PWJohn

Downloaded 12/11/12 to 129.128.207.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.
php
P Statistical Design
and Analysis
of Experiments
Downloaded 12/11/12 to 129.128.207.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
Statistical Design
and Analysis
of Experiments
Peter W. M. John
University of Texas at Austin
Austin, Texas
51IITL.
Society for Industrial and Applied Mathematics
Philadelphia
SIAM's Classics in Applied Mathematics series consists of books that were previously
allowed to go out of print. These books are republished by SIAM as a professional service
because they continue to be important resources for mathematical scientists.
Editor-in-Chief
Robert E O'Malley, Jr., University of Washington
Editorial Board
Richard A. Brualdi, University of Wisconsin-Madison
Herbert B. Keller, California Institute of Technology
Andrzej Z. Manitius, George Mason University
Ingram Olkin, Stanford University
Stanley Richardson, University of Edinburgh
Ferdinand Verhuist, Mathematisch Instituut, University of Utrecht
Classics in Applied Mathematics
C. C. Lin and L A. Segel, Mathematics Applied to Deterministic Problems in the Natura!
Sciences
Johan G. F. Belinfante and Bernard Kolman, A Survey of Lie Groups and Lie Algebras with
Applications and Computational Methods
James M. Ortega, Numerical Analysis: A Second Course

Anthony V. Fiacco and Garth P. McCormick, Nonlinear Programming: Sequential
Unconstrained Minirnization Techniques
F. H. Clarke, Optimization and Nonsmooth Analysis

George F. Carrier and Carl E. Pearson, Ordinary Differential Equations
Leo Breiman, Probability
R. Bellman and G. M. Wing, An Introduction to Invariant Inbedding
Abraham Berman and Robert J. Plemmons, Nonnegative Matrices in the
Mathematica) Sciences
Olvi L Mangasarian, Nonlinear Programming
*Carl Friedrich Gauss, Theory of the Combination of Observations Least Subject to Errors:
Part One, Part Two, Supplement Translated by G. W. Stewart
Richard Bellman, Introduction to Matrix Analysis
U. M. Ascher, R. M. M. Mattheij, and R. D. Russell, Numerical Solution of Boundary Value
Problems for Ordinary Differential Equations
K. E. Brenan, S. L Campbell, and L. R. Petzold, Numerical Solution of InitialValue Problems
in Differential Algebraic Equations
Charles L Lawson and Richard J. Hanson, Solving Least Squares Problems

J. E. Dennis, Jr. and Robert B. Schnabel, Numerical Methods for Unconstrained Optimization
and Nonlinear Equations
Richard E. Barlow and Frank Proschan, Mathematica) Theory of Reliability

Cornelius Lanczos, Linear Differential Operators
Richard Beliman, Introduction to Matrix Analysis, Second Edition
Beresford N. Parlett, The Symmetrie Eigenvalue Problem
Richard Haberman, Mathematical Models: Mechanica] Vibrations, Population Dynaniics, and
Traffic Flow
Peter W. M. John, Statistica! Design and Analysis of Experiments

*First time in nrint.
Copyright 1998 by the Society for Industrial and Applied Mathematics.

This SIAM edition is an unabridged republication of the work first published by The
Macmillan Company, New York, 1971.
10987654321
All rights reserved. Printed in the United States of America. No part of this book may be
reproduced, stored, or transmitted in any maner without the written permission of the
publisher. For information, write the Society for Industrial and Applied Mathematics, 3600
University City Science Center, Philadelphia, PA 19104.2688.
Library of Congress Cataloging-in-Publication Data
John, Peter William Meredith.

Statistical design and analysis of experiments / Peter W.M. John.
p. cm. - (Classics in applied mathematics ; 22)
"An unabridged republication of the work first published by the
Macmillan Company, New York, 1971 "--T.p. verso.
Includes bibliographical references (p. - ) and index.
ISBN 0-89871-427.3 (softcover)
1. Experimental design. I. Title. II. Series.
QA279j65 1998
98-29392
001.4'34-dc21
ELBJ1L is a registered trademark.
Contents
Preface
xi
Preface to the Classics Edition

References in the Preface
x111
xxiii
Chapter 1. Introduction
1.1 The Agricultural Heritage
1.2 An Example of an Experiment
1.3 Tests of Hypotheses and Confidence Intervals
1.4 Sample Size and Power
1.5 Blocking
1.6 Factorial Experiments
1.7 Latin Square Designs
1.8 Fractional Factorials and Confounding
1.9 Response Surfaces
1.10 Matrix Representation
2
3
5
7
8
10
12
12
14
14
Chapter 2. Linear Models and Quadratic Forms
17
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
17
18
19
19
20
21
22
23
26
Introduction
Linear Models
Distribution Assumptions
The Method of Least Squares
The Case of Full Rank
The Constant Term and Fitting Planes
The Singular Case
Pseudo-Inverses and the Normal Equations
Solving the Normal Equations
vi
Contents

2.10 A Theorem in Linear Algebra

2.11 The Distribution of Quadratic Forms

2.12 The Gauss-Markoff Theorem

2.13 Testing Hypotheses

2.14 Missing Observations
28
29
34
35
37
Chapter 3. Experiments with a Single Factor

3.1 Analysis of the Completely Randomized Experiment

3.2 Tukey's Test for Treatment Differences

3.3 Scheff's S Statistic

3.4 A Worked Example

3.5 Qualitative and Quantitative Factors and Orthogonal Contrasts

3.6 The Gasoline Experiment (Continued)

3.7 The Random Effects Model

3.8 The Power of the F Test with Random Effects

3.9 Confidence Intervaas for the Components of Variance


3.11 The Randomized Complete Block Experiment


3.13 Missing Plots in a Randomized Complete Block Design

3.14 The Analysis of Covariance

3.15 Analysis of Covariance for the Randomized Complete Block Design
Chapter 4. Experiments with Two Factors

4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
The Two-Way Layout with One Observation per Cell

Orthogonality with More Than One Observation per Cell
The Two-Way Layout with Interaction
The Subdivision of the Sum of Squares for Interaction
Interaction with One Observation per Cell
The Random Effects Model
The Mixed Model
The Two-Stage Nested or Hierarchic Design
The Diallel Cross Experiment
39
41
45
45
46
48
52
52
53
54
54
55
58
59
60
63
66
66
68
70
73
75
76
77
80
81
Chapter 5. Experiments with Several Factors
86
5.1 Error Terms and Pooling

5.2 Replicates
5.3 The Complete Three-Factor Experiment
5.4 Approximate F Tests
86
87
88
92
Contents vii
5.5
5.6
5.7
5.8
5.9
An Experiment with Crossed and Nested Factors

General Rules for Sums of Squares and EMS
Split-Plot Designs
Split-Split-Plot Designs
Computer Programs
94
96
98
101
102
Chapter 6. Latin Square Designs
105
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
106
108
110
111
114
115
117
120
Analysis
Limitations of the Model
Graeco-Latin Squares
Sets of Orthogonal Latin Squares
Experiments Involving Several Squares
Change-Over Designs
Change-Over Designs Balanced for Residual Effects
Designs for Experiments with Residual Effects
Chapter 7. Factors with Two or Three Levels; Confounding
123
7.1 Factors at Two Levels

7.2 Yates' Algorithm and the General 2n Design
7.3 The 3'' Series of Factorials
7.4 Confounding
7.5 Partial Confounding
7.6 Confounding 2n Designs in 2k Blocks of 2n - k Points Each
7.7 The Allocation of Treatment Combinations to Blocks
7.8 The Composition of the Defining-Contrast Subgroup
7.9 Double Confounding and Quasi-Latin Squares
7.10 Confounding in 3n Designs
7.11 Confounding with Factors at More Than Three Levels
7.12 Confounding in the 4 2 Design
7.13 Confounding in Other Factorial Experiments
124
127
130
132
134
136
137
139
140
143
144
144
145
Chapter 8. Fractions of 2" Factorial Designs
148
8.1
8.2
8.3
8.4
8.5
8.6
8.7
149
151
152
153
155
157
158
The 2n k Fractional Factorials

The Classification of Fractions
Some 2n k Fractions
Designs with Eight Points
Designs with Sixteen Points
Other 2n -k Fractions of Resolution V
Blocking Fractional Factorials
-
viii Contents
8.8 Analyzing 2n - k Fractions

8.9 Designs Obtained by Omitting 2n - k Fractions
8.10 Three-Quarter Replicates
8.11 The 3(25- 2) Designs
8.12 Some 3(2n - k ) Designs of Resolution V
8.13 Adding Fractions
8.14 Nested Fractions
8.15 Resolution III Designs
8.16 Designs of Resolution IV
173
Chapter 9. Fractional Factorials with More Than Two Levels
177
3n Designs
Latin Squares as Main Effects Plans
9.3 3n - k Fractions with Interactions
9.4 Orthogonal Main Effects Plans
9.5 The Plackett and Burman Designs
9.6 Addelman's Designs
9.7 Fractions of 2m3 n Designs of Resolution V
9.8 Two Nonorthogonal Fractions of the 43 Design
9.1
9.2
159
161
163
165
167
167
170
172
178
179
181
184
185
187
190
191
Chapter 10. Response Surfaces
193
10.1 First-Order Designs

10.2 Second-Order Designs
10.3 3k Designs and Some Fractions
10.4 Polygonal Designs for k = 2
194
10.5
Central Composite Designs
10.6 Orthogonal Blocking

10.7 Steepest Ascent and Canonical Equations
196
198
201
204
206
208
209
10.8 The Method of Steepest Ascent

10.9 The Canonical Form of the Equations
10.10 Other Surfaces
10.11 Evolutionary Operation
10.12 The Method of Calculation
211
213
213
Chapter 11. Incomplete Block Designs
219
11.1
11.2
11.3
11.4
220
Balanced Incomplete Block Designs

Some Balanced Incomplete Block Designs
The Analysis of Balanced Incomplete Block Designs
Resolvable Designs
216
221
223
226
Contents ix
11.5 The General Incomplete Block Design

11.6 The Adjusted Treatment and Block Totals
11.7 Solving the Intrablock Equations
11.8 Connected Designs
11.9 The Recovery of Interblock Information
11.10 Combining Intrablock and Interblock Estimates
11.11 Some Computational Formulae for Combined Estimates
11.12 Other Combined Estimates
11.13 Maximum Likelihood Estimation
11.14 The General Balanced Design
11.15 Extended Incomplete Block Designs
227
230
230
233
233
235
238
242
242
245
248
Chapter 12. Partially Balanced Incomplete Block Designs
250
The Group Divisible Association Scheme

Intrablock Analysis of GD Designs
The Triangular Scheme
The Latin Square Scheme (L )
The Cyclic Association Scheme
The Intrablock Equations for PBIB(2) Designs
Combined Intrablock and Interblock Estimates
The Simple Lattice Design
Partial Diallel Cross Experiments
Choosing an Incomplete Block Design
252
253
254
255
255
256
259
260
262
265
12.1
12.2
12.3
12.4
12.5
12.6
12.7
12.8
12.9
12.10
Chapter. 13. The Existence and Construction of Balanced Incomplete Block

Designs
268
13.1
13.2
13.3
13.4
13.5
13.6
13.7
13.8
13.9
13.10
270
271
274
275
276
277
281
282
283
286
The Orthogonal Series

Series of Designs Based on Finite Geometries
Designs Using Euclidean Geometries
Designs Derived from Difference Sets
Two Nonexistence Theorems
The Fundamental Theorems of Bose
Designs Derived from Symmetric Designs
Designs with r < 10
Series of Designs Based on Difference Sets
Resolvable Designs
Chapter 14. The Existence and Construction of Partially Balanced Designs 288
14.1 Duality and Linked Block Designs

14.2 The Elementary Designs
14.3 Group Divisible Designs
290
291
292
Contents
14.4
14.5
14.6
14.7
14.8
14.9
Singular Designs
Semiregular Designs
Regular Designs
Designs with the Triangular Scheme
The Nonexistence of Some Triangular Designs
Designs for the Latin Square Scheme
293
293
295
301
305
305
Chapter 15. Additional Topics in Partially Balanced Designs
310
Geometric Designs
Latent Roots of the Intrablock and Interblock Matrices
Cyclic Association Schemes
Partially Balanced Association Schemes with Three Associate
Classes
15.5 The Hierarchic Group Divisible Scheme
15.6 The Rectangular Scheme
15.7 The Cubic Scheme
15.8 The Triangular Scheme for m = 3
15.9 The LS3 Association Scheme
15.10 Cyclic Designs
15.11 Generalized Cyclic Designs
15.12 Designs for Eight Treatments in Blocks of Three
15.13 Conclusion
310
312
315
316
317
318
319
320
321
322
324
325
327
Appendix. Matrices and Quadratic Forms
329
Matrices
Orthogonality
Quadratic Forms
Latent Roots and Latent Vectors
Simultaneous Linear Equations
329
331
332
332
335
15.1
15.2
15.3
15.4
A.1
A.2
A.3
A.4
A.5
Bibliography
339
Index
351
Preface
This book, about the design of experiments, is for the mathematically oriented
reader. It has grown primarily from courses that 1 have given during the past
decade to first- and second-year graduate students in the statistics department at

Berkeley and in the mathematics departments at the Davis campus of the University of California and at the University of Texas.
My interest in experimental design began in 1957 when I became a research
statistician for what is now Chevron Research Company (Standard Oil of
California). I had the particular good fortune to have the benefit of Henry
Scheff as a consultant there, and subsequently as a colleague and friend. That
was the time when Scheff's book on analysis of variance was just about to appear,
when Box and Hunter were introducing the chemical engineers to response
surface designs, and when engineers were becoming interested in 2n designs.
At Davis 1 was able to teach a statistical methods course to graduate students in
agricultural sciences, and at the same time 1 became interested in the combinatorial problems of the construction of incomplete block designs.
The book reflects all these phases. Although the treatment is essentially
mathematical, I have endeavored throughout to include enough examples from
both engineering and agricultural experiments to give the reader the flavor of the
practical aspects of the subject.
The main mathematical prerequisite is a working knowledge of matrix algebra
and an introductory mathematical statistics course. Unfortunately, courses on
matrices nowadays often seem to die before reaching quadratic forms, so 1 have
included an appendix dealing with them. There is no need for a knowledge of
measure theory, and the reader who is prepared to accept Cochran's theorem on
faith will need no calculus at all.
xi
xii Preface
From a practical standpoint, many readers will find the second chapter more
difficult than those that follow it. Unfortunately, this can hardly be avoided,
whether one uses Cochran's theorem or Scheff's geometric approach.
The choice of topics for a one-year course will vary with the interests of the
instructor and the mathematical abilities of the students. The first twelve chapters
are of interest to researchera and form the basis of a one-year course, with the
chapter on Latin squares coming either at the end of the first semester or at the
beginning of the second. This allows the first semester to be devoted to analysis of
variance and the complete factorial experiment. The second semester then
contains 2n and 3n designs, fractional factorials, response surfaces, and incomplete
block designs.
The last three chapters call for considerably more mathematical maturity than
the others, and it is not necessary that the practitioner of experimental design
read them. They take the reader to the frontier of research in the construction of
incomplete block designs, and a course for graduate students with a theoretical
emphasis should include some of these topics. Alternatively, I have sometimes
used these chapters as the basis for an advanced seminar.
I am greatly indebted to my wife, Elizabeth, for her continued help and
encouragement.
Austin, Texas
P. W. M. J.
Preface to the Classics

Edition
1.
Introduction
In the 27 years since the initial publication of this book, much has changed
in both application and theory of the design of experiments. The prime
catalysts have been the personal computer (PC) and its increasingly sophisticated software, and the concomitant burgeoning of experimental design
in industry. There are key debates over fixed, mixed, and random models
in the analysis of variante and over the theory of optimal designs. This
second release gives me a welcome opportunity to comment about these
topics and how they might have changed this book had it been written
today.
2.
Computers
When I signed my contract with Macmillan in 1966, some, but by no means

all, statisticians had access to computers, which were big mainframes in
computer centers. One took one's problem to the center, together with
the necessary Fortran cards to suit the software package, and gave it to an
employee to have the data keypunched; the only package to which I can
now recall having ready access was the Biomed series. The whole thing
was handled through the bureaucracy of the computer center, who usually
charged for their services. Some statisticians in industry had to send their
data away to the computer at company headquarters. So most of us did
what we ourselves could do on our desk calculators. Who of my generation
does not recall the noise of the Friden desktop machine and such tricks as
using a machine with a large register so that one could accumulate x 2 on
the left, 2xy in the center, and y 2 on the right?
Three decades later, we each have PCs of our own with (relatively) userfriendly software packages that will do, in a few seconds, analyses that used
xiii
xiv
to take hours, and without the problems of making a charge to a budget.

Not only can we do many things that we did not have the computer power
to do, we can do them while sitting at our own desks and at our own pace.
I have not carried a job to the computer center in years. This has had an
enormous influence on statistics in several ways. Obviously, we can perform
more complex calculations than before and do our analyses more quickly,
but there is more.
The "more" is manifested in a change of attitude. Graphics are so good
and so speedy that there is no excuse for a scientist not to plot his or
her data and to spot obvious outliers or nonnormality. One can vet one's
data as a matter of routine. More people can look at their own data and
become interested in statistical analysis. They can display their analyses
in presentations to colleagues in tables and charts. This can be a mixed
blessing, however; someone who does not know what he or she is doing can
do a lot of damage. Turning a graduate student loose on a data set armed
with a computer package and manual that are new to him or her (and with
little knowledge of statistics) is a passport to disaster.
I will not take sides and recommend any particular computer package to
the reader, except to say that I use Minitab. I use Minitab for two reasons.
The first is that I started my statistical computing with Minitab's father,
Omnitab, many years ago, and I am too set in my ways to change, even
though I do not question that there are other packages that may do some
things better. The second is that I find it a very easy package to use for
routine teaching purposes. It is very user friendly, and my students, some
of whom are not strong on computer skills, can learn to do many standard
things very easily. The package does things at the click of a mouse that
I would never even have dreamed of attempting 30 years ago. I have not
had to take time to teach the abbreviated Doolittle method of inverting
symmetric matrices for so long that I have probably forgotten it.
One impact of the PC revolution is that, if this book were to be rewritten, I would leave some sections out or radically alter them. I particularly
have in mind Section 3.14 on analysis of covariance and Section 6.7 on
change-over designs with residual effects. The main thrust of these sections
was to develop formulae to make the analysis systematic, and, therefore,
relatively easy on a desk calculator. Today's reader does not need those
formulae and should be spared the pain of reading about them. The missing plot formulae of Section 3.13 and the first exercise in Chapter 6 are
obsolescent, if not actually obsolete. Modern computer routines make the
calculations for such cases easy.
Readers whose interests in computing go beyond the scope of this book
may be interested, for example, in such topics as programs to obtain solutions to the equations that arise when the method of maximum likelihood
is used to find estimates of components of variance. I shall mention later
the use of computers to find D-optimal experimental designs.
xv
3. The growth of experimental design in industry

In 1971, the vast majority of applications of analysis of variance and design
of experiments were still in agriculture. There had been relatively little impact in industry in the United States and Europe outside Imperial Chemical
Industries Ltd. in Great Britain, where George Box and his tolleagues did
pioneering work, and in those oil and chemical companies whose members
had learned about 2 n factorials and response surfaces from Box, often at
Gordon Research Conferences. Other engineers had little opportunity to
learn about that approach to experimentation. After all, Yates's (1937)
monograph was published by the Imperial Bureau of Soil Science and dealt
with agricultural experiments, and there was little communication between
aggies and engineers.
But in the 1980s, the growth of interest in experimental design accelerated as the emerging semiconductor industry became conscious of Japan's
industrial success. Much of that success was attributed to the use of statistics that William Edwards Deming had advocated in Japan after World
War II and, in particular, to the resulting emphasis on designed experiments. The Japanese engineer, G. Taguchi, who had learned about 272
designs from Mahalanobis in India, introduced to the semiconductor industry his methods for small complete factorials and orthogonal fractions.
He was primarily concerned with main effects and paid little attention to
such complications as interactions. Taguchi's method reduces design of
experiments to a relatively simple drill, to which some companies require
strict adherence. That has some merit, as long as it works, but it does not
lead to a very deep understanding. However, Taguchi went further. He
made us think not just about experiments to find conditions that optimize
yield, but also about experiments to control the variability of a manufacturing process.
At about the same time, the advent of the PC made the requisite calculations much easier. We no longer had to use Yates's algorithm to analyze
2n factorials. It became easier to analyze them, or their fractions, by regression, especially since the I.C.I. group had persuaded us to think of the
levels of the factors as 1, +1 instead of 0, 1, as Yates had used. Running nonorthogonal fractions was no longer so daunting because we could
actually analyze the data when it came in! The journal Techinometrics,
which had first appeared in 1959 and to which there are numerous references in the 1971 edition of this book, thrived, publishing numerous papers
on design of experiments.
Engineers came to realize that they no longer had to associate designed experiments solely with the large complete factorial experiments,
planted in the spring and gathered in the fall, that were the tradition of
xvi
the agronomists. They could think of series of small experiments: running

an exploratory fractional experiment and then, after seeing the results,
running further points to answer the questions that the analysis of the first
experiment raised.
Performing a set of small experiments in sequence quite logically leads
one to consider running the points of an individual experiment in a planned
sequence. Often they will be run one after another anyway, which itself
implies a sequence. Why not arrange the order so that the estimates of the
effects will be orthogonal to time trends? Cox (1951) was the first to write
about this. Hill (1960), Daniel and Wilcoxon (1966), Steinberg (1988),
and John (1990) all wrote about the topic in Technometrics. Dickinson
(1974) wrote about sequences from another point of view: How can we
order the points so as to minimize the cost of all the changes in factor
levels that a factorial experiment requires? Another relevant question is
how to order the points so that early termination will still leave us with a
viable, though incomplete, set of data? (Section 8.14 touches upon the last
of these questions.)
4. Fixed, mixed, and random models

