A Multiple Comparison Procedure for Comparing Several Treatments with a Control

Author(s): Charles W. Dunnett

Source: Journal of the American Statistical Association , Dec., 1955, Vol. 50, No. 272
(Dec., 1955), pp. 1096-1121
Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association

Stable URL: https://www.jstor.org/stable/2281208

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to
digitize, preserve and extend access to Journal of the American Statistical Association

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
A MULTIPLE COMPARISON PROCEDURE FOR COMPARING
SEVERAL TREATMENTS WITH A CONTROL

CHARLES W. DUNNETT*
American Cyanamid Company

I. INTRODUCTION

A COMMON problem in applied research is the com

I ments with a control or standard. Such a situatio
example, when an agronomist tests the effects on crop yield of the addi-
tion of chemicals to the soil, or when a pharmacologist assays drug
samples to determine their potencies. In designing an experiment to
measure the effects of such treatments, it is often desirable to include
in the experiment a control in the form of either a dummy treatment,
to measure the magnitude of the experimental response in the absence
of the treatments under investigation, or some recognized standard
treatment. Sometimes past experience with the control will suffice, but
often this cannot be relied upon due to altered environmental con-
ditions. Thus the agronomist may leave a few of his experimental plots
untreated for comparison with the treated plots, and the pharmacolo-
gist may measure the response to a standard drug preparation of
known potency concomitantly with the test samples in order to esti-
mate the potencies of the latter.
We will consider the case where the numerical results of an experi-
ment performed to compare p treatments with a control can be sum-
marized in the form of a set of numbers XO, X1, ... , Xp and s, where
the X's are means of p+l sets of observations which are assumed to be
independently and normally distributed, Xo referring to the control and
Xi to the i-th treatment (i= 1, * * *, p), and s is an independent esti-
mate of the common standard deviation of the p+1 sets of observa-
tions. This paper presents a procedure for making confidence state-
ments about the true (or expected) values of the p differences zi- Xo,
the procedure having the property that the probability of all p state-
ments being simultaneously correct is equal to a specified value, P.
Tables have been computed which enable the procedure to be used by
the experimenter for P =.95 or .99 and p=1(1)9. When the numbers
of observations in each set are equal, the tables enable the experimenter
* I should like to express my appreciation to Frank Wilcoxon for suggesting this problem and for
his constant encouragement throughout the course of the investigation. I am grateful for this oppor-
tunity to acknowledge also my indebtedness to Robert Bechhofer and to Milton Sobel. Were it not
for my good fortune to be associated with them while working on an Air Force contract at Cornell
University in 1952-3, this paper could not have been written.
I am indebted to Robert E. Bechhofer also for making the tables in reference [10] available to me.

1096

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1097

to set one-sided upper (or lower) confidence limits on the true values of
the p differences Xi- Xo such that the probability is P that all p true
values will actually be less than the upper limits set, or to set two-sided
confidence limits on the true values of the p differences Xi - Xo such
that P is a lower bound to the probability that all p true values will
actually be between the limits set. If the numbers of observations in
each set are unequal, the tables may still be used but the associated P
will be only approximate. The tables may also be used to set joint
confidence limits on the potencies of p drugs relative to a common
standard, the associated P being approximately equal to the prob-
ability that all statements will be correct when one-sided limits are set
and approximately a lower bound to this probability when two-sided
limits are set.
The problem of multiple comparisons with a control is a special case
of the more general multiple comparisons problem considered by
Tukey [17] and Scheff6 [16]. Tukey's procedure based on the Student-
ized range and Scheff6's procedure based on the F-distribution enable
the experimenter to make any number of comparisons among a set of
sample means with the assurance that the probability of all confidence
statements being correct will be equal to or greater than a specified
value. When the experimenter only wishes to make comparisons be-
tween one of the means and each of the others, as in the case when one
of the means represents a control, use of the Tukey or Scheff6 pro-
cedure would result in confidence limits which are wider than necessary.
The procedure described in this paper results in narrower confidence
limits for the p comparisons Xi- Xo than either the Tukey or the
Scheff 6 procedure.
In an earlier paper, Roessler [15] considered the problem of multiple
comparisons involving a control. However, he assumed that the p com-
parisons Xi - Xo were independent which is incorrect since they all
have Xo in common. In the present paper it is shown that, to obtain
simultaneous confidence limits on the Xi - Xo, the multivariate ana-
logue of Student's t-distribution defined by Dunnett and Sobel [4] is
encountered. This same distribution was involved in a multiple de-
cision procedure for ranking population means described by Bechhofer
et al., [2], to which Tables la and lb of the present paper are applicable.
A multiple decision procedure for comparing several experimental
categories with a control was formulated by Paulson [11]. Tables la
and lb of the present paper are also applicable to Paulson's procedure.
The procedure described in the present paper may also be considered
as a multiple decision procedure; it is compared with Paulson's in
Section VII below.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1098 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

For the benefit of those who may be interested primarily in applica-

tions, the procedure is illustrated by two examples in Section II. The
main part of the theory is given in Section III with a description of the
construction of the tables in Section IV. In Section V, the question of
the optimum allocation of available experimental resources between
the control and the p treatments is considered. In Section VI, the pro-
cedure is applied to the problem of estimating the potencies of p drug
samples relative to a common standard.

II. EXAMPLES

(a) The following example was adapted from one given by Villars
[18]. The data represent measurements on the breaking strength of
fabric treated by three different chemical processes compared with a
standard method of manufacture.

Breaking Strength (lbs.)

