1660 1690 1710
1730 1720 1690 1705 1645 1710 1735
1725 1725 1700 1680 1600 1705 1740
1770 1705 1680 1690 1615 1690 1740

Determine the best-fitting polynomial for these data.

Problem 1.4.3. Taking the data

x___ L
1 4
2 5
3 7
4 8
38 1. BASIC STATISTICAL CONCEPTS

(a) Demonstrate that for the nonorthogonal variable X, in

the model y = Bq + S^X + e the estimates of 8^ and 8^ will
change if either is deleted (assumed zero) from the model (by
actual calculations) and plot the results to show this.
(b) Show, by calculations, that the estimates of and
do not change for the orthogonal variable, Z, in the model
y - + ai z + e
when either or isdeleted (assumed zero) from the model.
Plot the appropriate results.
(c) What have you discovered about the estimates of the
coefficients obtained in (a) and (b) above?
(d) For each Z model, transform back to X.
(e) Do you prefer the X or Z model for predictive purposes?
Explain your reasoning.

1.5 REFERENCES
Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research
McGraw-Hill, New York, 1952.
Bancroft, T. A. Topics in Intermediate Statistical Methods, Iowa
State University Press, Ames, Iowa, 1968.
Bartlett, M. S. Proc. Roy. Soc.A, 160:268(1937).
Bartlett, M. S. Biometrics 3:39 (1947).
Bartlett, M. S. and Kendall, D. G. JRSS Suppl.8:128(1946).
Box, G. E. P. Ann.Math. Stat. 25:290 (1954).
Bozivich, H., Bancroft, T. A., and Hartley, H. 0. Ann. Math. Stat.
27:1017 (1956).
Bratcher, T. L., Moran, M. A., and Zimmer, W. J. J. Quality Tech.
2:156 (1970).
Burr, I. W. and Foster, L. A. A Test for Equality ofVariances,
Department of Statistics Mimeo Series No.282, Purdue University,
Lafayette, Indiana, 1972.
Carmer, S. G. and Swanson, M. R. JASA 68:66 (1973).
Cochran, W. G. Biometrics 3:22 (1947).
Dixon, W. J. and Massey, F. J., Jr. Introduction to Statistical
Analysis, 3rd ed., McGraw-Hill, New York, 1969.
Draper, N. R. and Smith, H. Applied Regression Analysis, Wiley,
New York, 1966.
Eisenhart, C. Biometrics 3:1 (1947).
1.5 REFERENCES 39

Kastenbaum, M. A., Hoel, D. G., and Bowman, K. 0. Biometrika 57:421

(1970a).
Kastenbaum, M. A., Hoel, D. G., and Bowman, K. 0. Biometrika 57:573
(1970b).
Keuls, M. Euphytica 1:112 (1952).
Myers, R. H. Response Surface Methodology, Allyn and Bacon, Boston,
Mass., 1971.
Newman, D. Biometrika 31:20 (1939).
Ostle, B. Statistics in Research, 3rd ed., Iowa State University
Press, Ames, Iowa, 1963.
Robson, D. S. Biometrics 15:187 (1959).
Shapiro, S. S. and Wilk, M. B. Biometrika 52:591 (1965).
Shapiro, S. S., Wilk, M. B., and Chen, H. J. JASA 63:1343 (1968).
Snedecor, G. W. and Cochran, W. G. Statistical Methods, 6th ed.,
Iowa State University Press, Ames, Iowa, 1967.
Stein, C. Ann. Math. Stat. 16:243 (1945).
 Chapter 2

SOME INTERMEDIATE DATA ANALYSIS CONCEPTS

In Chapter 1 we described a few ideas in data analysis using the

one-way classification from a completely randomized design. In this
chapter we again utilize only the completely randomized design and
discuss the two-way classification and its advantages over the one-way.
The concepts used in analyzing two-way classified data are more
involved than those encountered in the one-way.

For instance, in most research problems of any consequence, the

information on the combined effects of two factors controlled in the
experiment (two-factor interactions) on the response variable usually
contributes more to solving the problem than the information on the
individual effects of the factors themselves (main effects).

By extending the analytical procedures to many controlled factors

in an experiment, the investigator is able to obtain the information
on all the two-factor interactions simultaneously and additional
information on the three-factor interactions also. Only recently
(to our knowledge) have some research workers, for example, highway
research engineers Hadley et al. (1969), interpreted three-factor
interactions thoroughly enough to want to use them in the mathematical
models to help to explain the responses obtained. As a general rule,
however, the four-factor interactions and higher tend to be small
contributors to the underlying ANOVA models used to explain the
mathematical phenomenon. These high order interaction effects are
frequently assumed zero to allow an estimate of the error if only one
experimental unit has been used for each combination of the factor
levels, often called the treatment combination. Details of this
extended type experiment with many factors are given in Chapter 4
on factorial experiments.
40
2.1 ONE OBSERVATION PER CELL 41

In this chapter we are especially interested in expressing the

interaction concept. In the first section two factors are
investigated thoroughly, using only one experimental unit per
treatment combination, or one observation per cell, and in the
following section, concepts on more than one observation per cell
are covered.

2.1 TWO FACTOR EXPERIMENTS WITH ONE OBSERVATION PER CELL

When experiments are run with only two factors (which is

certainly more informative than running only one factor at a time
because no information on the two-factor interaction can be obtained
for one factor at a time experiments) and one observation per cell,
there is not a clear-cut way to separate the effect of the interaction
of the two factors from the experimental error.

In an experiment on stress-rupture life of material used to make

turbine blades in fanjet aircraft engines, two factors; blade
temperatures which might be encountered at takeoff ”of the aircraft11
and alloys or materials that are of interest to the research engineer,
are to be investigated. Let us assume that for part of the experiment
12 treatment combinations (4 materials x 3 temperatures) are the
only ones of interest and that these combinations can be run
completely at random in the experiment. The experimenter can then
analyze the data as if he has a fixed two-way ANOVA with only one
observation per cell. The model for such an analysis is:

y. . = y + M. + T. + MT. . + e r . (2.1.1)
ij i 3 13 (13)
i = 1, 2, 3, 4 j = 1, 2, 3

where

y.. = hours to stress-rupture of the piece of metal made of

th th
the i material at the j temperature

y = overall mean
42 2. INTERMEDIATE DATA ANALYSIS

= the effect of the i**1 material (fixed)

th
Tj = the effect of the j temperature (fixed)

MT.. = the effect of the interaction of the ith

th
material with the j temperature
2
S(ij) = t*ie exPer^mental error (random), NID (0, a )

In this model the subscripts for MT and e are the same, thus
indicating that the effects due to the interaction and the error
effects will be confounded. The brackets on the subscripts of e
indicate there is only one observation per cell or only one
experimental unit (piece of metal) for each material-temperature
combination. We could write the error symbol indicating one
observation per cell but the e , . i s sufficient if understood.

The data for this part of the experiment are given in Table
2. 1. 1.

TABLE 2.1.1

Stress-Rupture Life (hours) for Four Materials

and Three Temperatures

Temperatures

Materials 1 2 3 Total

a 185 182 182 549

b 175 183 184 542
c 171 184 189 544
d 165 191 189 545
Total 696 740 744 2180

Using a = 0.05 the ANOVA for these data is given in Table 2.1.2.
2.1 ONE OBSERVATION PER CELL 43

TABLE 2.1.2

ANOVA of Data in Table 2.1.la

F(0.05)
Source df SS MS F
crit

Material (M) 3 8.67 2.89 <1 4.76

Temperature (T) 2 354.67 177.34 3.7 5.14
MT and/or error j6 291.33 48.56
Total 11 654.67

SS materials = *>42 + .54S2 _ .(2180)2

12
= 8.67

6962 + 7402 + 7442 (2180) 2

SS temperatures -------- j
12

= 354.67

SS total = 18S2 + 1752 + ... + 1892 -

= 654.67

SS MT and/or
a SS total - SS materials - SS temperatures
error
= 654.67 - 8.67 - 354.67

= 291.33

The conclusion is that all effects are not significant at the

0.05 level. The F-tests for temperature and material effects are
based on using the MT and/or error mean square for the denominator
which is really assuming the interaction of temperatures by materials
is zero. To show this we need to derive the expected mean squares
for each source of variation so that the theoretical basis for the
F-tests can be explained. Suffice it to say here that we will write
44 2. INTERMEDIATE DATA ANALYSIS

these expected values down showing where the assumption of the

interaction being zero is made and later use an algorithm on a more
complicated model to derive the expected mean squares. A good
reference for the mathematical derivation of expected mean squares
is Anderson and Bancroft (1952, Chapters 18 and 20).
2
The expected mean squares (EMS) shown as a sum of a and an
unknown function of the fixed effects are given in Table 2.1.3.

TABLE 2.1.3

Expected Mean Squares for Table 2.1.2

Source EMS

Materials (M) o2 + 3<|>(M)

Temperatures (T) a2 + 4<KT)

MT and/or error a2 + KMT)

In order to make a test of significance on temperature using F,

we need a denominator that has no term or component of temperature
(T). The hypothesis that there is no temperature effect may be given
as:
Hq : +(T) = 0

where <|>(T) is a function of the effect of temperature. To test Hft

2
using the F-test we need a denominator that estimates a only, but
in this experiment because we took only one observation per cell we
cannot separate the interaction functional effect <j>(TM) from the
error variance, a . Refer back to the model (2.1.1) and recall that
the subscripts on MT and e were the same. This indicates complete
confounding.

In practice for testing temperatures in Table 2.1.2 we used

F - MS temperatures
2,6 ” MS MT and/or error
2.1 ONE OBSERVATION PER CELL 45

but to do this we must assume the interaction is zero or <j>(MT) = 0.

Similarly for terting for materials.

The conclusion of course, is that if the interaction, MT, may be

important one should take more than one observation per cell to
evaluate it. This concept is covered in Section 2.2. Now, however,
we wish to show various analyses to investigate a part of the
interaction separate from error where there is only one observation
per cell.

2.1.1 Both Factors Qualitative

Even if only two experimental units (pieces of material) were

used for each treatment combination, the expense of the experiment
would double if the experiment were completely randomized. Hence
sometimes it is worth examining the interaction or nonadditivity of
the model shown in Eq. (2.1.1) using a technique given by Tukey
(1949).

Table 2.1.4 shows the technique of splitting the sum of squares

291.33 for MT and/or error into one portion for MT which has one
degree of freedom and the remaining portion for error with 5 df.

TABLE 2.1.4

SS for One Degree of Freedom for Nonadditivity

Temperature

:erial i 2 3 <y±. - y,
*i.

a 185 182 182 183.000 1.333

b 175 183 184 180.667 -1.000
c 171 184 189 181.333 -0.333
d 165 191 189 181.667 0.000

174 185 186 y## = 181.667

i - y.J -7.667 3.333 4.333

46 2. •INTERMEDIATE DATA ANALYSIS

a [1 yii • y.J]2
aSS (for 1 df) i,j 13 1____________I _________
Nonadditivity ‘ [E (- . - _)2][I (- . -_,2,
i j J

[185(1.333) (-7.667) +... + 189(0.000) (4.333)12

((1.333)2 +...+ (0.000)2] [(-7.667)2 +..+ (4.333)2;

73.01

where is defined as in Eq. (2.1.1); y^# = the mean of the

th th
i material; y ^ * mean of the j temperature;

y*.. = overall mean.

Using the result of Table 2.1.4, the ANOVA in Table 2.1.5 is

appropriate to study.

TABLE 2.1.5

ANOVA including Nonadditivity Source

F P(0.0£
Source df SS MS EMS
crit

Materials (M) 3 8.67 2.89 o2 + 34>(M) 0.07 5.41

Temperatures (T) 2 354.67 177.33 a2 + 4<KT) 4.06 5.79

MT (nonadditivity) 1 73.01 73.01 a2 + <KMT) 1.67 6.61

Error J5 218.32 43.66 a2

Total 11 654.67

It still turns out that all F-tests are not significant at the
0.05 level, including the interaction of MT, an indication that the
MT term could be deleted from Eq. (2.1.1).
2.1 ONE OBSERVATION PER CELL 47

2.1.2 One Factor Qualitative and the Other Quantitative

There is more efficiency from a quantitative factor than from

a qualitative factor because the majority of the quantitative effect
is usually best described by a linear component with only one degree
of freedom (recall the concept of investigating trend in Section 1.4).

If one uses the concept of expanding any function into a power

series, he realizes that ordinarily the lower order terms (e.g.,
linear) contribute more than higher order ones in explaining the
true function. We use this concept here.

The levels of temperature are equally spaced quantitative

levels. Hence one can estimate the interaction of materials by
temperatures by using three degrees of freedom from materials by
linear effect of temperatures.

To obtain an estimate of nonadditivity, the interaction of

materials by linear effect of temperatures may be used. We set up
the orthogonal polynomial coefficients of temperatures for
each material as given in Table 2.1.6. For a discussion of this
procedure see Hicks (1973, p. 138).

TABLE 2.1.6

SS of Materials by Linear Temperaturesa

Temperatures

orthogonal polynomials
(coefficients)

Materials -1 0 1

a 185 182 182

b 175 183 184
c 171 184 189
d 165 191 189
48 2. INTERMEDIATE DATA ANALYSIS

a
SS of materials by linear temperatures:

= ------2----~ 2-------- 2 {[( - D ( 1 8 5 ) + (0) (182) + (1) (1 8 2 )]2

c-ir ♦ ( o r * u r 1

+ ... + [(-1) (165) + 0 (191) + (1) (189)]2)

r-185 + 182 + ... - 165 f 18912__________________

[(-l)2 + (0)2 + (l)2] + ... + [(-l)2 ♦ 02 + (l)2]

■ lt(-3)2 ♦ (9)2 ♦ (18)2 + (24)2] - = 207

The ANOVA for the data from Table 2.1.6 is given in Table 2.1.7.

TABLE 2.1.7

ANOVA Using Materials x Linear Temperatures

p(0.05)
Source df SS MS EMS F
crit

Materials (M) 3 8.67 2.89 a2 + 3<(>(M) <1 9.28

Temperatures (T) 2 354.67 177.33 a2 + 4<(>(T) 6.31 9.55

(Materials) by
(linear temp.) 3 207.00 69.00 a2 + <|>(TM) 2.45 9.28

Error J5 84.33 28.11 a2

Total 11 654.67

Hence there were no significant main effects or interaction in

this case.

2.1.3 Both Factors Quantitative

Referring back to the example presented in Section 2.1.1 we see

that the levels of temperature are equally spaced quamtitative levels.
2.1 ONE OBSERVATION PER CELL 49

In addition, let us now state, the materials which were presented as

a qualitative factor were actually a quantitative factor, this was a
result of each type of material being different as a result of each
containing different percentages of a new alloying element. Further,
these percentages were equally spaced, hence we can again use the
concept of orthogonal polynomials introduced in Section 1.4 to
investigate nonadditivity.

Theoretically if there is one degree of freedom that represents

the nonadditivity or interaction of two quantitative factors, the
most probable one is associated with the linear by linearinteraction.
The basis for this statement is that if the true functionusing
temperatures and materials to predict stress-rupture life could be
written and this function expanded into a power series function in
which six interaction terms of temperature and materials would appear,
the lowest order term, linear x linear, usually would be the largest
contributor.

With this concept in mind, the practitioner usually should

compute the linear by linear term to represent the nonadditive
portion if only one degree of freedom is to be used to represent
interaction. Using only one degree of freedom for nonadditivity is
a common practice because the investigator wants to retain as many
degrees of freedom for error as he can.

Utilizing the same data as was used in Sections 1.2.1 and 1.2.2
we use the orthogonal polynomials in Appendix 11 and refer to Section
1.4 as a reminder of the technique. Here, however, we must consider
the orthogonal polynomials for both variables and cross multiply them
to get the appropriate coefficients for each observation. The
procedure is shown in Table 2.1.8.

The ANOVA using this linear x linear SS estimate of

nonadditivity is given in Table 2.1.9.
50 2. INTERMEDIATE DATA ANALYSIS

TABLE 2.1.8

SS for 1 df for MT Linear x Lineara

Temperatures
Orthogonal
coefficients -1 0 1

-3 185(3) 182(0) 182(-3)

-1 175(1) 183(0) 184(-1)
Materials ^
171 (-1) 184(0) 189(1)
3 165(-3) 191(0) 189(3)

*SS for MT(£ x .) ■ 11IBW * - * W W l ’

= 202.5

TABLE 2.1.9

ANOVA of Data in Table 2.1.7

F(0.0S)
Source df SS MS EMS F
crit

Materials (M) 3 8.67 2.89 a 2 * 3<KM) <1 5.41

Temperature (T) 2 354.67 177.33 a2 + 4$(T) 10.0 5.79

MT(A x A) 1 202.50 202.50 a2 + <KMT) 11.4 6.61

Error J> 88.83 17.77 a2

Total 11 654.67

Therefore both temperatures and the interaction of temperatures

by materials are significant. The next step in the analysis of the
data after this ANOVA would be to examine the interaction carefully.
Since the linear x linear term is significant an exploration of the
response surface dealing with stress rupture life must be made.
Plotting the observations and making some prediction of maximum
life should be planned for the next part of the experiment.
2.2 MORE THAN ONE OBSERVATION PER CELL 51

2.1.4 Summary of Section 2.1

The problem of examining nonadditivity or interaction in a two

way classification with one observation per cell is not straight­
forward. We have presented three methods which cover the use of both
qualitative and quantitative factors. As more information is
available on the levels of the factors (that is as one has more
quantitative information on the factors) the better will be the
estimate of the interaction contribution or nonadditivity. There is
no reason to stop at linear x linear for the orthogonal polynomial
investigation except to keep the number of degrees of freedom for
error reasonably high. One can remove the linear by quadratic,
quadratic by linear and so on to find out what each degree of freedom
provides if he is not satisfied with the variability in y accounted
for by the linear x linear term.

As for the Tukey procedure, it seems to be the best method to

estimate nonadditivity if all factors are qualitative. The linear
of one quantitative factor by the qualitative factor should be used
if there is a mixture of quantitative and qualitative factors.

Problem 2.1.1. Using the following data from a similar

experiment as given in Section 2.1:

Temperatures
Materials 1 2 3
a 169 182 188
b 174 183 190
c 181 188 194
d 184 191 195

(a) Run all the analyses: Tukey, materials x linear on

temperatures, and linear x linear.
(b) Comment on the results from these three analyses.

2.2 TWO-FACTOR EXPERIMENTS WITH MORE THAN ONE OBSERVATION PER CELL

In this section it seems most reasonable to introduce the

52 2. INTERMEDIATE DATA ANALYSIS

concept of random as well as fixed factors and their consequences in

ANOVA. This concept broadens the usefulness of ANOVA to many
problems including multi-factor experiments to be covered in Chapter
4. To show how an experimenter should test for the various main
effects and interactions of these factors we must first derive the
expected mean squares. In this book we will merely show an algorithm
given by Bennett and Franklin (1954) and not actually derive expected
mean squares for certain combinations of random and fixed factors.
Anderson and Bancroft (1952, Chapters 18, 20, 22, and 23) provide a
procedure for those readers who wish to see the mathematical proof.

2.2.1 Expected Mean Square Algorithm

As indicated in Section 1.2, a factor is declared 11fixed" if

all levels of interest to the investigator are included in the
experiment. This concept, then, is that the experiment contains the
population of levels of a fixed factor. For this fixed case the
experimenter is primarily interested in analyzing means.

If, however, fewer than the population of levels of a factor

are to be examined in an experiment and those levels are selected at
random from all possible levels of interest to the investigator, the
factor is called "random." Many times the number of all possible
levels is finite and the sampling rate, or the ratio of the number
of levels of the factor to be included in the experiment divided by
the number of possible levels of interest, is a fraction greater than
zero. This concept is the basis for part of the expected mean squares
algorithm.

Consider the case in which there are Q levels of the factor in

the population and only q < Q are selected at random to represent the
Q levels in the experiment. It follows, Cochran (1963, Chapter 2),
that a finite population correction (1 - q/Q) is the coefficient to
account for the finiteness encountered in the variance estimate.

Observe that this coefficient is equal to zero in the case of

the fixed model and approaches one whenever Q is much larger than q
as is experienced in most applications of the random model.
2.2 MORE THAN ONE OBSERVATION PER CELL 53

In this book we will assume that the finite population

correction is zero if the factor is fixed and one if it is random.
The basis for this choice is a conservative one, for if the
population is truly finite where Q < °°, the variance would actually
be smaller than the one used assuming the population is infinite.
In extreme cases, for example if q/Q = 1/2, extra care must be taken
not to be too conservative and the finite adjustment should be taken
into account. We believe this ultra finite case happens so
infrequently in practice that it is ignored in the remaining part of
this book. The Bennett and Franklin (1954) presentation of the
algorithm utilizes the finite population concept.

The discussion of the expected mean square algorithm is divided

into three subsections; fixed, random, and mixed models. Ostle (1963,
Chapter 11), provides a rather detailed discussion of these models.

Example 2.1; Fixed Model. Consider the case where both factors
are fixed and the model is y.M = p + a. + B. + aB.- + i = 1,
ijk i j lj (ijjk
...,a j = 1, ..., b k = 1, ..., n where
til
y . a variable to be analyzed from the k experimental unit
th th
associated with the i level of factor a and j level
of factor B

y = overall mean
I*l.
ou * effect ofthe i level of factor a (fixed)

Bj = effect of the j level of factor B (fixed)

aB.. = effect of the interaction of the i ^ level of factor

th
a with the j level of factor B

er.... a experimental error of the k ^ experimental unit

th th
associated with the i level of factor a and j level
of factor B, within the ij cell, assume NID (0, o )

Set up the following table:

54 2. INTERMEDIATE DATA ANALYSIS

a b n
F F R
Source
0
0
0 0
1 1 1
£(ij)k

where

(1) The sources of variation include all the variable terms on

the right-hand side of the model together with their associated
subscripts
(2) All subscripts are placed as column headings next to the
source with F for fixed and R for random above the appropriate
subscript and the number of levels; a, b, and n placed above the
appropriate letter and subscripts
(3) Ones are entered in the slots where the bracketed subscripts
in the row headings match the subscripts in the column heading
(4) Zeros are entered in the slots where the row and column
subscripts match for the F type columns and ones for the R type
columns.
(5) Fill in the remaining slots with the number of levels shown
in the respective column heading.

The table then becomes

a b n
F F R
Source
0 b n
a 0 n
0 0 n
1 1 1
E(ij)k

and we are ready to derive the EMS.

Begin at the bottom and multiply the coefficients for e

2.2 MORE THAN ONE OBSERVATION PER CELL 55

The product turns out to be 1. The EMS for this source is then
2 2
l«o » a . Next to determine the EMS for aB interaction, we only
consider those rows which contain at least the same subscripts as
a3..> in this case only z.,. ... and aB... Then cover all columns with
13 (i3)k iJ
headings the same as those contained in the aB row heading, i.e.,
columns i and j. For each rcw under consideration, determine the
product of the remaining coefficients. In this case it is the number
1 for the e r . ... row and n for the aB.• row. These values then
become the coefficient for the corresponding component. The sum of
these terms is then the EMS for aB.

So far the algorithm has produced the following:

a b n
F F R
Source i j k EMS
a. 0 b n
l
a 0 n
63 2
aB 0 0 n a +
1 1 1 a2
6(ij)k

We will use the notation given by Cochran (1951) for fixed components
as <p which is a non-negative function dependent upon the effect
contained within the parentheses. Hence, following the same rules
of allowing only those components that have at least the same subscripts
as the source being investigated to enter the EMS and using the same
column covering rule with multiplication of coefficients we finally
complete the EMS as:

Source i j k EMS
2
a. 0 b n a + bn<J>(a)
l 2
a 0 n a + an<J>(B)
63
0 0 n + n<|>(aB)
2
1 1 1 a
e(ij)k
56 2. INTERMEDIATE DATA ANALYSIS

Example 2.2: Random Model. For this case all factors are
random and we place R!s in all column headings. Using the same rules
as before we obtain the following:
a b n
R R R
Source i P k EMS
2 2
a. 1 b n +
a2 na 0 + bna
l a$ a
a 1 n a2 + naa$ + ana2
6i
a$. . 1 1 n a2 + na2
ij aB
2
1 1 1 a

2
Notice that since all effects are random the a notation
(rather than any <J> for fixed) is used for all components. This
indicates sampling variance and the experimenter's interest is in
estimation of these components of variance as well as testing the
2 2 2
hypotheses that the components, a^, a^ and a ^ are zero.
2
It can easily be seen that the test for 0 requires the e
or within error mean square as the denominator and both main effects
are tested using the interaction (aB^j) mean square. The latter is
quite different from the previous tests for fixed models (the means
are of interest) in which the within error mean square is used as
the denominator for all tests.

Example 2.3: Mixed Model. Let us allow a to be random and B

to be fixed in the mixed model. The expected mean square algorithm
then becomes:
a b n
R F R
Source i i k EMS
2 , 2
a. 1 b n a + bna
l a
2 2
a 0 n a + na . +
aB
2.2 MORE THAN ONE OBSERVATION PER CELL 57

2
In this case <J> is used only for fixed and a is used for both
2
random and mixed (a^ is mixed here) components.

It is interesting to note that for the mixed case the fixed

source, 3, is tested by the interaction and the random source, a, is
tested by the within error mean square. This seems contradictory
because for the fixed model the fixed main effects were tested using
the within error mean square and for the random model they were
tested using the interaction mean square.

This can be explained by considering the appropriate inference

spaces. If both a and 3 are fixed, the inference space is confined
to only those levels a = A and b = B of the two factors and all the
sampling is carried out within each ij cell.

within the cells in a

fixed model.

Hence all tests made within the inference space are based on the
within error.

If both factors are random the inference is over all combinations

of levels of both a and 3. Hence the interaction is the basis for
both tests since intuitively one is required to use a larger mean
square in the denominator because the inference is to a larger
population than in the fixed model case.
58 2. INTERMEDIATE DATA ANALYSIS

I 2 .... j

Experim ent

Inferen ce-----
w here:a « A
Space and b « B

Note: S am pling within cells of the

experim ent is the same as
for the fixed model.

If a is random and 8 is fixed, the inferences for a is only

over the b = B levels of 8 and within the cells, but the inference
for 8 is over the A levels for a as well as the within cells.
Pictorially, it follows that the test for 8 should be against the

0
1 2 ... j b=B
1
2

i E xp erim en t

9 ------------ In fe re n c e ------------------------

Space

A» a
2.2 MORE THAN ONE OBSERVATION PER CELL 59

mixed interaction component since it must be valid for all the A

levels of a. But the test for a should be made using the within
error source only because the inference is only over those levels
of 3 which were included in the experiment.

Example 2.4: Fixed Model (One experimental unit per treatment

combination.) Referring back to Section 2.1 and Table 2.1.3, let us
derive the EMS for Eq. (2.1.1). Using the algorithm and letting
k = 1 for one observation per cell we have Table 2.2.1.

TABLE 2.2.1

ANOVA: Derived EMS for Table 2.1.3

4 3 1
F F R
df_____________ Source______________ i j k________ EMS
3 Ma 0 3 1 a2 + 3*(M)
2 T\ 4 0 1 a2 +
6 MT. . 0 0 1 a2 + <HMT)
0 i l l

We could use this result for the EMS, but since there are no
degrees of freedom for the within error, many authors prefer
dropping this term in the ANOVA. From Table 2.2.1 above it can be
seen that Table 2.1.3 combines the components of both MI\^ and
in the same source, MI\^. This consolidation of effects is necessary
since there are no degrees of freedom for error. Notice, however,
in Tables 2.1.5, 2.1.7 and 2.1.9 that there are separate estimates
of MT and e mean squares each of which has nonzero degrees of freedom.

Example 2.5: Fixed Model (More than one observation per cell).
If in the same example given in Section 2.1, there were two
experimental units per treatment combination, then the algorithm
method of obtaining the EMS would be that presented in Table 2.2.2.

Notice that the two main effects and interaction are tested
using the within error.
60 2. INTERMEDIATE DATA ANALYSIS

TABLE 2.2.2

ANOVA: Two Observations per Cell Extension of Table 2.1.1

4 3 2
F F R
df Source i j k EMS
0 3 2 a + 6<J>(M)
4 0 2 a2 + 8<j>(T)
0 0 2 a2 + 2<KMT)
2
■

1 1 a
—i

Example 2.6: Random Model. Using the same example but assuming
both factors are random we have Table 2.2.3.

TABLE 2.2.3

ANOVA: Random for Table 2.2.2

4 3 2
R R R
df Source i ) k EMS
3 M. 1 3 2 a2 +
l 2° m * “ m
2 4 1 2 a2 +
tj 2ow * *>?
a2 +
£

6 1 1 2
2°S r
•H
•r-»

12 1 1 1 a2
e(ij)k

The df and source are identical to those of Example 2.5, but

the remaining part of the table is different. The algorithm uses
all ones in the slots for matching subscripts in this case and the
results of the EMS are quite different from those of the previous
2
two cases, namely, the interaction component appears in both main
effect EMS*s. For the F-tests on the main effects, then, it follows
that the interaction mean square should be used as the denominator.

Example 2.7: Mixed Model (M fixed, T random). Continuing with

2.2 MORE THAN ONE OBSERVATION PER CELL 61

the same example but letting T be random and M be fixed we get

Table 2.2.4.

TABLE 2.2.4

ANOVA: Mixed for Table 2.2.2

4 3 2
F R R
df Source_____ i j k___________ EMS______
3 0 3 2 a2 + 2a^, + 6<|>(M)

2 T\ 4 1 2 a2 + 8a2

6 0 1 2 a2 + 20 ^,

12 .(..)k 1 1 1 «2

In this case the df and source are identical to both Examples

2.5 and 2.6, but the other parts of the table are different from both
of the previous examples. Referring back to Example 2.3, we can see
that the inference space for temperatures (T^) is only over the four
levels of materials (M^) and the test for T^ requires the mean square
for (the within error) only. On the other hand, it is easily
seen that the test for materials (M^) uses the mean square for
interaction, indicating the inference is broader. In fact the
inference space is over the T » 3 levels of the population of
temperatures from which the three in the experiment were chosen.

For demonstration purposes this example is satisfactory to show

the mechanism of the algorithm and indicate the various F-tests, but
for the practitioner, temperatures are almost always chosen at
definite levels and cannot possibly be random. Explanations to guide
the investigator in deciding whether a factor is random or fixed in a
given situation will be given throughout the book. The decision of
whether a factor is random or fixed is necessarily part of designing
an experiment.

The standard error of a fixed treatment mean when the model is

62 2. INTERMEDIATE DATA ANALYSIS

mixed includes the random components and is quite often impossible

to estimate from the data. In general for those cases in the
completely randomized design in which there is a direct test for the
fixed treatment in the ANOVA, the standard error of the difference of
two means is estimated using the same mean square that was used to
make the test in the ANOVA. This allows one to use the Newman-Keuls
test by using

/
/Mean square used to test the fixed treatment_______W1 2
INumber of observations in each mean of the treatment#

for the standard error of the mean. Refer to Appendix 12 for an

example of derivational techniques.

From now on throughout this book we will place indices on the

symbols under Source in the ANOVA tables only if we believe the
indices are needed to obtain the EMS or if we believe a correspondence
between tables warrants them.

2.2.2 ANOVA of Two-Way Classification with More than One

Observation per Cell

Utilizing the various appropriate techniques described prior to

this section, we now show data from a completely randomized designed
experiment on weld breaks in steel manufacturing. Two factors were
investigated; time of welding (time of automatic weld cycle) and gage
bar setting (distance weld die travels during automatic weld cycle).
Detailed description of weld process is given in Problem 9.1.4. The
measured variable was the breaking strength of the weld, where two
welds were made per treatment combination and the objective was to
find the treatment combination that produced maximum strength. The
coded variable to be analyzed and levels of both factors are given
in Table 2.2.5 where the circled number is the total of the two
observations for that cell.

Since the levels of both factors were selected purposely, the

model is completely fixed.
2.2 MORE THAN ONE OBSERVATION PER CELL 63

TABLE 2.2.5

Gage Bar Setting

1 2 3 Total

1 10,12
© 15,19 © 10, 8 © 74

Time 2 13,17 © 14,12 © 12, 9 © 77

of 3 21,30 © 30,38 @ 10, 5 © 134

Welding 4 18,16 © 15,11 © 14,15 © 89

5 17,21 © 14,12 © 19,11 © 94

Total 175 180 113 468

y. ., = y + T. + G. + TG. . + er. (2.2.1)

ljk l j 12 (ij)k

i = 1, 2, .., 5 j = 1, 2, 3 k « 1, 2 where

th th
y. = coded value of strength of the k weld at the i
1*J time of welding and jth gage bar setting

y = overall mean

T^ = effectof the i ^ time of welding (fixed)

Gj = effectof the j**1 gage bar setting (fixed)

TG^ = effectof the interaction of the i**1 time with

the j**1 gage bar setting
tlii th
er* "\v ~ within error of the k weld in the i time and
th 2
j gage bar setting, NID (0, a )

All assumptions for the ANOVA are assumed met, and the ANOVA
data are given in Table 2.2.6
64 2. • INTERMEDIATE DATA ANALYSIS

TABLE 2.2.6

ANOVA of Data in Table 2.2.5°

5 3 2
F F R
df Source i j k EMS MS F

4 Time (T^ 0 3 2 a2 + 6<MT) 96.4 8.8*

2 Gage (Gj) 5 0 2 a2 + 10<f>(G) 139.3 12.8*

8 T x G (TG..) 0 0 2 a2 + 2if>(TG) 74.6 6.8*

Within
2
15 error (e,. ...) 1 1 1 a 10.9
(ij)k
29 Total

*Significant at the a = 0.05 level

3ms time . SS time _ 1 r742 + 772 + 1342 + 892 + 942 4682

df = 4 - T4 t- 30

96.4

SS gage 1 r175 + 180" + 113 468"-, , ,

MS = df=~2 = 2 [-------10----------- 30~]= 139‘3

we, „ ^ SS Tx G 1 r222 + 302 + ____ + 302 4682

Mo 1 X b = —
df —=— 8
5— = 77 L-------------------
8 L 2 30

- SS time - SS gage] = 74.6

^within
SS 9 9
within _ error = 1 2 2_ 2+ 2_ J S N j
error “ df = 15 ~ 15LJ‘w T
Jerror 9
2 2 J

= 10.9

Of course all these numbers come from Table 2.2.5.

Since there is a significant interaction of time of welding and

gage bar setting, care must be taken to interpret the results. The
2.2 MORE THAN ONE OBSERVATION PER CELL 65

first step for further analysis of the results is usually to plot

the cell means. Table 2.2.7 displays the 15 means identified by the
combinations of the levels of time of welding and gage bar setting.
To obtain Fig. 2.2.1 Table 2.2.7 of means and standard errors (refer
to Section 1.2.1 for notation) is given.

TABLE 2.2.7

Means and Standard Errorsa

Gage bar setting

Time of Welding 1 2 3 Mean

1 11.0 17.0 9.0 12.3

2 15.0 13.0 10.5 12.8
3 25.5 34.0 7.5 22.3
4 17.0 13.0 14.5 14.8
5 19.0 13.0 15.0 15.7

Mean 17.5 18.0 11.3 15.6

a /in Q ~
Standard errors: cell (body) = s- =J — j- = 2.33
Xij*

Gage bar setting = s- = “ 1.00

X.j.

Time of welding = s- = = 1*34

yi--

Overall = s- = = 0.60
X... y 30

The individual comparisons of these means depends upon the

interests of the experimenter. The following subsections show a
few of the possible further analyses that could be carried out on
the means of Fig. 2.2.1 and Table 2.2.7.
66 2. INTERMEDIATE DATA ANALYSIS

Figure 2.2.1

Example 2.8: Both Qualitative Factors. If both factors are

qualitative and there is no preference from a cost viewpoint, quite
frequently the investigator wants to find out the combination of
levels that provides him with the best response. In this case the
interaction is significant and there is crossing over in Fig. 2.2.1.
Hence the experimenter usually would not be very interested in the
main effects even though they are significant (Table 2.2.6). In
fact under the assumption that he is only interested in comparing
the cell means he has really changed his model to

A - 1, 2, ..., 15 (2 . 2 . 2)
m = 1, 2
2.2 MORE THAN ONE OBSERVATION PER CELL 67

where

'ii. ■ I ' m «f <2-2-‘>

y = y of Eq. (2.2.1)

e,n = of Eq. (2.2.1)

(£)m (ij)k
and
1.1.
C0 = effect of the I combination of the i level of
th
time with j level of gage bar setting

In other words the model is really a one way classification

with 15 levels of the combinations of time of welding and gage bar
setting and two observations per cell. Marascuilo and Levin (1970)
point out that if one uses Eq. (2.2.2) as the descriptive model
rather than retaining Eq. (2.2.1) (as most experimenters do), the
analysis of differences between cell means includes both the
interaction effects and the main effects. We believe this is
precisely what many experimenters really want to do.

Using Eq. (2.2.2) we can rank the cell means and use the
Newman-Keuls test as we did in Section 1.2 (See Fig. 2.2.2).

Using Appendix 6 and the s- =2.33 from Table 2.2.7 the

yij*
least significant ranges for this example are:
= (15,15) s- = (5.65)(2.33) 13.2
R1S q0.05 y* .

R14 = q0.05 (14,15) s- = (5.58)(2.33) 13.1

yij-
■‘• j

= (13,15) s- (5.49)(2.33) = 12.9

R13 q0.05
^J
yij‘
= (12,15) = (5.40)(2.33) = 12.7
R12 q0.05
V
= (11,15) = (5.31)(2.33) = 12.4
R11 q0.05
V J
R10 = q0.05 (10,15) yij •
S- = (5.20)(2.33) = 12.2
= s- = (5.08)(2.33) = 11.9
R9 q0.05 ( 9,15) yij •
AJ
= s- (4.94)(2.33) = 11.5
R8 q0.05 ( 8,15)
Xij*
s- = (4.78)(2.33) = 11.2
R7 = q0.05 ( 7,15)
Rank Order the Combinations ( i time , j gage) as follows:
On
I 2 3 4 5 6 7 8 9 10 II 12 13 14 15 00

(i.J) (3^2) (3,1) (5^1) (1,2) (4,1) (2.1) (5.3) (4.3) (2.2) (4 .2 ) (5 .2 ) 0,1) (2 .3 ) (1,3) (3.3)
Mean 34 25.5 19 17 17 15 15 14.5 13 13 13 il 10.5 9 7.5

The Differences of the Means are then

NUMBER OF MEANS SPANNED
INTERMEDIATE DATA ANALYSIS

Figure 2.2.2
2.2 MORE THAN ONE OBSERVATION PER CELL 69

R6 = q0.05
(6,15) s- = (4.60) (2.33) = 10.7
yij-
AJ
(5,15) s- = (4.37) (2.33) = 10.2
RS = q0.05
yy.± •
r
R4 = q0.05
(4,15) s- = (4.08)(2.33) = 9.5
y- •
(3,15) s- (3.67) (2.33) = 8 . 6
R3 = q0.05 =
yir^j
(2,15) s- = (3.01)(2.33) = 7.0
R2 = q0.05 yn.

Hence, from the differences of means triangular table, it is

seen that the only conclusions are the following:
(1) The mean of combination (3,2) > all others
(2) The mean of combination (3,1) > all others except (4,3),
(5,3), (2,1), (4,1), (1,2), and (5,1)
(3) All others are not significantly different at the a = 0.05
level

From a practitionerfs view this means that time of welding of

3 in combination with gage bar setting of 2 would be preferred if
the cost of using this combination was no greater than any other.

If the cost of using time of welding of 3 in combination with

gage bar setting of 1 were less than (3,2), the investigator should
consider combination (3,1) for future production of this type of
steel.

Another analysis that may be considered when both factors are

qualitative and interaction is significant is to look at the means
of the different levels of one factor for just one level of the
other factor. For example, in this case, it may be of interest to
the experimenter to examine the three gage bar settings at only the
5 ^ time of welding. The best estimate of the error variance is
still the within error from Table 2.2.6, namely, 10.9, because the
assumption of homogeneity of variance was accepted indicating the
basic error variance is the same over all these treatment combinations
and this procedure provides us with 15 degrees of freedom for the
error estimate. If only the three levels of gage bar setting for
th
5 level of time of welding were used, only 3 degrees of freedom
would be available for the error estimate.
70 2. INTERMEDIATE DATA ANALYSIS

th
The basic model would now be only appropriate for the 5
level of time of welding:

y = y + G + e, x P - 1, 2, 3 (2,2.3)
pq p (p )q q - 1, 2

where
th
y = variable to analyze the q experimental unit in
the p gage bar setting

y = overall mean for those three levels of gagebar

setting

Gp = effectof the p ^ level of gage bar setting

z, . = within error of the experimental unit in the

(p:)q th 2
p level of gage bar setting NID (0, a ). The
error estimate used for analysis of these three
levels, however, is from all 15 treatment combinations
of the original experiment

Hence run a Newman-Keuls test on these three means as follows:

(1) Refer to Table 2.2.7 to get the means:

Gage bar setting

Means 19 13 15

(2) Rank the means as:

1 3 2
19 15 13

(3) Differences:

Levels

(4) Using Appendix 6 and s- = 2.33 as before

y5p*
2.2 MORE THAN ONE OBSERVATION PER CELL 71

R3 = q0.05 (3'15) Sy = (^*67)(2.33) = 8.6

R2 - q0 05 (2'15) sv = (3.01)(2.33) = 7.0

r5p*

Conclusion:

No significant differences exist at the a = 0.05 level for time

of welding of 5 among the three means for the gage bar settings.

Problem 2.2.1. Compare the means of gage bar settings for

level 3 of time of welding and give a full explanation of the
results. Include the model, the reasoning for the error
variance used and how the investigator can use the results.

Problem 2.2.2. Compare the means of time of welding for level

2 of gage bar setting and give a full explanation of the results.
Include the model and so on as in Problem 2.2.1 above.

Example 2.9: One Factor Qualitative and One Quantitative. When

the interaction is significant in a two-way ANOVA and one factor is
qualitative and the other quantitative, the experimenter may be
interested in the curvature over each level of the qualitative factor.

Let us assume gage bar setting is qualitative and time of

welding is quantitative. If the experimenter wants to perform a
trend analysis for level 2 of gage bar setting he may test for
significance with the error variance from the whole experiment
because homogeneity of variances was accepted. He may also use
orthogonal polynomials assuming the levels of time of welding are
equally spaced. Using this assumption the procedure to investigate
the trend is to use Appendix 11 for n = 5 and set up the following:

Time
i 2 3 4 5

Z1 -2 -1 0 1 2
2 -1 -2 -1 2
Z11
-1 2 0 -2 1
zm
y 34 26 68 26 26

Using the procedure given in Section 1.4.2 to analyze for trend

we obtain the following results:
72 2. INTERMEDIATE DATA ANALYSIS

SS ,
-K»l 1 . 12.8
[(-2r * (-1) + r + 2 ] • (2)

SS : P P 4 ) .-1(26? -2(68) -1(26) . 2 j2 6 )j2 . ,

[2 + (-I)*4 + (-2) + C - i r + 2 ] • (2)

SS ; = L-lCll). + .2(26)_7._2(26) + 1(,26X]2 = 3 2

alll (1 0 ) • (2 )

SS time over gage level 2only

2 2 2 2 2 2
34 + 26 + 6 8 + 26 + 26 (180)
= ----------- -2-------------------- = 664

TABLE 2.2.8

ANOVA for Trend in Time over Gage Level 2

Source df SS MS

Linear 1 12.8

Lack of fit 3, 651.2 217.1*

Quadratic 1 : 165.1
Lack of fit 2 jI 486.1 243.0*
V
Cubic lll) 3.2
Lack of fit 1 1 482.9 482.9*

Error 15 164.0 10.9

*Significant at a = 0.05 level

Conclusion:

The polynomial trend is higher order than cubic. Hence the

experimenter is almost certainly not interested in the polynomial
fit to the data. He may have a theoretical function in mind that
would not be polynomial or he may just run a Newman-Keuls test to
check for the better levels of time for that particular level 2 of
gage.
2.2 MORE THAN ONE OBSERVATION PER CELL 73

Problem 2.2.3. Run a trend analysis on level 3 of gage bar

setting and make conclusions about the results.

Example 2.10: Both Factors Quantitative. The general approach

to investigating the means of the two way table after the ANOVA has
shown a significant interaction when both factors are quantitative
is to generalize the approach given in Section 2.1.3. This means
that the experimenter usually wants to investigate the polynomial
breakdowns of the interaction.

By this we mean for the example on time and gage that the 8
degrees of freedom for interaction may be broken into the following
parts each with 1 df:
1. Time linear x gage linear
2. Time linear x gage quadratic
3. Time quadratic x gage linear
and so on to
8. Time quartic x gage quadratic

From these results the experimenter must decide whether or not

he is interested in this polynomial approximation for the interaction
parts. If he is, he can then introduce the main effect parts also
and investigate the response surface. Usually he will not want to
go to an order much higher than two for the whole inference or factor
space here since the model is completely fixed. He could, however,
discover that a certain part of the factor space is different from
the others and he may want to investigate this part only with a
polynomial approach.

Utilizing the information obtained inExample 2.9along with an

inspection of Fig. 2.2.1, it is suggested that inorder toobtain a
low order polynomial to approximate the response surface represented
by the data of Table 2.2.5, it is necessary to eliminate welding
times 1 and 5. The determination of the optimum conditions will then
be done utilizing the data shown in Table 2.2.9. Note that the
original values of welding times are now T*fs which are coded into
T fs for ease in using orthogonal polynomials. Bear in mind that
74 2. INTERMEDIATE DATA ANALYSIS

the within error mean square of Table 2.2.6 will be used for any
tests of hypothesis. Reference should also be made to Section 1.4
for methods of calculation.

TABLE 2.2.9

Subset of Data from Table 2.2.5a

Time of Welding Gage bar setting

T» T 1 2 3 Total

2 1 13,17 Qo) 14,12 ( I p 12, 9 (n) 77

3 2 21,30 Qp 30,38 (68) 10, 5 O p 134
4 3 18,16 (34) 15,11 (2p 14,15 (29) 89

Total 115 120 65 300

^ h e circle indicates the total of the two observations in that

cell.

The ANOVA for the data of Table 2.2.9 is given in Table 2.2.10
for reference purposes in the derivation of the response surface
approximation.

TABLE 2.2.10

ANOVA of Data in Table 2.2.9

Source df SS MS

Time (T^ 2 301.0 150.5*

Gage (Gj) 2 308.3 154.1*
T x G (TG. .) 4 460.7 115.2*
Within error _9 110.0 12.2
Total 17 1180.0

Within error
frdnuTable 2.2.6 15 164.0 10.9

♦Significant at a = 0.05 level.

2.2 MORE THAN ONE OBSERVATION PER CELL 75

A brief review of the MS values in Table 2.2.10 indicates that both

main effects as well as the interaction are contributing
significantly to the variation in the dependent variable. Consequently,
to determine the response surface equation, one must determine the
importance of the linear and quadratic terms for each main effect as
well as the significant portions of the interaction effect.

The procedures used in Section 1.4.2 may be used to calculate

the sum of squares and the estimate of the appropriate regression
coefficient for each of the main effects. As a check on calculations
it will be noted that the linear and quadratic sums of squares add
to the sum of squares given in Table 2.2.10 for the corresponding
factor. This type of check may always be applied whenever all factors
are equally spaced with equal number of observations per cell as
long as one computes as many sums of squares for linear, quadratic,
cubic, and etc., as there are degrees of freedom for that particular
factor.

The T x G interaction sum of squares may also be partitioned

into four independent pieces with one degree of freedom each. These
four pieces are denoted as Tlinear x GUnear (T* x G^), Tlinear x

^quadratic x Gq^ * ^quadratic x G1inear ^ q x GflP9 ^quadratic

x Gquadratic ^Tq x Gq^ * Tlie calculati°n f°r t*ie sum squares due
to T^ x is similar to that shown in Table 2.1.8 for MT (& x A).
The calculation for SS T. x G and the estimate of the coefficient
i q
of the corresponding orthogonal polynomial is given in Table 2.2.11.
The label given to this polynomial is Z^ where the single
subscript 1 indicates linear for the first factor Time of Welding (T)
and the double subscript 22 indicates quadratic for the second factor
gage bar setting (G). The symbol for the coefficient is <*^22' T^ie
sums of squares and the estimate for all coefficients are given in
Table 2.2.12.
76 2. INTERMEDIATE DATA ANALYSIS

TABLE 2.2.11

SS for 1 df T. x G
I q
G
Orthogonal q
coefficients 1 -2 1
-1 30 (-1) 26 ( 2) 21 (-1)
0 51 ( 0) 68 (0) 15 ( 0)
1 34 ( 1) 26 (-2) 29 ( 1)

CO T _ r _ [30C-1) + S1(0) +...+ 29(1)]2 _ 144

& 1{. X n " T
q 2[(-l) + (0)2 +...+ (l)2] 24

= 6.0

The leading 2 in the denominator of SS T0 x G

x, q
represents the number of observations in each cell.

: r 3 0 ( - l) + 51(0) +...+ 29(1)1 .12 a r

122 2 [C-1) 2 + (0)2 + ...+ (1) 2] 24

TABLE 2.2.12

SS and Coefficient Estimates for All Possible Polynomials

Factor Polynomial SS Coefficient Coefficient estimate

12.0 1.00
T* zi al
289.0 -2.83
Tq zn all
301.0

208.3 -4.17
ga Z2 a2
I CM
Nl

G 100.0 -1.67
CM

q a 22
308.3

2.0 0.50
t ji x G* Z1 Z2 a12
6.0 0.50
T* x G q Z1 Z22 a 122
140.2 2.42
Tq x G * Z11 Z2 a112
T x G 312.5 2.08
q q Z11 Z22 a 1122
460.7
2.2 MORE THAN ONE OBSERVATION PER CELL 77

Each of the terms in Table 2.2.12 has one degree of freedom so

the SS is also equal to the mean square. Comparing these against
the error mean square of 10.9 indicates the terms that should be
included in the regression model. The linear term for time is
included for completeness only. The resulting orthogonal model is

y = ao ♦ a1Z1 ♦ Z2 ♦ an 2 2 Zn Z22

where the various Z values are as defined in Section 1.4.1. For

example,

zn z 2 = x2 t(T ■ T) 2 ■ ^ n ] xi ( G " G)

= 3 t(T - 2)2 - |] (G - 2)

Substitution of all theseterms and theirrespective coefficients

into the above equation results in the following equation:

y = 138 - 157T - 210 G + 39.5 T2 + 57.5 G2 + 271 GT - 67.75 GT2

- 75 T G2 + 18.75 T2 G2

The determination of the point in the factor space which yield

the maximum value of y in the previous equation is rather difficult
without the use of a computer. One may attempt to find the optimum
operating condition by trial and error or resort to some elaborate
optimization procedures. In either case one would determine that
T = 2.023 and G = 1.746 is an estimate of the optimum operating
conditions which produces an estimated response value of 35.17.
Note that this value does not differ very much from the average
value of 34 observed in the cell T = 2 and G = 2.

Problem 2.2.4. An animal feed manufacturing company was

interested in investigating the influences of five different
amounts of a certain feed supplement and three temperatures at
which the supplements were put into the feed on the amount of
vitamin A retained by the feed. The data were transformed so
that the assumptions in the ANOVA were met.
From a completely randomized design with two randomly
chosen feed samples per treatment combinations the following
data are to be analyzed:
78 2. INTERMEDIATE DATA ANALYSIS

Temperatures

Supplements 40 80 120

2 33,37 60,64 0 46,50 0

4 50,56 U0a 79,73 U53 72,58 U3TO
6 72,64 Q3d 80,88 Q6a 72,80 H52)
8 77,73 Q5d 79,87 81,77
10 67,73 K * y 84,80 69,83

(a) Run the appropriate ANOVA, using the algorithm to show

the correct tests, and make the tests of significance on the
main effects and interaction.
(b) Determine the regression equation which is the best fit
for these data.

2.3 REFERENCES

Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research

McGraw-Hill, New York, 1952.
Bennett, C. A. and Franklin, N. L. Statistical Analysis in Chemistry
and the Chemical Industry, Wiley, New York, 1954.
Cochran, W. G. Sampling Techniques, 2nd ed., Wiley, New York, 1963.
Cochran, W. G. Biometrics 7:17 (1951).
Hadley, W. 0., Hudson, W. R., Kennedy, T. W. and Anderson, V. L.
Association of Asphalt Paving Technology 1:1 (Feb. 1969).
Hicks, C. R. Fundamentals in the Design of Experiments, 2nd ed.,
Holt, Rinehart and Winston, New York, 1973.
Marascuilo, L. A. and Levin, J. R. Am. Ed. Res. J. 7:397 (1970).
Ostle, B. Statistics in Research, 2nd ed., Iowa State University
Press, Ames, Iowa, 1963.
Tukey, J. W. Biometrics 5:(232 (1949).
 Chapter 3

A SCIENTIFIC APPROACH TO EXPERIMENTATION

Experiments are performed by people in nearly all walks of life.

The basic reason for running most experiments is to find out something
that is not known. Unfortunately, there are cases in which the sole
purpose of the experiment is to "prove11 what the experimenter already
"knew." This type of experiment frequently is conducted so that the
"known" result will occur no matter whether it should or not. This
type of experiment cannot be condoned by persons seeking the truth.
On the other hand, a worse condition may exist where people run
experiments fully intending to be honest but being completely unaware
of their incompetency in conducting experiments intelligently.
Frequently experiments are run so that the effect of the factor of
interest is disguised by the effect of another factor not considered.
This latter factor is then ignored or considered unimportant; yet,
in the long run, it is the real cause.

An example of this occurred a number of years ago in a large

manufacturing company. A man working in the production area set up
an experiment to test a new alloy, possibly one to replace an old
one in production. He ran one heat of metal with the new alloy and
another heat with the old one. Taking one ingot from each heat and
30 pieces of metal from each ingot, he proceeded to test each of the
60 pieces for the property in which he was interested. With the data
he made a one-way ANOVA on the alloys using the pieces within ingots
with 58 degrees of freedom as the error. The results showed that the
new alloy was "better" than the old one, and the experimenter
convinced the vice president in charge of production to change the

79
80 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

production procedures so that the new alloy would be used in the

future. Since the experimenter had used a "designed experiment11 and
had tested the data ’’statistically,” the vice president concluded
there could be no doubt that the new one was better.

The change cost the company $200,000, and after two years in the
field there was as much trouble with the product made from the new
alloy as there had been with the old product. The vice president was
disgusted and called one of the authors of this book to say he would
never allow his company to use designed experiments again. After
some discussion, the vice president allowed the author to talk with
the experimenter to find out how the experiment was conducted.

In wanting to keep the cost of the experiment low, the

experimenter did not consider the possibility that the property in
which he was interested varied considerably both from heat to heat
and from ingot to ingot within a heat. From an ANOVA point of view,
his EMS should have been what is shown in Table 3.1.1.

TABLE 3.1.1

ANOVA of Alloy Problem

Source df EMS

Alloys 1 Op + 30a^ + 30a^ + 30<J>(A)

2
Pieces in ingots 58 a
P

Hence, rather than testing that the alloy effect was zero
[<f>(A) = 0], he was really testing that the total effect for ingots,
2 2
heats, and alloy was zero [300^ + 30cfh + 30<J>(A) = 0]. Since the
long-run production of the new alloy did not produce the improvement
seen in the experiment, <KA) must equal zero. Thus, it must have
2 2
been that a j. and/or a were/was not zero.

When the results were explained, the vice president was willing
to use design of experiments again on this type of problem but
insisted upon having more than one heat for each alloy.
3.1 RECOGNITION THAT A PROBLEM EXISTS 81

The conclusion is that if experiments are to be run, and

assuredly they will be, a scientific approach to designing them is
needed. The following is an ordered list of requirements for
scientific experimentation. The requirements will be described in
more detail in the following sections.
Recognition that a problem exists.
Formulation of the problem.
Agreeing on factors and levels to be used in the experiment.
Specifying the variables to be measured.
Definition of the inference space for the problem.
Random selection of the experimental units.
Assignment of treatments to the experimental units.
Outline of the analysis corresponding to the design before the
data are taken.
Collection of the data.
Analysis of the data.
Conclusions.
Implementation.

3.1 RECOGNITION THAT A PROBLEM EXISTS

In many industrial plants the accountants are able to show which

projects are possibly having production trouble by arraying the costs
in the various departments (sometimes called a "pareto"). The
indication is of course, that the relative costs in these areas
either have changed recently or are large relative to some production
criterion.

In other plants the foreman may recognize trouble and bring it

to the attention of his superior. In any event, at least one area
always seems to need attention, and alert management recognizes that
a problem exists.

In research departments of universities or in industry, people

are continually encountering problems that can be attacked success­
fully by using scientific experimentation. Usually the recognition
phase in these cases is short.
82 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

3.2 FORMULATION OF THE PROBLEM

Committee action is the most successful method of formulating

the problem after management agrees that a problem exists. The word
"action" is most important because the members of a committee for
attacking a problem must be ones who know the technical parts involved
and be willing to state their thoughts. A person who knows the
requirements for scientific experimentation and keeps the thoughts
flowing is the best moderator for such a group. There must never be
reticence by any participants even if the ideas are in conflict with
those of a supervisor who may be present; hence, a moderator from
outside the plant who stimulates thought and prevents conflicts is
usually preferred to a knowledgeable in-plant participant.

Usually this discussion period lasts from 2 to 8 hours; every

committee member has a chance of expressing his view. Frequently
45 or 50 possible causes of the problem will be put forth. The next
step for the moderator is to get the participants to reduce these to
a reasonable number of causes. Hopefully, the list will be reduced
to eight or ten, or preferably four or five. Rarely is the ultimate
list only two or three. Of course, the greatest difficulty in this
actual formulation of the problem is to get the committee members to
agree on the ultimate causes to be tried in the forthcoming experiment.

An example is an investigation related to the fabrication of

men's synthetic felt hats. The manufacturer had experienced extreme
difficulty in producing these hats so that the flocking appeared on
the molded rubber base in a uniform fashion to simulate the real felt
hats. In order to approach this problem, a committee was formed
consisting of a development engineer, a manufacturing foreman, a
chief operator, a sales representative and a statistician. The
statistician's job was to obtain from these people all possible
causes of imperfect hats. Factors which were thrown out for
discussion were as follows: thickness of foam rubber base, pressure
of molding, time of molding, viscosity of the latex used to glue the
flocking to the molded rubber base, age of the latex, nozzle size of
3.3 FACTORS AND LEVELS IN AN EXPERIMENT 83

several different spray guns, direction of spraying, condition of the

flocking, speed of drying, and the effect of location within the
drying furnace.

After considerable discussion, the committee finally decided

that the most serious problems were probably connected with the
nozzle size and the pressure under which the latex was sprayed. In
arriving at these various factors the committee essentially forced
a review of the entire production process. This in itself led to a
better understanding of the production process and an eventual
solution to the problem.

3.3 AGREEING ON FACTORS AND LEVELS TO BE USED IN THE EXPERIMENT

In arriving at the most reasonable causes of defective hats in

the previous section, the committee action essentially required a
critical review of all production process factors. In this review
the chief operator brought out the standard operating levels of the
nozzle size as well as the pressure under which the latex was
sprayed. Talk with the chief operator revealed that the latex
pressure varied considerably due to the viscosity of the latex and
from this information the pressure levels were eventually obtained.
Additional inquiry ascertained that the manufacturing area had two
different nozzle sizes that had been used interchangeably;
consequently, the two sizes became the basis for these factor levels.

The authors feel that through committee action of this type one
is almost always able to find realistic levels for all major factors.
Occasionally considerable effort is needed to find out just how
shoddy one's manufacturing operation really is, and this example
shows that actual production operations will allow their process to
operate at various levels as long as it works. The determination
of the optimal levels is, of course, the desired end of the
experimental investigation.
84 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

3.4 SPECIFYING THE VARIABLES TO BE MEASURED

As one can imagine in the manufacturing of synthetic hats the

measuring of product quality is a very difficult task. Consequently,
the method that was used in the experiment referred to above was to
visually grade the finished hat on the following items: the hungry
appearance of the flocking, the starchy appearance of the flocking,
and the appearance of the brim. During the course of the
investigation these responses were found to be essentially independent
of each other and consequently could be treated as three separate
dependent variables, which could be investigated on a one-at-a-time
basis. The standards for grading each of these factors were arrived
at again through committee action which eventually resulted in a
visual display board and gave the inspectors a realistic means of
grading each of these dependent variables.

One of the most difficult things in certain types of industrial

experimentation is the specification of a dependent variable. It is
usually quite obvious what the variable should be; however, the means
of measurement is sometimes quite difficult. In ideal cases this
would be merely the measured value obtained by some simple inspection
tool while in other cases it can be a value that is almost impossible
to measure and will have to be graded by one or more inspectors.

3.5 DEFINITION OF THE INFERENCE SPACE FOR THE PROBLEM

After the experimenter has defined his problem he must make

decisions on the limits of the inferences to be made from the results.
Those limits within which the results will apply may be called the
inference space. For example, a chemist may want the results of his
experimentation to apply to all laboratories in the world. If this
is the case, the experiment should be conducted in such a way that a
sufficiently large random sample of the possible laboratories should
be included in the experiment. Great care should be placed at the
initial stages of formulating the experiment to assure the wide
3.6 RANDOM SELECTION OF THE EXPERIMENTAL UNITS 85

applicability desired by the experimenter (Cox, 1958, Section 1.2,

pp. 9-11).

In genetic studies using Drosophila (fruit flies), the variation

between bottles containing the flies is usually larger than the
variation within bottles and the inference space is usually over
bottles. Hence for certain simple laboratory experiments the desired
experimental error must contain between bottle variation as well as
variation from flies within bottles in order to make the results
apply to all bottles. The general ideas involving the correct error
terms have direct association with the inference space for any factor
or combinations of factors (or treatments).

3.6 RANDOM SELECTION OF THE EXPERIMENTAL UNITS

Throughout this book the term "experimental unit" will refer to

the type of experimental material that is being used to receive the
application of the various treatments and be representative of the
desired inference space. For example, if one is interested in
investigating the effects of various types of fertilizers on potted
plants, the potted plant would be the experimental unit. In this
case extreme care should be taken before extending the inference to
field conditions. If one is interested in investigating the effect
of a certain type of teaching device then the experimental unit
would be an individual student. In agronomy experiments, many times
one uses a plot of ground as an experimental unit. In this case the
size of the plot is frequently dependent upon the growing habits of
the particular plant under consideration because the results must
apply to the desired population.

Once the experimental unit has been selected, the investigator

usually has several of the units at his disposal. Consequently it
becomes necessary to randomly select a sufficient number of
experimental units to be utilized in the experiment. The random
selection is necessary at this point in order to protect against any
86 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

bias in our experiment which could be the result of some unknown

factor having had prior influence on the experimental units in some
systematic fashion. The number of experimental units is naturally
dependent upon the size of the desired standard error of individual
cell means. In certain investigations the available experimental
units will not be of the same homogenous material. Consequently,
an investigator will suspect that he is apt to get a different
response for his treatments as a result of these experimental units
being stratified together based on certain inherent characteristics.
In agronomy experiments these inherent conditions could be soil
conditions. In Section 3.5 we discussed a genetic study using
Drosophila. If the individual fruit fly is the experimental unit,
then as discussed before these fruit flies will be stratified
according to their respective bottle. This type of stratification
will be referred to as blocking in the forth coming chapters.

3.7 ASSIGNMENT OF TREATMENTS TO THE EXPERIMENTAL UNITS

Experimenters want assurance that the probability statements

they make are as nearly correct as possible and to give them this
assurance, laws of chance must enter into the experimentation. The
research worker wants each treatment to have as good a chance as any
other treatment to show its effect on the experimental unit through
the variable to be analyzed. The idea of complete randomization is
to allow every experimental unit an equal chance of receiving each
treatment. For example, in an electroplating experiment, magnetic
material was electroplated to material with two different substrates,
copper and brass. The variable to be analyzed was thickness of
plating after a given time of current flow. The design of the
experiment or randomization was associated with the order of the
test pieces with the different substrates (manufacturing was assumed
to have been handled correctly just as other extraneous variables).
If all material with copper substrates were plated before any of the
brass substrates, there may be thicker plating just because of the
3.8 OUTLINE ANALYSIS BEFORE DATA ARE TAKEN 87

environmental conditions at that time, not because of the metal of

the substrates. Hence, to insure against unknown peculiarities, the
order of the plating should be randomized.

As a result of randomization, the estimates of the parameters

such as average thickness of plating should be unbiased and the tests
of significance of the effects of treatments on the variable to be
analyzed should be valid (Fisher, 1966, pp. 62-64). In addition, the
manner in which randomization occurs (for example completely at random
or inside blocks of experimental units), determines the analysis to
be used on data (Ostle, 1963, p. 251). Randomization is then the
basis for the design of the experiment, but it never eliminates
variation caused by extraneous variables. References are Yates
(1964, pp. 317-321), and Finney (1960, Chapter 1).

The definition of design of experiments that will be used in

this book is the method by which the treatments are placed on the
experimental units. Many times due to the nature of the experiment,
restrictions on the randomization are inherent (sometimes rather
cleverly disguised) in the design. This was the case in the
synthetic hat example introduced in Section 3.2. In this case the
committee decided that it would not be economically feasible in terms
of time to interchange spray gun nozzles in a completely randomized
fashion. Consequently, when a certain configuration of nozzles was
set up several treatment combinations were run by adjusting various
gun pressures in a random fashion. Restrictions on the randomization
as discussed above gives rise to the term ’’restriction error” which
must be acknowledged in the analysis of the resulting data. This
acknowledgement is accomplished by inserting an appropriate term into
the model (Anderson, 1970).

3.8 OUTLINE OF THE ANALYSIS CORRESPONDING TO THE

DESIGN BEFORE THE DATA ARE TAKEN

The concept of committee action was introduced in Section 3.2.

We now come to another critical step in this committee action. At
88 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

this point, the statistician must write down the mathematical model
that has evolved as a result of the committee activity in the
preceding sections. This mathematical model will give rise to the
ANOVA table. The ANOVA table will now consist of the degrees of
freedom and the expected mean squares for each of the specific
factors selected in Section 3.3 above. These expected mean squares
will indicate to the statistician and experimenter all of the various
factors which will have tests available. In addition, if the test
is not directly available, it will be clear as to what assumptions
must be made in order to obtain some type of conservative test. Once
the statistician has reviewed this ANOVA table he must then present
his findings to the committee to see if these assumptions and
conclusions are realistic from a practical point of view. Should
the committee decide that some of the necessary assumptions (which
must be made in order to test the effects of various factors) are not
realistic, the design will have to be changed before proceeding
further in the experimentation process.

As one investigates various designs by comparing expected mean

squares from the respective analysis of variance tables, it will
become obvious to the statistician as well as to the experimenter
when they have to compromise (for economical reasons) the
availability of a clear-cut test on the effect of a certain factor.
It is possible that by a certain type of restriction on the
randomization, which would lead to a conservative-type test, that
one would be able to run an experiment in a much more economical
fashion. One must recall, however, that there is economics involved
in the cost of the experiment as well as in the cost of making an
incorrect decision as a result of running an experiment which does
not give an explicit test for the effect of an important factor.
These costs, of course, would arise when an incorrect manufacturing
process is implemented on a production basis.
3.9 COLLECTION OF THE DATA 89

3.9 COLLECTION OF THE DATA

To the inexperienced investigator this particular topic seems

like a rather simple and straightforward procedure. However,
extreme care must be taken at this particular point. In some cases,
carrying out of the experiment may require an organized team with
many observers. Team planning meetings may become quite extensive
and detailed instructions may have to be prepared for operators,
supervisors, and technical observers. The whole success of this
scientific investigation depends upon the validity of all data
obtained. Even if the experiment is not too complicated, the authors
have found that one of the most successful means of collecting data
is for the statistician to make a specific data form which would be
filled out by the inspector and further verified by the person in
charge of running the experiment, possibly the manufacturing foreman.
The data form itself must be easy to use, relatively understandable,
and in general foolproof with respect to any misinterpretations.
For a complicated experiment it is sometimes wise to pretest the
experimental plan and the data collection form. By this we mean to
actually process some experimental units that are designed
specifically for pretesting the inspection and data recording
procedures.

Considerable thought should be given to the collection of the

data. For example, if the data are going to be keypunched for use
in a particular computer program it is well known that keypunch
errors can be reduced by having an efficient data form. In case one
is collecting extreme amounts of data, possibly the data can be
obtained on either magnetic tapes, paper tape, punch cards, or even
possibly some type of mark sense cards. In general, the size and
the importance of this experiment will dictate the amount of care
that should be taken in the collection of the data.
90 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

3.10 ANALYSIS OF THE DATA

The initial phases of the analysis of data will be primarily

the statistician's responsibility. Naturally the analysis of the
data is highly dependent upon the design of the experiment. Modem
computing facilities have many so-called canned programs which will
analyze almost every type of design that is available. The primary
input to these canned programs is the experimental model that we
have discussed previously. Once the model has been written down
and the number of levels decided upon, it is a relatively simple
procedure to process the data through the computer. If a computer
program is not available then one must resort to the use of a desk
calculator or possibly even a slide rule and an adding machine.
Formulas for these calculations may be readily found in such
references as Hicks (1973), Ostle (1963), and Steel and Torrie
(1960).

Once the basic calculations have been performed the statistician

should review the data and possibly make graphs of the individual
cell means in order to display the results of this experiment. The
authors have found that graphical presentation of the data is the
easiest way to show management the important aspects of the experiment.

3.11 CONCLUSIONS

Once the statistician has completed the analysis of data it is

again time for various members of the committee to meet and formalize
their conclusions. In almost every case the experimenter will be
able to physically interpret the statistical findings that have been
brought to light by this experiment. Then based on the magnitude of
these findings and the cost of implementing any new procedures which
might be required, the committee must make recommendations to
management. These recommendations are usually in some form of
written report and/or oral presentation to the interested members of
management.
3.12 IMPLEMENTATION 91

3.12 IMPLEMENTATION

The implementation of the recommendations developed by the

committee is of course a management decision. The purpose of the
scientific investigation was to develop a realistic type of
experiment, assure proper running of the experiment and to find
dependable results. These findings of course have to be presented
in such a fashion that management can make an intelligent decision
upon the implementation of these findings.

Once management decides to make a process change which has been

recommended by the committee, the committee has another tremendous
task facing it. This task is to assure management that the
experimental procedure will be implemented as well as possible into
the production process. In many cases it is not a simple thing to
implement a procedure developed in a pilot study into a full-fledged
production process. In no case should a new procedure be put into
production without further evaluation by some type of statistical
study. It is the responsibility of the committee to assure management
that the revisions to the process have indeed improved the production
operation. Many times a prototype of the actual production set up
is devised and tried out on a small-scale production basis. This
gives an indication of how the results of the experiment may work in
production without using actual production equipment as a testing
device.

If results from the prototype are successful then the results

from the experiment may be put into production; however, many times
changes are indicated by the prototype before implementation.

3.13 SUMMARY

It should be pointed out that the actual design of experiments

part of the scientific approach to experimentation is covered in
Sections 3.5, 3.6, and 3.7. The sections prior to these three
prepare for the design, and Section 3.8 allows the experimenter to
modify the design before putting it into operation.
92 3. A SCIENTIFIC APPROACH TO EXPERIMENTATION

Since the following chapters will be concerned mainly with the

design of experiments we will summarize the topics covered in
Sections 3.5, 3.6, and 3.7.

1. Inference Space

Before the experiment is really designed the investigator must

decide how widely he wants his results to apply. The experimenter
must compare the cost of the experiment with the usefulness of the
results he intended to obtain. This demands that he thoroughly
understand what inferences he can make before he actually undertakes
the experiment. This stage, in practice, is the most neglected
because many people do not take the time to think the problem through
before setting up the design.

2. Randomization

This stage is necessary in order for the experimenter to

construct probability statements about the results. Randomization
allows for valid tests and unbiased estimates after the inference
space has been decided upon.

3. Replication

Before the design is finalized, the experimenter must decide on

the amount of replication to use in the experiment. Present-day
design practices utilize existing information to the fullest and
replication is sometimes not needed at all. In fact fractional
replication (which will be covered later) actually means there are
less than a complete representation of all the treatment combinations.
In industry, usually there will be pressure from management (in
universities, this pressure comes from the major professor or
investigator) to run the experiment as quickly as possible with a
minimum amount of expense. To protect the validity of the experiment,
the statistician must make it clear to the investigator what minimum
replication is demanded to obtain the estimates of the effects the
experimenter desires with an adequate error. This decision quite
3.14 REFERENCES 93

frequently is based on the number of degrees of freedom that will be

available for the error to test the various effects of interest.

The remainder of this book really utilizes these concepts in

different arrangements of the experimental units and systematizes
these layouts into the so called "designs.” In all instances the
expected mean square derivation (the analysis) is immediately
available after the design has been made and shows the experimenter
what he can cover in his inferences. The results of the design
will not be realized until after the data are analyzed.

3.14 REFERENCES

Anderson, V. L. Biometrics 26:255 (1970).

Cox, D. R. Planning of Experiments. Wiley, New York, 1958,
Chapters 1, 4, 5, and 8 .
Finney, D. J. An Introduction to the Theory of Experimental
Designs, University of Chicago Press, Chicago, 111., 1960,
Chapters 1, 8 and 9.
Fisher, R. A. The Design of Experiments. 8 th ed., Hafner, New York,
1966, Chapters I-IV.
Hicks, C. R. Fundamental Concepts in the Design of Experiments,
2nd ed., Holt, Rinehart and Winston, New York, 1973, Chapter 1.
Ostle, B. Statistics In Research, 2nd ed., Iowa State University
Press, Ames, Iowa, 1963, Chapter 10.
Steel, R. G. D. and Torrie, J. H. Principles and Procedures of
Statistics, McGraw-Hill, New York, 1960, Chapter 6 .
Yates, F. Biometrics 20:307 (1964).
 Chapter 4

COMPLETELY RANDOMIZED DESIGN (CRD)

Determination of the proper design of the experiment to be used

by a research worker depends upon the inference he desires and the
randomization procedures available to him. If the experimenter wants
to investigate the effects of a factor in a narrow area, say within
his laboratory, the design may require only simple randomization and
the completely randomized design is appropriate. If, on the other
hand, he wishes the results to be applicable over an extremely wide
range, say over the entire United States, the design may need to be
quite elaborate and the completely randomized design may not be
appropriate. References on these concepts include Cox (1958, Chapter
7), Cochran and Cox (1957, Chapter 4), and John (1971, Chapter 3).

4.1 ONE FACTOR

This is the simplest design and is characterized by the

experimental units being drawn as completely at random as possible.
One procedure for constructing a completely randomized design is to
allot numbers to the experimental units and draw the numbers at
random assigning the levels of the factor or treatments (which may
be random or fixed) in order of the draw. For example, in an
experiment on firing time of an explosive switch, let us assume that
packing pressure of the primary initiator is the only factor involved.
If there are three pressures (12,000 psi; 20,000 psi; 28,000 psi) to
be investigated, it would be desirable to show the design and
analysis of this experiment if there are 21 explosive switches
available and no restrictions on randomization in the experiment.