In 1971, the question of the expected values of the mean squares in the
analysis of factorial experiments seemed to have been settled by Scheff
(1956). In Section 4.7, I gave an account of Scheff's method of analyzing
the two-factor experiment in which one factor, A, is fixed and the other,
B, is random. One tests the mean square for A against the mean square
for interaction; on the other hand, one tests the mean square for B against
the mean square for error. An essential feature of Scheff's model is that,
since A is fixed, we should impose the side condition > (a/3)jj = 0 for each
j. In deriving E(SB) we can argue that we sum (average) the interactions
over all the levels of the fixed factor A; hence, they "cancel out," and no
interaction term appears in E(SB).
Hocking (1973) introduced an alternative model in which the interactions were freed of those side conditions. In Hocking's model the interaction
term appears in both E(SA) and E(SB), and so MA and MB are each tested
against MAB. Scheff's model has come to be called the restricted model
and Hocking's the unrestricted model. Practitioners are given a difficult
choice. Bach model has a logical basis, and the differences are subtle; each
model can be defended and each has its advocates. I had no difficulty in
1971. Hocking had not yet written his paper. The only approach of which
I was aware was Scheff's, and so that was the only one that I described.
In Section 5.3 I gave a set of rules for computing the expected values of
the mean squares (EMS) in a crossed classification for three factors (using
Scheff's model). One writes down the EMS as if all three factors were
xvii
random, and then one strikes out certain interaction terms that involve
fixed effects, following rule 4. The reader should now call this rule 4r (to
correspond to the restricted rule) and add the following rule 4u (if he or
she wishes to use the unrestricted rule).
4.1. Rule 4u
In the unrestricted rule, an interaction is called fixed if and only if all its
factors are fixed. Otherwise, it is random. In the expected mean square
for any effect, strike out any interaction that is fixed and leave the others.
Do not strike out an interaction from its own EMS.
In the case of three factors with A random and B and C fixed, the only
fixed interaction is BC. We strike it out of the expectations of the mean
squares for B and C, but we do not strike it out in the line for BC itself.
Because A is random, we do not strike out the ABC interaction from any
mean square.
The restricted rule is much more liberal than the unrestricted rule in
striking out terms.
These distinctions also apply to the general balanced design in which
some factors are crossed and others are nested. We can still begin by
following the first four steps of the procedure of Bennett and Franklin that
is described in Section 5.6.
Step 5 should be revised as follows.
Write the EMS, assuming that all the factors are random and following
the rules that (i) each line of the EMS table contains all the terms that have
as subscripts all the letters in the title of the line, and (ii) the coefficient of
any component is the product of the levels of all the factors that correspond
to absent subscripts. Then strike out terms using step 5r or step 5u.
4.2. Step 5r (the restricted rule)

If some of the factors are fixed, we strike out of the expectation of any
mean square all those terms that contain in their live subscripts any letters
(other than those in the name of the mean square itself) that correspond
to fixed effects.
4.3. Step 5u (the unrestricted rule)

We strike out
fixed effects.
an interaction term only if all its live subscripts belong to
xviii
5. Optimal designs
Kiefer (1959) presented to the Royal Statistical Society a paper about his
work on the theory of optimal designs. How do we choose between one
design for p factors in n points and another? How do we find the best design? This book mentions the criterion of minimizing the average variance
of the estimates of the effects, which came to be known as A-optimality or
trace-optimality.
Kiefer spoke of D-optimality, choosing the design that maximized the
determinant, D, of X'X, where X is the design matrix. How does one find
such a design?
The factors can be discrete or continuous. Users of 2n factorials have
the discrete case with x i = 1. Mitchell (1974) published his DETMAX
algorithm, enabling us to find a design that maximizes D when the experimental space has a finite number of points. His algorithm has since been
improved and is now available in some standard software packages, so that
an engineer has only to type in the values of n and p and the model (with
or without interactions), and out comes a design.
It is all very easy, but there are questions in addition to the obvious
possibility of local maxima. The criterion is very sensitive to the model.
Suppose that the model has q terms. If you drop one of the terms, a
design that was optimal for q terms in n runs is not necessarily optimal for
q 1 terms in n runs, nor does an optimal design for q factors in n points
necessarily contain a subset that is a D-optimal design for n 1 points.
Also, the solution in any case may not be unique. If one permutes the
columns of X, or if one multiplies any column by minus one, which corresponds to changing the level of that factor at every point, D is unchanged.
One can call such solutions isomorphic to one another. But there are nonisomorphic solutions in some cases. In the small problem of fitting main
effects and two-factor interactions for a 2 4 factorial in 12 points, there are
4 families of nonisomorphic solutions, all with the same maximum value of
D, and one of them is a foldover design! Furthermore, none of them is an
optimal plan for main effects only. In this case, an optimal design is given
by any 4 columns of the Plackett and Burman design for 11 factors in 12
runs.
The continuous case is beyond the scope of this book. Its mathematical
demands go further into measure theory and functional analysis than most
of our readers wish. It has become a flourishing field of research among the
more mathematically oriented statisticians. A survey of the field written by
Atkinson (1982) has more than 100 references, including books by Fedorov
(1972) and Silvey (1980). A follow-up survey by the same author in 1988
has many more.
xix
6. Incomplete block designs

I have written of some areas that have grown in interest since 1971. Sadly,
I must comment on an area which has quietly disappeared from the radar
screen of experimental design: incomplete block designs. In 1971, it was an
important topic, and it still provides opportunity for some interesting and
elegant mathematics. I have enjoyed working on some of its problems over
the years, but I must confess that it is an obsolescent topic, a victim to the
advance of computers. I contributed a paper to the 1982 commemorative
volume for Yates about running 2'factorials in incomplete block designs,
and published a follow-up paper in 1984, but applications of the results
seem to have been few and far between, if, indeed, there have been any!
7. Finis
By contemporary standards, mine has been an unusual career. Nowadays
youngsters are expecteded to set a straight path almost as soon as they
become upper-division undergraduates and pursue it doggedly to the end.
My path has zigged and zagged with the help of some outstanding teachers
to whom I am deeply grateful. I began as a pure mathematician and first
encountered statistics in 1948 in the postgraduate diploma (now M.Sc.)
course at Oxford, where I was taught by David Finney, P.A.P. Moran, and
Michael Sampford. Upon coming to the United States, I set statistics aside
for a few years while earning a Ph.D. in probability.
Then an extraordinary opportunity occurred in industry. In 1957 I
joined Chevron Research Corporation in Richmond, California, as their
first statistician. In effect, that gave me a 4-year postdoctoral period working with Henry Scheff, whom Chevron had retained as its consultant in
the previous year. Chevron sent me to Gordon Conferences on Statistics,
which were then dominated by George Box and his pioneering work in the
application of design of experiments to chemical engineering and, in particular, to response surfaces. The conferences had an electric atmosphere
of intellectual ferment. At the same time, Henry set me to read the Box
Wilson (1951) paper and the I.C.I. book on industrial experimentation,
edited by my fellow Welshman Owen Davies. I learned a lot from that
book. The level of exposition is excellent and, unlike the spring of 1949,
when Sampford was lecturing about design to a young mathematician who
had never seen a real experiment, I now knew what an experiment was.
Real ones were going on all around me.
After 4 invaluable years at Chevron, I chose to return to the academie
world, where I have been ever since. I greatly enjoy teaching. Indeed, at
Henry's instance, for the last 3 of those years I had taught a course in the
statistics department at the University of California, Berkeley, in addition
xx
to my work at Chevron. Now I also wanted the freedom to work on any

problem that attracted me, regardless of its importante to the company.
An ideal opportunity beckoned from the mathematics department at the
University of California, Davis, one of the world's foremost agricultural
research domains. This involved me in a remarkable new array of experiments, including the large factorials of the agronomists with their fixed,
mixed, and random models. Finally, in 1967, I moved to a large mathematics department at the University of Texas, which has no tradition of
responsibility for any statistical services to the campus community.
This book, begun in Davis, was largely written in the relative isolation
of Austin, where I have remained for personal reasons. Happily, continuing
involvement in the Gordon Conferences and consulting kept me in touch
with real-world stimuli until another remarkable opportunity emerged from
the new semiconductor industry.
In cooperation with the federal government, a consortium of companies
in that industry established in Austin a research facility called Sematech
Corporation. For several years I have been fortunate to work with members
of their statistical methods group. My collaboration with them has led to
work in such areas as nonorthogonal fractions, D-optimal designs, and
multivariate analysis of variance. I have learned that engineers do indeed
carry out large factorial experiments with crossed and nested factors and
mixed models; they call them gauge studies!
My path has zigged and zagged but, in retrospect, each turn ultimately
afforded stimulating new experiences and opportunity for growth. More
than ever, I share Henry Scheff's (1958) conviction that industry is a
valuable source of research problems. The first question that he was asked
as a consultant for Chevron led him to invent a new branch of experimental
design: experiments with mixtures. He told me repeatedly that there are
dozens of interesting problems out there in industry, just below the surface,
waiting to be solved, if only one is alert enough to spot them and run with
them. That brings me to the role of serendipity and the thesis that a lot
of useful research comes from a combination of luck, accident, and being
in the right place at the right time.
This is true of two problems, to which I have referred above, from my
days at Chevron. The idea of running the points of a factorial experiment
in a prescribed sequence so as to protect against early termination came
Erom an accident. A client wanted to do a 2 factorial on a pilot plant which
he had at his disposal for 2 weeks, enough to do one complete replicate.
I knew the "correct" thing to do. We carefully listed the 16 points and
arranged them in random order. Fate struck. After 8 points, the pump
quit, and could not be repaired within the 2-week period. Have you ever
seen an experiment that consists of 8 points chosen at random from all
16? It is a mess. With the benefit of hindsight, we realized that it would
have been a lot smarter if, instead of randomizing the 16 points, we had
xxi
arranged for the first eight points to be either a resolution IV half replicate
defined by I = +ABCD, or I = ABCD, in random order. Then we
would have had a design with good salvage value when the darned pump
broke. Clearly, it would have been better to modify the randomization.
The other example led to my interest in nonorthogonal fractions. I had
persuaded an engineer to try a 2 factorial. The raw material came from a
facility in Scotland by slow boat. They sent enough for only 6 runs. I knew
what to do with 8 runs; I knew what to do with 4 runs; but 6 was betwixt
and between. I had to work out what to do. That led me to 4 factors in
12 runs, and thus to three-quarter replicates.
Both problems arose by accident. They should never have happened.
A referee of a paper about one of them objected that the problem was
contrived. It certainly was not. It was a real-life situation that probably
would not have occurred to me had I just spent my time in an ivory-tower
office thinking about mathematical theorems in the design of experiments.
Henry Scheff was right. There are all kinds of practical research problems lurking in industrial applications, just waiting to be found and solved.
I advise students to keep their eyes open and spot them. Some can be
tremendously interesting and full of theoretical challenge. Above all, being
helpful to people is rewarding in itself.
I end with the sentence that concluded my original preface. It is still
true. I am greatly indebted to my wife, Elizabeth, for her continued help
and encouragement.
Austin, Texas
P. W. M. J.

ATKINSON, A.
C. (1982). "Developments in Design of Experiments." Int. Statist.
Rev., 50, 161-177.

A. C. (1988). "Recent Developments in the Methods of Optimum
and Related Experimental Designs." Int. Statist. Rev., 56, 99-116.
ATKINSON,
Box, G. E. P., AND K. J. WILSON (1951). "On the Experimental Attainment

of Optimum Conditions." J. Roy. Statist. Soc. Ser. B., 13, 1 -45.
Cox, D. R. (1951). "Some Systematic Experimental Designs." Biometrika, 38,

312-323.
F. WILCOXON (1966). "Factorial 2P -9 Designs Robust Against
Linear and Quadratic Trends." Technometrics, 8, 259-278.
DANIEL, C., AND
L. (1954). Design and Analysis of Industrial Experiments. Oliver and

Boyd, London.
DAVIES, O.
A. W. (1974). "Some Run Orders Requiring a Minimum Number of

Factor Level Changes for the 2 4 and 2 Main Effect Plans." Technometrics,
16, 31-37.
DICKINSON,
FEDOROV, V.
(1970). Theory of Optimal Experiments. Academic Press, New
York.
HILL,
H. H. (1960). "Experimental Designs to Adjust for Time Trends." Tech-
nometrics, 2, 67-82.
HOCKING,
R. R. (1973). "A Discussion of the Two-way Mixed Model." Amer.
Statist., 27, 148-152.

JOHN,
P. W. M. (1982). "Incomplete Block designs for the 2 Factorial." Utilitas
Math., 21, 179-199.

JOHN, P. W.
M. (1984). "Incomplete Block Designs for the 2 4 Factorial." Utilitas
Math., 26, 79-88.

JOHN,
P. W. M. (1990). "Time Trends and Factorial Experiments." Technomet-
rics, 31, 275-282.

KIEFER, J.
(1959). "Optimal Experimental Designs (with discussion)." J. Roy.
Statist. Soc. Ser. B, 21, 272-319.

(1974). "An Algorithm for the Construction of `D-Optimal"
Experimental Designs." Technometrics, 16, 203-210.
MITCHELL, T. J.
xxiii
xxiv
SCHEFF, H. (1956). "A Mixed Model for the Analysis of Variance." Ann. Math.
Statist., 27, 23-36.
SCHEFF, H. (1956).
"Alternative Models for the Analysis of Variance." Ann.
Math. Statist., 27, 251-271.

SCHEFF, H. (1958).
20,344-360.
"Experiments with Mixtures." J. Roy. Statist. Soc. Ser. B,
John Wiley and Sons, New York.

SILVEY, S. D. (1980). Optimal Design. ChapmanHall, London.
STEINBERG, D. (1988). "Factorial Experiments with Time Trends." TechnometSCHEFF, H. (1959). The Analysis of Variance.
rics, 30, 259-270.

YATES, F. (1937). The Design and Analysis of Factorial Experiments.
Bureau of Soil Science, Harpenden, England.
Imperial
CHAPTER
Introduction
It is the view of many statisticians, amongst whom the author counts himself,
that statistics is a branch of applied mathematics, and that its health and growth
depend upon the involvement of the statistician in practical problems. This book
deals with one of the interfaces between applied mathematics and scientific
investigation: the design of experiments.
To some scientists, the idea that a statistician can contribute substantially to
the design of an experiment is, if not patently absurd, at least dubious. A common
image of statisticians is one of people who are rather clever with figures and hence
useful for reducing data to tables; the statistician might be of some help in the
analysis, but only if there are reams of data to feed into computers. One hopes that
this false image is fading with the passage of time. Meanwhile, however, this
attitude frustrates the statistician and deprives the scientist of valuable help.
It is true that the statistician is not usually an expert in the particular areas of
biology or engineering in which the experiments are being made. It is true also
that, given a computer, or for that matter enough time, pencils, and paper, one can
fit a straight line or a plane to almost any hodgepodge of data whether relevant or
not. The bigger the computer, the more variables it can handle in a multiple
regression program, and, unfortunately, the more miles of useless numbers it can
print out. It is clear, therefore, that some planning before the data is obtained is in
order, and that the planning had best be made with the method of analysis of the
experiment as well as its objectives in mind.
It follows that the role of the statistician in an experimental program should be
not merely to analyze the results of the experiment, but also to help the scientist
plan his experiment in such a way as to produce valid results as efficiently as
possible. There is little sense in waiting until the end of the experiment and then
Introduction
[Ch. 1
visiting the statistician for the first time with a pile of data to deal with as best he
can. Every applied statistician can tell horror stories of having been asked after
the event to analyze data from lengthy and expensive experiments which were so
badly planned that no valid conclusions could be drawn from them. There are
also numerous experiments which, although valid results have been obtained,
would have yielded even more information if skilled knowledge of design had
figured in the planning, and others in which the results obtained could, with
efficient design, have been obtained with less effort and expense.
This book is written for the mathematically oriented statistician, scientist, and
engineer. The basis of all the theory of design and analysis will be the linear model
and the analysis of variance, and throughout we shall refer the reader to the book
by Henry Scheff (1959). It is the most authoritative book available on the
analysis of variance and is invaluable as a reference book. The examples that we
include are drawn from various fields; most of them are taken from the published
literature, and some of them come from experiments in which the author has
participated as a consultant.
1.1 The Agricultural Heritage

The pioneer in the use of mathematical statistics in the design of experiments
was the late Sir Ronald Fisher. He dominated the bistory of experimental design
between the First and Second World Wars, 1918-1939. During the early part
of that era he was in charge of statistics at the Rothamsted Agricultural Experiment Station near London. In 1933 he succeeded Karl Pearson as Galton
Professor at the University of London and later he moved to a professorship at
Cambridge University. It was Fisher who introduced the analysis of variance.
He wrote two books about experimental design, each of which has run into
several editions, even though neither is by any means easy to read: Statistical
Met hods for Research Workers, which first appeared in 1925, and The Design of
Experiments. Dr. F. Yates joined Fisher at Rothamsted in 1931, and throughout
this book the reader will find references to the major contributions of Yates and
his colleagues at the agricultural experiment station.
From this background it has naturally followed that the common terminology
of experimental design comes from agriculture, where they sow varieties in or
apply treatments to plots and often have the plots arranged in blocks. We shall
use this terminology except when we think in terms of engineering applications,
and then we shall talk about points or runs. The terms point and run will be
used interchangeably; we make a run on the plant at a predetermined set of
conditions, and the result will be an observation or a data point (pedants would
say datum point). It sometimes puts engineers off to hear a statistician talk
about plots when he discusces their experiments. Some engineers tend to jump
to the superficial conclusion that the statistician is locked into a rigid framework
of agricultural thinking and hence can contribute little to the solution of
engineering problems.
Sec. 1.2]
An Example of an Experiment 3
This ignores the essential fact that the statistician really thinks in terms of
linear mathematical models. He is distinguished from many experimenters, who
are deeply involved in their own specialized sphere, by mathematical training
with a certain amount of emphasis on abstract thinking. This makel it easier
for him to see the essential mathematical similarity between an experiment
comparing varieties of barley and an experiment comparing operating temperatures in a catalytic cracker at an oil refinery.
The process of designing an industrial experiment is not, however, just a
matter of taking a standard design from an agricultural handbook and substituting temperature for variety and run for plot. Indeed, differences do exist
between agricultural and industrial experimentation, and, with all the incumbent
risks of generalizing, we suggest a few. One is that the agronomist usually has to
sow his experimental plots in the spring and harvest them in the fall. Plots and
plants are relatively cheap, and the emphasis has to be on. designing relatively
complete experiments; if anything is omitted it will have to wait until next year.
On the other hand, much industrial experimentation can be carried out relatively
quickly, in a matter of days rather than months, but experimental runs are
often very expensive. The emphasis should, therefore, be on small experiments
carried out in sequence, with continual feedback of information, so that in
designing each stage the experimenter can make use of the results of the previous
experiments.
The agronomist is more interested in the tests of significance in the analysis
of variance than is the industrial experimenter. One of the reasons is that the
agronomist is more often concerned with uniformity trials. Can he produce
strains of seed that are essentially equivalent? Can he produce a new variety of
seed that will do well, not just on the experimental farm at Davis, California,
but all up and down the hot Central Valley and perhaps also in the coastal
region, in the other states where the erop is grown, and even in other countries?
He wants to be able to accept the null hypothesis of uniformity. The position
of the industrial experimenter often differs from that of the agronomist in two
ways: he frequently knows before starting the experiment that his treatments
are not all the same and is interested in finding out which ones differ and by
how much. His emphasis will be on estimation rather than hypothesis testing.
He will sometimes argue that failure to reject the null hypothesis is merely the
result of taking too small an experiment. Expecting to reject the null hypothesis,
he is more interested in confidence intervals.
In the sections that follow we shall introduce the topics of the various chapters
by considering a typical experiment.
1.2 An Example of an Experiment

Suppose that a chemical engineer hopes to improve the yield of some petrochemical in an oil refinery plant by comparing several catalysts. Crude oil is fed
into the plant, which is charged with the catalyst; some of the crude oil, or
4 Introduction
Ch. 1]
feedstock, passes through the plant unchanged; some is converted into the
petrochemical, or product. The liquid that comes out of the plant is separated
into product and unconverted feedstock, and the yield, or response, is the
percentage of feedstock converted into product.
An obvious procedure is to make one or more plant runs using each of the
catalysts and to compare the average yields on each catalyst. There are, however, some other considerations that enter the picture. How many catalysts?
How many runs? How do we compare these averages after we have obtained
them? Let us assume for the moment that the crude supply is plentiful and
consider just two catalysts, A and B, which make up the simplest situation.
We take r runs on each catalyst, N = 2r runs altogether, and we obtain average
yields yl on A and y2 on B.
The first of our mathematica) models now enters the picture. Let y ij denote
the yield on the jth run made with the ith catalyst; y ij is a random variable and
we assume y t; = m i + e ij where m i is the true value for the ith catalyst and e^ 5 is
a random component of error or noise. We further assume that on the average
the noise contribution is zero; sometimes it is positive and sometimes negative,
but it averages out over the long haul, in the sense that E(e 5 ) = 0, and so
E(y11) = m ; . Then yt will be an unbiased estimate of m i , and y l y2 will be an
unbiased estimate of m l m 2 .
The word random should not be passed over lightly. The use of randomization
is the keystone of the application of statistical theory to the design of experiments, and the validity of our deductions rests upon the principle of randomization. The very least that can be said as a recommendation for randomization is
that it is only prudent to avoid the introduction of systematic bias into an
experiment. An agronomist comparing two varieties of corn would not rationally
assign to one variety all the plots that were in the shade and to the other all the
plots that were in the sun. If he did, he would not be able to teil after the experiment whether any apparent difference in yields resulted from varietal differences
or from the fact that one variety had more sun. Furthermore, Fisher (1947) has
shown that the fact that the treatments have been randomly assigned to the runs
is itself an adequate basis for obtaining tests of significance and confidence
intervals.
The other maxim of experimental design emphasized by Fisher was replication
to provide a proper estimate of the error variance, V(e t; ). If we were to take
r = 1 and obtain y l = 80, Y2 = 79, we should not be able to make conclusions
about the relative potencies of the catalysts with much confidence. We could pick
the winner and recommend catalyst A because y l l > y21, but we should have no
idea whether the apparent difference was genuine or a result of noise (error). If
we made several runs with each catalyst, and found that the observations were
highly repeatable, i.e., that the range of the observations on A, and the range of
the observations on B were both very small, we could be happy about concluding
that m l > m 2 when we observed y l > y2 . On the other hand, if the observations
on both A and B were scattered between 60 and 90, a difference between the
means of only one would hardly be convincing.
Sec. 1.3]
Tests of Hypotheses and Confidence Intervals 5
1.3 Tests of Hypotheses and Confidence Intervals

Deciding whether or not an observed difference between the sample means
yl and y 2 is convincing evidence of a real difference between m l and m 2 is
equivalent to testing the hypothesis H o : m l = m 2 . There is no need for us to go
into the details of the NeymanPearson theory of hypothesis testing at this
time, or to discuss the merits of one-tailed or two-tailed tests. The reader will be
familiar with these topics already. Nor is this the place to rekindle the Hames of
controversy between the confidence intervals that we shall employ and fiducial
intervals; it suffices to say that in practice the intervals are the same and the
argument is of little import.
It is at this stage of the proceedings that we have to make some decisions about
what we shall assume to be the nature of the random errors e t; . We mention
three possibilities:
(i) We could assume only that the treatments were assigned to the runs, or
the plots, by some process of randomization. This would lead us to the
consideration of randomization models and permutation tests. We shall
not take this road because not very much has been done with this
approach for designs other than randomized complete blocks and Latin
squares. The reader may wish to refer to the discussions of the randomization models and tests given by Kempthorne (1952) and (1955), Scheff
(1959), Wilk and Kempthorne (1955) and (1957). Nor shall we consider
the nonparametric rank order tests of the Wilcoxon type.
(ii) We could assume that the errors are (a) uncorrelated random variables
with zero expectations and the same variance o 2 , or else (b) random
variables with zero means and a covariance matrix that is of known form.
(iii) We could add to assumption (ii) the additional assumption of normality.
For obtaining our estimates under assumption (ii), we shall rely upon the
method of least squares. For testing hypotheses we shall make the normality
assumption and use F tests and t tests.
Returning to our earlier model under assumptions (iia) and (iii), we now have
y i; = m t + e t; where the errors e i; are independent normal random variables,
each with mean zero and variance o 2 . The least squares estimates m t of the
parameters m i are obtained by minimizing the sum of squares of the residuals,
se = (Yii mt) 2 ^
t i
and they are, as we suspected all the time, m t = i . We can also obtain an estimate of a 2 by comparing repeated observations on the same treatment. Looking
first at A, J ; (y, ; y^) 2 is an unbiased estimate of (r 1)u 2 . If there are t
treatments, E(Se ) = t(r 1)a2.
[Ch. 1
6 Introduction
When two treatments are to be compared, the statistic z = (j y2 )/^/(2Q 2 /r)

has a normal distribution with mean m 1 m 2 and variance unity, which we
shall write as z N(m 1 m 2 , 1). If Q 2 is not known, we shall substitute for Cr 2
the estimate s 2 = SQ /(rt t). The resulting statistic, when m l = m 2 , has
Student's t distribution with t(r 1) degrees of freedom.
No great difficulties are introduced if the numbers of observations are not
the same for each treatment. Suppose that there are r ; observations on the
ith catalyst. As before, m ; = y; , and now
E ( Se)
(rt 1)a 2 = (N t) 2 ,
= E{^ (j)t1 yt) 2} _
where N = ^, r1 . The estimate of a 2 is thus 5 2 = Se /(N t). When the hth and
ith treatments are to be compared, the statistic
Yn Yt
_
_
t s(1/rh
+ 1/ri) 1I2
has, if m h = m i , Student's distribution with (N t) degrees of freedom.