Standard Process 1 Process 2 Process 3

55 55 55 50
47 64 49 44
48 64 52 41

Means 50 61 52 45
Variances 19 27 9 21

Here, p =3 and N =3. The average variance is S2 = 19, which is an

estimate of the common variance of the four sets with (p+ 1) (N-1) = 8
degrees of freedom. It could be calculated directly as follows:
552 + 472 + 482 + 552 + . + 412 - 3(502 + 612 + 522 + 452)
n2-
8
152
= = 19.
8

The standard deviation is s -/19 = 4.36 and the estimated standard

error of a difference between two means is s\2/N = 4.36-\2/3 = 3.56.
The quantity which must be added to and/or subtracted from the
observed differences between the means to give their confidence limits
has been called by Tukey [17] an "allowance" and is given by A
=tsV2/N, where t is obtained from Table 1 if one-sided limits are
desired or from Table 2 if two-sided limits are wanted. For p=3 and
d.f. =8, 1 = 2.42 for one-sided limits and t = 2.94 for two-sided limits for
P=95%. Analogous values of t can be determined from the tables if
P =99% confidence is required.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1099

For one-sided limits, the allowance is A (2.42) (3.56) = 9 and the

experimenter can conclude that:
(i) The breaking strength using process 1 exceeds the standard by
at least 61-50-9=2 lbs.
(ii) The breaking strength using process 2 exceeds the standard by
at least 52-50-9= -7 lbs.
(iii) The breaking strength using process 3 exceeds the standard by
at least 45-50-9 = -14 lbs.
The joint statement consisting of the above three conclusions has a
confidence coefficient of 95%, i.e., in the long run, 95% of such joint
statements will actually be correct. Upper limits for the three differ-
ences could be obtained in an analogous manner.
For two-sided limits, the allowance is A = (2.94) (3.56) =11 and the
experimenter can conclude that:
(i) The breaking strength using process 1 exceeds the standard by
an amount between 61-50?11=0 and 22 lbs.
(ii) The breaking strengtlh using process 2 exceeds the standard by
an amount between 52-50?11= -9 and 13 lbs.
(iii) The breaking strength using process 3 exceeds the standard by
an amount between 45-50?11= -16 and 6 lbs.
The joint confidence coefficient for these three statements is greater
than 95%. (Due to an approximation made in computing Tables 2a
and 2b, the tabulated values of t are somewhat larger than necessary
so that the actual P's attained are slightly greater than 95 and 99%.
No such approximation was made in computing Tables la and lb.)
(b) The following data are blood count measurements on three
groups of animals, one of which served as a control while the other two
were treated with two drugs. Due to accidental losses, the numbers of
animals in the three groups are unequal.

Blood Counts (millions of cells per cubic millimeter)

Controls Drug A Drug B

7.40 9.76 12.80

8.50 8.80 9.68
7.20 7.68 12.16
8.24 9.36 9.20
9.84 10.55
8.32

Sums: 49.50 35.60 54.39

N: 6 4 5
Means: 8.25 8.90 10.88

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1100 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

Computations:

49.502 35.602 54.392

7.402+8.502+ * +9.202+10.552- - -_ _
s26 4 5
(ENi) -(p+1)
16.566 16.566
= = = ~~1.3805
15-3 12

s =\/1.3805 = 1.175

For d.f. = ( EN)-(p+1) = 12 and P=95%, t=2.11 (one-sided) or

t = 2.50 (two-sided).
"Allowances" for differences from the control:

One-sided: drug A: (2.11) (1.175) V1/6 + 1/4 -1.60

drug B: (2.11) (1.175)V\/1/6 + 1/5- 1.50

Two-sided: drug A: (2.50) (1.175) V1/6 + 1/4 = 1.90

drug B: (2.50)(1.175)V/1/6 + 1/5 = 1.78

If the experimenter is interested only in upper one-sided limits for th

differences from the control, he can make the following statements:
(i) Drug A raises the blood count by at most 8.90-8.25+1.60
= 2.25 millions per cmm.
(ii) Drug B raises the blood count by at most 10.88-8.25+1.50
=4.13 millions per cmm.
The joint confidence coefficient for these two statements is approxi-
mately 95%. (Since the tables of t were computed for equal numbers
of observations per group, their use in the case of unequal numbers
results in the desired probabilities being only approximately achieved.)
Corresponding lower limits could be calculated in an analogous manner.
If the experimenter desires simultaneous upper and lower limits on
the differences, he should use the two-sided allowances as follows:
(i) Drug A raises the blood count by an amount between 8.90-8.25
? 1.90= -0.25 and 2.55 millions per cmm.
(ii) Drug B raises the blood count by an amount between 10.88
-8.25?1.78=0.85 and 4.41 millions per cmm.
An approximate lower bound to the joint confidence coefficient of
these statements is 95%.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1101

III. THEORETICAL BASIS

Suppose there are available No observations on the control, N1 ob-

servations on the first treatment, , N, observations on the p-th
treatment. Denote these observations by Xii (i =0, 1, *, p; j = 1,
2, * , Ne), and the i-th treatment mean, 7Nf$ X l/N, by Xi. We
make the assumptions usually made in the analysis of variance, namely,
that the Xij are independent and normally distributed with common
variance a2 and means mi. We assume also that there is available an
estimate s2 of a2, independent of the Xi, which is based on n degrees of
freedom. For example, we may take
p Ni

S2 = E E (Xj - )2/n (1)

i-O j=1

where n = ( = Ni) - (p+ 1). Our problem is to obtain separate con-

fidence limits for each of the differences mi - mo (i= 1, 2, , p), such
that the joint confidence coefficient, i.e., the probability that all p
confidence intervals will contain the corresponding mi-mo, is equal to
a preassigned value, P (0 <P < 1).
Consider first the case p= 1, where there is only one treatment to be
compared with the control. The method of obtaining confidence limits
for mi - m, as described in almost any statistics textbook, is based on
Student's t-distribution. If we write

X- Xo- (mi- MO)

/ 1 1

V/ NJ No

which is normally distributed with mean 0 and variance u2, then

t-z/s follows the Student t-distribution with n degrees of freedom. A
lower limit on mi-m 0 with the desired confidence coefficient will be
given by

/1 1
- Xo0 -d's + 1
N,No
if d' is chosen so that

Prob (t < d') P (2)