94
4.1 ONE FACTOR 95

Let us look at the inference space. The experimenter would

like to imply that if the effects of packing pressures show up in
this experiment, he can be reasonably sure that the same effects
will occur on all explosive switches from which these 21 were a
sample. This requires that these 21 switches be selected from the
population of interest, at random, even before the experiment is
begun and then numbered 1 through 21 (Hader, 1973).

The next step is to be sure these 21 switches are assigned at

random to the packing pressures in such a manner that there is an
equal number (in this case 7) of switches for each packing pressure.
To do this we can use a set of random numbers (Appendix 1) and draw
a number between 1 and 21, say 10. Then the number 10 switch is
assigned to packing pressure 12,000 psi and an explosive switch is
constructed with 12,000 psi. This procedure is continued so that
the next explosive switch drawn at random will be constructed with
20.000 psi, the following one with 28,000 psi, the one after that
with 12,000 psi and so on. This will finally give a design with
seven switches with 12,000 psi, seven with 20,000 psi and seven with
28.000 psi. The only restriction on randomization (which is minor
and ignored in the analysis) is that if the same number, say 10 ,
were drawn the second time it would be disregarded. In other words,
there is only one switch for each number. The sampling is really
without replacement but the correction for this in the analysis is
ignored. One reason for ignoring the sampling without replacement
is that the inference is over all switches from which these 21 are
drawn.

In some experiments only one randomization takes place, but

here one must consider the firing order of the constructed explosive
switches. If, by chance, there was some reason for the switches
fired early to have a longer firing time due to atmospheric or
environmental conditions and all the 12,000 psi were fired first,
it would not be possible to distinguish between the effect of packing
pressures and the effect of the environment. This demonstrates that
the effect of packing pressure would then be completely confounded
96 4. COMPLETELY RANDOMIZED DESIGN

with the environmental effect. Hence, to assure the experimenter

that the results are as near correct as he can make them, he must
randomize the firing order. In this case, he can renumber the 21
explosive switches as

Packing Pressure

12,000 psi_____ 2 0 , 0 0 0 psi_____ 28,000 psi

1 8 15
2 9 16
3 10 17
4 11 18
5 12 19
6 13 20

7 14 21

and draw a random number between 1 and 21 again. Say the number is
12 , then.the number 12 explosive switch (2 0 , 0 0 0 psi pressure) is
fired first. Continue in this manner until all 21 switches have an
order number to be fired. All of this randomization must be
completed before the firing is begun so the experiment can be carried
off smoothly. One possible firing order is

Firing order_____ Switch number_____ Pressure

1 10 20,000

2 2 12,000

3 11 20,000

4 20 28,000
5 15 28,000
6 3 12,000

7 9 20,000

8 17 28,000

21 8 2 0 ,0 0 0
4.1 ONE FACTOR 97

Notice the raixup in the order of the column headed "Pressure." Of

course order could be a classified variable and controlled like a
factor but then the experiment would not be run in a completely
randomized design (Prairie and Zimmer, 1964).

The model for the analysis of variance of this experiment is

yij = U + Pi + e(i)j = [i = 1» 2, 3; j = 1, 2, 7] (4.1.1)

where

y. . = observed firing time of the j**1 explosive switch with

the i packing pressure (assumed normally distributed
and used as the variable to be analyzed)

\i = overall mean
th
P^ = effect of the i packing pressure

£(i)j = e^ ect t^ie sw;*-tch within the i**1 pressure

assumed to be NID (0, a^)

The corresponding ANOVA for Eq. 4.1.1 is given in Table 4.1.1.

TABLE 4.1.1

ANOVA for Pressures

Source df EMS
2
Pressures 2 a + 7<|>(P)
2
Residual JJ* a
Total 20

The estimated standard error of each pressure mean is

Problem 4.1.1. Show how you would examine the trend of firing
time over packing pressures.
98 4. COMPLETELY RANDOMIZED DESIGN

4.2 MORE THAN ONE FACTOR

When more than one factor is used in an experiment and all

levels of each factor are represented in combination with all levels
of every other factor, the experiment is called a "complete
factorial." This property (complete factorial) allows the factorial
experiment to have inherent or hidden replication because any level
of one factor occurs as often as the product of the number of levels
in the other factors. Since, in reality, a given combination of
levels of all the factors acts as a treatment on an experimental
unit, these combinations are frequently called "treatment
combinations."

In this chapter on completely randomized designs it will be

assumed that all treatment combinations are placed completely at
random with the experimental units (regardless of the number of
factors), In addition, if there is more than one experimental unit
for each treatment combination, the designs of the experiments for
combinations of fixed and random factors are easily extended from
the one factor completely randomized design given in Section 4.1.
In this section designs with one experimental unit for each treatment
combination will be emphasized because this condition requires more
design techniques (to get the design efficient) than is required for
more than one experimental unit per treatment combination.

It should be understood that the analysis of a complete factorial

with two factors whose combination of levels have been completely
randomized is sometimes called a two-way classification, and with
q-factors is a q-way classification.

Problem 4.2.1. Show a completely randomized design and analysis

for a two factor factorial experiment when the model is mixed
if there is more than one experimental unit per treatment
combination.

4.2.1 All Fixed Factors

If all factors in a completely randomized design are fixed and

4.2 MORE THAN ONE FACTOR 99

there is only one experimental unit per treatment combination, there

may be no good estimate of the experimental error for investigation
of the main effects and interactions. If, however, higher factor
interactions are negligible these sources of variation may actually
be excellent estimates of the experimental error.

Consider a three-factor factorial experiment example on explosive

switches in which the three factors are fixed. The experimenter is
interested in two metals for pistons (2 s al and Teflon). This factor
is called qualitative. The other two factors are amount of primary
initiator with levels 5, 10, and 15 mg and packing pressures 12,000,
20,000, and 28,000 psi which are quantitative. The purpose of the
experiment is to investigate the effects of the factors on the
dependent variable, firing time.

An example of hidden replication can be shown in this explosive

switch experiment. There are nine observations of the firing time
for each metal and six observations of firing time for each amount
of primary initiator and for each packing pressure.

The design of the experiment or the randomization procedure is

to take the 18 treatment combinations, that is the combinations of
metals, initiators, and pressures, in an order such as that given in
the following tabulation.

Metals Initiator Pressure

1 . 2 s al 5 12,000
2. Teflon 5 12, 000
3. 2 s al 5 20,000
4. Teflon 5 20,000

18. Teflon 15 28,000

Draw a random number between 1 and 18, say 3. Then the first
combination or explosive switch to be fired will be number 3
100 4. COMPLETELY RANDOMIZED DESIGN

(2 s al, 5 mg of primary initiator at 20,000 psi packing pressure).

Continue the random procedure until all 18 are placed in firing
order, at random.

The model considering all 18 treatment combinations as separate

treatments could be written as:

y i - U ♦ Ti * ; i = 1, 2, ..., 18 (4.2.1)

where
y^ = firing time of the switch with the i treatment
(combination) considered the variable to be analyzed

y = overall mean

T^ = effect of the i ^ treatment (combination)

£(i)l = effect of the other extraneous variables (consider

random because of the randomization) on that experimental
th 2
unit with the i treatment, assumed NID (0 , a )

TABLE 4.2.1

ANOVA using Equation (4.2.1)

Source df EMS

Treatments 17 a2 + « (T)
Experimental Error 0 o2

The analysis in Table 4.2.1 obviously allows no tests on treatments

because there are zero df for experimental error (to be called
error from now on).

If, however, the factorial structure is utilized in the

analysis, the model for the 18 treatment combinations with one
experimental unit per treatment combination is:
4.2 MORE THAN ONE FACTOR 101

where y. ., = the firing time required for the switch made up of the
th J ^ ^
i metal, j initiator and k pressure, as indicated in the
treatment combination table, and all components on the right-hand
side of the equation are the main effects and interactions of metals,
initiators, and pressures plus the mean and error.

In Eq. (4.2.1) the term T^ is merely extended into more

meaningful components known in the experiment as effects of initiators,
pressures and so on, in Eq. (4.2.2). Since all factors are fixed
the analysis is as shown in Table 4.2.2.

TABLE 4.2.2

ANOVA using Equation (4.2.2)

Source df EMS
2
Metals (M) 1 0 + 9<j>(M)
Initiators (I) 2 a 2 + 6 <J>(I)
M x I 2 a2 + 3<|>(MI)
Pressures (P) 2 a2 + 6 *(P)
M x P 2 a2 + 3<j>(MP)
Ix P 4 a2 + 2*(IP)
M x Ix P 4 a2 + <#>(MIP)
2
Error __0 a
Total 17

In order to describe the inference space, the experimenter must

recognize that no true replication exists in this experiment and if
2
there is no prior estimate of a , no exact test of main effects or
interactions is possible. If the experimenter is willing to assume
that the three-factor interaction is negligible, <J>(MIP) = 0, then
the mean square for the source, MIP, may be used for the error
estimate. An additional possibility exists even if <|>(MIP) cannot be
assumed negligible. That is to use the mean square for MIP as the
denominator for testing all other main effects and interactions.
Since the true experimental error should be less than or equal to
102 4. COMPLETELY RANDOMIZED DESIGN

the MIP mean square the tests are conservative and those effects and
interactions that show significance should be significant.
2
If the experimenter had no prior knowledge of a and could not
assume <|>(MIP) = 0, he may consider repeating treatment combinations
on randomly chosen experimental units or repeating a few of the
treatment combinations in the experiment. The usual practice in
analyzing the experiment with only a few of the treatment combinations
repeated on other experimental units is to use the additional
observations to estimate the experimental error, but not to use these
additional observations in estimating the effects of the treatments.
If a program for a digital computer were available to use the
appropriate weights, it may be desirable to utilize all the
information; but if the investigator does not have the program and
is in a hurry for the results, he usually does not use these
additional observations in the analysis of effects.

Consider the example on explosive switches in which six of the

treatment combinations are repeated. The model is

rijw = * + Mi + ij + MIij + pk + MPik

♦ IP.k ♦ MIpijk * e(ijk)t (4.2.3)

where the components are the same as in Eq. (4.2.2) except that
I = 1 , 2 for only six of the treatment combinations, and I = 1 for
the remaining twelve treatment combinations. The analysis of
variance is the same as that given in Table 4.2.2 except that we use
only the first observation for the factorial analysis, but use the
two observations per treatment combination to calculate the error
sums of squares (SS). For each treatment combination with two
observations, calculate the SS with 1 df by squaring the difference
between the observations and dividing by 2 [see Example 7.2 for
proof]. Then merely add the six SS to obtain the SS for error. The
ANOVA is given in Table 4.2.3.
4.2 MORE THAN ONE FACTOR 103

TABLE 4.2.3

Error Analysis Using Equation (4.2.3)

Source df EMS

2
Error 6 a

If the levels of the factors are equally spaced and quantitative,

the experimenter may use a more efficient design. If he can assume
a second-order model then he can utilize the df from the higher
order terms to estimate the error and no repeats of treatment
combinations will be necessary. A model to describe the effects and
interactions of metals, initiators, and pressures on firing time
follows:

y = aQ + o1 Z1 + a2 Z2 + <*2 2 Z 2 2 + ai2 ZlZ 2 + a3 Z 3 + a3 3 Z 3 3

+ a13ZlZ3 + a23Z2Z3 + e (4-2'4)

where
y = firing time (assumed normally distributed)

ot0 = overall constant

cx^ = partial regression of firing time on metals

a^ = partial regression of firing time on linear part of

initiators

<*22 = partial regression of firing time on quadratic part of

initiators

a 12 = partial regression of firing time on the interaction of

metals by the linear part of initiators and so on to

a23 = Part*al regression of firing time on the interaction of

the linear part of initiators and the linear part of
pressures

Zy ..., Z2 Z3 = the corresponding orthogonal independent

variables
104 4. COMPLETELY RANDOMIZED DESIGN

e = residual from the higher order terms, assumed

NID (0, a2)

This model is obtained using orthogonal polynomials since the

spacings on the quantitative factors are equal.

To use this equation as a prediction equation the correct

transformed variables must be used for the Xfs as indicated in
Chapters 1 and 2. Davies (1971, p. 519 and 520) explains the
analysis through a corresponding model and Edwards (1962, Chapter
14), writes on trend analysis. All of this is useful in designing
an experiment which will have as few observations as possible, yet
obtain the information desired by the investigator.

Table 4.2.4 summarizes the 1 to 1 correspondence between the

equation using independent variables, Z, and the analysis of variance.

TABLE 4.2.4

ANOVA using Equation (4.2.4)

Regression Source df EMS
coefficient

ai
Metals (M) 1 a2 + <|>(M)
Linear initiators (1^) 1 *2 + *(1 *)
a2
Quadratic initiators (I ) 1 a2 + *(l )
a 22
2 "
ai2 M x I, 1 o + ♦(MI1)
Linear pressure (P^) 1 a2 + ♦(P1)
a3
Quadratic pressure (P ) 1 a2 + <t>CP )
a33
2 q
MxP, 1 + <KMPa)
ai3
a23 1 °2 * ♦(!,?,)

df
M x I
q
Residual (higher order M x P
q
terms from the P0 x I
interactions) _9 I q > a2
P x I0
17 q *
P X I
q q
MIP
4. 2 MORE THAN ONE FACTOR 105

The estimate of the standard error of the metal mean is

residual mean square_________________ \p

(number of observations for that mean = 91

)
and similarly for other means.

Hence, with the higher order terms in the model being zero,
there is an estimate of the error variance from the data with nine
degrees of freedom. If the assumptions are correct, the design of
this experiment is very "efficient,” that is, few observations have
provided the estimates desired for the model; therefore there is a
reasonably good error estimate.

If, on the other hand, the experimenter is uncertain that the

higher order terms are zero and feels that he must be assured that
the residual source with the 9 degrees of freedom is a good estimate
of the experimental error, a few treatment combinations, say six,
could be repeated (the same concept as given before). After this
better estimate of experimental error is obtained, say with the 6

degrees of freedom, the residual from regression (or the lack of fit)
could be tested using the estimate of the experimental error from
the six repeated treatment combinations as the error. It would be
advisable to use the probability of a Type I error of 0.25 before
pooling the residual error with the experimental error because under
these experimental conditions and reasonable alternative hypotheses
the probability of the Type II error, 3, is usually quite large for
small values of a. Pooling is accomplished by adding together the
SS of each source and the corresponding degrees of freedom.

References for pooling in analysis of variance are Bozivich,

et al. (1956) and Winer (1971, Section 5.16).

A less desirable approach is to obtain one degree of freedom

for nonadditivity given in Chapter 1 to break certain interaction
sources into nonadditive and error components.

Example 4.1: Calculation of the ANOVA table and investigation

of the relationship between an orthogonal and nonorthogonal
106 4. COMPLETELY RANDOMIZED DESIGN

regression model for a completely randomized design, and the data for
this two factor design with coded levels of temperature and calcium
content are given below. The measured response, y, is time-measured
in hours (cell totals are circled).

Model y..k - p ♦ C. ♦ T. ♦ CT.. ♦ e(..)k

i = 1, 2, 3 j = 1, 2, 3, 4 k = 1, 2

Temperature

1 2 3 4 Total

26 ^

c
1
: © ; © © 27 © T."*• = 117
1
o
-M
e
o
u 2
; © : © :: © © T 02 *• = 113

G)

©21© W
u
3
: © : 23
;; © T.3* • = 143
Total T.i. = 43 T 0 = 51 T •4*
A =169 T>-# = 373
•2 * T*3* " 110

The calculation of the sums of squares and the ANOVA table

follows:

Z E E y2 = 62 + 82 + ... + 352 = 7683

i j k 13K

CT
L = 2x3x4
-*■*— = = 5797 04
24 S/S/.U4

SS Calcium = (1172 + 1132 + 1432) - CT = 66.33

SS Temperature = i (432 + 512 + 1102 + 1692) - CT = 1721.46

SS Cells =(142 + 162 + ... + 672) -CT = 1857.46

4.2 MORE THAN ONE FACTOR 107

SS C x T = SS Cells - SS Calcium - SS Temperature * 69.67

SS Total = 62 + 82 + 72 + ... + 352 - CT = 1885.96

SS Error = SS Total - SS Cells = 28.50

Source df SS MS EMS F

Calcium (C) 2 66.33 33.17 a2 + 8 (f>(C) 13.96*

Temperature (T) 3 1721.46 573.82 02 + 6 <f>(T) 241.61*

C x T 6 69.67 11.61 a2 + 2<|>(CT) 4.89*

Error 12 28.50 2.38 a2

Total 23 1885.96

*Significant at the a = 0.05 level.

The 11 degrees of freedom for the above treatments will be

broken down into single degrees of freedom with the method of
orthogonal polynomials which was demonstrated in Chapter 2. This
will then be followed by various nonorthogonal regression models in
order to relate the correspondence between the two methods. The
method of calculation for the orthogonal polynomials was given in
Chapter 2 and will not be reproduced here. The nonorthogonal
regression equations were generated with a canned computer program
and so no method of calculation will be shown here. The sums of
squares and coefficient estimates for all possible orthogonal
polynomials are given in the following tabulation.
108 4. COMPLETELY RANDOMIZED DESIGN

Factor Polynomial SS Coefficient Coefficient es

42.25 1.62
Z1 al
24.08 0.71
cq Z 11 all
66.33

1591.41 3.64
Z2 a2

T 108.38 2.12
q Z 22 a 22

T 21.67 -0.42
c Z 222 a 222
1721.46

Cl x Tl 33.80 0.65
Z1 Z 2 a 12

0.00 0.00
C* X T q Z1 Z 2 2 a 122

0.45 -0.08
C* X T C Z1 Z 2 2 2 a 1222

35.27 0.38
Cq X T * Z1 1 Z 2 a 112

C x T 0.00 0.00
q q Z1 1 Z 2 2 all2 2
C x T Z1 1 Z 2 2 2 0.15 -0 . 0 2
q c a 11222
69.67

Fitting a regression equation to the calcium content factoi

produces the following equation and ANOVA table:

y = 19.38 - 6 .8 8 C + 2.12C2

Source____ df______ SS_______ MS

Regression 2 66.33 33.16
Residual 21 1819.63 86.65
Total 23 1885.96

Note that the sum of squares due to regression on calcium content

is the same as that obtained for the orthogonal fit and that
obtained for calcium content in the analysis of variance. A similar
result is obtained for the temperature effect as shown in the
following equation and ANOVA table:
4.2 MORE THAN ONE FACTOR 109

y = 23.072 - 27.376 T + 12.917 T2 - 1.439 T3

Source______df_____ SS_______ MS
Regression 3 1721.46 573.82
Residual 164.50 8.22
Total 23 1885.96

One might think that a regression equation with the cross

product terms might produce the sums of squares due to regression
equal to the CT interaction sum of squares (69.67). The equation
for all the cross products terms and the ANOVA table follow:

y = 19.830 - 21.902 CT + 9.495 CT2 - 0.953 CT3

+ 4.700 C2T - 1.828 C2 T2 + 0.168 C2 T3

Source_____ df_____ SS_______ MS

Regression 6 1773.00 295.50
Residual 17 112.96 6.64
Total 23 1885.96

As seen in the above ANOVA table, the sum of square due to regression
is 1773, not 69.67. The reason for this result is that the cross
product sum of squares includes some of the main effects of both
calcium content and temperature as well as their interaction effect
for this nonorthogonal regression case.

Fitting a regression model containing all 11 terms produces

the following:

y = 11.914 + 12.35 C - 2.91 C2 - 18.67 T + 11.31 T 2

- 1.23 T3 - 6.99 CT - 0.44 CT2 +0.06 CT3 + 1.15 C2T

2 2 2 3
+ 0.53 C T - 0.07 C T
110 4. COMPLETELY RANDOMIZED DESIGN

Source df______ SS_______MS

Regression 11 1857.46 168.86
Residual 12 28.50 2.37
Total 23 1885.96

The sum of squares due to regression is now the same as the sum of
squares due to the calcium content, temperature main effects and
the interaction combined.

In summary it should be pointed out that Marascuilo and Levin

(1970) show that one cannot examine interaction effects individually
without removing the influence of the main effects first. The
problem encountered here is of this general nature. Given that the
SS due to total regression for calcium content + temperature
+ CT interaction is 1857.46 and knowing the SS due to regression
for calcium content and temperature is 1787.79, we find that the SS
due to regression for CT interaction is the difference, 69.67. This
CT interaction SS cannot be obtained directly using nonorthogonal
regression.

Problem 4.2.2. (a) Utilizing the data of Problem 2.2.4,

determine by means of nonorthogonal regression procedures the
sums of squares due to individual effects of temperatures,
supplements, TS interaction, and TS cross-product terms.
(b) Interpret your results.
(c) If you answered Problem 2.2.4, explain what additional
information, if any, you obtained in (a) above.

Problem 4.2.3. If an experimenter ran an analysis of variance

on data from his experiment and all main effects and*interactions
turned out to be significant, describe the procedure you would
recommend before the results are written up for publication.

Problem 4.2.4. An agronomy experimenter was interested in the

amylose content of rice. Three varieties of rice each given
five different levels (equal spaced) of fertilizer were to be
grown on an agronomy experimental farm. The experimenter had
the option of using from 15 to 30 plots for the experiment and
sampling the plants as he desired from each plot before analyzing
the rice for amylose in the laboratory.
The laboratory analysis procedure involves gradually heating
starch granules under a microscope until they lose their
reflecting ability (birefringence) and recording the temperature
4.2 MORE THAN ONE FACTOR 111

at the time of birefringence. The variable to be analyzed is

a function of the recorded temperature which in turn is a
function of the amylose content.
(a) Discuss how you would design this experiment
describing the various places randomization should occur.
(b) Show the model and outline the analysis for your
experiment including inferences you want to make.
(c) If the experimenter wanted a second-degree regression
model, show the procedure you would use to handle the
qualitative factor and write out the model.
(d) Consider the role of locations and explain how the
experimenter can expand his inference space if more locations
were used.

Problem 4.2.5. Consider a two factor experiment in which factor

A has five equally spaced levels and factor B has four
qualitative levels. If the third and higher degree terms are
considered negligible, design a good experiment and show the
analysis for it, including appropriate means and estimated
standard errors.

Problem 4.2.6. Fabricate data for a three factor experiment

which has a significant three factor interaction but the sums
of squares for all main effects and two factor interactions are
equal to zero. For simplicity use only two observations per
cell and two levels per factor.

4.2.2 Some Factors Fixed and Some Random

For the explosive switch example consider the influence on the

design of the experiment if the factor, metals, is random. The
randomization procedure for the factorial experiment is the same as
that for the fixed model once the two metals are picked from a large
population of possible metals, but the inference space has changed.
Let us assume that there are at least one hundred possible metals so
that, for practical purposes, the model is infinite. Again we have
only one observation per treatment combination; the analysis is
given in Table 4.2.5.

For mixed models see Anderson and Bancroft (1952, Chapter 23),
Hicks (1973, Chapter 10), Winer (1971, Chapter 7), Edwards (1962,
Chapter 17) and John (1971, Section 4.7).
112 4. COMPLETELY RANDOMIZED DESIGN

TABLE 4.2.5

ANOVA

Source df EMS

2
Metals (M) 1 a +
9 °M
2
Initiators (I) 2 a + 3 ojjj * 64.(13

M x I 2 a2 +
" mi
Pressures (P) 2 a2 + 3 o9p • 6 *(P)

M x P 2 a2 +
3 4
Ix P 4 a2 +
°MIP * 2*<IP>
2
M x I x P 4 a2 +
°MIP
Error 0 a2

To test initiators one should use the interaction of metals x

initiators, to test pressures the correct test is against metals x
pressures, and to test the interaction of initiators by pressures
the correct error is the mean square of metals x initiators x
pressure. In general, for any factorial experiment with one factor
random and the remaining factors fixed, all the main effects and
interactions of fixed factors are tested by the corresponding
interaction of the fixed part and the random one. The random main
effect and the interactions of the random factor by the fixed parts
are tested against the error.

If experimenters are not interested in the random components

other than to provide an inference base for the fixed factors, there
usually is little need to have more than one experimental unit per
treatment combination. If however, there is interest in estimating
2
the variance component of the random main effect (here aM), then an
2
estimate of a must be provided. In this latter case a regression
model dealing only with the fixed factors may be desired and a
2
residual could be obtained to estimate a ; but it would be preferable
4.2 MORE THAN ONE FACTOR 113

2
to take a partial replication to obtain an estimate of a directly
(not assuming the higher order terms are zero for the residual).

Here, as in the fixed model, the design procedures demonstrated

are easily extended to any complete factorial; however, in the mixed
factorial, attention must be given to the various combinations of
random and fixed factors in the analysis (the expected mean squares)
before the assumptions on higher order terms are made and pooling
procedures tried.

Example 4.2: Calculation of the ANOVA table for the three

factor experiment on explosive switches, and the coded data for firing
time are given in the following table. The only difference between
this example and that shown in Table 4.2.5 is that in this example
there are two switches per treatment combination (cell totals are
circled) instead of one.

Metals (random)
1 2

Initiators (fixed) Initiators

1 2 3 1 2 3

1
44

39
© 27©
r \
20
35
30
© 13©
10
22

I S ©
r \
00

25
6©
©

Pressures ~ 12 /~\
(fixed) 1 I ? © 27 ©
O

21 3 2 © 11 ^ 22

3 43 /^\ 28

22
r \
31 ©
38 W
7 ©
12
“ ©
13
21©
19

Hotels YijkI ■ „ . M. . I. ♦ MI.. . Pk * MPlk . IPJk . MIP.jr . c(..k)t

i = 1,2 j = 1, 2, 3 k = 1, 2, 3 t = 1, 2
114 4. COMPLETELY RANDOMIZED DESIGN

Totals:
Metals Initiators Pressure Grand

T.1 *.. = 595 = 310 T. . = 276 T •••• = 854

T-i..
T0 = 259 = 221 T 0 = 292 n = 36
2 *•• T*2' • ••2 #
= 323 T . = 286
T*3* • ••3*

Metals by Initiators Metal by Pressure

TH.. - 255 T,. = 55 Tl.l. = 195

21 *• T2 *1 * - 81

T 78 T2.2. - 95
T1 2 .. * 143 22 *• t i .2. = 197
T = 126
T13*• = 197 23* • t i .3. = 203 T2.3. - 83

Initiators by Pressure

T.H. - 102 = 72 T = 102

T*2 1 * •31*
= 75 T = 112
T*1 2 * = 1°5 T*2 2 * •32*

T,13* " 103 T*23* = 74 T •33*

„ = 109

Z Z Z Z y2 . = 25196
i j k it

T2
CT = — = 20258.78
n

Sums of squares:

SS Total = 25196 - CT= 4937.22

SS Metals = (5952 + 2592) - CT = 3136.00

SS Initiators = yy (3102 + 2212 + 3232) - CT = 513.72

SS Pressure = yy (2762 + 2922 + 2862) - CT = 10.89

4.2 MORE THAN ONE FACTOR US

SS M x I = | l E T?.- C T - S S Metals - SS Initiators

6 . . ij ••
i 3
1 2 2 2 2 2 2
= ^ (255 + 143 + 197 + 55z + 78z + 126z)
D
- CT - SS M - SS I

= 24878.00 - 20258.78 - 3126.00 - 513.72

= 969.50

SS M x P = £ (1952 + 1972 + 2032 + 812 + 952 + 832)

- C T - S S Metals - SS Pressure

= 23419.67 - 20258.78 - 3136.00 - 10.89

= 14.00

SS I x P = 1 (1022 + 1052 + ... + 1092)

- C T - S S Initiators - SS Pressures

= 20788 - CT - SS I - SS P

= 4.61

S S M x I x P = i s £ ST?., - CT - SS M - SS I - SS P - SS M x I
i j k 1Jk*

-SSMxP-SSIxP

* i (832 + 882 + 842 + 472 + ... + 402) - 24907.50

= 24946 - 24907.50

= 38.50

SS Error = SS Total - SS M - SS I - SS P - SS M x I

- SS M x P - SS I x P - SS M x I x P

= 4937.22 - 4687.22

= 250.00
116 4. COMPLETELY RANDOMIZED DESIGN

The resulting ANOVA table is as follows:

Source df SS MS EMS F
2
Metals 1 3136.00 3136.00 a + 225.8*
18 4
2
Initiators 2 513.72 256.86 0 + 6 12 *(I) <1
’mi *
2
M x I 2 969.50 484.75 a + 6 34.9*
•2 ,
2
Pressure 10.89 5.44 a2 +
QMP + 12*(P)
2 6 <1

2
M x P 2 14.00 7.00 a2 + 6 <1
aMP
I x P 4 4.61 1.15 a2 +
aMIP + 4<j>(IP)
2 <1

M x Ix P 4 38.50 9.62 o2 + 2 <1

aMIP
2
Error 18 250.00 13.89 0

Total 35 4937.22

*Significant at the 0.05 level.

Problem 4.2.7. Interpret the results of Example 4.2.

Problem 4.2.8. Let metals and initiators be random and

pressures be fixed.
(a) Show the model and analysis with one experimental
unit for each treatment combination.
(b) Using (a) above, design an efficient experiment using
more than one experimental unit for each treatment combination
if it is warranted, and show the appropriate analysis.

4.2.3 All Random Factors

Using the same explosive switch example as given previously,

and assuming an infinite model (the levels of the quantitative
factors will not be evenly spaced if selected at random), we have
the analysis shown in Table 4.2.6.
4.2 MORE THAN ONE FACTOR 117

TABLE 4.2.6

ANOVA

Source df EMS

2 2 2 2
Metals (M) 1 a2 + + 3 3 + 9
°MIP aMI + aMP aM
Initiators (I) 2 a2 +
2
°MIP
+ 2
«?p *
3
■s, + 6
2
al
M x I 2 a2 + a2
MIP
+ 3
•2,
Pressures (P) 2 a2 + 2
MIP
2
+ 2
2
3
2
aMP
+ 6
<•1
M x P 2 a2 + + 3
°MIP MP
2
Ix P 4 a2 + + 2
aMIP alP
2
M x I x P 4 a2 +
°MIP
2
Error __0 a

Total 17

The two-factor interactions are all tested against the three factor
interaction, and all main effects require approximate F tests which
are sometimes called F 1 tests, Cochran (1951), Anderson and Bancroft
(1952, p. 350), Ostle (1963, p. 302). For example, to test
2
H: aw = 0 use
M
r, _ MS(metals) + MS(M x I x P)
“ MS(M x I) + MS(M x P)
2
since the EMS (metals) + EMS(M x I x P) contains only the term 9 aXi
M
in addition to the terms in EMS(M x I) + EMS(M x P).

To obtain the degrees of freedom for either the numerator or

the denominator, one may use the Satterthwaite (1946) approximation.
118 4. COMPLETELY RANDOMIZED DESIGN

where
k
M = 7 a.M.
i=i 11
th
M^ = i mean square in the numerator if the degrees of freedom
are being calculated for the numerator or similarly if the
denominator degrees of freedom are being calculated

a^ = coefficient of the i**1 mean square

f^ = degrees of freedom for the mean square

Hence all main effects and two-factor interactions can be tested

using only one experimental unit per treatment combination. This is
a very efficient design because no assumptions on residuals are
necessary. Of course, if the three factor interaction is of interest
2
there must be partial replication to provide an estimate of a .

Of all the models fixed, mixed, and random; the random model
allows a design with the fewest assumptions after the factors are
declared random. Thus it is absolutely necessary for the experimenter
to know whether each factor will be random or fixed before an
appropriate design is prescribed for factorial experiments.

If all factors are random there should be no regression trend

analysis. Usually the estimation of the variance components is more
important than the test of hypotheses for a random model. It is
clear that the above design provides estimates of the variance
2 2
components except for and a , separately, with one experimental
unit per treatment combination. A good reference for variance
component analysis is Anderson and Bancroft (1952).

As in the fixed model, the design methods demonstrated for the

random three factor-factorial are extended in a straight forward
manner to any sized complete factorial. Remember the design is
completely randomized.
4.2 MORE THAN ONE FACTOR 119

Problem 4.2.9. The random model example on explosive switches

had the variable to be analyzed as a transformation of firing
time. The data were analyzed with the following results:

ANOVA

Source df MS
Metals 1 530
Initiators 2 225
M x I 2 46
Pressures 2 35
M x P 2 28
I x P 4 10

M x Ix P 4 5

2 2 2 2
(a) Test the various hypotheses, = aj = = ap =
aMP = aIP = using the F f test and pooling
techniques where appropriate.
(b) Explain the results.
(c) Comment on a possible new model.

4.2.4 Spacing Factor Levels

For random factors there is no problem on spacing because the

levels are not purposely spaced and there is no interest in means.
For fixed qualitative factors it is obvious that there is no problem
on spacing, but there usually is interest in the means. Finally,
for fixed quantitative factors there is a problem on spacing, and
there usually is interest in the means.

Most experimenters think they should arithmetically equally

space the levels of fixed quantitative factors. For example, the
four levels of the factor or independent variable (X) for fertilizer
treatments may be the percent of nitrogen 0, 20, 40, 60. The main
reason for using equally spaced factor (X) levels is to use orthogonal
polynomials when the trend is not known in hopes that the trend may
be predicted from the data. If the trend is known to be linear, and
120 4. COMPLETELY RANDOMIZED DESIGN

the research worker is only interested in the slope of the line,

placing half the observations at either extreme of the factor space
(e.g., at zero and 60 for the nitrogen example) would give him

2 a2 - 2
maximum information because a, = — z and I(X.-X) is
1 I(Xi-X) 1

maximized when the V s are at theextremes. As more information is

available to indicate the function is other than polynomial, the
experimenter should depart from the usual arithmetic equal-spacing
idea.

If the trend is exponential in nature the polynomial may be

quite a poor predictor and arithmethic equal spacing is not the best.
Usually it is better to follow a geometric spacing rather than an
arithmetic one if an exponential relationship is suspected. In
general, there should be a method for testing departure fromthe
exponential. After transforming the independent variable (or factor
levels) to logarithms the usual orthogonal polynomial trend analysis
may be used, but the analysis is on the transformed independent
variable, Li (1964, pp. 205-209), and Finney (1960, pp. 170-173).

Consider the example in which the relationship between y, the

surviving offspring, and X, the dosage of X-rays to the parents, is
v b
thought to be of the general form e7 = aX .

The experimenter may select doses of 100, 200, 400, and 800
roentgens for a fixed time and examine the trend by testing a
departure from the exponential. Of course it can be seen that if
the logarithm of dosages are taken and the departure from a straight
line is then investigated, the experimenter has really examined the
departure from the exponential. If the trend is not found to depart
significantly from the exponential, the coefficients a and b can be
estimated using the transformed X variable.

Problem 4.2.10. In an experiment using flour beetles

(Tribolium castaneum), the effect of X-rays on progeny numbers
was to be investigated. The relationship between progeny
number (y) and dosage of X-ray (X) was thought to be
4.3 REFERENCES 121

The lower limit of X-rays was set at 100 roentgens for a

fixed time and the upper limit was set at 1 0 0 0 roentgens for
the same fixed time. If there were going to be five levels of
X-ray doses;
(a) What should the other three levels of roentgens be in
order to allow an easy test for departure from linearity on the
transformed scale?
(b) Using the levels of (a) above, the experimenter
obtained the following data:

Dosage level Average progeny number per family

1 25.8
2 22.5
3 21.2
4 19.0
5 17.5

Assuming there were 10 families for each dosage level in

a CR design and that the error mean square was 13.0 show the
analysis of the data including the prediction model for an
exponential investigation. That is, check for linearity on the
transformed scale. If you accept linearity estimate a and b
and then write down the appropriate exponential prediction
model.

4.3 REFERENCES

Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research,

McGraw-Hill, New York, 1952.
Bozivich, H., Bancroft, T. A. and Hartley, H. 0. Ann. Math. Stat.
27:1017 (1956).
Cochran, W. G. Biometrics 7:17 (1951).
Cochran, W. G. and Cox, G. M. Experimental Designs, 2nd ed., Wiley,
New York, 1957
Cox, D. R. Planning of Experiments, Wiley, New York, 1958.
Davies, 0. L. (Ed.). Design and Analysis of Industrial Experiments,
2nd ed., Oliver and Boyd, London, or Hafner, New York, 1971.
Edwards, A. L. Experimental Design in Psychological Research,
Rinehart, New York, 1962.
Finney, D. J. An Introduction to the Theory of Experimental Designs,
University of Chicago Press, Chicago, 111., 1960.
Hader, R. J. The Am. Statistician 27, 82 (1973).
122 4. COMPLETELY RANDOMIZED DESIGN

Hicks, C. R. Fundamental Concepts in the Design of Experiments, 2nd

ed., Holt, Rinehart and Winston, New York, 1973.
John, P. W. M. Statistical Design and Analysis of Experiments,
Macmillan, New York, 1971.
Kempthorne, 0. Design and Analysis of Experiments, Wiley, New York,
1952. Distributed by Krieger Pub. Co., Huntington, New York.
Li, Jerome C. R. Statistical Inference II. Edwards Brothers, Inc.
Ann Arbor, Michigan, 1964.
Mandel, J. Statistical Analysis of Experimental Data, Wiley, New
York, 1964.
Marascuilo, L. A. and Levin, J. R. Am. Ed. Res. J. 7:397 (1970).
Ostle, Bernard Statistics in Research, 2nd ed., Iowa State University
Press, Ames, Iowa, 1963.
Prairie, R. R. and Zimmer, W. J. JASA 59:1205 (1964).
Satterthwaite, F. E. Biometrics 2:110 (1946).
Winer, B. J. Statistical Principles in Experimental Design, 2nd ed.,
McGraw-Hill, New York,1971.
 Chapter 5

RANDOMIZED COMPLETE BLOCK DESIGN (RCBD)

Basically, the randomized complete block design is a group of

completely randomized designs (usually, however, each CRD has only
one experimental unit per treatment combination). Ordinarily each
member of the group contains a sufficient number of homogeneous
experimental units to accommodate a complete set of treatment
combinations. This collection of experimental units is referred to
as a block. One major reason for blocking is that the experimenter
does not have a sufficient number of homogeneous experimental units
available to run a completely randomized design with several
observations per treatment combination.

A realistic situation where blocking would be required would be

in an industrial type experiment where there is only sufficient time
to run one set of treatment combinations per day. In such an
experiment one would be blocking on time and the number of days that
the experiment is carried out would be the number of blocks. This
type of experiment points out another reason for blocking which is
the technique that is utilized to expand the inference space. Even
if it was possible to run a large number of each treatment combination
on one day the experimenter might still choose to block over time so
that he could make inferences over time rather than to just one
particular day.

If an experimenter were to run a RCBD and analyze it as a CRD,

then any effects which should have been attributed to blocks would
end up in the error term of the model. Thus another major reason
for blocking is to remove a source of variation from the error. It
should be observed that if the experimental unit variation among

123
124 5. RANDOMIZED COMPLETE BLOCK DESIGN

blocks is not larger than the variation within blocks, there is no

reason to run a RCBD rather than a CRD.

The group of blocks, which may be random or fixed, make up the

design called the randomized complete block design (RCBD). These
blocks are called Tlcomplete” because all treatments or treatment
combinations appear in each block. As a result of the construction
of this design, there is one restriction on the randomization in that
the randomization of treatments onto the experimental units is carried
out within each block separately, not over all the experimental units
at one time as with a CRD.

The single most important concept introduced in this text is the

usefulness of the random component called a "restriction error"
(Anderson, 1970). We place this component (where appropriate) into
the mathematical model in order to present a realistic picture of the
experimental situation. The component is common to a group of
experimental units isolated by the experimental procedure. It is
random in the sense that if the experimenter were to recreate the
same group of units, this component would be different.

By the very nature of the confinement of the experimental units

to a given group, any randomization of the treatments onto the
experimental units must be restricted to only those units that are
present in that grouping. This situation occurs in all RCBD, and
the restriction error component is completely confounded with blocks.

In many other experimental situations treatments cannot be

randomized onto the experimental units within the blocks or groupings.
In this nonrandom case the restriction error is still present because
the grouping of the experimental units still exists and the effect
of blocks is still completely confounded with the restriction error.

The restriction error is not estimable from the data but it is

placed in the model (appropriately indexed) and is allowed to appear
in the corresponding analysis of variance as a source of variation.
There are no degrees of freedom (df) and no sum of squares (SS) for
the error; however, since the restriction error appears in the
5.1 BLOCKS FIXED 125

theoretical linear model, the variance component for it does appear

in the expected mean squares (EMS). This variance component in the
EMS forces the experimenter to recognize the restriction on
randomization and account for it in the F-tests. The real power of
the restriction error is that it makes the experimenter think about
the restriction on randomization he has imposed on his design
(usually to save time and/or money) and to understand its effect on
the overall results of the experiment. Almost always that effect is
to decrease the number of degrees of freedom for the appropriate
error to test the most important factor.

It is interesting to note that the restriction error does not

disturb the algorithm presented in Chapter 2, and referenced in
Bennett and Franklin (1954) and more recently by Hicks (1973), to
derive the expected mean squares given the linear model. The EMS
are still obtained in the same straightforward manner and are
correct when restriction errors are in the model.

To explain the influences that the restriction error has on

analysis of data from a RCBD, treatments will be considered fixed and
the two cases, blocks fixed and random, will be covered in detail in
the following sections. It is left to the reader to use the material
from Chapter 2 to handle the cases in which treatments are random.

5.1 BLOCKS FIXED

Suppose that b levels of one fixed set of treatments and t

levels of another fixed set are run in all possible bt combinations
on bt experimental units completely at random. The design of such
an experiment is called a completely randomized design (CRD). The
linear model for the analysis of the data from such an experiment,
if the interaction of blocks and treatments is zero, is usually
expressed as:
b
(5.1.1)
t
126 5. RANDOMIZED COMPLETE BLOCK DESIGN

where
y ^ = the response from the experimental unit treated with the
i**1 level of treatment B and the j**1 level of treatment T

Vi = overall mean

Bi = theeffect of the itl1 level of (fixed) treatment B

Tj = theeffect of the j levelj i f (fixed) treatment T

e(ij) = t*ie(random) withinerror

An estimate of the error mean square may come from the interaction
source. Assume all the analysis of variance assumptions hold, so
2
that the is NID (0, a ), i.e., normally and independently
distributed with mean zero and variance a^. For this case the
appropriate analysis of variance (ANOVA) is given in Table 5.1.1.

To be able to write Eq. (5.1.1) one must assume there are no

restrictions on randomizations in the design of the experiment. This
may be called a two-way factorial completely randomized design.

Suppose, next, there are b blocks in an experiment, and t fixed

treatments are completely randomized onto t experimental units in
each of the b blocks such that there is a different randomization
in each block. The equation for analyzing the data from such an
experiment has been given by most authors and implied by many others
to be the same as Eq. (5.1.1). This design is certainly not
completely randomized because there is a different randomization of
the treatments in each block, not just one randomization over the
whole experiment as would be demanded for a CRD. In general,
statisticians agree that this is a randomized complete block design
(RCBD); however, they do not always agree on the model to analyze
the data. Wilk (1955), Addelman (1969), and others have suggested
using a model for a generalized randomized block design which includes
the usual CRD and RCBD. We wish to stress restriction errors and in
this section we deal with an equation that accounts for errors
between blocks in contrast to Eq. (5.1.1) that does not account for
the errors.
5.1 BLOCKS FIXED 127

TABLE 5.1.1

ANOVA for the Data from the CRD Using Equation (5.1.1)

Source df EMS

Treatments (B^) b - 1 a2 + t<KB)

Treatments (T^) t - 1 a2 + b<KT)

Within error (e,...) (b - 1 )(t - 1 ) a2

(ij )
(assuming interaction
BT is zero)

Since there is a different randomization of treatments within

each block, we recognize this peculiarity by making sure that the
error to test the hypothesis of equality among the treatment means
comes from within the blocks. On the other hand, if there is
interest in testing the hypothesis that block means are equal, we
insist that there must be an error among the blocks. The former
test (for treatment means) would be satisfied using the error
designated in Eq. (5.1.1), but the latter test (for block means) is
not satisfied using the same error.

One linear model that expresses all the sources of variation

mentioned above with the assumption of no interaction of blocks by
treatments and also provide intuitively correct tests is:

y. .. n = y + B. + + T. +
'ljk l l (i)k j (ijk)Z
(5.1.2)
i = l , . . . ,b j = 1 ,..., t k = 1 1 = 1

where
t»h
y - i v o = t*ie response from the % experimental unit in block
l j KXr th th
i, the k randomization, and given the j treatment

y = overall mean
th
B. = effect of the i block. In this case B. may be
i l 7
random or fixed, but we will assume it is fixed for
the ANOVA
128 5. RANDOMIZED COMPLETE BLOCK DESIGN

6 .... = the k**1 restriction error within the block,

(i)k ,
NID (0, c^), and completely confounded with since
k must always equal 1. In general one does not
include a subscript when the range is 1 ; however, it
is included here to indicate that there are many
errors possible but only one is realized. This term
will have zero degrees of freedom and no sum of
squares, but must appear in the model and expected
mean squares in order to distinguish this design from
the CRD related to Eq. (5.1.1). This term is the
result of the restriction on the randomization of the
th
treatments onto the i block's experimental units,

Tj = the effect of the j**1 treatment (fixed),

th th
., v = effect of the I random error associated with the ij
(ljKjx,
experimental unit and subjected to the k restriction
th
on the i block. Again, I must always equal 1, and
hence is usually not included in the model. It is
2 2
assumed that is NID (0, a ). The variance a
is estimated in the analysis of variance from the
usual source since we assumed that the block x
treatment interaction is zero in this experiment.

The analysis of variance for this model, along with the EMS
derivation is given in Table 5.1.2.

Ordinarily one does not include the subscripts k and I in a model

of this type since they have a range of only 1. Note, however, that
it is necessary to include these subscripts in order to make the EMS
algorithm function properly. For all models of this type from now on
we will not include the subscripts that have the range of one except
for cases that are more clearly described using them, but it will be
understood that the procedure used in Table 5.1.2 will be inherent
in the derivation of the EMS.

The corresponding model for Eq. (5.1.2) not using the subscripts
k and I is:
5.1 BLOCKS FIXED 129

TABLE 5.1.2

ANOVA for Data from RCBD Using Equation (5.1.2)

b 1 t 1

F R F R
df Source i k j Z EMS

b - 1 B. 0 1 t 1 a2 + ta2 + t<KB)
l
0 6 f ^i 1 1 t 1 a2 + ta2
(i)k
t - 1 T. b 1 0 1 a2 + bKT)
3
2
(b - 1 )(t - 1 ) 1 1 1 1 a
e(ijk)£

i = 1 , 2, .
,.,. + T.
, 2 , .;;; bt
y. . = U + B. + 6 cs.i.3)
0) J + £(ij) j = 1

l e r e the terms are defined as they were for Eq. (5.,1 .2 ) except that
k and Z are deleted.

To emphasize the effect of the restriction error on the various

F-tests in the ANOVA, we introduce a line immediately below each
source of variation called "restriction error."

It can be seen from Table 5.1.3 that the correct error for
blocks is the restriction error with zero df not the within error
with (b - 1)(t - 1) df as suggested in Table 5.1.1.

TABLE 5.1.3

ANOVA for Data from RCBD Using Equation 5.1.3

Source df EMS

Blocks (B^) b - 1 a2 + ta2 + t<j>(B)

2 2
Restriction error 0 a + to6

Treatments (T^) t - 1 a2 + b<J>(T)

2
Within error (e,. ..) (b - 1 ) (t - 1 ) a
Ojr
130 5. RANDOMIZED COMPLETE BLOCK DESIGN

Note: The restriction error is represented in the theoretical

model but cannot be estimated from the data in this experiment. All
of this is recognized in Table 5.1.3 by writing down the source
(restriction error), showing zero (0 ) df to depict the lack of data
and expressing the theory in the EMS; also, there is no sum of
squares or mean square. Of course, the other three sources have
sums of squares that can be computed from the data.

Since there are zero df and no sum of squares for the

restriction error, there is no test for blocks in this experiment.
Many authors* have indicated that there would be no test for blocks
for various reasons, but none has been specific to indicate it from
the model as we attempt to do here. It should be pointed out that
many times the experimenter is interested only in whether or not
blocks have been effective in reducing the estimated error to test
treatments. Under these circumstances the investigator may test the
combination of block effects and the block error (our restriction
error) to see if it is zero. That is, the hypothesis could be stated
2
from Table 5.1.2 as Ho : cr^ + <J>(B) = 0. It is obvious, then, that
the within error is appropriate for the test and it would make no
difference in the test whether Eq. (5.1.1) or (5.1.2) is considered.

It should be emphasized that the inference from such an

experiment is only for this set of treatments and blocks, and that
the estimated standard error of the j**1 treatment mean is

which is impossible to estimate from this RCBD. It is interesting

to notice, however, that the estimated standard error of the
difference of two treatment means (which is more important for
individual or multiple comparisons once treatments show significance
in the analysis of variance) is:

*
For example, Ostle (1963, p. 368).
5.1 BLOCKS FIXED 131

and is estimable.

The real danger in using Eq. (5.1.1) in analyzing data from

RCBD occurs when the experimenter is interested in the block means
in addition to the treatment means. To be more explicit, frequently
experimenters want to use treatments to represent the blocks in a
RCBD and restrict randomization of another set of treatments inside
the ones used as blocks. Recognition of this restriction on
randomization demands an equation other than Eq. (5.1.1). Our
suggestion is that Eq. (5.1.2) provides the basis for a more nearly
correct analysis of the data from such a design.

5.1.1 A Medical-Engineering Example

An example to demonstrate this danger occurred in a medical-

engineering problem (Beeson, 1965). An engineer constructed a
mechanical apparatus to simulate the circulatory system of human
beings. A storage tankwas used to control the pressure of the
liquid, simulating the blood, in the system. A pulse pump squeezed
flexible rubber tubing in the system to create the pulsing action
needed to simulate the heart actions. The motor on the pump could
vary the pulse rate between 0 and 220 beats per minute. In the
experiment 4 prosthetic cardiac valve types and 6 different pulse
rates were to be used. If a completely randomized design were run
so that each treatment combination would occur twice, the
experimenter would have needed 48 valves (12 of each type).

The purpose of the experiment was to select the best valve type
out of the four for all pulse rates and/or best valve type for
particular pulse rates. To a statistician this means that the main
effect of valve types and the interaction of valve types by pulse
rates should be examined carefully.
132 5. RANDOMIZED COMPLETE BLOCK DESIGN

One of the variables to be measured and analyzed was maximum

flow gradient (mmHg) and it was found to have reasonably good
statistical properties for analysis.

The experimenter thought it was too difficult to obtain as many

as 12 valves per valve type; and, too, the experiment would take too
long to run. After all, the CRD would require that the mechanism
holding the valve type would have to be dismantled and reassembled
47 times. An example of such a CRD is the following:

Run Valve type Pulse rate

1. 2 4
2. 4 3
3. 3 5
4. 2 1
5. 2 4
6. 1 2

48. 1 6

Note: In this design a new valve and reassembly operation or

set up of the machine would be necessary for each run. Even for run
5 it would be necessary to remove the valve for type 2 and replace
it with another valve of type 2 .

5.1.2 RCBD More Than One Observation Per Cell (Example)

The investigator wanted to run the experiment by drawing, at

random, one of the valve types and running, in a random order, all
12 pulse rates (6 actual pulse rates each repeated once) on only one
valve of this valve type. Next he wanted to select one of the
remaining three valve types at random and with a new random order,
run the 12 pulse rates as before on only one valve of this new valve
type. He wanted to continue this procedure for the other two valve
types and complete the experiment using this RCBD with two
5.1 BLOCKS FIXED 133

observations per treatment combination of valve type by pulse rate,

assuming that there is no correlation of the within treatment
combination errors. Since he has used six different treatments
(pulse rates) twice on each experimental unit (valve), this
assumption may be violated. Detailed discussion of this topic is
given in Section 6.2.1. Suffice it to say here that this assumption
of no correlated errors will be used for the remainder of this
chapter.

Pictorially an example of this RCBD is shown in Table 5.1.4.

TABLE 5.1.4

Layout for RCBD Cardiac Valve Example

Pulse rates

1 3 2 1 4 2 3 5 4 6 6 5
I
Run c i) ( 2 ) ( 3) ( 4) ( 5) ( 6 ) ( 7) ( 8 ) ( 9) (1 0 ) (ID (1 2 )

5 2 3 4 2 6 1 4 5 3 6 1
(/> 4
<D
Run (13) (14) (IS) (16) (17) (18) (19) (2 0 ) (2 1 ) (2 2 ) (23) (24)
4&->
<D
>*4
r- 4 2 1 6 4 3 5 2 1 6 5 3
a
> 1
Run (25) (26) (27) (28) (29) (30) (31) (32) (33) (34) (35) (36)

6 1 5 4 1 2 3 3 5 6 2 4
3

Run (37) (38) (39) (40) (41) (42) (43) (44) (45) (46) (47) (48)

The model for this example of a RCBD in Table 5.1.4, if there

are no correlations among errors, is:

I'ijk - » * V i . « (1) . Pj . V P ^ . , (..)k (5.1.4)

i = 1, 2, 3, 4 j = 1, 2, ..., 6 k = 1, 2
134 5. RANDOMIZED COMPLETE BLOCK DESIGN

where

th th
y . = k maximum flow gradient from the j pulse rate of
1 JK th
the i valve type

y = overall mean

Vj, = effect of the i ^ valve type (fixed)

6 ... = restriction error(random) caused by all 1 2 pulse

th
rates being run on the i valve type before running
any other valve type without reseating the valve type
(completely confounded with the effect of the i
2
valve type) NID (0, c^)

Pj = effect of the j**1 pulse rate (fixed)

* interaction effect of the valve type with the j**1

pulse rate
th th
er* = rand°m error caused by the k observation of the j
th 2
pulse rate in the i valve type, NID (0,a )

The corresponding analysis of variance for this experiment is

set out in Table 5.1.5. It is seen that the mean square for the
restriction error (which is really a between block error and
intuitively correct) would be the denominator to test valve types.
In this case, however, there are zero df and no sum of squares for
the restriction error;consequently there is no mean square and
there is no test for one of the most important factors, valve types,
when this design is used even if the correlation assumption is
correct. Hence if the experimenter uses Eq. (5.1.4) in his
preliminary outline of the analysis from this proposed design he
will know he should change the design of his experiment before he
has run any part of it.
5.1 BLOCKS FIXED 135

TABLE 5.1.5

ANOVA Using Equation (5.1.4)

Source df EMS

Valve types (V^) 3 a2 + 12a2 + 12<|>(V)

Restriction error 0 a2 + 12a2

Pulse rates (P^) 5 a2 + 8<KP)

Interaction (VP..) 15 a2 + 2* (VP)

^ ir
Within error (e ^jjjc) 24 a2

Total 47

Example 5.1: Calculation of the ANOVA table for the experiment

described by Eq. (5.1.4). Even though the observations of maximum
flow gradient for each pulse rate were taken randomly for each valve
type (block), the coded values will be shown in the following
ordered array.

Pulse rate (treatments)

(fixed blocks)
1 2 3 4 5 6 Totals

1
;® 7 © j© ,i© s® 104

2
3 ® 15® ‘?© :® 3 © a© 170

3
3 ® 3 ® ;® ;® s ® s® 174

4
!S@ ;s© !® 5© IS© n® 141

Totals 114 91 56 49 111 168 589

136 5. RANDOMIZED COMPLETE BLOCK DESIGN

The ANOVA table for these data follows.

Source df SS MS EMS F

Valve types (V) 3 261.90 87.30 a2 + 12a2 + 12$(V) None

Restriction error 0 None a2 + 12 a2

Pulse rate (P) 5 1192.35 238.47 a2 + 8 <KP) 28. T

Interaction (VP) 15 55.73 3.72 a2 + 2*(VP) < 1

Within error 24 199.50 8.31 a2

Total 47 1709.48

*Significant at the 0.05 level.

Thus one would conclude that there is a significant difference among

pulse rates but no valve-pulse rate interaction. A multiple
comparison test may be carried out to find out which pulse rates are
2
different from one another. The best estimate of a would be a
pooled mean square obtained as follows:

-2 199.50 + 55.73 255.23

a " 24+15 " 3 9

= 6.54

Thus the standard error to use for the multiple range test is

^6754
= /0782 = 0.91
8

The calculation of the sums of squares are as follows:

E J J y?., = 122 + 72 + 202 + ... + ,172 = 8937

i j k ^

T• •• = 589 CT = - = 7227.52
S.1 BLOCKS FIXED 137

SS total = £ E Zy?.k - CT= 1709.48

SS valves = t4 £ T? - CT = -rA (1042 + 1702 + 1742 + 1412)

12 i** 12 v '

- CT

= 7489.42 - CT = 261.90

SS pulse rates = 4 £ T2. - CT = 4 (1142 + ... + 1682) - CT

o *3 * °

= 8419.87 - CT = 1192.35

SS VP s i £ T?. - CT - SS V - SS P
2 lj*

= | (192 + 132 + ll2 + ... +382) - CT

- SS V - SSP

= 8737.50 - 7227.52 - 261.90 - 1192.35

= 55.73

SS error = SS total - SS V - SS P - SS VP

= 199.50

5.1.3 RCBD When Interaction Is Present and One Observation Per Cell

In some experiments there is only one experimental unit per

treatment and only one observation on that experimental unit in a
block. If, in addition, the interaction between blocks and treatments
cannot be assumed zero the model will be

y.. = ii + B. + + T. + BT. . + (5.1.5)

i (i) j (ij)

i = 1 , 2 , ..., b j = 1 , 2 , ..., t

where the usual definitions of the components hold. The corresponding

ANOVA is given in Table 5.1.6.
138 5. RANDOMIZED COMPLETE BLOCK DESIGN

TABLE 5.1.6

ANOVA Using Equation (5.1.5)

Source df EMS

Blocks (B^) (b - 1 ) CT2 + ta^ + t<J>(B)

Restriction error (6 ^ 0 ff2 +

Treatments (T^) (t - 1 ) a2 + b*(T)

Interaction (BT^) (b - 1 ) (t - 1 ) a2 + ♦(BT)

2
Within error 0 a

Total (bt - 1 )

If treatments are tested by the interaction mean square, the

test is conservative. By conservative we mean that the denominator
of the F test may be somewhat larger than it should be. As a result,
the actual a level of a conservative test may be smaller than the
selected a level. If treatments turn out to be not significant, it
would be desirable to use a nonadditivity test or to partially
replicate in the original design.

5.2 BLOCKS RANDOM

Many experimenters would like to have the blocks in the

experiment represent all possible blocks or be random. To do this
they must be careful in the selection of the blocks so that what
they want in their inference actually exists in the experiment.
For example, if a chemist wants the results from an experiment run
in his laboratory to apply to all chemistry laboratories in the
world, he must verify these results under a large enough random
sample of laboratories over the world. The chemist must be the one
to determine how many and which laboratories are to represent all
laboratories because he must be concerned about his inference space
before the scientific investigation is complete.
5.2 BLOCKS RANDOM 139

In this case (blocks random) even if interaction between blocks

and treatments exists there is a direct test for treatments when
there is only one experimental unit for each treatment in a block.
The general model for the analysis of b blocks and t treatments is

y. . = y + B. + 6 ... + T. + BT. . + e r . (5.2.1)

13 1( 1) 3 13 ( 13 )

where the symbols are defined as before except that B^ is random

and BT.. is mixed.

TABLE 5.2.1

ANOVA Using Equation (5.2.1)

Source df EMS

Blocks (B) b - 1 ' a2 + ta6 +

Restriction error 0 a2 +
ta6

Treatment (T) (t - 1 ) a2 +
°BT + b<KT)
2 2
B x T (b - 1 ) (t - 1 ) a +
aBT
Within error 0 a2

Total (bt - 1 )

2
where: cr^ is the ’’mixed” interaction component and cannot be
estimated with only one experimental unit per treatment, unless one
is willing to use a test for nonadditivity.

The estimate of the standard error of the treatment mean coming

from this mixed model is:

which is not estimable. The estimate of the standard error of the

difference, however, is estimable and is:
140 5. RANDOMIZED COMPLETE BLOCK DESIGN

which is the usual basis for making individual or multiple comparisons

after treatments are declared significant in the analysis of variance.
A derivation for this type of standard error is given in Appendix 12.

Let us now extend the mixed model case discussed in Example 2.7
to factorials. Consider the model

y. „ = y + B. + 6 ...
'ljkfc l (l)

+ p* + BPu + MP

where i = 1, 2, 3; j = 1, 2; k = 1, 2, 3, and Jl = 1, 2, 3

^ijk£ = firinS time t h e switch in the i ^ block or replicate

made up of the metal, k ^ initiator, and 5,^
pressure and all other symbols are defined as before,
except that blocks are random and all interactions
with blocks are similar to BT.^ in Eq. (5.1.5).

Allowing blocks to be random and metals, initiators and

pressures to be fixed, the appropriate denominator for the F-tests
for all fixed effects and their interactions will be the
corresponding interaction with blocks mean square. If only three
replicates are run there will be only 2 degrees of freedom for the
error to test metals. Hence the investigator must be careful to
outline the analysis before running the experiment. In this case
he almost surely would want more than two degrees of freedom for
error.

Some experimenters know before the experiment is run that all

the interactions with blocks or replicates will be about the same
size. If this is the case here, there would be 34 degrees of
freedom for a common pooled error. The degrees of freedom for the
corresponding pieces of the error are:
5.2 BLOCKS RANDOM 141

Source: Error = BM + BI + BMI + BP + BMP + BIP + BMIP

df: 3 4 = 2 + 4 + 4 + 4 + 4 + 8 + 8

If not as much information is available to the experimenter, an

optional procedure is to use the sometimes pooling procedure with
a = 0.25 (Bozivich et al., 1956), after the data are obtained hoping
enough mean squares will be poolable to provide sufficient degrees
of freedom for the various tests to give reasonable power.

Another example in which blocks are random occurred in a

quantitative genetics study. Drosophila were used to investigate
chromosomal effects on body weight. The entire experiment had four
genome samples drawn at random which formed the blocks or
replications. Within each genome sample or block, eight heterozygous
states of chromosomes, acting as treatments, were examined. The
model is:

y. . = y + G. + + H. + GH. . + z f . .v (5.2.3)
'ij i (i) j ij (ij) v J
i = 1, 2, 3, 4 j = 1, 2, 8

with the usual definitions of the terms in the equations.

The analysis is given in Table 5.2.2.

TABLE 5.2.2

ANOVA Using Equation (5.2.3)

Source df EMS

2
Blocks (genome sample) 3 a + 8 a2 ♦
8aG
Restriction error 0 a2 + 8 a2

a2 +
+

Treatments (heterozygous states) 7

Q

4*(H)

Interaction (genome sample by 9 o

heterozygous states) 21 a +
aGH
2
Within error J) a

Total 31
142 5. RANDOMIZED COMPLETE BLOCK DESIGN

The test on heterozygous states is against the interaction

2
source mean square with or without = 0 .

Example 5.2: Consider the case of Eq. (5.2.3) when blocks are
genome samples and treatments are heterozygous states. The coded
value of Drosophila body weight is as follows:

Heterozygous states (treatments)

(random block) 1 2 3 4 5 6 7 8 Totals

1 16 13 8 7 15 3 14 7 83
2 15 15 10 6 13 6 15 10 90
3 20 18 14 10 22 9 20 14 127
4 12 14 9 5 13 1 12 5 71

Totals 63 60 41 28 63 19 61 36 371

The ANOVA table for these data follows and :

is based on the
analysis shown in Table 5.2.2.

Source df SS MS EMS F
2
Genome sample 3 218.59 72.86 a + ♦ 8 a2 None

Restriction error 0 None a2 +

S° l
2
Heterozygous states 7 543.97 77.71 a2 + 46.4*
aGH + 4<|>(H)
2
Interaction 21 35.16 1.67 a2 + None
aGH
2
Within error _0 None a

Total 31 797.72

*Significant at the 0.05 level.

Thus one would conclude that there is a significant difference among

States but as indicated in the test there is no exact test for
Genome Samples.
5.3 ALLOCATION OF EFFORT 143

Multiple comparison tests among heterozygous states means may

be carried out using

s- = A . 67/4 = C 4 2 .

The sums of squares calculations are as follows:

Z E y?. = 162 + 152 + ... + 52 = 5099

. . 1 1
i J J
2
E J y.. = 371 CT = ^ I L - = 4301.28
. • /il 32
1 3 J

SS total = H y?. - CT = 5099 - 4301.28 = 797.72

• • ij
i I

4
4
SS blocks = 1 V T? - CT = 1 (832 + 902 + 1272 + 712)
8 .I
4-1 i* 8
i=l

- CT = 218.59

g
SS treatments = -j I T2^ - CT = i (632 + ... + 362)

- CT = 543.97

SS interaction = SS total - SS blocks - SS treatments = 35.16

5.3 ALLOCATION OF EXPERIMENTAL EFFORT IN RCBD

For all RCBD there may be more than one observation for each
treatment in each block if there is enough time (and/or money) and
a reason from an efficient design point of view to handle the extra
observations. An extra observation may just be another measurement
of the same experimental unit, possibly by another technician. This
type of observation may not have much value for the inference because
the variation measured may be only an operator and measurement error
variation. Under some circumstances it is of value to have this
144 5. RANDOMIZED COMPLETE BLOCK DESIGN

information, but to allow the experimenter to make wider inferences,

over all possible experimental units, the experimenter must repeat
the treatments on other experimental units for his additional
observations.

This concept of the experimenter knowing his inference space

before he decides the method of obtaining repeated observations is
basic for designing any experiment. An example of this idea occurred
in an agronomy experiment in which the cellulose content of alfalfa
was measured from plants in each pot (the experimental unit). The
blocks (locations in the greenhouse) were fixed and the treatments
were fixed so the inference would be made for those blocks and
treatments only. The repetitions within each treatment-block
combination could be just another reading for the same experimental
unit (that is dividing the material from the same experimental unit
and getting two observations) or randomly assigning the same
treatment to another pot in the same block and getting another
observation. The variation within (B - T) combinations for the
former design should be smaller than, or at most equal to that
variation for the latter design. The inference for the former
should be more limited than for the latter. The model for the
former design is

y + B. +5
yijkJl l

where

i = 1 , 2 , ..., b j = 1 , 2 , ..., t k = 1 £ = 1 , 2

th
y. 0 = cellulose content of the JI observation of the
ij t!h tlr
plants in the k pot from the j treatment in the
i location in the greenhouse
1.L
B^ = effect of the i location

^(i) = restr:*-ct:^on erro**

Tj = effect of j**1 treatment
5.3 ALLOCATION OF EFFORT 145

BT. . = effect of the interaction of the i**1 location with

13 th
the j treatment

pr.... = random error due to the pot in the i**1 block and
.th
j treatment
th th
nr* •V'w = random error due to the I observation in the k
th th
pot in the j treatment in the i location.

The analysis of the data from the former design (observations

from the same experimental unit) is given in Table 5.3.1.

TABLE 5.3.1

ANOVA When Pots Are Not Repeated or Using Equation (5.3.1)a

Source df EMS

Blocks (B) b - 1 a2 + 2 a2 + 2ta2 + 2t$(B)

n P
Restriction error 0 a2 + 2 a2 + 2tol
n P 6

Treatments (T) (t - 1 ) a2 + 2 a2 + 2b* (T)

n P
Interaction (B x T) (b - 1 ) (t - 1 ) a2 + 2 a2 + 2*(BT)
n P
Between pots within 9 9
(B - T) combinations 0 a2 + 2 a2
n P
Between observations 9
within pots bt a
n
Total 2 bt - 1

2
= variance component between samples within experimental
units (pots) (a type of sampling error)
2
= variance component between experimental units (pots)
treated alike
2
cjg = variance component due to extraneous variables on the
blocks
146 5. RANDOMIZED COMPLETE BLOCK DESIGN

Hence there is no estimate of the error to test treatment

effects and the interaction. Treatments can be tested against the
interaction mean square if <J>(BT) is zero or a conservative test may
be run if <#>(BT) is not negligible. The model for the latter design
is:

y. .. = y + B . + 6 , . + T. + BT.. + + i]r (5.3.2)

7ijkJl K l (l) j ij (ij)k (ljk) I

i = 1 , 2 , ..., b j = 1 , 2 , ..., t k = 1 , 2 1=1

and the components are defined the same as for Eq. (5.3.1) except
that k = 1 , 2 here and I = 1 only here.

The analysis for the latter design (two experimental units

randomly assigned to each treatment within each block) is given in
Table 5.3.2.

TABLE 5.3.2

ANOVA When Pots Are Repeated or Using Equation (5.3.2)

Source df EMS

2
Blocks (B) (b - 1 ) a + a2 + 2ta2 + 2ti(>(B)
n P
2 2
Restriction error 0 a + a2 + 2 ta^
n P 6

Treatments (T) (t - 1 ) a2 + a2 + 2b<J>(T)

n P v '
Interaction (B x T) (t - 1 )(b - 1 ) a2 + a 2 + 2<f>(BT)
n p 3
Between pots within
2 2
(B - T) combinations (bt) o + o
n p

Total 2 bt - 1

Hence the tests on treatments and the interaction of blocks by

treatments are possible since the appropriate error is estimated
from between experimental units within the block-treatment
2
combinations. If an estimate of a^ alone were wanted, more than one
sample per treatment combination must be taken on at least a few
experimental units.
5.3 ALLOCATION OF EFFORT 147

This general approach of evaluating the experiment through the

outline of the analysis before the experiment is run to guide the
experimenter in designing his experiments should always be carried
out, keeping the inference space clearly in mind.

An example of a RCBD in chemistry is that if chemists wish to

make inferences about five methods over all laboratories in the
United States, a random sample of the admissible laboratories may
be obtained.

It should be made clear to the reader at this point that this

is an example in which the experimenter is not interested in the
blocks, i.e., the laboratories per se, and emphasis is made on the
tests of treatments (here methods) only.

Let us say that twenty laboratories are selected over the

United States at random and the same CRD is used at each laboratory.
The effect of laboratories has location effects, the environmental
conditions around the experiment, personnel, and so on confounded
with it. It is impossible to enumerate all the causes of variation
from laboratory to laboratory, but it is assumed that the factor,
laboratories, has all these in it.

Within each laboratory the same five fixed methods to develop

a product are used. Each method is to be run ten times in each
laboratory. Assuming the 50 setups of equipment to obtain
information on the five methods each run ten times are completely
randomized, there is really only one place in the design that
randomization occurs once the twenty laboratories are known. Of
course that place is within each laboratory. The random laboratories
merely allows a broad inference space for the fixed methods.

The model for this experiment is:

148 5. RANDOMIZED COMPLETE BLOCK DESIGN

where
the variable to be analyzed from the setup of
yijk
the j method in the i laboratory

u = overall mean

L. the effect of the i ^ laboratory (random)

l
the i*'*1 laboratory or restriction error due to the
6 (i)
restriction on randomization of the methods in the
laboratories

M. the effect of the method (fixed)

3
LM. . the interaction effect of the i ^ laboratory with
11
the j**1 method

within error, the random effect of the error

£(ij)k
due to the k**1 setup within the i^*1 laboratory and
the j**1 method

TABLE 5.3.3

ANOVA Using Equation (5.3.3)

Source df EMS

2
Laboratories (L) 19 0 + 50o2 + 50a2
o L
Restriction error 0 a2 + 50a2

2
Methods (M) 4 a + 10a + 200<)>(M)

L x M 76 a2 +

Within error 900 a2

Total 999

As indicated in Chapter 2 on mixed models the test on the fixed

main effect uses the interaction mean square since in this case the
inference is made to all the laboratories from which the 20 were
5.3 ALLOCATION OF EFFORT 149

drawn. An inference that is broad must contain in its error a

reflection of the sampled material. In our previous example for
only one laboratory the within methods source of variation was used
as the experimental error reflecting a very narrow inference, namely
for that one laboratory, or if it is known there are no differences
between laboratories in their effect on the systems, the inferences
would be over all laboratories.

Another feature of this design is the allocation of time and

observations in experimentation. Is it necessary to have ten
observations per laboratory-method combination? Since it is so
expensive to investigate laboratories, is it permissible to reduce
the number of laboratories?

The answers to these questions must come from the interests of

the experimenter. If the experimenter is interested in estimating
2 2 2 2
o^, he must have an estimate of o and To estimate a with 900
degrees of freedom seems like a waste of observations unless it
costs almost nothing to make the experimental setups (which is most
2
unlikely). An estimate of a with about 100 degrees of freedom is
usually more than sufficient for most practical work. This means
that two observations per cell would be sufficient here. Of course
this assumes the variances are homogeneous because this estimate
comes from the one hundred combinations of laboratories and methods.
2
For estimating a^, there must be replication of the overall
experiment at each laboratory.

As far as the number of laboratories is concerned, there should

be an ample number to make the inferences on the methods for all
the laboratories in the population. If there is very little
variation from laboratory to laboratory and the interaction of
laboratory by methods is small, there is no need for a huge sample
of laboratories. In devising a sampling plan for selecting
laboratories, the experimenter should use as much information as he
can on differences between laboratories to select the sample size.
In this case, sequential experimentation seems like a very good
150 5. RANDOMIZED COMPLETE BLOCK DESIGN

procedure. That is, take a relatively small random sample of

laboratories, say seven, and estimate the variance components. If
the methods mean square seems large relative to the interaction mean
square but the experimenter can detect no statistical difference
among methods, the experimenter should take more laboratories. A
guide to the number of laboratories would be to find the number of
degrees of freedom needed in the interaction to show the F-ratio
obtained in the first experiment to be significant and calculate the
number of additional laboratories required to get a significant
effect.

Consider the following numerical example: Let us use five

laboratories, three methods and four experimental setups per
laboratory-method combination for the first part of the experiment.
The analysis is given in Table 5.3.4.

TABLE 5.3.4

ANOVA For Example Above

Source df MS EMS

Laboratories (L) 4 70 a2 + 12 a2 ♦
12aL
Restriction error 0 a2 + 12 o2

Methods (M) 130 a2 +

4^ m + 20<KM)
2

L x M 8 40 a2 +

Within error 45 25 a2

Total 59

The F-test for methods is:

F = = 3 25
2,8 40

Using a = 0.05, theoretical 8 = 4.46. T^e *atio in the experiment

5.4 RELATIVE EFFICIENCY 151

looks large, but it is difficult to draw conclusions. How many more

laboratories would be suggested? A theoretical = 3.23, so 16
more laboratories should be drawn at random which would make a total
of 21 laboratories in the experiment. If similar results were
obtained for the mean squares (there is no absolute assurance that
this will happen), the combined analysis would allow a conclusion
that the methods were different.