There is, however, a problem if the variance of ejj changes from treatment
to treatment. Consider again two treatments and suppose that the variances
of the random variables involved are V(e 11 ) = a and V(e 21 ) = 02. Then, if
there are r{ observations on the ith treatment,
V (Yl Y2) = r1 l al
+ r 2 1 az = v 2 ,
and (y1 Y2) N(m l m 2i v 2 ).

If cri and oZ are known, v 2 is easily computed, and the distribution of
(Yl Y2)/v is standard normal. If, however, neither oi nor a2 is known, the best
estimates of them from the data are s; = (y y ) 2 /(r, 1).
We let w; = rti 's?. Unfortunately, the random variable t' _ (y l Y2)/
(w 1 + w2 )" 2 does not have the Student's distribution and its distribution has to
be approximated. There are several approximations that can be used, and we
mention two of them. The disparity between the distributions of t and t' is least
when r 1 = r 2 .
Cochran and Cox (1957, page 101) give a procedure for obtaining an approximation to the critical value of t' for any level a. Let 0 { = r t 1. Then the
critical value is t* = (w l t l + w 2 t 2 )/(w l + w 2 ), where ti is the critical value, at
the desired level a, of Student's t with 0 t degrees of freedom.
Welch (1947) approximates the distribution of t' by a Student's t distribution
with 0 d.f., where
1
0 = 919'2(wl + W2) 2 /(6 1w1 + 2w2)
The assumption of equal variances is usually not unreasonable, especially

when two similar treatments are being compared. Scheff (1959) devotes a
Sec. 1.4]
Sample Size and Power 7
chapter to the consequences, in the analysis of variance, of departures from

assumptions (ii) and (iii). In some cases where the variance is believed to depend
upon E(y) in a known manner, the variance can be stabilized bytransforming
the data. These transformations do not necessarily eliminate the nonnormality,
but they leem to alleviate it somewhat.
If V(y) = {g(y)} 2 , and if z is a single-valued function of y, the variance of z is
approximated by
vz = (azl a ) 2 V(y).
If v is to be a constant, we therefore need to have
z = c f [g(Y)]
-1
dY,
where c is an arbitrary constant. If g(y) is directly proportional to y, the transformation z = In y is appropriate; in y denotes the natural logarithm of y;
log la y will do just as well. For binomial data the transformation z = Arcsin y
may be used, but this is often not done if the range of y is not large, especially
when p is about 0.5. When p = 0.5 we have pq = 0.25, and when p = 0.3 or
p = 0.7 the product pq only decreases to 0.21.
1.4 Sample Size and Power

If the experimenter has at his disposal a prior estimate of the variance o 2 , he
is in a position to make some approximate power calculations. Most readers will
already be acquainted with the procedures for computing r in the case of only
two treatments, when type I and II errors are given. If the experimenter has
decided to use a one-sided normai test with type I error a to compare two
means, he will reject the hypothesis Ho : m l = m 2 in favor of the alternative
HA : m l > m 2 if y l y2 > c. The cutof point c is given by c = z a a /(2/r) where
Za is the upper a point of the standard normal density, i.e., P(z > za ) = a.
Suppose now that he wants to choose r to be so large that if m l = m 2 + So
he will have probability t1 P) of deciding that m l > m 2 , i.e., of having
yl Y2 fall in the critical region. He must then have
P(Y 1 y 2 <c^m l m 2 =8a)=P.
It follows that
SQ c = z s u/(2/r)
and
8 = (za + zs )V(2/r).
This gives an approximate value

r=2(za+zs)2/S-2
8 Introduction
[Ch. 1
In practice, however, the sample size is often determined not by calculations

of power but by real or imagined considerations of the budget. If a prior estimate
of the variance is available, a variation of the previous calculation is still worth
carrying out. To revert to our earlier example, suppose that previous experience
in the plant gives a prior estimate az = 2 and that the engineer wishes to detect
that the new catalyst A is better than the old catalyst B if m l m 2 >_ 2 =
but that he is only prepared to make n = 4 runs with each of the catalysts. Let
us suppose further that he wishes to use a = 0.05. We have z a = 1.645, 6 = ^/2.
Then z s = 0.355 and P = 0.36, so that there is about one chance in three that
with n = 4 he will fail to reveal an actual improvement of 2 units with his
experiment. If 6 had been one, corresponding to m l m 2 >_ 1.4, the chance of
detection would have been only about four in ten (0.4). This kind of information,
even if it is a rough approximation, is salutary to the engineer. It should at least
serve to make his hopes for the outcome of his experiment more realistic. It may
even prompt him to abandon or to revamp completely an investigation that was
doomed to be unproductive.
When there are several treatments, similar calculations can be made involving
the power of the Ftest. This is discussed by Scheff (1959), Graybill (1961), and
Kempthorne (1952). Scheff uses the charts by Pearson and Hartley (1951) and
by Fox (1956). The other two books contain Tang's tables based upon the
incomplete Beta distribution.
1.5 Blocking
There are two ways in which we can reduce the variance of m l m 2 . We
can take more data points, i.e., increase r and thereby increase the cost of the
experiment, or we can somehow improve the design.
Some reduction can be made by refinement in experimental technique and by
improved methods of measurement and chemical analysis, but these improvements are limited in their achievement and there still remains what might be
called the natural noire of the system. Some engineers have difficulty accepting
the fact that a system or a procedure has a built-in variability. This is true for
analytical procedures, and the American Society for Testing Materials has tried
hard to make its members realize that in standard measurements a certain amount
of variability is to be expected because of the nature of the measurement
procedure. Hence, failure to have duplicate measurements identical is not
necessarily an indication of incompetence on the part of the operator. It is
regrettable that in some laboratories duplicate measurements are useless
because the atmosphere is such that the operator feels so strong a pressure to
have his second reading conform to the first reading that the second reading
can by no stretch of the imagination be considered as an independent observation. W. J. Youden has suggested that second readings obtained in this way
should be regarded not so much as duplicate readings as duplicity; they certainly
Sec. 1.5]
Blocking
give the scientist a false sense of confidence in his results. Genuine duplicates can
be obtained by sending the analytic laboratory coded samples, e.g., pouring a
gallon of product into eight pint bottles and calling for a single determination
on each of the eight "different" samples. It is a nuisance to have to do this, but
it is often worth the trouble, and the results can be quite revealing.
One way of improving the precision of estimates of differences, such as
m l m 2i is by blocking. If there are t catalysts, perhaps we can take a barrel of
crude oil, divide it into t parts, and use one part with each catalyst. The agricultural equivalent is to divide a field into blocks and then to subdivide each block
into t plots. This topic is discussed at the end of Chapter 3. When we have a
completely randomized design in which the runs are made in completely
random order, or in which the plots with any given treatment are scattered all
over the field, then comparisons between treatments contain additional variability from the fact that some runs are made from one barrel and some from
another, or that some are made on good land and some on stony land. The idea
behind the blocking is that by balancing the treatments over the blocks we make
the differences between the blocks cancel out when we compare the treatments.
The individual estimates of m l and m 2 are no better than they were before, in the
sense that V(y) has not been reduced, but V(y l y 2 ) has hopefully suffered a
considerable reduction.
This is the randomized (complete) block experiment, and the reader should
note that the word randomized appears again. We assign the treatments at
random to the runs within each block; we should perhaps be suspicious if we
always used the first sample drawn from each barrel for catalyst A and the last
sample for catalyst B that there might be some systematic bias introduced. We
modify our earlier model and, letting yl; denote the observation on the ith
treatment in the jth block, we write y tf = m ; + b1 + e{; where b; is the effect of
the jth block. We shall actually write the model a little differently later when we
come to look at it a little more closely.
This procedure works well so long as each barrel of crude oil is big enough
to handle t runs, and the randomized complete block design is said to be the
most commonly used experimental design. If, however, the blocks are not big
enough to accommodate all the treatments, we have to resort to incomplete
blocks, with k (<t) treatments being contained in each block. The topic of
incomplete block designs is introduced in Chapter 11. If we cannot arrange to
have every pair of treatments appear together in every block, perhaps we can
arrange the design so that every pair appears in A blocks where t is a constant.
Such a design is called a balanced incomplete block design. These designs were
introduced by Yates in 1936, but long before then such arrangements of symbols
had been of interest to pure mathematicians. Number theorists have for years
investigated the combinatorial properties of these designs without any care for
possible statistical applications. Chapter 13 is devoted to the problems of the
construction of balanced incomplete block designs. It is a chapter for mathematicians, and the experimenter who simply wants a particular design for his
experiment might be content to look it up in the Fisher and Yates tables.
10 Introduction
[Ch. 1
The restriction that each pair of treatments shall appear in exactly A blocks is
demanding, and designs do not exist for all values of t, k, and A. In 1939 Bose
and Nair introduced partially balanced incomplete block designs. These designs
are the topics of Chapters 12 and 14, the latter chapter being devoted to the
construction of the designs for different partially balanced schemes. In partially
balanced designs each treatment appears with n l of the other treatments ,1 1 times,
with n 2 of them d 2 times, and so on. There are also some further restrictions, but
they can wait until Chapter 12.
Just as one would not normally use an incomplete block design when a
suitable complete block design is available, so balanced incomplete block
designs are preferable to partially balanced designs because of their higher
efficiency. The measure of efficiency is the average of the variances of the
estimates m,, m i for all pairs of treatments; the measure is made scale free by
dividing it by the average that would be obtained for a complete block design
with the same number of observations, if such a design could have been used.
Kempthorne (1956a) conjectured, and J. Roy (1958) proved that for any given
set of numbers N, t, k (k < t) the most efficient incomplete block design is the
balanced design; that is, if one exists.
1.6 Factorial Experiments

We have just talked about blocking as a technique for improving the precision
of the estimates (m h m i ). The block effects b. canceled out of the comparison
between the treatments. We were thinking at th time of the block differences as
a nuisance to be eliminated. The idea was to avoid the differences between the
crude oils by balancing them out over the treatments. We could, however,
purposely choose one barrel of Venezuelan crude oil, one of Sumatran, one of
Californian, one of Texan, and so on. This would give us two factors: catalysts
and types of oil. The structure is very similar to that of the blocked experiment.
Just as each catalyst "sees" each barrel of oil, so each type of oil "sees" each
catalyst. The crude oil differences cancel out of comparisons between catalysts;
the catalyst differences cancel out of the comparisons between crude oils.
We can rewrite our model to take into account two factors, A and B, the latter
standing either for blocks or for another factor:
yti; =m+a i +b 5 +ejj .
Here m is a grand mean effect, a, is the main effect of the ith level of A, and bj is
the main effect of the jth level of *B. We talk of the several crude oils as being
levels of the factor "crude oil." The terminology sounds artificial in this context,
but its justification becomes more apparent when we think of a quantitative
factor such as temperature at levels of 100, 200, and 300 degrees.
The model that we have just written is an additive model. It may be
inadequate. It could happen that the first catalyst is clearly better than the
Sec. 1.6]
Factorial Experiments
11
others on all crude oils except the Sumatran, which might perhaps have a higher
sulfur content. This failure of the additive model, inasmuch as the expectation
of the difference between observations on different crude oils does not remain
constant from catalyst to catalyst, is an example of interaction between the
factors. Interaction is an added complication, and at this stage the experimenter
has to face up to a problem that he should have thought about earlier. From
what populations did the levels of the factors that were used come? If the
catalysts are some specific types having perhaps an alumina base, a silica base,
and so on, we have a fixed factor, meaning that the levels are fixed effects, and
the parameters a i are unknown constants. The objective would then be to
investigate comparisons between the parameters a i .
On the other hand, suppose that the barrels of crude oil were all drawn from
day-to-day production in the same California oil field. The effect b. would be a
random variable representing the day-to-day variation in the quality of the oil.
Then B would be a random factor, and the objective would usually be to obtain
an estimate of the component of variance V(b) = ob.
If both factors are fixed, we have a fixed effects model. If both are random, we
have a random effects model. If there is one of each, we have a mixed model.
The analysis, particularly in regard to the choice of denominators for the F
statistics, depends upon which of these three types of model is relevant. This is
discussed for the case of two factors in Chapter 4.
In Chapter 5 we turn to factorial experiments with several factors. Some of
the further development is a routine extension of the results of Chapter 4 to a
few more dimensions. There are, however, other complications and variations
that will be considered at that time. One of them merits mention now.
Let us return for a moment to the crude oils and the catalysts. Suppose that
there is indeed within each oilfield barrel-to-barrel variation in the quality of the
crude oil, and that we take three barrels from each of four fields and make one
run with each catalyst from each of the twelve barrels. This is an example of
what is called a split plot experiment. The catalysts see every barrel, and so
barrel differences cancel out in comparing catalysts. However, in comparing
crude oils each crude oil average contains the variability only of its own barrels,
and this does not cancel out.
The term split plot comes from the agricultural background at Rothamsted.
A modern example of split plots occurs with the use of crop-dusting aircraft.
One possibility is to take several fields; these will be the plots, or the whole plots.
We now divide the field into strips and plant each strip with a single variety of
barley; if there are enough strips we may have several strips to each variety.
These strips are the subplots. Suppose now that we also want to test some
fertilizers. We can then fertilize each complete field with the same fertilizer by
spraying it from an aircraft. Comparisons between fertilizers will be made
between whole fields. Comparisons between varieties of barley will be made
between strips in the same field. Conversely, one could apply the different
fertilizers by hand to the different strips in the fields, and then sow the seeds a
field at a time from an aircraft. A more complicated approach would be to sow
12 Introduction
[Ch. 1
strips running north-south from the air using different varieties for each strip in
each field, and then to spread different fertilizers over the field by having the
aircraft fertilize strips running in the east-west direction. There are numerous
possibilities.
1.7 Latin Square Designs

Chapter 6 is devoted to Latin square designs. A Latin square is a square of
side p in which p letters are written p times in such a way that each letter appears
exactly once in each row and each column. Latin square designs figure in two
contexts. They were first introduced by Fisher in agricultural experimentation as
designs for blocking in two directions (designs for the two-way elimination of
heterogeneity), with rows and columns representing the blocking and letters
standing for treatments.
Latin squares may also be used for fractional factorial designs. Suppose that
we have three factors each at p levels. To carry out a complete factorial would
need p x p x p = p 3 runs. If we use a Latin square we can let the rows represent
the levels of factor A, the columns the levels of factor B, and the letters the levels
of factor C. Then we shall have a design in which each level of each factor occurs
exactly once with each level of every other factor. This balancing enables us to
estimate the effects of all three factors (assuming that there is no interaction) with
only p 2 out of the p 3 runs, which represents a considerable saving. There are
occasions when such a design is forced upon us. This happens when time is one
of the factors. If we want to test p gasolines in p cars, and we cannot test more
than one gasoline in a car at a time, we must use a design like this. An interesting
variation occurs in nutritional experiments on dairy cows, where the different
rations are fed in sequence to the cows under test, and where there are residual
effects to be considered.
1.8 Fractional Factorials and Confounding

Chapters 7, 8, and 9 continue the investigation of the factorial experiments,
but now the emphasis is on several factors at two or three levels each, and the
principal area of application is moving away from agriculture to industry.
Again, however, the roots are in agriculture. It is to Yates (1937) that we turn
for the original work in Chapter 7, and the introduction of fractional factorials
by Finney (1945) dates from his days at Rothamsted. In Chapter 7 we introduce
2n and Y' designs; these are factorial designs for n factors each at two levels and
n factors each at three levels respectively, and because they are so often used
they merit special treatment. The main topic of the chapter is confounding.
If the number of points in a factorial design is so great that we cannot accommodate all the points in a single block, then we must run some of the points in
Fractional Factorials and Confounding 13
Sec. 1.8]
one block, and the remainder in one or more other blocks. One of the possible
contrasts between the data points will coincide with the differente between the
block totals. That degree of freedom is said to be confounded with blocks.
Clearly we should not like, for example, to include all the points at high temperature in one block and all the points at low temperature in the other block,
because then the temperature effect, if any, would be hopelessly confused or
confounded with the block effects. We should like to divide the design into
blocks in such a way that the division does as little damage as possible to the
design.
In Chapters 8 and 9 we turn to fractionation. We have already mentioned the
Latin square design as a 1/3 fraction of a factorial with three factors, each at
three levels. There are several reasons for the extensive use of fractional factorials. The relatively high cost of industrial experimentation makel it essential
that the designs used be as economical and as efficient as possible. Runs on a
refinery unit may cost several thousand dollars a time, and, even though the
eventual payoff may be many times more than that, management decision
makers are usually loath to spend more money on experimentation than they
feel is absolutely necessary.
The fact that it is possible to carry out experiments in relatively quick
succession means that it is feasible to consider in the initial stages of an experimental program a large number of factors in the hope of finding a few important
ones for further investigation. Such an initial experiment is called a screening
experiment, and it is customary to use as small a fraction as possible with hardly
any more degrees of freedom than there are factors. We shall consider such
designs for n 1 factors in n runs.
An example in which, on the other hand, the scientist hopes to find no
important factors is found in the food industry. Suppose that a company has
developed a new cake mix. When made in their laboratory ovens it is excellent.
Before marketing it, however, the company needs to be sure that it is robust
against the vagaries of the outside world, the uncertainties of the equipment in
many home kitchens, and the mistakes of some housewives. Will the new mix
produce an acceptable cake if the oven is ten degrees too hot or ten degrees too
cold, if it is left in too long or taken out a few minutes too Boon, if when the
cook adds the prescribed two eggs, she uses extra large eggs or the smaller
economy size? Here it is appropriate to run a screening experiment to check a
whole list of possible mistakes as factors with the hope than none, or at most only
a few, will turn out to be damaging to the cake.
Interest in fractional factorials has grown steadily during the last decade.
With factors at two levels, we shall present designs that allow us to estimate all
the main effects and two factor interactions if the higher order interactions are
ignored; smaller designs for estimating main effects but not the two factor
interactions; and even smaller designs for estimating all the main effects if all
the interactions are ignored. With factors at three or more levels, the number of
degrees of freedom taken up by the two factor interactions increases fast with n,
and most of the designs that are used are for the estimation of main effects only.
14 Introduction
[Ch. 1
1.9 Response Surfaces

Chapter 10 is industry's own chapter; the ideas were developed in industry
for industrial applications. The applications to agriculture are few and far
between. We shall be concerned with the basic problem of finding optimum
conditions for operating chemical processen. In its simplest form, we can
conceive of a chemical process as a black box with two knobs, one marked
temperature and the other pressure (with perhaps a third knob at the back for
noise). We envision a three-dimensional space with temperature plotted along
the x axis, pressure along the y axis, and the yield along the z axis, rising up
out of the paper. This response surface is like a mountain with a (hopefully)
single peak, and we have to turn the two knobs, i.e., carry out experiments
varying the levels of the two factors, until we find that peak. The original work
in this area was carried out by members of a group at Imperial Chemical
Industries, Great Britain, prominent among whom was G. E. P. Box. One of the
first papers on the topic is that of Box and Wilson (1951) on the experimental
attainment of optimum conditions. The work has been continued by Box and
his students since he left industry to enter the academic world. The chapter
concludes with a short section on evolutionary operation, which applies the
methods of the earlier part of the chapter to actual working plants in a refinery.
1.10 Matrix Representation

Throughout the book we shall make extensive use of matrices and quadratic
forms. This aspect of linear algebra is introduced in Chapter 2, together with a
discussion of the distribution of quadratic forms under the normality assumption. We shall conclude this chapter by sketching how the procedure that we
shall follow might be applied to the problem of comparing two means.
We shall rewrite the earlier model as
=m +t, +e11 ,
where yt; is thejth observation on the ith treatment (i = 1, 2; j = 1, 2, ..., n),

m is a grand mean, and t{ is the i th treatment effect. We arrange the observations
in a vector Y, the first n observations being those on the first treatment. Then
m
Y=X t l + e,
t2
where e is the vector of errors and X is a matrix with three columns and N = 2n
rows. The first column of X is 1 N , a column of N unit elements which corresponds
Matrix Representation 15
Sec. 1.10]
to the mean. The second column corresponds to t l and consists of n ones

followed by n zeros. The third column has n zeros followed by n ones. Since the
sum of the last two columns is equal to the first column, the rank of X is two.
We seek a vector (m, tl , t2 )' of estimates such that
Se
=
t
(Ys1 .Yti) 2
is a minimum where j = th + Z{ . Differentiating Se with respect to m, 11i 12 in

turn gives the three normal equations
1y, ; =Nm+ni,+n12 ,
n92 =
n9l=1Ylf=nm+ni,
X'Y = X'X(m, tt 1 , 1 2 ).
y2f = nm + n? 2i or
We cannot invert X'X because it is singular. Had we stayed with the earlier
formulation of the model, we would have avoided this crisis. However, we are
really concerned with (m + t1 ) (m + 2 ) = Zl 12 . Subtracting the third
normal equation from the second we have i l 12 = Yl 92, which comes as no
surprise; also, = y i .
We now consider four quadratic forms. One of them is the raw sum of squares
li Ir Y i = Y'Y = Y'IY. The second is N92 , where y is the mean of all N
observations. This form may be written as Y'J N Y/N, where J N denotes a square
matrix of order N with every element unity.
The identity
y4 r = NY 2 +
r
(j, 9) 2 = N9 2 + n(Yi Y2) 2 / 2 +

t
(Yj 9)2
{
may be written
Y'Y = Y'A 1 Y + Y'A 2 Y + Y'A 3 Y,
where
Al = N - i JN,
A 2 = (2n) - [
J" ij
Jn
I n - 'J,^
A3_
0
[
0
I, n Jj
These three matrices, A l , A 2 , and A 3 are all idempotent, i.e., A = A. We

shall show later that this implies that the quadratic forms, when divided by u 2 ,
have Xz distributions. Furthermore, A 1 A 2 = A 1 A 3 = A 2 A 3 = 0, and this will
be seen to imply that the forms are independently distributed. Hence, we may
test the hypothesis m l = m 2i or t, = t 2 , by the statistic
= (Y'A2Y)I(Y'A3Y/(N 2))
16 Introduction
[Ch. 1
(the ranks of the forms are equal to the numbers of degrees of freedom). Under
2) distribution, and the test is
the null hypothesis this statistic has the
the same as the usual t test.
Indeed, we have
F(1, N
n(91 Y2)
2
se _ n(Y1 Y2)2 t .
N-2

2s2
2
CHAPTER
Linear Models
and Quadratic
Forms
2.1 Introduction
In this chapter we derive some of the mathematica) results that will be used in
the subsequent chapters. The treatment makes considerable use of matrices and
quadratic forms; a short discussion of these topics appears in the appendix.
If we arrange the data in an experiment in a vector Y, the analysis of variance
procedure involves separating the sum of squares of the observations y 2 or
Y'Y into a set of quadratic forms
Y'Y=Y'A 1 Y+Y'A 2 Y+Y'A 3 Y+
If the observations Y are normally distributed, tests of hypotheses are made by
examining some of the ratios of these forms. The distribution of quadratic forms
is the topic of one of the later sections. This section is somewhat more demanding
mathematically than the rest of the chapter. Some readers may wish to omit it
provided they are prepared to accept our assurances that except where we say
otherwise:
(i) Y'A Y does, with the proper divisors, have a X 2 distribution with the
number of degrees of freedom equal to the rank of A,
(ii) The forms are independent, and, hence, (iii) The corresponding ratios
have F distributions.
{
We shall often use the abbreviation d.f. for degrees of freedom.