Similarly, if an upper confidence limit is required, it will be given by

X- Xo+ d's/' + 1
0'VN, No

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1102 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

On the other hand, if the experimenter wishes to have bounds on

ml -m, in both directions, he may take

X1i - Xo ?+ N0

where d" is chosen to satisfy

Prob(|t| <d")=P (3)

The constant d' or d" corresponding to the desired value of P can be

obtained from tables of the percentage points of Student's t-distribu-
tion, which are widely available.
Now consider the general case where there are p treatments and a
control. Write

Xi-X - (mi -MO)

zi
/ 1 1
'VNi No
and ti=zi/s, i=1, 2, , p. Then lower confidence limits with
joint confidence coefficient P for the p treatment effects mi-mo will
be given by

Xi -/ N1+ N~
Xi-Xo-di's 1 ' (i = 1, 2, *..., p),

if the p constants di' are chosen so that

Prob (t, < di', t2 < d2',**, tp < dp')-=P. (4)

Similarly, upper confidence limits will be given by

/ 1 1
Xi-Xo + dits N + -N

On the other hand, two-sided confidence limits having

confidence coefficient will be given by

Xi -Xo ?distt4 + A~ (i = 1, 2, ... A,)

1 1~~~

if the p constants di" are chosen to satisfy

Prob (I t, < dil, I t2 I< d2", ., tj I < d") = P. (5)

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1103

To find any set of constants di' or di" satisfy

joint distribution of the ti is required. It may be noted that the joint
distribution of the zi is a multivariate normal distribution with means
0 and variances a2 and where the correlation between zi and zj is given
by

Pij= No @+ No +1)

The joint distribution of the ti is thus the multivariate analogue of

Student's t-distribution defined by Dunnett and Sobel [4].
To find solutions to (4) and (5), we require a tabulation of the multi-
variate Student t-distribution. We shall show now that the problem of
tabulating the multivariate Student t-distribution can be reduced to the
problem of tabulating the corresponding multivariate normal distribu-
tion. Consider first equation (4) above. It can be written

P = Prob (z1 < di's, Z2 < d2's, * , zp < dp's)

r (6)
= j __ F(d/'s, d2's, * * *, dp's) p(s)ds
00

where F (z1, Z2, * , Zp) is the multivariate normal c.d.f. of the zi and
p(s) is the probability density function of s. Thus, if F were tabulated,
it would be fairly easy using a desk calculator to evaluate (6) by nu-
merical integration for any set of fixed di', and hence to find solutions
to (4) as functions of P, p and n.
Similarly, (5) can be written

P = Prob (I z' I < di"s, I Z21 <d2"s, * * *, I zPI < dp"s)

r+00 (7)
= f G(df"s, d2"s, * * *, dp"s)p(s)ds
-00

where G (zI, Z2, * ,Zp) is the c.d.f. of the I z i.

lated, we could also evaluate (7) and determine the solutions to (5) as
functions of P, p and n.
The functions F and G can be obtained for p =2 from K. Pearson's
tables of the bivariate normal distribution [13], although the tabulation
interval is not fine enough for numerical integration purposes. How-
ever, for p = 2, (4) and (5) can be evaluated directly from the results of
Dunnett and Sobel [4]. The function F has been tabulated for equal
values of its arguments and for p < 9 by the National Bureau of Stand-
ards [10] for the special case Pij = 1/2. As will be explained in the next

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1104 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

section, Table 1 of this paper was based on this tabulation. It would be

extremely useful for the purpose of the problem considered in this paper
to have a corresponding tabulation of the function G.
Until exact tables are available, it will be necessary in general to rely
upon approximations to the solutions of (4) and (5). From the results
given by Dunnett and Sobel [5], we can write, when

pzj = / 1 C + 1 _+ 1),
p

Prob (ti < d1', t2 < d2', p, t < dp') > T Prob (t < di') (8)
i=l

and

Prob (f t1I < d11, I t2| < d2", t, It < dp")

> ][I Prob t |It < di").

i=j

Thus, upper bounds to the constants di' and di" can be determined
by equating the right-hand sides of (8) and (9) to the desired value of
P. These calculations involve only the probability integral of the uni-
variate Student t-distribution, which has been tabulated most recently
by Hartley and Pearson [8, 12].
Since (4) and (5) are each increasing functions of the correlations
Pij, alternative lower bounds which are closer than the lower bounds
given in (8) and (9) may be obtained by taking pij = 0. Pillai and
Ramachandran [14] have tabulated solutions d'= d2'= ... dp' to (4)
and solutions d1"=d211= * =d," to (5) for P=.95 and pij=O. It
would be useful to have a tabulation of such equicoordinate percentage
points of the multivariate Student t-distribution in the important spe-
cial case where pij = 0 for other values of P.
Lower bounds based on the bivariate Student t-distribution can also
be obtained, as shown in [5]. Taking di' = = .. =dp'=d', di"
=d2 = *.** =dp"=d", and pij=p=p>O, these can be written

Prob (ti < d', t2 < d', .. * * tp < d' I Pij = P) (1

> [Prob (ti < d', t2 < d' I P12 = P)]p/2
and

Prob (I tiI < d", I t2| < d", * *, | t1I < d" I pijP)(
>!_ [Prob (I ti I. < d",It2 < d" I P 12t| . = P) IP/2

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1105

The probability within the square brackets can be determined from

the probability integral of the bivariate Student-t distribution [4] with
correlation p. The bounds obtained from these inequalities are sharper
than those given by (8) and (9), and in most cases will also be sharper
than those obtained by using the Pillai and Ramachandran tables [14].
There are, of course, infinitely many solutions to (4) and (5). For the
applications considered in this paper, we will take the constants to be
equal, viz., di'= d2'= d. =zdp' (= d', say) and di" = d21 = ... =dp"
(=dd", say). Besides greatly simplifying the computational problem,
there are some theoretical grounds for doing so in the case where
Pij= p >0, which occurs frequently in practice. For example, it c
shown that >E di2 and E di"' are each minimized by this cho
IV. CONSTRUCTION OF THE TABLES