If the results from part one of the experiment using five

laboratories were drastically different from the second part using
a different group of 16 laboratories, the experimenter must
thoroughly investigate the two sets of laboratories to explain the
reason for the large difference.

Another possibility from the sample of five laboratories exists,

and that is there is already a significant difference in methods
obtained with the five laboratories. If the research worker has
good reason for making the inferences to all laboratories, there may
be no reason to go further.

The final possibility for this randomized complete block design

with more than one experimental unit per combination is that the
mean square for methods is no larger than the interaction mean square
and the experimenter has good reason to believe that all assumptions
for the model are met. Under these circumstances the experimentation
would usually stop with the conclusion that the methods are not
different.

5.4 RELATIVE EFFICIENCY OF DESIGNS

After understanding the two basic designs discussed so far it

is important to compare the efficiency of one design relative to
the other. In this case let us compare the efficiency of the
randomized complete block relative to the completely randomized
design (Fisher, 1966; and Cochran and Cox, 1957, pp. 31-34). There
are various ways of doing this but we will use the ratio,
152 5. RANDOMIZED COMPLETE BLOCK DESIGN

(n1 + l)(n2 + 3) s2

(n2 + l)fn1 + 3) s2

where for our problem

?
is the estimate of the error mean square (from the
interaction mean square) to test treatments from the
randomized complete block design
2
s2 is the estimate of the error mean square to test
treatments from the completely randomized design

n^ is the number of degrees of freedom for the error in the

randomized complete block design

n^ is the number of degrees of freedom for the error in the

completely randomized design

Problem 5.4.1. A company wanted to increase light intensity

of its photoflash cartridge. The variable to be analyzed was
a function of light intensity.
The four treatments:
(1 ) 1 / 8 inch thick wall with the ignition point at the
end of the cartridge
(2 ) 1/16 inch thick wall with the ignition point at the
end of the cartridge
(3) 1/8 inch thick wall with the ignition point at the
center of the cartridge
(4) 1/16 inch thick wall with the ignition point at the
center of the cartridge
were used in the experiment.
Five random batches of the basic formulation used in the
cartridges were made up. From each batch three cartridges with
the same treatment were constructed so in all there were 1 2
cartridges made from each batch.
(a) Describe the various places in the experiment that
randomization should occur.
(b) Show the layout of the experiment.
(c) Carefully describe the inference space.
(d) Show the appropriate model and analysis resulting
from the randomizations and inference space. Describe each
symbol.
(e) Compare the inference space of this design with a
design in which only one batch with fifteen cartridges for
each treatment was used.
(f) Under what conditions would the one batch design be
preferred.
5.5 REFERENCES 153

Problem 5.4.2. An agronomist was interested in the effect of

four different sprays on control of weeds in com. The
variable to be analyzed was a function of the remaining weeds
1 week after application of the sprays.
Five locations on the research farm were used to run the
experiment and three randomly chosen plots in each location
were used for each spray.
(a) Describe the randomization procedures and inference
space for the experiment.
(b) Show the appropriate model and analysis, describing
the symbols.
(c) Comment on the design and discuss improvements, if
possible, using another inference space and model.

5.5 REFERENCES

Addelmen, S. Amer. Statistician 23:35 (1969).

Anderson, V. L. Biometrics 26:255 (1970).
Beeson, J. A simulator for evaluating prosthetic cardiac valves.
Unpublished M.S. thesis, Purdue University Library. (1965).
Bennett, C. A. and Franklin, N. L. Statistical Analysis in
Chemistry and the Chemical Industry. Wiley, New York, 1954.
Bozivich, H., Bancroft, T. A. and Hartley, H. 0. Ann. Math. Statist.
27:1017 (1956).
Cochran, W. G. and Cox, G. M. Experimental Design, 2nd ed., Wiley,
New York, 1957.
Fisher, R. A. The Design of Experiments, 8 th ed., Hafner, New York,
1966.
Hicks, Charles R. Fundamental Concepts in the Design of Experiments,
2nd ed., Holt, Rinehart and Winston, New York, 1973.
Ostle, Bernard Statistics in Research, 2nd ed., Iowa State
University Press, Ames, Iowa, 1963.
Wilk, M. B. Biometrika 42:70 (1955).
 Chapter 6

NESTED (HIERARCHICAL) AND NESTED FACTORIAL DESIGNS

The designs presented in Chapters 4 and 5 are more conventional,

a bit more easily recognized and used more frequently than the ones
to be presented in this chapter. For the designs in this chapter
the factorial structure is not complete and to recognize the features
and how to set up reasonable linear models for the analysis of the
data from these designs is not always obvious. Also, in this
chapter, we use Latin letters rather than Greek letters to represent
the well-defined random components in the mathematical models.

6.1 NESTED (HIERARCHICAL)

In some experiments the levels for a given factor are all

different across the levels of the other factors. Frequently the
levels are chosen at random for each factor, but this is not a
necessary condition for the so-called "nested design." In fact,
there can be random, mixed, or fixed models in the hierarchical
arrangement. It should be emphasized that this is really another
type of the CR design; or if there is a restriction, the associated
error is understood to be part of the appropriate component.

6.1.1 All Factors Random

An example of a nested design (often called multistage

sampling) involves human beings in the United States. States are
sampled at random, counties are sampled at random in the states,
towns and cities are sampled at random in the counties, and finally
households may be sampled at random in the towns or cities.

154
6.1 NESTED (HIERARCHICAL) 155

Actually, the sampling procedures are quite a bit more complicated

than this, but a linear model to analyze data from the above design
could be the following:

y. .. . = y + S. + .n, + H,. . (6.1.1)

M l (i)j (ij)k (ijk)*

i = 1 , 2 , ..., s j = 1 , 2 , ..., c k = 1 , 2 , ..., t I = 1 ,2 , ..., h

where
y = the overall mean

S. = the effect of theitk state

i
th th
C(i)j = the effect of the j county in the i state

T r.... = the effect of the k**1 town in the jthcounty inthe

.th
l state

*i mi = the effect of the household in the ktk town in

(ljk)* th th
the j county in the i state

Notice that the example shows the impossibility of the level of any
factor such as countiesbeing the samefor the levels of the factor,
state. Since there are never the same counties within the various
states in this investigation, it is impossible to obtain an
interaction between counties and states. This concept of never being
able to obtain an interaction is always present in a nested design.

Table 6.1.1 shows the ANOVA resulting from Eq. (6.1.1).

TABLE 6.1.1

ANOVA Using Equation (6.1.1)

Source df EMS

, 2
°l + hoT + htoc
2
States s - 1 itcas
, 2
Counties in states s(c - i) haT +
•2 * htac
Towns in counties in states sc(t - i)

Households in ... states sct(h - i)

4
Total scth- 1
156 6. NESTED AND NESTED FACTORIAL DESIGNS

In this table the number of levels are constant in the subclasses.

In most actual sampling and design of experiments problems the
number of levels may be quite unequal. In this case the calculations
of sums of squares is still straightforward, but the calculation of
the coefficients on the variance components may be quite involved
(Anderson and Bancroft, 1952, p. 327).

Example 6 .1: Determination of the components of variance for

the model

y. = u + S. + T r.. . + H r....
ijk H l (i);j (ij)k

i = 1, 2, 3 j = 1, 2 k = 1, 2, 3, 4

where the terms in the model are as defined in Eq. (6.1.1). Two
methods of calculation are presented: one gives the sum of squares
for nested factors directly, whereas the other utilizes the method
of calculation for a factorial design and then combines appropriate
sums of squares to obtain the correct values. The data for this
example are given in the following tabulation.

State

1 2 3

Town Town Town

1 2 1 2 1 2

Household
1 10 7 6 6 15 12

2 13 12 5 12 18 15
3 16 11 9 7 20 18
4 12 9 3 10 19 16

Totals 51 39 23 35 72 61

T. = 90 T0 = 58 T-
1 *• 2 *• 3* • = 133

T• • • = 281
6.1 NESTED (HIERARCHICAL) 157

The ANOVA table is:

Source df SS MS EMS
2 2
States 2 354.08 177.04
4 + 4 + *4
2
Towns/states 3 51.13 17.04
4 + °T
Households/towns/states 18 107.75 5.99
•2
Total 23 512.96

where
2
i
1 3
n 0
/ T•••
SS states = j ) T. - «--- =-
2x4 i** 2 x 4 x 3
1=1

= i (902 + 582 + 1332) - = 354.08

1 3 2 , 2
SS towns/states = j I I T. - £ T...
i=l j=l J i

= (i T2 . At2 )
>.*■4 J=1
1=1 i ij-
J
8 i-*J

= sum of squares for towns within state summed

over states

1 2 2 2 1 2 2
= j (51Z + 39 + ... + 6 1Z) - j (90Z + 58

+ 1332)

51.13

2 1 2
SS households/towns/states = Z EZ y...- 7 Z Z T,.
i .k ^ j k 4 . . ir

= 102 + 132 + ... + 182 + 162

1 2 2 2
- ~ (51 + 39 + ... + 61 )

= 3803 - 3695.25
= 107.75
158 6. NESTED AND NESTED FACTORIAL DESIGNS

2 •••
SS total * E n y...--
i j k 1JK
= SS states + SS towns/states + SS households/
towns/states

= 3803 - 3290.04

= 512.96

In order to determine the method of calculating the above

analysis of variance table using a factorial computing procedure we
combine sums of squares in the same manner that one has to combine
degrees of freedom for the model

y. ., = \i + S. + T. + ST. . +
ijk l j ij (ij)k

to obtain the correct degrees of freedom for the model

y. ... = u+ S. + T... . + H.. ...

J ijk H i ( 1)3 ( i 3) k

In the nested model it is impossible to have a state by town

interaction and so it seems reasonable to combine the degrees of
freedom for the fictitious "towns" with 1 df with the "town by state
interaction" with 2 df to obtain the 3 df for towns within states.
Checking this reasoning gives 1 df + 2 df = 3 df from the factorial
model which in turn is the correct degrees of freedom for towns
within states. Consequently we combine the same sums of squares.
One may prove this relationship algebraically, but this will not be
shown here. In order to demonstrate this we calculate the two sums
of squares for the factorial model.

T2
SS "towns" = [(51 + 23 + 72)2 + (39 + 35 + 61)2] -

= 3295.08 - 3290.04 = 5.04

SS "T x S" = I (512 + 232 + ... + 612) - CT - SS towns - SS states

= 3695.25 - 3290.04 - 5.04 - 354.08

= 46.09
6.1 NESTED (HIERARCHICAL) 159

Note that 46.09 + 5.04 =» 51.13, which is the same as that obtained
for the towns within states sum of squares. Calculation of sums of
squares for any nested type of sums of squares in the nested
factorial models may be done in a similar fashion.

The estimates for the individual components of variance are as

follows (refer to the above ANOVA table)

' 5'99

a* = j (17.04 - 5.99) = 2.76

of = i (177.04 - 17.04) = 20.00

b O

The estimate of the total variation of a single observation

becomes the total of these three variance components (Eisenhart,
1947). Hence ajotal = 28.75. It follows that = 0.70 of the
total variance is estimated to be due to variation among states.
This indicates that if further information is desired one should
concentrate more on states and less on towns in states.

Example 6.2: In a genetics study on fruit flies, the variance

components of body weight due to males, females, bottles in which the
flies are kept and offspring was investigated. The mating system
involved selecting four males and twelve females at random from each
of four genome samples, mating three females to each male in single­
pair matings to retain identity of the female and dividing the eggs
laid into two bottles for each mating. After the progeny became
adult flies, five picked at random from each bottle were weighed
individually.

Hence the design was completely hierarchical and a reasonable

linear model is

y. .. „ = y + G. + M.... + F,. ... + B,....,, + (6.1.2)

'ljk&m H i (i)j (ij)k (ijk)A (ijk£,)m

i = 1, 2, 3, 4 j = 1, 2, 3, 4 k = 1, 2, 3 I = 1, 2 m = 1,

where each symbol is easily identified from the explanation above.

160 6. NESTED AND NESTED FACTORIAL DESIGNS

Problem 6.1.1. Considering all factors random, show the

2 2 2 2 2
variance component estimates for a^, ap, and a^ in the
preceding example on fruit flies.

Problem 6.1.2. In the actual experiment there were only 3

males for genome 4; two females mated to the second male for
genome 1, first male from genome 2, fourth male from genome 3;
and only one female mated to the first male from genome 1. In
addition only three progeny existed for the mating of male 1
with female 1 in genome 1 and male 2, with female 3 in genome
3; and only one progeny existed for mating of male 1 with
female 3 in genome 4.
Write out the adjusted coefficients of the variance
components for the actual experiment (Anderson and Bancroft,
1952, p. 327).

6.1.2 Fixed and Random Factors

An example in which some factors are random and one is fixed

occurred in an aluminum alloy development problem. An aluminum
company had four (fixed) alloys in which the chemistry was different.
All the properties except strength had been evaluated and an
experiment to examine this variable was wanted.

One proposal was to make four ingots, one with each chemistry,
and compare the strengths of the finished product. If this were
done the inference could be made to this one ingot for each
chemistry made from only one heat (batch of molten aluminum). If
each ingot could possibly represent an adequate random sample of
ingots for each heat, the inference could be made to all ingots from
that one heat with that chemistry. To infer indefinitely over heats
one must replicate the experiment by having at least one more heat
run per chemistry or assume the variation between heats was no
greater than the variation between ingots within heats. This latter
assumption, however, is usually very poor for problems of this type.

The cost of heats was quite large and the research worker
decided to take three ingots per chemistry chosen at random from
each heat-chemistry combination. Further he took four random
samples of the material in each ingot to measure the variation
6.1 NESTED (HIERARCHICAL) 161

between pieces within ingots. The appropriate model and ANOVA for
this experiment is:

y..ln = u + C. + H r... + I,, .v. + e,..n . (6.1.3)

ijki H l (i)j (i3 )k (ijk)£

i = 1, 2, 3, 4 j = 1 k = 1, 2, 3 I = 1, 2, 3, 4

where
y. ., 0 = strength of the ft**1 piece of metal in the k ^ ingot
13 th th
from the j heat with the i chemistry

y = overall mean

C^ = effect of the i ^ chemistry (fixed)

H(i)j = effect of the j heat using the ichemistry. This

is completely confounded with chemistry [random,
NID (0 , a j ) ]

I,.... = effect of the ingot in the jt^1 heat in the i**1

(ij)k 6 2 J
chemistry [random, NID (0, a^)]

er. .0 = effect of the piece of metal in the k ^ ingot in

th th 2
the j heat in the i chemistry [random, NID (0, a^)]

TABLE 6.1.2

ANOVA Using Equation (6.1.3)

Source df EMS

Chemistries (C.) 3 a2 + 4a2 + 12a2 + 12<j>(C)

iJ £ 1 H
2 2 2
Heats in C (H^^) 0 a + 4a^ + 12a„
e I H
2 .2
Ingots in heats (I^j)k^ 8 a + 4a,
e I
2
Pieces in ingots U ^ ^ ) 36 a
e

Problem 6.1.3. Using the EMS results from Table 6.1.2 and the
information in the following tabulation
162 6. NESTED AND NESTED FACTORIAL DESIGNS

Source df MS
Chemistries 3 160
Ingots 8 20
Pieces 36 10

(a) Give a detailed report of the results to the research

worker.
(b) If it costs $1000 to run a heat, $300 to use an ingot
and $50 to examine a piece of metal, recommend a design for the
future and explain to the research worker in the aluminum plant
why you prefer your design over his. How much more would your
experiment cost than the one that was run?

Problem 6 .1.4. An agronomist was interested in finding out the

amount of potassium in corn leaves, not including the stems,
after four different fertilizers had been used in a field of
corn.
(a) If 5 plots of com were used per fertilizer in a
completely randomized manner, show the layout of an experiment.
• (b) If in the 5 plots per fertilizer, 2 stalks of corn
per plot were to be sampled from the experimental area, show
the ANOVA and state the assumptions necessary to test the
effects of fertilizer.
(c) Comment on this design and analysis.
(d) Show the standard error of the fertilizer means.
Refer to Section 2.2.1, Example 2.7, and Appendix 12.

6.2 NESTED FACTORIAL

The concept developed in Section 6.1 that some experiments are

designed such that the levels of a factor cannot be the same across
the levels of another is extended in this section. In addition,
however, another factor or factors may have the same levels across
other factors and be "factorial” to these factors in the experiment.
This mixture of nesting and factorial structure is sometimes called
nested factorial (see Hicks, 1973, Section 11.4).
6.2 NESTED FACTORIAL 163

6.2.1 Cardiac Valve Experiment

To appreciate the concept of this design let us return to the

example on prosthetic cardiac valves given in Chapter 5. A more
reasonable design, provided that the within cell errors are not
correlated, is to repeat the tests on the valve types at random and
to run each of the six pulse rates only once on each valve
(occurrence) for a given valve type. To further clarify this design,
consider the four valve types as treatments and completely randomize
the order that the valve types or treatments are run on the machine
so that each valve type occurs twice, each time with a different
valve of the same type. Notice that each occurrence (valve) is
still a block as defined in Chapter 5, but the blocks are now nested
within valve types. At this stage, we have a completely randomized
design and one possible randomization is:

Valve types

1 2 3 4

4 5 2 7 6 8 1 3

Order of run on machine

Using the above schematic, after a valve of type 4 is chosen for

the first run, all six pulse rates are run at random on the valve of
valve type 4. Next a valve of type 2 is inserted in the machine
because the #2 run in the schematic indicates this. Of course all
six pulse rates are then run in a new random order using a valve of
type 2. The experiment continues in this manner until the second
valve of the third type is used (#8 in the order of run on the
machine in the schematic), and the experiment is completed.

After the experiment is completed the order of running the

experiment can be ignored and the data may be put into a table such
as Table 6 .2.1.
164 6. NESTED AND NESTED FACTORIAL DESIGNS

TABLE 6.2.1

Arrangement for Cardiac Valve Data

from a Nested Factorial Experiment

Valve types

1 2 3 4

1 2 3 4 5 6 7 8

Pulse rates
1

3
4
5
6

The model for the analysis of the data from this experiment,
assuming the errors are not correlated, is:

= »*+ vi + °(i)j+ 5(ij) + pk + vpik <6-2-

+ op(i)jk+ eajk)
i = 1, 2, 3, 4 j = 1, 2 k = 1, 2, 6

where

^ijk = max*mum fl°w gradient (mmHg) obtained from the j**1

occurrence of the i**1 valve type with the k**1 pulse
rate

\i = overall mean

= effect of the i^*1 valve type (fixed)

th
0 ,.. . = effect of the j occurrence (random valve) in the
^ J th 2
i valve type NID (0, a )
6.2 NESTED FACTORIAL 165

S^j) = restriction error (Anderson, 1970) caused by the 6

pulse rates being run in the j**1 occurrence of the

i valve type. This has zero degrees of freedom
and no sums of squares. It is assumed that
NID (0, ffg)

= effect of the k1"*1 pulse rate (fixed)

th
V P ^ = effect of the interaction of the i valve type with
the pulse rate

OP..... = effect of the interaction of thejt *1 occurrence in

(l);,k th th 2
the i valve type by the k pulse rate, NID (0, a^)
2
e r . ... = within error, NID (0, a ), in this case since there
(ijk) th
is only one observation within the j occurrence of
th th
the i valve type and k pulse rate, there are zero
degrees of freedom

In this type experiment, where the same experimental unit is

used repeatedly for the six pulse rates, a multivariate analysis
may be preferred (Morrison, 1967, Chapter 5) if the errors,
are correlated. There is a conservative test that could be made
for pulse rates and the interaction of valve types by pulse rates
for the situation of correlated errors. In any event the test for
valve types is identical. Assuming that the effect of the
correlated errors is relatively trivial in this particular
experiment, the analysis is as shown in Table 6.2.2.

In Table 6.2.2 it can be seen that there is no test for

occurrences in valve types since there are zero degrees of freedom
for the appropriate source to test it, namely, restriction error.
This is consistent with the RCBD result.
166 6. NESTED AND NESTED FACTORIAL DESIGNS

TABLE 6.2.2

ANOVA Using Equation (6.2.1)

Source df EMS

2
Valve types (V^) 3 a2 + 6 a2 + 6 a2 + 12$(V)

Occurrences in valve 9 9 9
types (O^jj) 4 a2 +
<*« + 6%
Restriction error (6 ^ ^ ) 0 a2 +
6»s

Pulse rates (P^) 5 a2 + ♦ 8 $(P)

£
Interaction (VP^k) 15 a2 + + 2$ (VP)
•i
Interaction (OP^jfc) 20 a2 +
•i
Within error ( e , . 0 0 a2
v (I3kr

To provide information to the reader who will not use

multivariate analysis for correlated error data, one can obtain
conservative tests. The conservative tests require that the degrees
of freedom for the main effect given repeatedly to the same
experimental units [pulse rates (P^) with 5 df] and for the
interactions associated with that factor [(VP^) with 15 df and
(0 Pi( k)) with 20 df] all be divided by the degrees of freedom for
that main effect (5 df). The mean squares are not altered. Hence
the conservative tests for Table 6.2.2 are:

MS(Pk)
(1) Pulse rates using F = j
(i)3 kJ

MS (VP..)
and (2) (VP.k) using F3 = MS(OP
(i)jk)

(see Morrison, 1967, pp. 194-196; Winer, 1971, pp. 522-524; and
Greenhouse and Geisser, 1959).
6.2 NESTED FACTORIAL 167

The present tendency, however, is for experimenters to use the

tests indicated by Table 6.2.2 and not use the conservative tests
for those designs (including split plot type given in Chapter 7) in
which different treatments are used on the same experimental unit.
Even for moderate correlation among the errors, Monte Carlo studies
have indicated this conclusion (Winer, 1971, p. 524).

For those cases in which the experimenter is greatly concerned

about correlated errors because he has used more than one treatment
on each experimental unit, we recommend using both the conservative
and direct or usual ANOVA test indicated by the EMS assuming no
correlation (e.g., Table 6.2.2) in the following manner:

(1) Use the results of the conservative tests if they show

significance

(2) Use the results of the direct ANOVA tests if they show no
significance

(3) Use theory on the size of the correlation for those

results lying between (1 ) and (2 ) with a tendency to use the direct
ANOVA tests unless the correlation is known to be very high.

It has been our experience that whenever a process or

experimental unit is disturbed while the experimenter applies
another treatment, that the correlation of errors becomes small.
Hence we believe that the direct test from the ANOVA is usually
appropriate.

Example 6.3: Given that Table 6.2.1 is filled in with the

following data, determine the corresponding ANOVA table.
168 6. NESTED AND NESTED FACTORIAL DESIGNS

Valve types

1 2 3 4

Occurrence Occurrence Occurrence Occurrence

1 2 3 4 5 6 7 8 rp . ,
Pulse r a t e s ________________________ Totals
1 2 3 4 2 6 5 7 5 34
2 4 4 4 4 5 5 5 4 35
3 5 7 4 3 5 6 6 5 41
4 3 5 5 3 8 10 9 10 53
5 7 7 8 5 9 9 10 11 66

6 6 6 6 7 7 8 8 9 57

Totals 27 32 31 24 40 43 45 44 286
59 55 83 89

+ V. + 0 * VP., + OP r.v ., +£,....

yijk = y l (i)j + 6 (ij) + pk lk (i)jk (ijk)

i = 1 , 2 , 3, 4 j = 1 , 2 k = 1 , 2 , •••, 6

ANOVA

SS MS p(0.05)
Source df EMS F
crit

3 72.25 24.08 a2 + + 6 a2 + 12<j>CV) 13.76* 6.59

vi 6a6

4 7.00 1.75 a2 + 6afi + None

°(i)j 6ao
0 None a2 + 6 °5
6 (ii)

5 105.42 21.08 a2 + 28.11* 2.71

Pk aOP + 84»(P)
VP.1, 15 38.25 2.55 a2 + 2(f)(VP) 3.40* 2.20
lk £ ♦
2
20 15.00 0.75 a + None
0 PCi)jk a0P
2
J) None a
6 (ijk)
Total 47 237.92
6.2 NESTED FACTORIAL 169

where

2 2 2 2 2
no ir _ 59Z + 55 + 83Z + 89 286
i " 12 “ 48

= 1776.33 - 1704.08

= 72.25

2 2 2 2
oo ^ _ 34 + 35 + ... + 57z 286
Fk “ 8 “ 48

= 1809.50 - 1704.08

= 105.42

2 2 2 2 2 2 2
„ _ 21l + 32* + ... + 44* 59 + 55* + 83* + 89*
bb U (i)j 6 “ 12
= 1783.33 - 1776.33

= 7.00

SS vp.k - 52 • 62 ~ ....... 1 7 2 - - SS V. - SS p k

= 1920.00 - 1704.08- 72.25 - 105.42

= 38.25

SS O P (.)jk = t (SS OP(V.))

where
2 2 2
SS OP(V1) = 2 + 3 + ... + 6 -

f s 2 + 82 + . . . + 122 5921 [*272 + 322 5921

' [ 2 12J L 6 12J
= 2.42

2
2 2 2 55
SS OP(V2) = 4* + 2* + ... + 7* -

|*6 2 + 82 + ... + 132 5521 ^ l 2 + 242 552"|

“[ 2 12J I 6 12J
= 5.42
170 6. NESTED AND NESTED FACTORIAL DESIGNS

SS OP(V3) = 62 + 52 + ... + 82 -

fll2 + 102 +... + 152 8321 [*402 + 432 8321

I 2 12J I 6 12J
= 2.75

SS OP(V4) = 72 + S2 + ... + 92 -

J"l22 + 92 + ... + 172 8921 P452 + 442 8921

"I 2 12j ' 6 12j

= 4.41

Therefore

SS OP..... = I SS OP(V.) = 15.00

i= 1 1

SS Total = 323 + 285 + 611 ♦ 723 - 1704.08 (from our work

above)

= 237.92

Problem 6.2.1. Using the results from Example 6.3, investigate

the interaction thoroughly and interpret the overall results of
the experiment carefully. Be sure to consider theconsequences
of installing an incorrect valve into a patient who had a
particular range of pulse rates. In this case assume that the
data used in this analysis have the characteristic of a small
number being desirable.

Problem 6.2.2. Suppose that an experimenter sets up a

prosthetic cardiac valve designed experiment and the correct
model for the analysis is
y. ., * v + V. + 0,... + P. ♦ VP.. + OP,..., +
H 1 (1 ) 3 k lk (i)jk (ljk)t
with the ranges on the subscripts being
i = 1, 2, 3, 4 j = 1, 2 k = 1, 2, 3, 4, 5, 6 I = 1, 2
(a) How many valves (occurrences) were used in this
experiment?
(b) Illustrate the layout of this design and describe how
6.2 NESTED FACTORIAL 171

the randomization must be carried out in order to make this

model valid (consider correlations among errors and where they
could occur).
(c) Show the analysis for this design and compare it to
the analysis shown in Table 6.2.2.

6.2.2 Social Science Application

Nested factorial experiments were recognized early in the

social sciences where human beings are the experimental units. In
these designs frequently subjects (experimental units) are grouped
because of a common characteristic and then quite a few subjects are
used in each group. Since the same individuals cannot appear in
different groups, they are nested within groups. Also, many
experiments then have the subjects do certain tasks (treatments)
and there is a factorial arrangement of groups and tasks. The same
problem, possible correlated errors, exists here as was discussed
in the cardiac valve experiment, Section 6.2.1. In this case the
same subjects do all four tasks.

An example is that three randomly chosen groups of 10 students

each were given a different method of training. After the training
period four tasks were given the students at random (this gives rise
to the restriction error, The dependent variable, y, is the
time to perform the task correctly. One object of the experiment
was to find out whether or not one training method was superior to
the other two methods for all tasks. Another object was to find out
whether or not certain methods of training were better in adapting
students for certain tasks (interaction).

A model that recognizes the restriction error and assumes all

errors are not correlated is

y.
'ljk
= y + M. + Sr.v. +
H i (i)j (1 3 )
+ T. + MT., + ST,.*., +
k lk (1 )jk
.n
(ijk)
(6.2.2)
^ '

i = 1, 2, 3 j = 1, 2, 10 k = 1, 2, 3, 4
172 6. NESTED AND NESTED FACTORIAL DESIGNS

where
th
variable to be analyzed, time to perform the k task
yijk
th th
by the j student in the i training method

y = overall mean

M. = the effect of the i^*1 training method (fixed)

i
xj. i.L
the effect of the j student in the i method
s(i)j
(random)

restriction error caused by the four tasks being

(ij )
done by the j**1 student in the i ^ method

the effect of the k task (fixed)

Tk =
the effect of the interaction of the i**1 method with
^ik3
the k**1 task
th
ST,--,
0 )jk
= the effect of the interaction of the j student in
■hh •t’li
the i method with the k task

within error
e(ijk)

Note that the model and analysis for this example is the same as that
given in Section 6.2.1 (also see Winer, 1971).

6.2.3 Nutrition Application

Recently the nested factorial experiments have been used by

research workers in other areas. An example of such an experiment
occurred in a nutrition study where three (fixed) rations were used.
There were eight (random) hens assigned to each ration and the gain
in weight from week to week was measured on each hen at the end of
each of 5 (fixed) weeks. In addition, the hens were placed in three
cages, one cage for eachration and the number of eggs laid by each
group of eight was kept for the 5 weeks also.

The model, assuming the correlations among all errors are zero,
(Mandel, 1957), for the gain variable to be analyzed follows:
6.2 NESTED FACTORIAL 173

y. ... = y + R. + H... . + 6,. .. + W. + RW., + HW,.. .. + er.... (6.2.3)

7ijk K l (l) 3 (ij) k lk (i)jk (ljk) ^

i = 1, 2, 3 j = 1 , 2, ..., 8 k = 1, 2, ..., 5

where
y. = gain in weight of the j***1 hen on the -ration
th
during the k week

y = overall mean

R^ = the effect of the i*'*1 ration (confounded with cages)

f.u
H... . = the effect of the j hen on the i ration
(i)3
<Sr... = restriction error
(1 3 )
W, = the effect of the week
k
RW., = the effect of the interaction of the i ^ ration and
lk
k week

HW,.. .. = the effect of the interaction of the j**1 hen on the

(l)jk th th
i ration and the k week

z r . .,. = within error

(ijk)

The analysis of variance is given in Table 6.2.3.

TABLE 6.2.3

ANOVA Using Equation (6.2.3)

Source df EMS

2
Rations (R) 2 a2 + So2 + 5a„ + 40<j)(R)
n
2
Hens in Rations (H) 21 a2 + 5a'
♦ *>2
c 2
Restriction error 0 a2 + 6

2 2
Weeks (W) 4 a + °HW + 24<t>(W)
2 2
R x W 8 a + + 81)1 (RW)
°HW
2 2
H x W 84 a +
°HW
2
Within error 0 a
174 6. NESTED AND NESTED FACTORIAL DESIGNS

The estimated standard error of a ration mean is derived in

Part 4 of Appendix 12 and is shown to be

s-
y*.

which can be written as [-^ (mean square hens)]-

Problem 6.2.3. Find the estimated standard error of the

difference between ration means using the method described in
Appendix 12.

The model for the eggs laid variable follows:

yik = y' + R. ♦ «(i) ♦ Wk ♦ RW.k + e(ik) (6.2.4)

i = 1, 2, 3 k = 1, 2, ..., S

y., - the number of eggs laid by all hens on the i ^ ration

IK ^
in the i cage for the k week

y =overall mean

= the effect of the i**1 ration and/or i**1cageand/or i^

group of hens (Note: ration, cages and group ofhens
are completely confounded)

^(i) =restr^ct^on er*or

W, =the effect of the k ^ week
k
RW., = the effect of the interactions of the i ration
1 and/or cage and/or group of hens with the kth week

e,.,. = within error

(lk)

The analysis of variance for the eggs laid variable is given in

Table 6.2.4.
6.2 NESTED FACTORIAL 175

TABLE 6.2.4

ANOVA Using Equation (6.2.4)

Source df EMS

Rations and/or cages 9

and/or groups of hens (R) 2 a2 + 5a^ + 5<KR)
2 - 2
Restriction error 0 a + <5

2
Weeks (W) 4 a + 3*00

R x W 8 a2 + ♦ (MO
2
Within error 0 a

This is a very poorly designed experiment because the analysis

shows that rations cannot be tested even if it were not confounded
with cages and groups of hens. This example is a clear case of the
experimenter not outlining his analysis before the data were taken.

The research workers for this problem should have been aware of
the difficulty in the analysis of lfeggs laid” and have designed the
experiment differently. One major contribution at the design stage
to a good analysis would have been to 11trap nest” the hens and
record individual henfs performance. The analysis, then, would have
been similar to the analysis for the weight varaible, if the
assumption of no correlated errors holds.

6.2.4 An Application in Ammunition Manufacturing

In an ammunition depot an igniter was operating ineffectively

and an experiment was designed in an attempt to locate the trouble.
One variable of interest had to do with whether or not the igniter
functioned properly and another was related to the length of time
it functioned given that the igniter was satisfactory. After
thorough discussion a group of engineers agreed that the trouble
should be found if the following factors were examined together in
an experiment:
176 6. NESTED AND NESTED FACTORIAL DESIGNS

1. Thickness of encloser discs; as thickness increases the

pressure increases and the disc may blow out causing a malfunction.
On the other hand, if the disc is too thin, the pressure is too low
and could also cause a malfunction.

2. Powder lots from manufacturers.

3. Containers within powder lots.

4. Moisture content of ingredients in igniter.

In each of two powder lots (1 and 2) two containers were used,

but there were different containers for lot 1 (A and B containers)
and for lot 2 (C and D containers). The same two thicknesses of
discs were used for each igniter made from the powder. The moisture
content was controlled at 0.2%, 0.4%, and 0.8% for the experiment.
The usual level was 0.4%, but this may be a source of trouble.

The engineers presented this picture: (continued on next page)

Discs (thickness)

Powder lots

1 2

Container Container
A B C D

Moisture content (%)

0.2 0.4 0.8 0.2 0.4 0.8 0.2 0.4 0.8 0.2 0.4 0.8

1 6
2 7
3 8
4 9
5 10

igniters in the cells

6.2 NESTED FACTORIAL 177

Discs (thickness)

Powder lots

1 2

Container Container
A B C D

Moisture content (%)

0.2 0.4 0.8 0.2 0.4 0.8 0.2 0.4 0.8 0.2 0.4 0.8

116

119
120

but the nesting of containers repeated for discs causes trouble in

understanding the following model, because one cannot easily
understand where the restriction on randomization occurs:

y. ., = \i + D. + P. + DP. . + C,.., + DC.,.., + + DM. 0

7 i j k A m i j i j ( j ) k i ( j ) k Z l Z

+ PM. . + DPM. .. + CM.,.., + DCM. + e,..,.. (6-2-5)

J z i j Z M j ) k X C J ) k)L ( l j k J l ) m

i = 1, 2 j = 1, 2 k = 1, 2 Z = 1, 2, 3 m = 1, 2, 5

where
^ijk£m = var:*-a^le t0 analyzed if each igniter performed
properly and the length of time is used

y = overall mean
x.L
D^ = i thickness of disc effect

Pj = j**1 powder lot effect

DP — = effect of the interaction of the i**1 thickness with

the powder lot
178 6. NESTED AND NESTED FACTORIAL DESIGNS

th th
C(j)k = e^ ect ^
container in the j powder lot
th
DC.,.,*, = effect of the interaction of the i thickness of
1(;,)k th th
disc with the k container in the j powder lot

= effect of the moisture content

and so on
th th
= effect of the m igniter in the ijkfc cell assumed
"(ijkJl)m
NID (0, a 2)

In order to see the restriction on randomization within the

containers, the following arrangement seems to be better:

Powder lots

Container Container
A B C D
Moisture ^ ^
1, 2, 3, 4, 5
content
1 0.4
0.8
Discs
0.2
2 0.4
0.8 116, ..., 120

and the model is easily seen to be

yijklm = u * Pi + C (i)j + 6 (ij) * Dk * PDik * C D (i)jk + M t

* "it * “ ( t m * DMM ' * CDM(i)jU <6-2-6>

+ e(ijkil)m

where all the symbols are defined as before except that

6.... = restriction error.
(ij)
6.2 NESTED FACTORIAL 179

Of course if percentage failures are analyzed, the variable is

^ijk£ i-nsteac* o f y±jkAm ^ecause there is only one percentage computed
for the 5 observations of success or failure.

Problem 6.2.4. Referring to Eq. (6.2.6), if powder lots,

containers and igniters are random and discs and moisture
content are fixed, show the analysis of the experiment
(including source, degrees of freedom and expected mean
squares) if the variable to be analyzed is to be a function of
the percent failures per treatment combination. Comment on the
design of this experiment in which there are five experimental
units per treatment combination but only the percent of
failures of the five responses can be analyzed. Is there a
better design? If so, explain your reasoning and show your
design.

Problem 6.2.5. An advertising firm wanted to test the

effectiveness of three fixed sizes (S) and four fixed color
arrangements (C) on attractiveness of billboards in five
different fixed locations (L) of the country. The variable to
be analyzed was a score from each of five different random
judges (J) used in each of the 20 location-color arrangement
combinations. Hence there were 100 judges in this experiment.
Each judge looked at and scored the three sizes of billboards
for his particular location-color arrangement combination.
Recognizing that there were 300 scores in the experiment:
(a) Show the layout of the experiment.
(b) Show an appropriate model and analysis of the
experiment.
(c) Indicate the tests of the various main effects and
interactions.
(d) Show the standard errors for the size and the color
combination.
(e) What would be the appropriate standard error for a
Newman-Keuls test on color arrangements?
(f) How would you change the experiment so you could get
a good test on location?
(g) Suggest changes in other parts of the experiment to
make the inference space broader.

COMMENT
It has been our experience that to understand the restriction
on randomization (if one exists) in nested factorial experiments, the
experimenter should place the factor that has the nesting at the top
of the design layout (horizontally). This allows one to see the
repeated measures (if they exist) over the other factor, down the
layout (vertically). As a result of all the above one can write
180 6. NESTED AND NESTED FACTORIAL DESIGNS

down the model in a manner that provides an ANOVA (with the EMS)
that is easy to interpret.

There is an example of the nested factorial experiment with no

restriction on randomization in Hicks (1973), p. 199, Problem 11.5.

6.3 REFERENCES

Anderson, V. L. Biometrics 26:255 (1970).

Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research,
McGraw-Hill, New York, 1952.
Edwards, A. L. Experimental Design in Psychological Research, Holt,
Rinehart and Winston, New York, 1962.
Eisenhart, C. Biometrics 3:1 (1947).
Hicks, C. R. Fundamental Concepts of Design of Experiments, 2nd ed.
Holt, Rinehart and Winston, New York, 1973.
Greenhouse, S. W. and Geisser, S. Psychometrika 24:95-112 (1959).
Mandel, J. JASA 52:552 (1957).
Morrison, D. F. Multivariate Statistical Methods, McGraw-Hill,
New York, 1967.
Winer, B. J. Statistical Principles in Experimental Design, 2nd ed.
McGraw-Hill, New York, 1971, Chapter 7.
 Chapter 7

SPLIT PLOT TYPE DESIGNS

The concept of restriction on randomization for the blocks

introduced in Chapter 5 on the RCBD is extended in the use of
split plot designs. Actually there can be many restrictions on
randomization at various levels in experimentation that will cause
many "splits" in the design. These designs are sometimes called
split-split plot, split-split-split plot and so on. The general
split plot type designs with their associated restrictions on the
randomization are considered in this chapter. For all cases covered
in this chapter we assume that the effects of any correlation among
errors on the analyses given is negligible.

7.1 SPLIT PLOT DESIGNS

The major difficulty that investigators have in using the split

plot design is recognition of the restrictions on randomization in
their experimental setup. After they understand the restrictions
they must next understand the difference between the model for the
factorial experiment CRD and that for the split plot design caused
by the restrictions on randomization. In Chapter 4 on the CRD, the
factorial experiment was introduced where no restriction on
randomization existed between the treatment combinations and the
experimental units. This is not the case for a split plot type
experiment. Kempthorne (1952, Chapter 19, especially p. 375) gives
an excellent discussion of split plot models when interactions of
treatments with blocks are assumed zero.

181
182 7. SPLIT PLOT TYPE DESIGNS

If one has a factorial experiment with three fixed factors, say

A with 3 levels, B with 4 levels and C with 5 levels, the experiment
for one experimental unit per treatment combination would demand 60
experimental units. The complete randomization procedure may take
place by arraying the treatment combinations as:

A B C
1. 1 1 1
2. 1 1 2
3. 1 1 3

60. 3 4 5

and choosing a random number between 1 and 60, say 2. Then

assigning the second treatment combination (A at level 1, B at level
1 and C at level 2) to the first experimental unit. Continue in
this manner as expressed in Chapter 4 for a completely randomized
design.

A reasonable linear model for the analysis of this experiment

y. ., = m + A. + B. + AB. . + C,+ AC.,

7ijk l j ij k lk

+ B C + ABC. ., + z f . .|x
}k ljk (ljk)

i = 1, 2, 3 j = 1, 2, 3, 4 k = 1, 2, 3, 4, 5

where all effects and interactions are defined as usual. The

analysis for this model is given in Table 7.1.1.
7.1 SPLIT PLOT 183

TABLE 7.1.1

ANOVA Using Equation (7.1.1)

Source df EMS

A 2 a 2 + 20<KA)
B 3 a2 + 154i(B)
AB 6 a2 + 54(AB)
C 4 a 2 + 124(C)
AC 8 a2 + 4<j>(AC)
BC 12 a2 + 34(BC)
ABC 24 a2 + 4 (ABC)
Error 0 a2

Total 59

In the usual split plot design, the three levels of A act as

blocks and the four levels of B are called whole plot treatments.
Then the five levels of C are randomized within each of the twelve
combinations of A and B. A linear model for the analysis of this
experiment is:

Y... = y + A. + 6^.^ + B. + AB.. + u)f . . ^ + C. + AC.,

13k l (l) 3 13 (13) k lk
(7.1.2)
+ BC., + ABC... + £f ^
Jk 13 k (13k)

where
th th
= variable to be analyzed from the i level of A, j
level of B and k**1 level of C

y = overall mean

A. = effect of the i**1 level of A

1
6^ = first restriction error

B. = effect of the levelof B

3
th th
AB^ = interaction effect of i level of A with the j level
of B
184 7. SPLIT PLOT TYPE DESIGNS

u),. = whole plot error or second restriction error

(i j ) F
(Anderson, 1970), and so on to

e...,. = split plot error or error due to repeating the

UJKJ th th
combination of the i level of A with j level of
B and level of C

This model may be interpreted as a split-split plot model if

one calls
a whole plot error

a split plot error

a split-split plot error

This model approach essentially refers to a randomized complete

block design as a special case of a split plot design. In most
books, however, this design is called a split plot design because
the first restriction error is not recognized.

The analysis of variance for data from this design is given in

Table 7.1.2.

The tests for all main effects and interactions are not
straightforward in Table 7.1.2. Some assumptions on interactions
are needed or conservative tests must be made when this fixed type
design is used. If, however, it is known by the experimenter that
all interactions with A in this experiment are zero, the tests are
all obvious.

To demonstrate how the split plot design may be used in a given

experimental situation the prosthetic cardiac experiment discussed
in Chapters 5 and 6 is continued here.

The actual design of the experiment used by Beeson (1965) was

a split plot design. He set up 2 random blocks in which he forced
all 4 valve types to appear once before the second valve of any
type was used. Of course, he randomized the order of using the
valve types in the machine and randomized the 6 pulse rates for each
valve type.
7.1 SPLIT PLOT 185

TABLE 7.1.2

ANOVA Using Equation (7.1.2)

Source df EMS

A 2 a2 + 5a2 + 20a2 + 20<J>(A)

First restriction o 9
error 0 a2 + 5a + 20a*
0) 0

B 3 a2 + Sa2 + 15*(B)

AB 6 a2 + 5a2 + 5* (AB)

Second restriction 9
error 0 a2 + 5a2
w

C 4 a2 + 12*(C)

AC 8 a2 + 4* (AC)

BC 12 a2 + 3<f>(BC)

ABC 24 a2 + *(ABC)
2
Within error _0 a

Total 59

Some authors believe the analysis for such an experiment should

be as in Table 7.1.3, where the arrows on the mean squares indicate
that the main effect of valve types would be tested by the whole plot
error and that the main effects of pulse rates and the interaction
of valve types by pulse rates would be tested using the split plot
error assuming the correlation of errors caused by all six pulse
rates occurring on the same valve for each type is negligible. In
order to calculate the whole plot error sum of squares one would use
the blocks by valve types interaction sum of squares and to calculate
the split plot error sum of squares one would combine the blocks by
pulse rates and blocks by valve types by pulse rates sums of squares.
*
Examples are Yates (1967, p. 785) and some of his references and
Federer (1955, p. 274).
186 7. SPLIT PLOT TYPE DESIGNS

TABLE 7.1.3

ANOVA of Usual Split Plot Design

Source df MS

Blocks 1
Valve types
Whole plot error

Pulse rates
Valve types x pulse rates
Split plot error

Harter (1961), Chew (1958, p. 48), and others have doubts that
one should always obtain the split plot error by pooling. In
biological studies using the "sometimes pooling" technique described
by Bozivich et al. (1956) and Bancroft (1968, p. 11), there seems to
be good reason to pool "almost always." Some authors have used the
sometimes pooling criterion when the split plot treatments are
random.

We take the view that if the experimenter has worked in the

field of investigation for some time and "knows" the errors involved,
he should pool and use the analysis given in Table 7.1.3 if this
seems appropriate. If, however, he does not know whether or not
blocks interact with the various treatments, we believe he should
let the data guide him in this decision. This thinking is consistent
with our use of restriction error in the RCBD and that the following
linear model is appropriate and allows for the sometimes pooling
methods (a numerical example will be given later):

(7.1.3)

i = 1, 2 j * 1, 2, 3, 4 k = 1, 2, ..., 6
7.1 SPLIT PLOT 187

where

y . = maximum flow gradient (mmHg) obtained from, the i

rth
13 .th , ,,th
block, j Uil valve type, and kU11 pulse rate

U ® overall mean

B. = effect of the i block (random), NID (0, aD)

1 D
2
= first restriction error zero df, NID (0, a^)

V. = effect of valve type (fixed)

i.L
BV.. = effect of the interaction of the i block with the
1J th 2
j valve type, NID (0, afiV)
2
w (ij) = seconc* restriction error, zero df, NID (0, a^) (cf
Kempthome, 1952, p. 375, the of Eq. (13) is
similar to this

P^ = effect of k**1 pulse rate

til til
BP., = effect of the interaction of the i block and the k
1 2
pulse rate, NID (0, a^p)
til
V P ^ = effect of the interaction of the j valve type with
the kth pulse rate

BVP.., = effect of the interaction of the i ^ block with the

l] th th 2
j valve type with the k pulse rate, NID (0, agyp)
2
e(ijk) = w^-t:^ n error, zero df, NID (0, a )

The corresponding analysis of variance is given in Table 7.1.4. We

believe that the controversy on pooling or not pooling BP and BVP
may be resolved by a sometimes pooling procedure for the given
problem, unless the experimenter wishes to specify his model before
running his experiment.
188 7. SPLIT PLOT TYPE DESIGNS

TABLE 7.1.4

ANOVA Using Equation (7.1.3)

Source df EMS

2 2 2 2
Blocks (Bi) a + 6cT + 24a* + 24a^
a) 6 B

First restriction 2 2 2
error C5^ ) (T + 6<T + 24a*
(0 0

Valve types (V^) 3 a2 +

6a« + 6aBV + 12*(V)
Interaction (BV..) 3 a2 ♦ 6a2 + 6a2
u> BV
Second restriction
error 0 ^ ) ) 0 a + 6a2
0)

Pulse rates 5 a2 +
" b p * 8<KP)
Interaction (BPik) 5 a2 +
<
Interaction a2 +
<vv
15 °BVP + 2+(VP)
2
Interaction (Bw ijk) 15 a2 +
BVP
Within error 0 a2
(e(ijk)}

To give further explanation for our position in analyzing split plot

data in this manner, let us compare this analysis step by step with
the nested factorial (Table 6.2.2). The premise we use is that there
is a possibility of a block by treatment (BV) interaction if a split
plot design is used. This source has 3 df in Table 7.1.4. There is
no BV interaction, of course, if the nested factorial design is used.
Then the blocks and BV are not separable as in occurrences in valve
types with 4 df in Table 6.2.2.
7.1 SPLIT PLOT 189

If it is agreed that a block by treatment interaction may exist,

and this seems quite logical when block means are of interest as in
many engineering experiments, then there certainly may be a block by
treatment interaction at another stage in the design. For example
BP may exist, in which case we can see no reason for the possibility
that BVP cannot exist separate from BP.

All this is not to say, in many agricultural experiments where

block effects and block error are not separable or the experimenter
does not want to separate them, that he should not run his split plot
analysis as given in Table 7.1.3. In this case the experimenter has
used a model that he ’’knows” is correct when he began his experiment.

Our experience in various other fields is that experimenters

do not have this information and may gain information by using the
flexible Eq. (7.1.3) and test this model using the analysis in
Table 7.1.4. The use of the restriction errors at the various
stages allows the investigator to think about various sources of
variation that are covered up when only the analysis in Table 7.1.3
is carried out.

Example 7.1: Using the actual data from Beeson (1965), the
analysis turned out to be as that given in Table 7.1.5. When the
correct tests are made, only one source, pulse rates, is significant
at the a = 0.05 level. However, the mean squares of the
interactions with blocks at the various stages are about the same
size, indicating that the actual variance components with blocks
are possibly zero [from Eq. (7.1.3)]. Using a certain criterion
for pooling, the more liberal ’’sometime poolers” may even obtain
the analysis of Table 7.1.6. This shows that all effects and the
interaction are significant; however, this last analysis is valid
only if the errors are poolable and pooling is correct. In general,
for a correct analysis, all restrictions on randomization must be
considered before pooling procedures are undertaken.
190 7. SPLIT PLOT TYPE DESIGNS

TABLE 7.1.5

ANOVA for Data Using Eq. (7.1.3)

Source df MS

Blocks (B^) 1 17.52

First restriction
error 0

Valve types (Vp 3 420.41

Interaction (BV..) 3 77.19
ly
Second restriction
error 0

Pulse rates (Pp 5 427.64^

Interaction (BP^) 5 51.97

Interaction (Vpjp 15 133.79

Interaction (BVPjjp 15 56.84

Within error (£,..,0 0
^ (ij^)

♦Significance at a = 0.05.

TABLE 7.1.6

ANOVA After Obtaining "Pooled Error" from

All Sources with Blocks from Table 7.1.5

Source df MS

Valve types 3 420.41^

Pulse rates 5 427.64^
Valve types x pulse rates 15 133.79^
Pooled error 24 56.73

♦♦Significance at a = 0.01.

♦Significance at a = 0.05.
7.1 SPLIT PLOT 191

Frequently experimenters in industry do not recognize a split

plot design and analyze it as a factorial experiment CRD. An
example occurred in the steel industry in which the effect of
temperature and orientation on strength of alloys was investigated
using four furnaces in the laboratory. Each furnace had a different
temperature and within each furnace the investigator randomly placed
(orientation 1) two samples from each of three alloys and had an
aligned arrangement (orientation 2) of two other samples from the
same three alloys. In other words, he put 12 samples of alloys in
each furnace, six randomly oriented and six aligned arranged. He
analyzed the experiment as a completely randomized design using the
between samples within the temperature, alloy, orientation source
for the error. The analysis he presented for the factorial
experiment is given in Table 7.1.7.

TABLE 7.1.7

ANOVA for the CRD Factorial

Source df EMS

Temperatures (T) 3 o2 + 12*(T)

Alloys (A) 2 a2 + 16*(A)
T x A 6 a2 + 4* (TA)
Orientations (0) 1 a2 + 24*(0)
T x 0 3 ff2 + 6*(TO)
A x 0 2 a2 + 8*(AO)
T x A x 0 6 a2 + 2*(TAO)
Between samples 9
within (T-A-0) Cell 24 a
Total 47

It was shown that temperatures had a major effect and the best
temperature from this analysis was used in production. Some time
later in the plant production, it was shown that the new temperature
used in this experiment had no significant effect over the old one
on strength of the alloys and the experimenter could not understand
why his experiment had given him the wrong results.
192 7. SPLIT PLOT TYPE DESIGNS

If he had analyzed the experiment as he had designed it (as a

split plot), he would have had no error to test temperatures. Since
alloy-orientation combinations are completely randomized in
temperatures (furnaces), a more appropriate model for the analysis
of the data from this designed experiment is:

y a u = u + Ti + 6(i)+ Aj ♦ TAij + °k♦ TOik

(7.1.4)
+ AQjk + TAOiik + e(ijk)Jl

where

i = 1, 2, 3, 4 j = 1, 2, 3 k = 1, 2 £ = 1, 2

and all effects and interactions are defined in the usual manner.
The corresponding analysis of variance is given in Table 7.1.8.

TABLE 7.1.8

ANOVA Using Equation (7.1.4)

Source df EMS

Temperatures and/or furnaces 3 a2 + 12a2 + 124(T)

Whole plot error 0 a2 + U o 2s

Alloys 2 a2 + 164(A)

T x A 6 a2 ♦ 44 (TA)

Orientation 1 a2 + 244(0)
2 +
T x 0 3 a 64(TO)

A x 0 2 a2 + 84(AO)

T x A x 0 6 a2 + 24 (TAO)

Split plot error 24 a2

Total 47
7.1 SPLIT PLOT 193

Since the location of the 12 samples in each furnace was at

random, there was a completely randomized design in each furnace and
the split plot error came from between samples within the temperature,
alloy, orientation combination in each furnace and is appropriate for
testing the effects and interactions from alloys through the
T x A x 0 source. The temperatures were not replicated in this
experiment and were completely confounded with furnaces, which
formed the whole plot. There was a restriction on randomization in
that once a temperature was chosen for a furnace the alloys and
orientation level had to go in that furnace. There was no overall
randomization of the combinations of temperature, alloys, and
orientation for all 48 samples. Hence this is a split plot design
and has no whole plot error estimate.

To obtain a whole plot error estimate it would be necessary for

the experimenter to repeat the whole experiment at least once,
possibly more, or take only two temperatures and obtain a repeat of
the experiment using the remaining two furnaces. With furnaces
random and temperatures fixed a random assignment of temperatures to
the furnaces would give a RCBD at the whole plot level. Therewould
be no test on furnaces but the interacting source would test
temperatures as explained in Chapter 5. Again, the research worker
must determine his inference space before he decides how he should
run the experiment.

Example 7.2: In order to illustrate the calculational

procedure for the analysis of data from a split plot design, consider
the model presented in Eq. (7.1.4). The measured variable in this
case was alloy strength. The coded values for such an experiment
are shown below. The data are set up to correspond to ANOVA table
presented in Table 7.1.8.
194 •7. SPLIT PLOT TYPE DESIGNS

Temperature

675 700 725 750

Alloy Alloy Alloy Alloy

1 2 3 1 2 3 1 2 3 1 2 3

Orientation 2
15 20 27 35 48 47 55 62 23 33 25
i
i
19 28 26 40 39 55 55 63 58 34 39 17

15 25 25 48 48 55 48 63 68 28 33 14
2
23 31 33 55 62 64 60 68 62 37 38 12

The ANOVA for these data is as follows:

Source df SS MS EMS F

Temperature (T) 3 10865.83 3621.94 a2 + 12a2 ♦ 12*(T) None

Whole plot error 0 a2 + 12a2

Alloys (A) 2 434.29 217.15 a2 + 16<|>(A) 5.49*

T x A 6 1240.55 206.76 a2 + 4* (TA) 5.23*

Orientation (0) 1 468.75 468.75 a2 + 24<K0) 11.85*

T x 0 3 400.42 133.47 a2 + 6<j>(TO) 3.38*

A x 0 2 74.63 37.31 a2 + 8*(AO) 0.94

T x A x 0 6 75.20 12.53 a2 + 2<|>(TA0) 0.32

Split plot error 24 949.00 39.54 a2

Total 47 14508.67

♦Significant at the 0.05 level

The sum of squares calculations are as follows (note that the

calculation is identical to an ordinary three factor experiment for
a completely randomized design):
7.1 SPLIT PLOT 195

E H E y..,( = T = 1880 CT = ■ = 73633.33

. ., n J •••• 48
i j k A

S H I y* = 88142
i j k 4 1JJa

SS total = 88142 - CT = 14508.67

SS temp = -j| I T?>#> - CT = -ji (2622 + 5762 + 7092 + 3332)

i

- CT

= 84,499.16 - CT

= 10865.83

SS alloys = y| I T2^. - CT = (S612 + 6752 + 6442) - CT

= 74067.62 - CT

= 434.29

SS orientation = ^ £ T2 -CT = ^i (8652 + 10152) - CT

24 ^ #*k* 24 v J

= 74,102.08 - CT

= 468.75

The interaction sum of squares calculations are carried out by first

forming the three 2-way tables of totals as follows:

Alloy Orientation Orientation

1 2 3 1 2 1 2

675 59 99 104 675 110 152 1 247 314

700 170 184 222 700 244 332 2 307 368
Temp ?25 Alloy 3
725 210 249 250 340 369 311 333
750 122 143 68 750 171 162
196 7. SPLIT PLOT TYPE DESIGNS

Now,
S S T x A -ilS T ? . -CT-SSA-SST
4 . . li••
i 3 J

*j (592 + 1702 ♦ ... ♦ 2502 ♦ 682) - CT

- SS A - SS T

= 86,174 - CT - SS A - SS T

* 1240.55

SS T x 0 4 Z I T? , - CT - SS T - SSO
6 ., i*k*
ik

* i (1102 ♦ 2442 ♦ ... i- 1622) - CT - SS T - SS 0

* 85368.33 - CT - SS T - SS 0

* 400.42

s s a x o 4 z z t 2 ., - c t - s s a - sso
» j k ' 3K*

* 74,611.00 - CT - SS A - SS 0

= 74.63

As before, the three factorinteraction sum ofsquares is given by

S S T x A x O = i z Z Z T?..- CT - SS T - SS A - SS 0 - SS TxA
2 . . . ilk*
1 3 k J

-SSTxO-SSAxO

1 2 2 2 2
=j (21z ♦ 38^ + 43^ ♦ ...♦ 26 ) - CT

- SS T ..... S S A x O

« 87,193 - CT - SS T .......SS A x 0

* 75.20

SS error * SS total - S S T - S S A - ••• - S S T x A x O

7.1 SPLIT PLOT 197

ss error = i J k J y« “ " 2 j ; j TJjk.

= 88,142.00 - 87193.00

= 949.00

Since there areonly two observations per cell for thisproblem, the
error sum of squares may be calculated by summing the squareof the
differences of the two observations in each cell and dividing by
two, i.e.,

SS error = (172 + 82 + 132 + ... + 22)

=j (1898.00)

= 949.00

This last relationship may be explained by letting y^ and y2 be any

two observations in any one cell. Then this cell contributes
- 2 - 2
(yx - y) + (y2- y) to the error sum of squares where

- =-
y yl
- + y2 9 the cell mean. Note that

- 2 - 2 yl + y2 2 yl + y2 2
(yx - y) + (y2 - /) s Cxx - 3 + (y2 " " S " 3

2yr yr y2 2 2y2_yr y2 2
= +( 2 )

( y r y 2) 2 (y2- y p 2
4

cy r y 2)2
2

difference squared over 2

Summing over all cells gives the sum of squares for error for all
cells.
198 7. SPLIT PLOT TYPE DESIGNS

Problem 7,1,1. An agronomist investigated the effects of three

different perennial legumes on weight gain in cattle, A large
field was divided into four parts. Each random part (replicate)
had three fenced-in areas one for each perennial legume.
Thirty six steers of comparable age, weight and breed were
allowed to graze the 12 pastures with three randomly selected
steers per pasture (replication-legume combination) for a fixed
amount of time. Other food and water was kept as nearly equal
as possible. The gain in weight for the grazing period of time
was recorded on each steer the first year.
The next year the same pastures were used again with 36
different, but similar, steers. The same type randomization
of the cattle to the pastures was carried out. This experiment
was continued for a total of five years.
(a) Sketch the design of the experiment.
(b) Show an appropriate model, the corresponding analysis
and the tests on legumes, years and their interaction.
(c) How would you improve this experiment using the same
number of animals?

Problem 7.1.2. A research worker agreed to investigate four

types of heart valves. For the research, a mechanical apparatus
with a water storage and circulatory system with a pulsing pump
was constructed. Water was to be used to simulate the blood
and a pump attached to an electric motor was to simulate the
heart. The motor could vary the pulse rate from 0 to 220 beats
per minute and it was decided to use the six levels: 60, 80,
100, 120, 140, and 160 beats per minute in the experiment.
The variable to be analyzed was a function of a measured
efficiency variable. The experimental unit is a valve set in
the equipment. When more than one pulse rate is used on a
given valve there may be correlation of errors.
The purpose of the experiment was to select the best valve
type for all pulse rates and/or the best valve for particular
pulse rates.
Using the information above, design three different
experiments with conditions stated in a, b, and c, and show
their analyses:
(a) Three experimental units (valves) per treatment
combination with no restriction on randomization.
(b) Allow three valves per valve type or a total of
twelve valves to be used in the experiment. Choose one of the
twelve valves at random, insert it in the system and run all
six pulse rates at random. Next, choose one of the remaining
eleven valves at random. Remove the first valve and insert the
second one chosen. Run all six pulse rates rerandomized.
Continue in this manner until all four valve types have been
used three separate times (using a different valve each time,
completely randomly chosen) with the six pulse rates at random
each time.
(c) Over a six-months period, choose three days at random.
7.2 SPLIT-SPLIT PLOT 199

For the first randomly chosen day, choose a valve type at

random. Run all six pulse rates at random on the chosen valve.
Next choose one of the remaining three valve types at random.
Remove the first valve and insert the second randomly chosen
type. Run all six pulse rates rerandomized. Continue in the
same manner for the remaining two valve types. At this stage
only one valve has been used for each type.
Repeat the experiment as above for the other two randomly
chosen days, using different valves for each type.
(d) How would you prefer running the experiment [not
necessarily one of (a), (b) or (c)] to get the widest
inferences and the best tests using a maximum of seventy two
valves and/or observations? Why? Comment on costs. Consider
correlated errors.

7.2 SPLIT-SPLIT PLOT DESIGNS

Restrictions on randomization of the treatments with the

experimental units may be carried on at any number of stages in a
split plot type experiment. Each of these stages may be described
in terms of splits in the design, and a design with two stages of
restrictions on randomization is often called a split-split plot.

An example of such a design occurred in a company dealing with

metals. Three chemical formulations were examined using six heats
(two heats for each chemistry with only one randomly chosen ingot
for each heat). From each ingot three wheels of metal were cut,
one near the top of the ingot, one near the center and one near the
bottom.

From each wheel six specimens were cut at random. Three of

these specimens were placed in one environmental condition and the
other three were placed in another environmental condition. This
procedure was carried out for all combinations of heats and wheels.

In this case the heats (or ingots representing heats) acted as

whole plots, the wheels acted as sub plots because the ingots were
examined at particular spots from top to bottom (wheels). Finally
the sub-sub plots were represented by the two conditions (three
specimens placed in the same environmental condition).
200 7. SPLIT PLOT TYPE DESIGNS

A reasonable model for the analysis of this experiment is:

Yijk*m ° y + Ci + H(i)j + 6(ij) + Wk + ^ik + ^(ijjk

+ w(ijk) + c* + CcU + Hc(i) j)t + WcU + ^ikJ. C7-2-1)

+ HWcr.. .. „ + e r .
(i)jk£ (ijk£)m

i = 1, 2, 3 j = 1, 2 k = 1, 2, 3 Z = 1, 2 m = 1, 2, 3

where
^ijkfcm = var^-a^^e t0 analyzed
y s overall mean

« effect of the i**1 chemistry (fixed)

H(i)j =e^ ect t*ie ^eat (random) in the i**1

chemistry

6^^ «first restriction error

= effect of the k**1 wheel height (fixed)

CW., =effect of the interaction of the i**1 chemistry

th
with the k wheel height

HWr..., a effect of the interaction of the j**1 heat in the

th th
i chemistry with the k wheel height

w(ijk) * seconc* restriction error

c^ a effect of the condition (fixed)

Cc.0 a effect of the interaction of the i**1 chemistry

with the Z condition

Hcr v o a effect of the interaction of the j**1 heat in the

'*1^ th th
i chemistry with the Jl condition

Wc, 0 a effect of the interaction of the k**1 wheel height

th
with the Z condition

CWc., 0 a effect of the interaction of the i**1 chemistry

^ th th
with the k wheel height with the Z condition
7.2 SPLIT-SPLIT PLOT 201

HWc... n effect of the interaction of the j**1 heat in the

th th
i chemistry with the k wheel height with the
condition

within error of specimens within conditions,

wheels and heats in chemistries

The corresponding analysis of variance is given in Table 7.2.1.

TABLE 7.2.1

ANOVA Using Equation (7.2.1)

Source df EMS

Whole Plot
Chemistry (C) 2 a2 + 6a2 + 18a2 +18a2 + 36<|>(TV
C)
a) <5 H J

Heats and/or ingots 2 2 2 2

in C(H) 3 a + 6a + 18a* + 18a*
to 6 H
2 2 2
First restriction error 0 a + 6a*+ 18a*
CD 6

Split Plot
Wheel heights (W) 2 a2 + 6a2 + 6a ^ + 36<KW)

C x W 4 a2 ♦ 6a2 + 6a ^ ♦ 12* (Of)

2 2 2
H x W 6
• ♦ K ♦ 6* m
Second restriction error 0 a2 + 6a2

Split-Split Plot
2
Conditions (c) 1 a + 9a2c + 54*(c)
2
C x c 2 a + 9a2c + 18* (Cc)

H x c 3 a2 +
2
W x c 2 a + ' i * »«♦<*>
C x W x c 4 a2 +
3<THWc + 6^ CWc^
H x W x c 6 a2 +
3«mc
Within error 72 a2

Total 107
202 7. SPLIT PLOT TYPE DESIGNS

The tests are obvious from the expected mean squares. It is

interesting to notice, however, that the information obtained by
taking three specimens per condition is not necessary. It would
have been almost as informative to take only two specimens per
condition.

Problem 7.2.1. Would it have improved the experiment above

with the ANOVA given in Table 7.2.1 to have taken more than
one ingot per heat and taken fewer heats? Explain carefully.

To reduce the cost of the experiment, the investigators would

have liked to have run only one heat per chemistry, which would have
given no degrees of freedom for testing chemistries. Of course, this
would have given no information on the most important factor. To
examine the efficiency of the design one could look at the
combinations of the number of heats per chemistry and the number of
specimens per condition with the corresponding number of degrees of
freedom for the errors.

Consider the number of heats per chemistry as follows:

Number of heats Degrees of freedom

per Chemistry for testing chemistries
1 0
2 3
3 6
4 9

When considering efficiency there is not much value in running the

experiment if only one heat is used per chemistry. There is a 100%
gain in degrees of freedom for error going from two to three heats
per chemistry, but only 50% gain in going from three to four. The
increase in efficiency of the design decreases as the number of heats
per chemistry increases from there on.

Next, consider the number of specimens per condition. If only

one were used there would be zero degrees of freedom for the within
error, which may not make much difference.
7.2 SPLIT-SPLIT PLOT 203

It seems there is almost no advantage in taking more than one

specimen per condition if the variances are homogeneous, but usually
experimenters would like a few repeats in cells to check equipment
and operators.

Overall recommendations for the design of the experiment:

(a) If possible use three heats per chemistry and one specimen
within conditions, with a few repeats. Of course four heats would
be better but the experimenters were not allowed that much expense.
(b) If (a) is too expensive, use two heats per chemistry.
(c) If two heats per chemistry is too expensive and a test for
chemistry is desired, there is not much reason to run the experiment.

Problem 7.2.2. In a heat paper battery experiment the variable

to be analyzed was the energy output of 25 small pieces of
paper cut from a large square sheet of heat paper. The sheets
of heat paper were made up of combustible material spread over
a sheet of paper as evenly as the equipment would allow. The
experimental units (small circular pieces) were cut out of the
sheets in the following manner:

Columns

1 ooooo
2 ooooo
Rows 3 ooooo
4
ooooo
5 ooooo
The active ingredients are most important to the
experimenter and the effects of rows and columns in describing
the response surface of the sheets are less important. Each
sheet was selected at random for a run through the equipment
and was made up of one combination of three active ingredients
(A at 4 levels, B at 3 levels and C at 2 levels). This made a
total of 24 sheets ( 4 x 3 x 2 ) for one experiment.
(a) Show the design of the experiment, emphasizing the
randomization and inference space.
(b) Show the model and analysis of variance for your
experiment.
204 7. SPLIT PLOT TYPE DESIGNS

(c) Explain what results can be obtained (include the

response surface for the sheets).
(d) Did the design allow for the most important effects
to be estimated with the smallest variance? Comment.
(e) When the actual experiment was run and the data
analyzed, the error mean square within the sheets of heat
paper was 10 times larger than the error mean square between
the sheets. What would you suggest doing next?

Problem 7.2.3. An experiment was conducted on cold rolled

steel, and was repeated a week later. Two heats were run each
week and the same two cold rolled methods were used for each
of the four heats. All interaction with heats and weeks (both
considered random and no interest to the experimenter except
to provide the inference base) were impossible because there
must be different heats nested in the different weeks.
Hence there were eight coils of steel, four from each
rolling method, and each of these coils was cut in two. One-
half of each coil was run at one milling speed and the other
half at another speed. Each half coil was then examined at
two places; one place was at the end away from the cut that
divided the coil and the other was 50 feet from the end.
A schematic diagram of a half coil is shown in Fig. 7.1. Two
random samples of steel were taken at each of these two places
on each half coil to allow measurements of many variables.
Using a variable to be analyzed associated with magnetic
properties the analysis of variance (only the source and mean
squares) for this problem is shown in Table 7.2.2. These
results were obtained from a computer output that did not take
into account the nesting quality of heats.

Sam ples of Steel

h — ii
Cut

Fig. 7.1. A schematic diagram of a half coil.

(a) Show a diagram of the whole experiment.

(b) Assuming all factors are fixed except heats and weeks,
analyze the data appropriately showing the model and correct
ANOVA using the mean square data and degrees of freedom. (The
df and SS for all nested sources may be obtained in a fashion
similar to that presented in Example 6.1).
(c) Run the appropriate tests of significance and show
the standard errors necessary to run individual comparisons
where necessary. Explain the results.
 TABLE 7.2.2

ANOVA for Problem 7.2.3

Source MS

Weeks (W) 4.4

Heats (H) 34.7
H x W 20.6
Cold rolled methods (C) 116.8
C x H 1.4
C x W 0.4
C x H x W 11.4
Mill speed (S) 3.1
S x H 23.6
S x W 17.9
S x H x W 2.5
S x C 9.1
S x C x H 0.3
S x C x W 12.8
S x C x H x W 2.0
Positions (P)
(ends or 50 feet) 23.9
P x H 0.7
P x W 6.3
P x H x W 0.2
P x C 3.2
P x C x H 9.5
P x C x W 1.0
P x C x H x W 3.4
P x S 23.4
P x S x H 0.5
P x S x W 0.2
P x S x H x W 0.2
P x S x C 2.8
P x S x C x H 3.8
P x S x C x W 1.2
PxS xCxH xW 4.7
Within cells 2.6
206 .7. SPLIT PLOT TYPE DESIGNS

(d) Explain the role of the two experimental units for

each treatment combination. How would you have designed the
experiment? Why?

7.3 REGRESSION ANALYSES FROM SPLIT PLOT DATA

In Chapter 2 it was made clear that there could be only one

error in a regression analysis. This implies that a CRD is necessary
for obtaining the data. Of course, in actual practice most of the
regression analyses are run on unstructured data. That is, the
design of the experiment is ill-conceived or the investigator merely
uses data on hand rather than controlling levels of factors. In
general, however the regression analysis assumes there is only one
error and the design is completely randomized.

Many times experimenters want to use a regression analysis after

the ANOVA has been run on the split plot data. If the various errors
in the analysis are not poolable the whole plot effects and
interactions regression coefficients will have variances that are too
small and the split plot main effects and interactions regression
coefficients will have variances that are too large. A reference for
this concept is Daniel and Wood (1971, p. 59) on nested data. In
effect the regression error from split plot data or any data coming
from a design with restriction on randomization such as RCBD and
nested factorials, is a diluted average of the appropriate errors.

One proposal for correct regression analyses of split plot data

is to use the means of the whole plot data with a separate regression
analysis and individual observation data for the split plot part.
Thus forcing two regression analyses for split plot data. This, of
course, is not very satisfactory and it would be much better if the
errors could be pooled in the ANOVA before the regression analysis
is run.

Problem 7.3.1. A research organization ran the following

experiment to examine the compressive strength (y) of concrete
when steel fibers were used in the mix.
Three factors were used in the experiment.
7.3 REGRESSION FROM SPLIT PLOT DATA 207

(a) Mix types: (1) sand with cement; (2) 3/8 inch
rock with cement; (3) 3/4 inchrock with cement.
(b) Percent volumeof fibers: 0, 1, 1.5, 2, 2.5.
(c) Size of cylinders: 3” x 6”, 4”x 8", 6” x 12”.
The experiment was conducted in the following manner:

Day 1: One batch of mix (sand with cement) was made up and
split into 5 parts.
The first part had 0% fibers added,
the second part had 1% fibers added,
the third part had 1.5% fibers added,
the fourth part had 2% fibers added,
and the fifth part had 2.5% fibers added.
Each part was made into 6 cylinders, two each for the
three sizes: 3" x 6”, 4" x 8", 6” x 12”.
Day 2: Repeat of Day 1.
Day 3: The batch of mixwas made of 3/8” rock with cement and
the same procedure was followed for the other two
factors as in Days 1 and 2.
Day 4: Repeat of Day 3.
Day 5: The batch of mixwas made of 3/4” rock with cement and
the same procedure as above was followed for the other
two factors.
Day 6: Repeat of Day 5.

Each day's cylinders were tested at random after each day's

cylinders set for a given time.
Answer the following:
(a) Show the layout of the experiment. If the three factors
are fixed and days are random, describe the inference space.
(b) Describe briefly two weaknesses at the design stage and
what these could cause in the interpretation of the analysis.
(c) Write down the model assuming the weaknesses have not
prevented an analysis of data, show the limits on each index
and identify each symbol.
(d) Show the ANOVA (source, df and EMS) and indicate
appropriate tests using arrows. Indicate where pooling of
errors is likely.
(e) If the source and df printout from the computer are as
shown in the following tabulation, show what sources and df
must be pooled and where the restriction errors are to get the
correct analysis in (d).
208 7. SPLIT PLOT TYPE DESIGNS

Source df
Mixes (M) 2
Days (D) 1
M x D 2
Fibers (F) 4
MF 8
DF 4
MDF 8
S 2
MS 4
DS 2
MDS 4
FS 8
MFS 16
DFS 8
MDFS 16
within 90

(f) If it turned out that the main effects were all significant
and no interactions were significant show what cautions you
would make in the interpretations and what further analyses you
would do on each.
(g) What further analyses would be appropriate if the
interaction of mixes x fibers turned out to be significant?
Consider regression analyses.
(h) Consider the whole experimental procedure used and
recommend a better design explaining at each stage why you
think your design is better. (Keep the cost of running the
experiment in mind, so you do not run a completely randomized
design each day.)