17
18 Linear Models and Quadratic Forms
[Ch. 2
2.2 Linear Models

We shall be concerned throughout this book with linear models in which an
observation y may be represented by
y
Nix1 +/32x2 ++Pyxy + e,
where x 1 , .. xp are coordinates, f3, ..., Sp are coefficients, and e denotes

random error. If there is more than one data point under consideration, we may
denote the ith data point by y { and write
.,
yt =
Nixil
N2xi2
++
NP tD +
ei,
with xij denoting the value of the jth coordinate at the ith data point. We shall
customarily use matrix notation and write
Y=Xf3+e,
where Y is a vector of N observations, X is a matrix of known constants, (3 is a
vector of p parameters, and e is the vector of random errors.
The matrix X = (x j ) is sometimes called the design matrix; it has N rows and
p columns. In the experimental design situations that will be considered, x, i will
take only the values zero or one. That restriction is not, however, necessary for
the results of this preliminary chapter. The elements of the parameter vector (3
will usually be unknown constants; there will be some occasions in later chapters
when we shall consider the special case in which the parameters f3; are random
variables. Our main purpose will be to obtain estimates of these parameters f3,
or some functions of the parameters, and to test some hypotheses about them.
The hypotheses will usually be that some subset of the parameters P 1 , ... , ,8 are
all zero.
The model is raid to be linear because it expresses y as a linear combination
of the parameters P; . There is no restriction that the model has to be a linear
function of the x j or that the Jatter be independent. We can put the polynomial
Po + flix + j2 x 2 + 93x 3 into this form by setting x l = x, x2 = x 2 , x 3 = x 3 .
Some models that are not linear can be made so by transformations. For
example, y = AeBx becomes z = Po + , 1 x if we take z = in y; then Po = ln A
and S l = B. On the other hand, the two models
Y = N1 + 1 f(N2 + N3x)
and
y = AeBX + CeD"
are not linear and cannot be made linear by transforming y.
Sec. 2.4]
The Method of Least Squares 19
2.3 Distribution Assumptions

In the original linear model, the ith observation consists of two components:
I j x, j,S1 and e ; , the "error" in the ith observation. The following assumptions
are made about the errors e i .
(i) They have zero expectations: E(e i ) = 0 for all i;
(ii) They are uncorrelated: E(ee,) = 0 if i # i';
(iii) They have homogeneous variance: E(e;) = u 2 for all i.
These assumptions may be written E(e) = 0, E(ee') = Ia 2 , and it follows from
assumption (i) that E(Y) = X13.
(In general, if Z is a vector of random variables, E(Z) is the vector whose ith
element is E(z i ). The variance-covariance matrix of the random variables
constituting Z will be denoted by cov (Z). If cov (Z) = Va 2 , the covariance of z{
and z; is v if c 2 ; we shall sometimes omit the scalar a 2 . If A is a matrix such that
AZ exists, we have E(AZ) = AE(Z), and cov (AZ) = AVA', where cov
(Z) = V.)
In addition to the three assumptions given before, we shall often make a
fourth assumption.
(iv) The normality assumption: that the errors e have a multivariate normal
distribution with E(e) = 0, cov (e) = Ia 2 , which will be written e
N(0, Io 2 ).
Equivalently, we may say that Y has a multivariate normal distribution
Y N(Xp, Ia 2 ). This assumption is important in carrying out tests of significance or in interval estimation, but it is not essential to the subdivision of sums of
squares or for obtaining least squares estimates. Indeed, the method of maximum
likelihood under the normality assumption gives the same estimates of estimable
functions as the method of least squares, which we shall consider in the next
section.
2.4 The Method of Least Squares

In estimating the unknown constants, P f , by the method of least squares, we
choose as our estimates a set, P 1 , P2 , ... , ^ , which minimize the sum of the
squares of the residuals. The vector of the estimates is written as (3, and we shall
see later that in some situations it is not unique. The residuals are defined in
the following way. Let f, denote the estimate of E(y 1 ) obtained by substituting
0 for the unknown (d, i.e., yt = 1 x,,ft ; then yl y t is the ith residual, and we
seek to minimize the sum of squares 1(y, y ; ) 2 In matrix notation, we write
[Ch. 2
the sum of the squares of the residuals as Se = (Y ')'(Y Y). To minimize

Se , we solve a set of p equations obtained by equating to zero the partial deriva-
tives of Se with respect to each of the estimates P; . Dropping the common

factor 2, thejth of these equations is
N
p N
XisxijP, = o.
X^jYi
s=li=1
{=1
These are called the normal equations. In matrix form they may be written
X'Y = X'X.
The matrix X'X is a symmetric matrix with p rows and p columns. Its rank
is the same as the rank of X, which is the number of linearly independent
columns of X.
2.5 The Case of Full Rank

This is the usual case that occurs in multiple regression problems. It will,
however, be the exception in the experimental designs that will be considered
in later chapters. We shall examine it briefly as an introduction to the general
case. If the p columns of X are linearly independent, r(X) = r(X'X) = p. Thus,
(X'X) -1 exists, and the normal equations have a unique set of solutions, or
solution vector, = (X'X) -1 X'Y.
The vector A is an unbiased estimator of (3, for we have
E(0) = (X'X) X'E(Y) = (X'X) 'X'Xa = P.

-1
Furthermore,
coV ( f") = (X'X)-1X'Ia2X(X'X)-1 = (X'X) 1 a 2 .
-
In computing the sum of squares of residuals

Se =Y'YY'YY'Y+Y'Y,
we substitute ' = XA and note that

Y' = Y'Y = 'Y = Y'X(X'X) -1 X'Y,
whence,
Se = Y'(I X(X'X) - 'X')Y.
We denote the quadratic form Y'X(X'X) 'X'Y by SR and call it the sum of
squares for regression. We have a subdivision of the total sum of squares into
two components, Y'Y = S R + Se.
-
Sec. 2.6]
The Constant Term and Fitting Planes 21
We shall see later that the expectations of these two component quadratic
forms are E(SR ) = (3'X'Xp + pa 2 , E(Se ) = ( N p)a 2 . Since X'X is positive
definite, p'X'X(3 is nonnegative and is zero only when (3 = 0. Under normality,
SR /a 2 has a noncentral x 2 distribution with p d.f., and SQ /a 2 has a centra! X2
distribution with (N p) d.f.; furthermore, the two forms are distributed
independently. Thus, an appropriate test for the hypothesis H0 : (3 = 0, with
alternative HA : (3 s4 0, makes use of the test statistic F = p~'SR /(N
Under the null hypothesis, F has Snedecor's F distribution with p and (N p)
d.f. The critical region is the upper tail of the F distribution. This is a consequence of the following results about the distribution of quadratic forms in
normal variables. Let Y N(, I). Then Y'AY X 2 (m) if, and only if, A is an
idempotent matrix of rank m; Y'AY and Y'BY are independent if, and only if,
AB = 0.
2.6 The Constant Term and Fitting Planes

The linear model presented in section 2.2 does not appear to contain a
constant term. However, such a term is easily included within the scope of the
model. We add a coordinate x 0 , which always takes the value one, and write
E(y)
= P O + N1x1 +. ' ' + Nkxk = N0x0 + N1x1 + ' . + Nkxk-
We shall not attempt a comprehensive discussion of fitting planes and hyperplanes to data. The reader who wishes to look further at the practical aspects is
referred to the book by Draper and Smith (1966). It has become the custom to
refer to the variables x 1 , ..., Xk as independent variables and the observed
response y as the dependent variable. The use of the word independent should not
be taken seriously. It derives from the habit in beginning algebra courses of
labeling y as the dependent variable and x as the independent variable in the
formula y = x + 3, because one can choose any value of x, substitute in
the formula, and obtain the value of y. Nothing, however, is implied about the
independence of the x variables. Indeed, as we mentioned before, we can fit a
polynomial in x by letting x = x,.
In some situations we have no control over the elements of the design matrix
X. When we can choose the points at which observations are to be taken, we
note that cov () = (X'X) -1 a 2 , and choose the experimental points with this in
mind. One possibility could be to choose X in such a way as to minimize the
average variance of the Pi . This is equivalent to minimizing the trace of (X'X) -1 .
This procedure of fitting hyperplanes is commonly used to obtain equations
for predicting future values of the response given x 1 , ... , X. If f3 = 0, we might
conclude that x ; is of little use as a predictor variable, although one should be
careful about jumping to that conclusion when x i is highly correlated with other
independent variables. One might ask the following question: are the predictor
variables x 1i x 2 , ... , X k collectively useless as predictors, or, alternatively, does
[Ch.2
adding the set of x variables give a significant improvement over the simpler
prediction equation E(y) = P o ? More formally, we wish to test the hypothesis
Ho : p* = 0 where (3* = (P1, P2. Pi), against the alternative HA : F3* # 0.
We have two nested models:
L: E(y) = po i
Qi : E(y)
= Po +
fl,x1.
Under S? o Po = y so that S R = N9 2 = N 'Y'JY, and the sum of squares of

the residuals is 1 yt Ny 2 = 1 (y i y) 2 . The matrix J is a square matrix
with every element unity. Under 0 1 we have, as before, SR = Y'X(X'X) -1 X'Y.
The increase in S R is attributable to the addition of x 1 , ..., xk to the model.
We have
,
Y'Y = N 'Y'JY + Y'(X(X'X) -1 X' N 'J)Y + Se = Y'A 1 Y + Y'A 2Y+Se .

-
Under normality, the three quadratic forms Y'A 1 Y, Y'A Z Y, Se have, upon
division by o 2 , independent X 2 distributions with 1, k, and N k 1 d.f.
respectively. The test statistic for Ho is
= Y'A 2Y/k = Se /(N k 1),
which has, under the null hypothesis, the F(k, N k 1) distribution. Some
writers prefer to call Y'A 2Y the sum of squares for regression. We shall not do
this, but shall continue instead to let SR = Y'Y S. The generaI question of
testing subhypotheses will be considered later.
2.7 The Singular Case

If X is of less than full rank because the columns are not linearly independent,
X'X is singular and (X'X) -1 does not exist. Then we must modify the results of
the previous sections. This state of affairs can occur either inadvertently or by
design. It will, in fact, be true of almost all the experimental design models in
the subsequent chapters.
As an example of accidental singularity, suppose that we have a model
E(y) = Po + R1x1 + R2x2,
where x l + x2 - 100. This would occur if x l denoted the percentage of olefins
in a hydrocarbon and x 2 the percentage of nonolefins. The matrix X is singular
because the first column is a linear combination of the last two columns.
Suppose that y = 90 + 3x 1 + 5x 2 is a predicting plane. Since x l + x2 - 100,
we may equally welf have y = 190 + 2x 1 + 4x2 or more generally,
9=90 100t +(3 +t)xl +( 5 +t)x2.
Pseudo-Inverses and the Normal Equations 23
Sec. 2.8]
Obviously, in this simple example the difficulty can easily be avoided by

dropping either x l or x 2 (it does not matter which) from the equation. The three
parameters P o , S l , and P, cannot be (uniquely) estimated. However, we note that
for all values of t we have
p1
P2=(3+t)
(5+t)= 2,
P0
+50(P 1 +fl2)= 490.
These two parametric functions are said to be estimable.

An example in which the model was purposely chosen in such a way that X
is singular appears at the end of the previous chapter in the experiment to
compare two treatments. The model was E(Y) = X13, where (i' = (m, t l , t2 ).
If y, j denotes the jth observation on the ith treatment, we have yij = m +
t, + ei; . The first column of X consists entirely of ones. The elements of the
other columns are either zero or one, and the first column is the sum of the other
two; we have r(X) = r(X'X) = 2.
2.8 Pseudo-Inverses and the Normal Equations

We consider now the solution of the normal equations X'Y = X'X in the
general case where X has N rows and p columns and r(X) = k < p and
m = p k.
The normal equations are consistent. The augmented matrix (X'X, X'Y) may
be written as X'(X, Y). Then r(X'X, X'Y) < r(X') = r(X'X). However,
r(X'X, X'Y) >_ r(X'X) since all the columns of X'X are contained in the larger
matrix, and so r(X'X, X'Y) = r(X'X). Let p = PX'Y be a set of solutions to
the normal equations. We call a solution vector and P a solution matrix, or a
pseudo-inverse of X'X. P is not necessarily a symmetric matrix. If k = p, there
is only one solution matrix, namely P = (X'X) -1 . The use of pseudo-inverses
in the solution of the normal equations has been discussed by Rao (1962).
Substituting for (3 in the normal equations, we have X'Y = X'XPX'Y. This
is an identity for all Y, and so
X' = X'XPX'.
(2.1)
Multiplying both sides of this equation on the right by XP gives
X'XP = X'XPX'XP,
so that X'XP is idempotent. It may similarly be shown that XPX' and PX'X
are also idempotent. Furthermore, multiplying on the right by X gives
X'X(PX'X I) = 0, so that a general solution to the equations X'X (i = 0 is
(PX'X I)Z, where Z is an arbitrary vector. A general solution to the normal
equations is thus
= PX'Y + (PX'X I)Z.
[Ch. 2
If P = (X'X) -1 , the extra term vanishes; if k < p, it would appear that we might
find ourselves in the awkward position of having numerous solutions. If we seek
to estimate a linear function = a'(3 by substituting solution vectors (3 for [i,
we obtain a set of estimates
^ = X'PX'Y + X'(PX'X I)Z. (2.2)
We now define an estimable function [Bose (1944)]. A linear combination
= X'[3 of the parameters is said to be estimable if, and only if, there exists a
linear combination c'Y of the observations such that E(c'Y) = a'(3, i.e., such
that c'X(3 = ?'f3 for all vectors P. The condition is that a vector c exists such
that X'c = ).. A linear function 0 is estimable if, and only if, ais a linear combination of the rows of X. It follows that the set of estimable functions forms a
vector space. The dimension of the vector space, i.e., the number of linearly
independent estimable functions, is equal to the rank of X.
An equivalent necessary and sufficient condition for a linear combination of
the parameters to be estimable is that there exists a solution vector to the
system of equations X'Xr = a. If such a vector r exists, then we let Xr = c.
Conversely, if there is a vector c such that X'c = a, c is a solution vector to that
set of equations, and hence r(X') = r(X', l). It follows that r(X'X) and
r(X'X, X) are also equal, and so the set of equations X'Xr = a has a solution
vector r.
Multiplying Equation 2.1 on the left by r' and on the right by X gives
r'X'X = r'X'XPX'X or a' = X'PX'X. Hence the term X'(PX'X I)Z in
Equation 2.2 vanishes, and that equation becomes = X'PX'Y.
Let P* be any other solution matrix. From Equation 2.1 we have X'XPX' _
X'XP*X'. Multiplying on the left by r' gives a'PX' = X'P*X', whence _
a'PX'Y = a'P*X'Y.
Thus, we have shown that the estimates of estimable functions are unique.
They are the same, no matter which solution vector (i to the normai equations we
choose. Henceforth, we can denote ?'PX'Y by and call it the least squares
estimate of a'(3.
The estimate is unbiased; we have
E(^) = E(X'PX'Y) = X'PX'Xp = X'(3.
The variance of is given by
V() = V(a'PX'Y) = X'PX'XP'aa 2 = a'Paa 2 .
This means that we can, in calculating the variances of least squares estimates of
estimable functions, act as if Pa 2 is the variance-covariance matrix of 0. In
particular, for any solution matrix P we have
V(N1 Ni) = (Ptt + pij pui Pi1 )a2
Sec. 2.8]
Pseudo-Inverses and the Normal Equations 25
EXAMPLE. Consider the experiment mentioned earlier in which six

observations are made, three on each of two treatments. The usual model is
y;f = m + t, + eij where i = 1, 2 and j = 1, 2, 3. We have three parameters:
(3' = (m, t l , t2 ), and the transpose of the design matrix is
X'=
0 0 0;
0 0 0 1
X'Y = (G, T1 , T2 )' where Ti = y and G = I t I j yij . The rank of X'X is 2,

and the normai equations are
6 3 3
X'Y = 3 3 0 .
3 0 3
The usual method of solving the equations is to impose the side condition
t l + t 2 = 0 upon the parameters. This makes the first equation 6m = G,
whence 61 1 = 2T1 G and 612 = 2T2 G. The corresponding solution
matrix P l is unsymmetrical.
1 0 0
6P 1 = --1 2 0
1 0 2
If zG = t l t, we have X' = (0, 1, 1) and, if 3r' = (0, 1, 1), X'Xr = a.

Then 0 is an estimable function and
^ = (0, 1,
1)P1X'Y = (T1 T2)/ 3 ,
V(^) = a'P 1 a7 2 = 2a 2 /3.
A second solution matrix with t l + t2 = 0 is given by
1
12P 2 = 1
1 3
1 3
for which m = (G + Tl + T2 )/12 and 1 1 = (G + Tl 3T2 )/12. Again we have

+^ = (Ti T2 )/3 and V(^) = (1 + 3 + 3 + 1)0 2 /12 = 2a2/3.
[Ch. 2
Two other solution matrices having t 2 = 0 and t l = (T1 T2 )/3 are given by
00
3P 3 = 0 1 1 ,
0 0
3P 4 = 0
1 1
1 1
These give the same results as before for and V(^).

In this example the estimable functions are 0 1 = m + t 1 , 1s = m + t 2 , and
any linear combination of them, including in particular 0 3 = t l t 2 .
2.9 Solving the Normal Equations

In the example given in the previous section, we solved the normal equations
by imposing a side condition upon the parameters. The side condition t l + t 2 = 0,
or Zl + t2 = 0, added an extra equation to the normal equations, which then
gave us a set of p + 1 equations of which p were linearly independent. Any
nonestimable linear function of the parameters can be used as a side conditlon.
For example, m = 0 or t l = 0 or t 2 = 0 would each sufface. So, for that matter,
would a bizarre linear combination such as 72m + 13t 1 + 97t 2 = 0; the
equations would not be convenient to solve and the estimates of m, t 1 , and t2
would look a little peculiar, but it should be emphasized that the estimates ^
of the estimable functions 0 would be the same no matter which of the side
conditions was imposed. The usual procedure. is to choose a side condition
which enables us to obtain a solution matrix P and a solution vector (3 as
conveniently as possible.
We present now two standard methods of obtaining solution matrices in the
case where r(X) = k = p m and m > 0, with a set of m side conditions. For
those experiments in which the normal equations have a simple form, the solution matrix is easily obtained directly; the reader may, therefore, wish to
postpone this section until he reaches Chapter 11.
Let D be a matrix ofp rows and m columns such that r(D) = m and XD = 0.
Suppose allo that we add m further rows to X in the form of a matrix H of m
rows and p columns, such that the rank of (X', H') is p. Then HD is nonsingular.
If the normal equations are solved, subject to the side conditions H13 = 0, the
solutions ft are unique numerically. However, since D'X'Y = 0, the solutions
may differ in form by some linear combination of the rows of D'X'Y.
In the previous example, D' = (1, 1, 1) and D'X'Y = G Tl T2 - 0.
The solutions are obtained with H = (0, 1, 1). The solution matrix P l gives
= G/6; P 2 gives = (G + Tl + T2 )/12 = (G/6) (D'X'Y/12).
Let P and P* be two solution matrices under the side conditions Hp = 0.
Then PX'Y = P*X'Y. This is an identity in Y, and so (P P*)X' = 0. But
r(X) = (p m), and r(D) = m. Thus, if W is a matrix such that XW = 0,
there exists a matrix E such that W = DE', and so P P* = ED'. If P and P*
Solving the Normal Equations 27
Sec. 2.9]
are both symmetric, then P* = P + DCD', where C is a symmetric matrix of

order m. The two standard methods of solving the normal equations both give
symmetric solution matrices.
In the first method [Kempthorne (1952) and Graybill (1961)], the matrix
X'X is augmented to form the new matrix
A* = [XIX H'1
H 0
Then r(A*) = p + m, and A* -1 exists. Let
A* _1 = B11 B12
B21 B22]
The identity A*A* -1 = I gives four equations:

X'XB 11 + H'B 21 = I,
(2.3)
X'XB 12 + H'B 22 = 0,
(2.4)
HB 11 = 0,
(2.5)
HB 12 = I.
(2.6)
Multiplying Equations 2.3 and 2.4 on the left by D' and recalling that D'X' = 0
and that D'H' is not singular, we have D'H'B 21 = D', whence B12 = B'21 =
D(HD) -1 and D'H'B 22 = 0, whence B 22 = 0.
Multiplying Equation 2.3 on the left by B 11 gives B 11 X'XB 11 = B, and so
X'XB 11 is idempotent.
X'XB 11 = I H'B21 = I H'(D'H')
-1
D',
X'XB 11 X'Y = X'Y H'(D'H') -1 D'X'Y = X'Y.