Tables la and lb give solutions d'= d'= ... = dp' to (4) for P =-.95
and .99, respectively, for p < 9 and Pij = 1/2. They are applicable to t
situation where there are equal numbers of observations on the control
and the p treatments, viz, No = N1 = ... = Np. These tables were con-
structed by numerical evaluation of the integral in (6) using tables of
the function F computed by the National Bureau of Standards [10].
The method used was as follows: For n = 5, 10, 20, 25 degrees of free-
dom, the integral in (6) was calculated for three successive values of d'
differing by 0.1 such that the desired value of P was bracketed. The
required value of d' was then determined to 3 decimal places by 3-point
inverse interpolation. For n = oo degrees of freedom, the values given
in tables computed by Bechhofer [1] were used. For the intermediate
degrees of freedom, the values were obtained by interpolation with 1/n
as argument. An accuracy of 1 in the second decimal place should be
achieved by this method.
Tables 2a and 2b give solutions d" =d"= . . . = dp" to the right-
hand side of (11) for P = .95 and .99, respectively, for p < 9 and pij = 1/2.
The method of constructing these tables was as follows: For n =5, 10,
20, 40, oo degrees of freedom, the probability in the square brackets of
(11) was calculated for three successive values of d" differing by 0.1
such that the desired value of P was bracketed, using the expressions
developed in [4] for the probability integral of the bivariate Student
t-distribution. The required value of d" was then determined by 3-point
inverse interpolation to 3 decimal places. For the intermediate degrees
of freedom, the required values were obtained by interpolation using
1/n as argument. An accuracy of 1 in the second decimal place should
be achieved by this method.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1106 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

V. OPTIMUM ALLOCATION OF OBSERVATIONS BETWEEN

CONTROL AND TREATMENTS

The tables given in this paper were prepared to handle the case where
equal numbers of observations are available on each of the p treatments
and on the control. In many practical situations, the experimenter will,
in fact, wish to allocate the available number of observations equally
to each group. Where it is feasible, however, it may be advantageous
to do otherwise. In this section, we will consider the consequences of
allocating No observations to the control and N1 observations to each of
the other treatment groups, i.e., we will take N1 = N2= N =N, and
No$N1.
Lower confidence limits to the mi-mo will be given by

Xi- XO -d's4f + N~ (i = 1, 2) y P)l

where d' is chosen to satisfy

Prob (ti < d', i = 1, 2, .. * *,p) = P' (12)

the tt having the multivariate Student t-distribution with

P=i + 1

We will consider the allocation No/N1 optimum if it maximizes P for

fixed value of

1 1

N' t + No
and fixed total number of observations, No+pN,. It may be noted that
fixing the total number of observations also fixes n, the number of de-
grees of freedom associated with the multivariate t-distribution, if s is
defined by (1). Let

h = d' +No

Then (12) can be written

h
Prob ti < 1 1 ' = 1, 2, * * p P (13)
N1 No

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1107

From (13), it is evident that P can be increased for fixed h by de-

creasing

/1 1
'V NJ No

It is easy to show, as has been pointed out by several authors, see, for
example, Finney [7], that

/1 1

'V NJ No

attains a minimum for fixed No+ pNI when No/N, = V/

choice makes pij<1/2, which operates to decrease P.
In order to investigate numerically the effect of different allocations
on P, the curves shown in Figs. 1 and 2 computed. Fig. 1 shows P as a
function of No/N1 for p = 2 and n = 1, 2, 5, 10, oo, with h chosen in each
case to make P=. 95 when No/N1= 1. Fig. 2 shows P as a function of
No/N1 for p=2, 4, 9 and n= oo, with h again chosen to make P=.95
when No/N1= 1. The curves for finite n were computed from formulas
given in [4]. The curves for n = oo were computed by numerical integra-
tion using tables of the normal distribution [9], with certain points for
p = 2 checked against Pearson's bivariate normal tables [13].
Figs. 1 and 2 indicate that the optimum allocation occurs where
No/N1 is only slightly less than \/p except when the number of degr
of freedom is small. Some further computations carried out by the
author for the case of n = oo degrees of freedom indicated that the point
of optimum allocation becomes even closer to -\/p when h is chosen to
make P = .99 for No/N1 = 1. However, for h chosen to make P = .75 for
No/N, = 1, it was found that P <.75 for No/N, = Vlp and the optimum
value of No/N, was considerably less than \p. For practical purposes,
we can thus conclude that, if the experimenter is working with a joint
confidence coefficient in the neighborhood of P=.95 or greater, then
the experiment should be designed so that No/N1 = V/p approximately,
where No is the number of observations on the control and N1 the num-
ber on each of the p treatments.

VI. APPLICATION TO BIOLOGICAL ASSAY

An important example involving the multiple comparison of several

treatments with a control arises in the biological assay of several drug
samples relative to a common standard. In the "parallel line" type of
biological assay, regression lines Y=ao+bX and Y=al+bX are fitted

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1108 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

Relaionship Between Allocation Ratio and

Confidence Coefficient for p=2and Various n.