7.4 REFERENCES

Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research,

McGraw-Hill, New York, 1952, pp. 344-351.
Bancroft, T. A. Topics in Intermediate Statistical Methods, Iowa
State University Press, Ames, Iowa, 1968.
Beeson, J. Unpublished M.S. Thesis, Purdue University Library,
West Lafayette, Indiana, 1965.
Bennett, C. A. and Franklin, N. L. Statistical Methods (In Chemistry
and Chemical Industry), Wiley, New York, 1954, pp. 539-545.
Bozivich, H., Bancroft, T. A. and Hartley, H. 0. Ann. Math. Statist.
27:1017 (1956).
Bliss, C. I. Statistics in Biology, McGraw-Hill, New York, 1960,
Section 12.2.
7.4 REFERENCES 209

Chew, V. (Editor) Experimental Designs in Industry, Wiley, New York,

1958, pp. 48-52.
Cochran, W. G. and Cox, G. M. Experimental Designs, 2nd ed., Wiley,
New York, 1957, Chapter 7.
Daniel, C. and Wood, F. S. Fitting Equations to Data, Wiley, New
York, 1971.
Federer, W. T. Experimental Design (Theory and Application),
Macmillan, New York, 1955, Chapter X.
Fisher, R. A. The Design of Experiments, 8th ed., Hafner, New York,
1966.
Harter, H. L. Biometrics 17:144 (1961).
Kempthome, 0. Design and Analysis of Experiments, Wiley, New York,
1952, Chapter 19. Distributed by Krieger Pub. Co., Huntington,
New York.
Ostle, B. Statistics in Research, 2nd ed., Iowa State University
Press, Ames, Iowa, 1963.
Snedecor, G. W. and Cochran, W. G. Statistical Methods, 6th ed.,
Iowa State University Press, Ames, Iowa, 1967.
Steel, R. G. D. and Torrie, J. H. Principles and Procedures of
Statistics, McGraw-Hill, New York, 1960, Chapter 12.
Winer, B. J. Statistical Principles in Experimental Design, 2nd ed.,
McGraw-Hill, New York, 1971.
Yates, F. Fifth Berkeley Symp. Vol. IV, Biology and Problems of
Health, 777 (1967).
 Chapter 8

LATIN SQUARE TYPE DESIGNS

The completely randomized design (CRD), the randomized complete

block design (RCBD) and the split plot designs are used more
frequently in experimentation than the designs in this chapter. The
Latin square (LS) and other designs similar to it are used by some
research workers, but are misused by others more frequently. Hence
in this chapter we will try to point out the type of experimental
problems in which these designs are appropriate and the mistakes
that experimenters may make if these designs are not handled
correctly.

8.1 LATIN SQUARE

In the RCBD each block is large enough to contain all of the t

treatments. In the Latin square design "square” refers to the
2
existance of t experimental units, each of the t treatments are
assigned in a special way, as described below, to t of the
experimental units. The t treatments are labeled with Latin
letters. Hence the name Latin square.
2
The actual arrangement of the t treatments onto the t
experimental units may be handled by drawing a square where the
sides of the square are called rows and columns and the restrictions
on the drawn square are such that each treatment must be represented
once and only once in each row and once and only once in each column.
The assignment of treatments to the experimental units is done at
random as long as the two restrictions are observed. Thus the term

210
8.1 LATIN SQUARE 211

random square is used to refer to a specific design. The

randomization procedure and tables of many squares are given by
Fisher and Yates (1963, pp. 22-25 and 86-89).

In agronomy field experiments (where the inference is confined

to a specific field) frequently the experimenters find that there
is a fertility trend in the direction perpendicular to the blocks
used in a randomized complete block design. To account for the
effect of this additional extraneous variable and remove it from
the error in the model, blocks may be set up as

Columns (Blocks)
Fertility Trend ------------
1 2

Rows (Blocks)

In this case t plots are required for blocking in both directions.

This naturally forces the number of rows (r) and columns (c) to be
equal to the number of plots (t) and the t treatments are assigned
at random to each row and column combinations as discussed above.
The t treatments are sometimes given Latin letters, A, B, ..., to
distinguish them from the numbers in the rows and columns and this
constructed square is called a Latin square.

There are two restrictions inherent in this randomization scheme

and these restrictions are recognized as rows and columns in the
analysis (comparable to blocks in the RCBD). Using the same thought
process as was used on the RCBD with blocks fixed and assuming all
interactions are zero, the model could be written as:
212 8.- LATIN SQUARE TYPE DESIGNS

y. ., = y + R. + 6... + C. + r),.^ + T, + £r .n
•'ljk i (i) j Q) k (ljk)

j = k = 1, 2, ..., t
where
y.., = variable to be analyzed from the i row j column
1*J th
and k treatment

y = overall mean
.th
R. = effect of the i row
l
6... = row error, due to restriction on randomization
(1) (Anderson, 1970) of treatments in the i
th row
•f”L
Cj = effect of the j column

nr.N = column error, due to restriction on randomization of

^ th
treatments in the j column

T. = effect of the k treatment

k
e(ijk) = experimental error from the i ^ row, j**1 column and
k ^ treatment estimated by the residual.

The corresponding ANOVA is shown in Table 8.1.1.

TABLE 8.1.1

ANOVA Using Equation (8.1.1)

Source df EMS

Rows (R) t - 1 a2 + ta2 +

Row restriction error 0 a2 +

ta6
Columns t - 1 a2 + ta2 + t<KC)
n
Column restriction error 0 a2 + ta2
n
Treatments t - 1 a2 + tt(T)

Experimental error
2
(residual) (t - 1) (t - 2) a

Total t2 - 1
8.1 LATIN SQUARE 213

It should be noted that for the Latin square design, since it

is a type of fractional replication (to be discussed in Chapter 10),
the algorithm to obtain the EMS given in Chapter 1 cannot be used.
One must go through a proof of the EMS for each source. References
for deriving EMS are Anderson and Bancroft (1952), Wilk and
Kempthorne (1957), and Edwards (1962).

Since the agricultural experimenter is not interested in the

effects of rows and columns (essentially he knows they are large),
the analysis satisfies him (which indicates the design is
satisfactory). As shown in the above analysis, the variation
remaining (after removal of rows, columns and treatments variation)
was used as the experimental error. Under the usual experimental
conditions and if no interactions exist among these three factors
(rows, columns, and treatments) the test is quite valid. For many
experiments in agriculture the requirements for valid tests may
usually be met, but for many experiments in engineering, social
sciences and other research areas they are not met and the results
could be interpreted erroneously if the experimenter merely uses
this design without understanding the assumptions.

To illustrate some of the difficulties an experimenter may

encounter using this design, consider an experiment on explosive
switches. There were five fixed treatments, that is, five levels
of packing pressures (10,000, 15,000, 20,000, 25,000, and 30,000
psi), five fixed machines (which act as rows) to make the switches
and five fixed men (which act as columns) to fire them. Hence there
are 25 combinations of machines and men with a treatment in each
cell. The design of a selected random square is given in Table 8.1.2.

The 25 switches were fired in a completely random order. If we

do not assume the interactions are zero and if all treatment
combinations were run as a CRD, the model would be
214 8. LATIN SQUARE TYPE DESIGNS

TABLE 8.1.2

Explosive Switch Latin Square

Men (m)

Machines
1 2 3 4 5
(M)

1 20,000 15,000 10,000 25,000 30,000

2 15,000 25,000 30,000 20,000 10,000
3 10,000 30,000 20,000 15,000 25,000
4 25,000 10,000 15,000 30,000 20,000
5 30,000 20,000 25,000 10,000 15,000

= y + M. + m. + Mm.. + T. + MT.. + mT.,

^ijk l 1 il k ik jk
(8.1.2)
+ MmT. ..
ijk (ijk)

where

y — ^ = the firing time of the switch made by the i ^ machine,

by the jt^1 man with the k**1 packing pressure

y = the overall mean

th
M. = the effect of the i machine
l
th
iik = the effect of the j man
th
T^ = the effect of the k packing pressure

and all interactions, plus

e..... = error within the (iik)^ cell.

(ijk)

The ANOVA for the part of the factorial experiment actually run
from Table 8.1.2 and using a CRD as given for Eq. (8.1.2), is given
in Table 8.1.3.
8.1 LATIN SQUARE 215

TABLE 8.1.3

ANOVA Using Equation (8.1.2)

Source df EMS

Machine (M) 4 a2 + <|>(MmT) + <KmT) + 5<KM)

Men (m) 4 a2 + <f>(MmT) + <j>(MT) + 5<f>0)

Packing pressure (T) 4 a2 + <J>(MmT) + <J>(Mm) + 5<J>(T)

2
Residual 12 a + <KMmT) + <J>(Mm) + 4>(MT) + <J>(mT)

Within error 0 a2

There is no restriction error in this analysis because the

design is a CRD, taking only one-fifth of the 125 possible treatment
combinations. One should note that there has been a restriction
utilized in selecting 25 treatment combinations but for the analysis
given in Table 8.1.3 it is assumed that the restriction error is
negligible. References for developing the EMS are Wilk and
Kempthorne (1957) and Edwards (1962, p. 312). This design is not,
however, a usual fractional factorial. This concept will be
clarified in Chapter 10.

In the randomized complete block design, allowance was made in

the model for the possible interaction of the treatments and the
blocks. In the Latin square, no such allowance is usually made for
the tests, and if there happens to be an interaction present, the
estimates of the effects will be biased. For example, in Table
8.1.3, if men interact with machines, part of that effect goes into
the packing pressure effect which in turn may appear large when in
fact it is not.

Whether rows, columns and treatments are random or fixed does

not make much difference because all influences of the interactions
must be zero before the treatments can be analyzed properly.
216 8. LATIN SQUARE TYPE DESIGNS

Experimenters have used this design to reduce the number of

3 2
experimental units from t to t without understanding that they
were assuming the interactions were zero. When one looks at the
model and compares it with the analysis it is immediately known,
but the consequences of this assumption are not always appreciated.
To obtain a separate estimate of the interactions in one term and
still keep the number of observations rather small, Youden and
Hunter (1955) suggested repeating t of the combinations on the
diagonal.

The Latin square design is most valuable to those experimenters

who use only restrictions on randomization (not true factors of
interest to the experimenter) for the rows and columns and know
there are no interactions present. The design is also useful in
screening for major effects when all three factors are of interest
to the experimenter if the assumption is made that even if
interactions are present, they are small relative to the main
effects. Hence if the main effect mean square is large, the
experimenter~concludes it is due to the main effect more than the
interactions, and if an effect appears to be small it is concluded
that it has little influence on the variable being analyzed and
would not be used in future experimentation (see Cochran and Cox,
1957).

The Latin square design is not too useful for fewer than four
treatments nor more than eight (Kempthorne, 1952).

Example 8.1: In order to illustrate the calculational technique

for data from a Latin square design consider the coded data given
below which represent the firing time for the design given in Table
8.1.2. In order to eliminate confusion in the table of observations,
the packing pressures 10,000, 15,000, 20,000, 25,000, and 30,000 psi
are represented by the respective Latin letters A, B, C, D, and E.
8.1 LATIN SQUARE 217

Men

Machines Total

1 C 25 B 17 A 13 D 32 E 42 129
2 B 2 D 26 E 35 C 8 A 8 79
3 A 6 E 43 C 16 B 14 D 25 104
4 D 27 A 9 B 17 E 31 C 19 103
5 E 31 C 22 D 21 A __2 B 13 89
>tal 91 117 102 87 107 504

The treatment totals are obtained as follows:

A B C D E

13 17 25 32 42
8 2 8 26 35
6 14 16 25 43
9 17 19 27 31
_2 13 22 21 31
38 63 90 131 182

resulting ANOVA is shown in the following tabulation

Source df SS MS F
Columns (Men) 4 117.76 29.44
Rows (Machines) 4 284.96 71.24
Pressures 4 2598.96 649.74 42.44
Residual 12 183.68 15.31
Total 24 3185.36

where

Z I y. ., = 504
i j
CT="inr=10»160-64
Z Z y ? = 13,346
i j J
218 8. LATIN SQUARE TYPE DESIGNS

SS total = 13,346 - CT = 3185.36

SS columns = (912 + 1172 + 1022 + 872 + 1072) - CT

= 10278.4 - CT

= 117.76

SS rows = (1292 + 792 + 1042 + 1032 + 892) - CT

= 284.96

SS pressures = (382 + 632 + 902 + 1312 + 1822) - CT

= 2598.96

SS residual = SS total - SS columns - SS rows - SS pressure

= 3185.36 - 117.76 - 284.96 - 2598.96

= 183.68

8.2 ASSOCIATED DESIGNS

An associated design is the Graeco-Latin square which adds one

more restriction on randomization and uses Greek letters within the
2
t cells along with the Latin letters in such a manner that no
combination of Greek-Latin letters are repeated (Hicks, 1973,
pp. 79-80). In this case four factors can be investigated at once,
there may be restrictions on randomization, but if all combinations
4 2
were considered there is a reduction from t to t experimental
units. Another associated design is the Youden square (a form of
an incomplete Latin square design) which allows rectangular
arrangements (Hicks, 1973, Section 5.4). In addition, there are
systematic squares such as the Knut Vik, that have been used by
experimenters, but the experimental error is in question on these
designs even if there are no interactions (Kempthome, 1952, p. 330
Fisher, 1966, pp. 76-80; and Yates, 1964, pp. 318-319).
8.2 ASSOCIATED DESIGNS 219

A very useful design for experimenters confronted with the

problem that some of the experimental units are better than others
or time causes a change in the units is the cross over design.
Cochran and Cox (1957, pp. 127-141) show how a cross over design is
a special type of the Latin square design.

Example 8.2: An illustration of the cross over design occurred

in a problem for a drug manufacturing company. Two formulas of a
drug were given to 18 human beings selected at random from a
population to which the drug would be given. Nine random subjects
were given formula one and the other nine were given formula two.
The amount of the drug in the blood of each of the 18 subjects was
measured at specified intervals of time and the cumulative amount
was recorded.

After 7 days of ’’washout” period in which the drug supposedly

was excreted, another blood sample was taken and all 18 subjects
had zero amount of the drug in their blood samples. The experimenter
then had the group of nine who took formula one take formula two and
vice versa for the other group. This part of the experiment is
called the ’’cross over” part of the design.

A model for the analysis of this experiment is:

y. = y + S. + <$... + 0. + + F, + e,..,.
ijk i (i ) j '(j) k (ljk)

i = 1, 2, 18 j = 1, 2 k = 1, 2

where

th
^ijk = CUTnu^ative amount of drug in the blood of the i
subject after 8 hours given order j and formula k.

All other terms are comparable to those of Eq. (8.1.1).

The data from this experiment are given in the following

tabulation.
220 8. LATIN SQUARE TYPE DESIGNS

Subjects

Order 1 2 3 4 5 6 7 8 9

Formula One (351.7)

1 (Initial)
51.9 35.1 38.6 36.1 34.6 39.7 37.8 38.8 39.1
7 day "washout”

Formula Two (376.1)

2 (Crossover) 43.5 45.4 35.4 43.7 49.8 39.9 41.4 37.5 39.5

Total 95.4 80.5 74.0 79.8 84.4 79.6 79.2 76.3 78.6

Subj ects

Order 10 11 12 13 14 15 16 17 18

Formula Two (332.0)

1 (Initial)
50.8 41.1 39.1 35.7 33.7 31.2 34.3 31.8 34.3
7 day "washout"

Formula One (316.6)

2 (Crossover) 44.2 33.4 32.7 33.1 33.4 27.1 33.1 39.5 40.1

Total 95.0 74.5 71.8 68.8 67.1 58.3 67.4 71.3 74.4

ss order = Initial),2 .+ Crossover)2 _ fjGrand total^ = (cr)]

18 I 36 I

(683.7)2 + (692.7)2 (1376.4)2

18 36

SS formula = (Formula one)2 + (Formula two)2 _

18

, ( 6 6 8 V . (7 .8 .1 )2 . 52i62136 , ^

SS subjects - (,80.5)^ . . . . . (7S.4)2

- 52,624.36 = 717.79
8.2 ASSOCIATED DESIGNS 221

SS total = (51.9)2 + (35.1)2 + ... + (40.1)2

- 52,624.36 = 1,093.98

SS error = SS total - SS order - SS formulas - SS subjects

= 1093.98 - 2.25 - 44.00 - 717.79 = 329.94

The corresponding ANOVA is tabulated in the following.

Source df SS MS EMS

Subjects (Sp 17 717.79 42.22 a2 + 2a2 + 2a2

o b
Subj ect
2 0 2
restriction error 0 a + 2a6

Order (0^) 1 2.25 2.25 a2 + 18a2 + 184>(0)

Order restriction
2 2
error 0 aL + 18a
n
Formulas (F^) 1 44.00 44.00 a2 + 18<|>(F)
2
Residual (e,....) 16 329.94 20.62 o'
v (ljk)'
Total 35 1,093.98

To find the smallest error from these data we may test for
order

H : a + (j)(0) = 0
o n

2.25
< 1
*1,16 ” 20.62

Hence accept Hq at a > 0.25 and pool mean squares, which actually
results in a RCBD where subjects act as blocks and formulas are
treatments. The error mean square with 17 df is:

Error mean square = IUI2.2S).^7,(16),i(20.62) = ^

222 8. LATIN SQUARE TYPE DESIGNS

Subjects are random and no interest other than to provide the

inference space. The test on formulas is

F - = 2 25
1,17 19.54

ToKi^ c.
Tabled F a = 0.10
n = 3.03
1,17 a = 0.25 = 1.42

Hence the effect of formulas is not significant at a = 0.10 level.

Since the tabled ^(a = 0.10) = 2.71, the sample size

(number of subjects) could be increased indefinitely with these
results (F = 2.25) and there would not be a significant effect due
to formulas at the a = 0.10 level.

Problem 8.2.1. In a problem on processing of plastics, yield

strength of the material (coded for analysis purposes) was of
interest. The factors used in the experiment were:
Preconditioning temperatures 1, 2, 3,4
Preconditioning time 1, 2, 3,4
Materials a, b, c,d
The experimental layout and coded yield strengths are given in
the following tabulation.
Temperatures
Times 1 2 3 4
1 a 9.4 b 13.1 c 12.4 d 30.2
2 b 22.6 c 20.9 d 33.7 a 13.7
3 c 9.6 d 34.6 a 4.6 b 13.8
4 d 29.7 a 10.2 b 15.2 c 18.7

(a) Analyze the data (include the model, analysis of

variance, means of levels of all factors and standard errors).
Be sure to show your EMS.
(b) Discuss the design of the experiment and assumptions
involved for this design.
(c) Draw conclusions from the experiment.
(d) Assuming design is LS, determine the relative
efficiency of this Latin square design with the RCBD where
times are blocks. Also determine the relative efficiency of
this Latin square design with the CRD (refer to Section 5.4).
(e) Based on your results of part (d), how would you
design future experiments of the same type.

Problem 8.2.2. In a study comparable to that of Example 8.2

except that blood samples were not taken after the 7 days
8.3 REFERENCES 223

to check the "washout" period. The data are given in the

following tabulation.
Subjects

Order 1 2 3 4 5 6 7 8 9

Formula Two

1 Initial 51.9 35.1 38.6 36.1 34.6 39.7 37.8 38.8 39.1

Formula One

2 Crossover 58.5 60.4 50.4 58.7 64.8 54.9 56.4 52.5 54.5

Subjects

Order 10 11 12 13 14 15 16 17 18

Formula One

1 Initial 50.8 41.1 39.1 35.7 33.7 31.2 34.3 31.8 34.3

Formula Two

2 Crossover 62.2 51.4 50.7 51.1 51.4 45.1 51.1 57.5 58.1

(a) Analyze the data appropriately; showing the model and

ANOVA with df, SS, MS and EMS.
(b) Interpret the results.
(c) Does this design result in a RCBD? Explain.

8.3 REFERENCES

Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research,

McGraw-Hill, New York, 1952.
Anderson, V. L. Biometrics 26:255 (1970).
Cochran, W. G. and Cox, G. M. Experimental Designs, 2nd ed., Wiley,
New York, 1957.
Edwards, A. L. Experimental Design in Psychological Research, Holt,
Rinehart and Winston, New York, 1962, pp. 311-313.
Fisher, R. A. The Design of Experiments, 8th ed., Hafner, New York,
1966.
224 8. LATIN SQUARE TYPE DESIGNS

Fisher, R. A. and Yates, F. Statistical Tables for Biological,

Agricultural and Medical Research, 6th ed., Hafner, New York, 1963.
Hicks, C. R. Fundamental Concepts of Design of Experiments, 2nd ed.,
Holt, Rinehart and Winston, New York, 1973.
Kempthome, 0. The Design and Analysis of Experiments, Wiley, New
York, 1952. Distributed by Krieger Pub. Co., Huntington, New York.
Wilk, M. G. and Kempthome, 0. JASA 52:218 (1957).
Yates, F. Biometrics 20:307 (1964).
Youden, W. J. and Hunter, J. S. Biometrics 11:399 (1955).
 Chapter 9

2n FACTORIAL EXPERIMENTS (COMPLETE AND INCOMPLETE BLOCKS)

In Chapter 4, Section 4.2, a general discussion of factorial

experiments was presented. Whereas the meaning of factorial
experiments is that all possible combinations of the levels of the
factors exist in the experiment, this in no way describes the design
of the experiment as dictated by the randomization of the combinations
onto the experimental units. Hence, "factorial" refers only to the
type of experiment and not to the design or layout. In this chapter
the special case in which all factors of a factorial experiment have
only two levels is described and the design is assumed to be
completely randomized within each block.

In many experimental situations, the investigator would like

to examine the effects and interactions of many variables (n factors)
simultaneously on a dependent variable. The 2n factorial experiments
provide information on all possible two factor interactions if the
experiment is designed properly. In many cases the investigator
will follow an experiment of this type with another experiment with
more levels on a subset of the original factors in order to obtain
more detailed information on the most important factors. In order
to investigate non-linear trends one must have more than two levels
of the factors.

In the past it was felt that two leveled factorials were good
only for exploratory work by some research workers. Recently it
seems that the philosophy of experimenters has changed in this
regard. Many investigators believe it is better to introduce many
factors each with a high and low level into the experiment rather

225
226 9. 2n FACTORIAL EXPERIMENTS

than choosing arbitrarily a few factors and run many levels on each.
The reasoning for this whole approach seems to be that experimenters
who encounter many factors (independent variables) in their
investigations would prefer to have experimental, along with their
theoretical, evidence to discard independent variables rather than
using the theoretical approach only. The reason for this approach
is that the actual theory may be unknown but assumed based on past
experience. Thus, with the availability of rather efficient designs
of experiments involving two-leveled factorials this ’’theoretical
only” approach does not seem very scientific. In any case the
inference space must be carefully considered.

Other designs based on two-leveled factorials, including

composite designs, will be discussed in Chapter 13. This chapter
will deal mainly with techniques for this special case of two-leveled
factorials assuming CRD, and not discuss the randomization procedures
in detail as in former chapters.

9.1 COMPLETE BLOCKS

This section includes designs in which all the combinations of

the levels of the factors (treatment combinations) in the experiment
are assumed to be handled under similar conditions called a block.
Usually if no repeats of the treatment combination occurs and the
treatment combinations are completely randomized within a block,
the higher factor interactions are frequently assumed zero and the
resulting mean square is used as the error for testing the main
effects and lower factor interactions. As discussed in Chapter 4,
however, the usual approach is to repeat a few treatment combinations
on other experimental units and use the appropriate differences as
an estimate of error. References for all this material are
Kempthorne (1952, Chapter 13); Hicks (1973, Chapter 7), and John
(1971, Chapter 7).
9.1 COMPLETE BLOCKS 227

9.1.1 The 22 Factorial

This is the simplest of factorial experiments. It is hardly

ever used in actual experimentation but it is used here to
demonstrate many basic notions of two-leveled factorials.

Consider an example in metals. The variable to be analyzed

is y, heat loss per unit of time, and the treatments are factor a,
hardness of metals (soft, hard) and factor 6, metals (alloy, steel).
The interests are in finding out the effects of hardness of metals
over alloy and steel and of metals over both soft and hard on heat
loss per unit of time (main effects) or the effects of each of the
two factors over the levels of the other factor; plus the effect of
certain combinations of hardness and metals on heat loss per unit
of time (interaction of the two factors).

In the layout of the experiment small letters indicate the

treatment combinations and yields for the corresponding treatment
combinations depending on context, and in the analysis capital letters
indicate the main and interaction effects. In the 2 x 2 example a^
indicates soft and a^, hard; b^ indicates alloy and b^ steel. Hence
aQb0 indicates the treatment combination of soft alloy, and so on.
There are various ways to portray the treatment combinations compared
to aQbQ and the remaining treatment combinations. Two other
corresponding notations commonly used are as follows:

Notations

1 2 3
( 1) 00
a 10
b 01
ab 11

The association is obvious; however, it should be recognized that

notation 2 can be used only for two-leveled factorials (either
228 9. 2n FACTORIAL EXPERIMENTS

presence or absence concept), while notation 3 is general. That is,

if there are three levels one merely uses 112” for level three, and
so on for any number of levels.

The data may be organized as follows.

__________ b__________
a __________ 0____________ 1_____________ Total____
0 00 or (1) 01 or b (1) + b
1 10 or a 11 or ab a + ab

Total (1) + a b + ab (1) + a+ b + ab

where notation 3 is a matrix indicator. In all instances the

notation for the yields of the treatment combinations and the
treatment combinations are not distinguishable but it is hoped that
there will be no confusion for analysis purposes.

The analysis of variance is given in Table 9.1.1.

TABLE 9.1.1

ANOVA Using a 22 Factorial CRD

Source df SS

A 1 [a + ab - (1) - b]2/4
B 1 [b + ab - (1) - a]2/4
AB 1 [ab + (1) - a - b]2/4

The total effects are defined as

[A] = [ab - b + a- (1)]

[B] = [ab + b - a- (1)]
[AB] = [ab - b - a+ (1)]

See Kempthome(1952, Section 13.2.1). Notice the correspondence

of the total effects and the sums of squares in the analysis of
9.1 COMPLETE BLOCKS 229

variance. Really these are merely contrasts (defined in Section

1.4.1) or comparisons and they are orthogonal. The sum of squares
for each constrast is given by the square of the contrast divided
by the sum of the squares of the coefficients within the contrast,
e.g., SS A = [A]2/[(l)2 + (l)2 + (-1)2 + (-1)2]. Note that the
denominator is c? where the coefficient, c., is either (-1) or
(+1). Many times investigators are interested in average effects,
e.g., A = (1/2) [ab - b + a - (1)]. Kempthome (1952, p. 236)
explains average effects in detail.

One way to display the interrelationship of treatment

combinations and effects is given in Table 9.1.2.

TABLE 9.1.2

2
Treatment Combinations and Effects in 2

Treatment combination
Average effect
Effect (i) a b ab divisor
A - + + 2
B - + + 2
AB + - + 2
Mean + + + + 4

Problem 9.1.1. (a) Show that A, B, and AB are orthogonal

comparisons.
(b) Show that the sums of squares calculated using the
usual analysis method is equivalent to that given in Table
9.1.1.
(c) Could the comparison method of obtaining the sums of
squares be used if there were three levels? Explain.

9.1.2 The 25 Factorial

3
The eight treatment combinations for the 2 factorial using
factors a , b, and c are: (1), a, b, ab, c, ac, be, abc. Notice
how the treatment combinations are generated by multiplication. As
2
in the 2 factorial the relationship between the effects and
3
treatment combinations for the 2 factorial are shown in Table 9.1.3.
230 9. 2n FACTORIAL EXPERIMENTS

TABLE 9.1.3
3
Treatment Combinations and Effects in 2

Treatment combination
Effect d)_ a b ab c ac be abc Divisor
A + - + - + - + 4
B - - + + - - + + 4
AB + - - + + - - + 4
C - - - - + + + + 4
AC + - + - - + - + 4
BC + + - - - - + + 4
ABC - + + - + - - + 4
Mean + + + + + + + + 8

As before all main effects have a plus (+) sign when the small
letter is. present in the treatment combination and a minus (-) when
absent. To get the signs for the interactions multiply the
corresponding main effect rows. Multiplication of like signs is
defined to give a positive sign and multiplication of unlike signs
is defined to give a negative sign. The analysis of variance for a
3
2 factorial experiment is given in Table 9.1.4.

TABLE 9.1.4

ANOVA Using 2^ Factorial

Source df

A 1
B 1
AB 1
C 1
AC 1
BC 1
ABC __1
Total 7
9.1 COMPLETE BLOCKS 231

9.1.3 The 2n Factorial

The procedure for writing down the signs of the treatment

combinations that go into each main effect and interaction for the
2n factorial may be extended from the 2^ factorial. There is another
method for writing these down, however, from the average effects as
follows:

A = [(a-l)(b+1)(c+1)(d+1)(e+1) ....]

AB tCa_ 1) (b-l) (c+1) (d+1) (e+1) ....]

ABD = -jjty [(a-l) (b-l) (c+1) (d-1) (e+1) --- ]

and so on,where the products inside the square brackets will

produce the treatment combinations with the correct sign to obtain
the designated effect or comparison. A reference for this method
is Kempthome (1952, p. 246).

Analysis of variance for 2n factorial is given in Table 9.1.5.

TABLE 9.1.5

ANOVA Using 2n Factorial

Source df

Main effects n

2 factor interactions n Ci_-. y.

2
n(n - 1) (n - 2)
3 factor interactions
3!

n factor interaction 1

Total 2n - 1
232 9. 2n FACTORIAL EXPERIMENTS

Problem 9.1,2. Write out the treatment combinations with

appropriate signs for the interactions ABC, ABD, and ABCD in
4
a 2 factorial.

Example 9.1: An experiment which illustrates the 2n factorial

(completely randomized in each replication) arose in the finishing
of metal strips in a metallurgical process. The measured response
was a score for smoothness of the surface finish where a small
reading was desirable and a large number indicated roughness. The
factors (two levels of each) which were included for this experiment
were as follows.

Factor Levels
Solution temperature (T) Low High
Solution concentration CQ Low High
Roll size (R) 1 2
Roll tension (F) Low High
Replication (blocks) (B) I II

If all the assumptions for an ANOVA are met, the ANOVA in this
case is similar to that given in Table 9.1.5 except that one must
account for the restriction on randomization in each replicate. In
addition one will have a measurement of error from the block by
treatment interactions.

The data for this experiment are as follows :

(y = smoothness score).
9.1 COMPLETE BLOCKS 233

Treatment combination Replication

t c r f I II
L L 1 L 18 16
L L 1 H 10 12
L L 2 L 10 12
L L 2 H 8 19
L H 1 L 9 7
L H 1 H 9 7
L H 2 L 8 10
L H 2 H 14 10
H L 1 L 12 16
H L 1 H 21 14
H L 2 L 15 15
H L 2 H 21 21
H H 1 L 17 15
H H 1 H 24 18
H H 2 L 4 4
H H 2 H 13 13

Total 213 209

The calculation of the sums of squares may be carried out

following the usual procedure using blocks as a fifth factor, if
one has a computer program available. However, if the calculations
are to be done by hand the method of using orthogonal contrasts as
described in Table 9.1.1 will beeasier. To illustrate this
equivalence, consider the sum of squares for replicates. Thenormal
method of calculation yields

_ (213)2 + (209)2 (213 + 209)2 _ r

SS B ------- ^ ------------- T l ------- 0.5

The replicate contrast is [B] = 209 - 213 = -4 and SS B = (-4)^/32

= 0.5. The main effect contrast is easily displayed by recalling
that the low level of thefactor is associated with a negative sign
and the high level with a positive, e.g.,
234 9. 2n FACTORIAL EXPERIMENTS

[T] = 12 + 16 + 21 + 14 + ... + 13 + 13 - 18 - 16 - 10 - ...

- 14 - 10
= 243 - 179 = 64

and

SS T = (64)2/32 = 128.0

In order to illustrate the calculation of the two factor

contrasts and sums of squares the detailed calculation for the TC
interaction is displayed. The two-way table of totals for these
two factors is

C
L (-1) H (+1)
(+) 105 (-) 74

(-) 135 (+) 108

where the -1, -1 cell is obtained by the sum

18 + 16 + 10 + 12 + 10 + 12 +8 + 19 = 105
One always has a check on the celltotals as the totalfor allcells
should give the grand total of 422. The appropriate sign for the
cell total is obtained by multiplying the column sign times the row
sign and placing the result in the corresponding cell. Now

[TC] = (105 - 74 - 135 + 108) = 4

and
SS TC = (4) 2/32 = 0.5

The calculation of the three-way interaction follows in the

same fashion except that a three-way table must be formed, e.g.,
consider the TCF interaction. The three-way table is shown here.
9.1 COMPLETE BLOCKS 235

T C L(-l) H(+l)

(-) (18+16+10+12) (+)(10+12+8+19)

L(-l)
= -56 = +49
L(-l)
(+)(9+7+8+10) (-)(9+7+14+10)
H(+l)
= +34 = -40

(+) (12+16+15+15) (-)(21+14+21+21)

L(-l)
= +58 = -77
H(+l)
(-)(17+15+4+4) (+)(24+18+13+13)
H(+l)
= -40 = +68

Note that the sum of the values of the cells prior to the sign
alteration checks out to a total of 422. The TCF contrast is given
by

[TCF] = -56 + 49 + 34 - 40 + 58 - 77 - 40 + 68

= -4
and
SS TCF = (-4)2/32 = 0.5

The calculation for all the remaining effects may be carried

out in a similar fashion. The resulting ANOVA table is as follows:
236 9-. 2n FACTORIAL EXPERIMENTS

„ .Based on.
Source df MS F(verror J
J
Replicates (B) 1 0.5
Restriction error 0

Treatments 15
T 1 128.000 15.4*
C 1 105.125 12.6*
TC 1 0.500 < 1
R 1 24.500 2.9
TR 1 36.125 4.3
CR 1 32.000 3.8
TRC 1 136.125 16.4*
F 1 66.125 8.0*
TF 1 72.000 8.7*
CF 1 15.125 1.8
TCF 1 0.500 < 1
RF 1 40.500 4.9*
TRF 1 3.125 < 1
CRF 1 2.000 < 1
TCRF 1 6.125 < 1
B x treatments 15 8.833

^ /Reps and B x Trt.\

Error! or J 16 8.312
Within treatments/

*Significant at the a = 0.05 level.

The significance of the three factor interaction TRC should motivate

one to set up a table of means in order to visualize the physical
relationship among these factors. A Newman-Keuls test may be
considered to explore the individual means.

^Since the mean square for replicates is so small the design

may be considered CRD and the error is within treatments.
9.1 COMPLETE BLOCKS 237

Example 9.2: The material presented in Example 9.1 was

analyzed using the method of orthogonal contrasts. The calculational
procedure, even though it is easily understandable, is somewhat time
consuming. A faster technique for doing hand calculations is Yates*
(1937) method. Another reference for this procedure is Davies 1971.
This procedure, which only applies to 2n experiments, is illustrated
by utilizing the data of Example 9.1. The calculations are given
below. Replicates I and II are added together and the analysis is
4
carried out as if the experiment was a 2 . The explanation of the
method of calculation follows.

The first step in this procedure, which is illustrated below,

is to list the treatment combinations in the systematic fashion
using notation 2 described in Section 9.1.1. The appropriate
observation is listed next to the treatment combination. In this
example it is the total of r = 2 observations. This analysis could
have been carried out by listing twice as many treatment combinations,
taking the individual replications into account, and using r = 1.
The first half of the values in column (1) are obtained by adding
together successive pairs of observations from the previous column
e.g., 56 = 34 + 22 and 49 = 22 + 27. The last half of the values
in column (1) are obatined by subtracting the first member of a pair
from the second member in the observation column, e.g., -12 = 2 2 - 3 4
and 5 = 27 - 22. Column (2) is then generated from column (1) in the
same fashion that column (1) was generated from the observation
column. This process is continued until column n is generated from
column n - 1 for a 2n factorial. The calculation of the average
effect and sum of squares columns is self-explanatory. The source
column is written down by changing the small Latin letter in the
treatment combination column to a capital Latin letter.

The ANOVA shown in Example 9.1 could now be generated by merely

calculating the replicate and total sum of squares.
 Total Sum of
Average 238
Treatment Observation Effect squares
effect
Source combination sum of r=2 values Cl) (2) (3) (4) (4)/24-!r (4)2/24r
CD 34 56 105 179 422
F f 22 49 74 243 46 2.875 66.125
R r 22 32 135 -1 -28 -1.750 24.500
RF rf 27 42 108 47 36 2.250 40.500
C c 16 63 -7 3 -58 -3.625 105.125
CF cf 16 72 6 -31 22 1.375 15.125
CR cr 18 74 19 23 -32 -2.000 32.000
CRF erf 24 34 28 13 -8 -0.500 2.000
T t 28 12 -7 -31 64 4.000 128.000
TF tf 35 5 10 -27 48 3.000 72.000
TR tr 30 0 9 13 -34 -2.125 36.125
9.

TRF trf 42 6 -40 9 -10 -0.625 3.125

TC tc 32 7 17 17 4 0.250 0.500
TCF tcf 42 12 6 -49 -4 -0.250 0.500
TCR ter 8 10 5 -11 -66 -4.125 136.125
TCRF tcrf 26 18 8 3 14 0.875 6.125
2“ FACTORIAL EXPERIMENTS
9.1 COMPLETE BLOCKS 239

Problem 9.1.3. Analyze the data of Example 9.1 as if it were

a 2^* experiment using Yates procedure or a digital computer.
Calculate the error mean square by pooling the replicate and
all fifteen (replicate by treatment sums of squares). This
error sum of squares should agree with that shown in Example
9.1.

Problem 9.1.4. A steel company was interested in causes of

weld breaks in its plant and set up an experiment that had the
following factors and coded levels:
Levels
Factor Low High
w Tap Setting 1 2
(fc) Time of Weld 5 7
(e) Gage bar setting 3 4
GO Upset (real) L H
(e) Upset cycles 1 3

When welding steel in this company two strips to be welded

are placed in position so the edges to be welded are facing
each other. The thickness of the strips is measured and a
setting to control the voltage is determined. This setting is
called "tap setting.” The strips are then clamped in dies.
The two strips are manually brought together until the distance
between them is small enough to permit an electric arc to begin
(caused by the different polarity of the dies on each strip).
The start of the arc is termed the ”start of the flash time”
and initiates the following automatic cycle. The arc or flash
continues for a given time, and is followed by a heat cycle.
During the two cycles the strips are automatically traversed
towards one another and are literally melted together. The
final cycle, termed the ’’upset,” forcibly jams the strips
together and finishes the cycle. The force or "upset (real)”
in this experiment was set at L (low) and at H (high). The
number of the "upset cycles” was either 1 or 3.
The distance the strips move toward each other during the
flash and heat cycles is determined by a bar inserted in the
proper place on the welder, with various thicknesses of this
bar chosen as a function of the strip gage. The thickness
chosen, and the resultant distance of the strip traverse
during the weld cycle is termed and detemined by the "gage
bar setting.” The total time from start of flash through upset
is termed the "time of weld.” Everything from the start of
flash through upset is automatic, and determined by various
cycle times and welder settings.
A replicated CRD factorial experiment (really a RCBD) was
set up for a certain grade and gage of steel. Assume that
replications are random and all five factors are fixed. The
ultimate strength of the weld was measured and the following
coded data were obtained:
240 9. 2n FACTORIAL EXPERIMENTS

Replication I Replication II
CL b C d £ Strength (coded) a 6 c d £ Strength 1
1. 1 7 3H 1 62 3. 2 7 4 L 1 42
2. 2 7 4 L 3 45 6. 2 7 4H 3 39
3. 2 7 4 L 1 50 10. 2 5 4 L 3 62
4. 2 5 3 L 1 49 16. 1 7 4 L 3 48
5. 1 5 3H 3 60 21. 1 5 4H 1 47
6. 2 7 4H 3 61 29. 1 7 3H 3 61
7. 2 7 4 H 1 60 26. 2 5 3H 3 57
8. 2 5 4H 3 56 19.--1 5 4 L 3 61
9. 1 5 3H 1 59 13. 2 5 3 L 3 49
10. 2 5 4 L 3 61 2. 2 7 4 L 3 60
11. 1 7 4H 1 52 8. 2 5 4 H 3 43
12. 1 7 3 L 3 60 15. 1 7 3 L 1 61
13. 2 5 3 L 3 61 18. 2 5 4 H 1 41
14. 2 5 4 L 1 60 1. 1 7 3H 1 62
15. 1 7 3 L 1 62 20. 2 5 3H 1 61
16. 1 7 4 L 3 47 28. 1 7 4H 3 45
17. 2 7 3 L 3 58 31. 2 7 3 L 1 56
18. 2 5 4H 1 60 25. 1 5 3 L 1 59
19. 1 5 4 L 3 57 17. 2 7 3 L 3 61
20. 2 5 3H 1 61 7. 2 7 4H 1 61
21. 1 5 4H 1 61 4. 2 5 3 L 1 49
22. 2 7 3H 1 62 9. 1 5 3H 1 60
23. 2 7 3H 3 57 22. 2 7 3 H 1 56
24. 1 7 4 L 1 44 27. 1 5 4 L 1 60
25. 1 5 3 L 1 60 12. 1 7 3 L 3 61
26. 2 5 3H 3 61 14. 2 5 4 L 1 46
27. 1 5 4 L 1 58 23. 2 7 3 H 3 61
28. 1 7 4H 3 56 30. 1 5 4H 3 61
29. 1 7 3H 3 60 32. 1 5 3 L 3 60
30. 1 5 4H 3 61 24. 1 7 4 L 1 30
31. 2 7 3 L 1 53 5. 1 5 3H 3 60
32. 1 5 3 L 3 61 11. 1 7 4 H 1 50
9.2 INCOMPLETE BLOCKS 241

The usual ANOVA for this type of experiment (refer to Section

5.2) is:
______ Source___________ df
Replications (R) 1
Restriction error (<S)_____0
Treatments (T) 31
Pooled error (R x T) 31

(a) Using models, show what assumptions are required to

use the pooled error term in the above ANOVA.
(b) Analyze the data with the pooled error, but split up
the 31 df for treatments appropriately. Then set up tables of
means with the appropriate standard errors to allow Newman-
Keuls1 tests where needed.
(c) Discuss the inference space.
(d) Later in the investigation it was determined that
high carbon steel, impurities and grade differences were the
major causes of weld breaks. What, then, did the results of
this experiment contribute to solving the weld break problem?

9.2 INCOMPLETE BLOCK

If the total number of treatment combinations in a factorial

experiment is too large to be handled under similar conditions or
in one block, more blocks with only a fraction of the total number
of treatment combinations is used in each block. Assuming that the
treatment combinations that are chosen (purposely) for each block
are randomly assigned to the experimental units within the blocks,
we have what is called a randomized incomplete block design.

The main problem is to assign certain sets of treatment

combinations to blocks so that the effect or interactions confounded
with blocks is not important to the experimenter. It is immediately
obvious that one must lose information on as many effects and/or
interactions as there are degrees of freedom for blocks.

2
9.2.1 The 2 Factorial

Let us assume that only two treatment combinations of the four

2
in a 2 factorial can be run on a given day. If main effects are
242 9. 2n FACTORIAL EXPERIMENTS

more important than the interaction and there is an estimate of the

error variance cr“ , we should set up the design as

Day 1 (1) ab
Day 2 a b

To compare Day 1 vs Day 2 we could use

[(1) + ab - a - b]

which is exactly the comparison used to estimate the AB interaction

effect. Hence Days are completely confounded with AB. This leaves
us with complete or full information on the two main effects A and
B, but no information on AB.

The analysis of variance for this incomplete block design is

given in Table 9.2.1.

TABLE 9.2.1
2
ANOVA Using 2 Blocks of 2 in a 2 Factorial

Source df

AB and/or days 1
Restriction error 0
A 1
B J.
Total 3

Problem 9.2.1. Show the corresponding design if A is

confounded with days.

Kempthorne (1952, p. 244) shows the variance of the effects

and interactions.
9.2 INCOMPLETE BLOCKS 243

9.2.2 The 23 Factorial

Assume we can run only 4 observations of the 8 under similarly

3
controlled conditions, how can we design a 2 experiment? Usually
the three factor interaction is least important so we will confound
ABC with blocks. Using the general approach of putting the four
treatment combinations with (+) sign in one block and the four with
(-) sign in the other block we get the experiment

Block 1 (-) Block 2 (+)

(1) a
ab b
ac c
be abc

Note: Once we have block one we can take any remaining treatment
combination, say a, and multiply the treatment combinations
in block 1 by a to generate block 2;

a • (1) = a

a • ab = a^b = b

2
a • ac = a c = c

a • be = abc

The basis for this generation is given by Kempthome (1952) and

Winer (1971, p. 606). Since there are only two levels possible in
this design and we define the levels as 0 and 1, the design has a
basis of modulo 2. Modulo 2 means dividing the number obtained
(here 2) by 2 and using the remainder (here 0) as the appropriate
2 0
number. Hence a b = a b = b, and so on.
3
The analysis of variance for 2 in blocks of 4 is given in
Table 9.2.2.
244 9. 2n FACTORIAL EXPERIMENTS

TABLE 9.2.2
3
ANOVA Using 2 Factorial in 2 Blocks of 4 Each
Source df
ABC and/or Blocks 1
Restriction error 0
A 1
B 1
AB 1
C 1
AC 1
BC _1
Total 7

There is full information on all main effects and two factor

interactions but we lose the information on ABC completely if there
is a difference between blocks. Of course, one must have information
on the error variance for this design to be of practical value.

Consider the case of four blocks of two. There must be three

degrees of freedom confounded with blocks. It can be shown that if
A is confounded with blocks and B is confounded with blocks, then
AB is also confounded with blocks. This makes up one possible
design. Another possibility is if AB and BC are confounded with
2 2 o
blocks then AB C or AC is also confounded, since AB C = AB C =
A*1*C = AC, modulo 2.

_________ AB__________
BC - +

(b), (ac) (ab), (c)

+ (a), (be) (1), (abc)

Consider the interaction of this table, i.e., [the block with (b)
and (ac) plus the block with (1) and (abc)] minus [the block with
(a) and (be) plus the block with (ab) and (c)] is the same as
9.2 INCOMPLETE BLOCKS 245

[AC] = [(1) - a + b - a b - c + a c - b c + abc]

Hence AC is confounded with the blocks for the interaction. This

notion holds in general and AC is an example of what is called
generalized interaction. A good reference in addition to Kempthome
(1952) is Davies (1971, Chapter 9).

9.2.3 General Approach

Let us switch to a more general approach which can be used on

other factorials but will be demonstrated on a 2^ factorial
[Kempthome (1952) explains the procedure].

Consider a 25 in four blocks of eight. There must be three

degrees of freedom confounded with blocks. One procedure is to
confound ABD and ACE. It follows that the generalized interaction
is ABD x ACE = BCDE. Hence these three interactions are confounded
with blocks. Let all treatment combinations that satisfy the
following equations called defining equations

ABD : x, + x0 + x. = 0, mod 2
1 2 4
ACE : + x^ + X(- = 0, mod 2,

be in block 1.This block iscalled the intrablock subgroup since

it contains the treatment combination 00000 or control. In order
to determine the remaining composition of the intrablock subgroup
which gives rise to this confounding scheme one writes the two
defining equations
x. + x0 + x. = 0 mod 2
1 2 4
xl + Xg + x$ = 0 mod 2

where thepresence of x. in one of these equationsindicates the

th ^
presence of the i factor in the interaction term which is being
confounded e.g., for ABD we include x^, x2 > and x^, respectively.
The values of each x^ are either 0 or 1 depending on the level of
factor which corresponds to x^. Using the models and the above
procedure we arrive at the same result as expressed in block 1 that
246 9. 2n FACTORIAL EXPERIMENTS

follows in Table 9.2.3. Blocks 2, 3, and 4 are obtained in the

usual multiplication manner described previously.

TABLE 9.2.3

A Design of a 25 Factorial
Experiment in 4 Blocks of 8

ABD: 0 1 1 0
ACE: 0 1 0 1
a b c d e Block 1 Block 2 Block 3 Block 4
0 0 0 0 0 = (1) a b c
0 0 1 0 1 = ce ace bee e
0 1 0 1 0 = bd abd d bed
0 1 1 1 1 = bcde abcde cde bde
1 0 0 1 1 = ade de abde acde
X 0 1 1 0 = acd cd abed ad
1 1 0 •o 1 = abe be ae abce
1 1 1 0 0 = abc be ac ab

The analysis of variance for the design given in Table 9.2.3 is

given in Table 9.2.4.

TABLE 9.2.4

ANOVA Using the Design in Table 9.2.3

Source df

Blocks and/or (ABD, ACE, BCDE) 3

Restriction error 0

Main effects 5
2 factor interactions 10
Error 13
(assume 3 factor and higher
interactions are zero)

Total 31
9.2 INCOMPLETE BLOCKS 247

In many experiments the blocks must be small to retain

homogeneity of the experimental units but sometimes there is enough
experimental material that many replications of the whole experiment
may be made. In this case the experimenter would like to know how
best to use the experimental material he has. In most instances,
where information on main effects and/or interactions is lost between
blocks, partial information is desired on these effects and/or
interactions and, if possible, equal information is quite desirable.

Let us consider an example of a design that has equal

information on all main effects and interactions which is called a
balanced incomplete block design (BIBD). There are two factors
each at two levels and three replicates are possible, but there can
be blocks of two experimental units only. If information on main
effects and interactions are of equal interest, the design in Table
9.2.5 seems best.

TABLE 9.2.5

BIBD of a 22 Factorial Using 3

Replications of 2 Blocks of 2

Replication

I II III
Block 1 Block 3 Block 5

(1) b (1) a (1) ab

Block 2 Block 4 Block 6

a ab b ab a b

The corresponding analysis of variance is given in Table 9.2.6.

248 9. 2n FACTORIAL EXPERIMENTS

TABLE 9.2,.6

ANOVA of the BIBD in Table 9.2.5

df
Replications 2
Blocks in Replications 3
Restriction error 0

A 1 Two-thirds information on
each. The sum of squares
B 1
are computed from those
AB 1 replicates in which the
effect or interaction is
Error _3_
not confounded.
Total 11

Of course any multiple of three replicates will give equal

information on the main effects and interactions if the design is
’’balanced1’ as that given for the three replicates.

(Note: In all systems of confounding, care must be taken to

assure the experimenter that the effects or interactions of least
interest are the ones confounded with blocks. It is not necessary
that the effects or interactions to be confounded are small, only
that they are of least interest.)

One must remember that the requirement for a negligible effect

or interaction is necessary for using that source of variation for
the estimate of the error but not for confounding purposes. This
idea is sometimes confusing to the novice in designing experiments,
but is never a difficulty if the reasons are understood.

Example 9.3: An example from the area of Animal Science

concerning the gain in weight of lambs over a two month period will
illustrate the necessary calculations required for computing the
sums of squares from data for a BIBD design. The design is the
same as that shown in Table 9.2.5. The use of 6 pairs of twin
lambs provide the 12 experimental units and each set of twins
constitute the blocks. The two factors A and B are low and high
9.2 INCOMPLETE BLOCKS 249

amounts of vitamin A and protein respectively. The coded data for

these results follow:

Replication

II III
Block 1 Block 3 Block 5

(1) = 6 b = 12 (1) = 6 a = 7 (1) = 6 ab = 12

Block 2 Block 4 Block 6

a =8 ab = 17 b = 13 ab = 17 a =9 b = 10

The calculations of the various sums of squares is shown below.

The treatment sum of squares must be modified so that, e.g., the
sum of squares for A is calculated by using only the replicates
that do not have A confounded with blocks, i.e., replicates II and
III.

I y = 123

£ y2 = 1437

CT = ^12* = 1260-75

TSS = 1437 - 1260.75 = 176.25

SS „ ps . («)2 > (43)2 ♦ (37)2 _ CT

= 1266.75 - 1260.75 = 6.00

2 2 2 2 2 2
SS ui ir m• reps = --------------
cc blocks 18 + 25 + 13 + 30z + 18 + 19
*--------------- 10„
1266.75

= 1351.50 - 1266.75

= 84.75
250 9. 2n FACTORIAL EXPERIMENTS

(7 + 17 + 12 +• 9)2 + (6 + 13 + 6 + 10)2 (45 + 35)2

SS A =
4 8

= 812.5 - 800 = 12.5

(12 + 1 7 + 1 2 + 10)2 + (6 + 8 + 6 + 9)2 (51 + 29)2

4 8

= 860.5 - 800 = 60.5

(6 + 17 + 6 + 17) 2 + (8 + 12 + 13 + 7) (46 + 40)2

4 8

= 929 - 924.50 = 4.50

SS error = 176.25 - 6.00 - 84.75 - 12.50 - 60.50 - 4.50

= 8.00

The resulting ANOVA table is as follows:

Source df SS MS F
Replication 2 6.00
Blocks in reps 3 84.75
Restriction error 0
A 1 12.50 12.50 4.68
B 1 60.50 60.50 22.66*
AB 1 4.50 4.50 1.69
Error _3 8.00 2.67
Total 11 176.25

♦Significant at a = 0.05

Problem 9.2.2. Repeat the exercise shown in Example 9.3 using

only replicates I and II. Comment on the results of this
problem as compared to the results of Example 9.3.

Problem 9.2.3. Consider the effects of lubrication, and other

factors on life of gears. The factors to consider are weight
of load: light and heavy; type of load: shock and steady;
speed of operation: low and high; and application of
lubrication: spray and bath.
9.3 REFERENCES 251

If the experimenter can run only eight treatment

combinations under similar conditions at one time, design the
most efficient experiment, including a suggested number of
replications, and show your analysis of variance. Assume
three factor and higher interactions are negligible and that
the most interest is on the main effects and two factor
interactions.

9.3 REFERENCES

Davies, 0. L. (Ed.) Design and Analysis of Industrial Experiments,

2nd ed., Oliver and Boyd, Edinburgh, 1971.
Hicks, C. R. Fundamental Concepts of Design of Experiments, 2nd ed.
Holt, Rinehart and Winston, New York, 1973.
John, P. W. M. Statistical Design and Analysis of Experiments,
Macmillan, New York, 1971.
Kempthorne, 0. Design and Analysis of Experiments, Wiley, New York,
1952. Distributed by Krieger Pub. Co., Huntington, N.Y.
Winer, B. J. Statistical Principles in Experimental Design, 2nd ed.
McGraw-Hill, New York, 1971.
Yates, F. Design and Analysis of Factorial Experiments. Imperial
Bureau of Soil Sciences, Harpenden, England, 1937.
 Chapter 10

FRACTIONAL FACTORIAL EXPERIMENTS FOR TWO-LEVELED FACTORS

In almost all experiments the investigator would like to reduce

the number of observations required for a complete factorial. If
certain assumptions can be met the use of fractional factorials is
a most efficient technique to reduce the number of observations and
still obtain the desired information. The usual fractional
factorial is still orthogonal, which means that certain effects are
estimated independently of one another; however, fewer effects can
be estimated and these may be biased. The usual procedure given by
many authors including Kempthorne (1952, Chapter 20), Hicks (1973,
Chapter 15, Davies (1971, Chapter 10), and Peng (1967, Chapter 7)
is the one that will be discussed in this chapter.

The usual fractions given in textbooks are powers of 1/2 for

two-leveled factorials. These will be described briefly.
Statisticians have developed a set of other fractions such as 3/4
which have experimental possibilities because of their flexibility
(see John, 1961, 1966, and 1971, Chapter 7).

This chapter will cover two-level factors only, but one of the
methods of fractional replication given here can be applied to any
prime number equal leveled factorial such as 3, 5, 7, and so on.
Besides, if a nonprime number such as 4, 8, 9, and so on occur then
2 3 2
the experimenter can convert the factorial into a 2 , 2 , 3 , and
so on, which can be fractionated easily (see John, 1961 and 1966).

252
10.1 THE 1/2 REPLICATION 253

10.1 THE 1/2 REPLICATION

The method of obtaining a 1/2 fractional replicate is merely

to choose one of the two blocks that are generated by the techniques
of Section 9.2. There are certain procedures available that will
show quickly whether the proposed design is very good or not. The
techniques discussed here can easily be applied to higher leveled
factorials and will be shown in a Chapter 11.
3
Using the incomplete block design from the 2 factorial for
demonstration purposes, let us assume that the experimenter has
three factors ( a , b and c) each at 2 levels and only four
observations are to be used in an experiment. Hence, we have a 1/2
3
replication of 2 factorial. If we assume the interactions are
zero what is the best design?
Let us go back to Table 9.1.3 of main effects and interactions
to see how the treatment combinations are used in a fractional
factorial. This table is reproduced as Table 10.1.1 for convenience
purposes.

TABLE 10.1.1

Treatment Combinations and Effects

3
in a 1/2 Replication of a 2

Treatment combination

Effect a) a b ab c ac be abc

V + + + + + + + +
A - + - + - + - +
B - - + + - - + +
AB + - - + + - - +
C - - - - + + + +
AC + - + - - + - +
BC + + - - - - + +
ABC - + + - + - - +
254 10. FRACTIONALS FOR TWO-LEVELED FACTORS

The randomized incomplete block design from Chapter 9 for 2 blocks

of 4 when the interaction ABC is confounded with blocks is

(-) (+)
Block 1 Block 2
(1) a
ab b
ac c
be abc

If we throw away Block 1, we would have our required four

observations. What do the contrasts or comparisons look like?
(Note Table 10.1.1 and select only the treatment combinations
a, b, c, and abc which are reproduced in Table 10.1.2.)

TABLE 10.1.2

Treatment Combinations and Effects

for the 1/2 Fractional Replication

Treatment combination

Effect a b c abc

A + - - +
BC + - - +
B - + - +
AC - + - +
C - - + +
AB - - + +
U + + + +
ABC + + + +

Hence, the effect for A will be calculated the same as the BC effect;
similarly for B and AC; C and AB; and y and ABC. This feature is
the basis for saying A is an alias of BC, and so on to y is an alias
of ABC, because in this particular fractional replication whenever we
measure the effect of A we also measure the effect of BC and so on.
10.1 THE 1/2 REPLICATION 255

Now that we see what a fractional replicated experiment looks

like and we assume the treatment combinations are completely
randomized onto the experimental units for the design, let us
develop a procedure that will allow us to examine the value of the
design quickly.
3
Given we want a 1/2 replicate of 2 , assuming interactions are
zero, let us try to design an experiment with the most information.
Since ABC is the highest factor interaction, in most cases, it
should be confounded with blocks and this automatically (notice
previous development) makes the mean, y, an alias of ABC. Using
our modulus 2 concept we may write the defining relation or identity
relationship (Kempthome, 1952, p. 395) or defining contrast
(Davies, 1971, p. 393) as:

y = I = ABC

Where I means the identity and all effects and interactions may be
multiplied by it to show the results of the design. Thus we would
have the following display of confounded effects:

A = A2BC = BC

B = AB2C = AC

C = ABC2 = AB

The analysis of variance is given in Table 10.1.3.

TABLE 10.1.3

ANOVA of 1/2 Fractional Replicated 2^ Factorial

Source df

A and/or BC 1
B and/or AC 1
C and/or AB J.

Total 3
256 10. FRACTIONAL^ FOR TWO-LEVELED FACTORS

It should be pointed out that if the block with the control

(1), that is,
(1)
ab
ac
be

had been used, the relationship between A and BC would have been
negative as given in Table 10.1.4.

TABLE 10.1.4

A vs BC and the Treatment Combination

Treatment combinations
Effect (1) ab ac be

A + +
BC + +

The design for this experiment would be CR and have aliases

I = -ABC B = -AC

A = -BC C = -AB

For a study of this feature see Davies (1971, p. 512).

The sign is not important in the fractional replicated

experimentsper sebecause the sums of squares erase thesign, but
this will be of value to us in Chapter 13 indiscussing thesteepest
ascent methods used to obtain optimum levels of factors because
signs are of interest for the regression coefficients.

Example 10.1: In an experiment on asphalt concrete using

cylindrical specimens, a civil engineer wanted to determine the
effects and interactions of certain factors on a coded index of
goodness of the material. The factors and their levels are given in
the following tabulation:
10.1 THE 1/2 REPLICATION 257

Level
Factor 1 2
a. Aggregate gradation Fine Coarse
b. Compaction temperature 250°F 300°F
c. Asphalt content 5% 7%
d. Curing condition Unwrapped Wrapped
e. Curing temperature 45°F 72°F

The experimenter wanted to run only 16 treatment combinations

and was willing to assume three factor and higher interactions were
zero. Hence a one-half replicate of a 2^ factorial was constructed
and the 16 treatment combinations were run completely at random.

The identity used was

I = ABCDE
so that all main effects have aliases of four factor interactions
and all two factor interactions have three factor interactions as
aliases. Since there are zero degrees of freedom for error in this
experiment, the investigator had to rely on a previous estimate of
the error mean square. This estimate was 200.

The treatment combinations run in the experiment and their

yields are the following:
Treatment combination Yield
(1) 13
ae 54
be 44
ab 49
ce 13
ac 14
be 18
abce 85
de 41
ad 73
bd 79
abde 17
cd 82
acde 58
bede 10
abed 29
Total 679
258 10. FRACTIONALS FOR TWO-LEVELED FACTORS

From these 16 yields we can obtain the SS for all 5 main

effects and the 10 two factor interactions. For the interactions,
we can set up two-way tables.

AB:
B

149 = 13+13+41+82 44 + 18 + 7 9 + 1 0 = 151

(1) + ce + de + cd be + be + bd + bcde

199 = 54 + 14 + 73 + 58 49 + 85 + 17 + 29 = 180
+
ae + ac + ad + acde ab + abce + abde + abed

SS AB . - ■151 - ™ ) 2 , ilflji. . 27.6

lo lo ---

Notice that 149 + 180 + 151 + 199 = 679 (the total)

AC:
C +

Here, too, 177 + 186 + 123 + 193 = 679.

Problem 10.1.1. Set up the two-way tables for CD and DE and

calculate the SS in a comparable manner as shown for AB and AC.
Check your answers with the ANOVA in Table 10.1.5.

Of course one can see that the SS for the main effects are
easily found from the marginal totals of the two-way tables. For
A we obtain

SS A . (1« • >80 - M 9 - i s i ) 2 , i 9 0 1
lo
10.1 THE 1/2 REPLICATION 259

for B

s s i s - t 1 5 1 * . 18.p . - .6 14 » .- . A 9a . i , 18 ,1

and for C

SS c = (123 + 186 - 177 - 1 9 3 )2 = 232>(.

16

Problem 10.1.2. Find the SS for D and E using the technique

above. Check your answers with the ANOVA in Table 10.1.5.

TABLE 10.1.5

ANOVA for the Asphalt Experiment

Source df MS F

A 1 390.1 1.95
B 1 18.1 < 1
C 1 232.6 1.16
D 1 612.6 3.06
E 1 76.6 < 1
AB 1 27.6 < 1
AC 1 138.1 < 1
AD 1 1387.6 6.94^
AE 1 1105.6 5. 53^
BC 1 68.1 < 1
BD 1 3052.6 15.26^
BE 1 0.6 < 1
CD 1 0.1 < 1
CE 1 410.1 2.05
DE 1 3570.1 17.85^
Error 00 200

Tabled F^CO.05) = 3.84

F^CO.Ol) = 6.63

♦Significant at 0.05 level.

♦♦Significant at 0.01 level.
260 10. FRACTIONALS FOR TWO-LEVELED FACTORS

For further analyses the experimenter would probably want to

look at all two-way tables of means that are significant and
consider running Newman-Keuls test.

10.2 A MORE COMPLICATED DESIGN (1/4 REPLICATION)

As the number of factors gets larger, it is more efficient to

3
run an experiment with a smaller fraction. For example, the 2
completely randomized design should hardly ever be fractionated and
the previous discussion was to show the meaning, not a good design
g
example; however, a 2 factorial could be run as a 1/4 replicated
experiment.
9
Let us take a 2 experiment and assume we have no information
on the experimental error. Then we may try a 1/4 replication or
have 2 =128 treatment combinations. Factors a, b, c, d, e, g,
h, j , ar e available and one good completely randomized design
results if the defining relation or identity is:

I = ABCDEF = DEFGHJ = ABCGHJ where ABCDEF and DEFGHJ are

selected interactions called defining effects and their product
ABCGHJ is the generalized interaction. A few of the aliases are:

A = BCDEF = ADEFGHJ = BCGHJ

AB = CDEF = ABDEFGHJ = CGHJ
ABC = DEF = ABCDEFGHJ =GHJ

It can be seen that all main effects have five factor

interactions or higher as aliases, all two factor interactions have
four factor or higher interactions as aliases, and no information
can be obtained from three factor interactions because they are
aliases of three factor interactions. It would be best if we could
assume three factor and higher interactions are zero so the three
factor interaction source could be used as error.

The analysis of variance for this experiment is given in

Table 10.2.1 assuming three factor interactions and higher orders
10.2 THE 1/4 REPLICATION 261

are zero. The details on obtaining the treatment combinations to

be used will not be described here.

TABLE 10.2.1

9
ANOVA for 1/4 Fractional Replicated 2 Factorial

Source df

Main effects 9
Two factor interactions 36
(Remainder) error 82

Total 127

Another method of analysis is the half normal plots given by

Johnson and Leone (1964, Vol. II, pp. 188-191) and Daniel (1959).
The use of half normal plots provides a graphical method of
displaying the average effect for each degree of freedom and a
method of testing effects for significance. This method of
significance testing essentially produces the same results as
obtained with the classical ANOVA table approach. For this reason
we do not reproduce the half normal plot procedure in this text.
One advantage of displaying all effects on a half normal plot is
that it gives the experimenter some feeling for the validity of the
assumption of normality in each cell. This follows since all of
the nonsignificant effects should be normally distributed and
display a straight line on the half normal plot.

Problem 10.2.1. Using results of Problem 9.1.3 in Chapter 9,

run a half normal plot analysis (see Johnson and Leone, 1964,
pp. 188-191).

Example 10.2: Utilizing a 1/4 replicate of a 2^ to demonstrate

the Yates (1937) method for obtaining total effects and sums of
squares for fractional factorial experiments, we set up the
following identity:

I = ABC = CDE = ABDE

262 10. FRACTIONALS FOR TWO-LEVELED FACTORS

Using the intrablock subgroup

Treatment combination Yield

(1) 6
ab 10
de 4
abde 5
bed 9
acd 3
bee 8
ace 0

we leave it as a problem for the reader to show that

I = - ABC = -CDE = ABDE

A rule that may be used for determining the signs of the effects in
the identity relation is that a minus sign is placed before all
effects which have an odd number of letters provided the intrablock
subgroup is used for the experiment. If we are interested in all
main effects we can obtain the following aliases for the 7 degrees
of freedom from the experiment:

A = -BC = -ACDE = BDE

B = -AC = -BCDE = ADE
C = -AB = -DE = ABCDE
D = -ABCD = -CE = ABE
E = -ABCE = -CD = ABD
AD = -BCD = -ACE = BE
AE = -BCE = -ACD = BD

The procedure for calculating the sum of squares for each of

the above aliases is not unique. One method would be that given in
Example 10.1 where one would construct a table of values for each
of the seven aliases given above. A second method would be write
down all 32 treatment combinations of the 2^ design in the usual
order and use zeroes for yields of those treatment combinations not
10.2 THE 1/4 REPLICATION 263

in the experiment together with the actual yields for those

treatment combinations utilized in the experiment. Yates method
can now be used to determine the effects due each source of
variation. Note that there will be a sources of variation for each
main effect and interaction and it will turn out, for example, that

[A] = -[BC] = -[ACDE] = [BDE]

All effects will turn out as shown in the above aliases.

The second method described above becomes rather tedious and

too large for many computer programs as n becomes large and the
fraction (1/2^) gets smaller. Berger (1972) gives another method
of analysis which allows one to use Yates procedure as if there were
only n-p factors in the experiment. The data of this example will
be analyzed using this procedure. In general, one must find a
nonunique set of n-p letters, e.g., a, c, e, ... which will be used
to generate the entire set of 2n”^ treatment combinations in an
order such that Yates algorithm can be applied to them in the
normal fashion and the generated effects are properly identified.
The remaining p letters are referred to as dead letters.

The defining effects are used to determine the dead letters.

For this example the defining effects are ABC and CDE. First select
a dead letter from the first defining effect which is preferrably
not in any other defining effect, CDE in this case. Suppose B is
selected, since I = ABC, set B = AC, so that if B should appear in
some future defining effect AC can be substituted for it. The
second dead letter could be either D or E as neither appear in ABC.
Another way of selecting D or E is that C is related to the dead
letter B by the identity relation B = AC, so we do not select C.
If E, say, is selected as the second dead letter then we see that
E = CD. Thus the letters A, C, and D are admissible and B and E
are dead.

The admissible letters are used to develop the following table

which is used to implement Yates method using the above data. The
first column is generated in the usual fashion using the admissible
 264
Admissible set Dead Augmented Total Aliased
in Yates order letters letters Yield 1 2 effect effects

(1) C1) 6 16 24 45

a b ab 10 8 21 -9 A = BC = BDE*

c be cbe 8 9 -4 -5 C = AB = DE

ac e ace 0 12 -5 -19 AC = B = ADE

d e de 4 4 -8 -3 D = CE = ABE
10.

ad be abde 5 -8 3 -1 AD = BCD = ACE = BE

cd bee = b bed 9 1 -12 11 CD = ABD = E

acd ee = 1 acd 3 -6 -7 5 ACD = BD = AE = BCE

*
Fourth and higher order interactions are omitted.
FRACTIONALS FOR TWO-LEVELED
FACTORS
10.2 THE 1/4 REPLICATION 265

letters. The second column is a group of dead letters developed by

making entries opposite the terms which have a single admissible
letter. These entries are made by using the dead letter identity
relations B = AC and E = CD. Since A appears with the dead letter
B but no other dead letters, b is placed in the dead letter column
opposite the admissible letter a. The same type of reasoning places
b and e opposite the admissible letter c. The dead letter entries
b and be are multiplied together to obtain the entry e opposite the
admissible value ac. Since D appears only with E in the identity
relations an e is placed opposite d and the remaining dead letters
are formed by multiplication.

The dead letters are augmented to the admissible letters to

form the third column. The third column must be the complete list
of treatment combinations, hence the associated yield is placed in
5-2
the next column. Yates method is now carried out on the 2
observations in the usual manner. The aliased effect column is
obtained by multiplying the letters in the first column, the
admissible set, times the identity relation

I = ABC = CDE = ABDE

Notice from the signed identity relation

I = -ABC = -CDE = ABDE

we obtain

[A] = -9
[BC] = 9
[ACDE] = 9
[BDE] = -9

Next we need to show the sums of squares for the main effects
and two factor interactions, AD and AE. These sums of squares can
now be obtained as follows:
266 10. FRACTIONALS FOR TWO-LEVELED FACTORS

SS - (Total effect)2
?L 2
c-
i=l 1

In this case, n = 8 and c. = ± 1. Hence the denominator is 8 for

5 l
all cases.

Sums of squares:

A - — - 10.1 AD = - 0.1

45.1 E.iiiii.is.i

C.i^.5,1 AE - - 3.1

In general, if one has a digital computer and a good factorial

program available he can use zeroes for these treatment combinations
not in the experiment and multiply the sums of squares from the
computer printout by the denominator of the fraction involved to
get the correct sums of squares. In this case the computer would
divide by 32 to get the SS. Hence, all seven sums of squares on
the computer output need to be multiplied by 4 in order to obtain
the correct sums of squares.

Problem 10.2.2. Determine the intrablock subgroup for a 1/8

replication of 9 factors using the identity relationship
I = ABEGHJ = ACFGJ = BCEFH = ABCD = CDEGHJ = BDFGJ = ADEFH.
Use the Berger (1972) procedure to find the order of the 64
treatment combinations that allows the Yates procedure to be
used to identify the corresponding main effects and
interactions.
.2 THE 1/4 REPLICATION 267

An approach for solving Problem 10.2.2:

(1) Select 3 dead letters out of the three interactions:

ABEGHJ; ACFGJ; ABCD (defining effects)

as follows:
a) Select a letter in the first defining effect that is
not in the other two, say E. Since I = ABEGHJ
E = ABGHJ.
b) Similarly for second defining effect, F = ACGJ
c) Similarly for the third, D = ABC
(2) Hence D, E and F are the dead letters and associate with
the treatment combination letters. Admissible letters are
associated with the dead letters to obtain augmented column
as follows:
Admissible Dead Augmented
a def adef
b de bde
c df cdf
g ef efg
h e eh
j ef efj

(3) Arrange the admissible set of 64 treatment combinations

in Yates order and generate the corresponding augmented
treatment combinations.

4
Problem 10.2.3. Given a 1/2 replicate of a 2 factorial,
(a) Discuss the quality of the best design.
(b) Set up the identity and aliases.
(c) Show your intrablock subgroup.
(d) If the yields for the corresponding treatment
combinations are (1) =6; ab = 10; ac = 4, be = 5, ad = 9,
bd = 3; cd = 8 and abed = 0, use the Yates method and the
Berger procedure to find the best ANOVA assuming the error
mean square is known to be 5.0.
268 10. FRACTIONALS FOR TWO-LEVELED FACTORS

10.3 BLOCKING WITHIN THE FRACTIONAL REPLICATION

In many fractional replicated experiments there are too many

treatment combinations to be handled under homogeneous conditions
so that blocking is required, even within the fraction.
g
Consider a 1/4 replicate of a 2 or 64 treatment combinations
in four blocks of 16. One design is established from the identity
relationship.

I = ABCDE = ABFGH = CDEFGH

From the identity, I, it is seen that all aliasesof main

effects and two factor interactions are three factor or higher
interactions. Then, finding that all aliases of ACF and BDG are
four factors and higher and their generalized interaction, ABCDFG,
has three factor or higher interactions confounded with it, we can
use these as our blocking basis. These aliases are:

ACF = BDEF = BCGH = ADEGH

BDG = ACEG = ADFH = BCEFH
ABCDFG = EFG = CDH = ABEH

To write out the blocks we may use the four defining equations
(see Kempthorne, 1952, Chapter 20).