Thus, (3 = B 11 X'Y is a solution to the normal equations, and cov () _
B 11 X'XB 11 a 2 = B 1l a 2 .
In the second method [Scheff (1959) and Plackett (1960)], the solution
matrix is P* = (X'X + H'H) -1 , with 0* = P*X'Y as the corresponding
solution vector (see particularly Scheff, p. 17). Numerically, P*X'Y = B 11 X'Y,
but it does not follow that (X'X + H'H) -1 = Bl1.
cov (*) = (X'X + H'H) -1 X'X(X'X + H'H) - lo2 = cov (^) = B ll v2 .
But (X'X + H'H)(I D(HD) -1 H) = X'X, and so
(X'X + H'H) -1 X'X(X'X + H'H) -1 = (I D(HD) -1 H)(X'X + H'H) -1 = B11.
[Ch. 2
This gives us a method of obtaining B 11 from (X'X + H'H) -1 . We cannot,

however, obtain (X'X + H'H) -1 from B 11 with this result because
(1 D(HD) -1 H)
is idempotent, and thus is not of full rank.
We now obtain a different relationship between the two solution matrices.
We write (X'X + H'H) -1 = B 11 + DCD'. Then
I = (X'X + H'H)(B ll + DCD') = X'XB ll + H'HDCD'
= I H'(D'H') - 'D' + H'HDCD',
and so H'(D'H') -1 D' = H'HDCD'. Multiplying both lides on the left by
(D'H') - 'D' and on the right by H'(D'H') -1 reduces this equation [John (1964)a]
to
(D'H')-1 = HDC,
whence
C = (HD) - '(D'H') - ' = (D'H'HD) -1 ,
so that
(X'X + H'H) -' = B 11 + D(D'H'HD) - 'D'.
2.10 A Theorem in Linear Algebra

The following theorem will be used later in proving Cochran's theorem.
Alternative proofs of the theorem and other relevant results are given by James
(1952), Lancaster (1954), Graybill and Marsaglia (1957), and Banerjee (1964).
Theorem 1. Let A l , A k be symmetrie matrices of order n such that
I = î= 1 . Each of the following conditions is necessary and sufficient for the
other two:
(i) All the Ai are idempotent;
(ii) A,A ; = Ofor all pairs i, j, i s4 j;
(iii) Ii n, = n where n i = r(A i ).
PROOF .
There is an orthogonal matrix P such that

k
P'IP = I = P'A 1 P + P'A JP,

t=z
where
P'A'P = (0 0)
and
P'A,P = (B; Cl )
i
The Distribution of Quadratic Forms 29
Sec. 2.111
where A is the diagonal matrix whose diagonal elements are the n l nonzero
latent roots of A l , and B ; is a square matrix of order n l . Throughout this proof,
j takes the values 2 < j 5 k.
Suppose all the A. are idempotent. To establish condition (ii), it is enough to
show that A,A ; = 0 for all j. We have A = I and 1, tr (Bi ) = 0; but P'AfP
is idempotent, and so the diagonal elements of B ; cannot be negative. Hence
B1 = 0 and C, = 0 for each j. It follows that A 1 A, = PP'A 1 PP'A1P = 0.
Since P'A;P is idempotent, E; is idempotent. Then n = tr (In ) = n l +
1 tr (E; ) = n l + 1 n f , and condition (iii) is established.
We now show that conditions (ii) and (iii) each imply condition (i). Suppose
A,A 1 = 0 for all pairs i, j where i j. Then for each i,
A;= AtiI= A{(A1+ ..+Ai+...+A,j=A,.
.
If condition (iii) holds, we can show that A l is idempotent. We write B, C, E, Q

for 1 Bj , 1 C j , E5 , 1 A; . Then, since P'A 1 P + P'QP = I, we have E = In_n1,
C = 0. But r(P'QP) = r(Q) = n n l = r(E), and so B = 0 and A = I, so
that A l is idempotent.
2.11 The Distribution of Quadratic Forms

If Y is a vector of N random variables, the quadratic form Y'AY is also a
random variable. Most of the results in this section about the distribution of
quadratic forms hold only when Y has the multivariate normal distribution.
The following lemma is true even when the yl are not normally distributed.
Lemma. Let E(Y) = p and cov (Y) = Io 2 . Then E(Y'AY) = 'A + a 2 tr (A).
PROOF.
Y'AY =
a, t yi + 2
t
E(y) = i
+ a 2 ,
at,yty>>
<1
E(ytyf) = tlif
Hence,
E(Y'AY) _ a41 + at1Q 2 + 2 a115

t
= &' AN.
+ a2 tr (A).
t<1
The next result concerns the normal distribution.
Theorem 1. Let Y N(, I). The moment generating function of Y'AY is

My.Ay(t) = II 2AtJ -1/2 exp { II' '( I 2At) -1g)}.
[Ch. 2

PROOF .
Suppose X
N(U, V). The probability density function of X is
.1x (X) = (2 1r)N12 [V 1 112 exp{ 1(X
U)' V
(X
U)},
and the corresponding multiple integral over the range 00 <X < +00 is
unity. The moment-generating function of Y'AY is
E[e<Y"A Y ] = (2^.) - Nj 2
exp{ [(Y )'(Y ) 2tY'AY]/2} dY.
Writing the exponent as Q/2, we have

Q=Y'(I-2tA)YY' .'Y+'...
Completing the square and writing U for (I 2At) - 'g,

Q=(YU)'(I-2At)(YU)+'(I(I-2At) -1 ),
so that
M(t) = (21T) - N 12 f 00 exp{I(Y U)'(I 2At)(Y U)}

dY exp['(I (1 2At) ')K/2]
-
= Î 2At^ 112 exp(+[i'p. (I 2At) fit]}

-
If A = I, we have for the moment-generating function of Y'Y

My , y (t) = (1 2t) - N 12 exp[,1 + a(1 2t) -1 ].
This is the moment-generating function of the noncentral X 2 distribution with

N d.f. and noncentrality parameter A = '/2. If A = I and = 0, M(t) =
(1 2t) - N 12 , which is the moment-generating function of the centra! X2
distribution with N d.f.
If cov (Y) = V, we may prove in similar fashion that
M(t) = II 2tAVI -1 / 2 exp{I['V - l 'V -1 (I 21VA) - l]}.
The next theorem points out the importante of idempotent matrices in the
theory of the distribution of quadratic forms.
Theorem 2. Let Y N(, I). A necessary and sufficient condition that Y'AY have
a X 2 distribution with m < N d.f. is that A be an idempotent matrix of rank m.
Sec. 2.11]
PxooF. We give a proof for the case = 0 when the distribution is central
X2.
The proof consists of equating the moment-generating function of Y'AY with

that of a X 2 distribution. The proof in the noncentral case is similar:
MY.nx(t) = ]I 2Atj -112.

A is a symmetrie matrix; hence, there exists an orthogonal matrix P such that
P'AP = A. Then
I 2At I = IP'I II 2At HPI = II 2At I.
Sufficiency: Let A be idempotent of rank m. Then A is a diagonal matrix with m
of the diagonal elements unity and the remainder zero; I 2At is a diagonal
matrix with m diagonal elements (1 2t) and the others unity. Hence,
]I 2At = (1 2t)m and M(t) = (1 2t) m/ 2 , which is the momentgenerating function of a x 2 distribution with m d.f.
-
Necessity: In general, I 2At = 1I(1 21,t). Suppose that Y'AY has a Xa

distribution with m d.f. Then we have (1 2t)m = 1I(1 2),t). This is an
identity for all t, and it follows that the latent roots of A are A = 0, 1 with
multiplicities N m and m, so that A is an idempotent matrix of rank m.
Two extensions of this theorem are of interest: we state them without proof as
corollaries.
Corollary 1. 1f Y N(, Ia 2 ), the quadratic form Y'AY/a 2 has a noncentral X 2
distribution with m d.f. and noncentrality parameter A = 'A/2a 2 if, and only if,
A is idempotent of rank m.
Corollary 2. 1f Y N(, V), the form Y'AY has a noncentral x 2 distribution
with m d.f. and noncentrality parameter A = 'A/2 if, and only if, AV is an
idempotent matrix of rank m.
The following theorem is concerned with the independence of quadratic forms
in normal variables. There are at least five proofs in the literature, mostly for the
case = 0 [Craig (1943), Hotelling (1944b), Ogawa (1949), Aitken (1950), and
Lancaster (1954)]. The validity of the first two proofs has been questioned.
Hotelling questioned Craig's proof, and his proof in turn was questioned by
Ogawa. The proof that we shall give is a modification by Lancaster of Aitken's
proof.
Theorem 3. Let Y N(1r., I). Two quadratic forms Y'AY, Y'BY are independent
ifandonlyifAB = 0.
PROOF. [Lancaster (1954)]. This proof is for the case = 0. We have shown
that the moment-generating functions of Y'AY and Y'BY are 11 2AsI -112 and
[Ch. 2
11 2Bt1 -112 . By a similar argument, the joint moment-generating function of

the two forms is seen to be 1I 2As 2Bt 1 -1 " 2 . We have to show that the
joint function is the product of the two individual functions, i.e.,
Î 2As j]I-2BtI=Î 2As 2BtI,
(2.3)
for all values of s and t in an interval around the origin if, and only if, AB = 0.
The sufficiency part is trivial since, if AB = 0, we have
Î-2AsJ I-2BtI = Î 2As 2Bt+4ABst = Î-2As-2BtI.
Conversely, suppose Equation 2.3 is true. Let P be an orthogonal matrix such
that
P'AP =
`0
(A 0
0)
P'BP =

C^^ C^2
C21
C22
where A is diagonal and A and C 11 are square matrices of order r(A). It follows
from Equation 2.3 that for any ,1 0 we may replace s by s/A and t by t/)) to
obtain
IJ-2AsH Ju-2BtI = a"IM-2As-2BtI
or
^1I 2P'APs
2P'BPtI
2P'APs 2P'BPtI.
Equating coefficients of st( ) on each side of the equation gives

Al 2 P'BPt 1= PA) 1 M- 2 C 22 t 1.
Thus P'BP and C 22 have the same nonzero latent roots. This implies that the
sum of the squares of elements of P'BP is equal to the sum of the squares of the
elements of C 22 , each being the sum of the squares of the latent roots. Since all
the elements of B are real, this implies that C l Cr2, and C 21 are all zero matrices
so that P'APP'BP = 0 and AB = 0.
In the case g 0, we must, in addition, take into account the exponents in the
moment-generating functions. Thus we must show that
..'{I (I 2As) -1} + N.'{I (I 2Bt) - '}p. = 1.'{I (I 2As 2Bt) -1 }g
for all s and t in some interval containing s = t = 0, i.e.,
I(I-2As) -1 +I(I-2Bt) -1 =I(I-2As-2Bt) -1 ,
if, and only if, AB = 0.
Sec. 2.11 ]
For sufficiently small s we have
I (I 2As) - = I (I + 2As + (2 As) 2
+ ... + ( 2As)"
+ ... ),
and the required condition becomes

2As +(2As) 2
+ 2Bt +(2Bt) 2
= (2As + 2Bt) + (2As + 2Bt) 2 +...

Equating coefficients, this is true if, and only if, AmBnsmtn = 0 for all m, n, and
for all s, t, which in turn is true if, and only if, AB = 0.
3
In the analysis of variance, Y'Y is subdivided into several quadratic forms
(sums of squares). The next theorem is concerned with the joint distribution of
these sums of squares. It is due to Cochran (1934) and is essentially an algebraic
result expressed in a statistical context. He originally proved it in the centra!
case. The noncentral case is due to Madow (1940).
Theorem 4. Let Y N(p., 1) and let Y'Y = Y'A 1 Y + + Y'A,,Y where Y

has n observations and r(A 1 ) = n i . A necessary and sufficient condition for the
forms Y'A,Y to be distributed independently with Y'A,Y X 2 (ni ) is that 1 n, = n.
PROOF. We have already shown that n { = n if, and only if, each A, is idempotent and if, and only if, AAf = 0 where i s4 j. But A i idempotent is equivalent to
Y'A,Y
2 (n i ), and AA, = 0 is equivalent to Y'A,Y and Y'A 1 Y independent.
We may rephrase the statement of the theorem in the following way. Let
Y N(p., I), and let Y'Y = 1 Y'A,Y where r(A i ) = n i . Then any one of the
following five conditions implies the other four:
_X
(i) n i = n;
(ii) Each A t is idempotent;
(iii) A,A1 = 0 for all pairs i, j, where i 0 j;
(iv) Y'A 1Y x 2 (ni) for each i;
(v) Y'A {Y, Y'A 5Y are independent for all i, j where i # j.
Each of the forms Y'A,Y may be written as Y'Bi'B1Y where Bt is a matrix of n,
rows and n columns such that B,B! = I74 i.e., such that the columns of Bi form
an orthonormal set of ni vectors, It follows from the fact that A tA; = 0 that
BBI' = 0 also, for we have
A1A5 = BiB,BJBJ = 0,
and, multiplying on the left by B i and on the right by Bi,
IB,B f'I = B{B! = 0.

Let B be the matrix defined by B' = (B'1 , Bi,. . . ). Then B is an orthogonal
matrix, and the transformation Z = BY is an orthogonal transformation taking
Y'Y into Z'Z.
[Ch.2
2.12 The GaussMarkoff Theorem

We have shown that the method of least squares provides unique unbiased
estimates of the estimable functions. Apart from stating that under normality
the least-squares estimates are the same as the maximum likelihood estimates,
we have made no claim that these estimates are in any other way particularly
desirable. The GaussMarkoff theorem shows that, in a certain sense, the leastsquares estimates are optimal and justifies our choice of this procedure.
Definition. An estimate of a function of the parameters is called a linear estimate
if it is a linear combination of the observations, i.e., a'Y where a is a vector.
Theorem. (GaussMarkoff). Let Y be distributed with E(Y) = X(3 and cov (Y) _
Ia2 ; let 0 = a'(3 be an estimable function. Then, in the class of all unbiased linear
estimates of b, the least squares estimate = ai has the minimum variance and is
unique.
PROOF .
Suppose that 0 = PX'Y where P is any solution matrix. Let 0* be

an unbiased estimate of 0, other than ; * = a'Y. Let a' = )'PX' 4- b'.
Then
E(O*) = E[(a'PX' + b')Y] = :G + b'Xp.
Since :G* is unbiased, E(*) = 0 and b'Xp = 0. for all parameters [i so that we
have b'X = 0.
Then
V(O*) = (a'PX' + b')(a'PX' + b')'a 2
= (a'PX'XPa + b'XP'a + )^'PX'b + b'b)a 2
= V(^) + b'bo2 = V(^) + bi a 2 .
Thus V(/*) >_ V(+) with equality only if b = 0. We have already shown that
is unique. Hence, is the unique unbiased estimate of 0 with the minimum
variance.
u
We now show that the least squares procedure produces an unbiased estimate
of a2 . The expected values of the observations are estimable functions with
estimates Y = Xî, and we have minimized the sum of squares for error Se =
(Y Y). The subdivision of Y'Y into rums of squares, Y'Y = SR + Se ,
was sketched briefly at the beginring of this chapter for the case in which
r(X'X) = p. In the general case, recalling that X' = X'XPX', we have
yl
' = î'X'XO = Y'XP'X'XPX'Y = Y'XPX'Y = Y'Y = Y'Y,
The GaussMarkoff Theorem 35
Sec. 2.12]
and
Se =Y'YY"'i'Y+' 't = Y'(I XPX')Y,
SR = Y'XPX'Y = ^3'X'Y.
Since E(Y) = X(d and cov (Y) = Ia',
E(Y'Y) = (3'X'1X13 + a 2 tr (I) = f3'X'X(3 + Na 2 ,
and
E(SR) = E(Y'XPX'Y) = E3'X'XPX'XR + a 2 tr (XPX').
XPX' is idempotent, and so tr (XPX') = r (XPX'). Now r (XPX') < r (X) = k;
but X' = X'XPX', and so r(X) = r(X') < r(XPX'). Hence, r(XPX') _
r(X) = k, and we have
E(SR) = f'X'X13 + ka 2 ,
and, by subtraction,
E(Se) = (N k)a 2 .
Thus, we see that the sum of squares for error provides an unbiased estimate
S 2 = Se /(N k) of the error variance a 2 .
The matrices XPX' and (I XPX') of the quadratic forms SR , SQ are both
idempotent. Under normality, the conditions of Cochran's theorem are satisfied
if we replace Y by o - 1 Y, and so SR /a2 and Se /o 2 have independent X 2 distributions with k and (N k) d.f. respectively.
2.13 Testing Hypotheses

Throughout this section we shall assume that Y has a multivariate normal
distribution. The distribution of SR /aa is noncentral x 2 with noncentrality
parameter f'X'X f /(2o 2 ). Thus, as we have already mentioned, an appropriate
test statietic for the nul! hypothesis H o : (i = 0 against the general alternative
p # 0 (i.e., at least one of the elements of p is not zero) is = k 'SR /
(N k) 'Se , where k = r(X). Under the nuli hypothesis, . has Snedecor's F
distribution with k and (N k) d.f. The critical region is the upper tail of
F(k, N k). We shall not discuss here the optimum properties of the F test
in this situation. The reader who wishes to pursue the matter further is referred to
Scheff (1959), especially Chapter 2.
When a sum of squares is divided by the number of degrees of freedom in the
corresponding x 2 distribution, the quotient is called a mean square. Thus, S 2 is the
mean square for error, and SR /k is the mean square for regression.
-
[Ch. 2
It may be argued that in some practical situations the hypothesis t = 0 is not

especially interesting, and the experimenter might be more interested in testing
either the hypothesis that some particular estimable function is zero or the
hypothesis that all the parameters in some subset of P are zero.
Consider first a particular estimable function 0 = ?'[3, which is estimated by
= a = >.'PX'Y. Then 2 = Y'XP'XX'PX'Y, and 2 /(a'Pao2) has a X2
distribution with one d.f. Furthermore, X'(I XPX') = 0, and so 2 /(a'Paa 2)
and Se /a 2 have independent X 2 distributions. Thus, the hypothesis H 0 : 0 = 0
may be tested by the F test using the test statistic F = 2 /(X'PXs 2 ), or, equivalently, by the t test with t 2 = F. Confidence intervals for 0 can be constructed
using Student's t in the usual way. For example, a two-sided interval could be
obtained as
ti t*()'Ps2)-1I2 ^ ,/,
+ t*('P%ls2)-1I2
More will be said on this topic later.

We next consider testing hypotheses of the type Ho : P l = 0 where p 1 is a
subvector of P. Suppose that (3 is divided into two subvectors, P = 1, 0 2), and
that X is similarly partitioned into (X 1 , X 2 ). Let r(X i ) = k l , r(X 2 ) = k2 . The
complete model for which we have just estimated the parameters is
Y = (X1, X2)(P1, f3 2) + e.
Under the hypothesis Ho : (3 1 = 0, we have the reduced model

Y=X202+e.
We write S R (P 2 ), SR (P) for the sums of squares for regression under the
reduced and complete models, respectively, and SR(P1IP2) for the difference
SR (P) SR (P 2 ). Let M be a solution matrix for the reduced normal equations
X2Y = X' X 2 (3 2 i X 2 MX2 is idempotent and SR (P 2 ) = Y'X 2 MX2Y.
We now have a subdivision of Y'Y into three quadratic forms:
Y Y = SR\N2) + SR(P1I P2) + Se,
'
where SR(P11 r2) = Y'(XPX' X 2MX2)Y.