.953a-

CL951

0'

947

/~~~~~i I
.96
0.8 1.0 1.2 1.4 1.6 1.8 2.0
Allocation Rcatio) N0
FIG. 1.-In plotting these curves, d'V(-/
so that P =.95 for No =N1, where d's V(1/N1)+(1/NO) is the difference between
X,-Xo and its lower confidence limit. The dotted curve indicates where the
optimum occurs for each n; the vertical dashed line is drawn at No/N = =p-=\2.

to data representing observed responses Y at several log-dose levels X

of a standard drug preparation S and a test sample U. The estimated
log-potency of U relative to S is represented by the horizontal distance
between the two regression lines,

m l a-a = b XO-Xi + bY (14)

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1109

Relationshi p Between Allocation Ratio and Confidence

Coefficient; for n?=0o ond Various p
.972
p=9

..968

(D

Zt.964
0

.948
1.0 1.5 a.0 2.5 3.0 3.5 4.0
Allocation Ratio, N

FIG. 2.-In plotting these curves,

so that P =.95 for No = N1, where d
X,-XO and its lower confidence limit. The black arrows indicate where the
optima occur; the white arrows indicate the points where No/N, = V/P.

where X0 and X1 are the mean log-dose levels, and Yo and Y7 are the
mean observed responses, of S and U respectively. On the other hand
the true log-potency may be represented by

M = Xo-X+ miM o (15)

where mo, ml and ,B are the expected values of Yo

Assuming that the responses Y are independent and normally dis-
tributed with common variance U2 and means mo+,B(X-Yo) and
ml+fl(X-71) for S and U respectively, we can obtain confidence
limits for M by Fieller's [6] method by considering

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1110 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

71-0o-(M-Yo + Xi)b

__ + 1 + (M-Xo + Y)2c2]

Here, No and N1 are the numbers of responses observed on S and U

respectively, and C2o-2 is the variance of the common slope b. Then z is
normally distributed with mean 0 and variance 2. If S2 is an estimate
of a-2 independent of z, based on n degrees of freedom, then z/s has
Student's t-distribution with n degrees of freedom. Hence, confidence
limits for M with confidence coefficient P can be obtained by equating
z2/S2 to t2, where t is the P-percentage point of Student's t, and solving
the resulting quadratic equation for M. The solution may be found, for
example, in Bliss [3].
Now suppose there are p unknown drug samples: U1, U2 *, U, to
be compared with a common standard, S. Extending the above nota-
tion in an obvious way, we define the p variables,

Yi-Y0- (M -Xo + X)b

Zi 1 _ _- 111 . I p (17)
+ + (M i-Xo + X)2c21

The zi have a joint p-variate normal distribution with means 0,

common variance a-2 and correlation between zi and zi given by

Pii =(l+ Ej) N/o( + 1+ 6ei2) (NO+ 1+ 6j2)

where e = M- Xo+ X)cVNO. For the p pairs of confidence limits on

the Hi to have a joint confidence coefficient equal to P, we are led to
consider the joint distribution of the zi/s, which is the multivariate
Student t-distribution as in Section III, with this important difference:
the correlations Pij are no longer known exactly since they involve the
unknown parameters Mi and Mj. Fortunately, ei may be expected to be
fairly small in most cases since it is common practice in designing a
biological assay experiment to try to arrange the dose levels so that
O- Xi is close to Mi. Thus, we can obtain confidence limits for the Mi
which have approximately the desired confidence coefficient by assum-
ing the ei to be negligible, whence

Pii = 1/ 1/( N + 1) ( + 1). (18)

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1111

Contour Curves for Correlation

in Bioloqical Assay Case

1.05

-0.5

E4

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

FIG. 3

This takes the value 12 in the usual situation where N = Ni (i-=1,

2, * , p), in which case the tables in this paper are applicable.
The contour curves in Fig. 3 show the effect of ei and e, on p*u
No = Ns-N. In most practical situations, the log-dose levels chosen by
the experimenter for each drug will cover a wide enough range so that

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1112 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

Effect of CorrelcQtion Coefficieent on

.958 Confidence Coefficienb

.956
0L

;.954

30.952

U
c
..D .9 50 0000 00 000*00000#00
4-~ ~ ~ ~ -
0 -

-94 8 t X

.946

.944 -
.3 .4 .5 .6 .7
Correlation Co
FIG. 4

the ei are numerically small, in which case they do not exert much
influence on the value of pij. In Fig. 4, curves are drawn for degrees of
freedom equal to 5 and oo showing the actual P attained as a function
of p in the case of estimating the log-potencies of two drugs (p=2)
when No = N1 = N2 and the tables are used applicable to p = - and

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1113

P =95%. It may be seen that for a wide range of values of p around

p2=, P does not differ much from the value 95 %.

VII. MULTIPLE DECISION PROCEDURES FOR COMPARING

TREATMENTS WITH A CONTROL

A multiple decision procedure for selecting the "best" of p+l cate-

gories when comparing p experimental categories with a control has
been developed by Paulson [11]. Paulson gives a method for making
one of the following p + 1 decisions:

JDo: select the control as best

|Di: select the i-th category as best (i= 1, 2, *, p)

If Xo, X1, * * *, X, represent the p+l means, each based on N observa-

tions, Xo being the control, then Paulson's method consists in picking
out X* = max (X1, * * *, Xp) and if X*-Xo - > X1,s 2/N the decision D*
corresponding to Y* is made, whereas if X* - Xo <XosV2/N the de-
cision Do is made. The constant XA is chosen so that the probability of
making decision Do, when all p treatments are equivalent to the control,
is equal to a pre-assigned value P. Clearly, di'= X. must be a solution
of (4). Thus, Tables la and lb of this paper give the required values of
X,, for P=.95 and .99.
In some situations it may be appropriate to consider the following set
of possible decisions from which a choice is to be made:
Do: select the control as best
Di: select the i-th treatment (i = 1, 2* *, p) as the only one better
than the control
D,j: select the i-th and j-th treatments (i, j= 1, 2, . .. , p; i j),
without ordering them, as the only two better than the control

D1,2,...,,: select all p treatments, without ordering them, as better

than the control

The following procedure is proposed for choosing one of the above 2-

decisions on the basis of the p+ l observed means XO, X, * * , p,, each
based on N observations, and the independent estimate s of the com-
mon standard deviation:

Accept Do if X,- YO <ds\V2/N for all i

Accept Di if Xi-Xo > dsV 2/N and Xj-Xo <ds /2/N for all j i
Accept Di1 if X_ -Xo>?dsV2/N, Xj-Xo > dsV\2/N and Xk-XO
< ds \2/N for all k Hci or j

\Accept D12.X, if Xi-Xo _ ds-v2/N for all i

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1114 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