ABCDE : X. + X0+ X7 + X. + Xc =0 mod 2

-
From T
I: 1 2 3 4 5
ABFGH : X- + X0 + X. + X- + XQ = 0 mod 2
1 Z 0 / 0

ACF : X, + X7 + X. =0, 1, mod 2

„ l o o
From
Block’s: BDG : X2 + X4 + X? =0, 1, mod 2

and using levels such that the equations are all set equal to zero
gives the intrablock subgroup. Since we will deal only with the
intrablock subgroup of the 1/4 replicate, the levels of ABCDE and
ABFGH will always be zero, but the levels of the ACF and BDG will
be 0 and 1 to make up the four blocks of 16 in the 1/4 replicate.
In other words, the four blocks may be represented as:
10.3 BLOCKING IN FRACTIONAL 269

ABCDE : 0 ACF :: 0
ABFGH : 0 BDG : 0

ABCDE : 0 ACF : 0
ABFGH : 0 BDG : 1

ABCDE :: 0 ACF :: 1
ABFGH :: 0 BDG : 0

ABCDE : 0 ACF :: 1
ABFGH : 0 BDG : 1

The analysis of variance for this entire experiment of 64

observations is given in Table 10.3.1.

TABLE 10.3.1
o
ANOVA for 1/4 Fractional Replicated 2 in Blocks of 16

Source df

Blocks and/or (ACF, BDG, EFG, CDH,

BDEF, BCGH, ADEGH, ACEG, ADFH,
BCEFH, ABCDFG, ABEH) 3
Main effects 8
Two factor interactions 28
Error (assuming all 3 factors and
higher interactions are zero) 24

Total 63

Designs of experiments, including blocking in the incomplete blocks

of fractional replications for 2n factorials are given in the
booklet: fractional Factorial Experiment Designs for Factors at
Two Levels,11 issued by National Bureau of Standards Applied
Mathematics Series 48 (1957).
270 10. FRACTIONALS FOR TWO-LEVELED FACTORS

Problem 10.3.1. A company was making hat shells out of foam

plastic by molding it and spraying the molded hat shells with
latex. There were six spray guns set so that they would hit
different regions of the hat.
The problem facing the experimenter was to adjust the
latex flow rate and the air pressure on each gun in order to
produce a uniformly coated hat. A lower and upper level for
each latex flow rate and a lower and upper level for each air
pressure was selected on each gun. He also had two different
foams, one sparkly and the other dull. Hence there were 13
factors each at 2 levels for the experiment.
He could run only 35 treatment combinations per day
(there were day to day differences) and wanted to know the
minimum number of days he would need to run the experiment to
get information on main effects and two factor interactions
assuming that three factor and higher interactions were zero.
(a) Design a good experiment, that is, show the identity
and its aliases for his conditions (do not write out the
treatment combinations).
(b) Show your analysis of variance and your assumptions.

10.4 SOME OTHER VERSIONS AND DEVELOPMENTS IN FRACTIONAL FACTORIALS

10.4.1 Parallel Fractional Replicates

If many responses or yields (y^, i = 1, ..., k) are measured

on the same experimental unit and one factor is known not to have
an effect on one of the yields, one can utilize this information in
knowing that certain main effects and interactions are not present
for that yield. Then if another factor is known not to have an
effect on another yield, the same experimental units may be analyzed
another way assuming other main effects and interactions are zero
for this new variable.

Daniel (1960) uses an example on flavor (y^) and texture (y^)

in a new food experiment in which factor a is oil, b is wheat
product and c is branlike material. Factor a does not influence
y2> b may influence both y^ and y^ and c does not influence y^.
3
Given that a 1/2 replicate of the 2 was run, the aliases are:
10.4 OTHER FRACTIONALS 271

I = ABC
A = BC
B = AC
C = AB

The analysis of variance for y^ (knowing factor c. has no influence)

is given in Table 10.4.1.

TABLE 10.4.1

ANOVA of y

Source df

A 1
B 1
AB 1

The analysis of variance for y2 (knowing a. has no influence) is

given in Table 10.4.2.

TABLE 10.4.2

ANOVA of y2

Source df

B 1
C 1
BC 1

These analyses are run from the same experiment.

More general type designs are given in the paper by Daniel

(1960).
272 10. FRACTIONALS FOR TWO-LEVELED FACTORS

10.4.2 The 2n~P Fractional Designs

Box and Hunter (1961, pp. 317-319) introduce the fractional

factorial notions and application areas of the 2n”^, where
n = number of factors and 1/2^ = the fraction. These types of
designs, classified according to the degree of fractionation, are
given the following names:

(a) Resolution III: No main effect is confounded with another

main effect but main effects are confounded with two factor
interactions, and two factor interactions are confounded with one
3-1
another. An example is 2

(b) Resolution IV: No main effect is confounded with any

other main effect or two factor interaction, but two factor
4-1
interactions are confounded with one another. An example is 2

(c) Resolution V: No main effect or two factor interaction

is confounded with any other main effect or two factor interaction,
but two factor interactions are confounded with three factor
interactions. An example is 2^~*.

Resolutions III and IV designs are discussed in great detail

in this paper. The list of references at the end of the paper is
most appropriate here.

Box and Hunter (1961, p. 449) discuss the Resolution V designs

in detail.

10.4.3 Symmetrical and Asymmetrical Fractional Factorial Plans

Addelman (1962, p. 47) gives the following synopsis:

"Procedures for constructing plans which permit uncorrelated
estimation of all main effects and some specified two factor
interaction effects are developed for symmetrical factorial
arrangements. These plans can then be adjusted to yield plans with
similar properties for asymmetrical factorial experiments.”

These plans are for three-level factorials, as well as two-level,

10.4 OTHER FRACTIONALS 273

and follow up on the idea of the Resolution V type designs. The

references at the end of Addelman's paper are most relevant
material.

10.4.4 The 3/4 Fractional Factorial

4
Consider a 2 factorial. If a 1/2 replicate were run on the
4
2 , ordinarily one would use the identity relationship

I = ABCD

This would cause complete confounding of all two factor interactions

with other two factor interactions. One method of eliminating this
problem of confounding two factor interactions is with a 3/4
fractional factorial which is described below.

Using four blocks of four treatment combinations each and

confounding BD, ABC and ACD with blocks we would have the following
blocks:

I II III IV
(1) a b d
ac c abc acd
bed abed cd be
abd bd ad ab

A three quarters (3/4) replicate may be obtained by combining three

half-replicates: blocks II and III; blocks II and IV; and blocks
III and IV. Notice that block I is not used in the experiment.
Hence the name 3/4 replicate.

The half-replicate made up of blocks

(a) II and III, have the identity relationship of I = ABC
(b) II and IV, have the identity relationship of I = ACD
(c) III and IV, have the identity relationship of I = BD

If three and four factor interactions are zero, main effects

and two factor interactions can be estimated from the 3/4
274 10. FRACTIONALS FOR TWO-LEVELED FACTORS

replicated experiment. The aliases and blocks to be used for

estimating them are given in Table 10.4.3. It must be understood
in the actual experiment that all 12 treatment combinations given
in blocks II, III, and IV are run completely randomized in one
block of the 12 treatment combinations. Since there are no degrees
of freedom remaining for estimating error it must be estimated from
an outside source. The reference for this design is John (1961),
which also provides a 3/4 replicate for a 2^ factorial to obtain 8
degrees of freedom for error assuming three factor and higher
interactions are zero.

TABLE 10.4.3

Aliases for the 3/4 Replicate

and Blocks Estimating the Aliases

A = ABD in I I and IV
B = ABCD in I and IV
C = BCD in I I and IV
D = ABCD in I and III
AB = BCD in I and IV
AC = ABCD in I I and IV
AD = BCD in I and III
BC = ABD in I and IV
BD = ACD in I and III Use the mean of these
or ABC in I and IV two estimates for BD
CD = ABD in I and III

John (1966) generalizes the results on 3/4 replication by

augmenting the usual fractions to improve some estimates. John
(1971, Chapter 7), is an excellent overall reference for these
designs.

Addelman (1961) showed plans for 3/2n replicates (irregular

fraction plans) some of which permit the estimation of main effects
and two factor interactions with fewer trials then is required with
an orthogonal plan. Although there is some correlation between some
10.4 OTHER FRACTIONALS 275

of the estimates there is a procedure given to estimate the main

effects and certain interactions. With present-day regression
computer programs these designs should be considered for the special
type problems.

10.4.5 Case When Some Three Factor Interactions Cannot Be Assumed

Negligible
g
If in a 1/4 replicate of a 2 the identity relationship:

I = ABCDE = DEFGH = ABCFGH

is used, the following three factor interactions would be confounded

with four factor interactions or higher:

ADF BEF
ADG BEG
ADH BEH
AEF CDF
AEG CDG
AEH CDH
BDF CEF
BDG CEG
BDH CEH

If the experimenter knew before the experiment was conducted that

he could not assume some three factor interactions were negligible,
he should force the factors to be labeled so that these questionable
three factor interactions would be among the 18 above. If he did
not force the factors to be handled in this manner there may be some
three factor interaction effects mixed up with the two factor
interactions and wrong conclusions regarding the importance of these
two factor interactions could be drawn. A reference for this
situation occurred in an experiment described by Hadley et al (1969).
Figure 10.4.1 pictorially displays an example of an interpretable
three factor interaction from Hadley et al. (1969).
 Indirect Tensile Stress (psi)
Fig. 10.4.1.
Three factor interaction.

SHOJDVd aaiHAai-OMJL HOd STVNOUDVHd *01 9LZ

10.5 GENERAL 2n PROBLEMS 277

In general, statisticians and experimenters should be more

aware of the possibilities that three factor interactions may be
important to research workers than is usually assumed by almost all
statisticians working in the design of experiments area. If there
is doubt that certain three factor interactions are zero, this
information should be utilized in the design of the experiment by
confounding all main effects, all two factor interactions and three
factor interactions with four factor interactions and higher. It
may turn out that the interpretation of the three factor interactions
is the most essential result. In some cases experimenters have
even interpreted four factor interactions by changing the model and
using the treatment combinations of all four factors simultaneously
as the treatments, thereby ignoring the original main effects,
two factor interactions and three factor interactions. A good
reference for this concept is Marascuilo, L. A. and Levin, J. R.
(1970).

10.5 GENERAL 2n PROBLEMS

10.5.1. In a 2^ factorial experiment an investigator could

handle only 17 treatment combinations under one environmental
condition, and he had enough money to run 34 treatment
combinations for the whole experiment. What is a good design
to use? Show your analysis of variance (including source and
degrees of freedom) and state your assumptions on interactions
being negligible.

10.5.2. In an experiment investigating wear and life of gears,

there were 15 factors each at two levels. Let us assume four
factor and higher interactions are negligible. Design (without
blocking within the fraction) a good experiment for this
condition. Show the analysis of your experiment.

10.5.3. In a 2** factorial, what is the minimum number of

observations required to get information on main effects and
two factor interactions if we assume three factor interactions
and higher are zero and we use the maximum number of blocks.
Show the identity.

10.5.4. If, in a 29 factorial experiment a research worker did

not want to assume any three factor interactions were negligible,
278 10. FRACTIONALS FOR TWO-LEVELED FACTORS

design a fractional factorial experiment with the smallest

fraction and show your analysis if:
(a) three factor interactions are of no interest
(b) three factor interactions are of interest
(c) comment about the two designs and their analyses

10.5.5. Find a reasonably good design for a 3/4 replicate of

a 2^ factorial with no blocking in the fraction. State your
assumptions on three factor interactions. Show your analysis
of variance. Compare your results with a 1/2 replicate of a 2^
factorial with no blocking in the fraction.

10.6 REFERENCES

Addelman, S. Technometrics 3:479 (1961).

Addelman, S. Technometrics 4:47 (1962).
Berger, P. D. Technometrics 14:971 (1972).
Box, G. E. P. and Hunter, J. S. Technometrics3:311 (August1961);
ibid.,•3:449 (November 1961).
Daniel, C. Technometrics 1:311 (1959).
Daniel, C. Technometrics 2:263 (1960).
Davies, 0. L. (Ed.) Design and Analysis of Industrial Experiments,
2nd ed., Oliver and Boyd, Edinburgh, 1971.
Hadley, W. 0., Hudson, W. R., Kennedy, T. W.,andAnderson, V. L.
Association of Asphalt Paving Technology 1:1 (1969).
Hicks, C. R. Fundamental Concepts of Design of Experiments, 2nd ed.,
Holt, Rinehart and Winston, New York, 1973.
John, P. W.M. Biometrics 17:319 (1961).
John, P. W. M. Technometrics 8:469 (1966).
John, P. W.M. Statistical Design and Analysisof Experiments,
Macmillan, New York, 1971.
Johnson, N. L. and Leone, F. C. Statistics and Experimental Design,
Vol. II, Wiley, New York, 1964.
Kempthorne, 0. Design and Analysis of Experiments, Wiley, New York,
1952. Distributed by Krieger Pub. Co.,Huntington, New York.
Marascuilo, L. A. and Levin, J. R., Am.Ed. Res. J.7:397 (1970).
Peng, K. C. The Design and Analysis of Scientific Experiments
Addison-Wesley, Reading, Massachusetts, 1967.
Statistical Engineering Laboratory, ’’Fractional Factorial Experiment
Designs for Factors at Two Levels”, U. S. Dept, of Com. Nat’l. Bur.
of Standards, Applied Math. Series 48, (1957).
10.6 REFERENCES 279

Yates, F. Design and Analysis of Factorial Experiments, Imperial

Bureau of Soil Sciences, Harpenden, England, 1937.
 Chapter 11

THREE-LEVEL FACTORIAL EXPERIMENTS

One must recognize that the 2n factorial experiment is a

special case of the general equal prime-number leveled factorial
experiments. In the quantitative two-leveled factorial no mention
is made of investigating for trends because one cannot examine the
concept of deviations from linear. In even the equally spaced
2
quantitative 3 factorial (which is the simplest case with more
than two levels for each factor in the quantitative equal prime
number leveled factorial), one can examine the results for
departures from the linear concept. The reader may want to refer
to Chapters 1 and 2 to reacquaint himself with the usefulness of
quantitative levels and model development. In general, it will be
seen that this chapter on three-leveled factorials will form the
basis for easily extended procedures not only on equally spaced
quantitative equal prime-number leveled factorials but also on the
qualitative equal prime-number leveled factorials. Also, this
chapter deals only with CRD in each block discussed.

When only two levels occur in each factor there is only 1

degree of freedom for each main effect and interaction. For any
other case the interaction degrees of freedom are larger than those
2
of the main effects. For example, in the simplest case of a 3 ,
the number of degrees of freedom for the two factor interaction is
four, whereas the degrees of freedom for each main effect is only 2.

Since the interactions have the same number of degrees of

freedom as the main effects in the two-leveled factorials, complete
flexibility in using any interaction or main effect in confounding

280
11.1 CONFOUNDING IN 3n 281

systems is possible there. In all equal prime-number leveled

factorials this type of flexibility is desired. The procedure,
then, for the factorials with the number of levels greater than two
is to split the interactions into orthogonal pieces with the same
number of degrees of freedom as the main effects. In this case two
of the degrees of freedom for the interaction may be used to
estimate the interaction and the other 2 degrees of freedom can then
be used for confounding purposes.

As indicated in the introduction of Chapter 10 the present

method of experimentation is to include many factors with only two
levels each early in the investigation. Later, however, many levels
of few factors may be used to obtain a thorough description of the
response surface associated with the few factors. Since three
levels really is in between these two extremes, designs of this
nature do not seem to be used as often. There are occasions when
three-leveled designs are useful and a general analysis for them is
given by Margolin (1967), Kempthorne (1952), Davies (1971), and
Winer (1971).

In this chapter a short description of how these designs may

be set up, especially for incomplete blocks and fractional designs,
is given. A thorough description is given by Kempthorne (1952) and
Davies (1971).

11.1 CONFOUNDING IN A 3n SYSTEM

The system of confounding that was used in the 2n factorials

given in Chapter 9 will be continued here. To get the pieces of
interactions to have the same number of degrees of freedom as the
main effects we will use the system given by Kempthorne (1952,
Chapter 6) and Hicks (1973). A breakdown of the interactions that
follows the 2n approach given in Chapter 9 is adaptable to incomplete
blocks and fractional factorial experiments for three-leveled
factorial experiments.
282 11. THREE-LEVEL FACTORIALS

11.1.1 The 3 Factorial

Following the method given by Kempthome (1952, p. 293) and

2
Hicks (1973, Chapter 9), the 3 factorial may be examined as follows

Main effects: 2 degrees of freedom each on margins

b
0 1 2

II
0 00 01 02-- 0, mod 3

H-X
1
a l 10 11 12-- mod 3
— Xl = 1,
2 20 21 22-- mod 3
1 | 1 — Xl = 2,
I
X2 = fo x2 = 1
X2 = 2
mod 3 mod 3 mod 3

Interactions: (a) 2 degrees of freedom for right diagonals or AB

2, mod 3

(b) 2 degrees of freedom for left diagonals or AB

11.1 CONFOUNDING IN 3n 283

The analysis of variance is given in Table 11.1.1.

TABLE 11.1.1

ANOVA for 3^ Factorial

Source df

A 2
B 2 This breakdown of the
AB 4 I 2 Hinteraction has no )
*AB (2 df)* *experimental meaning.*

Interactions can be split up in the above manner so that the

parts have the same number of degrees of freedom as the main effects
and all four pieces (two main effects and the two pieces of
interactions) are orthogonal to each other. There is no known
experimental value in splitting up AB with 4 df into AB (2 df) and
2 2
AB (2 df). That is, there is no value in comparing AB and AB .
The only purpose for this scheme is for use in confounding. Each
2
part of the interaction, AB or AB , is equally good for an estimate
of the interaction between A and B. There is twice as much
information on the AB total interaction from both parts as from
either part.

The sums of squares may be described as:

284 •11. THREE-LEVEL FACTORIALS

where
th
A^ = sumof observations in i level of a (i = 0, 1, 2)
th
Bj = sumof observations in j level of b (j= 0, 1, 2)

G = sumof all observations

ABg = total of observations where + X2 = ^ moc* ^

00 + 12 + 21)

AB^ = total of observations where X^ + X2 = m°d 3 (i«e»>

10 + 01 + 22)

AB2 = total of observations where X^ + X2 = 2, mod 3 (i.e.,

20 + 11 + 02)
2
ABq = total of observations where X^ + 2X2 = 0, mod 3 (i.e.,
00 + 11 + 22)
2
AB^ = total of obse:
observations where X^ + 2X2 = 1, mod 3 (i.e.,
02 + 10 + 21)
2
AB2 = total of observations where X^ + 2X2 = 2, mod 3 (i.e.,
12 + 01 + 20)

Margolin (1967) shows an algorithm for computing effects for

three leveled factorials if the levels are equally spaced
quantitative. A general discussion using an algorithm for 2n3m
factorial experiments is also given by Margolin (1967). Of course
when n = 0 the 3m algorithm results and when m = 0 the 2n algorithm
(usually called the Yates method) results.

11.1.2 Other Three-Leveled Factorial Experiments

4
In a 3 factorial experiment:

2 2
(AB C D)2 = total of the observations from the defining
equation X^ + 2X2 + 2X^ + X^ = 2, mod 3
2
(AC )1 = total of the observations from the defining
equation X^ + 2X^ = 1, mod 3
11.1 CONFOUNDING IN 3n 285

4
Problem 11.1.1. For a 3 factorial experiment, show the sum
2 2
of squares for AB CD . (Write out the equation, establish the
correct treatment combinations for the various totals and
indicate the way to obtain the sum of squares.)

For all designs with prime and equal number of levels, the
first letter of any part of an interaction may have an exponent of
unity. This condition is easily seen for the three-leveled
2 2
factorials. For example: AB = A B since the treatment combinations
for each interaction are combined in the same manner as shown below

(AB2)0 = 00 + 11 + 22 (a2b)0 = 00 + 11 + 22
(AB2)1 = 02 + 1 0 + 2 1 (A2B)2 = 0 2 + 1 0 + 21

(AB2)2 = 0 1 + 1 2 + 2 0 (A2B)1 = 01 + 12 + 20

In order to obtain an exponent of unity for the first letter

of an interaction part in a 3n factorial, square the interaction
part. For examp1e:

(a) (A2B)2 = A4B2 [A4, mod 3 = A]. Hence A2B = AB2

2 2 3 2 2 2 422 22
(b) ABC x AC D = A BC D = A BD = (A BD) = A B D = AB D
since (C^, mod 3) = = 1

11.1.3 Method of Confounding

Consider what can be confounded in a 3n factorial experiment.

Of the (3n - 1) total degrees of freedom there are 2 degrees of
freedom for each main effect and each piece of interaction, therefore
the number of possible pieces (including the main effects) to
confound are (3n - l)/2. In a 3^ factorial experiment Table 11.1.2
shows there are 13 possible pieces to confound.
286 11. THREE-LEVEL FACTORIALS

TABLE 11.1.2
3
ANOVA of a 3 Factorial Experiment
Pieces that can
Source df
be confounded
A A 2
B B 2
AB 2
AB
AB 2
C C 2
AC 2
AC
AC 2
BC 2
BC
BC 2
ABC 2
ABC2 2
ABC 2
AB C 2
a b 2c 2 2

In general there has to be a one-to-one correspondence between

the degrees of freedom for blocks and the degrees of freedom of
effects and/or interactions confounded. One way of summarizing
this is the following:

Number of df Main effect and/or part of the

confounded with blocks_____ interactions confounded (2 df each)
2 X
8 X Y XY XY2
Once X and Y are confounded XY and
2
XY will also be confounded
26 X Y XY XY2 Z XZ
XZ2 YZ YZ2 XYZ XY2Z
? 2 2
XYZ XY Z
Once X, Y, and Z are confounded the
other effects are also confounded

and so on
11.1 CONFOUNDING IN 3n 287

To show the method of handling the generalized interaction concept

for three levels consider the example:

X Y XY XY

(a) AB BC a b 2c AB3C2 = AC2

(b) AB AC a 2bc « a b 2c 2 BC2

One explanation of this is:

3 2 0 2 2 2
(a) Since modulo 3 is used here, AB C = AB C = A1C = AC .

(b) Since the first letter in an interaction part may be with

exponent 1, (A2BC) = (A2BC)2 = A4B2C2 = AB2C2. The
squaring merely provides the exponent (mod 3) to be 1.
It is interesting that as this method of dividing
interactions into orthogonal parts is used on larger prime-
numbered leveled factorials, the same procedure of forcing
the first letter to have exponent 1 is used but, of
course, the squaring goes to cubing and so on.
3
Let us consider a 3 factorial experiment or 27 observations
in nine blocks of three. There must be eight degrees of freedom
confounded with blocks. Show a design that provides full information
on main effects and as much information on two factor interactions
as possible. Try:

X Y XY XY
2 2
(a) ABC a b 2c A C = AC B
2 2 5 2
(b) ABC AB2 A C = AC B^C = BC

Results:

(a) This design has a main effect, B, confounded with blocks,

so, in most cases, it is not satisfactory.

(b) Only 1/4 of the three factor interaction is confounded,

only 1/2 of each of AB, AC and BC is confounded and the
three main effects are completely free of confounding.
This is usually considered a good design.
288 •11. THREE-LEVEL FACTORIALS

For the layout of this design we may use the equations

ABC : Xx + X2 + X3 = 0, 1, 2 mod 3

AB : Xx + 2X2 =0, 1, 2 mod 3

The blocks may be formed as follows:

levels of ABC
levels of AB2
Blocks

0 0 0 0 0 1 0 0 2 0 2 1 2 0 0
1 1 1 1 1 2 1 1 0 10 2 0 1 1
2 2 2 2 2 0 2 2 1 2 10 1 2 2

The corresponding analysis of variance is given in Table 11.1.3.

TABLE 11.1.3
3
ANOVA for 3 Factorial in Blocks of 3

Source df

Blocks and/or ABC, AB2, AC2, BC2 8

Restriction e r t o r 0

A 2
B 2
AB 2
C 2
AC 2
BC 2
Error and/or AB2C, ABC2, AB2C2 _6
Total 26

In general in designing incomplete block experiments, one should

confound the interaction of least interest, and use the interactions
which are assumed zero for error.
11.2 SYSTEM OF CONFOUNDING 289

11.2 SYSTEM OF CONFOUNDING

2
11.2.1 A3 Factorial Experiment in Blocks of 3

(a) An example of complete confounding:

One may confound AB (2 df) with three blocks. The
equation = 0, 1, 2; mod 3 is appropriate to indicate the
confounding procedure:

Block 1 Block 2 Block 3

x2 =
X
+
ii
1 + 1 + <N
X2 =

0 0 1 0 2 0
1 2 2 2 0 2
2 1 0 1 1 1

The corresponding analysis of variance is given in Table 11.2.1.

TABLE 11.2.1

ANOVA for 32 in Blocks of 3a

Source df

Blocks and/or AB 2
Restriction error 0

A 2
B 2
AB (actually AB2) 4-2 = 2

a 2
AB (actually AB ) is the estimate of the whole interaction
2
AB coming from the part, AB .

(b) An example of partial confounding and balanced incomplete

blocks design (BIBD).

One may consider the replications of the experiment above.

Confound A" in replication I, B in replication II, AB in replication
2
III, and AB in replication IV. The corresponding analysis of
290 11. THREE-LEVEL FACTORIALS

variance for the data from this design is given in Table 11.2.2.
Refer to Example 9.3 for the method of calculation of SS.

TABLE 11.2.2

ANOVA of 4 Replications of 3^ in Blocks of 3

Source df Information
Replications (R) 3
Blocks in R 8
Restriction error 0

A 2 3/4 Balanced
B 2 3/4 information
. AB 3/4
AB 4 9
3/4 on all effects
AB 3/4

Error 16

Problem 11.2.1. Given the following data:

Replications
I II

B1 1 00 = 2 01 = 4 02 = 3 B1 4 00 = 3 10 = 5 20 = 6

B1 2 10 = 3 11 = 6 12 = 5 B1 5 01 = 5 11 = 7 21 = 8

B1 3 20 = 7 21 = 6 22 = 5 B1 6 02 = 2 12 = 7 22 = 3

III IV

B1 7 00 = 1 12 = 5 21 = 9 B1 10 00 = 4 11 = 6 22 = 1

B1 8 02 = 3 11 = 6 20 = 6 B1 11 01 = 6 12 = 5 20 = 4

B1 9 01 = 4 10 = 7 22 = 4 B1 12 02 = 7 10 = 8 21 = 7

(a) Tell what is confounded in replications I, II, III

and IV. What kind of design is this?
(b) Run ANOVA on these data. Show the Source, df, SS,
MS, and test the effects of A, B, and AB.
11.2 SYSTEM OF CONFOUNDING 291

(c) Show how you calculated the SS for all sources and
give the amount of information on each of A, B, and AB.

Problem 11.2.2. Delete replication IV in Problem 11.2.1 and

reanalyze the data. Compare the results of Problem 11.2.1
with the results in this problem.

3
11.2.2 Two 3 Factorial Experiments

(a) If one considers a design in which there are three blocks

of nine treatment combinations in each, there will be two degrees
2 2
of freedom confounded with blocks and either ABC, AB C, ABC , or
2 2
AB C may be confounded with blocks.

(b) 9 blocks of 3; 8 df confounded with blocks.

If there were nine blocks of three treatment combinations in

each block, then one possible design is:

_X _Y XY XY2

AB BC AB2C AC2

4
11.2.3 Two 3 Factorial Experiments

(a) If there were three blocks of 27 treatment combinations

in each block then there would be two degrees of freedom confounded
2 2 2 2
with blocks and either ABCD, AB CD, ..., or AB C D may be confounded
with blocks.

(b) If there were nine blocks of nine treatment combinations

in each block then one possible design is:

X Y XY XY2
2 2 2 2 2
ABC AC D AB D BCV

and the corresponding analysis of variance is given in Table 11.2.3.

292 •11. THREE-LEVEL FACTORIALS

TABLE 11.2.3

ANOVA for a 3^ Factorial When

2
ABC and AC D Are Confounded

Source df
2
Blocks and/or ABC, AC D,
2 2 2 2
AB D , and BC D 8
Restriction error

Main effects 8
2-factor interactions 24
3-factor interactions 24
4-factor interactions 16

Problem 11.2.3. In an experiment conducted by a social

scientist, the variable to analyze is the score on a
citizenship attitude scale. The factors in the experiment
are:
(a) schools in the university called: 0, 1, 2
(b) home environments are: low = 0, medium = 1,
high = 2
(c) scholastic indices are: low = 0, medium = 1,
high = 2
Tests could be given by only one instructor in one room. The
room held 45 students and it was desired that there be five
students for each treatment combination. The design allowed
confounding in blocks of nine and a record was kept on the
total of five students for each treatment combination only.
The following scores are totals of five scores each and the
total will be used as the unit in this analysis. The blocks
are different days in which different students took the tests.
Block I Block II Block III
a b£ Score a b c Score £ b£ Score
0 0 0 174 0 0 2 193 0 0 1 186
1 0 1 183 1 0 0 207 1 0 2 175
2 0 2 185 2 0 1 205 2 0 0 191
0 1 2 163 0 1 1 164 0 1 0 184
1 1 0 168 1 1 2 176 1 1 1 193
2 1 1 213 2 1 0 216 2 1 2 215
0 2 1 175 0 2 0 187 0 2 2 197
1 2 2 186 1 2 1 168 1 2 0 182
2 2 0 201 2 2 2 221 2 2 1 207
11.3 FRACTIONAL REPLICATIONS 293

(a) Analyze the experiment and draw conclusions.

(b) Comment on the design and devise an alternative
specifying the assumptions involved.

Problem 11.2.4. Design an experiment for blocks of size three

in which the treatment combinations are the same as those in
2
Problem 11.2.3 but AB and BC are confounded. Write out the
treatment combinations for each block. Discuss this design
relative to the design in Problem 11.2.3.

11.3 FRACTIONAL REPLICATIONS

In two-level factorial experiments dealing with fractional

replications it is easy to see at a glance whether or not a
design is satisfactory or not because there is only one exponent on
the letters of the main effects and interactions. This is not true
for the three-level case and peculiarities in this case have
generalities not visible in the two-level case.

It was noticed that if there were only three blocks in a

three-level case the highest factor interaction was confounded with
blocks. Using this idea as a basis for the aliases and setting the
identity or mean to be completely confounded with a highest factor
interaction part, it can be seen that the mean has only one degree
of freedom but the part has two. This peculiarity can be handled
by recognizing that a reflection or the square of that part will
generate aliases with two degrees of freedom each just as the
original part will. The result is that there is one degree of
freedom for each part, the original and the squared which provide
a basis for generation of aliases and a technique to follow for
smaller fractional experiments. Kempthome (1952) Chapter 21 is an
excellent reference for the general case.

11.3.1 A 1/3 Replicate of 3^ Factorial

A 1/3 replicate of a 3^ factorial experiment could be handled

as follows:
294 11. THREE-LEVEL FACTORIALS

I = ABCDE = (ABCDE)2
2 2 2 2
A = AB C D E = BCDE

B = AB2CDE = ACDE

AB = ABC2D2E2 = CDE
and so on.

The assumption that three factor and higher interactions are

zero would give the analysis of variance in Table 11.3.1.

TABLE 11.3.1

ANOVA of 1/3 Replicate of a 35 Factorial

Source df

Main effects (A,B,C,D,E) 10

Two factor interactions 40
Error 30
Total 80

This analysis shows that one obtains full information on the main
effects and two factor interactions from the one block (this is the
1/3 fraction).

11.3.2 A General Approach

If a 1/9 replicate of a 3n were used, we may set up a general

approach
X Y XY XY2 (X)2 (Y)2 (XY)2 (XY2)2

to investigate the aliases. For example, take a 1/9 replicate of a

35; one identity relationship is:

I = AB2CD = BCDE2= AC2D2E2= ABE = (AB2CD) 2=(BCDE2)2=(AC2D2E2)2 = (ABE) 2

A = ABC2D2 = ABCDE2 = ACDE = AB2E2 = BC2D2 = AB2C2D2E = CDE = BE

and so on for 13 main effects and interactions which have aliases.

11.4 SPECIAL FRACTIONALS 295

The reasoning for there being 13 main effects and interactions to

investigate is:

I has four aliases with two degrees of freedom each or using

the reflections or squares there are 8 aliases with one degree
of freedom each. Hence there is a total of nine degrees of
freedom on line I.

A has 8 aliases and there are 18 degrees of freedom on line

A. These can be generated by multiplication of A with I and
all of the aliases of I. There are 243 or 35 degrees of
freedom to be accounted for

243 = total df
-__9 = df confounded on line I
234 = df for main effects and interactions remaining.

234 df
= 13, the number of main effects
18 df for each set of aliases

and interactions with aliases in addition to the mean (I).

A reference that gives a thorough coverage of fractional

factorials at three levels with blocking of the fraction is Conner
and Zelen (1959).

11.4 A SPECIAL USE OF FRACTIONAL FACTORIALS

In the past few years research workers have recognized the

importance of fractional replicated experiments and have been
willing to assume three factor and higher interactions are usually
smaller than the two factor interactions, but have been hesitant in
assuming that the interactions used for the error are actually zero.
As a check on the latter, experimenters have been running duplicates
on various treatment combinations (usually three or four) to obtain
a "pure" estimate of the error. A test on the residual is usually
made and then a decision can be made on whether or not the average
of the interactions included in the residual are ,fnear enough" zero
296 11. THREE-LEVEL FACTORIALS

to use the residual as an error estimate. In order to maintain the

homogeneity of the variance of the estimates, these additional
observations are not used in the estimates of the effects or in the
analysis of variance.

In addition to utilization of a few observations to estimate

the error, experimenters are frequently finding that they may be
able to utilize information in a sequence of experiments by using
low fractions as blocks of larger fractions, obtaining information
at the completion of each block. This allows flexibility for the
research worker in that he may discard factors at various stages or
keep them in the design as he chooses.

For example, an experimenter finds that he can obtain only 28

observations under homogeneous conditions, but he is willing to
run the experiment at three different times for a total of 84
observations from five factors each at three levels. If he can
obtain information from the first set of 28 observations that would
give him insight into large main effects, he may be willing to
discard a factor or two. How can such a design be set up so that
intermediate analyses will give him some information, yet he will
have a sound overall analysis if he uses the 84 observations as
originally stated?

A 1/9 replicate of a 35 requires 27 observations and an

experimenter would have one observation to duplicate one of the
treatment combinations in order to check the error. Three repeats
of a 1/9 replicate would, however, be much less efficient in
estimating main effects and two factor interactions than considering
three blocks of a 1/3 replicate and run one block at a time. To
2
show this, let us take I = (ABCDE) = (ABCDE) for the mean, and
2 2 2 2 2
AB C = ACD E = BD E to be confounded with blocks within the 1/3
replicate. The overall analysis of variance (using the extra
observation in each block on a chosen treatment combination) is
given in Table 11.4.1.
11.4 SPECIAL FRACTIONALS 297

TABLE 11.4.1

ANOVA for 1/3 Replicate of 35 Factorial3

Source df

Blocks and/or
(a b 2c , bd 2e 2, ac d 2e 2) 2
Restriction error 0

Main effects 10
Two factor interactions 40
Residual (error) 28 May be broken into components
if the mean square is much
larger than within, then look
at the larger and smaller
pieces of these interactions
separately
Within duplicate error _3

Total 83

Within duplicate error (3 df) is used to test residual (error)

to check on the assumption that three and higher factor interactions
are zero.The extra three observations are not usedfor the
remainingpart ofthe analysis, totaling 80 df.

This is most satisfactory for the overall analysis, but is the

information from one of the blocks any value if one were run before
the other two?

The defining equations to consider for the 1/3 replicate with

3 blocks are:

ABCDE : x. + x0 + x_ + x. + xc = 0 or 1 or 2 mod 3 : Fraction

1 2 3 4 5
2
AB C : Xj + 2x 2 + x ^ = 0,1, and 2 mod 3 : Blocks

If*the experimenter assumes he will use the intrablock

subgroup, he will use zero level for ABCDE for all blocks and zero
2
level for AB C for the first block. Hence for the first block he
298 11. THREE-LEVEL FACTORIALS

has a 1/9 replicate with the identity

I = AB2C = ABCDE = ACD2E2 = BD2E2

2 2
= (AB C) = (ABCDE)
2= 222 222
(ACD E ) = (BD E )

It follows that

2 2222
A = ABC = A B C D E
2
= AC DE = ABD E
22 2
= BC = BCDE = CD^E
22 2
= AB DE

and similarly for the other main effects. In general, ifthe two
factor interactions are considered smaller than the main effects,
a conclusion may be made to discard some factors from the next two
blocks if the main effects are quite small.

If, however, no conclusion can be reached from the one block

regarding discarding factors the experiment can continue efficiently
by taking another block and possibly both additional blocks from
the three blocks inside the 1/3 replicate of the 3^ factorial
experiment.

This is more efficient than taking three replicates of a 1/9

replicate of the 3^ factorial.

11.5 SPECIAL LATIN SQUARE DESIGNS AS FRACTIONAL FACTORIALS

The Latin square design is an excellent design to remove

sources of variation in two directions when only one set of
treatments is of interest. This is the intent of the design.

In many scientific fields the Latin square design has been

used as a method of reducing the number of observations when three
factors are of interest. The danger in using this design for a
fractional replication is that the main effects are confounded with
parts of two factor interactions. To demonstrate this let us take
a 3 x 3 Latin square:
11.5 SPECIAL LATIN SQUARE FRACTIONALS 299

Columns
1 2 3

1 A B C
Rows 2 B C A
3 C A B

The design can be written in terms of a block of nine from a

factorial experiment,

abc
where:
0 0 0
Xx + X2 + 2X3 = 0, mod 3
10 1
2 0 2 satisfies the block. Hence
0 11
I = ABC2 = (ABC2)2
1 1 2
A = AB2C = BC2
2 1 0
2 2 2
0 2 2 B = AB C = KC
12 0 C = AB = ABC
2 2 1
AB = AC = BC

and it can be seen that all main effects are confounded with parts
of two factor interactions. If interactions are zero, this
fractional factorial is of value to investigate main effects. In
large investigations, the fractional setup as a Latin square has
advantages as a screening device if the lower order terms express
the larger source of variation of the variable being analyzed. In
other words, even if there is confounding of the two factor
interactions with main effects, the main effects ought to be
larger and indicate the major source of variation.

It should be pointed out that the Latin square must have ia

pattern, such as
A B C D E
B C D E A
C D E A B
D E A B C
E A B C D
300 11. THREE-LEVEL FACTORIALS

before an identity can be written down for the fractional factorial.

In all cases, however, the mixup* of the main effects and two factor
interactions exist. In a nonsystematic Latin square the identity
relationship cannot be explicitly shown.

11.6 EXTENSION OF 3n TO pn, WHERE p IS A PRIME NUMBER

3
Consider a 1/5 replicate of a 5 factorial experiment for
which the identity relationship is:

I = ABC = (ABC)2 = (ABC)3 = (ABC)4

2
A = A BC = A B C
322 433
= ABC = ABC,
544 mod 5

To get the first letter of each interaction to have an exponent of

one, it is necessary to raise the interaction to appropriate powers.
For example for A we have:

A = (A2BC)3 = (A3B2C2)2 = (A4B3C3)4 = a °b 4c 4

= a b 3c 3 = a b 4c 4 = a b 2c 2 = BC

and so on for the other aliases.

If the experiment is 1/25 replicate, there would be 24

aliases for I or y. As in the three level case where the square
was the reflection, here there are squares; cubes, and quartics to
handle in order to develop the whole set of aliases.

This same procedure is expansible to seven levels and so on

for any p, prime number of levels.

When dealing with a large number of levels and factors in a

small fractional replicated experiment, the problem of handling the
aliases is enormous and usually experimenters resort to computers
to develop the block and aliases for the experiment.

Problem 11.6.1. (a) Show that there does not exist a

fractional factorial with a well defined identity with
aliases for the example in Chapter 8 (the 5 x 5 Latin square).
(b) Show for (a) above the arrangement of the treatments
within the rows and columns that would give a fractional
11.7 REFERENCES 301

replication with a well-defined identity. State the identity

for your design and show the correct ANOVA indicating the
assumptions that must be made on interactions.

11.7 REFERENCES

Conner, W. S. and Zelen, M, Fractional Factorial Experiment Designs

for Factors at Three Levels, National Bureau of Standards, Applied
Math. Series 54, 1959.
Davies, 0. L. (Ed.) Design and Analysis of Industrial Experiments,
Oliver and Boyd, Edinburgh, 1971.
Hicks, C. R. Fundamental Concepts of Design of Experiments, 2nd ed.
Holt, Rinehart and Winston, New York, 1973.
Kempthorne, 0. Design and Analysis of Experiments, Wiley, New York,
1952. Distributed by Krieger Pub. Co., Huntington, New York.
Margolin, B. H. Technometrics 9:245 (1967).
Winer, B. J. Statistical Principles in Experimental Design, 2nd ed.
McGraw-Hill, New York, 1971.
 Chapter 12

MIXED FACTORIAL EXPERIMENTS AND OTHER INCOMPLETE BLOCK DESIGNS

There are many designs available for factorial experiments

with unequal levels. These designs are not handled the same as the
equal leveled factorials, so special emphasis must be made on these
so called "mixed" factorial completely randomized incomplete block
designs (Kempthorne, 1952, Chapter 18). These are the designs that
are needed for experiments in which there are large numbers (25 or
more) levels of one factor only investigated simultaneously. For
these designs it is sometimes easy to force the number of levels to
be a perfect square and then use a special type "lattice" design to
compare the levels or treatment means. If the number can only be
forced to be one number times another, "rectangular" lattice
designs may be used. Federer (1955) is a very good reference for
lattice designs.

Many other incomplete block designs exist and special plans are
available for some of them in Cochran and Cox (1957). The treatment
of all incomplete block designs can follow a different procedure
from the one presented here but the results are essentially the
same. In all instances in this chapter it is assumed that there is
complete randomization in the blocks considered.

12.1 DESIGNS FOR FACTORIAL EXPERIMENTS WITH THE NUMBER OF

LEVELS THE SAME PRIME POWER

If a factorial experiment has the number of levels of each of

the factors equal to k = pm, where p is a prime number and m is an

302
12.1 SAME PRIME POWER 303

integer greater than one, special incomplete block designs can be

used.

12.1.1 One Factor (to Demonstrate Pseudofactor)

Take an example in which there are eight levels of a factor or

call these levels treatments. Then one can consider this a 2
pseudofactorial where there are three pseudofactors each at two
levels and handle the design as previous two-leveled factorials
except that the experimenter must eventually discuss the results
in terms of the eight treatments. One way of renaming the eight
treatments is the following:

Treatment Pseudofactor
tf
CM

a ai *3
1 0 0 0
2 0 0 1
3 0 1 0
4 0 1 1
5 1 0 0
6 1 0 1
7 1 1 0
8 1 1 1

It should be understood that the correspondence between the

treatment and the pseudofactor is arbitrary, but once it is decided
upon for that design the correspondence must be maintained
throughout the experiment.

Remember this subsection is for demonstration of the

pseudofactor technique and does not include a design to provide an
error estimate.
304 12. MIXED FACTORIALS

12.1.2 Two Factors

If there are two factors each with a number of levels equal to

a multiple of prime numbers, another pseudofactorial type design
may be used. Consider two factors a and b each with eight levels,
one can call these factors six pseudofactors each with two levels.
A possible naming method follows:

Factor 1 Factor 2

a b
*1 *2 bi fa2 b3
1 0 0 0 1 0 0 0
2 0 0 1 2 0 0 1
3 0 1 0 3 0 1 0
4 0 1 1 4 0 1 1
5 1 0 0 5 1 0 0
6 1 0 1 6 1 0 1
7 1 1 0 7 1 1 0
8 1 1 1 8 1 1 1

If a complete 8 x 8 factorial were run, the analysis of the

completely randomized factorial and of the pseudofactorial is
given in Table 12.1.1.
From this analysis it is seen that if a main effect of a
pseudofactor is confounded this is no worse than if an interaction
of the same pseudofactor (e.g., A^A2 ) is confounded because all of
the main effects and interactions from the same factor are really
main effects of the factor of interest (e.g., A here).
12.1 SAME PRIME POWER 305

TABLE 12.1.1

ANOVA for 8 x 8 Factorial Experiment

Original Pseudofactorial

Source df Source df

A 7 1
Ai
1
A2
1
A3
1
A1A2
1
A1A3
1
A2A3
1
A1A2A3

B 7 1
B1
1
*2
1
B3
1
B1B2
1
B1B3
1
B2B3
B,B„B„ 1
12 3

AB 49 1
AiBi
A1B2 1

1
A1A2A3B1B2B3

Total 63 63
306 12. MIXED FACTORIALS

For the above experiment consider the case in which only 16

treatment combinations can be run under homogeneous conditions but
that four blocks may be run.

One possible design (1) is:

X Y XY

A1A2A3 B1B2B3 A1A2A3B1B2B3

and the analysis of variance is given in Table 12.1.2.

TABLE 12.1.2

ANOVA for 4 Blocks of 16 Using Design (1)

Source df Information

Blocks 3
Restriction n
error u
A 6 6/7
B 6 6/7
AB 48 48/49

:2) is:

X Y XY
A1B1 A
2B2 A1A2B1B2

and the analysis of variance is given in Table 12.1.3.

TABLE 12.1.3

ANOVA for 4 Blocks of 16 Using Design (2)

Source df Information

Blocks 3
Restriction o
error
A 7 Full
B 7 Full
AB 46 46/49

Conclusion: Usually the second design would be preferred.

12.2 DIFFERENT PRIME LEVELS 307

Regardless of the design actually used, the pseudolevels must

be converted back to the original levels before the experiment can
be run. Consider the second design, the two defining equations
(if the variables for a^, a^, b^, b2, b^ are X^, X^, X^, X^, X,.,
X^) are:
X1 + X4 = 0, 1 mod 2

x2 + x5 = °, 1 mod 2

The corresponding intrablock subgroup is:

Pseudofactor Factor

1 X2 X3 X4 X5 X6

0 0 0 0 0 0 1 1
1 0 0 1 0 0 5 5
0 1 0 0 1 1 3 4
1 1 1 1 1 0 8 7

The experimenter is interested only in the block that contains

the treatment combinations in terms of the actual factor a and 6.
He has no interest in the pseudofactor block. It is merely a ’’means
to an end.”

12t2 DESIGNS IN WHICH THE NUMBER OF LEVELS ARE

DIFFERENT PRIME NUMBERS

Consider an example of a 2 x 2 x 3 x 3 factorial. Since all

the factor’s levels are prime numbers, no pseudofactorial is needed.
Let the factors be a, 6, c, and d in that order.

The analysis of a completely randomized 2 x 2 x 3 x 3 factorial

experiment and the corresponding confounding pieces of main effects
and interactions is given in Table 12.2.1.
308 12. MIXED FACTORIALS

TABLE 12.2.1

ANOVA for 2 x 2 x 3 x 3 Factorial

Source df Pieces to Confound

A 1 A
B 1 B
AB 1 AB
C 2 C
AC 2 AC
BC 2 BC
ABC 2 ABC
D 2 D
AD 2 AD
BD 2 BD
ABD 2 ABD
CD 4 CD and CD2
ACD 4 ACD and ACD2
BCD 4 BCD and BCD2
ABCD __4 ABCD and ABCD2

Total 35

The possible block sizes for this experiment are 2, 3, 4, 6, 9, 12,

and 18.

When choosing interactions to confound, all factors must have

the same prime number of levels within a defining equation such as
X or Y (e.g., the interaction AB may be selected as a defining
equation since both a and b have the same number of prime levels)
for the procedure to work. This procedure will not work if AD is
used as a defining equation. The resulting interaction need not be
one of the defining equations to establish the intrablock subgroup
but it is necessary to understand which generalized interaction is
confounded with blocks in order to analyze the data properly.
12.2 DIFFERENT PRIME LEVELS 309

12.2.1 Design with 9 Blocks of 4 for the 2 x 2 x 3 x 3

X Y XY XY
c D CD CD'
df = 2 2 2 2

This is a very poor design because all main effects of C and D are
lost and so is their interaction.

12.2.2 Design with 6 Blocks of6 for the 2 x 2 x 3 x 3

X Y XY
AB CD ABCD
df = 1 2 2

This is a fairly good design because all main effects and 1/2 of CD
are estimable. The informationon AB,however, is lost.

12.2.3 Design with 4 Blocks of9 for the 2 x 2 x 3 x 3

X Y XY
A B AB
df = 1 1 1

This is a poor design for the same reasons as those given in

Section 12.2.1.

12.2.4 Design with 3 Blocks of 12 for the 2 x 2 x 3 x 3

X
CD or CD2
df = 2

This is not as bad a design as the one given in Section 12.2.3 since
one obtains half information on CD, and full information on all main
effects and the other interactions.
310 12. MIXED FACTORIALS

12.2.5 Design with 2 Blocks of 18 for the 2 x 2 x 3 x 3

X
AB
df = 1

This design is not as good as the one given in Section 12.2.4

because all information on AB is lost.

12.2.6 A Partially Balanced Incomplete Block Design

(Four Replications of 6 Blocks of 6)

df — 1 2 2
JC Y XY
Replication 1 AB C ABC
Replication 2 A CD ACD
Replication 3 B CD2 BCD-
Replication 4 AB D ABD

This design provides 3/4 information on all main effects, 1/2

information on AB and 3/4 information on CD (with 4 degrees of
2
freedom since one obtains 1/2 information (CD ) in replication 2 and
another 1/2 (CD) in replication 3). In mixed factorial incomplete
block design it is very difficult to obtain balanced designs without
many replications.

12.3 DESIGNS THAT HAVE THE NUMBER OF LEVELS AS

PRODUCTS OF PRIME NUMBERS

In some mixed factorial experiments the factors have the number

of levels that are products of prime numbers and other factors with
prime numbers. Consider the example 2 x 2 x 6 x 6. One can make
4 2
this into an experiment that is a 2 3 by allowing
12.3 LEVELS, PRODUCT OF PRIMES 311

a k ci c2 ^1 ^2
2 2 2 3 2 3

The pseudofactors could be

c
C1 C2
1 0 0
2 0 1
3 0 2
4 1 0
5 1 1
6 1 2

Various incomplete block designs could result from such a factorial

experiment.

Example 12.1: An experiment in industrial engineering was set

up to investigate the effects of various factors on reducing user
dissatisfaction with microforms. Microforms consist of photographs
of hardcopy originals reduced in size for efficient storage. The
measured variables include time measurements of reading,
comprehension, and subjective preferences.

The factors and their levels include:

Factors Levels

(a) Screen angles from

horizontal 0°, 45°, 90°, 105°
(b) Screen luminance
(foot lamberts) 30, 40, 50
(c) Ambient illumination
(foot candles) 20, 40, 60
(d) Visual tasks Reading (R), Information
Searching (I)
(e) Type of projection Front (F), Back (B)

A subject can comfortably perform 24 conditions (treatment

combinations) at one sitting. This takes about one and one half
312 12. MIXED FACTORIALS

hours. Since there are 4 x 3 x 3 x 2 x 2 = 144 treatment

combinations, there must be six sittings for each subject to
complete the experiment. This will demand six blocks within each
subject (replication). There are twelve random subjects available
for the experiment. The usual assumption that there is no
correlation among errors within a subject is made.

The design of the experiment, then, must be an incomplete

pseudofactorial, and preferably partially balanced to obtain as
much information on all main effects and two factor interactions as
possible. Use the letter notation [a., b, c., d, e) for designating
the factors and the usual capital letters (A, B, C, D, E) to indicate
the effects.

Since the first factor, a, has four levels we will regard it as

2
a 2 pseudofactorial where the pseudofactors a^ and a 2 each has two
levels as follows:

a
f
t
CM I
1

0° 0 0
45° 0 1

90° 1 0
105° 1 1

With five degrees of freedom to confound with blocks (sittings)

for each subject, we may construct the following partially balanced
incomplete block design for the 12 subjects:

if — ► 2 1 2 Number of
Condition X Y XY subjects
I. BC A1DE A1BCDE 2
II. BC A2DE A2BCDE 4
III. BC2 A^DE AjBC2DE 2
IV. BC2 A^DE A1A2BC2DE 4
12.3 LEVELS, PRODUCT OF PRIMES 313

Hence we can see that we get 1/2 information on BC interaction

since the 2 df BC is represented by 1/2 of the subjects and the
2
other 2 df BC is represented by the other six subjects. In a
similar manner there is 2/3 information on ADE because A^DE is
calculated using eight subjects, A 2 DE is calculated using eight
subjects and AJA 2 DE is calculated using eight subjects with full
information on all main effects and all other two and three factor
interactions. Since we must confound something with two degrees of
freedom in order to get 5 df confounded with blocks, the best
(highest factor interaction of factors with 3 levels) we can do is
the design above.

The reader must understand that for each subject blocks are
chosen at random to run first, second and so on. Then, given a
block the treatment combinations within that block are chosen at
random to find out which will be run first, second and so on for
each sitting.

To indicate how the treatment combinations are selected for

each block, let us use the condition I : BC, A^DE, A^BCDE and the
intrablock subgroup. Then, we must solve the following defining
equat ions:

BC : X3 + X4 = 0 mod 3

A.DE : X. + X. + X. = 0 mod 2
1 1 5 6

The resulting treatment combinations follow:

314 12. MIXED FACTORIALS

Psuedoblock Real 'or usable

X3 block for experimenter

X1 X2 X4 X5 X6
6 c d e a b C d G
*1 a2
1. 0 0 0 0 0 0 1. 0° 30 20 R F
2. 0 1 0 0 0 0 2. 45° 30 20 R F
3. 0 0 0 0 1 l 3. 0° 30 20 I B
4. 0 1 0 0 1 l 4. 45° 30 20 I B
5. 0 0 1 2 0 0 5. 0° 40 60 R F
6. 0 0 2 1 0 0 6. 0° 50 40 R F
7. 0 1 1 2 0 0 7. 45° 40 60 R F
8. 0 1 2 1 0 0 8. 45° 50 40 R F
9. 0 0 1 2 1 1 9. 0° 40 60 I B
10. 0 1 1 2 1 1 10. 45° 40 60 I B
11. 0 0 2 1 1 1 11. 0° 50 40 I B
12. 0 1 2 1 1 1 12. 45° 50 40 I B
tO
Oo

13. I 0 0 0 1 0 13. 30 20 I F
14. 1 0 0 0 0 1 14. 90° 30 20 R B
15. 1 0 1 2 1 0 15. 90° 40 60 I F
16. 1 0 1 2 0 1 16. 90° 40 60 R B
to
oo

17. 1 0 2 1 1 0 17. 50 40 I F
o
o

18. 1 0 2 1 0 1 18. 50 40 R B
19. 1 1 0 0 1 0 19. 105° 30 20 I F
20. 1 1 0 0 0 1 20. 105° 30 20 R B
21. 1 1 1 2 1 0 21. 105° 40 60 I F
22. 1 1 1 2 0 1 22. 105° 40 60 R B
23. 1 1 2 1 1 0 23. 105° 50 40 I F
24. 1 1 2 1 0 1 24. 105° 50 40 R B

Assuming that all 24 different blocks have been set up for the
whole experiment we next outline the ANOVA showing the source, df
and amount of information for each source.
.3 LEVELS, PRODUCT OF PRIMES 315

TABLE 12.3.1

ANOVA for Microform Experiment

Source df Amount of Information

Subj ects 11
Blocks and/or interactions 60
Restriction error 0

A 3 Full
B 2 Full
AB 6 Full
C 2 Full
AC 6 Full
BC 4 Half
ABC 12 Full
D 1 Full
AD 3 Full
BD 2 Full
ABD 6 Full
CD 2 Full
ACD 6 Full
BCD 4 Full
ABCD 12 Full
E 1 Full
AE 3 Full
BE 2 Full
ABE 6 Full
CE 2 Full
ACE 6 Full
BCE 4 Full
ABCE 12 Full
DE 1 Full
ADE 3 2/3
BDE 2 Full
ABDE 6 Full
CDE 2 Full
ACDE 6 Full
BCDE 4 Full
ABCDE 12 5/6
Error (Residual) 1513
Total 1727
316 12. MIXED FACTORIALS

Problem 12.3.1. (a) Find the usable block in the microfilm

experiment if we use the defining equations
BC : X3 + X4 = 1, mod 3
AXDE :Xx + X5 + X6 = 0, mod 2
(b) Indicate where correlated errors may exist in this
experiment.

Problem 12.3.2. Find the usable block for the intrablock

2 2
subgroupusing Condition III; BC , A^DE, A^BC DE.

Problem 12.3.3. (a) If only 6 subjects were used in the

microfilm experiment (Example 12.1), write out the ANOVA
showing source, df and amount of information.
(b) Discuss the value of using 12 subjects rather than
only 6 and give your recommendations plus your reasoning.
Consider the inference space and correlated errors.

Problem 12.3.4. In a farm cardiac study a problem

investigating the effects of certain farm tasks upon heart
action were studied. The variable measured was oxygen uptake
when men carried one or two pails at six different heights
using three different loads. Three men were used in the
experiment and the complete factorial would have been a
3 x 2 x 6 x 3 experiment.
Each man could perform only six tasks in any one day,
only one man could work on any one day, and eighteen days
were allowed for the experiment. All two factor interactions
are of interest but three factor and higher are assumed zero.
Men are random and may be considered replications. The other
three factors are fixed.
(a) Design a good experiment identifying the treatment
combinations as the experimenter would want them.
(b) Show the analysis including the amount of information
on each main effect and two factor interaction for the three
fixed factors (Remember men act only as replications and are
not of interest per se.)
(c) Describe the randomization procedure.

12.4 DESIGNS IN WHICH THE PSEUDOFACTORS HAVE DIFFERENT POWERS

ON THE PRIME NUMBER OF LEVELS

An example is 9 x 8 or

ai *2 61 b2 b3
3 3 2 2 2
12.6 FRACTIONALS 317

2 3
which results in a pseudofactorial experiment of 3 2 if incomplete
block designs are required.

12.5 DESIGNS IN WHICH THE NUMBER OF LEVELS ARE PRODUCTS

OF POWERS OF PRIME NUMBERS

An example of this type design is the case in which a , b, c ,

and d are each at 6 levels and e., g , and k are each at 10 levels.
The factorial experiment can be written as
4 4
6 10

but to handle incomplete block designs it may be broken into

ai a2 bl b2 C1 C2 dl d 2 ei £2 h <2 91 92 fel k 2
2 x 3 x 2 x 3 x 2x3x 2x3x 2x5x 2x5x 2x5x 2x5
8 4 4
or a pseudofactorial experiment 2 3 5 .
3
Problem 12.5.1. In a 6 experiment, show a good system of
confounding blocks of 36.

Problem 12.5.2. Using Problem 8.2.1 as a reference:

(a) Rearrange the materials a, b, c, d within the rows
and columns so that an identity can be found(using
pseudofactors) that will be appropriate for all 16 treatment
combinations in the experiment. This will show the fractional
replication used to obtain that particular Latin square.
(b) Show the correct analysis for fractional replication
in (a). Indicate the confounding.

12.6 FRACTIONAL MIXED FACTORIALS 2m3n

The designs in the booklet by Conner and Young (1961) are

constructed so that all main effects and two factor interactions
are estimated under the condition that three factor and higher
interactions are zero. A design is constructed for 39 pairs (m, n)
included from (m + n = 5) through (m + n= 10) where (m, n ^ Q).

The objective of the designs was to keep thenumber of required

treatment combinations small while retaining as much orthogonality
among the estimates as possible.
318 12. MIXED FACTORIALS

A brief example shows the basic method of construction of the

3 2
design: Consider a 2 3 experiment with 72 possible treatment
combinations. If one desires a 1/2 replicate of this experiment he
must confound either one degree of freedom for a three factor
interaction or use a nonorthogonal design. If three factor
interactions are of interest to the experimenter he would attempt
to confound a degree of freedom from the five factor interaction
^1^2^3R1R2* as mentioned earlier one cannot use an interaction
from factors with different numbers of levels as a defining contrast.
A very good design, however, is obtained by structuring a design by
working with the two and three level factors separately and then
combining the results.
3
Working with the 2 part of the experiment and the A^^A^
interaction for confounding purpose we form the defining equations
Xi + X2 + X3 = 0, 1, mod 2, which produces the two following
groups of treatment combinations:

ai *2 *3 *1 *2 *3
0 0 0 1 1 1
1 1 0 1 0 0
1 0 1 0 1 0
0 1 1 0 0 1

2
Similarly using B2 for confounding the 3 part, l
Wx + W2 = 0, 1,
2, mod 3 (using Wfs insteadI of Xfs to identity the equation for
mod 3), the three resulting groups of treatment combinations are

r2
Ri R3
b 6, b, 6 bn
1 2 1 2 1 2

1 0

0 1
CM

2
12.7 LATTICE DESIGNS 319

To form the 36 treatment combinations given in Conner and Young (1961)

one may use the following:

S1R1 S2R2 S2R3

al a2 a3 bl b2 al a2 a3 bl b2 ai a2 a3 bl b2
1. 0 0 0 0 0 13. 1 1 1 1 0 25. 1 1 1 2 0
2. 0 0 0 1 2 14. 1 1 1 0 1 26. 1 1 1 0 2
3. 0 0 0 2 1 15. 1 1 1 2 2 27. 1 1 1 1 1
4. 1 0 0 0 0 16. 1 0 0 1 0 28. 1 0 0 2 0

12. 0 1 1 2 1 24. 0 0 1 2 2 36. 0 0 1 1 1

The analysis of an actual example with data is given in Conner and

Young (1961, pp. 9-12). The analyses could be programmed on
computers, since, of course, the least square estimates of the
effects do become difficult to compute. (These computations assume
b*s are equally spaced quantitative factors and trends are desired.)
Regression programs may be used for analyses on the data from such
experiments, too.

12.7 LATTICE DESIGNS (PSEUDOFACTORIAL)

The general class of lattice designs are really a special case

of the pseudofactorial incomplete block designs with one factor
(discussed briefly in Section 12.1 in this chapter). The number of
levels of the factor is usually quite large since the most common
usage of the design is in comparing varieties in agronomic studies.
Of course, the design can be used for various conditions where
there are many treatments, all of equal interest, to be compared.

Since there are too many treatments or varieties to be placed

in one block, the main problem is to group the treatments in
incomplete blocks in such a manner that maximum information is
obtained on the average yields. The various groupings actually
320 12. MIXED FACTORIALS

make up the various designs. In all designs the interest is in

retaining information on the treatments or varieties from between
and within the incomplete blocks. Elaborate designs (using partial
confounding methods) confounding in different directions and
utilizing information from many sources, have been evolved. Three
of the best sources of information on such designs are in the books
by Federer (1955, Chapters XI and XII), by Kempthorne (1952,
Chapters 22-26), and Cochran and Cox (1957, Chapter 10).

There is a general notion concerning the effects confounded

with blocks that must be stressed. In all previous designs there
were more than one factor concerned, so that higher factor
interactions were of less value to the investigator than lower
factor interactions and main effects. When there is only one
factor (in lattice designs) each treatment is as important as any
other, and the confounding of effects is completely artificial.

What is meant by ’’artificial" in this case is the effects

k-1 2
A, B, AB, ..., and AB (if there are k treatments) have no real
meaning as effects, also they are equally valuable because all
represent main effects or comparisons of the treatment means.
Hence, in all cases in lattice designs, the pseudomain effects are
usually confounded first, and the designs are built up from these
main effects and simple interactions.

A few of the lattice designs will be discussed briefly, more

to explain the type of design than to give an account of the minute
details of the layout and analysis of each. Details are given so
well by Federer and Kempthorne that there seems no reason to cover
them here.

12.7.1 One-Restrictional Designs

(a) Simple Lattice

This design has one restriction on randomization (that is,

2
blocking in one direction). Consider the case of k varieties or
12.7 LATTICE DESIGNS 321

treatments in blocks of k varieties. We can consider this

experiment as a pseudofactorial in which pseudofactor a is at k
levels and pseudofactor b is at k levels. If main effect A is
confounded with blocks in replicate one and B is confounded with
blocks in replicate two, there is blocking in only one direction
and the two pseudo main effects are each confounded in the two
replications. This basic design structure is called a simple
lattice.

If there were a repeat in which two more replications of the

experiment were run, in one of the new replications A was confounded
and in the other new replication B was confounded, the design would
be called a double simple lattice. The analysis for such a design
is given in Federer (1955, pp. 319-333).

It can be seen that even in this simple lattice design there

will be information in one replication lost due to blocking, but
that same information will be available in the other replication.
In other words this is a partially confounded type design with
blocks nested in replications and information on every treatment
will be available even if there are large block to block differences.
In addition, it can be seen that if the block to block variation is
very small or negligible the data may be analyzed as a randomized
complete block design where the replications act as the blocks in
this RCBD. For the simple lattice design analyzed as a randomized
complete block design, the analysis is given in Table 12.7.1.

TABLE 12.7.1

ANOVA of a RCBD from a Simple Lattice

Source df

Replications 1
Restriction error 0
Treatments k2 - 1
R x T (error) k2 - 1
322 12. MIXED FACTORIALS

Similarly for the double simple lattice the analysis for RCBD is
given in Table 12.7.2

TABLE 12.7.2

ANOVA for a RCBD from a Double Simple Lattice

Source df

Replications 3
Restriction error 0
Treatments k2 - 1

R x T (error) 3(k2 - 1)

Federer (1955, p. 324) shows a method to compare the efficiency of

double lattice design to randomized complete blocks using the ratio
of the two effective error variances. A quicker, but cruder,
method (described in Chapter 5) is to use the Fisher*s amount of
information method to determine the efficiency of the lattice
design relative to the randomized complete block design. The
appropriate ratio is

(nx + l)(n2 + 3) s2
_ 2
(nx + 3)(n2 + 1) s j

where
n^ = degrees of freedom for the error in.the lattice
2
s^ « error mean square for the lattice

n2 = degrees of freedom for the error in the randomized

complete block
2
$2 ~ error mean square for the randomized complete block.

The techniques of utilizing intrablock information and

interblock information will not be discussed here. Of course, both
Federer (1955) and Kempthorne (1952) discuss methods thoroughly.
12.7 LATTICE DESIGNS 323

(b) Triple Lattice

If three replicates or multiple of three replications can be

2
run in a k experiment, we can confound A, B and AB in the various
replications equally.

Of course quadruple lattices, and so on can be used depending

on the number of replications and type of confounding in each set.

(c) Balanced Lattice

If k + 1 replications can be taken

A, B, AB, AB2, ..., ABk_1

can be confounded with blocks and there will be equal information

on each pseudo effect and interaction. As a result there is equal
information on all treatment or varietal means. This design is an
example of a balanced incomplete block design (BIBD).

This completes the designs called one-restrictional, one­

dimensional.

(d) Other One-Restrictional

In addition to one-dimensional, there are two-dimensional

designs called rectangular lattices (k p) discussed by Federer
(1955, pp. 343-348). In this design the number of varieties or
treatments are not a square. Further discussion on three
3 4
dimensional, cubic lattices (k ), four dimensional (k ) and so on
to n-dimensional (kn) is given by Federer (1955, p. 348-368) and
the references given by Federer (1955) and Kempthome (1952). This
completes the one-restrictional designs.

12.7.2 Two-Restrictional Designs

(a) Lattice Square

This is the first design with two restrictions on randomization

in the lattices and compares, quite naturally, to the Latin square
for complete blocks.
324 12. MIXED FACTORIALS

In this case A is confounded in one direction and B is

2
confounded in another direction for the same replication for k
varieties or treatments, Federer (1955, Chapter XII).

(b) Semibalanced Lattice Square

As more replications are introduced into the design, more

interactions are confounded with blocks. For example

Effect Replication
confounded
Wlth______ 1 2 3 ... (k + lj/2
Rows A AB AB3 ... ABk"2
Columns B AB2 AB4 ... ABk_1

(c) Balanced Lattice Square (BIBD)

Continuation of the replication pattern established by the

semi-balanced lattice square by reversing the confoundings in rows
and columns give a complete balanced system.

Effect Replication
confounded ,
with 1 ... — ^— (— 2 — + ^ ... (k + 1)

Rows A ... ABk'2 B ... ABk_1

Columns b____ ABk~* A ... ABk~2

A reference for the two-restrictional designs is Federer (1955,

pp. 377-412).

12.7.3 Designs with More Than Two Restrictions

A good reference is Federer (1955, pp. 413-414).

12.8 ADDITIONAL INCOMPLETE BLOCK DESIGNS

Designs that have only one factor with more than one level and
all the levels of the factor cannot be handled in a single block have
12.9 REFERENCES 325

been outlined in Chapter 11 of Cochran and Cox (1957). They have

plans and analyses for these designs that are most useful for cases
not covered in this material. The reference on pages 468 and 469
of Cochran and Cox (1957) are excellent for this type problem.

A calculus for factorials is also available. A recent paper

on applications of this concept is given by Kurkjian and Woodall
(1972).

12.9 REFERENCES

Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research,

McGraw-Hill, New York, 1952.
Cochran, W. G. and Cox, G. M. Experimental Designs, Wiley, New York,
1957.
Conner, W. S. and Young, S. Fractional Factorial Designs for
Experiments with Factors at Two and Three Levels, Nat11. Bureau
of Standards, Applied Mathematics Series 58 (1961).
Federer, W. T. Experimental Design Theory and Application,
Macmillan, New York, 1955.
Kempthorne, 0. Design and Analysis of Experiments, Wiley, New York,
1952. Distributed by Krieger Pub. Co., Huntington, New York.
Kurkjian, B. and Woodall, R., HDL-TR-1596, U. S. Army Materiel
Command, Harry Diamond Laboratories, Washington, D. C., 20438
(May 1972).
 Chapter 13

RESPONSE SURFACE EXPLORATION

In the preceding chapters the designs had one common

characteristic, that of finding out the influences of the various
factors on the variable analyzed. In this chapter that characteristic
still exists, but the emphasis includes finding that particular
treatment combination which causes the maximum (or minimum) response,
yield, or variable analyzed. In addition to finding the treatment
combination that causes the optimum yield, response surface
exploration includes investigation of the response surface near the
optimum yield. Since one can determine the minimum of a response
surface by finding the maximum of the negative of the actual
response surface we will only discuss the concept of maximum. In
this chapter it should be recognized that, at most, a second degree
polynomial will be investigated. In general, the response surface
will be some nonlinear function which may be sufficiently approximated
with a quadratic polynomial in a small region near the optimum
operating condition.

Not all of the response surface or sometimes called ’’optimum”

designs will be given here and even those given will not be
described in depth. The purpose of presenting the designs in this
manner is to expose the reader to some of the most useful designs,
describe the strengths and weaknesses of each method and provide
references if further reading is desired. The approach is to divide
the designs into fixed (nonsequential) type, in which the decision
of the maximum response is made with a given set of observations,
and into sequential, in which the first design is used to provide
information on how to set up the next design and so on until the

326
13.1 FIXED DESIGNS 327

experimenter is satisfied he has reached the maximum or ’’near

enough11 maximum response. In addition, an attempt has been made to
list the designs within each classification of fixed and sequential
from least to most desirable (from a statistician’s view).

13.1 FIXED DESIGNS

The notion of fixed factors has been used in factorial

experimentation with the understanding that all of the levels of
interest of that factor were in the experiment. The fixed or
nonsequential designs used to seek the optimum levels of the factors
in a given experiment include previously discussed designs and
additional designs that have been used successfully by experimenters.
The additional designs include those which have fewer observations
than factors and mixture designs that require the sum of the levels
in a treatment combination to add to a constant.

13.1.1 Random Balance

The random balance design was suggested by Frank E. Satterthwaite

(1956, 1959). The design has been used by various industrial
organizations since then. The purpose of this design is to use very
few observations in an easy manner to find the maximum response and
also to try to describe the response surface.

There has been quite a bit of controversy among statisticians

concerning the design and analysis of the random balance experiment.
To examine the peculiarities of this design, let us consider an
example. An experimenter can name 16 factors or independent
variables (X^, X^) that may influence the response (y).
Let us say that the number of levels of factors varies from 2 to 7
as follows:

Factor X2 X3 X4 X$ Xfi X? Xg XQ X1Q Xn X12 X13 XM X1£ XJ6

Levels 3 2 7 4 2 3 6 5 2 3 5 6 3 4 5 2
328 13. RESPONSE SURFACE EXPLORATION

The total number of treatment combinations in the population is the

product of the levels or in this case 653,184,000. Hence the usual
fractional replicated experiment seems impossible, especially when
the experimenter wants to look at only 25 treatment combinations.

The method of obtaining the 25 treatment combinations out of the

possible 653,184,000 by the random balance method follows:

for take a random number between 1and 3, say 3

for X2 take a random number between 1and 2, say 1
for X^ take a random number between 1and 7, say 6
and so on, until finally,
for X ^ take a random number between 1 and 2, say 2

Then the first treatment combination for (X., X0, X,, ..., X,.) to
1 Z O ID
be used in the experiment is (3, 1, 6, ..., 2). For the next
treatment combination take a random number between 1 and 3, say 2
for Xj. Continue as above to obtain the second treatment combination.
This method is continued for all 25 treatment combinations with the
restriction that the number of treatment combinations containing
each level of each factor be as equal as possible. For example,
consider factor X ^ each of the three levels will be represented
at least eight times, but one level must be represented nine times;
whereas for factor Xg, each of the five levels must be present in
the experiment five times. This completes the design.

One part of the analysis is to declare that treatment

combination with the highest response the winner. If the number of
treatment combinations used in the experiment is more than the
number of factors, a multiple regression analysis using only the
linear main effects may be used assuming all the other effects and
interactions are zero.

In this case y = 8Q + 81 X^^ + 82 *2 + 33 X3 + ... + 816 X16 + e.

If the number of treatment combinations used in the experiment is
less than the number of factors, it has been suggested that graphing
of effects and interactions may still give good information. One
13.1 FIXED DESIGNS 329

experimenter even declared that he could evaluate the higher factor

interactions with fewer observations than factors. If one considers
the design as a fractional replicated experiment with an unknown
identity, he must acknowledge the difficulty in such a claim.

The following statement given by James H. Stapleton in the

October, 1959, issue of the American Statistician summarized the
feelings of the authors about this design:

The major error committed by those who advocate

the ’’random balance” approach is in their viewpoint,
centering on the results to be expected from a random
experiment and not on the results to be expected from
the particular experiment chosen. Let us suppose that
there are k factors of interest, that a complete factorial
would require M experiments, and that n experiments are
to be made, with n much smaller than M. The problem is
to choose these n experiments from the M possible in
some optimum fashion. The solution of the ’’random
balancists” is to choose these n at random under possibly
a few restrictions. It is hoped that these n experiments
will form an optimum design in the sense that the important
effects will be measured as well as possible. Why leave
it to chance? The only possible answer to this is that
it is felt that somehow chance will come up with a
better design than the statistician can find deterministicly.
If this is so, and I very much doubt that it is, it only
points out the inadequacy of the theory of the design of
experiments. Certainly if M is much larger than n, the
design determined in a nonrandom manner must include much
complete and partial confounding, but so must the random
design. As a statistician I feel much better if I have
control over this confounding.
330 13. RESPONSE SURFACE EXPLORATION

13.1.2 Systematic Supersaturated

Supersaturated designs are characterized by there being more

factors (q) than treatment combinations or observations (n) and no
repeats of any treatment combination so there must be an external
error estimate. The effects of certain factors must be assumed
predominant and all interactions negligible. These designs look
similar to the Plackett and Burman (1946) designs for the 2^
fractional factorials. In fact two of the systematic supersaturated
designs described by Booth and Cox (1962) are regular Plackett and
Burman designs.