Now X' = X'XPX', and so X2 = X' XPX', and (XPX' X 2 MX2) is idempotent. Thus, SR(P2)/a 2 , SR(P1 1 3f 2)/a 2 , Se /a 2 have independent x 2 distributions
with k 2 , k k 2 and N k d.f. The hypothesis H 0 : [i l = 0 is tested by comparing the sum of squares for error under the reduced model, S e + SR(al I
to the sum of squares for error, Se , under the complete model, ad asking ourselves whether the reduction in error, S R([3 1 1(3 2), made by including P l in the
model is large enough to be regarded as significant. The test statistic F =
(k k2) -1 SR(P11P2)/s 2 has the F[k k 2 , N k] distribution under the null
hypothesis.
Missing Observations 37
Sec. 2.14]
2.14 Missing Observations

It occasionally happens that in an experiment one or more of the observations
is spoiled or omitted. One might have planned a setup in which there were
several cars and several gasolines, each gasoline to be used in each car (a nicely
balanced experiment) only to find that something went wrong. Perhaps one of
the cars broke down before being used to test all the gasolines and could not be
repaired in time; perhaps, because of a failure in communications, not enough
of one of the fuels was prepared and the supply ran out before it could be tested
in every car; perhaps one of the observations is so clearly incompatible with the
others that it should be rejected as containing gross error. What can be done?
If possible, the experimenter should try to take another observation duplicating
the experimental conditions prescribed for the missing observation. If this
cannot be done, we ask the following question: can we obtain a convenient
formula or recipe for generating from the actual data an estimate for what the
missing value should have been? Because this question first arose in agricultural
experimentation [Yates (1933)] where the experimental units are called plots,
these formulae are called missing plot formulae in the literature. We shall discuss
the general case in this section. Specific formulae for lome of the standard
designs will be given in later chapters, either in the text or as exercises.
Suppose that the experimental plan called for N + 1 observations or plots.
We have N observations, Y, for which the design matrix is X (N x p). The
(N + l)th observation is missing; its design matrix is a single row x. We shall
choose for our estimate of the missing observation the value y that minimizes the
sum of squares for error when we carry out the least squares procedure for all
(N + 1) points (the N actual data and y).
Writing 0 for the vector of estimates obtained in this way, we minimize
Se =(YXO)'(YX3)+(yx0) 2 .
The first term is minimized if we take 0 to be any solution vector to the normal
equations for the N actual data points, and the second term is zero if we take
y=x0.
From an abstract point of view we appear to have gained nothing except to
say that the best estimate y is obtained by solving the normal equations for the
actual data and substituting 0 for (3 to obtain y = xî. In practice, however, the
design matrix (X', x')' is usually chosen in such a way that the augmented
normal equations can be solved very easily, whereas the equations X'X[3 = X'Y
may be more awkward to handle computationally. Since y x^3 = 0, an
equivalent procedure, which we shall use later, is to choose for y the value that
will fit the model with zero residual when the least squares procedure is carried
out on the set of (N + 1) points consisting of the N data points and the missing
plot (x, y).
[Ch. 2
Exercises
1. Let y l and y a be independent random variables each N(, 1). Then Z =
(Y Y2)/V2 N(0, 1) and Z 2 _ X 2 (1). Writing Z 2 = Y'AY where Y' =
(Yi, Ya) show that the expression for the moment-generating function of Z 2 :
M(t) = II 2Atj -1 j 2 exp '{I (I 2At) -1 }/2 reduces to (1 2t) -112 .
2. Let Y N(, I). We proved the theorem that Y'AY X 2 (n,), if, and only if, A is
an idempotent matrix of rank n, for the case = 0. Prove it for the case ^ 0.
3. Prove that if Y N(0, V) then Y'AY X 2 (n) if, and only if, AV is an idempotent
matrix of rank n.
4. If Y N(0, V) show that the two forms Y'AY, Y'BY are independent if, and only
if, AVB = 0.
5. If A and B are nonnegative definite matrices and if Y N(0, I), show that the
forms Y'AY, Y'BY are independent, if and only if, tr (AB) = 0.
6. Let Y N(0, D) where Dis diagonal, and let
A = D - ' (D - 'JD -1 /1'D -1 1).
Show that Y'AY X 2 (n 1).
7. Let Y be a vector of n independent normal random variables with E(y) = , and
V(y ) = wr 1 , and let y = 1 w,y,/> w,. Show that
1 ws(Y1 9) 2
X 2 (n 1).
CHAPTER
Experiments
with a
Single Factor
If an agronomist wishes to compare the yields of five varieties of wheat, an

obvious procedure would be to mark off twenty plots of land and to sow variety
A on four of them, variety B on four more of them, and so on. The varieties are
assigned to the plots by some random procedure, and from each plot comes a
single observation, the yield on that plot. This simple experiment is called a
completely randomized experiment because of the method of assigning the
varieties. We shall begin our discussion of experimental designs by considering
this kind of experiment.
An industrial example would be the comparison of four different operating
temperatures on the yield of a product from a chemical reactor. Twenty runs
could be made, five at each temperature, and the yield of product at each run
recorded. Again, the temperatures should be assigned to the runs at random.
Suppose that the process involves the use of an acid catalyst which deteriorates
with time. If the experimenter were to make the runs at the highest temperature
first, then the runs at the next highest temperature, and finally the runs at the
lowest temperature, he would have difficulty interpreting his results. If the yields
at the high temperatures were greater than the yields at low temperatures, could
he be sure whether the apparent drop in yield resulted from the reduction in the
operating temperature or from the deterioration of the acid? The random assignment of the temperatures to the runs, or of the varieties to the plots, is essential to
our analysis of the experiment. The models for the experiments will be linear
models like those of the last chapter. In the model for each observation there
will be an error term, e; these errors are assumed to be uncorrelated.
In general, a set of different treatments are applied to N experimental units,
one treatment to each unit, and an observation is made on each unit. The
39
40 Experiments with a Single Factor
[Ch. 3
theory of experimental design was first developed in agriculture and much

of the jargon is taken from that field. We shall refrain from using the expression
experimental unit. Instead we shall most often use the agricultural term plot,
or, when talking of industrial experiments, the terms run or point. An observation
will commonly be referred to as a yield or a response.
Some of our examples will be concerned with octane numbers. Producers of
gasoline are interested in the octane number of their product. If a driver uses a
gasoline with too low an octane number, the engine of his car will knock. Larger
cars with engines that have higher compression ratios need higher octane than do
compact cars to prevent knocking. Thus, high octane number is a desirable characteristic of a gasoline. The octane number may be determined in the laboratory or
on the road by a procedure that may be described in a simplified fashion as follows.
There is a standard engine on which the compression ratio may be varied.
The engine is run with the gasoline to be tested. The compression ratio is increased until knocking occurs, and the reading on the meter is noted. Then the
procedure is repeated using reference fuels. These are mixtures of normal
heptane and iso-octane. A reference fuel composed of X per cent iso-octane and
(100 X) per cent normal heptane is defined as having octane number X. Two
reference fuels are used, one with a slightly higher octane number than the unknown gasoline and one with a slightly lower octane number. The readings on
the compression ratio meter at which the engine knocks using each of the
reference fuels are noted, and the octane number of the gasoline being tested is
obtained by interpolation. An observed octane number is a random variable.
A standard deviation of about 0.3 ON is not unusual.
A similar procedure can be followed using automobiles on the road to obtain
"road-octane numbers." The manufacturers of both cars and gasolines conduct
tests under road conditions to find both the road-octane numbers of commercial
gasolines and the octane requirements of commercial automobiles. Because the
modern automobile engines differ somewhat from the standard engines, and
from each other, and also because the components in commercial gasoline are
not iso-octane and normal heptane, it often happens that even apart from
random error a test fuel will have different octane ratings in different cars and
these ratings will differ from those obtained in the laboratory. The octane number
of a gasoline can be increased by the use of special additives such as tetra-ethyl
lead or tetra-methyl lead, or by adding some of the aromatic hydrocarbons
found in crude oil. The latter, however, are also in demand for petrochemicals,
and the managers of an oil company have to decide how much of their supply
should be used for each purpose.
An engineer wishing to compare five gasolines might take twenty cars, assign
each gasoline to four cars, and observe the octane number for each cargasoline
combination. The choice of cars is important to the analysis of the experiment
and to the use of the results. One should not experiment solely on Volkswagens
if the results are to be used to make decisions about Cadillacs, nor should one
expect crops to yield the same amounts in the Central Valley of California as
they do in the coastal area.
Sec. 3.1 ]
Analysis of the Completely Randomized Experiment 41
3.1 Analysis of the Completely Randomized Experiment

In either of the two experiments just mentioned we can take as a model
yij=
i= 1, 2,...,t;j= 1,2,...,r,
s +e15,
where y{ f denotes the jth observation on the ith gasoline, or the yield on the jth
of the plots that were sown with the ith variety of wheat, , denotes the " true "
value of the octane number or the yield of the ith gasoline or variety; i.e.,
E(y 15 ), and e15 is the random error of the (ij)th observation or plot. We assume
that the , are unknown constants, and that the e11 are uncorrelated random
variables, each distributed with the same mean, zero, and the same variance Q 2 .
We use the general term treatments for the gasolines or the varieties and speak of
an experiment with t treatments.
In addition we may make the normality assumption, namely that the elf are
normally distributed, i.e., the vector e of the e15 has a multivariate normal
distribution with mean vector 0 and covariance matrix Ia 2 .
The least squares estimate of p-, which is the maximum likelihood estimate
under the normality assumption, is the average of the observations on the ith
treatment. Writing
G
Ti = > y1j,
y11,
y,. = Yt = Ti/r,
we have
(3.1)
11 = yt.,
Se = t(r 1 )s 2
( Yti
y1.) 2 =
2u
T2 /r.
(3.2)
If the number of observations is not the same for each treatment, we denote the
number of observations on the ith treatment by r, and let y,. = Tlr,. .Then
Equation 3.2 becomes Se = I i Ii yi'i l { T,2 /r1 .
We shall use N to denote the total number of observations, G for the grand
total 1, I i ytj of the observations, and y.. or y for the grand mean G/N. The case
where ri = r for all i will be called the equireplicate case. (Some people like to say
that the ith treatment is replicated r, times; we shall do so when we consider
incomplete block designs in a later chapter).
Under the normality assumption the hypothesis , = ,. can be tested by
Student's t test. In the equireplicate case the test statistic
t = (yi Y'i) (r /25 2 )
has, under the hypothesis, Student's t distribution with t(r 1) d.f.
[Ch. 3
An alternative formulation of the model commonly used is

yi=
ri + eu ,
where yi and e15 are the same as before, is a general mean, and T{ is an unknown constant parameter associated with the ith treatment. We shall call -Ti the
main effect of the ith treatment. Neither nor T{ is estimable; if we were to add
any arbitrary constant c to and subtract c from each T,, the expected value of
each observation y ij would remain unchanged.
The model may also be written E(y 5) = + l h XhTh where x, = 1 if the
observation is made on the ith treatment and zero otherwise. The design matrix
X has N rows and t + 1 columns; assuming that there is at least one observation
on each treatment, the rank of X is t. The first column, corresponding to , is 1 N ,
a vector with every element unity. The remaining columns, corresponding to
treatments, consist of zeros and ones, and in any row exactly one of these
columns has unity as its element. Thus, the first column is equal to the sum of the
last t columns, and the last t columns are linearly independent.
If, therefore, we wish to use this model, we must impose a side condition on
the parameters. Three possibilities come to mind.
(i) = 0. This sends us back to the original model with Ti instead of ,.
T l = 0. This is useful in some computer programs. It amounts to designating treatment 1 as a base treatment. If treatment 1 is a control or
standard treatment, this has some merit.
(iii) 1 t ri ri = 0. This is the most commonly used side condition. It reduces to
1, -r, = 0 in the equireplicate case. An attractive consequence is that
with this side condition = 1, I f yij /N = y. We shall use this condition.
(ii)
Although and Tt are not estimable, ( + Ti) is estimable. Its least squares
estimate is y i ., which is the same as i was before. Similarly, the hypothesis
= . is the same as the hypothesis T, = -r,., and can also be tested by y 1 . yi ...
One might ask what, if anything, is gained by introducing the additional
complication of singular design matrices. It will actually turn out that when we
consider several factors, or experiments with blocking, we shall be faced with
singular design matrices anyway. The advantage in this simple experiment lies in
the fact that the new model gives a sum of squares for treatments.
The simplest model that could have been chosen for the data would have been
ysi=+eu,
which is equivalent to assuming that all the i are equal, or that with the side
condition 1 Ti = 0 all the -r, are zero. Under this model = G/N, and the error
sum of squares is
S=
G=yii
t
G 2 /N.
Analysis of the Completely Randomized Experiment 43
Sec. 3.1]
When the treatment parameters r, are added to the model the estimate of is
stil! . = y; the estimate of Tt is Tt = y,. y, and the new sum of squares for
regression is
SR = G + Tf zt = G 2 /N +
i
TT(yc.
9) _ T2 /rr .
t
The reduction in the sum of squares for error or the increase in the sum of
squares for regression because of the addition of the treatment parameters to the
model is thus 1 T,a/r G 2 IN. This is called the sum of squares for treatments
and is denoted by S. The sum of squares for error, S e , is the same as it was with
the first model.
We now have the subdivision of the sum of the squares of the observations
{
Y'Y = G 2 /N
+ [> T?/r G 2 /N] +
(y{r y,) 2
or
Y'Y =C+St + Se .
The term G 2 /N is called the correction for the mean and is often written as
C. Y'Y is called the raw sum of squares; the corrected sum Y'Y C is called
the total sum of squares, and we shall occasionally denote it by ST .
The quadratic forms C, St , and Se have ranks 1, t 1 and N t respectively.
Under normality they have, upon division by a2 , independent chi-square
distributions. The expected value of C is a2 + N -1 (N + r,-r,) 2 . 1f we write
rr r1 = NT, this becomes a2 + N( + T) 2 , and if we take the side condition
r,-r, = 0, E(C) reduces to u2 + N2 . To compute E(St) we note that
E(Ti2 /r1 ) = rt 'E(> e, ! + ri u + ri T,) = a + rt(N, + Tt) .

2
Then
E(S) = (tal + Ng + 2
2
rt T, + rir2) E(C)
_ (t 1)a 2 +
r,(_, T) a .
This information is summarized in Table 3.1.

TABLE 3.1
ANALYSIS OF VARIANCE
Source
SS
EMS
d.f.
Total
Y'Y C
N 1
Treatments
Error
7 /r, C
Y'Y 7 /r,
t 1
a2 + ( t
N t
aa
1)' r ( r, T)z
{
[Ch. 3
There has been a tendency in the last few years to replace the words analysis of
variance by the single word, Anova. In the column headings SS stands for sum
of squares and EMS for the expected value of the mean square, the Jatter being
the sum of squares divided by the number of degrees of freedom. The mean
squares for error and treatments may be denoted by Me and Mt respectively,
although we shall usually use s 2 rather than Me .
The hypothesis Ho : T l = T 2 = = Tt , which is the hypothesis that there
are no treatment differences, is tested by the statistic F = Mt /s 2 . Under the
normality assumption F has, when the null hypothesis is truc, a central
F(t 1, N t) distribution; when the null hypothesis is false, the distribution
of F is a noncentral F with a noncentrality parameter ,1 = r{ Ti /2a 2 . The
test has for its critical region the upper tail of the F distribution, and the null
hypothesis is rejected if F exceeds the upper a value of F. Some authors add a
fifth column to the analysis of variance table, in which they enter values of F;
these entries are indicated by one asterisk if they exceed the 5 per cent value of F
and with two asterisks if they exceed the one percent value. We mean by the
upper a value of F( l , 2 ) the value F* defined by fF f(u) du = a where f(u)
is the probability density function of a random variable, u, having the F( 1 , 0 2 )
distribution.
This hypothesis, which may also be expressed as r = 0 where T is the vector
of the i, is of interest in uniformity trials where, for example, an agronomist
wants to ascertain whether several strains of wheat are essentially equivalent
or whether the yield of a particular variety is reasonably constant under varying
conditions. (Power calculations for the F test involve integrals for the noncentral distribution. Tang (1938) has prepared a set of tables for use in power
calculations. Subsequently, charts were made by Pearson and Hartley (1951) and
by Fox (1956). These charts are somewhat easier to use than the Tang tables.)
In many industrial situations, however, the hypothesis that there are no
treatment differences is of little interest; the scientist expects to find differences
and is more concerned with finding which treatments differ or in investigating
contrasts among the treatments.
We shall confine ourselves now to the equireplicate case with the normality
assumption and consider contrasts in the treatment effects, = ^ c,T, where
1 ct = 0 (or, equivalently 0 = c'T where c'1 = 0). The difference between two
effects, T i i-, will sometimes be called a comparison. The least squares estimate
of 0 is = 1 c iT, = c'T = 1 C,yt .; ^ is normally distributed with variance c'co 2 /r,
and an obvious test for the hypothesis 0 = 0 is Student's t test.
To carry out a t test for a hypothesis such as ri = T,. we can use a least
significant difference (LSD). We have Tt T t . = y t . y.. Under the null
hypothesis, (yi . y1 ..)/(2s 2 /r)" 2 has Student's distribution with (N t) d.f.
The hypothesis is rejected in the two-tailed test if 1 y1 . y i .. > t *(2s 2 /r) 112
where t * is the critical value of t with the desired significance level a.
The quantity t*(2s 2 /r) 112 is the LSD. The multiple t test is carried out by
comparing the treatment means and by declaring any pair of treatments whose
means differ by more than the LSD to be different.
Sec. 3.3]
Scheff's S Statistic 45
This is a simple and convenient procedure, but the criticism has been made
that when this LSD is applied to all possible pairs of treatments there is an
increased chance of type-one errors. The argument of the critics is essentially
that whereas it is true that the probability is 0.95 that any two observations
made at random from a normal population will differ by less than 1.96a, if we
take a sample of n > 2 observations, the probability that the largest and the
smallest of them differ by more than 1.96a is greater than 0.05 and increases
as n increases. This has led to the development of tests with the property that
the probability of a type-one error per experiment (rather than per comparison)
is a. Three such tests in common use are due to Newman (1939), Duncan (1955),
and Tukey (1953); Newman's test was revived by Keuls (1952) and is more
commonly known as the Newman-Keuls test. We shall derive Tukey's test.
Accounts of the other two tests are found in the references cited and in the books
by Federer (1955) and Miller (1966).
3.2 Tukey's Test for Treatment Differences

The Studentized range is defined as follows. Suppose that x 1 , x2 , ... , x n is a
sample of n observations from a normal population with mean and variance
a 2 and that 5 2 is a quadratic estimate of a2 with 0 d.f., i.e., s 2 /a 2 is distributed as
chi-square with 0 d.f. The range R(x) is the difference between the largest and
smallest members of the sample, and we suppose further that R(x) and s 2 are
independent. Then the random variable R(x)/s is said to be distributed as the
Studentized range with 0 d.f. With probability (1 a) we have R(x) < gn;ms,
where qn;m is the upper a point of the Studentized range.
In Tukey's test we replace the x, of the previous paragraph by the treatment
means y i . and argue as follows. Under the null hypothesis, the probability is
(1 a) that R(y,.) < qt ms/^/r where 0 = N t. Thus the probability is
(j: a) that Ti Ti . < q^ m s/i/r for all pairs i, i'. We may therefore use q^ m s/-,/r
as the least significant difference and declare that T t # Ti , if, and only if,
IY{ Yi'.1 > sgiN-t/\/r. Tukey's test is somewhat too conservative to suit
some experimenters. The tests of Newman and Keuls and of Duncan provide
compromises between Tukey's test and the t test.
3.3 Scheff's S Statistic

The three tests mentioned before are useful for testing comparisons. The
experimenter usually has a few specific contrasts in mind when he plans his
experiment, and it is appropriate to use the t test for these. However, if he is
going to wait until after the data have been collected and then test those contrasts that strike his eye during the analysis, the criticisms of blanket use of the t
test for the whole set of t(t 1)/2 comparisons are no less valid for the set of all
contrasts. Scheff's statistic provides a general hunting license for the experimenter who wishes to test any estimable function that takes his fancy after he has
[Ch. 3
taken a look at the data. The derivation of this statietic may be found in Scheff's
original paper (1953) or in his book (1959). The procedure is the following.
Let = c'T be an estimate of an estimable function 0 = c'T, and let s 2 be an
independent quadratic estimate of the variance based on a X 2 distribution with
0 d.f. We write & = c'cs 2 . Scheff's result is as follows: The probability is
(1 - a) that, for all possible estimable functions,
^-S& fr <
+S&*,
where S 2 = (t - 1)Fa (t - 1, 0), 0 = N - t and Fa (t - 1, 0) is the upper

value of F. This test gives longer intervals than the t test when there are more
than two treatments, but that is the price that has to be paid to have an error
rate of a per experiment. For simple comparisons, Tukey's method gives
shorter confidence intervals than Scheff's.
We have mentioned only a few of the numerous alternatives to the t test in the
analysis of variance. The alternatives to which we have referred all assume
normality. There are also nonparametric procedures available, such as those of
Kruskal and Wallis (1952) and Steel (1960). The subject of multiple comparisons
in the analysis of variance is a topic that fills a book on its own [Miller (1966)],
and we shall pursue it no further here. From now on we shall have in mind the
t test when we think of tests of significance for contrasts. A great deal has also
been written about the robustness of the t and F tests in the face of departures
from normality, notably by Box (1953) and by Scheff (1959, Chap. 10); lens is
known about the robustness of the other procedures that are based upon the
assumption of normality.
3.4 A Worked Example

Consider the following data from an experiment with five gasolines A, B, C,
D, E and four observations on each gasoline. Note that G = 1834.2 and
y = 91.71.
Observations
Gas
A
B
C
D
E
91.7
91.7
92.4
91.8
93.1
91.2
91.9
91.2
92.2
92.9
90.9
90.9
91.6
92.0
92.4
Total Mean
90.6
90.9
91.0
91.4
92.4
364.4
365.4
366.2
367.4
370.8
91.10
91.35
91.55
91.85
92.70
In computing the sums of squares in Table 3.2 the arithmetic is simplified if

we subtract 90.0 from each observation. If this is done we have Y'Y = 67.96,
C = 58.482, and 1 T,z/4 = 64.59. The analysis of variance table follows.
A Worked Example 47
Sec. 3.4]
TABLE 3.2
Source
SS.
d.f.
MS.
19
Total 67.96 58.482 = 9.478
4 1.527 = Mt 6.78
St 64.59 58.482 = 6.108
Se 67.96 64.59 = 3.370 15 0.225 = s 2
The upper 5 per cent and 1 per cent points of F(4, 15) are 3.06 and 4.89. The
ratio Mt /Me = 6.78 exceeds both these values, and so the hypothesis Ho : r = 0
is rejected. We could put two stars after 6.78 in the F column to show that 6.78
exceeds the upper one per cent value.
It should be noted that the fact that S t + Se adds up to the total sum of
squares does not in itself provide much of a check on the calculations, because it
provides no check that the quantities Y'Y, C and T'T have been computed
correctly. If we had erroneously calculated Y'Y as 77.96, we should still have
obtained an analysis of variance table which, superficially at least, would look
valid; on the other hand, if we had mistakenly calculated T'T/4 as 54.59, that
would have given a negative value of St , which would have indicated that either
T'T/4 or C was wrong. A check can, however, be made by obtaining the sum of
the squares of the deviations Ij (yt5 y,.) 2 for each treatment and then checking
that S. = 1 Ij (yu
{
Having rejected the hypothesis of no treatment effects, we now look at the

treatment comparisons. The estimated variance of a mean is s 2 /4.
Using the t test we have, at a = 0.05, ta,15 = 2.131, and the LSD is
2.131 //(2s 2 /4) = 2.131 x/0.1125 = 0.714. Comparing differences between means
to 0.71 we should conclude that gasoline E has a higher octane number than
any of the other gasolines and that D exceeds A.
For the more conservative Tukey test, the upper 5 per cent value of the
Studentized range for five samples with 15 d.f. for error is 4.37 [see, for example,
Pearson and Hartley (1954)] and the LSD is 4.37\/(S 2 /4) = 1.036. Using this
test, we should only conclude that E> A, B, C.
For an example of the use of Scheff's S statistic, consider the estimable
contrast 0=T 1 +'r2 T 4 T 5 .We have +,i=1.10+1.35-1.85-2.70=
2.10; V(î) = 4a 2 /4 = a 2 . With a = 0.05, S a = 4Fa (4, 15) = 12.24, and
S = 3.5. Then
(3.5)(0.474) 2.10 < < (3.5)(0.474) 2.10,
3.76 < 0 < 0.44,
and the hypothesis 0 = 0 would be rejected.
[Ch. 3
3.5 Qualitative and Quantitative Factors and Orthogonal