By choosing d to satisfy (4), we will be assured that the probability of

accepting Do when all the treatments are equivalent to the control is
equal to a pre-assigned value P. In fact, d will then be identical with
Paulson's X1n. Table 1 gives the required values of d for P =.95 and .99.
It is also possible to determine the size N of sample required to
achieve a specified, probability of accepting some other decision when
all the treatments are not equivalent to the control. For example,
suppose all the treatments are equivalent to the control except one,
say the first one, which is better than the control. The correct decision
to make would then be D1. Suppose mO=m2=M3= m .. =mp=ml-8
Then the probability of making decision D1 can be written

Pi = Pr (Di |mO=m2 = *... = m = ml-5)

= Pr (X1-Xo >ds/2/N, X2-Xo

____ ~~(19)
< dsA2/N, *.* X, -Xo < ds8/2/N)

=Pr (ti > d, t2< dy I t* *t<d)

where

ti- xi /O (i =1, 2,*,p)

We note that t2, *, are Student t-variates, but t1 is a non-central
t-variate.
To obtain bounds for Pi, write

Pi = Pr (ti <d)P t2 <d, *.* *, tp<d) (20)

> Pr (ti < d) Pr (t2 < d,**, tp < d)

This inequality follows from Dunnett and Sobel [5]. Since

P, + P, = Pr(t2 < d, ... ** tp < d), (21)

it follows that an upper bound for P1 is given by

P, _ Pr (ti > d) Pr (t2 < d, * * ., tp < d). (22)

On the other hand, an obvious upper bound on P5 is

1P1 < Pr (ti < d) (23)

From (21) and (23), we get the following lower bound for P1,

P1 > Pr (t2 <d, * * *,tp < d)-Pr(t1 < d). (24)

The difference between the bounds given by (22) and (24) is

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1115

Bounds on Pi for n 00 and p 2

1.0

to~/
.8-

.7
.9~ ~~~Z

.6

p1 .5 -D 1

.3 D 1

FIG. 5

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1116 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

Pr (t1 <d) * [1- Pr(t2 d, * * *, tp, <d)] which will be small in

interest where P1 is fairly large. In Fig. 5, the two bounds for Pi are
plotted as functions of V(N/2) 6/l- for d.f. = oo, p = 2 and 9, and where
d is chosen in each case so that P=.95. To compute the probability
involving t2, - * *, t,, in (22) and (24), tables of the multivariate nor-
mal distribution computed by the Bureau of Standards [10] were used.
If it were desired to compute the bounds on P1 for a finite number of
degrees of freedom, the probability involving t2, * * *, t, could be com-
puted by numerical integration on the Bureau of Standards tables.
However, since in most practical situations the number of degrees of
freedom will be large, it may be sufficient to assume d.f. = oo and use
the Bureau of Standards tables directly.
Using Fig. 5, it is possible to determine the required value of N cor-
responding to any given values of b/l- and P1. The following table shows
the value of V\N6/- corresponding to p = 1(1)9 to achieve Pi= .80 for
P=.95 and d.f.= oo.

p V/N6/'

1 3.52
2 4.05
3 4.30
4 4.46
5 4.58
6 4.67
7 4.74
8 4.81
9 4.86

For example, suppose that the experimenter has p =5 treatments to

compare with a control. Table la will provide the value of d to give a
probability of P =.95 that none of the treatments will be declared su-
perior to the control when, in fact, all are equivalent to the control. If,
in addition, the experimenter wants to achieve a probability of
P1= .80 of correctly selecting a superior treatment, if there is one wh
is, say, one standard deviation better than the control, the above table
shows that \VN6/- =4.58 whence N = 21 on substituting b/o-= 1. Thus
21 observations should be taken on each of the treatments and on the
control. This will provide (p+ 1) (N-1) = 120 degrees of freedom for
estimating the variance, so that d = 2.26 from table la. While the table
used above to determine N is based on d.f. = oo, the result should not be
much different for d.f. = 120. The experimenter would then take 21 ob-

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1117

servations in each group, observe XO and Xi (i= 1, 2, *, 5), calcu-

late 82 from equation (1), and declare any treatment superior to the
control which gives a mean Xi greater than Xo+2.26sV/2/21.

TABLE la*

TABLE OF t FOR ONE-SIDED COMPARISONS BETWEEN p TREAT-

MENT MIEANS AND A CONTROL FOR A JOINT
CONFIDENCE COEFFICIENT OF P =95%

P, NUMBER OF TREATMENT MEANS (EXCLUDING THE CONTROL)

d.f. 1 2 3 4 5 6 7 8 9

5 2.02 2.44 2.68 2.85 2.98 3.08 3.16 3.24 3.30

6 1.94 2.34 2.56 2.71 2.83 2.92 3.00 3.07 3.12
7 1.89 2.27 2.48 2.62 2.73 2.82 2.89 2.95 3.01
8 1.86 2.22 2.42 2.55 2.66 2.74 2.81 2.87 2.92
9 1.83 2.18 2.37 2.50 2.60 2.68 2.75 2.81 2.86

10 1.81 2.15 2.34 2.47 2.56 2.64 2.70 2.76 2.81

11 1.80 2.13 2.31 2.44 2.53 2.60 2.67 2.72 2.77
12 1.78 2.11 2.29 2.41 2.50 2.58 2.64 2.69 2.74
13 1.77 2.09 2.27 2.39 2.48 2.55 2.61 2.66 2.71
14 1.76 2.08 2.25 2.37 2.46 2.53 2.59 2.64 2.69

15 1.75 2.07 2.24 2.36 2.44 2.51 2.57 2.62 2.67

16 1.75 2.06 2.23 2.34 2.43 2.50 2.56 2.61 2.65
17 1.74 2.05 2.22 2.33 2.42 2.49 2.54 2.59 2.64
18 1.73 2.04 2.21 2.32 2.41 2.48 2.53 2.58 2.62
19 1.73 2.03 2.20 2.31 2.40 2.47 2.52 2.57 2.61