While the Plackett and Burman designs are obtained by

systematically cycling the first member of the basic complete
factorial, Booth and Cox acquire their two-leveled designs (when
n < q) by using an even number of factors (q) and have a computer
try all possible treatment combinations.To obtain an admissible
design the factors are arrayed as column headings and the complete
factorial of treatment combinations can be listed under the headings
(one treatment combination per row) as input. The admissible
designs require that the number of pluses be equal to the number of
minuses in each column so that each effect can be estimated (not
necessarily uncorrelated). Hence there are 1/2 n negative ones
and 1/2 n positiveones in each column.

If the designwere orthogonal all of the sums of cross

products of all the pairs of columns would be zero, but since there
are fewer treatment combinations (n rows) than factors (q columns)
there cannot be complete orthogonality and some of the sums of
cross products must not be zero. To obtain the best design Booth
and Cox require that the set of treatment combinations (the design)
must provide a minimum value of the maximum absolute value of the
sum of cross products of all the pairs of columns and if there is
more than one design meeting this requirement select that design in
which the number of pairs of columns attaining this minimum value
for the maximum absolute value of the sum of cross products is a
minimum.
13.1 FIXED DESIGNS 331

An example given by Booth and Cox (1962) for q = 16, n = 12,

follows:

Factors (q)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

1 + + + + + + + + + + + - - - - -

2 + - + + +

3 - + .+ + - - - + - - + + + - + +

4 + + + - - - + - - + - - + + + +
g
5 + + - - - + - - + - + + + - + -
cd
C
•H + + + + + +
6 + - - - - - + - + -
1
CJ
+ + + +
7 - - - + - - + - + + + -
P
c
<D
a
p
8 - - + - - + - + + + - + - + - +
cd
a) + + +
U 9 - + - - + - + + + - - + -
H
10 + - - + - + + + - - - - + + - -

11 - - + - + + + - - - + - - - + +

12 _ + + + + _ _ _ + _ _ _ _ -

Any factor goes in the experiment (-) as often as (+). This

is similar to balance in the random balance design. Notice that
there are six pluses and six minuses in each column. The sum of
products between column 1 and column 2 is zero, but between column
15 and column 16 is 4. If all possible pairs of columns are
investigated 67 pairs have zero and the other 53 pairs have 4 for
the absolute value of the sum of cross products.

Booth and Cox (1962) provide designs for the following:

332 13. RESPONSE SURFACE EXPLORATION

i i* Number of pairs
|s|
Designs q n
of columns = (^)
0 2 4 6 8

I (Plackett-Burman) 16 12 67 53 120
II 20 12 75 115 190
III 24 12 101 175 276
IV 24 18 193 83 276
V 30 18 281 154 435
VI 36 18 385 245 630
VII (Plackett-Burman) 30 24 295 81 59 435

*|s| = absolute value of the sum of cross products.

The analyses of the data from an experiment may be to find the

optimum treatment combination, i.e., the one that produces the
maximum response, or to look at the effects of the factors. As in
the random balance design, it is difficult to interpret a regression
analysis because of the confounding even of main effects (since
q > n).

Booth and Cox (1962) compare the systematic supersaturated

designs given above with random balance designs by looking at the
variance of the measure of nonorthogonality. This variance is
obtained from the frequency of the sum of cross products given in
the previous table. An example using the design q = 16, n = 12
follows:
The probability is 67/120 that |s| = 0 and 53/120 that |s| = 4.

Booth and Cox (1962) indicate that since the number of -4!s and +4*s
are equally likely over the set of designs from which this particular
design was evolved, the E(s) = 0 and the variance of s is
E(s^) = (53/120)(4^) + (67/120)(0^). Hence the variance for this
design is 7.07.

Further, for the random balance design the experimenter is

2
dealing with finite sampling and the variance of s is [n /(n-1)] or
2
for this case the variance is (12 /ll) = 13.1. A table summarizing
comparisons of the two designs follows:
13.1 FIXED DESIGNS 333

Systematic
Design q n Random balance supersatured

I 16 12 13.1 7.07
II 20 12 13.1 9.68
III 24 12 13.1 10.1
IV 24 18 19.1 13.6
V 30 18 19.1 15. 3
VI 36 18 19.1 17.4
VII 30 24 25.0 11.4

This design includes only 2-leveled factors and is not as

practical as the random balance design which allows factors with
varying levels such as 3 , 2 , 7 , 6 , 4 .

13.1.3 Random Method

Brooks (1958) describes a procedure for seeking optimum

combinations of levels of factors or of independent variables with
maximum yield of the dependent variable. The procedure is very
similar to the random balance design discussed in 13.1.1 except
that the treatment combinations are chosen completely at random.
The random method described by Brooks allows the experimenter to
choose the sample size, n (number of treatment combinations in the
experiment), so that he has probability, S, of obtaining at least
one treatment combination which will fall in the optimum subregion.
This optimum subregion is that proportion, a, of the factor space
(all possible treatment combinations from which n is a random
sample) which produces the highest responses.

The probability that at least one of the n trials falls in the

optimum subregion is S = 1 - (1 - a)n,

where: (1 - a) = the probability that a given trial will

fall outside the optimum subregion.

Solving for n (for certain values of a and S) Brooks obtains the

data in Table 13.1.1.
334 13. RESPONSE SURFACE EXPLORATION

TABLE 13.1.1

Number of Observations (n)

S
a
0.80 0.90 0.95 0.99

0.100 16 22 29 44
0.050 32 45 59 90
0.025 64 91 119 182
0.010 161 230 299 459
0.005 322 460 598 919

If the experimenter requires an estimate of experimental error he

must take some extra trials at the same points.

The procedure for running the experiment is to set up the

factor space or population of factors and levels of each, then
select the n combinations of the levels of the factors, that is a
treatment combination at random and conduct the experiment at these
n points. The analysis is merely to select the treatment or factor
combination that gave maximum response and call these conditions
optimum. Of course, one could examine the response surface by a
regression analysis as described for the random balance design.

The characteristics of this design are:

(1) The number of trials does not depend on the number of

factors or levels of each factor.

(2) The number of trials does not depend on the function

of the surface.

(3) The design is fixed, not sequential.

Brooks explains a stratified random design in which the

factors and/or levels of factors are stratified before the treatment
combinations are drawn. Then there is only one treatment
combination drawn at random from each stratum.
13.1 FIXED DESIGNS 335

13.1.4 Mixture Designs

In experimentation dealing with mixtures of ingredients (as in

many chemical-type experiments), the sum of the levels of the
factors for each treatment combination adds to a constant. For
example, in manufacturing a marker flare the ingredients magnesium,
sodium nitrate, strontium nitrate, and a binder make up 100% of
the ingredients, while in stainless steel many times the amount of
iron is fixed so that the sum of the variable ingredients such as
carbon, chromium, and others is 100% minus the percentage of the
iron.

The design of these mixture experiments is most unusual because

the experimenter must restrict the treatment combinations to sum to
a constant. This causes an entirely different design approach from
the usual factorial procedure.

(1) Simplex Lattice. A mathematical definition of a simplex

is a convex hull of any setof n + 1 points from a Euclidean n
space (En) which do not lieon a hyperplane in En.

A comprehensive description of the simplex lattice design is

given by Scheffe (1958). When the simplex lattice design is used
in mixture problems, the responses are measured at the simplex
lattice composition points. Polynomial equations having a special
correspondence with the simplex lattice points arethen used to
represent the response surface. This polynomial equation is quite
easily determined, the response surfaces can then be drawn and the
optimum proportion of components may be determined.

The following tabulation, taken from an article by Gorman and

Hinman (1962), relates the number of response points to the number
of factors and the number of spacings per factor:
336 13. RESPONSE SURFACE EXPLORATION

Quadratic Special Cubic Cubic Quartic

Number of spacings (m) -► 2 2 3 4

Number of factors (q) Number of response points (k)

3 6 7 10 15
4 10 14 20 35
5 15 25 35 70
6 21 41 56 126
8 36 92 120 330
10 55 175 220 715

In general,

, _ (m + q - 1) !
K " m! Cq - 1) I

where q is the number of factors and m is an integer related to the

spacing of points between 0 and 1. There are (m + 1) equally
spaced points from 0 to 1. Hence if m = 3, the responses will be
measured at the proportions 0, 1/3, 2/3, and 1 for each component.
This would be a simplex cubic lattice. The above formula does not
hold for the special simplex cubic lattice which is a simplex
quadratic lattice with additional center points on the two-
dimensional face (s). The number of response points in the special
simplex cubic lattices is given by:

k = + q ( q - 1) (g - 2)
2 6

The special simplex cubic lattice is sometimes referred to as the

augmented quadratic.

Graphical aids facilitate the calculations associated with

simplex lattice designs and as an aid to see where the responses
are measured and what simplex lattice designs look like, the figure
below shows typical 3-factor designs.
13.1 FIXED DESIGNS 337

Quadratic Sp e cial Cubic

yi yi

y2 y23 y3
338 13. RESPONSE SURFACE EXPLORATION

The polynomials associated with each of these 3-component designs

are given below.

Quadratic: y = + Y2X2 + Y3X3 + Y12X1X2 + Y13X1X3 + Y23X2X3

Special Cubic: y - Y1X1 + r2x2 + Y3 X3 + Y12X1X2 + Y13X1X3 + Y23X2X3

+ Y123X1X2X3
Cubic: y = YjXj^ + Y2x2 + Y3X3 + Xi2XlX2 + Y13XjX3 + Y23X2X3

♦ n 12x 1x 2 ( x 1- x 2) ♦ n 13x 1x 3 ( x 1- x 3) + n 23x 2x 3 ( x 2- x 3)

+ Y 123X 1X 2X 3

Quartic: y =Y ^ + Y2X2 + Y3X3 + Y ^ X ^ - Y ^ X ^ + Y23X2X3

+ n 12x 1x 2 (X 1- x 2) + n 13x 1x 3 cxl - x 3) ♦ n 23x 2x 3 ( x 2- x 3)

. * W s < V X/ * 52 3 W X 2 - V

+ V 1123X 1X2X3 * V 1223X 1X2X3 * V 1233X 1X 2X3

where the coefficients are special functions of the usual regression

terms. The usual second-order polynomial with three independent
variables is

y = 3Q + 31X1 + 62X2 + 83X3 + 812X1X2 + 3 ^ ^ + 323X2X3 + 3-^Xj

+ e 22x 2 + e33x 3

where y is the dependent variable and the X^, X2, X3 are the
independent variables. It is required that X1 + X2 + X3 = 1 in the
simplex lattice design. We can form the following four identities:
13.1 FIXED DESIGNS 339

2
Substituting these identities into the general equation for 30, X-,
2 2
X2, and X^, we will get, after simplification,

y = (6 0 ♦ Si ♦ 6 n )X1 ♦ (B0 + e2 ♦ e22)X2 ♦ (6 0 * 63 ♦ B3 3 )X3

+ ^B12 " B11 ' B22^X1X2 + ^B13 ' B11 ' B33^X1X3

+ ^B23 " B22 “ B33^X2X3

Letting
>-
II

B0 + Bl + B11
I—1
>-
II
CM

B0 + B2 + B22

Y3 = B0 + B3 B33
Y12 = B12* Bll‘ B22

Yi13
z = B13 ■M
7 ' i 1 - ^33
B11 B,

and

Y23 = B23 ‘ B22 ‘ B33

it follows that the quadratic polynomial for the simplex lattice

design with three factors is

y = y1X1 + y2X2 + 1r3X3 + y12XlX2 + Y13X1X3 (13.1.1)

' T23* 2*3

Problem 13.1.1. Let and use

X1 = 1 * X2 X3

y = 60 + 62x2 + 63X3 + 623X2X3 + * SSJX3

to show that this is the same equation as Eq.

If the design of the experiment is

-2L xi X2 X3
i 0 0
yi
0 1 0
y2
0 0 1
y3
1/2 1/2 0
y12
1/2 0 1/2
y13
y23 0 1/2 1/2

(The y's correspond to the y's in the diagram for quadratic.)

340 13. RESPONSE SURFACE EXPLORATION

the least square estimates are

= yv Y2 = y 2 , Y3 = y 3, Y12 = 4y12 - 2yx - 2y2

Y13 = 4y13 - 2yl - 2y3 and Y23 = 4y23 ’ 2y2 ' 2y3

Using these estimates of the parameters (y's) in the prediction

equation for three independent variables we obtain the predicted
value (y) as

y = Xl(2Xl - * X2(2X2 - l)y2 ♦ X3(2X3 - l)y3

+ 4XxX2y12 + 4XxX3y13 + 4X2X3y23

th
where X^(i = 1, 2, 3) is the value of the i independent variable
to be used as a predictor of y; y\ or y\^ (i < j = 1, 2, 3) is the
mean of the observations at the designated treatment combination
(i or ij). It follows that the variance of the estimate is

2 r i [xi(2xi-i)]2 2 t4xixi)2i
v(y)=a2 I * I — ^ — \
Li=l i (i<j)=l ij J
2
where a = error variance from a previous experiment or estimated
from the within treatment combinations: r. or r..
l ij
(i < j = 1, 2, 3) = the number of observations at the
designated treatment combination (i or ij).

To test the adequacy of the quadratic model, the center point

can estimated from the model of the simplex lattice and
observed with the point added to the design and the t-test used if
an error variance is available.

After the experimenter obtains the response surface equation,

he can examine the contours of the surface. Gorman and Hinman
(1962, pp. 474-479) discuss a procedure for doing this. They have
examined the surfaces, by the signs and sizes of the interaction
yfs for the three factor quadratic model as follows:
13.1 FIXED DESIGNS 341

Case Y12 Y13 Y23

A 0 0 0
B + + +
C + 0
D + + 0
E + +

where 0 indicates linearity; (+) indicates synergism; (-) indicates

antagonism.

This specific type design in which each factor makes up 100%

of one of the treatment combination is easily handled mathematically,
but for many experiments it is impossible to have any treatment
combination made up completely of one factor. Scheffe (1958, p. 354)
discusses a modification that allows one factor to have an upper
bound less than 1. In the analysis he uses a ,!pseudocomponent,f
(coding of the original variables) which allows the use of a
regression equation in terms of the coded variables.

(2) Simplex Centroid. To provide experimental coverage of

the response surface in the center of the planes and hyperplanes,
Scheffe (1963) introduced the simplex centroid design. One way to
visualize the design is to consider the structure of a three factor
experiment, next a four factor experiment, etc. The four factor
simplex centroid has one center point in each if its four faces
(sides) of the design plus one in the center of the tetrahedron.
This procedure is extended to any q-factor simplex centroid design.

The center points (centroids for a regular q-factor simplex

lattice) are located in the design by finding the mean level for
all factors involved. For example, the center point for a three
factor simplex or any face of a q-factor simplex is found to be
(1/3, 1/3, 1/3), while the center point of a four factor simplex or
tetrahedron is (1/4, 1/4, 1/4, 1/4) and so on.

The analysis of a simplex centroid is similar to that of a

simplex lattice since the restriction that the sum of the levels
342 13. RESPONSE SURFACE EXPLORATION

of the factors

i - i 1

must be used in the regression analysis for the simplex centroid

also. If an experimenter is interested in optimizing the levels of
the factors, he will probably use a second degree model. In the
case that the simplex lattice design allowed only an incomplete
model, the simplex centroid may allow a complete second degree
regression model because of the additional center points.

(3) Extreme Vertices. In most experimenter problems dealing

with mixtures, some portion of all ingredients are necessary for
the experimental unit to work. For example, in a certain type
flare some of each ingredient: magnesium, sodium nitrate,
strontium nitrate, and binder must be present for the flare to
function correctly. In addition, many experimenters will know the
lower and upper bounds of each ingredient that they wish to examine
before the experiment is designed.

The extreme vertices design (McLean and Anderson, 1966) takes

into account certain aspects of the simplex centroid design and
allows upper and lower bounds on every factor that are inside the
usual simplex boundaries. If one considers q-factors (ingredients)
represented by a proportion of the total mixture, then

q
I X. = 1,
i=l
and
0 < a. < X. < b. < 1
— i — l— l—

where i = 1, ..., q and a^ and b., are constraints on the X^ imposed

by the experimenter.

Once a. and b. are determined for all q-factors the extreme

i i n
vertices experiment can be designed as follows:

Write down all possible two-level treatment combinations using

13.1 FIXED DESIGNS 343

the a. and b. levels for all but one factor which is left blank,
1 l
e«g»> (a!> b2, a3* * as* b^), for a six factor experiment. This

procedure generates possible treatment combinations with one

factor!s level blank in each.

Then, go through all possible treatment combinationsand

fill in those blanks that are admissible, i.e., that level
(necessarily falling within the constraints of the missing factor)
which will make the sum of the levels for that treatment combination
equal to 1. Each of the admissible treatment combinations is a
vertex; however, some vertices may appear more than once.

The hyperpolyhedron so constructed contains a variety of centroids.

There is one located in each bounding 2-dimensional face,
3-dimensional face, ..., r-dimensional face (r _< q - 2), and the
centroid of the hyperpolyhedron. The latter point being the
treatment combination obtained by averaging all the factor levels
of the existing vertices. The centroids of the 2-dimensional faces
are found by isolating all sets of vertices for which each (q - 3)
factor levels remains constant within a given set and by averaging
the factor levels for each of the three remaining factors. All
remaining centroids are found in a similar fashion using all
vertices which have (q - r - 1) factor levels constant within a set
for an r-dimensional face where 3 ^ r _ < q - 2. It should be noted
that under the assumptions given above the dimensionality of the
hyperpolyhedron formed by the extreme vertices will always be q - 1.

Example 13.1; In manufacturing one particular type of flare

the chemical constituents are magnesium (x^), sodium nitrate (x2),
strontium nitrate (x^), and binder (x^). Engineering experience
has indicated that the following constraints on a proportion by
weight basis should be utilized:

0.40 ^ xx _< 0.60

O.iO < x2 < 0.50
0.10 < x3 < 0.50
0.03 —< x.
4 —< 0.08
344 13. RESPONSE SURFACE EXPLORATION

The problem is to find the treatment combination (x^, *4 )

which gives maximum illumination as measured in candles.

The vertices of the polyhedron consisting of all the admissible

points of the factor space are found by applying the above procedure.
See the following tabulation which indicates eight admissible vertices
and six faces. These eight treatment combinations are shown in
Fig. 13.1.1.

Treatment Treatment
>ination combination
X1 X2 X3 X4 X1 X2 X3 X4

0.40 0.10 0.10 CD 0.40 0.10 0.47 0.03

0.40 0.10 0.50 (2 ) 0.40 0.10 0.42 0.08

0.40 0.50 0.10 0.40 0.50 0.03

0. 40 0.50 0.50 0.40 0.50 0.08

0.60 0.10 0.10 (3) 0.60 0.10 0.27 0.03

0.60 0.10 0.50 (4) 0.60 0.10 0.22 ,0.08

0.60 0.50 0.10 0.60 0.50 0.03

0.60 0.50 0.50 0.60 0.50 0.08

(5) 0.40 0.47 0.10 0.03 0.10 0.10 0.03

(6 ) 0.40 0.42 0.10 0.08 0.10 0.10 0.08

0.40 0.50 0.03 0.10 0.50 0.03

0.40 0.50 0.08 0.10 0.50 0.08

(7) 0.60 0.27 0.10 0.03 0.50 0.10 0.03

(8 ) 0.60 0.22 0.10 0.08 0.50 0.10 0.08

0.60 0.50 0.03 0.50 0.50 0.03

0.60 0.50 0.08 0.50 0.50 0.08

13.1 FIXED DESIGNS 345

Fig. 13:1.1. Extreme vertices for flare experiment.

In order to complete the design, one must determine the six

centroids for each and the centroid for the polyhedron. To do this
we must isolate all existing faces. These are made up of at least
3 vertices which have 4-3 or 1 factor level which is constant. For
example, vertices (1), (3), (5), and (7) all have the binder (x^)
level equal to 0.03. All possible faces and their associated
centroids are as follows:
346 13. RESPONSE SURFACE EXPLORATION

Treatment Treatment combination

combination Centroid which form the face
(9) (0.50, 0.1000, 0.3450, 0.055) (1), (2), (3), (4)
(10) (0.50, 0.3450, 0.1000, 0.055) (5), (6), (7), (8)
(ID (0.40, 0.2725, 0.2725, 0.055) (1), (2), (5), (6)
(12) (0.60, 0.1725, 0.1725, 0.055) (3), (4), (7), (8)
(13) (0.50, 0.2350, 0.2350, 0.030) (1), (3), (5), (7)
(14) (0.50, 0.2100, 0.2100, 0.080) (2), (4), (6), (8)

and the final centroid of the polyhedron, of course,, comes; from the
average of all eight treatment combinations and is

(15) (0.50, 0.2225, 0.2225, 0.055)

Fifteen flares assembled at each of the above treatment

combinations produced, respectively, the following amounts of
illumination (measured in 1000 candles):

(1) 75 (6) 230 (ID 190

(2) 180 (7) 220 (12) 310
(3) 195 (8) 350 (13) 260
(4) 300 (9) 220 (14) 410
(5) 145 (10) 260 (15) 425

Standard least squares techniques were used on the above data

to obtain the complete quadratic model.

y = -l,558Xj - 2,351x2 - 2,426x^ + 14,372x^ + 8,300Xj,x 2 + 8,076x^x3

- 6,625x1x4 + 3,213x2x3 - 16,998x2x4 - 17,127x3x4

2
The squared multiple correlation coefficient (R ) for this
model is 0.9833, with only five degrees of freedom for residual.
2
If only Xj, x y x i x 2> xix3 > xlx4* X2X3 were u s e &> t*ie R would
be' 0.9829, with eight degrees of freedom. Since all four variables
still appeared in the latter model, the authors decided to retain
the full model. The reader should recognize that, as in any model
13.1 FIXED DESIGNS 347

development problem, one must have stopping rules for evolving

models from data. The purpose of this example, however, is merely
to demonstrate the use of the regression model to determine the
optimum treatment combination not to elaborate on model development
per se.

In order to consider all the necessary constraints, a quadratic

programming routine was used to calculate the optimum treatment
combination
(0.5233, 0.2299, 0.1668, 0.0800)

which is the desired solution to the problem. The predicted value

of y for this optimum point is 397.48. It should be noted that
this procedure only guarantees an optimum in the case where the
response surface is a concave function.

It is quite feasible that one would like to further verify the

initial estimate of the optimum condition. This could be done by
applying another extreme vertices design to a localized region
containing this initial point.

Snee and Marquardt (1974) developed an algorithm for selecting

a set of extreme vertices for linear models which tend to minimize
the average variance of the estimated regression coefficients.
This article also includes an extended list of references on the
subject of mixture designs. Snee (1975) also discusses construction
of efficient designs for constrained mixture systems in which the
response can be described by a quadratic model.

Problem 13.1.2. An experiment on steel alloys was needed to

find the optimum combination of carbon, nickel, chromium,
manganese and iron to maximize the strength of the alloy
(keeping other desirable properties).
Using the extreme vertices designs, design experiments
and show approach for analyses for the following conditions.
Low High
Carbon 0.0004 0.0010
Nickel 0.0800 0.1200
Chromium 0.1200 0.2000
Manganese 0.0050 0.0200
keep Iron at 0.7000 (fixed).
348 13. RESPONSE SURFACE EXPLORATION

(b)Use same levels of carbon through manganese but use

Low High
Iron 0.65 0.75
(Indicate the difficulty in handling this part but do not try
to solve for the centroids).

13.1.5 Rotatable Designs

The rotatable designs require the usual assumptions for

regression and the sum of the levels of the factors for any
treatment combination does not necessarily add to 1. The treatment
combinations are located equidistant from the originally estimated
maximum point, designated by (0, 0, 0, ..., 0), so that the
standard error of the estimate of y is the same for all points on
the circumference of the design. Box and Hunter (1957) have done
extensive work on these designs and Cochran and Cox (1957) have
given many of these designs.

(1) First-Order Design, Two Factors X^, X^ and No Interaction

(Nonorthogonal).

The simplest experiment possible to obtain information on both

main effects requires three observations. The rotatable design for
such an experiment may be pictured as:

where points (1), (2), and (3) make up the design of the experiment,
p = distance from the center point (0, 0) in standardized units
which the experimenter believes will give information on
the response surface; 8 = the angle from one point,
treatment combination to another. (In this case 0 = 360°/3
= 120°), and the resulting design is an equilaterial triangle
13.1 FIXED DESIGNS 349

The regression equation forthis experiment is

y = e0 + eiXl + 62X2
where the usual assumptions apply to the nonorthogonal regression
analysis except that there is no estimate of error. To obtain an
error estimate, at least one of the three treatment combinations
must be repeated in the experiment. Usually forthis type of
experiment previous knowledge ofthe error is available.

(2) Second-Order Design, Two Factors with Interaction

(Nonorthogonal)

In this rotatable design there must be a minimum of six

treatment combinations one located at the center and the arrangement
to give equal distance from the center point is a regular pentagon.
In this case 0 = 360°/5 = 72° between each point and the distance
again can be p, the length specified by the experimenter.

The regression equation for this experiment is

y = 60 ♦ ♦ e2X2 ♦ 3n x2 + ^ ♦ B^X,

with no error estimate from the data unless there are repeated
observations for at least one of the treatment combinations. Box
and Wilson (1951, pp. 8 and 9), compare pentagonal and factorial.
There is little to choose between the two. Quadratic effects are
estimated more accurately with pentagonal and interactions less
accurately.

(3) General Second-Order Two Factor (Nonorthogonal)

Cochran and Cox (1957, p. 346) recommend that to obtain a

general second-order model one may choose n^ points in the center
(0, 0) and n ^ points on the circumference of the circle at distance
p, and 0 = 360°/n2 apart such that the standard errors of estimate
of y are approximately the same. One can locate the treatment
combinations by using the following:
350 13. RESPONSE SURFACE EXPLORATION

Treatment
combination X^ X2

1 p COS 6 p sin 0
2 p COS 20 p sin 20
•
.
•

CD
p COS sin n 2 0

c
p

CM
n2

n2+l 0 0

0 0
V ni

With three or more factors, regular figures can be obtained

with additional observations at the center, Box and Hunter (1957).

13.1.6 Fractional Factorials (Orthogonal)

As discussed in previous chapters, fractional factorials allow

estimation of certain effects and interactions if other interactions
are negligible. Great care is taken to assure the experimenter of
exactly which interactions are assumed negligible before the
experiment is even laid out. Hence the analysis of variance is
clearly understood.

The important feature in using fractional factorial type

experiments for finding maximum points or regions is the extraction
of the regression model from the data. Consider an example of a
3
1/2 replicate of a 2 factorial experiment in which all factors are
considered to have only -1 and +1 levels. There are only four
treatment combinations, the two factor interactions and the three
factor interaction are assumed negligible. This concept can be
expressed through the aliases
13.1 FIXED DESIGNS 351

y = I = Z lZ 2 Z 3

Z1 = Z2Z3

Z2 = Z1Z3

Z3 " Z1Z2

where Z 2^3 s ^*2Z3 = Z1Z3 ~ Z1 Z 2 =

Then it follows that the regression equation (with no error

estimate from the experiment) is:

y = a0 * OjZj ♦ a2 Z2 ♦

As in previous regression equations dealing with orthogonal

experiments it is quite easy to transform back to

y = e0 + »!*! + 62X 2 + B3x3

using the actual levels of the factors X^, X2 and X^.

Continuing the same example, the design can generally be

described as

Z1 Z2 Z3
-1 -1 -1
Same as I = -ABC
1 1 -1
from fractional replicated
1 -1 1 factorial experiments
- 1 1 1

If the levels of X^ are really 2 and 14, of X2 are 15 and 23

and of X^ are 5 and 27, the design could be written as:

^2 ^3
2 15 5
14 23 5
14 15 27
2 23 27
352 13. RESPONSE SURFACE EXPLORATION

13.1.7 Complete Factorials (Orthogonal)

The complete factorials (as discussed previously) allow

estimation of all the effects and interactions with no confounding.
Of course one can estimate only linear effects and interactions if
there are only two levels for each factor and can estimate up to
the quadratic effects and corresponding interactions if there are
three levels of each factor and so on. This concept was discussed
thoroughly in previous chapters.

Here, as in the section on fractional factorials immediately

preceding this section, the important feature in using factorial
experiments to investigate the maximum points or regions is to
obtain a regression model to describe the response surface. Using
3
the 2 complete factorial here, the design may be given by

z
h fs h h
-1 -l -1 2 15 5
-1 -l 1 2 15 27
l -1 2 23 5
-1 l 1 2 23 27
or
1 -l -1 14 15 5
1 -l 1 14 15 27
1 l -1 14 23 5
1 l 1 14 23 27

The corresponding regression equations (with no estimate of

error) are:
13.1 FIXED DESIGNS 353

Of course if the three factor interaction is assumed zero, one can

estimate the error from that component.

It should be noticed that the two-leveled complete factorial

experiment allows estimates of interactions, but not of the squared
terms. The three-leveled factorials would allow the estimates of
squared terms, but too many observations are required for these
estimates relative to number of observations required by the
composite design (to be discussed in the next section).

13.1.8 Composite Design

The composite design Box and Wilson, (1951); Myers, (1971, Ch. 7;
and Davies, (1971, Ch. 11) has three parts: a basic two-leveled
factorial or fractional factorial, an extra point at the center of
the entire design and 2k (where k = number of factors) extra points,
one at either extreme of each factor and at the center of all other
factors. Hence in a composite design with a complete factorial
experiment in it there are 2 + 2k + 1 treatment combinations.
Correspondingly, if there was a fractional factorial instead of a
k
complete factorial experiment in the design, the 2 would be reduced
as required. The particular type of composite design depends on the
location of the extreme points. If the extreme points are located
at the same standardized distance from the center point as the
factorial points, the design is called a rotatable composite design
(sometimes the word “central" is included in the title of these
designs to indicate that there is a center point). If the extremes
are located at a distance that makes the squared terms in the model
orthogonal to each other, the design is called an orthogonal composite
design. Any other location of the extreme points may be used, but
the analysis is just a nonorthogonal regression analysis.

The advantage of a composite design over the fractional or

complete three-leveled factorial is in the reduction of the number
of treatment combinations required to estimate the squared terms in
a second-order model. This idea may be summarized in the following
tabulation :
354 13. RESPONSE SURFACE EXPLORATION

Number of Treatment Combinations

Number of Three-leveled
v
factors k factorial, 3 Composite, 2 + 2k + 1
2 9 9
3 27 15
4 81 25
5 243 43
5 (1/3 fractional) 81 (1/2 fractional) 27
6 729 77
6 (1/3 fractional) 243 (1/2 fractional) 45

As the number of factors increase in an experiment the savings

in number of treatment combinations using the composite design
instead of three factor factorial increases rapidly.

Two disadvantages in using the composite design instead of the

three leveled factorial are (1 ) estimating effects with unequal
variances (to be shown later), and (2 ) having fewer degrees of
freedom for error. The model for a composite design is:

B0 * B1 X1 * B 2X2 * ••• * BkXk * B 1 2 X 1 X2 * ••• *

B (k-l)kXk - lXk * 611X1 ♦ B 22X2 * - • * »kk^ * ‘

Consider k = 3, then the composite model is:

y ■ e„ . CjXj . e 2* 2 . 6 jX 3 . 6 3 3 X 3 X 3 . 6 3 3 X 3X 3 . 6 3 3 X 3X 3

♦ 633 X 3 . 6 3 3 X 3 * 6 3 3 X 3 . k

where the error has 5 degrees of freedom, and the corresponding

model for the three-leveled factorial is
13.1 FIXED DESIGNS 355

1 2 3 W 3 + 61123X1X2X3

1233X1X2X3 +

612233X1X2X3

In a second-order model, the terms from ®n2XlX2 tlirou8*1

61199--X.X9X- are assumed zero and could be put into the error
estimates. The factorial model would then allow 17 degrees of
freedom for error.

In general for optimum designs, the five degrees of freedom in

the composite design for the error estimate is adequate and the
composite design is preferred over the three-leveled factorial. In
the following two subsections two types of composite designs,
rotatable and orthogonal, are described.

(1) Rotatable Composite (Partly Orthogonal)

The rotatable composite design has all points equidistant from

the center point. The basic factorial terms are orthogonal to one
another and the squared terms are orthogonal to the linear and
interaction terms but not to each other. In other words the star
points, the 2k points used to estimate the squared terms in the
model, do not form an orthogonal design. As a result one may run
straight multiple regression analysis on all the data or run an
analysis of variance on the factorial part before running the
regression analysis. The analysis of variance on the factorial
part allows the experimenter to examine the linear main effects and
the linear by linear interaction parts before examining the overall
model.
356 13. RESPONSE SURFACE EXPLORATION

Here, as in other rotatable designs, the variance of the

estimates of y are approximately the same, but the variance of the
estimates of the regression coefficients may be quite different.
Of course these latter variances can be obtained by the usual least
squares method.

(2) Orthogonal Composite

This design requires that the distance from the center point
to each of the star points be such that all effects and interactions
estimated in the second order model are orthogonal to one another.
This distance, d, is

1/2 1/4
[(F + T) - F1/2]2
4n

where F = the number of treatment combination in the factorial run

(fractional or full); and T = the number of additional
points multiplied by the number of observations per
treatment combination (n).

In the three factor case; F = 8, T = 7 if there is only one

observation for each treatment combination (n = 1). If follows that

1/4
[(8 ♦ 7)
1/2
- 81 / 2 ] 2 1.215
■i

An example using the three factor case follows:

The design or the 15 treatment combinations for a three factor

orthogonal composite design is:
13.1 FIXED DESIGNS 357

Factor (Independent variable)

z. Treatment comt
Z2 Z3
1 -1 -1 -1 (1)
2 1 -1 -1 zi
3 -1 1 -1 Z2
4 1 1 -1 Z1Z2
5 -1 -1 1
Z3
6 1 1
Z1Z3
7 -1 1 1
Z2Z3
8 1 1 1
Z1Z2Z3
9 0 0 0
10 1.215 0 0
dl
11 -1.215 0 0
-dl
12 0 1.215 0
d2
13 0 -1.215 0
-d2
14 0 0 1.215
d3
15 0 0 -1.215
*d3

The analysis in model form is:

Y - “ oZ0 + a i z i + a2Z2 + “llZll + °22Z22 + a12ZlZ2 + a3Z3

+ a33Z33 + “l3ZlZ3 + a23Z2Z3 + a123ZlZ2Z3 + e

There are four degrees of freedom for error in this design,

unless is zero; then there are five degrees of freedom. It
is impossible to estimate (linear x quadratic) interactions
separately because Z^Z^ = Z^Zjj (note table below). The layout of
the effects and interactions is given in Table 13.1.2.
 TABLE 13.1.2

Treatment Combination

Effect 1 2 3 4 5 6 7 8 9 10 r 11 12 13 14 15

z1z2z3

*Note equality and hence are completely confounded.

13.1 FIXED DESIGNS 359

The levels of each factor (-1.215, -1, 0, 1, 1.215) are not

equally spaced and the usual orthogonal polynomials do not work
directly; however, if in this special case the quadratic variable

then the analysis is orthogonal. The same procedure is used for

Z22 and Z33*

All effects and interactions are orthogonal and are estimated
2
in the usual manner a = Z Z y /lZ . The regression coefficients in
the model are not correlated, and the terms are additive. As can
be seen by the table of coefficients, however, the effects and
2
interaction coefficients (a's) have different variances because Z
are different for certain coefficients. The variance of a is

V(a)

2
where a is the error variance, and Z's are as described in the
layout. As examples consider the variance of a^:

a2 n2
v(« ) = - 2 j = ----- 2 ------
ZZX 8 + 2(1.215)

and the variance of a^:

2 2
XTf > _ a________________a
^alP 2 2 2 2
ZZn 8(. 27) + 5(-. 73) + 2(.746)Z

whereas in a factorial experiment the variances of the a's are all

constant.

If an experimenter is interested in investigating only

second-degree polynomial surfaces in a fixed experiment the
recommendation is that he consider the two-level factorial plus
additional points to obtain a composite design rather than a
360 13. RESPONSE SURFACE EXPLORATION

complete or fractional three-level factorial. The important reason

for the recommendation is that the reduction in the number of
treatment combinations is quite large, especially when the number
of factors is four or higher. In the composite design, however, the
variances of the coefficients for effects and interactions are not
all equal and should always be considered before deciding on the
exact design of the experiment.

Problem 13.1.3. Suppose an experimenter wanted to estimate

the parameters in the equation

y = 60 + 61X1 + 62X2 + ei2XlX2 + h A + B22X2 + BU 2 X1X2

and wanted to use a design that would have the fewest number
of observations, could be nonorthogonal and have the standard
errors of estimate as near equal as possible.
Set up the design specifying the location of each
experimental unit relative to coordinates (0, 0) of (X^, X2 ).

Problem 13.1.4. In a weapons evaluation section of a large

corporation, an experiment was needed to investigate factors
affecting firing time of a weapon. The factors and their
lower and upper limits of the levels were:
(1) Primary igniter 5 to 15 mg
(2) Primary initiator 5 to 15 mg
(3) Packing pressure 12,000 to 28,000 psi
(4) Amount of igniter 80 to 120 mg
(5) Delay column pressure 20,000 to 40,000 psi
Eventually an equation of the response surface would be desired
in which interactions of linear x linear and squared terms may
be important. The number of observations should be held to a
minimum and all effects from a third degree term and higher
are considered negligible.
Design an orthogonal experiment showing the levels of each
factor to be used and explain why you used this particular
design.

Problem 13.1.5. Show the variance of a^, and in a

3-factor orthogonal composite design.
13.2 SEQUENTIAL DESIGNS 361

13.2 SEQUENTIAL DESIGNS

13.2.1 Univariate Method or One-Factor-at-a-Time

The one factor at a time method of finding an optimum

treatment combination has been used for many years. If there are
only two independent variables, the experimenter selects one
variable to hold constant and decides in which direction to go
after the first observation plus the amount of the jump to make for
the other variable. The first observation is taken at the ’’guessed
at” maximum. If the response on the second observation was larger
than the first, take the third observation in the same direction
with the same increment jump. If the response on the second
observation is smaller than the first, go in the opposite direction
from the first with the same increment. If the response on the
third observation is larger than the second continue in that
direction; but if smaller, go back to the point for which the
response was the greatest and follow a similar procedure in a
perpendicular direction by keeping the variable constant that was
being varied and change the variable that was held constant. Repeat
this procedure until no appreciable increase in yield results and
call the maximum response point the optimum for those independent
variables.

Brooks (1959) describes a slightly different procedure called

’’Univariate,” but the method is nearly the same. Using Brooks*
example in which 16 trials are allotted for the experiment, take
four trials equally spaced around the ’’guessed at” maximum in one
direction, say X2* Then a curve

y = 60 + 6 2X2 + e 22X 2 + B222X2

is fitted and the point in X2 that maximized this expression along

the line of the experiment is used as the fixed point to try four
more trials in the X^ (perpendicular to X2) direction. This
procedure of fitting is repeated for X^, and four more trials are
taken in the X2 direction in a similar manner except that the
362 13. RESPONSE SURFACE EXPLORATION

increments are decreased between points as described by Brooks

(1958). The final four trials are made in the other direction and
the maximum response for the final four gives the optimum factor
combination. Of course one could fit a polynomial over the surface
examined, but the factor space may not be covered adequately to
provide a very good overall response surface for the limits of X^
and X2 used in the experiment.

Brooks (1958) gives a numerical example of the univariate

method in his paper.

13.2.2 Simplex (Sequential)

The sequential application of simplex designs presented here

do not require that the sum of the levels of the factors add to one
for each treatment combination. The number of factors, n, then, is
equal to the dimension of the design. For the two factor experiment
the regular simplex design is an equilateral triangle in two
dimensions. It is not necessary for this design to be regular but
the scaling can be done such that the same unit is used for all
dimensions.

To portray the procedure we will use a two dimensional design

centered around a ’’guessed at” maximum and call this point the
centroid (0, 0). The adjacent points in our design will be one
unit apart. The layout for this design is as follows.

where only the three extreme points are used in the design.
13.2 SEQUENTIAL DESIGNS 363

The general n-dimensional design has (n + 1) points and can be

arrayed as a (n + 1) by (n) matrix Sf:

Point or Independent Variable (X)

treatment
combination

1 1
2/3 2/6 2/l0 /2n(n+l)

1 1 __1_
2/3 2 /6 2/10 /2n(n+l)

2 1
2/3 2/6 2/10 /2n(n+l)

3
2/6 2/fi ) /2n(n+l)

ij

(n + 1)
/2n(n+l)

Notice that the two-dimensional design iii the following

tabulation as given previously in the upper left-hand comers o tS fc

Independent variable

Point
364 13. RESPONSE SURFACE EXPLORATION

Next we must show how one moves sequentially to the maximum

yield or response. Let us assume that the point [0, 2/(2/3)]
gives the minimum yield of the three points in the two-dimensional
design. The sequential method is to delete that point and form
another simplex with the remaining two points and a point on the
opposite side. In this case the new point would be
[0, -4/(2/3)]. The algebraic method for finding the new point is
to delete the minimum point from the array giving the remaining
array as follows.

Independent variable

Point
1 . JL
2 2/3

1 . JL
2 2/3

The next step is to add the levels of each factor and multiply each
by 2/n (where n = the number of X's) or 2/2 = 1 for this case, and
subtract the level of the factor deleted. For variable 1 we have

| c- i + = °

and subtract 0 giving 0. For variable 2

| (. J —) = - —
2/3 2/3 2/3

and subtract 2/(2/3) which gives - 4/(2/3).

Using the X^ notation of the (n + 1) by n matrix given

previously to denote the level of the j***1 variable of the i ^
treatment combination, we can find the new level X?\ by using

X*. = [- {X.. + X_. + ... + X. , . + X. , . + . . . + X . - X..]

lj Ln lj 2j l-l,j 1+1,1 n+1,j 11J

The procedure to find the optimum region of the treatment

combinations is to continue to delete the point or treatment
13.2 SEQUENTIAL DESIGNS 365

combination that has the lowest response in the simplex, replacing

this point by a point opposite to it as described above. This
procedure is followed until an optimum region is located. The
optimum region is indicated by several new points, say n, failing
to produce a response that is greater than that obtained at a
previous treatment combination. At this point the optimum
treatment combination should be repeated before experimentation is
stopped in order to assure that this point is truly optimum.

If in the process of stepping through the factor space with

successive simplexes, the yield of a new point is less than any
other point in the current simplex, do not go back to the point that
was previously vacated. In this case determine the next lowest
point and move opposite it. This makes up a different simplex which
is the basis for the next decision.

This concludes thedesign discussion. The analysis, in

addition to findingthe optimum point, may include running a
regression analysis on those points near the optimum.

An example of the procedure is the following electrical

engineering problem in which the response y is output power and the
inputs are (X^) voltage, (X^) relative humidity, and (X^)
temperature. The present system for X^ is 110 V, is 43%, X^ is
86°F with y = 58%. Hence the point

(0, 0, 0) = (110, 43, 86)

has a yield of 58. For X^, X2, and X^ the experimenter felt that
increments of 10, 2, and 0.5, respectively, were equally important.
Using these increments to code about the centroid, we obtain the
four treatment combinations

"l. 105, 42.4, 85.8*

2. 115, 42.4, 85.8
3. 110, 44.2, 85.8
4. 110, 43.0, 86.6m

from the following matrix:

366 13. RESPONSE SURFACE EXPLORATION

1 _ JL_
’ 2/6

1 _ JL_
2/J " 2/6
2 _ _ l_
2/3 2/6

3
o,
2/6

The yields for the four treatment combinations were 52, 62, 61, 57,
respectively. Hence delete treatment combination 1 and replace it
by the three levels:

Level 1 = | (115 + 110 + 110) - 105 = 118.3

Level 2 = | (42.4 + 44.2 + 43.0) - 42.4 = 44.0

Level 3 = | (85.8 + 85.8 + 86.6) - 85.8 = 86.3

This new treatment combination (118.3, 44.0, 86.3), along with the
remaining three treatment combinations, forms the new simplex.
After running the new treatment combination, the sequential
procedure continues as discussed previously.

The efficiency of sequential designs is measured by the number

of points required to reach the maximum. For a given surface, one
can determine the rate of advance to the maximum. Spendley, Hext,
and Himsworth (1962) show that as the standard deviation (variation
of the responses for the same treatment combination) increases the
rate of advance decreases as {1/(standard deviation)}. For example,
the rate of advance in an experiment with standard deviation of 3
is twice that of an experiment with standard deviation of 6.

On comparing achievement of sequential procedures to reach

maximum, Spendley, Hext, and Himsworth (1962) indicate that the
simplex design is second only to the steepest ascent (regression)
13.2 SEQUENTIAL DESIGNS 367

procedure if there is no error and somewhat poorer (but not too bad)
when error is present.

Box and Behnken (1960) describe a "simplex-sum" design, a type

of second order design, that combines the vectors defining the
points of the initial simplex in pairs, threes and so on. One can
use the midpoints, or points off the midpoints, of the vectors
joining the vertices to estimate the curvature. The concept of
using intermediate points for estimation of curvature was given in
Section 13.1.4.

Problem 13.2.1. Consider the electrical engineering example

in Section 13.2.2 in which y is output power and the inputs
are (Xp voltage, (X2) relative humidity, and (X^) temperature.
If the new treatment combination gave a yield of 67, the next
one gave a yield of 69, and the next 65, then 62, 68, 69, and
finally 69.
(a) Show the new treatment combinations for each new
simplex.
(b) What is the treatment combination that gives
approximately the maximum yield?
(c) Run a nonorthogonal regression analysis on these
data, using the points you believe are applicable. Explain
your model.

13.2.3 Computer Aided Designs

Given a set of conditions for a factorial type experiment with

the lower and upper bounds on all factors established by the
experimenter, the problem is to find out the specific order to run
the possible treatment combination such that the most information
for that number of sequential observations may be obtained over the
factor space whenever the experimenter decides to stop the
experiment. "Most information" in this case usually refers to
minimum variance of the estimates of the parameters in the response
surface equation over the defined factor space for the data
obtained.

In this case Kennard and Stone (1969) use quantitative levels

and linear associations to find the design points. This allows
368 13. RESPONSE SURFACE EXPLORATION

them to compute distances (actually squared distance is used) between

points in the factor space using the computer. The technique is to
choose two points that are farthest apart (these may not be unique so a
criterion of choosing the smallest index is used for uniqueness) as
the first two treatment combinations to be used in the design.

To obtain the next best (in the sense of giving most information
using the distance criterion) point from the remaining (n - 2)
points, one must find the point that is farthest from the first two.
This point may not be unique either (as for the first two points),
and the same criterion of taking the point with the smallest index
is utilized.

This process of choosing succeeding points that are farthest

from the ones already in the design is continued until the
experimenter decides to stop or all points in the factor space are
used. In general this design attempts to spread the observations
evenly over the factor space with emphasis on using outermost
points first and come into the center part next. Then go out from
the center but not necessarily to the edge of the factor space and
so on.

The basic idea that the observations may be taken at these

points, an analysis run and a decision made as to whether or not
more observations are needed allows much flexibility in this
sequential type design.

Example of the Kennard and Stone Procedure

2
Consider a 5 equally spaced factorial. The points are
numbered for the 25 possible combinations of X^ and as follows:

-2 -1 0 1 2
+2 1 2 3 4 5
+1 6 7 8 9 10
0 11 12 13 14 15
-1 16 17 18 19 20
-2 21 22 23 24 25
13.2 SEQUENTIAL DESIGNS 369

The pair of points (1, 25) and the pair (5, 21) are tied for
being farthest apart. Since the index on the pair (1, 25) is
smaller, namely 1, this pair is used in the design.

Now there are 23 points (n - 2) remaining. The squared

distances to each of these points from 1 and 25 are computed. For
2
example to point 3 from point 1 the squared distance is (-2-0) or
4 and to 3 from 25 the squared distance using the Pythagorean
theorem is (2-0)2 + (-2-2)2 = 20.

The same type of calculations are made for all 23 points and
it is found that 5 and 21 are tied for being farthest from design
points 1 and 25. Choosing the smaller index point, 5, the third
design point is 5.

Continuing the process, design point four is 21 and design

point five is 13. This is a basic "EVOP" design to be discussed
later for a two factor design. That is, it is a two by two
factorial with the points at the extreme vertices and one point in
the center.
2
Continued application of the procedure gives a 3 for the first
9 design points since design point six is 3, design point seven is
11, design point eight is 15, and design point nine is 23.

One can continue this until all 25 points are in the design.
This concludes the demonstration of the design.

It should be clear by now that this procedure allows a

sequential design and a stopping point may be decided upon when the
experimenter feels he has enough information. He may use a
regression analysis at any time he has enough points to estimate
the parameters of interest and has enough information on the
experimental error.

This procedure is most flexible in that it is programmed to

handle many factor factorials with varying levels, orthogonal and
nonorthogonal, taking into account design points that are
experimentally impossible to run and accounting for points already
run.
370 13. RESPONSE SURFACE EXPLORATION

Additional procedures for similar sequential type designs are

discussed by Dykstra (1971) and Hebble and Mitchell (1972).

13.2.4 Steepest Ascent Method

The general mathematical concepts of utilizing regression

techniques to examine yield surfaces has been known for many years.
Hotelling (1941) discussed the problem using experimental procedures,
but it was not until after the Box and Wilson (1951) paper that
experimenters fully appreciated the usefulness of this tool. Other
references are Myers (1971, Chapter 5), and John (1971, Chapter 10).

The general principle that is followed in the steepest ascent

method is to establish an initial design, say a 2n factorial, with
a center point at some f,guessed at” maximum in the factor space.
A regression analysis is performed on the data of this initial
design in order to determine which direction, if any, one should
move from the center point of the design in order to increase the
yield in the most efficient manner, i.e., along the path of
steepest ascent. One usually follows along this path with equally
spaced observations until a decline or leveling off occurs in the
response variable. This region may then be used for a new initial
design and the process repeated, if necessary. Another alternative
at this point is to construct a second-order design, take the
observations for this design, analyze the data using a regression
analysis and finally use calculus techniques to determine an
optimum treatment combination.

A numerical example given by Davies (1971, pp. 511-518) is*

reproduced here for illustrative purposes. The experiment was a
laboratory study of a chemical reaction of the type

A + B + C — ► D + other products

where the reaction was taking place in the presence of solvent E.

Object: To maximize the yield of D for a given amount of A

(kept constant in the experiment), the most expensive material.
13.2 SEQUENTIAL DESIGNS 371

Given: About 45% of the theoretical yield (the yield which

would be obtained if all of A were changed to D), but it was
thought that a 75% yield was possible. The experimental error was
expected to be 1% or less.

The five factors and their levels were as follows:

Factor level
-1 +1

Amount of E, cc. 200 250

V
k2, Amount of C, mole/mole of A 4.0 4.5
Concentration of C, % 90 93
h*
Time of reaction, hour 1 2
Amount of B, mole/mole of A 3.0 3.5

The center (00000) of the design was chosen in the best-known

operating region. Since there seemed to be large possible gains,
and interactions would probably not be as important as main effects,
a quarter replicate was carried out assuming all two factor
interactions except one, were negligible. The design allowed
for this as is indicated by the following identity:

I = ZjZ^Z, = -z2z3zs = -ZlZ4Z5

The block containing the following treatment combinations was

used, and the corresponding yields obtained are as follows:
Treatment Yield
Z. z4
h. ^3 ^5 combination (x)
1. -1 -1 -1 (1) 34.4
2. -1 -1 1 1 1 cde 51.6
3. -1 1 1 1 bde 31.2
4. 1 1 -1 be 45.1
5. 1 -1 1 -1 ad 54.1
6. 1 -1 1 -1 1 ace 62.4
7. 1 1 -1 -1 1 abe 50.2
8. 1 1 1 1 -1 abed 58.6
372 13. RESPONSE SURFACE EXPLORATION

The following estimates

obtained using
ZZy
a =~2
ZZ

-*■ + (+an + a22 + a33 + ot^ + a5$) 48.5

ao ao
-> + 7.9
“l al (“a45^
-> + -2.2
“2 a2 (-*35>
+ 6.0
“3 a3 ^’a25^
-► + 0.4
®4 a4 f-al5^
->■ * -0.4
‘“5 a5 (+ai4 + a23}
->■ + -1.8
a13 al3 f+°24)
-+■ + (+a34) 0.2
“12 a12

where -*■ means estimates. Hence the model considered is:

y = 48.5 + 7.9Z.- 2.2Z0 + 6.0Z- + 0.4Z. + 0.4ZC - 1.8Z.Z- + 0.2Z.Z-

J 1 2 3 4 5 13 12
The approximate standard error of estimate of each a is

assuming a = 1.

In general, the steepest ascent procedure is easy to calculate

and visualize when only first order effects are used. Hence the
ci^ term will be dropped out of the calculations even though it is
significant. Note that this term was not as large as three of the
first order effects (further support for omitting it). Ordinarily,
the first order effects will be sufficient when one is relatively
far from the optimum treatment combination. As the experimenter
comes closer to optimum and second order terms become more important,
he will need to change the method of surface exploration.

Coming back to the example on steepest ascent methodology, the

calculation of the direction in which maximum gain would be expected
is given in the following table:
13.2 SEQUENTIAL DESIGNS 373

Calculation of Path of Steepest Ascent and

Subsequent Trials on the Path*

xi X2 X3 X4 xs

cc mole % hour mole

(1) Base level 225 4.25 91.5 1.5 3.25
(2) Unit 25 0.25 1.5 0.5 0.25
(3) Estimated slope a
(Change in Yield Per Unit) 7.9 -2.2 6.0 0.4 +0.4
(4) Unit x a [(2) x (3)] 197.5 -0.55 9.0 0.2 0.1
(5) Change in level per 10 cc
change in X^ 10 -0.028 0.456 0.010 0.005
(6) Path of steepest 225 4.25 91.5 1.5 3.25
ascent represented 235 4.22 92.0 1.5 3.25
by a series of possible 245 4.19 92.4 1.5 3.26
trials on it 255 4.17 92.9 1.5 3.26
265 4.14 93.3 1.5 3.27
Trial (9) 275 4.11 93.8 1.6 3.27
285 4.08 94.2 1.6 3.28
Trial (10) 295 4.06 94.7 1.6 3.28
305 4.03 95.1 1.6 3.29

*Data from Design and Analysis Industrial Experiments, edited

by Owen L. Davies, published by Oliver and Boyd, Hafner, New York,
1971, Table 11-22, p. 513.

In this calculation line (1) shows the base level for the
factors, line (2) shows the unit, i.e., the change in level for the
X-variable corresponding to the change from 0 to 1 for the
Z-variable, and line (3) shows the estimated slopes calculated on
the assumption that three and higher factor interactions are
negligible. The levels of the factors are then changed in proportion
to these slopes. For every 7.9 units increase in Z^, should be
decreased 2.2 units, Z^ increased 6.0 units, and so on. Hence, for
each 25 x 7.9 cc. that is changed, X2 should be changed by (.25)
x (-2.2) mole. These changes are given on line (4) and define the
374 13. RESPONSE SURFACE EXPLORATION

direction of steepest ascent at the point given by the base levels

in line (1). The path leading from the origin in this direction
can be obtained by taking a convenient increment in one variable
(10 cc in xp, and calculating the proportionate changes to be made
in the other variables. These are given in line (5). These
quantities may then be successively added to the base level to give
path (6).

At this point we deviate slightly from the example given in

Davies (1971). We recommend that the next observation be taken
along the path of steepest ascent slightly outside the limits of
the original design, say at (255, 4.17, 92.9, 1.5, 3.26). The next
observation would be taken one or two steps further along the path.
This would continue until no further increase in yield is observed.

Now, additional first order designs may be implemented to

arrive in the region of the maximum yield. Let us assume that this
region has been reached. The experimenter will then want to explore
the region near the maximum using a second order design in order to
investigate the curvature in the response surface.

The main weakness of the steepest ascent method is that it is

not invariant to scale changes (see Box and Draper, 1969, p. 166);
and Myers, 1971, p. 93).

13.2.5 Canonical Analysis

After the experimenter is satisfied that he has located the

region near the maximum response, he will usually want to explore
this region thoroughly before recommending the exact treatment
combination to be used in practice. Even if he knew the exact
optimum treatment combination, he will want to know the shape of
the response surface near that point so that he has the option of
picking alternative ’’near" optimum treatment combinations that may
be less expensive to use with little loss in yield.
13.2 SEQUENTIAL DESIGNS 375

In general, the design to explore the near maximum response

region should be second order to allow for the investigation of
curvature. Assume that a second order design has been used, and a
quadratic regression equation evolved from the data. To interpret
this response surface using the regression equation alone is
sometimes most difficult. If one obtains the canonical form of the
equation, however, interpretation of the results is much easier.

The following example gives the reader an approach to canonical

analysis so that he may use the concept in practice. If the
regression equation for two variables is

y = 6 + 3 Xj + 5 X2 - 4 Xj - 3 Xj - 2 Xj X2 (13.2.1)

one may find the contour for y = 8 by setting Eq. (13.2.1) equal to
8 and substituting a value for X^. This makes the equation a
quadratic in X2 which allows one to solve for two points. Next,
change the value of X^ and solve for two new X2 points. Continue
this until one obtains the contour of 8 and the picture shown in
Fig. 13.2.1.

Fig. 13.2.1. Surface contours.

376 13. RESPONSE SURFACE EXPLORATION

To find the canonical form of Eq. (13.2.1) differentiate with

respect to X^ and X2. Then set the equations equal to zero to find
the maximum point

f£ - . 3 - 8 h - 2 X2 ■ 0

F jq " 5 ' 2 xi ' 6 x2 ■ o

2 17
Solve and obtain the point of maximum, (X^ = — , X2 = -jj ), and
refer to it as M. Substituting these solutions into Eq. (13.2.1),
we obtain the maximum y^ = 8.2.

To determine the coefficients in the canonical form set up the

determinant from Eq. (13.2.1)

bu - » 1 /2 b,2 -4-A -1
r<

-1 -3-A
1
CM
CM

1/2 b12

and the characteristic equation becomes

X2 + 7X + 11 = 0.

The characteristic roots of this equation are A =-4.62 and -2.38.

The generalform of the canonical equation is

BU *i * B22^2 <1 3 - 2 - 2 >

where and B22 are the characteristic roots of the characteristic

equation and are the coefficients of a new set of axes having the
optimum treatment combination as their origin. The exact relationship
between the old and new coordinate systems may be found by a
procedure given in Myers (1971, p. 83).

In this example the canonical equation becomes

y - 8.2 = -4.62^2 - 2.38^2 (13.2.3)

where the and axes are labeled by observing that the contours
drop off faster in the direction than in the SC^ direction.
13.2 SEQUENTIAL DESIGNS 377

The interpretation of Eq. (13.2.2) is that the change in yield

on moving away from M to some point is given by the
right-hand side of Eq. (13.2.2). In the example, Eq. (13.2.3),
the coefficients B ^ and B22 are both negative. Hence there is
always a loss in yield whichever way one goes from M. This
concludes the example.

If one found in his canonical analysis that B ^ was negative

and B22 = 0, there would be no unique center, and

y - yM =

This is a stationary ridge.

An example:

If

y - yM = B11& 1 * B2 & 2

where B ^ is negative; B2 is positive and measures the rate of

increase in yield along the axis, and M is on the axis at
infinity, one has a rising ridge. An example:
378 13. RESPONSE SURFACE EXPLORATION

If B ^ is negative and B^ is positive, the contours are

hyperbolas. In the example, we have a saddle as
13.2 SEQUENTIAL DESIGNS 379

13.2.6 Evolutionary Operation (EVOP)

In the previous chapters and sections of this chapter various

designs have been presented and primarily intended for use by
experimenters associated with research laboratories, experimental
or developmental organizations. In this section, a design is
presented which is intended for use on a production line. This
concept is fairly new, having first been exposed in a paper by Box
(1957). This subject was later elaborated on by Box and Draper
(1969).

This method of experimentation is appealing to production

personnel because the basic idea is to select treatment combinations
in a relatively narrow range of each factor under consideration.
This type of selection of the factor levels has a tendency not to
upset the equilibrium of the production process but whenever the
complete design is replicated a sufficient number of times the
design will indicate conditions which will increase productivity.

Since each factor is constrained to a relatively narrow band

large differences in the response variable are not normally
expected among treatment combinations. For this reason a 2n design
with a center point is usually employed. A 2n design lends itself
to easy analysis; hence standard forms may be utilized to aid
production people to do the analysis, A form for this use is
illustrated in the following example. The term phase on this form
is used to record the design location in use, i.e., whenever the
data from the first phase indicates that the design location should
be shifted, then a new phase of the experiment is implemented.
Individual replications of the experiment in any one phase is
referred to as a cycle.

In the example that follows, three cycles are illustrated and

the method of calculation is that presented by Box and Hunter (1959)
in Technometrics. It should be pointed out that a more efficient
analysis may be carried out using the straight forward ANOVA
techniques presented earlier in this text.
380 13. RESPONSE SURFACE EXPLORATION

Example 13.2: The data from three cycles of an EVOP

application in a steel mill will be analyzed. A method for
determining the location of the second phase will be shown. The
location of design for Phase 1 and the data obtained from these
cycles is as follows:

50-

Q>
E

40- •2

_L_ _L
1220 1260
Temperature

Operating condition 1 2 3 4 5

1 27.0 28.6 25.1 26.0 26.8

Cycle 2 26.2 29.1 24.0 26.4 27.8
3 26.5 29.2 24.2 25.6 27.0

The form for Cycle 1 has very little filled in since there was
no previous information and no estimate of the standard deviation
can be made with only one set of observations. The calculations
shown for Cycle 2 in Table 13.2.2 shows the form much more complete.
All data shown under the heading of "Previous” have been transferred
from the form used for the calculations of Cycle 1. The value of
the range in line (iv) is calculated by finding the range of the
five differences shown on that same line. In order to find the
New s use the value of f^ n given in Table 13.2.4. Note that the
two sigma error limits For New Effects is 0.89 which is smaller than
the temperature and time effects indicating that there is
13.2 SEQUENTIAL DESIGNS 381

opportunity for improving the location of the design. This fact is

also supported by the relatively small change in mean effect, 0.10
as compared to the error limit of 0.79. Note that a significant
change in mean effect would indicate curvature in the response
surface over the region of design of the phase under consideration.
In order to display the method the remainder of the results of
Cycle 3 are given in Table 13.2.3. Again, notice that the
temperature and time effects are significant indicating that a new
phase should be implemented.

To locate a new set of operating conditions for Ph&se 2 one

can apply the calculational procedure given in the steepest ascent
method, Section 13.2.4. To do this denote the factor of temperature
and time by and X^, respectively; and

; = b l from Section 13.2.4, but in this example

* EX.y re i
^ = ~2 (temperature effect).

Using the data of Cycle 3 then, the calculations are the following:

= (temperature effect) == -1.45

* (time effect) = -0.85

X1 X2
Base level 1240 45
Unit 20 5
Estimated slope -1.45 -0.85
Unit x slope -29.0 -4.25
Amount of change per 20° -20.0 -2.93
 TABLE 13.2.1 382

EVOP Calculation Form

Cycle: n = 1 Phase: 1
Time 1
9 . Response: Tensile strength (1000 lb) Date: 5/20 - 5/22
Temp 4
Calculation of averages Calculation of standard deviation s
Operating Conditions (1) (2) (3) (4) (5)

(i) Previous cycle sum Previous sum s s

(ii) Previous cycle average Previous average s =
(iii) New observations 27.0 28.6 25.1 26.0 26.8 New s * range x f. =
K,n
(iv) Differences [(ii)-(iii)] Range of (iv)
13.

(v) New sums [(i)+(iii)] 27.0 28.6 25.1 26.0 26.8 New sum s =
(vi) New average y^[(v)/n] 27.0 28.6 25.1 26.0 26.8 New average s = (New sum s)/(n-l) =

Calculation of effects Calculation of error limits

Temperature effect = -j (yj+y^-y^- = -2.15

*5>
For new averages 2 s//n =
Time effect =j (y3+y5-y2-y4) as -1.35
For new effects 2 s//n =
Temp x time interaction = i f^+y^-^- = 0.45
For change in mean 1.78 s//n =
Changes in mean effect = y (y^+y2+^3+^4+^5"^ = -0.30
RESPONSE SURFACE EXPLORATION
 TABLE 13.2.2
13.2
EVOP Calculation Form

Cycle: n = 2 Phase: 1
Time l
9 . Response: Tensile strength (1000 lb) Date: 5/23 - 5/25
Temp 4
Calculation of averages Calculation of standard deviation s
Operating Conditions CD (2) (3) (4) (5)
SEQUENTIAL DESIGNS

(i) Previous cycle sum 27.0 28.6 25.1 26.0 26.8 Previous sum s =
(ii) Previous cycle average 27.0 28.6 25.1 26.0 26.8 Previous average s =
(iii) New observations 26.2 29.1 24.0 26.4 27.8 New s = range x f^ n = 2.1 x .30 = 0.63
5*,2
(iv) Differences [(ii)-(iii)] 0.8 -0.5 1.1 -0.4 -1.0 Range of (iv) * 1.1 -■ (-1.0) = 2.1
(v) New sums [(i)+(iii)] 53.2 57.7 49.1 52.4 54.6 New sum s = 0.63
(vi) New averages y^[(v)/n] 26.6 28.8 24.6 26.2 27.3 New average s = (New sum s)/(n-l)= 0.63

Calculation of effects Calculation of error limits

Temperature effect = ^ 3 +^4- 9 2- ? 5) a -2.65

For new averages 2 s / J n = 0.89
Time effect =j (y ^ s
rW * ) = -1.55
For new effects 2 s//n = 0.89
Temp x time interaction = ( ^ 2^ 1 = -0.05
rh -h "*
For change in mean 1.78 s//n = 0.79
383

Changes in mean effect =j ^ s '^ = 0.10

 TABLE 13.2.3 384

EVOP Calculation Form

Cycle: n = 3 Phase: 1
Time 1
^ A Response: Tensile strength (1000 lb) Date: 5/27 - 5/29
1 Temp 4
Calculation of averages Calculation of standard deviation s
Operating conditions (1) (2) (3) (4) (5)

(i) Previous cycle sum 53.2 57.7 49.1 52.4 54.6 Previous sum s =0.63
(ii) Previous cycle average 26.6 28.8 24.6 26.2 27.3 Previous average s =0.63
(iii) New observations 26.5 29.2 24.2 25.6 27.0 New s = range x f^ = 0.35
(iv) Differences [(ii)-(iii)] 0.1 -0.4 0.4 0.6 0.3 Range of (iv) = 1.0
13.

(v) New sums [(i)+(iii)] 79.7 86.9 73.3 78.0 81.6 New sum s =0.98
(vi) New averages y^[(v)/n] 26.6 29.0 24.4 26.0 27.2 New average s = (New sum s)/(n-l) = 0.49

Calculation of effects Calculation of error limits

Temperature effect = -j C 9^*9^-92'• y j -2.90

For new averages 2 s//n = 0.56
Time effect = -j (,9^*95- 9 2 ■y4i = -1.70
For new effects 2 s/Jn = 0.56
Temp x time interaction = -j s 0.10
For change in mean 1.78 s//n = 0.50
Changes in mean effect = -jr (y^+^2+^3+^4+^5~5 y,)- 0.04
RESPONSE SURFACE EXPLORATION
13.2 SEQUENTIAL DESIGNS 385

Thus the new center point for Phase 2 could be

Temperature 1240 - 20 = 1220

Time 45 - 2.93 * 42

Some of the mathematical rationale behind the method of

calculation is given in Hicks (1973, Chapter 16) and Box (1957).
The value of f, is shown to be
k, n

where d^ is the quality control symbol representing the factor that

must be divided into the range of a set of observations in order to
estimate the standard deviation. Values of f^ are given in
Table 13.2.4.

TABLE 13.2.4

Values of f,
k,n
n
2 3 4 5 6 7 8 9 10
k
4 0.34 0.40 0.42 0.43 0.44 0.45 0.45 0.46 0.46
5 0.30 0.35 0.37 0.38 0.39 0.40 0.40 0.40 0.41
6 0. 28 0.32 0.34 0.35 0.36 0.37 0.37 0.37 0.37
7 0.26 0.30 0.32 0.33 0.34 0.34 0.35 0.35 0.35
8 0.25 0.29 0.30 0.31 0.32 0.32 0.33 0.33 0.33
9 0.24 0.27 0.29 0.30 0.31 0.31 0.31 0.32 0.32
10 0.23 0.26 0.28 0.29 0.30 0.30 0.30 0.31 0.31
386 13. RESPONSE SURFACE EXPLORATION

13.3 REFERENCES

Booth, K. H. V. and Cox, D. R. Technometrics 4:489 (1962).