Contrasts
We have spoken so far of an experiment with t treatments. We may also talk
about an experiment with a factor at t levels. Thus, the gasoline experiment
could have been described either as an experiment with five gasolines or as an
experiment with one factor, gasoline, at five levels. The Jatter description may
be more appropriate if, for example, the gasolines differ only by the percentage
of olefins present in their compositions. It has already been remarked that if an
experimenter is concerned with a specific set of contrasts which have been
decided upon beforehand, it is reasonable for him to use the t test or, equivalently, the F test. It is convenient if these contrasts are orthogonal, but it is not
essential.
In the previous chapter we mentioned that when Y'Y is split into sums of
squares Y'A {Y these quadratic forms can in turn be expreseed as sums of the
squares of orthogonal linear combinations, z ti , of the observations, corresponding
to an orthogonal transformation Z = BY. If one of these linear combinations
is taken to be y/^/N, corresponding to the mean, the others must be contrasts
in the observations. In the equireplicate case a contrast in the treatment totals
is also a contrast in the observations. Suppose that a'Y and b'Y are two linear
forms. We say that they are orthogonal if a'b = 0. Similarly, two linear forms
c'T and d'T in the treatment totals (T being the vector of the treatment totals) are
orthogonal if, and only if, r ci d; r; = 0. If r = r for all i this becomes c'd = 0.
We shall again confine ourselves to the equireplicate case.
The correction for the mean G 2 /N may be written T'JT/N where J is a (t x t)
matrix with every element unity. Then St = r -1 T'T G 2 /N is actually a
quadratic form in the treatment totals, and the (t 1) linear combinations z t in
the transformation St --> Z'Z are orthogonal linear combinations of the treatment totals. Any set of orthogonal contrasts in the treatment totals will suffice,
and the experimenter can choose a set that satisfies his particular purposes.
If there are four treatments, we may take the set of orthogonal contrasts:
z l = (T, T2)/1/( 2 r);
z2 = (Ti + T2 2 T3)/1/( 6 r);
z3 = (T1 + T2 + T3 3 T4)/\/( 12 r),

or, as we shall see a little later in this section, we could also take, with the
proper divisors, the set
zl = ( 3 T1 T2 + T3 + 3T4),
z2 = (+ Tl T2 T3 + T4),
z3 =(T1 +3T2-3T3+T4).
In either case, St = zi + z + z3.
Sec. 3.5]
Qualitative and Quantitative Factors and Orthogonal Contrasts 49
We do not need to restrict ourselves to contrasts z = c'T with c'c = 1, i.e.,

to vectors c which have unit length. If z = c'T is any contrast we can always
divide each coefficient by \/c'c to obtain a coefficient vector with unit norm.
The sum of squares for any contrast is thus z'z/rc'c. Upon division by o 2 it has
a X 2 distribution (central, if E(z) = 0) with one degree of freedom.
Consider an experiment involving six cars, Plymouth, Ford, Chevrolet,
Dodge, Mercury, and Buick, and denote the corresponding treatment totals by
T1 , T2 ,. . ., T6 , respectively. One possible set of orthogonal contrasts follows.
T1
+1
+1
0
+1
0
Tz
T3
T4
T5
Ts
+ 1 + 1 1 1 1
0
0
2
0
1
0
0+ 1
0
+1
0 +1
0
0
0
2
0
1
0
+1
Low priced cars vs. medium priced cars

Plymouth vs. Chevrolet
Dodge vs. Buick
Ford vs. other low priced cars
Mercury vs. other medium priced cars
The engineer can also test the Ford vs. the Plymouth if he wishes, even though
that contrast is not orthogonal to the others.
The factor used in this experiment is a qualitative factor. The experimenter is
interested in establishing and estimating the differences, if any, between the
various makes of cars. The prediction of the performance in some subsequent
run at some intermediate level of the factor, such as a hybrid of Plymouth and
Mercury, is not relevant to the experiment. However, with a quantitative factor
such as temperature, which may in a particular experiment have such levels as
100, 200, 300, or 400 degrees, we are usually interested in the whole range
covered. We might wish to use the data to predict what the response would be if
a subsequent run were to be made at 150, or even to predict where in the range
is the "best" temperature in the sense of the temperature at which the response is
either a local maximum or a local minimum.
This leads to the fitting of linear, quadratic, and higher order trend lines,
y = f(t), for predicting the response as a polynomial function of the temperature.
Consider, for example, the cubic polynomial
y = a + bt + ct 2
+ dt .
3
One approach is to fit this model to the data by least squares, test the hypothesis
d = 0, and then, if that hypothesis is accepted, to fit the reduced model
y=a+bt+ct 2 .
In the reduced model we now test the hypothesis c = 0. If that is accepted, we

fit the straight line y = a + bt, and so on.
The difficulty with this procedure is that unless the least squares estimates of
a, b, c, d are orthogonal contrasts, the equation has to be refitted anew at each
[Ch. 3
stage. We can test in the analysis of the cubic model whether the cubic term is
necessary, but unless the contraets are orthogonal we cannot test simultaneously
c = 0 and b = 0.
An alternative approach is to make use of orthogonal polynomials. This is
conveniently done if the levels of the factor are equally spaced. We illustrate the
procedure by an example. Suppose that one observation is made at each of the
five temperatures, 100, 200, 300, 400 and 500 degrees, and that we wish to fit
the cubic function
E(y)
= R O + Nlx + N2x2 + N3x 3 ,
where x denotes the temperature and y is the response.

Instead we fit the polynomial
E(y) = Youo +
Y1u1 +
Y2u2
+ Y3u3,
where u0 = 1, and the ui (i = 1, 2, 3) are ith degree polynomials in x such that

1 u,ui = 0 for each pair 1, j (i j), the summation being made over all the
values taken by x. In practice it is convenient to take the values of u 1 to be small
integers and to derive u 2 and u3 as polynomials in u 1 rather than x.
Considering first u 1 , let u 1 = a l + b l x. The condition u o u l = 0 gives
5a 1 + 1500b 1 = 0.
This is a single equation in two unknowns, and the solution which gives the
smallest integer values of u 1 is a 1 = - 3, b l = 0.01, so that u 1 takes the values
(x - 300)/100 or -2, -1, 0, +1, +2.
For u 2 = a2 + b2 u 1 + c2 u1, the conditions uo u2 = 0, 1 u l u2 = 0 give
5a2 + 0b 2 + 10c 2 = 0,
10b 2 = 0,
whence, b2 = 0, a2
= - 2c 2 .
u2
The choice a2
= - 2, c 2 =
1, or u 2
u1 - 2, gives
=+ 2,-1,-2,-1,+2.
For u3 = a 3 + b 3 u 1 + c3ui + d3 ui, we have three conditions: uo u3 = 0,

u 1 u3 = 0, 1 u 2 u3 = 0. The last of these may be written as a 2 u 3 + 1 b 2 ul u3 +
c2 uiu 3 = 0. By virtue of the first two, this reduces to uiu3 = 0, giving the
three equations:
101, 3 + 34d3 = 0,
5a3 + 10e3 = 0,
10a3 + 34e 3 = 0,
whence a3 = c3 = 0 and b3 = -17d3 /5. The choice d3 = 5, b 3 = - 17, or

u3 = 5ui - 17u 1 , gives values u 3 = - 6, + 12, 0, -12, + 6. The usual choice is,
therefore, (5ui - 17u 1 )/6, giving
u3
= - 1,
+2,0, -2, +1.
Sec. 3.5]
Qualitative and Quantitative Factors and Orthogonal Contrasts 51
The estimates of y, are now easily obtained as yt = u, y/ ui with cov

(y^, y{ .) = 0 (i 0 i') and V( f ) = a 2 / ui'. Thus, writing the responses as y l ,
Y2, ... , y5 in ascending order of temperature and letting the vector Y be

(Y^, Y2, , Ys)', we have
yo=1y15=y
y1=(-2,-1,0,+1,+2)Y/10,
Y2 = (+2, 1, 2, 1, +2)Y/14
y3 = (-1, +2,0, 2, + 1)Y/10.
The contrasts
(-2, 1,0, +1, +2)Y,
(+2, 1, 2, 1, +2)Y,
and
(-1,+2,0,-2,+1)Y
are called the linear, quadratic, and cubic contrasts.

Once the estimates y i have been obtained, we can, if we wish, compute
estimates of the coefficients fl. For purposes of prediction it is easier to work in
terms of y, and u i . Thus, if it is desired to extrapolate and to predict the response
at a temperature of 600 degrees, we have u l = 3, u2 = 7, u 3 = 14, and y = y o +
32 1 + 72 2 + 142 3 . The sizes of the cubic and quadratic coefficients give an
indication of the dangers involved in extrapolation.
The following example illustrates this point. The earnings per share of stock
of American Telephone and Telegraph Company for the years 1963-1967 were
reported to be $3.03, $3.24, $3.41, $3.69, and $3.79. Suppose that we try to
predict from this data the earnings per share in 1968. We have y o = y = 3.432,
y l = 0.197, y 2 = 0.008, and y 3 = -0.014. With these coefficients we can
obtain linear, quadratic, and cubic estimates of the earnings for 1963-1967 and
projections for 1968.
Year
1963 1964 1965
1966 1967 1968 (Estimated only)
$3.03
$3.24
$3.41
$3.69
$3.79
Linear estimate
$3.04
Quadratic estimate $3.02
Cubic estimate
$3.04
$3.24
$3.24
$3.22
$3.43
$3.45
$3.45
$3.63
$3.64
$3.66
$3.83
$3.81
$3.80
Actual earnings
$4.02
$3.97
$3.77
Notice that throughout the range of actual reported data there are only slight
differences among the linear, quadratic, and cubic estimates; the range of the
three estimates is never more than three cents. At the extrapolated point, however, the question of whether the higher order terms are included becomes more
important. Adding the quadratic term to the linear estimate reduces the estimate
by five cents; the cubic term makes a difference of twenty cents (about 5 per
cent) in the extrapolated value.
[Ch. 3

Suppose now that the gasolines used in the experiment discussed Barlier in
this chapter differed only in the amount of some quantitative factor; in particular,
suppose that gasolines A, B, C, D, and E were identical except that they contained x = 0, 1, 2, 3, 4 cubic centimeters, respectively, of some additive per
gallon. We can subdivide St into linear, quadratic, cubic, and quartic components. The sum of squares for the linear contrast is
S l = (-2T1 T2 + T4 + 2T5 ) 2 /(10 x 4) = (14.8) 2 /40 = 5.476;
the sums for the quadratic and cubic contrasts are

S2 = (2T1 T2 2T3 T4 + 2T5 ) 2 /(14 x 4) = (5.2) 2 /56 = 0.483,
S3 = (T1 + 2T2 2T4 + T5 ) 2 /(10 x 4) = (2.4) 2 /40 = 0.144.
The quartic contrast is (1, 4, 6, 4, 1)Y, and the corresponding sum of

squares is (1.2) 2 /280 = 0.005.
These four sums of squares add up to St , which provides a check on our
calculations. If the corresponding coefficient in the polynomial is zero, S,/s2 has
an F distribution with 1 and 15 d.f. The upper 5 per cent value of F(1, 15) is
4.54. The only component that exceeds 454s 2 = 4.54(0.225) = 1.022 is S 1 . We
conclude that the higher order coefficients can be omitted from the polynomial.
The estimated line is
Y=Yo+Yiui,
wherey o =y= 91.71, y 1 = 14.8/40 = 0.37, and u 1 = x 2,ory=90.97+

0.37x. The linear estimates and their deviations from the treatment means are
x
y,
90.97
91.34
91.71
92.08
92.45
y,. y, +0.13 +0.01 0.16 0.23 +0.25
We note that (y,. Y,) = 0 and 4 (y,. 9t) 2 = 4(0.1580) = St S,.

In the previous model the levels of the factor were associated with unknown
constants T,. In the random effects model we have
yti=F^+at+e,1,
Sec. 3.8]
The Power of the F Test with Random Effects 53
where a, is a random variable drawn from a (normal) population with mean

zero and variance a. The random variables a, are uncorrelated with each other
and with the eij . It is convenient in this model to denote the variance of the
errors et j by a.
Thus, one might envision making n l batches of a chemical product by the
same process and drawing n 2 samples from each batch, in which case yij would
denote the response of the jth sample from the ith batch. We assume that the
process gives a mean response ; o denotes the batch-to-batch variation; o
measures the variability of samples in the same batch. Similarly, a geneticist
might pick n l dams at random from a population under study and then choose
n 2 offspring from each dam. This model is also called the components of
variance model, a and a being the components of variance.
The N observations are no longer all independent. Observations made on the
same batch are now correlated. We take the same subdivision into sums of
squares that we used in the fixed effects case. The expectation of the sum of
squares between batches, St , is (c + n 2 a)(n 1 1), and the hypothesis a l = 0
is tested in the usual way by the F statistic.
It is convenient to write q 1 for (n l 1) and 0 2 for n 1 n 2 n l . The
estimates of the components of variance are given by Q = 2 1 Se = s 2 and
Q + n 2 & = 'j 'S = M. Under normality, the two sums of squares Se fo
and St /(a + n 2 a) have independent chi-square distributions with 1 and 0 2
degrees of freedom, respectively.
In this relatively simple experiment we can obtain the maximum likelihood
estimates of aA and a without too much diffiiculty. However, in more complex
experiments involving random effects the solutions to the maximum likelihood
equations are not so easily obtained. We shall adhere in the future to the
procedure that we used this time. We shall take the same breakdown of Y'Y
into sums of squares that we get in the fixed effects case and obtain our estimates
of the components of variance from them.
3.8 The Power of the F Test with Random Effects

The hypothesis a = 0 is tested by the statistic 9 = Mt /s 2 Under normality,
St /(a + n2o) has a X 2 distribution with n l 1 d.f., and Se /Q has an independent x 2 distribution with n l (n 2 1) d.f., so that
.
.F = (1 + n2a fa)F(nj 1 , n1(n2 1 ))

We reject the hypothesis o = 0 if JF > Fa (n l 1, n l (n 2 1)), i.e., if
F(n l 1, n l (n 2 1)) > Fa(ni 1, n1(n2 1 ))

1 + n2 0
where 0 = a /oe. Thus, in the random effects case the power of the F test
depends only upon the tentral F distribution.
[Ch. 3
EXAMPLE. Suppose that n l = 4, n 2 = 6, 0 = 2.5, and that a is taken to be

0.05. The upper 5 per cent value of F(3, 20) is 3.10. Then the power of the test
is P[F(3, 20) > 3.10(1 + 6(2.5)) '] = P[F(3, 20) > 0.19], which is approxi-
mately 0.90.
3.9 Confidence Intervals for the Components of Variance

A confidence interval for or can be obtained in the usual way using the X2
distribution
[Se'Xu(42)] < ae < [+SeI XL (02)],
where for confidence, (1 a), Xu, and XL are the upper and lower a/2 points of
the chi-square distribution with 2 d.f.
An approximate confidence interval for a can be obtained in the same way
using St and assuming that 52 = o. We should then have
[St/Xu(6 1)] s 2
< n2QA < [st/XL(6 1)] s 2
Unfortunately, this method can give negative lower limits even when the F test
rejects the hypothesis a = 0. A preferable approach is to argue, as in the
previous section, that F/(1 + n 2 0) has a central F(o 1 , 0 2 ) distribution. Then
we can obtain a confidence interval with confidence (1 a) by writing
(1 + n 2 0)FL < .F < (1 + n 2 0)FLI ,
and manipulating it to obtain

n2 1 (. /Fv 1)s 2 < QA < n2 1 (^/FL 1)s 2 ,
where Fu and FL are the upper and lower a/2 points of F(O 1i 0 2 ). FL can be
obtained from the usual tables by the identity Fl , 2 (q 1 , i2) = 1 /F12(02, 0^)The general problem of confidence intervals for the components of variance
has been investigated by Bulmer (1957) and by Scheff (1959). There are also
interesting problems involving the allocation of observations. For a fixed
number N = n 1 n 2 of observations, what is the best ratio n l /n 2 for the experimenter to take in designing his experiment? Should he perhaps choose an
unbalanced design having unequal numbers of observations on the batches?
For the case in which the same number of observations is made on each batch,
the problem was solved by Hammersley (1949) [see also Scheff (1959)].

Suppose that instead of being fixed effects the gasoline effects were random.
Suppose that the gasolines were drawn from five different tanks, all filled with
Sec. 3.11]
The Randomized Complete Block Experiment 55
the product from the same plant, and that apart from random variation the
contents of the tanks were essentially the same. We have
Yu=i'
+tt+ei3,
where tt N(0, Q). The hypothesis Q = 0 is rejected by the F test. The point
estimates of o and a are 8 = Me = 0.225 and & = (Mt Me )/4 = 0.326.
A two-sided 95 per cent confidence interval for a is obtained in the customary
way using X 2 with 15 d.f. as follows:
3.370/27.49 < a < 3.370/6.262,
0.123 < a < 0.538.
When we try the first method of obtaining an approximate confidence interval

for a by putting s 2 for a and using X 2 , we run into trouble. The method gives for
a 95 per cent interval with 4 d.f. for Xa,
(1.527/11.14) s 2 < 4o < (1.527/0.4844) s 2
0.022 < a < 0.732,
and even though Ho : o = 0 was emphatically rejected, the lower limit turns
out to be negative.
The second procedure, which was the one we recommended, has for 95 per
cent confidence Fr, = F0025 (4, 15) = 3.80 and FL = 1 /F0 025(15, 4) = 0.1155.
Then the desired confidence interval is
0.04<v<3.25.
3.11 The Randomized Complete Block Experiment

At the beginning of this chapter we considered an experiment with gasolines
and cars. There were five gasolines and we wanted four observations on the
octane number of each gasoline. The experimental design suggested was to
choose twenty cars (presumably at random from whatever automobile population was germane to the experiment), assign four cars at random to each
gasoline, and then to take one observation in each car. The model then suggested
can be rewritten
Yi=+
Ti
+fr,
where y ij is the observation on the ith gasoline in the jth of the cars assigned to
it, T, is the effect of the ith gasoline, and f f is the random error in the observations. The f i are uncorrelated random variables with zero means and variance
o. Let us eiamine these f, f terms a little more closely.
Our assumptions imply that if we were to take several cars and observe the
octane number of gasoline A in each of them, the observations would be
56
Experiments with a Single Factor
[Ch. 3
scattered about the " t r u e " value of the octane number of gasoline A in a
random manner with variance a2. The scatter in the observations comes from
two sources: first, different cars rate gasolines differently; second, even in the
same car, repeated observations on the same gasoline will differ because of what
might be called the "noise" in the system or the general uncooperativeness of
nature. We can rewrite the model to take account of these two sources of
variation by splitting ftj into two terms:
yu = I* + rt + cm + e^.
Here cm is a " c a r " component which corresponds to the^'th of the cars that
were assigned to the ith gasoline.
Suppose that cw is a random variable with mean zero and variance a2, while
etj as usual has mean zero and variance a2. We have a2 = a2 + a2. We assume
also that the population from which the cars were chosen is so large that for
sampling purposes it can be regarded as infinite. Then if we were to take an
observation on a gasoline in each of r cars and average them, the variance of the
mean of the observations would be (a2 + a2)jr. The situation is analogous to
that in which d dams are assigned to a sire and a single offspring is observed
from each litter.
What would happen if, instead of taking twenty cars, we were to take four cars
and to use each gasoline once on each car? The experiment might take a little
longer because each car would now be used for five gasolines, but on the other
hand there would be fewer cars to worry about. Are there more pressing arguments in favor of this new design ? Does it give data that are in some sense better ?
As before, we estimate /* by the grand mean G/20 and (/n + T4) by the average
yt- = 2 / W 4 (to be derived shortly). Neither /*. nor T, is estimable, but, as
before, we can estimate T( rf- by yt. yv.. In the original experimental design
we had V(yt.) = of/4, and V(yt. - yv.) = 2CT?-/4. In the new design, V(yt.) is
still aj/4 so that as far as estimating p. + T( is concerned there has been no
improvement in precision in the sense that there has been no reduction in the
variance. However, V(y(. - yr.) is reduced to 2CT2/4. This reduction becomes
apparent when we rewrite the model
yu = /* +
T,
+ Cj + eu,
because car j is now they'th car for each of the gasolines, and we no longer need
to write cm. Then
Ti - TV = 2 y - 2 y*i = 2 ta> ~ yi'd

= 2 &
+ T(
= 4(n -
T|.)
+ ci + e ~ V- -
TY
- Cj - eVj]
+ 2 ( - eVi) = 4(T, i
and Vft - tt.) = 8<r2/16 = a 2 /2.
T(.) +
2 ew ~ 2 *<'
i
The Randomized Complete Block Experiment 57
Sec. 3.11]
What has happened is that the car effects have canceled out and the precision
of the estimates of treatment comparisons has been thereby increased. The same
procedure is commonly used in agronomy. Instead of dividing a field into N = rt
plots and. assigning the treatments to the plots at random, the field is first split
up into blocks that are as homogeneous as possible, and each variety is sown on
one plot in each block. Thus, each variety gets a plot on the good land, and each
gets a plot on the bad land, and so on. The term block has been retained in
standard use in all areas of application. In our experiment, each car constituted
a block. This kind of experiment is called a randomized complete block experiment or design: complete because each block contains each treatment, and
randomized because the plots are assigned to the treatments at random in each
block.
We now derive the analysis. The model for t treatments, each appearing
exactly once in each of b blocks, is
yt =+T{ +Pf+e,
;
where y { 3 is the observation on the ith treatment in the jth block. We shall
assume that the block effects, fl, are random normal variables, independently
distributed with zero means and variance ab. It does not matter, so far as our
derivation of estimates is concerned, whether the /3 are fixed or random. The
distinction is of interest when we consider applying the results of the analysis.
If the blocks can be considered as chosen at random from some population, it is
not unreasonable to suppose that the conclusions about treatment contrasts will
hold throughout the population of blocks bf which those used are regarded as
"typical. However, if the cars chosen come from a small group or if they are to
be regarded as fixed effects, there are problems when we wish to apply our results
elsewhere. Of what population are the cars used typical? Is it the population of
low priced 1969 cars? Or of all 1969 cars? Or the population of cars driven to
work by engineers in the research laboratory during the third week in July?
Just as in the case of the random samples from random batches, we have two
choices, under normality, when the blocks are random effects. We can solve the
maximum likelihood equations for the estimates of the parameters, noting that
the observations are no longer uncorrelated because cov (yij, yir.) = ab (j # j')
That approach will be taken up in the exercises at the end of this chapter.
The second approach is to act as if the / are fixed effects to obtain least squares
estimates. We shall follow the Jatter course. (Indeed, if we did not know the
distribution of the flf it would be the only course open to us.)
In vector notation the model can be written
Y = XO + e,
where 0' = (, r', (3'). The design matrix X has 1 + t + b columns. The first
column is 1 N , a column of ones, corresponding to the mean; the next t columns
are associated with treatments, and the othrs with blocks. The row for y j has
[Ch. 3
one in the first column, the (i + l)th column and the (1 + t + j)th column;
the other entries are zero. Columns 2 through t + 1, and columns t + 2 through
1 + t + b both add up to 1 N , and so X is of rank t + b 1.
The oormal equations are
G=N.+b f,+t>f,
4
Ti =b11+bfj +^p1 ,
B =t+^ +
{
where Bf = li yj is the sum of the observations in the jth block. Imposing the
side conditions 1 r, = 0 and 1 f f = 0, we immediately have
=G/N,
Tt=TilbA,
Pr =Bi /tA.
We note that the estimates T, are the same as they would have been without the
/3 in the model, and the nf are the same as they would have been without th Ti .
We therefore say that treatments are orthogonal to blocks. This property is a
consequence of the balance in the experiment inasmuch as every treatment
appears exactly once in every block. We shall have more to say about this
when we discuss two-factor experiments. The sum of squares for regression
SR = C + + 1 PA may be split into three forms, C, St , which we have
seen before, and a similar sum of squares S b = B C, the sum of squares
for blocks. The analysis of variante table becomes:
TABLE 3.3
Source
Total
Treatments
Blocks
Error
SS.
Y'Y C
EMS
d.f.
N-1
Tz/b C
t 1
If B, /t C
b 1
o2 + b 1, T;/(t 1)
02 + tob
Y'Y C St Sb (t 1)(b 1) a 2
The result of blocking has thus been to remove from the error sum of squares,
Se , a portion with b 1 d.f.; hopefully the amount of residual removed has
been appreciable, at least enough to compensate for reduction of d.f. for error.
With fewer d.f., the critical values of F and t will be increased.