20 1.72 2.03 2.19 2.30 2.39 2.46 2.51 2.56 2.60

24 1.71 2.01 2.17 2.28 2.36 2.43 2.48 2.53 2.57
30 1.70 1.99 2.15 2.25 2.33 2.40 2.45 2.50 2.54
40 1.68 1.97 2.13 2.23 2.31 2.37 2.42 2.47 2.51
60 1.67 1.95 2.10 2.21 2.28 2.35 2.39 2.44 2.48

120 1.66 1.93 2.08 2.18 2.26 2.32 2.37 2.41 2.45
inf. 1.64 1.92 2.06 2.16 2.23 2.29 2.34 2.38 2.42

* Table la gives a solution di'=t to equation (4) in the tex

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1118 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

TABLE lb*

TABLE OF t FOR ONE-SIDED COMPARISONS BETWEEN p TREAT-

MENT MEANS AND A CONTROL FOR A JOINT
CONFIDENCE COEFFICIENT OF P =99%

p, NUMBER OF TREATMENT MEANS (EXCLUDING THE CONTROL)

d.f. 1 2 3 4 5 6 7 8 9

5 3.37 3.90 4.21 4.43 4.60 4.73 4.85 4.94 5.03

6 3.14 3.61 3.88 4.07 4.21 4.33 4.43 4.51 4.59
7 3.00 3.42 3.66 3.83 3.96 4.07 4.15 4.23 4.30
8 2.90 3.29 3.51 3.67 3.79 3.88 3.96 4.03 4.09
9 2.82 3.19 3.40 3.55 3.66 3.75 3.82 3.89 3.94

10 2.76 3.11 3.31 3.45 3.56 3.64 3.71 3.78 3.83

11 2.72 3.06 3.25 3.38 3.48 3.56 3.63 3.69 3.74
12 2.68 3.01 3.19 3.32 3.42 3.50 3.56 3.62 3.67
13 2.65 2.97 3.15 3.27 3.37 3.44 3.51 3.56 3.61
14 2.62 2.94 3.11 3.23 3.32 3.40 3.46 3.51 3.56

15 2.60 2.91 3.08 3.20 3.29 3.36 3.42 3.47 3.52

16 2.58 2.88 3.05 3.17 3.26 3.33 3.39 3.44 3.48
17 2.57 2.86 3.03 3.14 3.23 3.30 3.36 3.41 3.45
18 2.55 2.84 3.01 3.12 3.21 3.27 3.33 3.38 3.42
19 2.54 2.83 2.99 3.10 3.18 3.25 3.31 3.36 3.40

20 2.53 2.81 2.97 3.08 3.17 3.23 3.29 3.34 3.38

24 2.49 2.77 2.92 3.03 3.11 3.17 3.22 3.27 3.31
30 2.46 2.72 2.87 2.97 3.05 3.11 3.16 3.21 3.24
40 2.42 2.68 2.82 2.92 2.99 3.05 3.10 3.14 3.18
60 2.39 2.64 2.78 2.87 2.94 3.00 3.04 3.08 3.12

120 2.36 2.60 2.73 2.82 2.89 2.94 2.99 3.03 3.06
inf. 2.33 2.56 2.68 2.77 2.84 2.89 2.93 2.97 3.00

* Table lb gives a solution di'=t to equation (4) in the t

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
 COMPARING SEVERAL TREATMENTS WITH A CONTROL 1119

TABLE 2a*

TABLE OF t FOR TWO-SIDED COMPARISONS BETWEEN p TREAT-

MENT MEANS AND A CONTROL FOR A JOINT
CONFIDENCE COEFFICIENT OF P=95%

P, NUMBER OF TREATMENT MEANS (EXCLUDING THE CONTROL)

d.f. 1 2 3 4 5 6 7 8 9

5 2.57 3.03 3.39 3.66 3.88 4.06 4.22 4.36 4.49

6 2.45 2.86 3.18 3.41 3.60 3.75 3.88 4.00 4.11
7 2.36 2.75 3.04 3.24 3.41 3.54 3.66 3.76 3.86
8 2.31 2.67 2.94 3.13 3.28 3.40 3.51 3.60 3.68
9 2.26 2.61 2.86 3.04 3.18 3.29 3.39 3.48 3.55

10 2.23 2.57 2.81 2.97 3.11 3.21 3.31 3.39 3.46

11 2.20 2.53 2.76 2.92 3.05 3.15 3.24 3.31 3.38
12 2.18 2.50 2.72 2.88 3.00 3.10 3.18 3.25 3.32
13 2.16 2.48 2.69 2.84 2.96 3.06 3.14 3.21 3.27
14 2.14 2.46 2.67 2.81 2.93 3.02 3.10 3.17 3.23

15 2.13 2.44 2.64 2.79 2.90 2.99 3.07 3.13 3.19

16 2.12 2.42 2.63 2.77 2.88 2.96 3.04 3.10 3.16
17 2.11 2.41 2.61 2.75 2.85 2.94 3.01 3.08 3.13
18 2.10 2.40 2.59 2.73 2.84 2.92 2.99 3.05 3.11
19 2.09 2.39 2.58 2.72 2.82 2.90 2.97 3.04 3.09

20 2.09 2.38 2.57 2.70 2.81 2.89 2.96 3.02 3.07

24 2.06 2.35 2.53 2.66 2.76 2.84 2.91 2.96 3.01
30 2.04 2.32 2.50 2.62 2.72 2.79 2.86 2.91 2.96
40 2.02 2.29 2.47 2.58 2.67 2.75 2.81 2.86 2.90
60 2.00 2.27 2.43 2.55 2.63 2.70 2.76 2.81 2.85

120 1.98 2.24 2.40 2.51 2.59 2.66 2.71 2.76 2.80
inf. 1.96 2.21 2.37 2.47 2.55 2.62 2.67 2.71 2.75