Box, G. E. P. Applied Statistics 6:81 (1957).
Box, G. E. P. and Behnken, D. W. Ann. Math. Stat. 31:838 (1960).
Box, G. E. P. and Draper, N. R. Evolutionary Operation, Wiley,
New York, 1969.
Box, G. E. P. and Hunter, J. S. Ann. Math. Stat. 28:195 (1957).
Box, G. E. P. and Hunter, J. S. Technometrics 1:77 (1959).
Box, G. E. P. and Wilson, K. B. J. Royal Stat. Soc. B, 13:1 (1951).
Brooks, S. H. Operations Research 6:244 (1958).
Brooks, S. H. Operations Research 7:430 (1959).
Cochran, W. G. and Cox, G. M. Experimental Designs, 2nd ed., Wiley,
New York, 1957.
Davies, 0. L. (Ed.), Design and Analysis of Industrial Experiments,
2nd ed., Oliver and Boyd,Edinburgh, 1971, Chapter 11.
Dykstra, 0. Technometrics 13:682 (1971).
Gorman, J. W.and Hinman, J. E. Technometrics 4:463 (1962).
Hebble, T. L. and Mitchell, T. J. Technometrics 14:767 (1972).
Hicks, C. R. Fundamental Concepts in Design of Experiments, 2nd ed.,
Holt, Rinehart and Winston, New York, 1973, Chapter 16.
Hotelling, H. Ann. Math. Stat. 12:20 (1941).
John, P. W. M. Statistical Design and Analysis of Experiments,
Macmillan, New York, 1971.
Kennard, R. W. and Stone, L. A. Technometrics 11:137 (1969).
McLean, R. A. and Anderson, V. L. Technometrics 8:447 (1966).
Myers, R. Response Surface Methodology, Allyn and Bacon, Boston,
Massachusetts, 1971.
Plackett, R. L. and Burman, J. P. Biometrika 33:303 (1946).
Satterthwaite, F. E. f,New Developments in Experimental Design."
Talk given in Boston, (1956).
Satterthwaite, F. E. Technometrics 1:111 (1959).
Spendley, W., Hext, G. R., and Himsworth, F. R. Technometrics 4:441
(1962).
Scheffe, H. J. Roy. Stat. Soc. Series B, 20:344 (1958).
Scheffe, H. <J. Roy. Stat. Soc. Series B, 25:235 (1963).
13.3 REFERENCES 387

Snee, R. D. Technometrics 17:149 (1975)

Snee, R. D. and Marquardt, D. W. Technometrics 16:399 (1974).
Stapleton, J. H. American Statistician, 13:4 October, 1959.
 A p pen d ix 1

Random Numbers

0 0 -0 4 0 5 -0 9 1 0 -1 4 1 5 -1 9 2 0 -2 4 2 5 -2 9 3 0-3 4 3 5 -3 9 4 0 -4 4 4 5 -4 9

00 01826 72696 67261 13748 57834 27748 47472 43428 85524 19311
01 70731 12890 90395 45245 71282 15960 02749 86763 80564 02651
02 46616 84522 17249 78172 14197 84272 53226 96719 83462 05628
03 46384 26607 53444 68780 00458 27488 04949 30717 14304 60222
04 58645 90453 83413 15983 42345 27118 29425 42172 58412 31268

05 41874 83916 07454 42647 09410 12882 80001 98932 63277 94255
06 26032 68402 68176 87347 52572 79056 22703 83175 95807 60134
07 56142 74950 98878 36160 28616 77908 04908 63013 30555 51831
08 57893 57918 56290 00612 53884 76761 47934 20050 84977 11053
09 46729 47281 30473 94911 69061 87486 96782 90636 91405 01219

10 14107 46068 15859 99140 29872 70750 35757 25344 84845 95688
11 27713 13273 09015 42262 24580 30925 96900 56246 10613 17230
12 49886 62560 23023 09812 91948 92265 77407 93047 45352 61689
13 86651 47164 79270 45746 73141 67388 08454 68246 39046 57933
14 96133 52026 25837 64698 06911 21730 86390 19749 25859 80005

15 00926 46212 00204 58304 73907 97914 06786 47324 81225 50754
16 89036 39245 74371 87236 92131 38908 22146 15710 36858 35904
17 79592 66982 87984 30503 86953 56490 20414 07662 11122 95517
18 25084 83491 79984 64583 33924 36146 54692 44989 71603 31274
19 90261 79382 24483 75975 14765 49897 75468 66940 82187 68875

20 94937 36303 85514 76537 86864 88600 51843 18255 49660 87326
21 10178 16103 33283 38239 32402 02950 90428 24405 38580 61198
22 05983 63553 93546 66079 18389 57705 86746 62498 71642 60219
23 80832 36191 32042 76015 51331 40592 04887 50771 23810 64650
24 85889 08924 41542 04758 98753 52695 28165 87031 00479 95974

25 77032 15760 83026 92461 51806 65495 32148 40714 49107 05758
26 84080 67254 89239 13272 50218 17737 00242 15203 69060 39047
27 41022 70272 73710 17625 51325 69525 63464 89526 76738 80210
28 72398 03426 36476 62922 43624 98779 63289 34550 94270 01263
29 52224 04363 92979 26520 91076 13849 17740 25964 22169 23110

30 26223 68525 82483 15232 93800 78903 18831 19546 50469 74328
31 57872 36109 88383 83512 58763 57230 35952 75716 57094 53951
32 44529 90455 96666 31804 24979 51863 42983 13367 14111 38541
33 13586 36649 80041 54602 17281 22392 63074 53831 85782 94644
34 49084 08133 45510 21472 14644 38592 04490 52187 33945 78491

35 42309 57412 33314 07820 31345 45074 93547 30023 47632 93222
36 08174 49231 02588 36639 53978 40177 45572 44117 97946 44214
37 58679 57695 94704 16260 41928 27300 14053 93050 87103 34434
38 46847 17986 19347 74125 64945 97496 00465 65830 56564 46884
39 91053 64137 26104 29911 01242 58960 60582 83119 35911 68859

40 98796 24540 75578 79529 63199 78959 60734 73433 91995 81783
41 66824 16251 63616 91787 77519 14592 15333 63525 88097 79340
42 35492 57116 78056 72780 44502 97058 25342 53590 29196 06552
43 53794 57593 27187 99355 12652 77611 95229 41602 99946 94801
44 86332 05856 07786 78241 13584 29243 81609 55078 25937 90051

15 56530 97998 92028 18839 05879 43900 91769 87557 50750 05667
46 22966 67453 46806 57010 49278 99735 84996 74079 77917 48474
47 17068 92733 58021 81102 19871 17442 45375 27328 58055 20323
48 21570 52311 25131 76662 18619 91141 89382 41329 91702 39975
49 09699 95392 78594 40396 83605 01008 48694 85571 64427 23547

388
 A p pen d ix 2

C u m u lative Normal F req u en cy D i s t r ib u t i o n

(A rea un d er th e sta n d a rd norm al cu rv e from 0 t o Z)

z 0.00 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 0 .0 7 0 .0 8 0 .0 9

0 .0 0 .0 0 0 0 0 .0 0 4 0 0 .0 0 8 0 0 .0 1 2 0 0 .0 1 6 0 0 .0 1 9 9 0 .0 2 3 9 0 .0 2 7 9 0 .0 3 1 9 0 .0 3 5 9
0.1 .0 3 9 8 .0438 .0 4 7 8 .05 1 7 .0 5 5 7 .0 5 9 6 .0 6 3 6 .06 7 5 0 .7 1 4 .0 7 5 3
0 .2 .0793 .0 8 3 2 .0871 .0 9 1 0 .0 9 4 8 .0 9 8 7 .1 0 2 6 .1 0 6 4 .1 1 0 3 .1141
0 .3 .1 1 7 9 .1 2 1 7 .1 2 5 5 .1 2 9 3 .1331 .1 3 6 8 .14 0 6 .1 4 4 3 .1 4 8 0 .1 5 1 7
0 .4 .1 5 5 4 .1591 .1628 .16 6 4 .1 7 0 0 .1 7 3 6 .1 7 7 2 .1 8 0 8 .1 8 4 4 .1 8 7 9

0 .5 .19 1 5 .1 9 5 0 .19 8 5 .2 0 1 9 .2 0 5 4 .20 8 8 .21 2 3 .2 1 5 7 .2 1 9 0 .22 2 4

0 .6 .2 2 5 7 .2291 .2 3 2 4 .23 5 7 .2 3 8 9 .2 4 2 2 .2 4 5 4 .2 4 8 6 .2 5 1 7 .2 5 4 9
0 .7 .2 5 8 0 .26 1 1 .2 6 4 2 .2 6 7 3 .2 7 0 4 .2 7 3 4 .2764 .2 7 9 4 .2 8 2 3 .2 8 5 2
0 .8 .2881 .2 9 1 0 .2 9 3 9 .2 9 6 7 .29 9 5 .3 0 2 3 .3051 .30 7 8 .3 1 0 6 .3 1 3 3
0 .9 .3 1 5 9 .3 1 8 6 .3 2 1 2 .32 3 8 .32 6 4 .3 2 8 9 .3315 .3 3 4 0 .33 6 5 .3 3 8 9

1 .0 .3 4 1 3 .3 4 3 8 .34 6 1 . 3485 .3508 .3531 .3 5 5 4 .3 5 7 7 .3 5 9 9 .3621

1 .1 .3 6 4 3 .36 6 5 .3 6 8 6 .3 7 0 8 .3 7 2 9 .3 7 4 9 .3 7 7 0 .3 7 9 0 .3 8 1 0 .3 8 3 0
1 .2 .3 8 4 9 .3 8 6 9 .3 8 8 8 .39 0 7 .39 2 5 .3 9 4 4 .39 6 2 .3 9 8 0 .3 9 9 7 .4015
1 .3 .4 0 3 2 .4 0 4 9 .4 0 6 6 .4 0 8 2 .4 0 9 9 .4 1 1 5 .4131 .4 1 4 7 .4162 .41 7 7
1 .4 .4 1 9 2 .42 0 7 .4 2 2 2 .4 2 3 6 .42 5 1 .4 2 6 5 .4 2 7 9 .4 2 9 2 .4 3 0 6 .4 3 1 9

1 .5 .4332 .43 4 5 .4 3 5 7 .4 3 7 0 .4 3 8 2 .4 3 9 4 .4 4 0 6 .4 4 1 8 .4 4 2 9 .44 4 1

1 .6 .44 5 2 .4 4 6 3 .4 4 7 4 .4484 .44 9 5 .45 0 5 .45 1 5 .4 5 2 5 .4 5 3 5 .4 5 4 5
1 .7 .45 5 4 .4 5 6 4 .4 5 7 3 .4 5 8 2 .4591 .4 5 9 9 .4 6 0 8 .4 6 1 6 .4 6 2 5 .4 6 3 3
1 .8 .46 4 1 .4 6 4 9 .4 6 5 6 .4 6 6 4 .4671 .4 6 7 8 .4 6 8 6 .4 6 9 3 .4 6 9 9 .4 7 0 6
1 .9 .4 7 1 3 .4 7 1 9 .4 7 2 6 .47 3 2 .47 3 8 .47 4 4 .4 7 5 0 .4 7 5 6 .47 6 1 .4 7 6 7

2 .0 .4 7 7 2 .47 7 8 .4 7 8 3 .4 7 8 8 .4 7 9 3 .4 7 9 8 .4 8 0 3 .4 8 0 8 .4 8 1 2 .4 8 1 7
2.1 .4821 .4 8 2 6 .4 8 3 0 .4 8 3 4 .4838 .4 8 4 2 .4 8 4 6 .4 8 5 0 .4 8 5 4 .4 8 5 7
2.2 .48 6 1 .4 8 6 4 .4 8 6 8 .4871 .48 7 5 .4 8 7 8 .48 8 1 .4 8 8 4 .4 8 8 7 .4 8 9 0
2 .3 .4 8 9 3 .4 8 9 6 .4 8 9 8 .4901 .4904 .4 9 0 6 .4 9 0 9 .4911 .49 1 5 .4 9 1 6
2 .4 .4 9 1 8 .4 9 2 0 .4 9 2 2 .49 2 5 .4 9 2 7 .4 9 2 9 .4931 .49 3 2 .49 3 4 .4 9 3 6

2 .5 .4 9 3 8 .4 9 4 0 .4 9 4 1 .4 9 4 3 .4 9 4 5 .4 9 4 6 .4 9 4 8 .4 9 4 9 .4951 .4 9 5 2
2 .6 .4 9 5 3 .4 9 5 5 .4 9 5 6 .49 5 7 .4 9 5 9 .4 9 6 0 .4961 .4 9 6 2 .4 9 6 3 .4 9 6 4
2 .7 .4 9 6 5 .4 9 6 6 .4 9 6 7 .49 6 8 .4 9 6 9 .4 9 7 0 .4971 .4 9 7 2 .4 9 7 3 .4 9 7 4
2.8 .4 9 7 4 .49 7 5 .4 9 7 6 .4 9 7 7 .4 9 7 7 .4 9 7 8 .4 9 7 9 .4 9 7 9 .4 9 8 0 .4981
2 .9 .4981 .49 8 2 .49 8 2 .4 9 8 3 .4 9 8 4 .4 9 8 4 .4 9 8 5 .4 9 8 5 .4 9 8 6 .4 9 8 6

3 .0 .49 8 7 .4 9 8 7 .4 9 8 7 .4 9 8 8 .4988 .4 9 8 9 .4 9 8 9 .49 8 9 .4 9 9 0 .4 9 9 0

3 .1 .4 9 9 0 .4 9 9 1 .4991 .4991 .4992 .4 9 9 2 .4 9 9 2 .4 9 9 2 .4 9 9 3 .4 9 9 3
3 .2 .4 9 9 3 .4 9 9 3 .4 9 9 4 .4 9 9 4 .4 9 9 4 .4994 .4 9 9 4 .4 9 9 5 .4 9 9 5 .4 9 9 5
3 .3 .4 9 9 5 .4 9 9 5 .4995 .4 9 9 6 .4 9 9 6 .4 9 9 6 .4 9 9 6 .4 9 9 6 .4 9 9 6 .49 9 7
3 .4 .4 9 9 7 .49 9 7 .4 9 9 7 .49 9 7 .4 9 9 7 .4997 .4 9 9 7 .4 9 9 7 .4 9 9 7 .49 9 8

3 .6 .4998 .4 9 9 8 .4 9 9 9 .4 9 9 9 .4 9 9 9 .4 9 9 9 .49 9 9 .4 9 9 9 .4 9 9 9 .4 9 9 9
3 .9 .5 0 0 0

389
 A p pen d ix 3

P e r c e n ta g e p o i n t s o f th e t - d i s t r i b u t i o n

Q=0. 4 0 .2 5 0. 1 0 .0 5 0 .0 2 5 0.01 0 .0 0 5 0 .0 0 2 5 0 .0 0 1 0 .0 0 0 5
V
2Q=0. 8 0 .5 0.2 0.1 0 .0 5 0.02 0 .0 1 0 .0 0 5 0.002 0 .0 0 1

1 0 .3 2 5 1 .0 0 0 3 .0 7 8 6 .3 1 4 1 2 .7 0 6 3 1 .8 2 1 6 3 .6 5 7 1 2 7 .3 2 3 1 8 .3 1 6 3 6 .6 2
2 .2 8 9 0 .8 1 6 1 .8 8 6 2 .9 2 0 4 .3 0 3 6 .9 6 5 9 .9 2 5 1 4 .0 8 9 2 2 .3 2 7 3 1 .5 9 8
3 .2 7 7 .7 6 5 1 .6 3 8 2 .3 5 3 3 .1 8 2 4 .5 4 1 5 .8 4 1 7 .4 5 3 1 0 .2 1 4 1 2 .9 2 4
4 .2 7 1 .7 4 1 1 .5 3 3 2 .1 3 2 2 .7 7 6 3 .7 4 7 4 .6 0 4 5 .5 9 8 7 .1 7 3 8 .6 1 0

5 0 .2 6 7 0 .7 2 7 1 .4 7 6 2 .0 1 5 2 .5 7 1 3 .3 6 5 4 .0 3 2 4 .7 7 3 5 .8 9 3 6 .8 6 9
6 .2 6 5 .7 1 8 1 .4 4 0 1 .9 4 3 2 .4 4 7 3 .1 4 3 3 .7 0 7 4 .3 1 7 5 .2 0 8 5 .9 5 9
7 .2 6 3 .711 1 .4 1 5 1 .8 9 5 2 .3 6 5 2 .9 9 8 3 .4 9 9 4 .0 2 9 4 .7 8 5 5 .4 0 8
8 .2 6 2 .7 0 6 1 .3 9 7 1 .8 6 0 2 .3 0 6 2 .8 9 6 3 .3 5 5 3 .8 3 3 4 .5 0 1 5 .0 4 1
9 .2 6 1 .7 0 3 1 .3 8 3 1 .8 3 3 2 .2 6 2 2 .8 2 1 3 .2 5 0 3 .6 9 0 4 .2 9 7 4 .7 8 1

10 0 .2 6 0 0 .7 0 0 1 .3 7 2 1 .8 1 2 2 .2 2 8 2 .7 6 4 3 .1 6 9 3 .5 8 1 4 .1 4 4 4 .5 8 7
11 .2 6 0 .6 9 7 1 .3 6 3 1 .7 9 6 2 .2 0 1 2 .7 1 8 3 .1 0 6 3 .4 9 7 4 .0 2 5 4 .4 3 7
12 .2 5 9 .6 9 5 1 .3 5 6 1 .7 8 2 2 .1 7 9 2 .6 8 1 3 .0 5 5 3 .4 2 8 3 .9 3 0 -4 .3 1 8
13 .2 5 9 .694 1 .3 5 0 1 .7 7 1 2 .1 6 0 2 .6 5 0 3 .0 1 2 3 .3 7 2 3 .8 5 2 4 .2 2 1
14 .2 5 8 .6 9 2 1 .3 4 5 1 .761 2 .1 4 5 2 .6 2 4 2 .9 7 7 3 .3 2 6 3 .7 8 7 4 .1 4 0

15 0 .2 5 8 0 .6 9 1 1 .3 4 1 1 .7 5 3 2 .1 3 1 2 .6 0 2 2 .9 4 7 3 .2 8 6 3 .7 3 3 4 .0 7 3
16 .2 5 8 .6 9 0 1 .3 3 7 1 .7 4 6 2.120 2 .5 8 3 2 .9 2 1 3 .2 5 2 3 .6 8 6 4 .0 1 5
17 .2 5 7 .6 8 9 1 .3 3 3 1 .7 4 0 2.110 2 .5 6 7 2 .8 9 8 3 .2 2 2 3 .6 4 6 3 .9 6 5
18 .2 5 7 .68 8 1 .3 3 0 1 .7 3 4 2.1 0 1 2 .5 5 2 2 .8 7 8 3 .1 9 7 3 .6 1 0 3 .9 2 2
19 .2 5 7 .68 8 1 .3 2 8 1 .7 2 9 2 .0 9 3 2 .5 3 9 2 .8 6 1 3 .1 7 4 3 .5 7 9 3 .8 8 3

20 0 .2 5 7 0 .6 8 7 1 .3 2 5 1 .7 2 5 2 .0 8 6 2 .5 2 8 2 .8 4 5 3 .1 5 3 3 .5 5 2 3 .8 5 0
21 .2 5 7 .68 6 1 .3 2 3 1 .7 2 1 2 .0 8 0 2 .5 1 8 2 .8 3 1 3 .1 3 5 3 .5 2 7 3 .8 1 9
22 .2 5 6 .6 8 6 1 .3 2 1 1 .7 1 7 2 .0 7 4 2 .5 0 8 2 .8 1 9 3 .1 1 9 3 .5 0 5 3 .7 9 2
23 .2 5 6 .6 8 5 1 .3 1 9 1 .7 1 4 2 .0 6 9 2 .5 0 0 2 .8 0 7 3 .1 0 4 3 .4 8 5 3 .7 6 7
24 .2 5 6 .6 8 5 1 .3 1 8 1.7 1 1 2 .0 6 4 2 .4 9 2 2 .7 9 7 3 .0 9 1 3 .4 6 7 3 .7 4 5

25 0 .2 5 6 0 .6 8 4 1 .3 1 6 1 .7 0 8 2 .0 6 0 2 .4 8 5 2 .7 8 7 3 .0 7 8 3 .4 5 0 3 .7 2 5
26 .2 5 6 .6 8 4 1 .3 1 5 1 .7 0 6 2 .0 5 6 2 .4 7 9 2 .7 7 9 3 .0 6 7 3 .4 3 5 3 .7 0 7
27 .2 5 6 .6 8 4 1 .3 1 4 1 .7 0 3 2 .0 5 2 2 .4 7 3 2 .7 7 1 3 .0 5 7 3 .4 2 1 3 .6 9 0
28 .2 5 6 .6 8 3 1 .3 1 3 1 .7 0 1 2 .0 4 8 2 .4 6 7 2 .7 6 3 3 .0 4 7 3 .4 0 8 3 .6 7 4
29 .2 5 6 .6 8 3 1 .3 1 1 1 .6 9 9 2 .0 4 5 2 .4 6 2 2 .7 5 6 3 .0 3 8 3 .3 9 6 3 .6 5 9

30 0 .2 5 6 0 .6 8 3 1 .3 1 0 1 .6 9 7 2 .0 4 2 2 .4 5 7 2 .7 5 0 3 .0 3 0 3 .3 8 5 3 .6 4 6
40 .2 5 5 .6 8 1 1 .3 0 3 1 .6 8 4 2 .0 2 1 2 .4 2 3 2 .7 0 4 2 .9 7 1 3 .3 0 7 3 .5 5 1
60 .2 5 4 .6 7 9 1 .2 9 6 1 .6 7 1 2.000 2 .3 9 0 2 .6 6 0 2 .9 1 5 3 .2 3 2 3 .4 6 0
120 .2 5 4 .6 7 7 1 .2 8 9 1 .6 5 8 1 .9 8 0 2 .3 5 8 2 .6 1 7 2 .8 6 0 3 .1 6 0 3 .3 7 3
OO .2 5 3 .6 7 4 1 .2 8 2 1 .6 4 5 1 .9 6 0 2 .3 2 6 2 .5 7 6 2 .8 0 7 3 .0 9 0 3 .2 9 1

i s t h e u p p e r - t a i l a r e a o f th e d i s t r i b u t i o n f o r v d e g r e e s o f fre e d o m ,
a p p r o p r i a t e f o r u se in a s i n g l e - t a i l te st. F or a t w o - t a i l t e s t , 2Q must be u s e d .

390
 A p p e n d ix 4
Number o f O b s e r v a t io n s p e r S am ple:
U sing t f o r d i f f e r e n c e o f means

S in g le -S id e d T e st a = 0 .0 0 5 a * 0.01 a = 0 .0 2 5 0 .0 5
D o u b le -S id e d T e s t a = 0 .0 1 a = 0 .02 a = 0 .0 5 0. 1

0 .0 1 0 .0 5 0 .1 0 .2 0 .5 0 .0 1 0 .0 5 0 .1 0 .2 0 .5 0 .0 1 0 .0 5 0 .1 0 .2 0 .5 0 .0 1 0 .0 5 0 .1 0 .2 0 .5

0 .3 0 123 87 61 0 .3 0
0 .4 0 85 70 100 50 108 78 35 0 .4 0
0 .5 0 96 55 106 82 45 106 86 64 32 88 70 51 23 0 .5 0
0 .6 0 101 85 67 39 90 74 58 32 104 74 60 45 23 89 61 49 36 16 0 .6 0
0 .7 0 100 75 63 50 29 90 66 55 43 24 76 55 44 34 17 66 45 36 26 12 0 .7 0

0 .7 5 88 66 55 44 26 79 58 48 38 21 67 48 39 29 15 57 40 32 23 11 0 .7 5
0 .8 0 77 58 49 39 23 70 51 43 33 19 59 42 34 26 14 50 35 28 21 10 0 .8 0
0 .8 5 69 51 43 35 21 62 46 38 30 17 52 37 31 23 12 45 31 25 18 9 0 .8 5
0 .9 0 62 46 39 31 19 55 41 34 27 15 47 34 27 21 11 40 28 22 16 8 0 .9 0
0 .9 5 55 42 35 28 17 50 37 31 24 14 42 30 25 19 10 36 25 20 15 7 0 .9 5

1.0 0 50 38 32 26 15 45 33 28 22 13 38 27 23 17 9 33 23 18 14 7 1.00
V alu e o f 1.1 42 32 27 22 13 38 28 23 19 11 32 23 19 14 8 27 19 15 12 6 1 .1
1 .2 36 27 23 18 11 32 24 20 16 9 27 20 16 12 7 23 16 13 10 5 1 .2
D - ia 1 .3 31 23 20 16 10 28 21 17 14 8 23 17 14 11 6 20 14 11 9 5 1 .3
1 .4 27 20 17 14 9 24 18 15 12 8 20 15 12 10 6 17 12 10 8 4 1 .4

1 .5 24 18 15 13 8 21 16 14 11 7 18 13 11 9 5 15 11 9 7 4 1 .5
1.6 21 16 14 11 7 19 14 12 10 6 16 12 10 8 5 14 10 8 6 4 1.6
1 .7 19 15 13 10 7 17 13 11 9 6 14 11 9 7 4 12 9 7 6 3 1 .7
1 .8 17 13 11 10 6 15 12 10 8 5 13 10 8 6 4 11 8 7 5 1.8
1 .9 16 12 11 9 6 14 11 9 8 5 12 9 7 6 4 10 7 6 5 1 .9

2.0 14 11 10 8 6 13 10 9 7 5 11 8 7 6 4 9 7 6 4 2 .0
2 .5 10 8 7 6 4 9 7 6 5 4 8 6 5 4 6 5 4 3 2 .5
3 .0 8 6 6 5 4 7 6 5 4 3 6 5 4 4 5 4 3 3 .0
3 .5 6 5 5 4 3 6 5 4 4 5 4 4 3 4 3 3 .5
4 .0 6 5 4 4 5 4 4 3 4 4 3 4 4 .0
 Appendix 5
F-Distribution
Upper 25%

4 5 6 7 8 9 10 12 15 20 24 30
V2
1 5,,83 7..50 8,.20 8.58 8.82 8.98 9. 10 9.19 9. 26 9.32 9.41 9. 49 9. 58 9. 63 9. 67
2 2.,57 3,.00 3,.15 3.23 3.28 3.31 3.34 3.35 3. 37 3.38 3.39 3.41 3.43 3. 43 3. 44
3 2,.02 2,.28 2,.36 2.39 2.41 2.42 2.43 2.44 2.44 2.44 2.45 2.46 2. 46 2. 46 2. 47
4 1.,81 2,.00 2,.05 2. 06 2.07 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2. 08
5 1.,69 1..85 1,,88 1,.89 1,.89 1,.89 1.,89 1.,89 1..89 1,.89 1.,89 1,.89 1,.88 1..88 1,.88
6 1.,62 1.,76 1,.78 1..79 1,.79 1,.78 1.,78 1,,78 1.,77 1,.77 1.,77 1,.76 1,.76 1,.75 1..75
7 1.,57 1.,70 1.,72 1..72 1,.71 1,.71 1,.70 1.,70 1.,69 1,.69 1.,68 1,.68 1,.67 1..67 1,.66
8 1..54 1..66 1..67 1,.66 1,.66 1,.65 1..64 1.,64 1.,63 1,.63 1.,62 1,.62 1,.61 1,.60 1,.60
9 1.,51 1.,62 1,.63 1..63 1,.62 1,.61 1..60 1,,60 1.,59 1,.59 1.,58 1,.57 1,.56 1..56 1,.55
10 1,.49 1.,60 1,.60 1,.59 1,.59 1 .58 1..57 1,,56 1..56 1,.55 1,.54 1,.53 1,.52 1,.52 1,.51
11 1..47 1.,58 1..58 1,.57 1,.56 1 .55 1.,54 1.,53 1.,53 1,.52 1,.51 1,.50 1,.49 1,.49 1,.48
12 1..46 1,.56 1,.56 1,.55 1 .54 1 .53 1..52 1,.51 1,.51 1,.50 1,.49 1,.48 1,.47 1,.46 1,.45
13 1..45 1..55 1,.55 1,.53 1,.52 1,.51 1,.50 1,.49 1.,49 1 .48 1,.47 1 .46 1,.45 1,.44 1 .43
14 1..44 1..53 1,.53 1,.52 1,.51 1 .50 1.,49 1,.48 1,.47 1 .46 1,.45 1 .44 1 .43 1 .42 1 .41
15 1,.43 1,.52 1,,52 1..51 1,.49 1,.48 1..47 1..46 1..46 1,.45 1.,44 1,.43 1,.41 1..41 1,.40
16 1..42 1..51 1,.51 1,.50 1,.48 1,.47 1..46 1,.45 1.,44 1,.44 1,.43 1,.41 1,.40 1,.39 1,.38
17 1.,42 1.,51 1,.50 1.,49 1..47 1,.46 1..45 1,.44 1.,43 1,.43 1..41 1,.40 1,.39 1,.38 1,.37
18 1..41 1..50 1,.49 1,.48 1,.46 1 .45 1.,44 1,.43 1.,42 1 .42 1,,40 1,.39 1,.38 1,.37 1,.36
19 1..41 1,.49 1,.49 1,.47 1,.46 1,.44 1.,43 1,.42 1.,41 1,.41 1.,40 1,.38 1,.37 1,.36 1,.35
20 1,.40 1..49 1,.48 1,.47 1 .45 1,.44 1..43 1,.42 1..41 1,.40 1..39 1 .37 1,.36 1 .35 1,.34
21 1.,40 1,.48 1,.48 1,.46 1,,44 1,.43 1.,42 1,.41 1..40 1,.39 1.,38 1,.37 1,.35 1,.34 1,.33
22 1..40 1..48 1 .47 1 .45 1,.44 1 .42 1,.41 1,.40 1,.39 1,.39 1..37 1 .36 1,.34 1,.33 1,.32
23 1,.39 1,.47 1,.47 1,.45 1,.43 1,.42 1,.41 1,.40 1,.39 1 .38 1,.37 1,.35 1 .34 1 .33 1,.32
24 1,.39 1..47 1,.46 1,.44 1,.43 1,.41 1..40 1,.39 1..38 1,.38 1,.36 1,.35 1,,33 1,.32 1,.31
25 1.,39 1..47 1.,46 1..44 1,.42 1,.41 1..40 1,.39 1.,38 1,.37 1.,36 1,.34 1,.33 1..32 1,.31
26 1..38 1..46 1,.45 1,.44 1,.42 1 .41 1..39 1,.38 1.,37 1,.37 1..35 1,.34 1,.32 1,.31 1,.30
27 1..38 1,.46 1,.45 1,.43 1,.42 1,.40 1.,39 1,.38 1,.37 1,.36 1,.35 1,.33 1,.32 1,.31 1,.30
28 1,.38 1,.46 1,.45 1,.43 1 .41 1,.40 1..39 1,.38 1,.37 1 .36 1,.34 1 .33 1,.31 1,.30 1 .29
29 1..38 1,.45 1 .45 1,.43 1,.41 1 .40 1,.38 1,.37 1.,36 1,.35 1,.34 1,.32 1,.31 1,.30 1,.29
30 1..38 1,.45 1 .44 1 .42 1 .41 1,.39 1,.38 1,.37 1.,36 1,.35 1,.34 1 .32 1 .30 1,.29 1,.28
40 1,.36 1,.44 1,.42 1 .40 1,.39 1,.37 1,.36 1,.35 1.,34 1,.33 1.,31 1 .30 1,.28 1..26 1,.25
60 1..35 1,.42 1,.41 1,.38 1,.37 1,.35 1,.33 1,.32 1.,31 1,.30 1,,29 1,.27 1,.25 1,.24 1,.22
120 1,.34 1..40 1,.39 1,.37 1,.35 1,.33 1.,31 1,.30 1.,29 1,.28 1.. f 6 1,.24 1,.22 1..21 1,.19
00 1..32 1..39 1,.37 1,.35 1,.33 1,.31 1.,29 1,.28 1.,27 1..25 1.,24 1,.22 1,.19 1..18 1,.16
 Appendix 5 (Continued)
F-Distribution
Upper 10%
V1 1 2 3 4 5 6 7 8 9 10 12 15 20 24 30 40 60 120
V2 ____________
1 3 9 .8 6 4 9 .5 0 5 3 .5 9 5 5 .8 3 5 7 .2 4 5 8 .2 0 58..91 59.,44 59.,86 60.,19 60.,71 61.,22 61. 74 62.,00 62.,26 62.,53 62. 79 63.,06 63.,33
2 8 .5 3 9 .0 0 9 .1 6 9 .2 4 9 .2 9 9 .3 3 9.,35 9.,37 9.,38 9.,39 9.,41 9.,42 9. 44 9. 45 9.,46 9.,47 9. 47 9.,48 9.,49
3 5 .5 4 5 .4 6 5 .3 9 5 .3 4 5 .3 1 5 .2 8 5.,27 5.,25 5.,24 5.,23 5.,22 5.,20 5. 18 5. 18 5., 17 5.,16 5. 15 5.,14 5. 13
4 4 .5 4 4 .3 2 4 .1 9 4 .1 1 4 .0 5 4 .0 1 3.,98 3.,95 3.,94 3.,92 3..90 3.,87 3. 84 3. 83 3.,82 3.,80 3.,79 3.,78 3. 76

5 4 .0 6 3 .7 8 3 .6 2 3 .5 2 3 .4 5 3 .4 0 3.,37 3.,34 3..32 3.,30 3.,27 3,,24 3.,21 3.,19 3.,17 3.,16 3.,14 3. 12 3. 10
6 3 .7 8 3 .4 6 3 .2 9 3 .1 8 3 .1 1 3 .0 5 3.,01 2 .,98 2 .,96 2 .,94 2 ..90 2 .,87 2 .,84 2 .,82 2 .,80 2 .,78 2 ..76 2 .,74 2 .,72
7 3 .5 9 3 .2 6 3 .0 7 2 .9 6 2.88 2 .8 3 2 .,78 2 .,75 2 .,72 2 .,70 2 .,67 2 . 63 2 .,59 2 .,58 2 ..56 2 .,54 2 .,51 2 .,49 2 .,47
8 3 .4 6 3 .1 1 2 .9 2 2 .8 1 2 .7 3 2 .6 7 2 .,62 2 . 59 2 .,56 2 .,54 2 . 50 2 , 46 2 .,42 2 . 40 2 .,38 2 .,36 2 .,34 2 . 32 2 .,29
9 3 .3 6 3 .0 1 2 .8 1 2 .6 9 2 .6 1 2 .5 5 2 ..51 2 .,47 2 ..44 2 .,42 2 .,38 2 , 34 2 .,30 2 .,28 2 .,25 2 , 23 2 .,21 2 .,18 2 .,16

10 3 .2 9 2 .9 2 2 .7 3 2 .6 1 2 .5 2 2 .4 6 2 .,41 2 .,38 2 .,35 2 ..32 2 .,28 2 ..24 2 ..20 2 ,,18 2 ., 16 2 . 13 2 .,11 2 .,08 2 .,06
11 3 .2 3 2.86 2.66 2 .5 4 2 .4 5 2 .3 9 2 .,34 2 ..30 2 .,27 2 .,25 2 . 21 2 ..17 2 .,12 2 .,10 2 ,,08 2 .,05 2 . 03 2 .,00 1.,97
12 3 .1 8 2 .8 1 2 .6 1 2 .4 8 2 .3 9 2 .3 3 2 .,28 2 .,24 2 ,. 21 2 .,19 2 .,15 2 ,. 10 2 .,06 2 .,04 2 .,01 1..99 1.,96 1.,93 1.,90
13 3 .1 4 2 .7 6 2 .5 6 2 .4 3 2 .3 5 2 .2 8 2 ..23 2 .,20 2 .,16 2 .,14 2 .,10 2 ., 05 2 . 01 1.,98 1.,9 6 1.,93 1.,90 1.,88 1 .,85
14 3 .1 0 2 .7 3 2 .5 2 2 .3 9 2 .3 1 2 .2 4 2 .,19 2 .,15 2 .,12 2 .,10 2 ., 05 2 ,. 01 1.,96 1.,94 1 ,.91 1.,89 1.,86 1.,83 1 .,80

15 3 .0 7 2 .7 0 2 .4 9 2 .3 6 2 .2 7 2 .2 1 2 .,16 2 .,12 2 ,,09 2 ..0 6 2 ., 02 1,.97 1.,92 1.,90 1.,87 1,.85 1..82 1,.79 1..76
16 3 .0 5 2 .6 7 2 .4 6 2 .3 3 2 .2 4 2 .1 8 2 .,13 2 .,09 2 ,. 06 2 ,. 03 1..99 1,.94 1.,89 1.,87 1.,84 1,.81 1,,78 1.,75 1,.72
17 3 .0 3 2 .6 4 2 .4 4 2 .3 1 2 .2 2 2 .1 5 2 ..10 2 .,06 2 ,. 03 2 ,. 00 1..96 1.,91 1.,86 1,.84 1,.81 1,.78 1,.75 1.,72 1..69
18 3 .0 1 2 .6 2 2 .4 2 2 .2 9 2.20 2 .1 3 2 .,08 2 ..04 2 ,. 00 1.,98 1..93 1.,89 1,,84 1.,81 1,.78 1,.75 1,.72 1.,69 1,.66
19 2 .9 9 2 .6 1 2 .4 0 2 .2 7 2 .1 8 2 .1 1 2 ..06 2 .,02 1.,98 1 ,. 9 6 1..91 1,.86 1.,81 1.,79 1,.7 6 1,.73 1,,70 1..67 1,.63
20 2 .9 7 2 .5 9 2 .3 8 2 .2 5 2 .1 6 2 .0 9 2 .,04 2 .,00 1.,96 1.,94 1 ,,89 1,.84 1.,79 1.,77 1,.7 4 1,.71 1,.68 1,.64 1,.61
21 2 .9 6 2 .5 7 2 .3 6 2 .2 3 2 .1 4 2 .0 8 2 .,02 1.,98 1..95 1.,92 1,.87 1,.8 3 1.,78 1,.75 1,.72 1,.69 1..66 1,.62 1,.59
22 2 .9 5 2 .5 6 2 .3 5 2 .2 2 2 .1 3 2 .0 6 2 .,01 1.,97 1..93 1,.9 0 1,.86 1,.81 1.,76 1..73 1,.7 0 1,.67 1,.6 4 1,.60 1,.57
23 2 .9 4 2 .5 5 2 .3 4 2 .2 1 2 .1 1 2 .0 5 1.,9 9 1.,95 1.,92 1,.8 9 1..84 1,.8 0 1..74 1.,72 1,.6 9 1,.66 1..62 1..59 1,.55
24 2 .9 3 2 .5 4 2 .3 3 2 .1 9 2.10 2 .0 4 1.,98 1.,94 1.,91 1.,88 1,.83 1,.78 1,,73 1,.70 1,.6 7 1,.64 1,.61 1.,57 1,.53
25 2 .9 2 2 .5 3 2 .3 2 2 .1 8 2 .0 9 2 .0 2 1.,97 1 ,,93 1..89 1.,8 7 1,.82 1,.7 7 1,.72 1..69 1,.66 1.6 3 1,.5 9 1.56 1 .52
26 2 .9 1 2 .5 2 2 .3 1 2 .1 7 2 .0 8 2 .0 1 1.,9 6 1.,92 1.,88 1.,86 1.,81 1,.7 6 1,.71 1,.68 1 ,. 65 1,.61 1,.58 1,.54 1,.5 0
27 2 .9 0 2 .5 1 2 .3 0 2 .1 7 2 .0 7 2.00 1..95 1.,91 1..87 1.,85 1,.8 0 1 ,. 75 1,.70 1,,67 1,.64 1,.6 0 1,.5 7 1,.53 1,.49
28 2 .8 9 2 .5 0 2 .2 9 2 .1 6 2 .0 6 2.00 1..94 1.,90 1..87 1.,8 4 1,.79 1 ,. 7 4 1,.69 1..66 1..6 3 1,.5 9 1,.5 6 1,.52 1,.4 8
29 2 .8 9 2 .5 0 2 .2 8 2 .1 5 2 .0 6 1 .9 9 1..9 3 1.,89 1,.86 1.,8 3 1..78 1,.7 3 1,,68 1.,65 1,.62 1,.5 8 1,,55 1,.51 1,.4 7
30 2.88 2 .4 9 2 .2 8 2 .1 4 2 .0 5 1 .9 8 1..93 1.,88 1..85 1.,82 1..77 1,.72 1..67 1,.64 1,.61 1,.5 7 1,.5 4 1,.50 1,.4 6
40 2 .8 4 2 .4 4 2 .2 3 2 .0 9 2.00 1 .9 3 1.,87 1.,83 1,.79 1.,7 6 1,.71 1,.66 1,.61 1,.57 1.5 4 1,.51 1,.4 7 1.42 1.38
60 2 .7 9 2 .3 9 2 .1 8 2 .0 4 1 .9 5 1 .8 7 1.,82 1.,77 1..74 1..71 1,.66 1,.6 0 1,.54 1,.51 1,.4 8 1,.4 4 1,.4 0 1,.35 1,.2 9
120 2 .7 5 2 .3 5 2 .1 3 1 .9 9 1 .9 0 1 .8 2 1.,77 1.,72 1..68 1..65 1,.60 1,.55 1..48 1,.45 1,.41 1.3 7 1,.32 1.26 1.1 9
00 2 .7 1 2 .3 0 2 .0 8 1 .9 4 1 .8 5 1 .7 7 1..72 1.,67 1..63 1.,6 0 1,.55 1,.4 9 1.,42 1.,38 1,.3 4 1,.3 0 1,.2 4 1.17 1.00
 Appendix 5 (Continued)
F-Distribution
Upper 5%

V2
l 8 10 12 15 20 24 30 40 60 120
1 1 61. 200. 216. 225. 230 . 234. 237 . 239 . 241. 242. 2 44 . 246. 248. 2 49 . 2 50 . 2 51 . 252. 2 53 . 254 .
2 1 8 .5 1 1 9 .0 0 1 9 .1 6 1 9 .2 5 1 9 .3 0 1 9 .3 3 1 9 .3 5 1 9 .3 7 1 9 .3 8 1 9 .4 0 19.,41 19.,43 1 9 .4 5 19., 45 19.,46 19,.47 19..48 19.,49 19..50
3 1 0 .1 3 9 .5 5 9 .2 8 9 .1 2 9 .0 1 8 .9 4 8 .8 9 8 .8 5 8 .8 1 8 .7 9 8 .,74 8 .,70 8 .6 6 8 ..64 8 ..62 8 ,. 5 9 8 ..57 8 .,55 8 ,. 53
4 7 .7 1 6 .9 4 6 .5 9 6 .3 9 6 .2 6 6 .1 6 6 .0 9 6 .0 4 6 .0 0 5 .9 6 5.,91 5..86 5 .8 0 5,.77 5,.75 5,.72 5,.69 5..66 5,.6 3
5 6 .6 1 5 .7 9 5 .4 1 5 .1 9 5 .0 5 4 .9 5 4 .8 8 4 .8 2 4 .7 7 4 .7 4 4.,68 4..62 4 .5 6 4..53 4.,50 4..4 6 4..4 3 4..40 4,.3 6
6 5 .9 9 5 .1 4 4 .7 6 4 .5 3 4 .3 9 4 .2 8 4 .2 1 4 .1 5 4 .1 0 4 .0 6 4..00 3..94 3 .8 7 3..84 3.,81 3..77 3,.74 3.,70 3,.67
7 5 .5 9 4 .7 4 4 .3 5 4 .1 2 3 .9 7 3 .8 7 3 .7 9 3 .7 3 3 .6 8 3 .6 4 3.,57 3..51 3 .4 4 3..41 3..38 3,.34 3,.3 0 3..27 3..2 3
8 5 .3 2 4 .4 6 4 .0 7 3 .8 4 3 .6 9 3 .5 8 3 .5 0 3 .4 4 3 .3 9 3 .3 5 3..28 3..22 3 .1 5 3,.12 3,.08 3 .04 3,.01 2 .,97 2 ,. 9 3
9 5 .1 2 4 .2 6 3 .8 6 3 .6 3 3 .4 8 3 .3 7 3 .2 9 3 .2 3 3 .1 8 3 .1 4 3.,07 3,.01 2 .9 4 2 ..90 2 . 86 2 ,. 8 3 2 ,. 79 2 ..75 2 .71
10 4 .9 6 4 .1 0 3 .7 1 3 .4 8 3 .3 3 3 .2 2 3 .1 4 3 .0 7 3 .0 2 2 .9 8 2 ..91 2 ,.85 2 .7 7 2 ,. 74 2 ,. 70 2 ,. 66 2 ,. 62 2 ..58 2 ,.54
11 4 .8 4 3 .9 8 3 .5 9 3 .3 6 3 .2 0 3 .0 9 3 .0 1 2 .9 5 2 .9 0 2 .8 5 2 ..79 2 .,72 2 .6 5 2 ,. 61 2 ,. 57 2 ,. 53 2 ,. 49 2 ..45 2 .40
12 4 .7 5 3 .8 9 3 .4 9 3 .2 6 3 .1 1 3 .0 0 2 .9 1 2 .8 5 2 .8 0 2 .7 5 2 ..69 2 ,. 62 2 .5 4 2 ,.51 2 . 47 2 ,. 4 3 2 ,. 38 2 ..34 2 .30
13 4 .6 7 3 .8 1 3 .4 1 3 .1 8 3 .0 3 2 .9 2 2 .8 3 2 .7 7 2 .7 1 2 .6 7 2, .60 2 ,. 53 2 .4 6 2 ., 42 2 ..38 2 .34 2 ,. 3 0 2 ,.25 2 ,. 21
14 4 .6 0 3 .7 4 3 .3 4 3 .1 1 2 .9 6 2 .8 5 2 .7 6 2 .7 0 2 .6 5 2 .6 0 2 .,53 2 ..4 6 2 .3 9 2 ..35 2 ..31 2 ,. 2 7 2 ,. 22 2 .,18 2 ,. 1 3
15 4 .5 4 3 .6 8 3 .2 9 3 .0 6 2 .9 0 2 .7 9 2 .7 1 2 .6 4 2 .5 9 2 .5 4 2 .,48 2 .,4 0 2 .3 3 2 .,29 2 ,. 25 2 .,20 2 ..1 6 2 .,11 2 . 07
16 4 .4 9 3 .6 3 3 .2 4 3 .0 1 2 .8 5 2 .7 4 2.66 2 .5 9 2 .5 4 2 .4 9 2 .,42 2 ,. 35 2 .2 8 2 ..24 2 ..19 2 ,. 15 2 ,. 11 2 .,06 2 . 01
17 4 .4 5 3 .5 9 3 .2 0 2 .9 6 2 .8 1 2 .7 0 2 .6 1 2 .5 5 2 .4 9 2 .4 5 2 . 38 2 ..31 2 .2 3 2 .,19 2 .,15 2 ,. 10 2 ,. 0 6 2 ..01 1 ,. 9 6
18 4 .4 1 3 .5 5 3 .1 6 2 .9 3 2 .7 7 2.66 2 .5 8 2 .5 1 2 .4 6 2 .4 1 2 .,34 2 ..27 2 .1 9 2 ,. 15 2 , 11 2 ,. 06 2 ,. 02 1 ..97 1 .92
19 4 .3 8 3 .5 2 3 .1 3 2 .9 0 2 .7 4 2 .6 3 2 .5 4 2 .4 8 2 .4 2 2 .3 8 2 .,31 2 .,23 2 .1 6 2 ..11 2 .,07 2 ,. 0 3 1 ..98 1 .,93 1 ..88
20 4 .3 5 3 .4 9 3 .1 0 2 .8 7 2 .7 1 2 .6 0 2 .5 1 2 .4 5 2 .3 9 2 .3 5 2 . 28 2 ,. 20 2.12 2 . 08 2 ,. 04 1 ,. 9 9 1 .95 1 ..90 1 .84
21 4 .3 2 3 .4 7 3 .0 7 2 .8 4 2.68 2 .5 7 2 .4 9 2 .4 2 2 .3 7 2 .3 2 2 . 25 2 ..18 2.10 2 ,. 05 2 .,01 1 .9 6 1 .92 1 ..87 1 ,.81
22 4 .3 0 3 .4 4 3 .0 5 2 .8 2 2.66 2 .5 5 2 .4 6 2 .4 0 2 .3 4 2 .3 0 2 ,. 23 2 ,. 15 2 .0 7 2 ,. 03 1 ,. 98 1 .9 4 1 .8 9 1 ,.84 1 .7 8
23 4 .2 8 3 .4 2 3 .0 3 2 .8 0 2 .6 4 2 .5 3 2 .4 4 2 .3 7 2 .3 2 2 .2 7 2 ,. 20 2 ,. 1 3 2 .0 5 2 ,. 01 1 ,. 96 1 .91 1 ,. 86 1 ..81 1 .7 6
24 4 .2 6 3 .4 0 3 .0 1 2 .7 8 2 .6 2 2 .5 1 2 .4 2 2 .3 6 2 .3 0 2 .2 5 2 ;.18 2 ,. 11 2 .0 3 1 ,.98 1 ,. 94 1 .8 9 1 .84 1 ,. 79 1 . 73
25 4 .2 4 3 .3 9 2 .9 9 2 .7 6 2 .6 0 2 .4 9 2 .4 0 2 .3 4 2 .2 8 2 .2 4 2 ,. 16 2 ..0 9 2 .0 1 1 ,. 9 6 1 ..92 1 .87 1 .82 1 ,.77 1 .71
26 4 .2 3 3 .3 7 2 .9 8 2 .7 4 2 .5 9 2 .4 7 2 .3 9 2 .3 2 2 .2 7 2.22 2 ..15 2 ,.0 7 1 .9 9 1 ,. 95 1 .,90 1 ,. 85 1 ,. 8 0 1 .,75 1 ,. 6 9
27 4 .2 1 3 .3 5 2 .9 6 2 .7 3 2 .5 7 2 .4 6 2 .3 7 2 .3 1 2 .2 5 2.20 2 ,. 13 2 .,0 6 1 .9 7 1 ,. 93 1 . 88 1 ,. 84 1 ,. 7 9 1 .,73 1 ,. 6 7
28 4 .2 0 3 .3 4 2 .9 5 2 .7 1 2 .5 6 2 .4 5 2 .3 6 2 .2 9 2 .2 4 2 .1 9 2 ..12 2 ,.04 1 .9 6 1 . 91 1 ..87 1 .82 1 ,. 77 1 ..71 1 .65
29 4 .1 8 3 .3 3 2 .9 3 2 .7 0 2 .5 5 2 .4 3 2 .3 5 2 .2 8 2 .2 2 2 .1 8 2, .10 2 ,. 0 3 1 .9 4 1 ,. 9 0 1 ,. 85 1 .81 1 .7 5 1 ,. 7 0 1 .64
30 4 .1 7 3 .3 2 2 .9 2 2 .6 9 2 .5 3 2 .4 2 2 .3 3 2 .2 7 2 .2 1 2 .1 6 2. ,09 2 ..01 1 .9 3 1 ,. 8 9 1 .,84 1 ,. 7 9 1 ,. 7 4 1 ..68 1 .62
40 4 .0 8 3 .2 3 2 .8 4 2 .6 1 2 .4 5 2 .3 4 2 .2 5 2 .1 8 2 .1 2 2 .0 8 2, .00 1 ,.92 1 .8 4 1 ,. 79 1 ,. 74 1 ,. 6 9 1 ,.64 1 ..5 8 1 ,.51
60 4 .0 0 3 .1 5 2 .7 6 2 .5 3 2 .3 7 2 .2 5 2 .1 7 2.10 2 .0 4 1 .9 9 1 .,92 1 ..8 4 1 .7 5 1 .,7 0 1 ..65 1 ,. 5 9 1 ,. 5 3 1 .,47 1 ,. 3 9
120 3 .9 2 3 .0 7 2.68 2 .4 5 2 .2 9 2 .1 7 2 .0 9 2.02 1 .9 6 1 .9 1 1 .,83 1 .,75 1.66 1 ..61 1 .,55 1 ,. 5 0 1 ,. 4 3 1 ..35 1 ,.25
3 .8 4 3 .0 0 2 .6 0 2 .3 7 2.21 2.10 2 .0 1 1 .9 4 1 .8 8 1 .8 3 1 .,75 1 .,67 1 .5 7 1 .,52 1 .,46 1 ,. 3 9 1 ,.32 1 .,22 1 ..00
 Appendix 5 (Continued)
F-Distribution
Upper 1%
V1 r
1 2 3 4 5 6 8 9 10 12 15 20 24 30 40 60 120 00
V2
1 4 052. 5 00 0 . 5 40 3 . 5625. 5 764. 5859. 5 928. 5 982. 6 02 2 . 6056. 6106. 615 7 . 6209. 6235. 6261. 6287. 631 3 . 633 9 . 6366.
2 9 8 .5 0 9 9 .0 0 9 9 .1 7 9 9 .2 5 9 9 .3 0 9 9 .3 3 99. 36 9 9 .3 7 9 9 .3 9 9 9 .4 0 9 9 .4 2 9 9 .4 3 9 9 .4 5 9 9 .4 6 9 9 .4 7 9 9 .4 7 9 9 .4 8 9 9 .4 9 99.,50
3 3 4 .1 2 3 0 .8 2 2 9 .4 6 2 8 .7 1 2 8 .2 4 2 7 .9 1 27. 67 2 7 .4 9 2 7 .3 5 2 7 .2 3 2 7 .0 5 2 6 .8 7 2 6 .6 9 2 6 .6 0 2 6 .5 0 2 6 .4 1 2 6 .3 2 2 6 .2 2 26.,13
4 21.20 1 8 .0 0 1 6 .6 9 1 5 .9 8 1 5 .5 2 1 5.2 1 14. 98 1 4 .8 0 1 4 .6 6 1 4 .5 5 1 4 .3 7 1 4 .2 0 1 4 .0 2 1 3 .9 3 1 3 .8 4 1 3 .7 5 1 3 .6 5 1 3 .5 6 13..46
5 1 6 .2 6 1 3 .2 7 1 2 .0 6 1 1 .3 9 1 0 .9 7 1 0 .6 7 10 .,46 1 0 .2 9 1 0 .1 6 1 0 .0 5 9 .8 9 9 .7 2 9 .5 5 9 .4 7 9 .3 8 9 .2 9 9 .2 0 9 .1 1 9.,02
6 1 3 .7 5 1 0 .9 2 9 .7 8 9 .1 5 8 .7 5 8 .4 7 8 .,26 8 . 1 0 7 .9 8 7 .8 7 7 .7 2 7 .5 6 7 .4 0 7 .3 1 7 .2 3 7 .1 4 7 .0 6 6 .9 7 6 .,88
7 1 2 .2 5 9 .5 5 8 .4 5 7 .8 5 7 .4 6 7 .1 9 6 . 99 6 .8 4 6 .7 2 6 .6 2 6 .4 7 6 .3 1 6 .1 6 6 .0 7 5 .9 9 5 .9 1 5 .8 2 5 .7 4 5..65
8 1 1 .2 6 8 .6 5 7 .5 9 7 .0 1 6 .6 3 6 .3 7 6 .,18 6 .0 3 5 .9 1 5 .8 1 5 .6 7 5 .5 2 5 .3 6 5 .2 8 5 .2 0 5 .1 2 5 .0 3 4 .9 5 4.,86
9 1 0 .5 6 8 .0 2 6 .9 9 6 .4 2 6 .0 6 5 .8 0 5. 61 5 .4 7 5 .3 5 5 .2 6 5 .1 1 4 .9 6 4 .8 1 4 .7 3 4 .6 5 4 .5 7 4 .4 8 4 .4 0 4..31
10 1 0 .0 4 7 .5 6 6 .5 5 5 .9 9 5 .6 4 5 .3 9 5. 20 5 .0 6 4 .9 4 4 .8 5 4 .7 1 4 .5 6 4 .4 1 4 .3 3 4 .2 5 4 .1 7 4 .0 8 4 .0 0 3.,91
11 9 .6 5 7 .2 1 6 .22 5 .6 7 5 .3 2 5 .0 7 4.,89 4 .7 4 4 .6 3 4 .5 4 4 .4 0 4 .2 5 4 .1 0 4 .0 2 3 .9 4 3 .8 6 3 .7 8 3 .6 9 3,.60
12 9 .3 3 6 .9 3 5 .9 5 5 .4 1 5 .0 6 4 .8 2 4.,64 4 .5 0 4 .3 9 4 .3 0 4 .1 6 4 .0 1 3 .8 6 3 .7 8 3 .7 0 3 .6 2 3 .5 4 3 .4 5 3.,36
13 9 .0 7 6 .7 0 5 .7 4 5 .2 1 4 .8 6 4 .6 2 4.,44 4 .3 0 4 .1 9 4 .1 0 3 .9 6 3 .8 2 3 .6 6 3 .5 9 3 .5 1 3 .4 3 3 .3 4 3 .2 5 3.,17
14 8 .86 6 .5 1 5 .5 6 5 .0 4 4 .6 9 4 .4 6 4.,28 4 .1 4 4 .0 3 3 .9 4 3 .8 0 3 .6 6 3 .5 1 3 .4 3 3 .3 5 3 .2 7 3 .1 8 3 .0 9 3,.00
15 8 .68 6 .3 6 5 .4 2 4 .8 9 4 .5 6 4 .3 2 4.,14 4 .0 0 3 .8 9 3 .8 0 3 .6 7 3 .5 2 3 .3 7 3 .2 9 3 .2 1 3 .1 3 3 .0 5

395
2 .9 6 2 ,. 87
16 8 .5 3 6 .2 3 5 .2 9 4 .7 7 4 .4 4 4 .2 0 4.,03 3 .8 9 3 .7 8 3 .6 9 3 .5 5 3 .4 1 3 .2 6 3 .1 8 3 .1 0 3 .0 2 2 .9 3 2 .8 4 2 ,. 75
17 8 .4 0 6.11 5 .1 8 4 .6 7 4 .3 4 4 .1 0 3. 93 3 .7 9 3 .6 8 3 .5 9 3 .4 6 3 .3 1 3 .1 6 3 .0 8 3 .0 0 2 .9 2 2 .8 3 2 .7 5 2 ..65
18 8 .2 9 6.01 5 .0 9 4 .5 8 4 .2 5 4 .0 1 3.,84 3 .7 1 3 .6 0 3 .5 1 3 .3 7 3 .2 3 3 .0 8 3 .0 0 2 .9 2 2 .8 4 2 .7 5 2.66 2 ,. 57
19 8 .1 8 5 .9 3 5 .0 1 4 .5 0 4 .1 7 3 .9 4 3.,77 3 .6 3 3 .5 2 3 .4 3 3 .3 0 3 .1 5 3 .0 0 2 .9 2 2 .8 4 2 .7 6 2 .6 7 2 .5 8 2 , 49
20 8.10 5 .8 5 4 .9 4 4 .4 3 4 .1 0 3 .8 7 3.,70 3 .5 6 3 .4 6 3 .3 7 3 .2 3 3 .0 9 2 .9 4 2.86 2 .7 8 2 .6 9 2 .6 1 2 .5 2 2 ,. 42
21 8.02 5 .7 8 4 .8 7 4 .3 7 4 .0 4 3 .8 1 3.,64 3 .51 3 .4 0 3 .3 1 3 .1 7 3 .0 3 2.88 2 .8 0 2 .7 2 2 .6 4 2 .5 5 2 .4 6 2 ,,36
22 7 .9 5 5 .7 2 4 .8 2 4 .3 1 3 .9 9 3 .7 6 3.,59 3 .4 5 3 .3 5 3 .2 6 3 .1 2 2 .9 8 2 .8 3 2 .7 5 2 .6 7 2 .5 8 2 .5 0 2 .4 0 2 ..31
23 7 .8 8 5 .6 6 4 .7 6 4 .2 6 3 .9 4 3 .7 1 3.,54 3 .4 1 3 .3 0 3 .2 1 3 .0 7 2 .9 3 2 .7 8 2 .7 0 2 .6 2 2 .5 4 2 .4 5 2 .3 5 2 ,. 26
24 7 .8 2 5 .6 1 4 .7 2 4 .2 2 3 .9 0 3 .6 7 3.,50 3 .3 6 3 .2 6 3 .1 7 3 .0 3 2 .8 9 2 .7 4 2.66 2 .5 8 2 .4 9 2 .4 0 2 .3 1 2 .,21
25 7 .7 7 5 .5 7 4 .6 8 4 .1 8 3 .8 5 3 .6 3 3. 46 3 .3 2 3 .2 2 3 .1 3 2 .9 9 2 .8 5 2 .7 0 2 .6 2 2 .5 4 2 .4 5 2 .3 6 2 .2 7 2 .,17
26 7 .7 2 5 .5 3 4 .6 4 4 .1 4 3 .8 2 3 .5 9 3.,42 3 .2 9 2 .1 8 2 .0 9 2 .9 6 2 .8 1 2 .66 2 .5 8 2 .5 0 2 .4 2 2 .3 3 2 .2 3 2 ,. 13
27 7 .6 8 5 .4 9 4 .6 0 4 .1 1 3 .7 8 3 .5 6 3.,39 3 .2 6 3 .1 5 3 .0 6 2 .9 3 2 .7 8 2 .6 3 2 .5 5 2 .4 7 2 .3 8 2 .2 9 2 .2 0 2 ., 10
28 7 .6 4 5 .4 5 4 .5 7 4 .0 7 3 .7 5 3 .5 3 3. 36 3 .2 3 3 .1 2 3 .0 3 2 .9 0 2 .7 5 2 .6 0 2 .5 2 2 .4 4 2 .3 5 2 .2 6 2 .1 7 2 .,06
29 7 .6 0 5 .4 2 4 .5 4 4 .0 4 3 .7 3 3 .5 0 3. 33 3 .2 0 3 .0 9 3 .0 0 2 .8 7 2 .7 3 2 .5 7 2 .4 9 2 .4 1 2 .3 3 2 .2 3 2 .1 4 2 ,. 03
30 7 .5 6 5 .3 9 4 .5 1 4 .0 2 3 .7 0 3 .4 7 3. 30 3 .1 7 3 .0 7 2 .9 8 2 .8 4 2 .7 0 2 .5 5 2 .4 7 2 .3 9 2 .3 0 2 .2 1 2.1 1 2 ,. 01
40 7 .3 1 5 .1 8 4 .3 1 3 .8 3 3 .5 1 3 .2 9 3.,12 2 .9 9 2 .8 9 2 .8 0 2.66 2 .5 2 2 .3 7 2 .2 9 2.20 2 .1 1 2.02 1 .9 2 1 ,. 80
60 7 .0 8 4 .9 8 4 .1 3 3 .6 5 3 .3 4 3 .1 2 2 . 95 2 .8 2 2 .7 2 2 .6 3 2 .5 0 2 .3 5 2.20 2.12 2 .0 3 1 .9 4 1 .8 4 1 .7 3 1 ,. 60
120 6 .8 5 4 .7 9 3 .9 5 3 .4 8 3 .1 7 2 .9 6 2 .,79 2.66 2 .5 6 2 .4 7 2 .3 4 2 .1 9 2 .0 3 1 .9 5 1.86 1 .7 6 1.66 1 .5 3 1 ..38
00 6 .6 3 4 .6 1 3 .7 8 3 .3 2 3 .0 2 2 .8 0 2 .,64 2 .5 1 2 .4 1 2 .3 2 2 .1 8 2 .0 4 1 .88 1 .7 9 1 .7 0 1 .5 9 1 .4 7 1 .3 2 1 ..00
 Appendix 6
Upper Percentage Points of the Studentized Range
a = .10
1 3 4 5 6 7 8 9 10 12 L5
1 8 .9 2 9 1 3 .4 4 1 6 .3 6 1 8 .4 9 2 0 .1 5 2 1 .5 1 2 2 .6 4 2 3 .6 2 2 4 .4 8 2 5 .9 2 2 7 .6 2
2 4 .1 3 0 5 .7 3 3 6 .7 7 3 7 .5 3 8 8 .1 3 9 8 .6 3 3 9 .0 4 9 9 .4 0 9 9 .7 2 5 1 0 .2 6 1 0 .8 9
3 3 .3 2 8 4 .4 6 7 5 .1 9 9 5 .7 3 8 6 .1 6 2 6 .5 1 1 6 .8 0 6 7 .0 6 2 7 .2 8 7 7 .6 6 7 8.1 2 0
4 3 .0 1 5 3 .9 7 6 4 .5 8 6 5 .0 3 5 5 .3 8 8 5 .6 7 9 5 .9 2 6 6 .1 3 9 6 .3 2 7 6 .6 4 5 7 .0 2 5
5 2 .8 5 0 3 .7 1 7 4 .2 6 4 4 .6 6 4 4 .9 7 9 5 .2 3 8 5 .4 5 8 5 .6 4 8 5 .8 1 6 6.101 6 .4 4 0

6 2 .7 4 8 3 .5 5 9 4 .0 6 5 4 .4 3 5 4 .7 2 6 4 .9 6 6 5 .1 6 8 5 .3 4 4 5 .4 9 9 5 .7 6 2 6 .0 7 5
7 2 .6 8 0 3 .4 5 1 3 .9 3 1 4 .2 8 0 4 .5 5 5 4 .7 8 0 4 .9 7 2 5 .1 3 7 5 .2 8 3 5 .5 3 0 5 .8 2 6
8 2 .6 3 0 3 .3 7 4 3 .8 3 4 4 .1 6 9 4 .4 3 1 4 .6 4 6 4 .8 2 9 4 .9 8 7 5 .1 2 6 5 .3 6 2 5 .6 4 4
9 2 .5 9 2 3 .3 1 6 3 .7 6 1 4 .0 8 4 4 .3 3 7 4 .5 4 5 4 .7 2 1 4 .8 7 3 5 .0 0 7 5 .2 3 4 5 .5 0 6
10 2 .5 6 3 3 .2 7 0 3 .7 0 4 4 .0 1 8 4 .2 6 4 4 .4 6 5 4 .6 3 6 4 .7 8 3 4 .9 1 3 5 .1 3 4 5 .3 9 7

11 2 .5 4 0 3 .2 3 4 3 .6 5 8 3 .9 6 5 4 .2 0 5 4 .4 0 1 4 .5 6 8 4 .7 1 1 4 .8 3 8 5 .0 5 3 5 .3 0 9
12 2 .5 2 1 3 .2 0 4 3 .6 2 1 3 .9 2 2 4 .1 5 6 4 .3 4 9 4 .5 1 1 4 .6 5 2 4 .7 7 6 4 .9 8 6 5 .2 3 6
13 2 .5 0 5 3 .1 7 9 3 .5 8 9 3 .8 8 5 4 .1 1 6 4 .3 0 5 4 .4 6 4 4 .6 0 2 4 .7 2 4 4 .9 3 0 5 .1 7 6
14 2 .4 9 1 3 .1 5 8 3 .5 6 3 3 .8 5 4 4 .0 8 1 4 .2 6 7 4 .4 2 4 4 .5 6 0 4 .6 8 0 4 .8 8 2 5 .1 2 4
15 2 .4 7 9 3 .1 4 0 3 .5 4 0 3 .8 2 8 4 .0 5 2 4 .2 3 5 4 .3 9 0 4 .5 2 4 4 .6 4 1 4 .8 4 1 5 .0 7 9

16 2 .4 6 9 3 .1 2 4 3 .5 2 0 3 .8 0 4 4 .0 2 6 4 .2 0 7 4 .3 6 0 4 .4 9 2 4 .6 0 8 4 .8 0 5 5 .0 4 0
17 2 .4 6 0 3 .1 1 0 3 .5 0 3 3 .7 8 4 4 .0 0 4 4 .1 8 3 4 .3 3 4 4 .4 6 4 4 .5 7 9 4 .7 7 4 5 .0 0 5
18 2 .4 5 2 3 .0 9 8 3 .4 8 8 3 .7 6 7 3 .9 8 4 4 .1 6 1 4 .3 1 1 4 .4 4 0 4 .5 5 4 4 .7 4 6 4 .9 7 5
19 2 .4 4 5 3 .0 8 7 3 .4 7 4 3 .7 5 1 3 .9 6 6 4 .1 4 2 4 .2 9 0 4 .4 1 8 4 .5 3 1 4 .7 2 1 4 .9 4 8
20 2 .4 3 9 3 .0 7 8 3 .4 6 2 3 .7 3 6 3 .9 5 0 4 .1 2 4 4 .2 7 1 4 .3 9 8 4 .5 1 0 4 .6 9 9 4 .9 2 4

24 2 .4 2 0 3 .0 4 7 3 .4 2 3 3 .6 9 2 3 .9 0 0 4 .0 7 0 4 .2 1 3 4 .3 3 6 4 .4 4 5 4 .6 2 8 4 .8 4 7
30 2 .4 0 0 3 .0 1 7 3 .3 8 6 3 .6 4 8 3 .8 5 1 4 .0 1 6 4 .1 5 5 4 .2 7 5 4 .3 8 1 4 .5 5 9 4 .7 7 0
40 2 .3 8 1 2 .9 8 8 3 .3 4 9 3 .6 0 5 3 .8 0 3 3 .9 6 3 4 .0 9 9 4 .2 1 5 4 .3 1 7 4 .4 9 0 4 .6 9 5
60 2 .3 6 3 2 .9 5 9 3 .3 1 2 3 .5 6 2 3 .7 5 5 3 .9 1 1 4 .0 4 2 4 .1 5 5 4 .2 5 4 4 .4 2 1 4 .6 1 9
120 2 .3 4 4 2 .9 3 0 3 .2 7 6 3 .5 2 0 3 .7 0 7 3 .8 5 9 3 .9 8 7 4 .0 9 6 4 .1 9 1 4 .3 5 3 4 .5 4 3
2 .3 2 6 2 .9 0 2 3 .2 4 0 3 .4 7 8 3 .6 6 1 3 .8 0 8 3 .9 3 1 4 .0 3 7 4 .1 2 9 4 .2 8 5 4 .4 6 8
 Appendix 6 (Cont'd .)
Upper Percentage Points of the Studentized Range
a = .05

2 3 4 5 6 7 8 9 10 12 15 30 60
1 1 7 .9 7 2 6 .9 8 3 2 .8 2 3 7 .0 8 4 0 .4 1 4 3 .1 2 4 5 .4 0 4 7 .3 6 4 9 .0 7 5 1 .9 6 5 5 .3 6 6 5 .1 5 7 3 .9 7
2 6 .0 8 5 8 .3 3 1 9 .7 9 8 10.88 1 1 .7 4 1 2 .4 4 1 3 .0 3 1 3 .5 4 1 3 .9 9 1 4 .7 5 1 5 .6 5 1 8 .2 7 20.66
3 4 .5 0 1 5 .9 1 0 6 .8 2 5 7 .5 0 2 8 .0 3 7 8 .4 7 8 8 .8 5 3 9 .1 7 7 9 .4 6 2 9 .9 4 6 1 0 .5 3 12.21 1 3 .7 6
4 3 .9 2 7 5 .0 4 0 5 .7 5 7 6 .2 8 7 6 .7 0 7 7 .0 5 3 7 .3 4 7 7 .6 0 2 7 .8 2 6 8 .2 0 8 8 .6 6 4 10.00 1 1 .2 4
5 3 .6 3 5 4 .6 0 2 5 .2 1 8 5 .6 7 3 6 .0 3 3 6 .3 3 0 6 .5 8 2 6 .8 0 2 6 .9 9 5 7 .3 2 4 7 .7 1 7 8 .8 7 5 9 .9 4 9

6 3 .4 6 1 4 .3 3 9 4 .8 9 6 5 .3 0 5 5 .6 2 8 5 .8 9 5 6.122 6 .3 1 9 6 .4 9 3 6 .7 8 9 7 .1 4 3 8 .1 8 9 9 .1 6 3
7 3 .3 4 4 4 .1 6 5 4 .6 8 1 5 .0 6 0 5 .3 5 9 5 .6 0 6 5 .8 1 5 5 .9 9 8 6 .1 5 8 6 .4 3 1 6 .7 5 9 7 .7 2 8 8 .6 3 2
8 3 .2 6 1 4 .0 4 1 4 .5 2 9 4 .8 8 6 5 .1 6 7 5 .3 9 9 5 .5 9 7 5 .7 6 7 5 .9 1 8 6 ,1 7 5 6 .4 8 3 7 .3 9 5 8 .2 4 8
9 3 .1 9 9 3 .9 4 9 4 .4 1 5 4 .7 5 6 5 .0 2 4 5 .2 4 4 5 .4 3 2 5 .5 9 5 5 .7 3 9 5 ,9 8 3 6 .2 7 6 7 .1 4 5 7 .9 5 8
10 3 .1 5 1 3 .8 7 7 4 .3 2 7 4 .6 5 4 4 .9 1 2 5 .1 2 4 5 .3 0 5 5 .4 6 1 5 .5 9 9 5 .8 3 3 6 .1 1 4 6 .9 4 8 7 .7 3 0

11 3 .1 1 3 3 .8 2 0 4 .2 5 6 4 .5 7 4 4 .8 2 3 5 .0 2 8 5 .2 0 2 5 .3 5 3 5 .4 8 7 5 .7 1 3 5 .9 8 4 6 .7 9 0 7 .5 4 6
12 3 .0 8 2 3 .7 7 3 4 .1 9 9 4 .5 0 8 4 .7 5 1 4 .9 5 0 5 .1 1 9 5 .2 6 5 5 .3 9 5 5 .6 1 5 5 .8 7 8 6 .6 6 0 7 .3 9 4
13 3 .0 5 5 3 .7 3 5 4 .1 5 1 4 .4 5 3 4 .6 9 0 4 .8 8 5 5 .0 4 9 5 .1 9 2 5 .3 1 8 5 .5 3 3 5 .7 8 9 6 .5 5 1 7 .2 6 7
14 3 .0 3 3 3 .7 0 2 4 .1 1 1 4 .4 0 7 4 .6 3 9 4 .8 2 9 4 .9 9 0 5 .1 3 1 5 .2 5 4 5 .4 6 3 5 .7 1 4 6 .4 5 9 7 .1 5 9
15 3 .0 1 4 3 .6 7 4 4 .0 7 6 4 .3 6 7 4 .5 9 5 4 .7 8 2 4 .9 4 0 5 .0 7 7 5 .1 9 8 5 .4 0 4 5 .6 4 9 6 .3 7 9 7 .0 6 5

16 2 .9 9 8 3 .6 4 9 4 .0 4 6 4 .3 3 3 4 .5 5 7 4 .7 4 1 4 .8 9 7 5 .0 3 1 5 .1 5 0 5 .3 5 2 5 .5 9 3 6 .3 1 0 6 .9 8 4
17 2 .9 8 4 3 .6 2 8 4 .0 2 0 4 .3 0 3 4 .5 2 4 4 .7 0 5 4 .8 5 8 4 .9 9 1 5 .1 0 8 5 .3 0 7 5 .5 4 4 6 .2 4 9 6 .9 1 2
18 2 .9 7 1 3 .6 0 9 3 .9 9 7 4 .2 7 7 4 .4 9 5 4 .6 7 3 4 .8 2 4 4 .9 5 6 5 .0 7 1 5 .2 6 7 5 .5 0 1 6 .1 9 5 6 .8 4 8
19 2 .9 6 0 3 .5 9 3 3 .9 7 7 4 .2 5 3 4 .4 6 9 4 .6 4 5 4 .7 9 4 4 .9 2 4 5 .0 3 8 5 .2 3 1 5 .4 6 2 6 .1 4 7 6 .7 9 2
20 2 .9 5 0 3 .5 7 8 3 .9 5 8 4 .2 3 2 4 .4 4 5 4 .6 2 0 4 .7 6 8 4 .8 9 6 5 .0 0 8 5 .1 9 9 5 .4 2 7 6 .1 0 4 6 .7 4 0

24 2 .9 1 9 3 .5 3 2 3 .9 0 1 4 .1 6 6 4 .3 7 3 4 .5 4 1 4 .6 8 4 4 .8 0 7 4 .9 1 5 5 .0 9 9 5 .3 1 9 5 .9 6 8 6 .5 7 9
30 2.888 3 .4 8 6 3 .8 4 5 4 .1 0 2 4 .3 0 2 4 .4 6 4 4 .6 0 2 4 .7 2 0 4 .8 2 4 5 .0 0 1 5 .2 1 1 5 .8 3 3 6 .4 1 7
40 2 .8 5 8 3 .4 4 2 3 .7 9 1 4 .0 3 9 4 .2 3 2 4 .3 8 9 4 .5 2 1 4 .6 3 5 4 .7 3 5 4 .9 0 4 5 .1 0 6 5 .7 0 0 6 .2 5 5
60 2 .8 2 9 3 .3 9 9 3 .7 3 7 3 .9 7 7 4 .1 6 3 4 .3 1 4 4 .4 4 1 4 .5 5 0 4 .6 4 6 4 .8 0 8 5 .0 0 1 5 .5 6 6 6 .0 9 3
120 2 .8 0 0 3 .3 5 6 3 .6 8 5 3 .9 1 7 4 .0 9 6 4 .2 4 1 4 .3 6 3 4 .4 6 8 4 .5 6 0 4 .7 1 4 4 .8 9 8 5 .4 3 4 5 .9 2 9
CO 2 .7 7 2 3 .3 1 4 3 .6 3 3 3 .8 5 8 4 .0 3 0 4 .1 7 0 4 .2 8 6 4 .8 3 7 4 .4 7 4 4 .6 2 2 4 .7 9 6 5 .3 0 1 5 .7 6 4
 Appendix 6 (Cont1d .)
Upper Percentage Points of the Studentized Range
ols .01
2 3 4 5 6~ 7 8 9 10 12 15 30 60 100
1 90,.03 135..0 1 6 4 .3 1 8 5 .6 2 0 2 ..2 215.,8 2 27.,2 237..0 245..6 260..0 277..0 326 ..0 3 70 .,1 400..1
2 14,.04 19,.02 2 2 .2 9 2 4 .7 2 26,.63 28..20 29..53 30..68 31..69 3 3 ..4 0 3 5 ..43 41..32 4 6 .,70 50..38
3 8 ,. 261 10, . 6 2 1 2 .1 7 1 3 .3 3 14,.24 15,.00 15..64 16..20 16,.69 17,.53 18,.52 2 1 ,. 44 24..13 25,.99
4 6 ,. 512 8 . 120 9 .1 7 3 9 .9 5 8 1 0 ,. 58 1 1 ..10 1 1 ..55 1 1 ,. 93 1 2 ,. 27 1 2 ,.84 13,.53 15,.57 17..46 18,.77
5 5,.702 6 ,. 9 7 6 7 .8 0 4 8 .4 2 1 8 ,.913 9..321 9.,669 9,.972 1 0 ,. 24 1 0 ,. 7 0 1 1 ,. 24 12 .87 14..39 15 .45

6 5,.243 6 ..331 7 .0 3 3 7 .5 5 6 7,.973 8 ..318 8 ..613 8 ..8 6 9 9,.097 9,.485 9,.951 1 1 ..34 1 2 .,65 13,.55
7 4,.949 5..9 1 9 6 .5 4 3 7 .0 0 5 7..373 7..6 7 9 7..939 8 ..166 8 ,. 368 8 ,. 711 9,.1 2 4 1 0 ,. 3 6 1 1 ..52 1 2 ,.34
8 4,.7 4 6 5..6 3 5 6 .2 0 4 6 .6 2 5 6 ..960 7,.237 7..474 7,.681 7,.863 8 ,. 1 7 6 8 ,. 552 9,.678 1 0 ,.75 1 1 ,. 49
9 4,.5 9 6 5,.428 5 .9 5 7 6 .3 4 8 6 ,.658 6 ..915 7..134 7..325 7,.495 7,.784 8 ,. 132 9 .177 1 0 .,17 1 0 ,. 87
10 4,.482 5..2 7 0 5 .7 6 9 6 .1 3 6 6 ,.428 6 ..6 6 9 6 .,875 7,.055 7,.213 7,.485 7..812 8 .794 9..726 1 0 ,. 39

11 4..392 5..1 4 6 5 .6 2 1 5 .9 7 0 6 .,247 6 .,4 7 6 6 ..672 6 ..842 6 ..992 7,.2 5 0 7,.5 6 0 8 ,.491 9..377 1 0 ,. 00
12 4,.3 2 0 5,.0 4 6 5 .5 0 2 5 .8 3 6 6 ..101 6 ..321 6 ..507 6 ..6 7 0 6 ,. 814 7..0 6 0 7..3 5 6 8 ,. 246 9..094 9,.693
13 4,.2 6 0 4..964 5 .4 0 4 5 .7 2 7 5,.981 6 ..192 6 ..372 6 ,. 528 6 . 667 6 ,.9 0 3 7..188 8 ,. 043 8 ..859 9 .436
14 4..21 0 4..8 9 5 5 .3 2 2 5 .6 3 4 5..881 6 ..085 6 ..258 6 ., 4 0 9 6 ,. 543 6 ,. 772 7,.047 7,.873 8 ..661 9 .219
15 4..168 4..836 5 .2 5 2 5 .5 5 6 5..796 5.,994 6 ..162 6 ..3 0 9 6 ,. 4 3 9 6 ..6 6 0 6 ,. 927 7,.728 8 ..492 9,.035