* Table 2a gives a solution di" =t which makes the right-han

equal to .95 for the case p =1/2. This may be used as an approximate solution to equation (5) in the
text for P -.95 for the case Pij =1/2.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
1120 AMERICAN STATISTICAL ASSOCIATION JOURNAL, DECEMBER 1955

TABLE 2b*

TABLE OF t FOR TWO-SIDED COMPARISONS BETWEEN p TREAT-

MENT MEANS AND A CONTROL FOR A JOINT
CONFIDENCE COEFFICIENT OF P=99%

p, NUMBER OF TREATMENT MEANS (EXCLUDING THE CONTROL)

d.f. 1 2 3 4 5 6 7 8 9

5 4.03 4.63 5.09 5.44 5.73 5.97 6.18 6.36 6.53

6 3.71 4.22 4.60 4.88 5.11 5.30 5.47 5.61 5.74
7 3.50 3.95 4.28 4.52 4.71 4.87 5.01 5.13 5.24
8 3.36 3.77 4.06 4.27 4.44 4.58 4.70 4.81 4.90
9 3.25 3.63 3.90 4.09 4.24 4.37 4.48 4.57 4.65

10 3.17 3.53 3.78 3.95 4.10 4.21 4.31 4.40 4.47

11 3.11 3.45 3.68 3.85 3.98 4.09 4.18 4.26 4.33
12 3.05 3.39 3.61 3.76 3.89 3.99 4.08 4.15 4.22
13 3.01 3.33 3.54 3.69 3.81 3.91 3.99 4.06 4.13
14 2.98 3.29 3.49 3.64 3.75 3.84 3.92 3.99 4.05

15 2.95 3.25 3.45 3.59 3.70 3.79 3.86 3.93 3.99

16 2.92 3.22 3.41 3.55 3.65 3.74 3.82 3.88 3.93
17 2.90 3.19 3.38 3.51 3.62 3.70 3.77 3.83 3.89
18 2.88 3.17 3.35 3.48 3.58 3.67 3.74 3.80 3.85
19 2.86 3.15 3.33 3.46 3.55 3.64 3.70 3.76 3.81

20 2.85 3.13 3.31 3.43 3.53 3.61 3.67 3.73 3.78

24 2.80 3.07 3.24 3.36 3.45 3.52 3.58 3.64 3.69
30 2.75 3.01 3.17 3.28 3.37 3.44 3.50 3.55 3.59
40 2.70 2.95 3.10 3.21 3.29 3.36 3.41 3.46 3.50
60 2.66 2.90 3.04 3.14 3.22 3.28 3.33 3.38 3.42

120 2.62 2.84 2.98 3.08 3.15 3.21 3.25 3.30 3.33
inf. 2.58 2.79 2.92 3.01 3.08 3.14 3.18 3.22 3.25

* Table 2b gives a solution di" =t which makes the right-

equal to .99 for the case p = 1/2. This may be used as an approximate solution to equation (5) in the
text for P =.99 for the case pij = 1/2.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms
COMPARING SEVERAL TREATMENTS WITH A CONTROL 1121

REFERENCES

[1] Bechhofer, Robert E., "A single-sample multiple decision procedure for
ranking means of normal populations with known variances," Annals of
Mathematical Statistics, 25 (1954), 16-39.
[2] Bechhofer, Robert E., Dunnett, Charles W., and Sobel, Milton, "A two-
sample multiple decision procedure for ranking means of normal populations
with a common unknown variance," Biometrika, 41 (1954), 170-6.
[3] Bliss, C. I., The Statistics of Bioassay. New York, Academic Press, 1952.
[4] Dunnett, Charles W., and Sobel, Milton, "A bivariate generalization of
Student's t-distribution, with tables for certain special cases," Biometrika,
41 (1954), 153-69.
[5] Dunnett, C. W., and Sobel, M., "Approximations to the probability integral
and certain percentage points of a mnultivariate analogue of Student's
t-distribution," Biometrika, 42 (1955), 258-60.
[6] Fieller, E. C., "A fundamental formula in the statistics of biological assay,
and some applications," Quarterly Journal of Pharmacy and Pharmacology,
17 (1944), 117-23.
[7] Finney, D. J., Statistical Method in Biological Assay. New York: Hafner
Publishing Company, 1952.
[8] Hartley, H. O., and Pearson, E. S., "Table of the probability integral of the
t-distribution," Biometrika, 37 (1950), 168-72.
[9] National Bureau of Standards, Tables of Normal Probability Functions.
Washington: U. S. Government Printing Office, 1953.
[10] National Bureau of Standards, Unpublished tables.
[11] Paulson, Edward, "On the comparison of several experimental categories
with a control," Annals of Mathematical Statistics, 23 (1952), 239-46.
[12] Pearson, E. S., and Hartley, H. O., Biometrika Tables for Statisticians,
Vol. I. Cambridge: Cambridge University Press, 1954.
[13] Pearson, K., Tables for Statisticians and Biometricians, Part II. Cambridge:
Cambridge University Press, 1931.
[14] Pillai, K. C. S., and Ramachandran, K. V., "On the distribution of the ratio
of the ith observation in an ordered sample from a normal population to an
independent estimate of the standard deviation," Annals of Mathematical
Statistics, 25 (1954), 565-72.
[15] Roessler, E. B., "Testing the significance of observations compared with a
control," Proceedings of the American Society for Horticultural Sciencc, 47
(1946), 249-51.
[16] Scheff6, Henry, "A method for judging all contrasts in the analysis of vari-
ance," Biometrika, 40 (1953), 87-104.
[171 Tukey, John W., "The problem of multiple comparisons. " Unpublished
notes in private circulation. Princeton University.
[18] Villars, Donald Statler, Statistical Design and Analysis of Experiments for
Development Research. Dubuque, Iowa: Wim. C. Brown Company, 1951.

This content downloaded from

200.144.62.89 on Sun, 07 Apr 2024 23:39:38 +00:00
All use subject to https://about.jstor.org/terms