16 4..131 4 ..7 8 6 5 .1 9 2 5 .4 8 9 5.,722 5..915 6 .,079 6 . 222 6 ,.3 4 9 6 . 564 6 ..823 7,.602 8 ..347 8 ,. 874
17 4,.099 4 ..742 5 .1 4 0 5 .4 3 0 5..659 5..847 6 ..007 6 ,. 147 6 ,. 2 7 0 6 . 480 6 ..734 7,.493 8 ..219 8 ,. 735
18 4..071 4..703 5 .0 9 4 5 .3 7 9 5,.603 5,.788 5..944 6 ,. 081 6 ,. 201 6 ,. 407 6 ..655 7,.398 8 ,. 107 8 .611
19 4..046 4..6 7 0 5 .0 5 4 5 .3 3 4 5,.554 5..735 5.,889 6 . 022 6 ..141 6 ,. 342 6 ..585 7,.313 8 ,. 008 8 .502
20 4,.024 4.,6 3 9 5 .0 1 8 5 .2 9 4 5,.510 5..688 5..8 3 9 5..9 7 0 6 ,. 087 6 ,. 285 6 ,. 523 7,.237 7,.919 8 .404

24 3..956 4..5 4 6 4 .9 0 7 5 .1 6 8 5..374 5..542 5.,685 5..8 0 9 5,.9 1 9 6 ..1 0 6 6 ..3 3 0 7,.001 7..642 8 .097
30 3..889 4,.455 4 .7 9 9 5 .0 4 8 5,.242 5.,401 5..536 5..653 5,.756 5,.932 6 ..143 6 ,. 772 7,.370 7,.796
40 3..825 4,.367 4 .6 9 6 4 .9 3 1 5,.114 5..265 5..392 5..502 5,.5 9 9 5..764 5,.961 6 .547 7,.104 7 .5 0 0
60 3..762 4 ,.2 8 2 4 .5 9 5 4 .8 1 8 4..991 5..133 5..253 5..3 5 6 5,.447 5..601 5,.7 8 5 6 ,. 3 3 0 6 . 843 7 .207
120 3,,702 4..2 00 4 .4 9 7 4 .7 0 9 4,.872 5..005 5..118 5..214 5,.2 9 9 5..443 5,,614 6 ,. 117 6 ,.588 6 .9 1 9
00 3..643 4..1 20 4 .4 0 3 4 .6 0 3 4,.757 4..882 4..987 5..078 5,.157 5..2 9 0 5..4 4 8 5 .911 6 ,.338 6 .636
 A p p e n d ix 6 ( C o n t 'd . )
U pper P e r c e n ta g e P o i n t s o f t h e S t u d e n t iz e d Range
a = .0 01

d f/k 2 3 4 5 6 7 8 9 10 12 15 30 60 100
1 9 0 0 .3 1351. 164 3 . 1856. 2022. 2158. 2272. 2370. 2455. 2600. 2770. 3260. 3701. 4002.
2 4 4 .6 9 6 0 .4 2 7 0 .7 7 7 8 .4 3 8 4 .4 9 8 9 .4 6 9 3 .6 7 9 7 .3 0 1 0 0 .5 1 0 5 .9 1 1 2 .3 1 3 1 .0 1 4 8 .0 1 5 9 .7
3 1 8 .2 8 2 3 .3 2 2 6 .6 5 2 9 .1 3 3 1 .1 1 3 2 .7 4 3 4 .1 2 3 5 .3 3 3 6 .3 9 3 8 .2 0 4 0 .3 5 4 6 .6 8 5 2 .5 1 5 6 .5 3
4 1 2 .1 8 1 4 .9 9 1 6 .8 4 1 8 .2 3 1 9 .3 4 2 0 .2 6 2 1 .0 4 2 1 .7 3 2 2 .3 3 2 3 .3 6 2 4 .5 9 2 8 .2 4 3 1 .6 5 3 4 .0 0
5 9 .7 1 4 1 1 .6 7 1 2 .9 6 1 3 .9 3 1 4 .7 1 1 5 .3 5 1 5 .9 0 1 6 .3 8 1 6 .8 1 1 7 .5 3 1 8 .4 1 21.01 2 3 .4 5 2 5 .1 5

6 8 .4 2 7 9 .9 6 0 1 0 .9 7 1 1 .7 2 1 2 .3 2 1 2 .8 3 1 3 .2 6 1 3 .6 3 1 3 .9 7 1 4 .5 4 1 5 .2 2 1 7 .2 8 1 9 .2 2 2 0 .5 8
7 7 .6 4 8 8 .9 3 0 9 .7 6 8 1 0 .4 0 1 0 .9 0 1 1 .3 2 11.68 1 1 .9 9 1 2 .2 7 1 2 .7 4 1 3 .3 2 1 5 .0 5 1 6 .6 9 1 7 .8 5
8 7 .1 3 0 8 .2 5 0 8 .9 7 8 9 .5 2 2 9 .9 5 8 1 0 .3 2 1 0 .6 4 1 0 .9 1 1 1 .1 5 1 1 .5 6 1 2 .0 6 1 3 .5 7 1 5 .0 1 1 6 .0 2
9 6 .7 6 2 7 .7 6 8 8 .4 1 9 8 .9 0 6 9 .2 9 5 9 .6 1 9 9 .8 9 7 1 0 .1 4 1 0 .3 6 1 0 .7 3 1 1 .1 8 1 2 .5 3 1 3 .8 2 1 4 .7 4
10 6 .4 8 7 7 .4 1 1 8 .0 0 6 8 .4 5 0 8 .8 0 4 9 .0 9 9 9 .3 5 2 9 .5 7 3 9 .7 6 9 10.11 1 0 .5 2 1 1 .7 5 1 2 .9 4 1 3 .7 8

11 6 .2 7 5 7 .1 3 6 7 .6 8 7 8 .0 9 8 8 .4 2 6 8 .6 9 9 8 .9 3 3 9 .1 3 8 9 .3 1 9 9 .6 3 0 10.01 1 1 .1 6 1 2 .2 5 1 3 .0 4
12 6 .1 0 6 6 .9 1 7 7 .4 3 6 7 .8 2 1 8 .1 2 7 8 .3 8 3 8 .6 0 1 8 .7 9 3 8 .9 6 2 9 .2 5 4 9 .6 0 6 10.68 1 1 .7 1 1 2 .4 5

399
13 5 .9 7 0 6 .7 4 0 7 .2 3 1 7 .5 9 5 7 .8 8 5 8 .1 2 6 8 .3 3 3 8 .5 1 3 8 .6 7 3 8 .9 4 8 9 .2 8 1 1 0 .2 9 1 1 .2 7 1 1 .9 7
14 5 .8 5 6 6 .5 9 4 7 .0 6 2 7 .4 0 9 7 .6 8 5 7 .9 1 5 8.1 1 0 8 .2 8 2 8 .4 3 4 9 .6 9 6 9 .0 1 2 9 .9 7 2 1 0 .9 1 1 1 .5 7
15 5 .7 6 0 6 .4 7 0 6 .9 2 0 7 .2 5 2 7 .5 1 7 7 .7 3 6 7 .9 2 5 8 .0 8 8 8 .2 3 4 8 .4 8 3 8 .7 8 6 9 .7 0 3 1 0 .5 9 1 1 .2 3

16 5 .6 7 8 6 .3 6 5 6 .7 9 9 7 .1 1 9 7 .3 7 4 7 .5 8 5 7 .7 6 6 7 .9 2 3 8 .0 6 3 8 .3 0 3 8 .5 9 3 9 .4 7 5 1 0 .3 4 1 0 .9 5
17 5 .6 0 8 6 .2 7 5 6 .6 9 5 7 .0 0 5 7 .2 5 0 7 .4 5 4 7 .6 2 9 7 .7 8 1 7 .9 1 6 8 .1 4 8 8 .4 2 7 9 .2 7 7 10.10 1 0 .7 0
18 5 .5 4 6 6 .1 9 6 6 .6 0 4 6 .9 0 5 7 .1 4 3 7 .3 4 1 7 .5 1 0 7 .6 5 7 7 .7 8 8 8.0 1 2 8 .2 8 3 9 .1 0 6 9 .9 0 4 1 0 .4 8
19 5 .4 9 2 6 .1 2 7 6 .5 2 5 6 .8 1 7 7 .0 4 9 7 .2 4 2 7 .4 0 5 7 .5 4 9 7 .6 7 6 7 .8 9 3 8 .1 5 6 8 .9 5 5 9 .7 3 0 1 0 .2 9
20 5 .4 4 4 6 .0 6 5 6 .4 5 4 6 .7 4 0 6 .9 6 6 7 .1 5 4 7 .3 1 3 7 .4 5 3 7 .5 7 7 7 .7 8 8 8 .0 4 4 8 .8 2 1 9 .5 7 5 10.12

24 5 .2 9 7 5 .8 7 7 6 .2 3 8 6 .5 0 3 6 .7 1 2 6 .8 8 4 7 .0 3 1 7 .1 5 9 7 .2 7 2 7 .4 6 7 7 .7 0 1 8 .4 1 1 9 .1 0 0 9 .5 9 6
30 5 .1 5 6 5 .6 9 8 6 .0 3 3 6 .2 7 8 6 .4 7 0 6 .6 2 8 6 .7 6 3 6 .8 8 0 6 .9 8 4 7 .1 6 2 7 .3 7 5 8.021 8 .6 4 7 9 .0 9 6
40 5 .0 2 2 5 .5 2 8 5 .8 3 8 6 .0 6 3 6 .2 4 0 6 .3 8 6 6 .5 0 9 6 .6 1 6 6 .7 1 1 6 .8 7 2 7 .0 6 7 7 .6 5 1 7 .2 1 4 8 .6 1 8
60 4 .8 9 4 5 .3 6 5 5 .6 5 3 5 .8 6 0 6.022 6 .1 5 5 6 .2 6 8 6 .3 6 6 6 .4 5 1 6 .5 9 8 6 .7 7 4 7 .2 9 9 7 .8 0 2 8 .1 6 1
120 4 .7 7 1 5 .2 1 1 5 .4 7 6 5 .6 6 7 5 .8 1 5 5 .9 3 7 6 .0 3 9 6 .1 2 8 6 .2 0 6 6 .3 3 9 6 .4 9 6 6 .9 6 6 7 .4 1 1 7 .7 2 6
00 4 .6 5 4 5 .0 6 3 5 .3 0 9 5 .4 8 4 5 .6 1 9 5 .7 3 0 5 .8 2 3 5 .9 0 3 5 .9 7 3 6 .0 9 2 6 .2 3 4 6 .6 5 1 7 .0 4 1 7 .3 1 4
 A p pen d ix 7

C u m ulative D i s t r i b u t i o n o f C h i-S q u a re

D e g re e s o f P r o b a b i l i t y o f a G r e a t e r V alu e
Freedom
0 .9 9 5 0 .9 9 0 0 .9 7 5 0 .9 5 0 0 .9 0 0 0 .7 5 0 0 .5 0 0 0 .2 5 0 0 .100 0 .0 5 0 0 .0 2 5 0.010 0 .0 0 5 0 .0 0 1
1 0.02 0.10 0 .4 5 1 .3 2 2 .7 1 3 .8 4 5 .0 2 6 .6 3 7 .8 8 1 0 .8 3
2 0 .01 0 .02 0 .0 5 0.10 0.21 0 .5 8 1 .3 9 2 .7 7 4 .6 1 5 .9 9 7 .3 8 9 .2 1 1 0 .6 0 1 3 .8 2
3 0 .0 7 0.11 0.22 0 .3 5 0 .5 8 1 .2 1 2 .3 7 4 .1 1 6 .2 5 7 .8 1 9 .3 5 1 1 .3 4 1 2 .8 4 1 6 .2 7
4 0.21 0 .3 0 0 .4 8 0 .7 1 1 .0 6 1 .9 2 3 .3 6 5 .3 9 7 .7 8 9 .4 9 1 1 .1 4 1 3 .2 8 1 4 .8 6 1 8 .4 7
5 0 .4 1 0 .5 5 0 .8 3 1 .1 5 1 .6 1 2 .6 7 4 .3 5 6 .6 3 9 .2 4 1 1 .0 7 1 2 .8 3 1 5 .0 9 1 6 .7 5 2 0 .5 2
6 0.6 8 0 .8 7 1 .2 4 1 .6 4 2 .20 3 .4 5 5 .3 5 7 .8 4 1 0 .6 4 1 2 .5 9 1 4 .4 5 1 6 .8 1 1 8 .5 5 2 2 .4 6
7 0 .9 9 1 .2 4 1 .6 9 2 .1 7 2 .8 3 4 .2 5 6 .3 5 9 .0 4 12.02 1 4 .0 7 1 6 .0 1 1 8 .4 8 2 0 .2 8 2 4 .3 2
8 1 .3 4 1 .6 5 2 .1 8 2 .7 3 3 .4 9 5 .0 7 7 .3 4 10.22 1 3 .3 6 1 5 .5 1 1 7 .5 3 2 0 .0 9 2 1 .9 6 2 6 .1 2
9 1 .7 3 2 .0 9 2 .7 0 3 .3 3 4 .1 7 5 .9 0 8 .3 4 1 1 .3 9 1 4 .6 8 1 6 .9 2 1 9 .0 2 2 1 .6 7 2 3 .5 9 2 7 .8 8
10 2 .1 6 2 .5 6 3 .2 5 3 .9 4 4 .8 7 6 .7 4 9 .3 4 1 2 .5 5 1 5 .9 9 1 8 .3 1 2 0 .4 8 2 3 .2 1 2 5 .1 9 2 9 .5 9
11 2 .6 0 3 .0 5 3 .8 2 4 .5 7 5 .5 8 7 .5 8 1 0 .3 4 1 3 .7 0 1 7 .2 8 1 9 .6 8 2 1 .9 2 2 4 .7 2 2 6 .7 6 3 1 .2 6
12 3 .0 7 3 .5 7 4 .4 0 5 .2 3 6 .3 0 8 .4 4 1 1 .3 4 1 4 .8 5 1 8 .5 5 2 1 .0 3 2 3 .3 4 2 6 .2 2 2 8 .3 0 3 2 .9 1
13 3 .5 7 4 .1 1 5 .0 1 5 .8 9 7 .0 4 9 .3 0 1 2 .3 4 1 5 .9 8 1 9 .8 1 2 2 .3 6 2 4 .7 4 2 7 .6 9 2 9 .8 2 3 4 .5 3

400
14 4 .0 7 4 .6 6 5 .6 3 6 .5 7 7 .7 9 1 0 .1 7 1 3 .3 4 1 7 .1 2 2 1 .0 6 2 3 .6 8 2 6 .1 2 2 9 .1 4 3 1 .3 2 3 6 .1 2
15 4 .6 0 5 .2 3 6 .2 6 7 .2 6 8 .5 5 1 1 .0 4 1 4 .3 4 1 8 .2 5 2 2 .3 1 2 5 .0 0 2 7 .4 9 3 0 .5 8 3 2 .8 0 3 7 .7 0
16 5 .1 4 5 .8 1 6 .9 1 7 .9 6 9 .3 1 1 1.9 1 1 5 .3 4 1 9 .3 7 2 3 .5 4 2 6 .3 0 2 8 .8 5 3 2 .0 0 3 4 .2 7 3 9 .2 5
17 5 .7 0 6 .4 1 7 .5 6 8 .6 7 1 0 .0 9 1 2 .7 9 1 6 .3 4 2 0 .4 9 2 4 .7 7 2 7 .5 9 3 0 .1 9 3 3 .4 1 3 5 .7 3 4 0 .7 9
18 6 .2 6 7 .0 1 8 .2 3 9 .3 9 10.86 1 3 .6 8 1 7 .3 4 2 1 .6 0 2 5 .9 9 2 8 .8 7 3 1 .5 3 3 4 .8 1 3 7 .1 6 4 2 .3 1
19 6 .8 4 7 .6 3 8 .9 1 10. 1 2 1 1 .6 5 1 4 .5 6 1 8 .3 4 2 2 .7 2 2 7 .2 0 3 0 .1 4 3 2 .8 5 3 6 .1 9 3 8 .5 8 4 3 .8 2
20 7 .4 3 8 .2 6 9 .5 9 1 0 .8 5 1 2 .4 4 1 5 .4 5 1 9 .3 4 2 3 .8 3 2 8 .4 1 3 1 .4 1 3 4 .1 7 3 7 .5 7 4 0 .0 0 4 5 .3 2
21 8 .0 3 8 .9 0 1 0 .2 8 1 1 .5 9 1 3 .2 4 1 6 .3 4 2 0 .3 4 2 4 .9 3 2 9 .6 2 3 2 .6 7 3 5 .4 8 3 8 .9 3 4 1 .4 0 4 6 .8 0
22 8 .6 4 9 .5 4 1 0 .9 8 1 2 .3 4 1 4 .0 4 1 7 .2 4 2 1 .3 4 2 6 .0 4 3 0 .8 1 3 3 .9 2 3 6 .7 8 4 0 .2 9 4 2 .8 0 4 8 .2 7
23 9 .2 6 10.20 1 1 .6 9 1 3 .0 9 1 4 .8 5 1 8 .1 4 2 2 .3 4 2 7 .1 4 3 2 .0 1 3 5 .1 7 3 8 .0 8 4 1 .6 4 4 4 .1 8 4 9 .7 3
24 9 .8 9 10.86 1 2 .4 0 1 3 .8 5 1 5 .6 6 1 9 .0 4 2 3 .3 4 2 8 .2 4 3 3 .2 0 3 6 .4 2 3 9 .3 6 4 2 .9 8 4 5 .5 6 5 1 .1 8
25 1 0 .5 2 1 1 .5 2 1 3 .1 2 1 4.6 1 1 6 .4 7 1 9 .9 4 2 4 .3 4 2 9 .3 4 3 4 .3 8 3 7 .6 5 4 0 .6 5 4 4 .3 1 4 6 .9 3 5 2 .6 2
30 1 3 .7 9 1 4 .9 5 1 6 .7 9 1 8 .4 9 2 0 .6 0 2 4 .4 8 2 9 .3 4 3 4 .8 0 4 0 .2 6 4 3 .7 7 4 6 .9 8 5 0 .8 9 5 3 .6 7 5 9 .7 0
40 2 0 .7 1 2 2 .1 6 2 4 .4 3 2 6 .5 1 2 9 .0 5 3 3 .6 6 3 9 .3 4 4 5 .6 2 5 1 .8 0 5 5 .7 6 5 9 .3 4 6 3 .6 9 6 6 .7 7 7 3 .4 0
50 2 7 .9 9 2 9 .7 1 3 2 .3 6 3 4 .7 6 3 7 .6 9 4 2 .9 4 4 9 .3 3 5 6 .3 3 6 3 .1 7 6 7 .5 0 7 1 .4 2 7 6 .1 5 7 9 .4 9 86.66
60 3 5 .5 3 3 7 .4 8 4 0 .4 8 4 3 .1 9 4 6 .4 6 5 2 .2 9 5 9 .3 3 6 6 .9 8 7 4 .4 0 7 9 .0 8 8 3 .3 0 8 8 .3 8 9 1 .9 5 9 9 .6 1
70 4 3 .2 8 4 5 .4 4 4 8 .7 6 5 1 .7 4 5 5 .3 3 6 1 .7 0 6 9 .3 3 7 7 .5 8 8 5 .5 3 9 0 .5 3 9 5 .0 2 1 0 0 .4 2 1 0 4 .2 2 1 1 2 .3 2
80 5 1 .1 7 5 3 .5 4 5 7 .1 5 6 0 .3 9 6 4 .2 8 7 1 .1 4 7 9 .3 3 8 8 .1 3 9 6 .5 8 101.88 1 0 6 .6 3 1 1 2 .3 3 1 1 6 .3 2 1 2 4 .8 4
90 5 9 .2 0 6 1 .7 5 6 5 .6 5 6 9 .1 3 7 3 .2 9 8 0 .6 2 8 9 .3 3 9 8 .6 4 1 0 7 .5 6 1 1 3 .1 4 1 1 8 .1 4 1 2 4 .1 2 1 2 8 .3 0 1 3 7 .2 1
100 6 7 .3 3 7 0 .0 6 7 4 .2 2 7 7 .9 3 8 2 .3 6 9 0 .1 3 9 9 .3 3 1 0 9 .1 4 1 1 8 .5 0 1 2 4 .3 4 1 2 9 .5 6 1 3 5 .8 1 1 4 0 .1 7 1 4 9 .4 5
 A p pen d ix 8

PERCENTILE POINTS FOR Q-TEST, FOR EQUAL DEGREES OF FREEDOM v , AND FOR p SAMPLES

v = 1 v = 2 v = 3 v = 4

p .9 9 .9 9 9 .9 9 .9 9 9 .9 9 .9 9 9 .9 9 .9 9 9

3 * * .8 6 3 * .7 5 7 .9 1 9 .6 8 4 .8 2 8

4 .9 2 0 * .7 2 0 .8 9 8 .6 0 5 .7 5 4 .5 4 9 .675

5 .8 2 8 * .6 0 8 .7 7 3 .5 1 2 .6 4 4 .4 4 3 .5 5 2

6 .7 4 4 .9 4 9 .5 3 9 .6 9 0 .4 3 0 .5 4 6 .3 6 9 .461

7 .671 .865 .4 6 9 .6 0 6 .3 7 2 .471 .3 1 8 .3 9 4

8 .6 0 9 .7 9 3 .412 .5 3 7 .325 .411 .2 7 6 .342

9 .5 7 6 .7 5 0 .371 .481 .2 8 7 .3 6 3 .2 4 4 .3 0 0

10 .5 2 8 .6 9 4 .3 3 3 .4 3 3 .2 5 7 .3 2 4 .2 1 8 .2 6 7

12 .4 4 8 .5 9 8 .2 7 6 .358 .211 .2 6 5 .1 7 9 .2 1 7

14 .391 .522 .234 .3 0 3 .1 7 8 .22 2 .151 .181

15 .3 6 5 .4 9 0 .2 1 7 .2 8 0 .1 6 5 .205 .1 4 0 .167

16 .3 4 3 .4 6 0 .202 .261 .1 5 4 .1 9 0 .1 3 0 .1 5 5

18 .3 0 4 .4 0 9 .1 7 8 .228 .135 .1 6 5 .1 1 4 .135

20 .2 7 3 .3 6 7 .1 5 8 .2 02 .12 0 .1 4 6 .101 .1 1 9

22 .2 4 6 .332 .1 4 2 .1 8 0 .1 0 8 .1 3 0 .0 9 0 .1 0 6

24 .2 2 4 .3 0 2 .1 2 9 .1 6 2 .0 9 8 .1 1 7 .0 8 2 .0 9 6

26 .2 0 6 .2 7 6 .1 1 8 .148 .0 9 0 .1 0 7 .0 7 5 .0 8 7

28 .1 9 0 .2 5 4 .1 0 8 .135 .0 8 2 .0 9 8 .0 6 9 .0 8 0

30 .1 7 6 .234 .10 0 .1 2 4 .0 7 5 .0 9 0 .0 6 4 .0 7 4

32 .1 6 3 .2 1 8 .0 9 3 .1 1 5 .0 7 0 .0 8 3 .0 6 0 .0 6 8

36 .1 4 3 .1 8 9 .0 8 2 .1 00 .0 6 2 .0 7 2 .0 5 2 .0 6 0
40 .1 2 7 .1 6 7 .0 7 2 .0 8 8 .0 5 5 .0 6 4 .0 4 7 .0 5 3
45 .111 .1 4 5 .0 6 3 .0 7 6 .0 4 8 .0 5 5 .041 .0 4 6

50 .0 9 8 .1 2 7 .0 5 6 .0 6 7 .0 4 3 .0 4 9 .0 3 7 .041

60 .0 8 0 .10 2 .045 .0 5 3 .0 3 5 .0 3 9 .0 3 0 .0 3 3
64 .0 7 4 .094 .0 4 2 .0 4 9 .0 3 3 .0 3 7 .0 2 8 .031

♦These e n t r i e s e x c e e d e d 1 u s in g t h e a p p ro x im a tin g d i s t r i b u t i o n . S in c e
Q ^ 1, th e y a r e o m it t e d .

401
 A p pen d ix 8 (C o n tin u e d )

v = 5 v = 6 v = 8 v = 10

p .9 9 .9 9 9 .9 9 .9 9 9 .9 9 .9 9 9 .9 9 .9 9 9

3 .631 .7 6 0 .5 9 3 .7 0 8 .5 3 9 .6 3 3 .5 1 2 .5 9 6
4 .4 9 8 .6 0 8 .4 6 1 .5 5 8 .4 1 3 .4 9 0 .3 8 3 .4 4 6
5 .3 9 9 .4 9 0 .3 6 8 .4 4 6 .3 2 8 .3 8 8 .3 0 3 .3 5 1
6 .3 3 4 .4 0 7 .3 0 7 .3 6 8 .271 .318 .2 5 0 .2 8 8
7 .2 8 4 .3 4 5 .261 .3 1 1 .2 3 0 .2 6 8 .21 2 .2 4 2
8 .2 4 6 .2 9 8 .2 2 6 .2 6 8 .1 9 9 .231 .1 8 4 .2 0 9

9 .2 1 7 .261 .1 9 9 .2 3 5 .1 7 6 .20 2 .1 6 2 .1 8 3
10 .1 9 4 .2 3 2 .1 7 8 .2 0 8 .1 5 7 .1 7 9 .1 4 5 .1 6 3
15 .1 2 3 .1 4 5 .1 1 3 .1 3 1 .101 .1 1 3 .0 9 4 .1 0 3

20 .0 9 0 .1 0 4 .0 8 3 .0 9 4 .0 7 4 .0 8 2 .0 6 9 .0 7 5
30 .0 5 8 .0 6 5 .0 5 3 .0 5 9 .0 4 8 .0 5 2 .0 4 5 .0 4 8

40 .0 4 2 .0 4 7 .0 3 9 .0 4 3 .0 3 5 .0 3 8 .0 3 3 .0 3 5

50 .0 3 3 .0 3 6 .031 .0 3 3 .0 2 8 .0 3 0 .0 2 6 .0 2 8

60 .0 2 7 .0 2 9 .0 2 5 .0 2 7 .0 2 3 .0 2 4 .02 2 .0 2 3

v = 12 v = 14 v = 16 v = 20

P .9 9 .9 9 9 .9 9 .9 9 9 .9 9 .9 9 9 .9 9 .9 9 9

3 .4 8 6 .5 5 8 .4 6 6 .5 3 0 .4 5 1 .5 0 8 .4 2 9 .4 7 6
4 .3 6 2 .4 1 5 .3 4 7 .3 9 3 .3 3 5 .3 7 5 .3 1 9 .351
5 .2 8 7 .3 2 6 .2 7 5 .3 0 8 .2 6 5 .2 9 5 .2 5 2 .2 7 6
6 .2 3 6 .2 6 7 .2 2 7 .2 5 3 .2 1 9 .2 4 2 .2 0 9 .2 2 6
7 .201 .2 2 5 .1 9 2 .2 1 3 .1 8 6 .2 0 4 .1 7 8 .191
8 .1 7 4 .1 9 4 .1 6 7 .1 8 4 .1 6 2 .1 7 6 .1 5 4 .1 6 6
9 .1 5 4 .1 7 0 .1 4 8 .1 6 2 .1 4 3 .1 5 5 .1 3 6 .1 4 6
10 .1 3 7 .1 5 2 .1 3 2 .1 4 4 .1 2 8 .1 3 8 .12 2 .1 3 0
15 .0 8 9 .0 9 7 .0 8 6 .0 9 2 .0 8 3 .0 8 9 .0 8 0 .0 8 4
20 .0 6 6 .0 7 0 .0 6 3 .0 6 7 .0 6 2 .0 6 5 .0 5 9 .0 6 2
30 .0 4 3 .0 4 5 .0 4 2 .0 4 3 .0 4 0 .0 4 2 .0 3 9 .0 4 0
40 .0 3 2 .0 3 3 .031 .0 3 2 .0 3 0 .0 3 1 .0 2 9 .0 3 0
50 .0 2 5 .0 2 6 .0 2 4 .0 2 5 .0 2 4 .025 .0 2 3 .0 2 4
60 .021 .022 .02 0 .021 0.20 .0 2 0 .0 1 9 .02 0

2
For v > 60, c a l c u l a t e p v ( p q - l ) and com pare w ith x w ith (p -1 ) degrees o f
freed om in A p pen d ix 7.

402
 Appendix 9
Coefficients {a^ for the W test for normality,

f o r n = 2 ( 1 )5 0

X 2 3 4 5 6 7 8 9 10

1 0 .7 0 7 1 0 .7 0 7 1 0 .6 8 7 2 0 .6 6 4 6 0 .6 4 3 1 0 .6 2 3 3 0 .6 0 5 2 0 .5 8 8 8 0 .5 7 3 9
2 .0 0 0 0 .1 6 7 7 .2 4 1 3 .2 8 0 6 .30 3 1 .31 6 4 .3 2 4 4 .3291
3 .0 00 0 .0 8 7 5 .1401 .1 7 4 3 .1 9 7 6 .2141
4 .0 0 0 0 .0561 .0 9 4 7 .1 2 2 4
5 .00 0 0 .0 3 9 9

X 11 12 13 14 15 16 17 18 19 20

1 0 .5 6 0 1 0 .5 4 7 5 0 .5 3 5 9 0 .5 2 5 1 0 .5 1 5 0 0 .5 0 5 6 0 .4 9 6 8 0 .4 8 8 6 0 .4 8 0 8 0 .4 7 3 4
2 .3 3 1 5 .3 3 2 5 .3 3 2 5 .3 3 1 8 .3 3 0 6 .3 2 9 0 .3 2 7 3 .3 2 5 3 .3 2 3 2 .3 2 1 1
3 .2 2 6 0 .2 3 4 7 .2 4 1 2 .2 4 6 0 .2 4 9 5 .2521 .2 5 4 0 .2 5 5 3 .25 6 1 .2 5 6 5
4 .1 4 2 9 .1 5 8 6 .1 7 0 7 .1 8 0 2 .1 8 7 8 .1 9 3 9 .1 9 8 8 .2 0 2 7 .2 0 5 9 .2 0 8 5
5 .0 6 9 5 .0 9 2 2 .1 0 9 9 .1 2 4 0 .1 3 5 3 .1 4 4 7 .1 5 2 4 .1 5 8 7 .16 4 1 .1 6 8 6
6 0 .0 0 0 0 0 .0 3 0 3 0 .0 5 3 9 0 .0 7 2 7 0 .0 8 8 0 0 .1 0 0 5 0 .1 1 0 9 0 .1 1 9 7 0 .1 2 7 1 0 .1 3 3 4
7 .0 0 0 0 .0 2 4 0 .0 4 3 3 .0 5 9 3 .0 7 2 5 .0 8 3 7 .0 9 3 2 .1 0 1 3
8 .00 0 0 .0 1 9 6 .0 3 5 9 .0 4 9 6 .0 6 1 2 .0 7 1 1
9 .00 00 .0 1 6 3 .0 3 0 3 .04 2 2
10 .00 0 0 .0 1 4 0

X" 21 22 23 24*' 25 26 27 28 29 30

1 0 .4 6 4 3 0 .4 5 9 0 0 .4 5 4 2 0 .4 4 9 3 0 .4 4 5 0 0 .4 4 0 7 0 .4 3 6 6 0 .4 3 2 8 0 .4 2 9 1 0 .4 2 5 4
2 .3 1 8 5 .3 1 5 6 .3 1 2 6 .3 0 9 8 .3 0 6 9 .3 0 4 3 .3 0 1 8 .2 9 9 2 .2 9 6 8 .2 9 4 4
3 .2 5 7 8 .2571 .2 5 6 3 .2 5 5 4 .2 5 4 3 .2 5 3 3 .2 5 2 2 .2 5 1 0 .2 4 9 9 .2 4 8 7
4 .2 1 1 9 .2131 .2 1 3 9 .2145 .2 1 4 8 .21 5 1 .2 1 5 2 .2 1 5 1 .2 1 5 0 .2 1 4 8
5 .1 7 3 6 .17 6 4 .1 7 8 7 .1 8 0 7 .1 8 2 2 .1 8 3 6 .18 4 8 .1 8 5 7 .1 8 6 4 .1 8 7 0
6 0 .1 3 9 9 0 .1 4 4 3 0 .1 4 8 0 0 .1 5 1 2 0 .1 5 3 9 0 .1 5 6 3 0 .1 5 8 4 0 .1 6 0 1 0 .1 6 1 6 0 .1 6 3 0
7 .1 0 9 2 .1 1 5 0 .1201 .1245 .1 2 8 3 .1 3 1 6 .1 3 4 6 .13 7 2 .13 9 5 .1 4 1 5
8 .0 8 0 4 .0 8 7 8 .0941 .0997 .1 0 4 6 .1 0 8 9 .1 1 2 8 .11 6 2 .11 9 2 .1 2 1 9
9 .0 5 3 0 .0 6 1 8 .0 6 9 6 .0 7 6 4 .0 8 2 3 .0 8 7 6 .0 9 2 3 .0 9 6 5 .10 02 .1 0 3 6
10 .0 2 6 3 .0 3 6 8 .0 4 5 9 .0 5 3 9 .0 6 1 0 .0 6 7 2 .0 7 2 8 .0 7 7 8 .08 2 2 .0 8 6 2
11 0.0000 0.0 1 2 2 0 .0 2 2 8 0 .0 3 2 1 0 .0 4 0 3 0 .0 4 7 6 0 .0 5 4 0 0 .0 5 9 8 0 .0 6 5 0 0 .0 6 9 7
12 .0 0 0 0 .0 1 0 7 .020 0 .0 2 8 4 .0 3 5 8 .0424 .0 4 8 3 .05 3 7
13 .00 00 .0 0 9 4 .0 1 7 8 .0 2 5 3 .0 3 2 0 .03 8 1
14 .00 0 0 .0084 .0 1 5 9 .02 2 7
15 .000 0 .0 0 7 6

403
 Appendix 9 Continued
Coefficients (a^ for the W test for normality,

f o r n « 2 (1 )5 0 (c o n t.)

31 32 33 34 35 36 37 38 39

1 0 .4 2 2 0 0 .4 1 8 8 0 .4 1 5 6 0 .4 1 2 7 0 .4 0 9 6 0 .4 0 6 8 0 .4 0 4 0 0 .4 0 1 5 0 .3 9 8 9
2 .2921 .2898 .2 8 7 6 .2 8 5 4 .2 8 3 4 - .2 8 1 3 .27 9 4 .2 7 7 4 .2 7 5 5
3 .2 4 7 5 .2 4 6 3 .2451 .2 4 3 9 .2 4 2 7 .2 4 1 5 .2 4 0 3 .2391 .2 3 8 0
4 .2 1 4 5 .2141 .2 1 3 7 .2 1 3 2 .2 1 2 7 .2121 .2 1 1 6 .2110 .2104
5 .1 8 7 4 .18 7 8 .1 8 8 0 .1 8 8 2 .1 8 8 3 .1 8 8 3 .1 8 8 3 .18 8 1 .1 8 8 0
6 0 .1 6 4 1 0 .1 6 5 1 0 .1 6 6 0 0 .1 6 6 7 0 .1 6 7 3 0 .1 6 7 8 0 .1 6 8 3 0 .1 6 8 6 0 .1 6 8 9
7 .1 4 3 3 .1 4 4 9 .1 4 6 3 .1 4 7 5 .1 4 8 7 .1 4 9 6 .1505 .1 5 1 3 .1 5 2 0
8 .1 2 4 3 .1 2 6 5 .1 2 8 4 .1301 .1 3 1 7 .13 3 1 .1 3 4 4 .1 3 5 6 .1 3 6 6
9 .1 0 6 6 .1 0 9 3 .1 1 1 8 .1 1 4 0 .1 1 6 0 .1 1 7 9 .1 1 9 6 .1 211 .1 2 2 5
10 .0 8 9 9 .0931 .0961 .09 8 8 .1 0 1 3 .1 0 3 6 .1 0 5 6 .1 0 7 5 .1 0 9 2
11 0 .0 7 3 9 0 .0 7 7 7 0 .0 8 1 2 0 .0 8 4 4 0 .0 8 7 3 0 .0 9 0 0 0 .0 9 2 4 0 .0 9 4 7 0 .0 9 6 7
12 .05 8 5 .0 6 2 9 .0 6 6 9 .0 7 0 6 .0 7 3 9 .0 7 7 0 .07 9 8 .0 8 2 4 .0 8 4 8
13 .0 4 3 5 .0 4 8 5 .0 5 3 0 .0 5 7 2 .0 6 1 0 .0 6 4 5 .06 7 7 .0 7 0 6 .0 7 3 3
14 .0 2 8 9 .03 4 4 .0 3 9 5 .0441 .0 4 8 4 .0 5 2 3 .0 5 5 9 .0 5 9 2 .0 6 2 2
15 .0 1 4 4 .0 2 0 6 .0 2 6 2 .0 3 1 4 .0 3 6 1 .0 4 0 4 .0 4 4 4 .0481 .0 5 1 5
16 0.0000 0 .0 0 6 8 0 .0 1 3 1 0 .0 1 8 7 0 .0 2 3 9 0 .0 2 8 7 0 .0 3 3 1 0 .0 3 7 2 0 .0 4 0 9
17 .0 0 0 0 .00 6 2 .0 1 1 9 .0 1 7 2 .02 2 0 .0 2 6 4 .03 0 5
18 .0 00 0 .0 0 5 7 .0 11 0 .0 1 5 8 .0 2 0 3
19 .00 00 .0 0 5 3 .0101
20 .0 0 0 0

41 42 43 44 45 46 47 48 49

i 0 .3 9 4 0 0 .3 9 1 7 0 .3 8 9 4 0 .3 8 7 2 0 .3 8 5 0 0 .3 8 3 0 0 .3 8 0 8 0 .3 7 8 9 0 .3 7 7 0
2 .2 7 1 9 .2701 .2 6 8 4 .2 6 6 7 .2651 .26 3 5 .2 6 2 0 .2 6 0 4 .2 5 8 9
3 .23 5 7 .23 4 5 .2 3 3 4 .2 3 2 3 .2 3 1 3 .2 3 0 2 .2291 .2 2 8 1 .2271
4 .20 9 1 .2085 .20 7 8 .20 7 2 .20 6 5 .2 0 5 8 .2 0 5 2 .2 0 4 5 .2 0 3 8
5 .1 8 7 6 .1 8 7 4 .1871 .1 8 6 8 .18 6 5 .1 8 6 2 .1 8 5 9 .1 8 5 5 .1851
6 0 .1 6 9 3 0 .1 6 9 4 0 .1 6 9 5 0 .1 6 9 5 0 .1 6 9 5 0 .1 6 9 5 0 .1 6 9 5 0 .1 6 9 3 0 .1 6 9 2
7 .1531 .15 3 5 .1 5 3 9 .1 5 4 2 .1 5 4 5 .1 5 4 8 .1 5 5 0 .1551 .15 5 3
8 .13 8 4 .1392 .1 3 9 8 .14 0 5 .1 4 1 0 .1 4 1 5 .1 4 2 0 .1 4 2 3 .1 4 2 7
9 .1 2 4 9 .1 2 5 9 .1 2 6 9 .1 2 7 8 .1 2 8 6 .1 2 9 3 .1 3 0 0 .1 3 0 6 .1 3 1 2
10 . 1123 .1 1 3 6 .1 1 4 9 .1 1 6 0 .1 1 7 0 .1 1 8 0 .1 1 8 9 .1 1 9 7 .1 2 0 5
11 0 .1 0 0 4 0.1020 0 .1 0 3 5 0 .1 0 4 9 0 .1 0 6 2 0 .1 0 7 3 0 .1 0 8 5 0 .1 0 9 5 0 .1 1 0 5
12 .0891 .0 9 0 9 .0 9 2 7 .0 9 4 3 .0 9 5 9 .0 9 7 2 .0 9 8 6 .0 9 9 8 .10 1 0
13 .0 7 8 2 .0 8 0 4 .0 8 2 4 .08 4 2 .0 8 6 0 .0 8 7 6 .0892 .0 9 0 6 .0 9 1 9
14 .0 6 7 7 .0701 .0 7 2 4 .07 4 5 .07 6 5 .0 7 8 3 .0801 .0 8 1 7 .0 8 3 2
15 .0 5 7 5 .0602 .0 6 2 8 .0651 .0 6 7 3 .0 6 9 4 .0 7 1 3 .07 3 1 .0 7 4 8
16 0 .0 4 7 6 0 .0 5 0 6 0 .0 5 3 4 0 .0 5 6 0 0 .0 5 8 4 0 .0 6 0 7 0 .0 6 2 8 0 .0 6 4 8 0 .0 6 6 7
17 .0 3 7 9 .0411 .0 4 4 2 .0471 .0 4 9 7 .0 5 2 2 .0 5 4 6 .0 5 6 8 .0 5 8 8
18 .0 2 8 3 .0 3 1 8 .03 5 2 .0 3 8 3 .04 1 2 .0 4 3 9 .0 4 6 5 .0 4 8 9 .0511
19 .0 1 8 8 .0227 .0 2 6 3 .0 2 9 6 .03 2 8 .0 3 5 7 .03 8 5 .0 4 1 1 .0 4 3 6
20 .0 0 9 4 .0 1 3 6 .0 1 7 5 .02 11 .0 2 4 5 .0 2 7 7 .0 3 0 7 .0 3 3 5 .0361
21 0.0000 0 .0 0 4 5 0 .0 0 8 7 0 .0 1 2 6 0 .0 1 6 3 0 .0 1 9 7 0 .0 2 2 9 0 .0 2 5 9 0 .0 2 8 8
22 .0 0 0 0 .0 0 4 2 .0081 .0 1 1 8 .0 1 5 3 .0 1 8 5 .02 1 5
23 .00 00 .0 0 3 9 .0 0 7 6 .0111 .0 1 4 3
24 .00 0 0 .0 0 3 7 .0071
25 .00 0 0

404
 A p pen d ix 10

_______________ P e rc e n ta g e p o i n t s o f th e W t e s t f o r n = 3 ( 1 ) 5 0 ____________
L evel
n 0.01 0.02 0.05 0.10 0.50 0.90 0.95 0.98
3 0 .7 5 3 0 .7 5 6 0 .7 6 7 0 .7 8 9 0 .9 5 9 0 .9 9 8 0 .9 9 9 1 .0 0 0
4 .687 .7 0 7 .7 4 8 .792 .9 3 5 .9 8 7 .9 9 2 .9 9 6
5 .6 8 6 .7 1 5 .7 6 2 .8 0 6 .9 2 7 .9 7 9 .9 8 6 .991

6 0 .7 1 3 0 .7 4 3 0 .7 8 8 0 .8 2 6 0 .9 2 7 0 .9 7 4 0 .9 8 1 0 .9 8 6
7 .7 3 0 .7 6 0 .8 0 3 .8 3 8 .9 2 8 .972 .9 7 9 .9 8 5
8 .7 4 9 .7 7 8 .8 1 8 .851 .932 .972 .9 7 8 .9 8 4
9 .764 .791 .8 2 9 .8 5 9 .9 3 5 .9 7 2 .9 7 8 .9 8 4
10 .781 .8 0 6 .8 4 2 .8 6 9 .9 3 8 .9 7 2 .9 7 8 .9 8 3
11 0 .7 9 2 0 .8 1 7 0 .8 5 0 0 .8 7 6 0 .9 4 0 0 .9 7 3 0 .9 7 9 0 .9 8 4
12 .8 0 5 .8 2 8 .8 5 9 .8 8 3 .9 4 3 .9 7 3 .9 7 9 .9 8 4
13 .8 1 4 .8 3 7 .8 6 6 .8 8 9 .9 4 5 .9 7 4 .9 7 9 .9 8 4
14 .825 .8 4 6 .8 7 4 .895 .9 4 7 .9 7 5 .9 8 0 .9 8 4
15 .835 .8 5 5 .8 8 1 .901 .9 5 0 .975 .9 8 0 .9 8 4
16 0 .8 4 4 0 .8 6 3 0 .8 8 7 0 .9 0 6 0 .9 5 2 0 .9 7 6 0 .9 8 1 0 .9 8 5
17 .851 .8 6 9 .8 9 2 .910 .9 5 4 .9 7 7 .981 .985
18 .8 5 8 .874 .8 9 7 .9 1 4 .9 5 6 .9 7 8 .9 8 2 .9 8 6
19 .8 6 3 .8 7 9 .901 .9 1 7 .9 5 7 .9 7 8 .9 8 2 .9 8 6
20 .8 6 8 .8 8 4 .9 0 5 .9 2 0 .9 5 9 .9 7 9 .9 8 3 .9 8 6
21 0 .8 7 3 0 .8 8 8 0 .9 0 8 0 .9 2 3 0 .9 6 0 0 .9 8 0 0 .9 8 3 0 .9 8 7
22 .878 .8 9 2 .911 .9 2 6 .961 .9 8 0 .9 8 4 .9 8 7
23 .881 .8 9 5 .9 1 4 .9 2 8 .9 6 2 .981 .9 8 4 .9 8 7
24 .884 .8 9 8 .9 1 6 .9 3 0 .9 6 3 .981 .9 8 4 .9 8 7
25 .8 8 8 .901 .9 1 8 .931 .9 6 4 .981 .9 8 5 .9 8 8
26 0 .8 9 1 0 .9 0 4 0 .9 2 0 0 .9 3 3 0 .9 6 5 0 .9 8 2 0 .9 8 5 0 .9 8 8
27 .8 9 4 .9 0 6 .9 2 3 .935 .9 6 5 .9 8 2 .9 8 5 .9 8 8
28 .8 9 6 .9 0 8 .9 2 4 .9 3 6 .9 6 6 .9 8 2 .985 .9 8 8
29 .8 9 8 .9 1 0 .9 2 6 .9 3 7 .9 6 6 .982 .9 8 5 .9 8 8
30 .9 0 0 .9 1 2 .9 2 7 .9 3 9 .9 6 7 .9 8 3 .9 8 5 .9 8 8
31 0 .9 0 2 0 .9 1 4 0 .9 2 9 0 .9 4 0 0 .9 6 7 0 .9 8 3 0 .9 8 6 0 .9 8 8
32 .9 0 4 .9 1 5 .9 3 0 .941 .9 6 8 .9 8 3 .9 8 6 .988
33 .9 0 6 .9 1 7 .931 .942 .9 6 8 .9 8 3 .9 8 6 .9 8 9
34 .9 0 8 .9 1 9 .9 3 3 .9 4 3 .9 6 9 .9 8 3 .9 8 6 .9 8 9
35 .9 1 0 .9 2 0 .9 3 4 .9 4 4 .9 6 9 .984 .9 8 6 .9 8 9
36 0 .9 1 2 0 .9 2 2 0 .9 3 5 0 .9 4 5 0 .9 7 0 0 .9 8 4 0 .9 8 6 0 .9 8 9
37 .9 1 4 .9 2 4 .9 3 6 .9 4 6 .9 7 0 .9 8 4 .9 8 7 .9 8 9
38 .9 1 6 .9 2 5 .9 3 8 .9 4 7 .971 .9 8 4 .9 8 7 .9 8 9
39 .9 1 7 .9 2 7 .9 3 9 .9 4 8 .971 .9 8 4 .9 8 7 .989
40 .9 1 9 .9 2 8 .9 4 0 .9 4 9 .9 7 2 .9 8 5 .9 8 7 .989
41 0 .9 2 0 0 .9 2 9 0 .9 4 1 0 .9 5 0 0 .9 7 2 0 .9 8 5 0 .9 8 7 0 .9 8 9
42 .922 .9 3 0 .9 4 2 .951 .972 .985 .9 8 7 .9 8 9
43 .9 2 3 .9 3 2 .9 4 3 .951 .9 7 3 .9 8 5 .9 8 7 .9 9 0
44 .9 2 4 .9 3 3 .9 4 4 .9 5 2 .9 7 3 .9 8 5 .9 8 7 .9 9 0
45 .9 2 6 .9 3 4 .945 .9 5 3 .9 7 3 .9 8 5 .9 8 8 .990
46 0 .9 2 7 0 .9 3 5 0 .9 4 5 0 .9 5 3 0 .9 7 4 0 .9 8 5 0 .9 8 8 0 .9 9 0
47 .9 2 8 .9 3 6 .9 4 6 .9 5 4 .9 7 4 .9 8 5 .9 8 8 .9 9 0
48 .9 2 9 .9 3 7 ,9 4 7 .9 5 4 .9 7 4 .9 8 5 .9 8 8 .9 9 0
49 .9 2 9 .937 .9 4 7 .955 .9 7 4 .985 .9 8 8 .9 9 0
50 .9 3 0 .9 3 8 .9 4 7 .955 .9 7 4 .9 8 5 .9 8 8 .9 9 0

405
 A p pen d ix 11

C o e f f i c i e n t s o f O r th o g o n a l P o ly n o m ia ls

n P o ly n o m ia l X=1 2 3 4 5 6 7 8 9 10 IZ2 X

2 L in e a r -1 1 2 2

3 L in e a r -1 0 1 2 1
Q u a d r a tic 1 -2 1 6 3

4 L in e a r -3 -1 1 3 20 2
Q u a d r a tic 1 -1 -1 1 4 1
C u b ic -1 3 -3 1 20 1 0 /3

5 L in e a r -2 -1 0 1 2 10 1
Q u a d r a tic 2 -1 -2 -1 2 14 1
C u b ic -1 2 0 -2 1 10 5 /6
Q u a r t ic 1 -4 6 -4 1 70 3 5 /1 2

6 L in e a r -5 -3 -1 1 3 5 70 2
Q u a d r a tic 5 -1 -4 -4 -1 5 84 3 /2
C u b ic -5 7 4 -4 -7 5 180 5 /3
Q u a r t ic 1 -3 2 2 -3 1 28 7 /1 2

7 L in e a r -3 -2 -1 0 1 2 3 28 1
Q u a d r a tic 5 0 -3 -4 -3 0 5 84 1
C u b ic -1 1 1 0 -1 -1 1 6 1 /6
Q u a r tic 3 -7 1 6 1 -7 3 154 7 /1 2

8 L in e a r -7 -5 -3 -1 1 3 5 7 168 2
Q u a d r a tic 7 1 -3 -5 -5 -3 1 7 168 1
C u b ic -7 5 7 3 -3 -7 -5 7 264 2 /3
Q u a r tic 7 -1 3 -3 9 9 -3 -1 3 7 616 7 /1 2
Q u in t ic -7 23 -1 7 -1 5 15 17 -2 3 7 2184 7 /1 0

9 L in e a r -4 -3 -2 -1 0 1 2 3 4 60 1
Q u a d r a tic 28 7 -8 -1 7 -20 -1 7 -8 7 28 2772 3
C u b ic -1 4 7 13 9 0 -9 -1 3 -7 14 990 5 /6
Q u a r tic 14 -2 1 -11 9 18 9 -11 -21 14 2002 7 /1 2
Q u in t ic -4 11 -4 -9 0 9 4 -1 1 4 468 3 /2 0

10 L in e a r -9 -7 -5 -3 -1 1 3 5 7 9 330 2
Q u a d r a tic 6 2 -1 -3 -4 -4 -3 -1 2 6 132 1/2
C u b ic -4 2 14 35 31 12 -12 -3 1 -3 5 -1 4 42 8580 5 /3
Q u a r t ic 18 -22 -1 7 3 18 18 3 -1 7 -2 2 18 2860 5 /1 2
Q u in t ic -6 14 -1 -1 1 -6 6 11 1 -1 4 6 780 1/10

406
 A p pen d ix 12

V a r ia n c e o f Means and C o n t r a s t s
fo r

M ixed M odels

A ssum ing t h e c o r r e l a t i o n s among a l l th e com ponents in th e m od els a r e z e r o , th e

f o l l o w i n g d e r i v a t i o n s a r e a p p r o p r ia t e [A nd erson and B a n c r o ft (1 9 5 2 ), p . 3 4 0 -3 4 4 ]:

1. V a r ia n c e o f th e colum n mean f o r a m ixed two way c l a s s i f i e d a n a l y s is

C o n s id e r th e la y o u t

1 2 j . . . . c
n
yi..

yi..
Rows (p )
hi-
1
yr ..

r.i. y. 2 . ••• y.y -------------- y -c- y...

i n w hich th ere is a two way c l a s s i f i c a t i o n w ith more than one o b s e r v a t i o n per c e l l ,

n. If y is fix e d and p i s random in th e f o l l o w i n g m o d e l:

Y. = y + p. + y . + (p y ) . . + e . .. ,
13k l *3 v" 1 3 13k ’

le t pa be ( 0 , a R) ,
2 (p y )^
2
b e ( 0 , a RC) and e . j k b e ( 0 ,a ) .
2

The E (Y j) = Yj s i n c e y . i s f i x e d and E (y ) = y . We w ish t o f i n d th e

v a r ia n c e o f t h e colum n ( f i x e d ) mean. From t h e e q u a tio n a b o v e , th e j th colum n mean

is :
r n
I X Y. n ][ p .
Y . = i=l k=l
■ ljk = p + — i------1 + r n Yi
L
•3 • rn m rn

n \ (p y ) • • I \ e.
i= l l> .i = l k=l

S in c e E (Y j) * Yy E (y ) = y , and E ( a l l random v a r i a b l e s in th e m o d e l) = 0

E(CP) * 0

E (Y -j.) = M + Yj

407
The variance of the estimate of the column mean is:

Z p. Z (p y ) . . Z Z £.
- 2 (i * 1A i . . ljk 2
'J ik
V (Y -j-) = E[Y
• r - E(v,j,)1 =E| T - (i- 7 - * V |
bu t E(p?) = a2, E (p y )?j = o ^ c > and E (e i ;j k) 2 = ^ •
Hence // 7
2 . 02 (C PI)
CCP1) v
+ ... + p + C ro ss p r o d u c t s 1\
V(Y
■i ■ -?- - - - - - - )
(PY)j j ♦ . . . + (PY)^ ♦ CP2
2
r

( i i + ... + e2 . + CP A x
*( — 7 ^ — 7 * cp<l

^r a R r 0 RC r n a2 aR aRC a2
and V(Y . ) » - V 1 + — ~ + ■■ ■ = -
*J • I 2 2 22 r r rn
J yr r r n
2. V a r ia n c e o f th e mean i n s i d e th e t a b l e .

N ext we want t o f i n d t h e v a r i a n c e o f t h e mean i n s i d e t h e t a b l e (w ith

th e a ssu m p tion t h a t (P Y )jj i s random , th e v a lu e o f t h i s mean i s q u e s t i o n a b l e ) :

n n
\ Y. ., . f \ e . ..
k = i 1Jk npi nYj n (P Y )i j k=i 1Jk
Y.i .r =---------------
^ = u + ------ + — ** + ■ — +---------

E (Y ..J = v + Yj

and

V (Y . . ) = E [Y . . - E (Y . . ) ]
- 2 (
= E I p . + ( p y ) . . + — -------- J
k=lEi:ik {
1 1i j3 • 1 ij • v ( i VKWi j n j

E {pi * (p Y )i j * ~ i± 1— * CP6 )

2 2 no2
°R + °RC + I "
n
F in a lly
2
2 2 .a

408
 3. Variance of £ contrast.

C o n s id e r t h e com p a rison i n v o lv i n g o n l y two colum n means Y*p* and

Y «q * , then th e d i f f e r e n c e i s r e a l l y th e c o n t r a s t :
E p. Z (p y )• Z Z z . .
( i 1 i ip i k ip k
(Y•p. - Y
-q.) = lu + + y + +-— ----------
v ( r p r rn

Z p. Z ( py) . Z Z z. .
i 1 i it 1Rki
r “ Yq r ” rn j
2 (P Y )ip E (P Y )i q
(Y
•p.
- Y
»q»
) = f CY
r Tp
- Y ) + -
q

Z Z e. , Z Z z. .
i k ip!c ■k iqk
rn ~ rn j

The v a r ia n c e o f t h i s c o n t r a s t i s

V(Y - Y ) = E[ (Y - Y ) - E(Y - Y ) ] 2.
1 *p* - q* 1 -p * -q- - p* #q* J
Now

E (V .p . - V .q . ) - (Yp - Yq )

and
/Z ,r (p y ) • y (p y ) • v
E \ //v
z r
Z c
z. , Z r
r Z c
z-
. , \\ ^

V(Y - Y
•p* -q

i (py),; M py^ *1 .^ I I «lqj|

■ E{
1
i— r j— + i— r r - * l Jrtn j - + iJT
rn
T " ♦ CP •
I t fo llo w s th a t 2 2

„fv
V(V -v
V i
)-' r°R
—C * r°RC * ~T~2
2~ rn°2 . ~2~2
rn<y2
r n r r r n r n

_ 2gRC + 2o^
r m
2 2
Knowing t h a t E.M .S. (R .C .) = naRC + a , th e e s tim a t e o f t h i s v a r ia n c e is

V(Y y •)2 (mean squ are RC)

»p* ~ *q»J “ m

The e x p e r im e n te r can a lw ays h a n d le th e f i x e d m odel p rob lem b e ca u s e th e

s
✓n
and d ep en d in g upon what i s f i x e d and what i s random , th e sta n d a rd e r r o r may a c t u a l l y

409
be q u i t e d i f f e r e n t from th e v a r ia n c e o f th e c o n t r a s t . To b e s a f e f o r m ixed m o d e ls ,

th e e x p e r im e n te r must d e r i v e t h e a p p r o p r ia t e s ta n d a rd e r r o r and v a r ia n c e o f th e

c o n t r a s t from t h e m od el.

O f c o u r s e t h e r e i s no i n t e r e s t in means f o r th e random m o d e l; c o n s e q u e n t ly

t h e r e i s no p rob lem f o r th e c o m p le t e ly random c a s e .

When i n v e s t i g a t i n g i n d i v i d u a l com p a rison s in a m ixed m o d e l, t h e r e r e s u l t s a re

e s p e c ia lly a p p lic a b le . C o n s id e r th e r e p e a t e d t - t e s t s

| y.p . - y . q. I > t a ( t f s - ) : ( f o r a f i x e d m o d e l),

(/2 s -) i s r e a l l y th e s ta n d a rd e r r o r o f th e c o n t r a s t and we m e r e ly r e p l a c e t h i s

com ponent by

f o r t h e m ixed m od el. O f c o u r s e th e e x p e r im e n te r w i l l t h o r o u g h ly i n v e s t i g a t e th e

i n t e r a c t i o n b e f o r e exa m in in g and e v e n t u a l l y i n t e r p r e t i n g t h i s c o n t r a s t among th e

tre a tm e n t means.

A s im i l a r p r o c e d u r e may be used on th e v a r io u s o t h e r t e s t s . For exam ple th e

Newman-Keuls t e s t u ses

Rk = <*f) ( s - ) •
’ ( f o r a f i x e d m o d e l).

But th e sta n d a rd e r r o r f o r t h e c o n t r a s t i s

Hence

qa (k, d f)
(N ote t h a t when k = 2 , = t ) and, f o r th e m ixed m o d e l, s u b s t itu tin g
./T a

(mean sq u a re RC)

fo r / 2 s - , we g e t

qa ( k , d f )
sq u a re RC)
7T~
or

410
 RC
Rk • rn

and / mean sq u a re RC ca n fce u je d

r rn

d i r e c t l y to r e p la c e s - . (N ote t h i s i s n o t th e u s u a l s - f o r a f i x e d m odel and

2
r e s u l t s o n l y b e ca u s e o f t h e c o n t r a s t c a u s in g aR t o d ro p o u t ) .

N ested F a c t o r i a l f o r S u b s e c t io n 6 . 2 . 3 .

D e r iv a t io n o f e s tim a t e d sta n d a rd e r r o r o f r a t i o n means from

E q u a tion (6 .2 .3 ):

y... = y R. + Hr v + + W. + RW., + H W ♦ e .......

ijk p l (i)j (lj) k lk (i)jk (ljk )

w h ere: R^ and a re f i x e d and a re random ,

jh HW i , j^(U)
Yi.. 8 8
5 55 8 5
I \k I RW.k
v D“l k I I HW I I e 1
k=l k k=l j =l k=l j =l k=l UJKJ I
+ 5 * 40 + 40 J
U 7~0
S in c e weeks (W^) a r e f i x e d and hens [H ^ ^ ] a re random , th e summation

o f t h e m ixed i n t e r a c t i o n o v e r t h e s u b s c r ip t k i s d e fi n e d as z e r o .

THu* (j i Ji^cnjkV 0 •
E (y . ) = y + R. 8
u „
8 ~
8 5 ,

v t/i..) ■ ElXj.. - EC/i..)]* ■ eI ^ * *J■ To

411
 5. E x t e n s io n s .

T h is a p p roa ch o f f i n d i n g th e v a r ia n c e o f means and c o n t r a s t s can be

u sed f o r a l l f i x e d o r m ixed ANOVA m od els i f t h e a ssu m p tion o f no c o r r e l a t i o n is

a p p r o p r ia t e f o r a l l o f t h e com ponents in a l l o f t h e ANOVA e q u a t i o n s .

412
 AUTHOR INDEX

Addelman, S., 126, 153, 272, Conner, W. S., 295, 301, 317,
274, 278 319, 325
Anderson, R. L., 27, 28, 29, 38, Cox, D. R., 85, 93, 94, 121, 330,
52, 78, 111, 117, 118, 121, 331, 332, 386
156, 160, 180, 213, 223 Cox, G. M., 94, 121, 151, 153,
Anderson, V. L., 40, 78, 87, 93, 209, 216, 219, 223, 302, 320,
124, 153, 165, 180, 184, 212, 325, 348, 349, 386
223, 275, 276, 278, 342, 386
Daniel, C., 206, 209, 261, 270,
Bancroft, T. A., 10, 27, 28, 29, 271, 278
36, 38, 52, 78, 105, 111, 117, Davies, 0. L. (Ed.), 104, 121,
118, 121, 141, 153, 186, 208, 237, 245, 251, 252, 255, 256,
213, 223 278, 281, 301, 353, 370, 373,
Bartlett, M. S., 19, 20, 25, 27, 386
38 Dixon, W. J., 7, 23, 38
Beeson, J., 131, 153, 184, 189, Draper, N. R., 27, 28, 38, 374,
208 379, 386
Behnken, D. W., 367, 386 Dykstra, 0., 370, 386
Bennett, C. A., 52, 78, 125,
153, 208 Edwards, A. L., 104, 111, 121,
Berger, P. D., 263, 278 180, 213, 215, 223
Bliss, C. I., 208 Eisenhart, C., 7, 38, 159, 180
Booth, K. H. V., 330, 331, 332,
386 Federer, W. T., 185, 209, 302,
Bowman, K. 0., 7, 39 320, 321, 322, 323, 324, 325
Box, G. E. P., 17, 38, 272, 278, Finney, D. J., 87, 93, 120, 121
348, 349, 350, 353, 367, 370, Fisher, R. A., 87, 93, 151, 153,
374, 379, 385, 386 209, 211, 218, 223, 224
Bozivich, H., 36, 38, 105, 121, Foster, L. A., 20, 38
141, 153, 186, 208 Franklin, N. L., 52, 78, 125,
Bratcher, T. L., 7, 38 153
Brooks, S. H., 333, 361, 362,
386 Geisser, S., 166, 180
Burman, J. P., 330, 386 Gorman, J. W., 335, 340, 386
Burr, I. W., 20, 38 Greenhouse, S. W., 166, 180

Carmer, S. G., 10, 38 Hader, R. J., 95, 121

Chen, Mrs. H. J., 25, 39 Hadley, W. 0., 40, 78, 275, 276,
Chew, V. (Ed.), 186, 209 278
Cochran, W. G., 19, 21, 38, 39, Harter, H. L., 186, 209
52, 55, 78, 94, 117, 121, 151, Hartley, H. 0., 36, 38, 105, 138,
153, 209, 216, 219, 223, 302, 141, 153, 186, 208
320, 325, 348, 349, 386

413
414 AUTHOR INDEX

Hebble, T. L., 370, 386 Newman, D., 10, 39

Hext, G. R., 366, 387
Hicks, C. R., 47, 78, 90, 93, Ostle, B., 25, 27, 29, 39, 53, 78,
111, 122, 125, 153, 162, 180, 87, 90, 93, 117, 122, 130, 153,
218, 224, 226, 251, 252, 281, 209
282, 385, 386
Himsworth, F. R., 366, 387 Peng, K. C., 252, 278
Hinman, J. E., 335, 340, 386 Plackett, R. L., 330, 386
Hoel, D. G., 7, 39 Prairie, R. R., 97, 122
Hotelling, H., 370, 386
Hudson, W. R., 40, 78, 275, 276, Robson, D. S., 28, 39
278
Hunter, J. S., 216, 224, 272, Satterthwaite, F. E., 117, 122,
278, 348, 350, 379, 386 327, 386
Scheffd, H., 335, 341, 386
John, P. W. M., 94, 111, 122, Shapiro, S. S., 25, 39
226, 251, 252, 274, 278, 370, Smith, H., 27, 28, 38
386 Snedecor, G. W., 21, 39, 209
Johnson, N. L., 261, 278 Snee, R. D., 347, 387
Spendley, W., 366, 386
Kastenbaum, M. A., 7, 39 Stapleton, J. H., 329, 386
Kempthorne, 0., 122, 181, 187, Steel, R. G. D., 90, 93, 209
209, 213, 215, 216, 218, 224, Stein, C., 2, 5, 39
226, 228, 229, 231, 242, 243, Stone, L. A., 367, 386
245, 251, 252, 255, 268, 278, Swanson, M. R., 10, 38
281, 282, 293, 301, 302, 320,
322, 323, 325 Torrie, J. H., 90, 93, 209
Kendall, D. G., 20, 38 Tukey, J. W., 45, 78
Kennard, R. W., 367, 386
Kennedy, T. W., 40, 78, 275, Wilk, M. B., 25, 39, 126, 153,
276, 278 213, 215, 224
Keuls, M., 10, 39 Wilson, K. B., 349, 353, 370, 386
Kurkjian, B., 325 Winer, B. J., 105, 111, 122, 166,
167, 172, 209, 243, 251, 281,
Leone, F. C., 261, 278 301
Levin, J. R., 67, 78, 110, 122, Wood, F. S., 206, 209
277, 278 Woodall, R., 325
Li, Jerome, C. R., 120, 122
Yates, F., 87, 93, 185, 209, 211,
Mandel, J., 122, 172, 180 218, 224, 237, 251, 261, 278
Marascuilo, L. A., 67, 78, 110, Youden, W. J., 216, 224
122, 277, 278 Young, S., 317, 319, 325
Margolin, B. H., 281, 284, 301
Massey, F. J., Jr., 7, 23, 38 Zelen, M., 295, 301
McLean, R. A., 342, 386 Zimmer, W. J., 7, 38, 97, 122
Mitchell, T. J., 370, 386
Moran, M. A., 7, 38
Morrison, D. F., 165, 166, 180
Myers, R. H., 27, 29, 39, 353,
370, 374, 376, 386
 SUBJECT INDEX

Algorithm, expected mean Central composite design, 353

squares, 52 Centroid, 341, 343
Alias, 254 Comparisons
Analysis contrasts, 29
canonical, 374 individual (Newman-Keuls), 10,
outline before final design, 67
87 least significant range, 11
Additivity, non, 45, 51 multiple, 10
ANOVA (Analysis of Variance) Completely randomized design, 1,
calculations 94
balanced incomplete block, Composite design, 353
249 Confidence interval, 12
completely randomized, 106, Confounding, 3, 42, 44, 95, 248
113 Confounding interactions, 254,
cross over designs, 220 285
one-way, 8 Conservative tests, 102, 138
nested, 158 Contrasts, 29
nested factorial, 168 Correlated errors, 165
randomized complete block, Cross over design, 219
142 Cycle (for EVOP), 379
split plot, 194
two-way, 64 Data analysis, 90
partitioning sums of squares, Data collection, 99
35 Defining
Arc Sin Sp transformation, 25 contrasts, 255
Asymmetrical factorial effects, 260
experiments, 272 equations, 268, 315
Design
Balanced information, 290 efficiency of, 151
Blocks efficient, 105
balanced incomplete, 247 experimental, 225
complete, 124 layout, 179
fixed, 125 of experiment, 97
generation for 2n, 243, 246 outline analysis, 87
generation for 3n, 286, 288 sequential efficiency, 366
incomplete, 241, 288, 302 Designs
random, 138 balanced incomplete block, 247
Blocking concept, 86 central composite, 353
completely randomized, 94
Calculations of interaction SS, composite, 353
195 cross over, 219
Canonical analysis, 374 EVOP, 369

415
416 SUBJECT INDEX

Designs (continued) whole plot estimate, 184, 193

extreme vertices, 342 Type I (a), 4
first order, 348 Type II (6), 4
fixed, 327 Expected mean square algorithm,
fractional factorial, 269 52
Graeco-Latin, 218 Experiment, complete factorial,
Hierarchical, 154 98, 225
incomplete block, 241 Experimental
Knut Vik square, 218 unit, 1, 2, 94, 98
Latin square, 210, 298 error, 99, 105
lattice, 302, 319 Experimentation
lattice, double simple, 322 purpose, 79
lattice (one restriction), 320 requirements, 81
lattice, rectangular, 323 EVOP, 369, 379
lattice, simple, 321 Extreme vertices design, 342
lattice square (two
restrictions), 324 Factor
mixture, 335 fixed, 7, 52, 61
nested, 154 level spacing, 119
nested factorial, 162 random, 7, 52, 61
nonsequential, 327 Factorial
orthogonal composite, 353, 356 asymmetrical experiment, 272
random, 333 block generation for 2n, 243,
randomized complete block, 123 246
repeated measures, 165 block generation for 3n, 286,
resolution, 272 288
rotatable, 348 block generation for mixed, 312
rotatable composite, 353, 355 complete experiment, 98, 225,
second order, 349 352
sequential, 327, 361 experiments, 2n, 225
simplex, 362 experiments, 3n, 280
simplex centroid, 341 mixed, 302, 312
simplex lattice, 335 symmetrical, 272
simplex-sum, 367 First order designs, 348
split plot, 181 Fixed
split-split plot, 199 designs, 327
supersaturated, 330 factors, 7, 52, 61
univariate, 361 Fractional
Youden square, 218 factorials, 252, 350
Dot Notation, 13 factorial 2n booklet, 269
factorial 3n booklet, 295, 301
Equations, defining, 245 factorial mixed booklet, 319,
Errors 325
correlated, 165 factorial 2m3n, 269, 317
definition, 100
experimental, 99, 105 Generalized interaction, 245,
high order interaction, 260
estimate of, 226 Generation
pure, 34 blocks for 2n, 243, 246
pooling, split plot, 186 blocks for 3n, 286, 288
restriction, 87, 124 blocks for mixed factorials,
312
SUBJECT INDEX 417

Graeco-Latin square, 218 Main effects, concept, 40

Mean squares, expected, 44
Half-normal plots, 261 Mixed factorials, 302
Hidden replication, 98 Mixture designs, 335
Hierarchical design, 154 Model
High order interaction, an fixed, 55
estimate of error, 40, 226 mixed, 55
Homogeneity of variance (tests) random, 56
Bartlett-Kendall log s2, 19,
21 Nested design, 154
Bartlett test, 19, 20 Nested factorial design, 162
Burr-Foster Q test, 20, 22 Newman-Keuls test, 10,68
Hyperpolyhedron, 343 Non-additivity, 45, 51
Non-sequential design, 327
Identity relationship, 255 Notation
Incomplete block design, 241 dot, 13
Inference space, 7, 57, 84, factorial, 227
92, 94, 95, 101, 123, 144, 147 orthogonal polynomial, 75
Information
balanced, 290 Orthogonal
equal, 277 composite designs, 353, 356
most, 367 contrasts, 30
on effects, 242, 244 polynomials, 28
partial, 247
Interaction Partial information, 247
calculation sum of squares, Partitioning interaction, 73
195 Phase (for EVOP), 379
calculation sum of squares Polyhedron, hyper, 343
(linear x linear), 49, 51, Pooling
76, calculations, 105, 136
generalized, 245, 260 concept, 105, 141
high order, 40, 226 split plot error, 186, 190
interpretation of three factor, Power of a test (1-8), 10
275 Pseudocomponent, 341
partitioning, 73 Pseudofactors, 303
two factor concept, 40 Pure error, 34
Intrablock subgroup, 245
R2, 32
Knut Vik square, 218 Random balance design, 327
Random design, 333
Lack of fit, 34 Random factor, 7, 52, 61
Layout, nested factorial, 179 Randomization, 86, 87, 92, 132
Latin square design, 210, 298 Randomized complete block design
Lattice design, 302, 319 (RCBD), 123
Levels Reciprocal transformation, 25
different primes, 307 Regression
product of primes, 310 analysis, 27
same primes, 302 R2, 32
Linear x linear interaction, Repeated measures, 165
calculations, 49, 51, 76 Replication, 92, 98, 272
Log transformation, 25 Resolution (design), 272
Response surface, 27, 73, 326
418 SUBJECT INDEX

Response variable, 84 homogeneity of variances

Restriction error, 87, 124, 125, Bartlett, 19, 20 2
128, 130, 149, 184 Bartlett-Kendall log s , 19,
Robust for a, 17 21
Rotatable composite design, 353, Burr-Foster Q, 20, 22
355 Newman-Keuls, 10
Rotatable design, 348 Shapiro-Wilk W, for normality,
Rules for transformations, 17 25
Three factor interaction,
Sample size, 2, 5 interpretation, 275
Satterthwaite1s approximation, Transformations
117 arc sin i/p, 25
Second order design, 349 log, 25
Sequential designs, 327, 361 negative correlation between
Significance tests mean and variance, 25
F, 15 reciprocal, 25
F’, 117 rules for, 17
t, 2 square root, 24
Simplex Treatment
definition, 335 combinations, 40, 98, 124
centroid design, 341 combinations, generation of,
design, 362 229
lattice design, 335 effects, 228, 231
sum design, 367 Two factor interaction concept,
Spacing factor levels, 119 40
Split plot design, 181 Type I (a) error, 4
Split-split plot design, 199 Type II (3) error, 5
Square root transformation, 23
Standard error Unit, experimental, 1, 2, 95, 98
of a mean, 11 Univariate design, 361
of a treatment mean, 62
Steepest ascent method, 370 Whole plot error estimate, 184,
Sum of squares, calculation for 193
interaction, 195
Supersaturated design, 330 Within error, 2 levels, 197

Tests Yates method, 237

conservative, 102 138 Youden square design, 218