We return now to the gasoline data, assuming that the effects are fixed and
that the experiment was a randomized complete block experiment. There were
four blocks. The jth block consists of the jth observation on each treatment;
Sec. 3.13]
Missing Plots in a Randomized Complete Block Design 59
thus block 1 consisted of observations 91.7 on A, 91.7 on B, 92.4 on C, 91.8 on D

and 93.1 on E. The block totals (after 90.0 has been subtracted from each observation) are 10.7, 9.4, 7.8, 6.3.
The analysis of variance table is as follows:
TABLE 3.4
Source
SS.
d.f.
M.S.
Total
Treatments
Blocks
Error
9.478
6.108
60.676 C = 2.194
1.176
19
4
3
12
1.527
0.731
0.098
15.6
7.46
The new estimate of the error variance is 5 2 = 0.098 and s = 0.313.

Using the t test, the new LSD is 2.179V0.049 = 0.482, and we conclude that
E> A,B,C,DandD > A,B.
For Tukey's test, go.05 with 12 d.f. is 4.51. The new LSD is 0.71, and we
conclude that E> A, B, C, D and D > A.
The S test for 0 still accepts 0 = 0, but barely. SZ = 4F(4, 12) = 13.04 and
Ss = 1.13, so that the confidence interval becomes
2.23<0<0.03.
3.13 Missing Plots in a Randomized Complete Block Design

The analysis of the randomized complete block design depends upon having
each treatment appear once in each block. What happens if something goes
wrong and one of the plots is missing? Perhaps somebody makes a gross mistake
and plants the wrong variety, or harvests it too early. In some industrial situations, it is possible to obtain a repeat observation fairly easily. If one of the cars
breaks down before all the gasolines have been tested, it might be feasible to
wait until the car is repaired and then finish the experiment, or to find another
car and, having tested all the gasolines on it, ignore the data on the old car.
However, the agronomist has to wait another growing season to obtain a
repeat observation, and we can easily think of industrial situations in which
a repeat observation would be impossible. An industrial example which is not
uncommon occurs when an experiment is conducted in a pilot plant which is
available to the project only for a limited period. If the observations require
rather lengthy chemical analyses, it is quite likely that by the time the scientist
discovers that an observation has been spoiled or lost, the test equipment will
already have been dismantled and the pilot plant will be in use on some other
project.
The statistician may calculate from the data what the missing observation
would have been if only he had it; then he will have a complete set of data and
[Ch. 3
he can perform the usual analysis (with some modification). There is the alternative of analyzing the data as an incomplete block experiment with unequal
block Bizes and unequal numbers of replications of the treatments. That, as will
be seen in a subsequent chapter, calls for messy calculations, and is considerably
less convenient than the procedure suggested here.
We discussed the general problem of missing plot formulae in the previous
chapter. Suppose that y jj is missing. Let T' denote the sum of the b 1 actual
observations on the ith treatment, B' the sum of the t 1 observations on the
jth block, and G' the sum of all N 1 observations. We take as our estimate of
the missing plot or observation the value u which minimizes S e when the
ordinary analysis of variance procedure is carried out on the N points (the
N 1 actual points and u). We showed in the last chapter that this was equivalent to choosing u so as to give an exact fit. In that case
G'+u T'+u G'+u B' +u G'+u

u=+T,+fl f = N + b
N + t N
and, solving,
_ tT' +bB' G
u (b 1)(t 1)
We are mainly concerned with comparisons y,. y,... The missing plot value
is a linear combination of the observations, and so Ti is correlated with the
other treatment totals. It can be shown that, when there is a missing observation
in the ith treatment,
V(yi .yt- .)=b-2V(T,T;.)=a2lb+b(b 1)(t 1)]

If there are two missing plots, we have two choices. Either we can repeat the
procedure just used with two simultaneous equations to be solved, or we can
adopt an iterative procedure of guessing one value and fitting the other, then
going back and fitting the first value, and so on.
The formula that we have derived was first obtained by Allan and Wishart
(1930). The F test will be slightly biased, and the reader is referred to the paper
by Yates (1933) for a discussion of this point.
3.14 The Analysis of Covariance

We have just considered the use of randomized complete blocks to improve
the precision of treatment comparisons. Another method that may be used
sometimes is the analysis of covariance. This procedure involves adjusting the
observed responses for the linear effect of another factor (a concomitant
variable). In a chemical experiment, where we have for the treatments different
temperatures, we might wish to adjust the yields for the percentage of some
impurity in the feed stock at each run. In an experiment to investigate the effect
The Analysis of Covariance 61
Sec. 3.14]
on the growth of young pigs of some change in the standard diet, we may take
weight at six months as the response, y, and adjust it for the initial weight, x.
The adjustment is linear, and we consider here only the case of a single concomitant variable. The model that we shall take is
f: Yti = F^ + Ti + 1xt7 +
eti,
where x 15 is the measurement made on the concomitant variable on the ijth run,
or for the ijth datum point.
We introduce the following notation
E . =
(x{i xt ) 2 ;
t
E., =
E,
T, =
. =
vv =
(x,i x..) 2 ;
S, =
(Yu yt.) 2 ;
Txv = G
(x,. x..) 2 ;
(xt1 x.)(Yt, Yt);

t
(xt . x..)(Yt.
(yt. y..) 2 ;
r
S.. =
(x1, x..)(Yts Y..);
S = G > (Ytt y) 2 .
{ r
where S = T + E. We do not assume that there are necessarily the same number
of observations at each level of the factor being considered; x t . is the average of
the x measurements for the ith treatment, and x.. is the average of all N
measurements of the concomitant variable.
The most convenient computational procedure for calculating these quantities
is to carry out a one-way analysis of variance for x and for y and a parallel
calculation for xy. We then obtain, writing X.. for Nx.. and Y.. for Ny.., the
following table:
TABLE 3.5
ANALYSIS OF COVARIANCE
Source
d.f.
xy
Raw total
Mean
Treatments
Error
1, J i 4
l s I x,5Yi
l i l t Y%
1
t1
X^/N
T.
X..Y../N
T.,
Y.^/N
T.
N t
Exx
Exy
Ey
T+E
N1
S...=T...+Exx.
Sz7
Sw
The entries in the x and y columns cannot be negative; entries in the xy

column may, however, be negative.
No te:
[Ch. 3
We can now proceed in two ways, and it is constructive to consider both.

When we introduce blocks, we shall set up and solve the normal equations. At
this time we shall use the other approach.
Suppose that we have no treatment effects. Our model would be
Oo: yit=ii+Pxit+eit.
Under this model P is estimated by = S./Sxz , and the sum of squares for
error is S = S, (S) 2 /S.
Suppose now that we have treatment effects and that the model is 0 1 . This
model implies that there is a family of parallel lines, one for each treatment.
Some of the lines may coincide; S2 o corresponds to the case in which all of
the lines coincide. The data for the ith treatment alone give an estimate bi of ,
where
b`
s
I f (xif xi.)(Yif Yi.) Q
and
V(bi) =
=
1i (xi5
xi.)2
(xi> xi .) 2
The several estimates, bi , are clearly uncorrelated because they involve different
data. The minimum variance unbiased combined estimate from them is P =
l i wibi , where the weight, wi , given to any bi is inversely proportional to V(bs ),
and wi = 1. This gives w, = I i (x 5 xt.) 2 IE. and
p = Ex /E
x.
The ith line is thus yi , = y,. + P(xiJ xi .).

The sum of the squares of the deviations about the several lines is
Se =
(Yi yi P(xii xi)) 2
E (E.v) 2 /E
We can now write down the analysis of variance table. The sum of squares
for treatments is the reduction in error made by fitting individual Tines rather
than a common line, i.e., SB Se
.
TABLE 3.6
Source
Total
d.f.
Fitting a single line
N 1
1
Treatments
Error
t 1
N t 1
SS
S.
(S.Y)2 /Sxx
(Sy, (SS) 2/Sxz) (E
Ey, ( E.xy) 2 /Ezx
(Exy) 2 /Exx)
The error variance, o2 , is estimated by S 2 = (N t 1) - '(E22 (Ex )2 /Exx)
Sec. 3.15] Analysis of Covariance for Randomized Complete Block Design 63

If the concomitant variable had not been added to the model, the error sum
of squares would have been E^ v . The reduction resulting from the inclusion of
the concomitant variable is, therefore, (E)2IExx, and the hypothesis Ha : S = 0
is tested by the statistic (Ex)2/(s2Exx), which has, under the null hypothesis, the
central F distribution with one and N t 1 degrees of freedom.
It should be noted that if the same x array is used for each of the treatments,
we are really back to a complete block design with the blocks corresponding to
the values of x.
3.15 Analysis of Covariance for the Randomized Complete

Block Design
We confine our discussion to the case in which there is a single observation on
each treatment in each block; y t , denotes the observation on the ith treatment in
the jth block, and x, f is the corresponding value for the concomitant variable.
The complete model is
ytf =g+ r +jxtf+yj+eti,
where y, is the jth block effect and i = 1, ..., t; j = 1, ..., r. We shall use
capital letters X, Y to denote totals. Thus Y{ . = y, and so on. The normai
equations are, with the side conditions
= 0, I f yf = 0,
Y..= tri +px..,

Yt . = r + rT t + Xt .,
Y., = t + ty, + X.,,
xtiyu = X.. + ^ Xt T t + ^ XrYr +

11
Eliminating , T,, and y, from the last equation, we have

xuyti t 1 > X. 1 Y. 5 r'
t
_ {^ ^ xti r
X. YY . + N
-1
X.. Y..
t
1
X.1 t 1
X a + N 1 X }.
The coefficient of P is the sum of squares for error obtained when an analysis of
variance, including blocks, is carried out on the x, 5 . The left side of the equation
is the corresponding sum obtained from the xy terms. We write Ex x , Ezj for
these quantities and define similarly E^, Sx x , Sx v , S. We then obtain
xv xx
[Ch. 3
The other parameters are estimated by

T, = {t Y{ . Y.. P(tX{ . X..)}IN,
= (Y.. $X..)/N,
y1 = {r Y., Y.. P(rX. j X..)}IN.
The sum of squares for regression (including the mean) is

YY .Ti +
Y.. +
Y.,
whence Se = Ey y PEx, = E^ (Exy)ZI Eax

If P = 0, we have a straightforward randomized complete block situation with
Se = S ' . (Sxy) Z I Sxx
The "recipe" for the calculations is thus to proceed in the same way as for the
unblocked case, using S, ,, etc., which are calculated, in the same manner used
before by analysis of variance procedures for x, y and xy, including the terms
for blocks in the computations.
Exercises
1. Six octane number determinations were made on each of five gasolines. The data
are given as follows. The columns correspond to six blocks. Analyze the data as a
randomized complete block experiment. Which gasolines differ?
Blocks
I
A
B
C
D
E
57
65
59
55
54
II III IV V VI
61
65
60
54
50
62
67
65
53
57
59
64
60
51
55
58
65
60
50
54
59
63
59
53
57
2. This data comes from a randomized complete block experiment with three blocks
and twelve treatments. The responses are the yields, in ounces, of cured tobacco
leaves. Compare the control to the average of the other treatments. [Petersen
(1952)].
Block
Treatments
1
2 3 4 5 6 7
8 9 10 11 12 (Control)
76 82 76 70 76 70 82 88 81 74 67 79
II
III
Means
70 70 73 74 73 83 74 65 67 67 67 78
80 73 77 62 86 84 80 80 81 76 79 63
75 75 75 69 78 79 79 78 76 72 71 73
Exercises 65
3. An experiment is to be designed to compare t new treatments with a control
treatment. A one-way layout is to be used; there are to be ro observations on the
control and r observations on each of the new treatments. Show that for any given
value of N the variante of the comparisons between the control treatment and
the new treatments, V(T0 - T{), is minimized if r o = rVt.
4. Let x take the values 1, 2, 3,..., n, and let y = x2 . Show that the correlation
coefficient of x and y is given by
15(n + 1) 2
r2=
16n2 + 30n + 11'
which is a monotone decreasing function of n with limit 15/16.

5. Two observations are made at x = 0 and one each at x = 1, 2, 3, and 4. Find a
set of orthogonal linear, quadratic polynomials.
6. If x takes the values 1, 2,..., n once each, show that the polynomials u 1 = x
z, u 2 = 12u1 (na 1) are orthogonal.
7. Find a pair of orthogonal linear and quadratic polynomials when x takes the
values 0, 1, 2, 4, 8.
8. Find the expected value of St for a randomized complete block design when one
of the plots is missing and its value has been calculated by the formula.
9. Derive the formula for V(yh yt ) if one of the observations on the hth treatment
has been calculated by the missing plot formula.
10. Find the expected value of Se in the random effects case when there are r, observations on the ith variety and the r, are not all equal.
11. In the random effects model for the one-way layout with r observations on each
variety, use the method of maximum liklihood under the normality assumption to
obtain estimates of o and oa,, and to test the hypothesis H0 : Q = 0.
12. Assume that in a randomized complete block experiment the errors e i and the
block effects S, are normal random variables. Use the method of maximum
likelihood to derive estimates of r and a.
13. Prove that S t and Se are independent in the random effects model under normality
in the balanced case. Is it true even when there are unequal numbers of observations on the varieties?
14. Suppose that in a randomized complete block design the observation on the first
treatment in the first block and the observation on the second treatment in the
second block are missing. Obtain formulae for calculating the two missing
values.
15. Repeat problem 14 for the case in which the two missing plots are in the same
block.
16. Scheff (1959, p. 70) points out that the F test will reject Ho if, and only if,
there is at least one estimable y which has a significant estimate by his S test.
Is it true that if, in the fixed effects case, the F ratio is significant, then there is at
least one contrast that will be declared significantly different from zero by the
t test? Is the converse true? If not, give an example.
CHAPTER
Experiments
with
Two Factors
In the previous chapter we considered as an example of a randomized block

experiment an experiment in which t treatments (gasolines) were tested in r cars.
The cars were chosen at random from some appropriate population and they
constituted the blocks of the design. The purpose of the blocking was to improve
the precision of the comparisons between gasolines. The differences between
cars were eliminated in estimating the gasoline contrasts. In some presentations
the analysis of variance tables contain under the heading, "Expected mean
square for the sum of squares between blocks," the single word discard. The
variation between blocks is considered as something to be "backed out" of the
experimental analysis.
Suppose, however, that instead of having r nondescript vehicles, we had been
at pains to choose a Chevrolet, a Plymouth, a Ford, and so on, with the result
that cars could be considered as a second factor occurring at r levels. The two
factors are said to be crossed. Each factor appears at least once with each level
of the other factor. Indeed, each gasoline sees each car exactly once, and each
car sees each gasoline exactly once. How, if at all, does this experiment differ
from the randomized block experiment of the previous chapter? We consider
first the simple experiment in which each gasoline is tested only once in each car.
This experimental setup is called the two-way layout (or two-way crossed
classification) with one observation per cell.
4.1 The Two-Way Layout with One Observation per Cell

We shall find it convenient now to think about a factorial experiment involving
two factors, A and B, with a and b levels respectively. In our previous example,
the factors were gasolines and cars. The factors can be qualitative or quantitative.
ri
The Two-Way Layout with One Observation per Cell 67
Sec. 4.1]
The levels of the car factor could be Plymouth, Ford, etc.; the levels of the
gasoline factor might be one ml/gal, 2 ml/gal, etc., of some additive like lead.
Our model is the simple additive model Y = XO + e, or
yi=+ at+ilj +et1,

where y{ 1 is the observation made on (or the response at) the ith level of A and
the jth level of B, and et , is the random error with the usual assumptions. The
other terms in the model are unknown constante; represents a grand mean, a, is
the main effect of the i th level of A, and f is the main effect of the j th level of B.
The design matrix X has N = ab rows and 1 + a + b columns. The row
corresponding to the observation y ij has unit elements in the first, the (i + 1)th
and the (1 + a + j)th columns, and zeros elsewhere. The first column is 1 N
The sum of the next a columns is also 1 N , and the sum of the last b columns is
also 1 N ; it follows that r(X) < a + b 1. This implies that we have two
parameters too many. We could get rid of one parameter by putting = 0,
which is the same as leaving out of the model, but we would still be faced with a
singular matrix X'X. The most common procedure is to leave in the model and
to impose the two side conditions 1'a = 0 and 1'(3 = 0, where a and p are the
veetors of the main effects a{ and Pi . The parameters , a,, and / are not estimable. The expected values of observations, contrasts in the parameters a,, and
contrasts in the f, are estimable.
As far as the derivation of the rums of squares is concerned, there is no change
from the randomized complete block design. We write T i . = y 1 = byt . and
T 1 = l ; yi; = ay. f . Then % = y, , = y,. y and P1 = y. ; y. Writing a. for
i, aja and P. for I; fj /b, we have
E(i ) = + ai +
P. a. P. = ai a..
This brings out the concept of the main effect of the ith level of A as the difference between the average response at the i th level of A and the average response
at all levels of A.
The comparison a a ; . is estimated by y i . y ; ... We point out again that
just as was the case in the randomized block experiment, the B effects are
canceled out in the expectation of the contrast because of the balance, or
orthogonality, that was built into the experiment by having each level of A
appear exactly once with each level of B. Similarly, the comparison fj /1 . is
estimated by y. j y. j -. Table 4.1 follows:
TABLE 4.1
Source
SS
d.f.
Total
Between levels of A
Between levels of B
Residual
1, J f y G 2 /N
Tlb G 2 /N
J J T.fla G 2 /N
1, I j y, 7-, T,?/b J t T.J/a + G 2 /N
ab 1
a 1
b 1
(a 1)(b 1)
68 Experiments with Two Factors
[Ch. 4
Tests of hypotheses about the a and about the /3 are carried out just as in a
randomized complete block design.
{
EXAMPLE. In their studies on the habits of Zebu cattle, Rollison et al.

observed a group of cattle and recorded every minute which of the cattle were
grazing. The same group of five observcrs was used throughout the work carried
out at Entebbe, Uganda. The authors (1956) reported upon an experiment carried
out to compare these observers. They watched a group of ten cattle in the same
paddock for eighty-eight minutes one afternoon. The numbers recorded are the
numbers of minutes in which observer i reported that he saw animal j grazing.
Animals
Observers
A
B
C
D
E
Total
1
2
3
34 76 75
33 76 72
35 78 76
34 77 71
33 77 70
169 384 364
5
7
4
6
31 61 82 82
29 60 82 84
30 65 86 88
29 60 78 83
27 59 81 82
146 305 409 419
8
67
67
66
67
67
334
9 10 Total
72 38
618
72 36
611
76 37
637
72 37
608
70 33
599
362 181 3073
TABLE 4.2
Source
d.f.
MS
Animals
Observers
Residual
Total
9
4
36
49
2149.42
20.33
2.06
4.2 Orthogonality with More Than One Observation per Cell

Suppose that we relax the requirement that there be exactly one observation at
each level of A and at each level of B, and allow several observations. The model
remains the same, except that we could write yU, to denote the mth observation
at the ith level of A and the jth level of B, i.e., in the (ij)th cell. If there are an
equal number of observations in each cell, the analysis will be a trivial extension
of that of the previous section. If the cell numbers are not all the same, the
solution of the normai equations is complicated by the fact that we cannot find
two side conditions that will separate the unknowns in a simple fashion.
We are not going to discuss the solution of the equations at this time. We have
already spoken in Chapter 2 about methods of finding pseudo-inverses, and the
matter will come up again in more detail in our discussion of incomplete block
Orthogonality with More Than One Observation per Cell 69
Sec. 4.2]
designs. We address ourselves in this section to the concept of orthogonality.

What do we mean when we say that the two factors A and B are orthogonal, and
what are necessary and sufficient conditions for orthogonality? Two definitions
of orthogonality of factors (or, in block designs, orthogonality of blocks and
treatments) commonly occur in conversation among practitioners of analysis of
variance. We shall show that they are equivalent.
Definition 1. The factors A and B are said to be orthogonal if, and only if, any
comparison between averages of observations at two levels of A is orthogonal
(when considered as a linear combination of the observations) to any comparison
between the averages of observations at two levels of B.
Definition 2. A and B are orthogonal if, and only if, the estimate of any contrast
in the parameters a i is the same whether or not the parameters /3 are included in
the model, and vice versa.
We shall consider in particular two levels, h and i, of A and two levels, j and k,
of B. The corresponding totals of the observations at these levels are T h ., T,.,
T. 1 , T. k ; the sum of the observations at the hth level of A and the jth level of B
is Thf . Let there be n,^ 5 observations in the (hj)th cell with marginal totals n,^. and
n. 5 , and N observations altogether.
Suppose that A and B are orthogonal by Definition 1. Then yh . yi . is orthogonal to y., y. k . Consider the equivalent contrasts n { .Th . n h .T{ . and
n. k T. j n. I T. k . The only observations that appear in both contrasts are those in
the cell totals Thj , Thk, T, j , Tik , and we rewrite the contrasts
n { .Thj + n i .Thk nh .Tlj nh.Tik,
and so on, and

n.kThi n. jThk + n.kTl j n.lT{k
and so on.
These are orthogonal as contrasts in the observations if, and only if,
n { n kn hf n i .n. jn hk n h n. knt1 + n h .n. 1nik = 0,
or,
n{.(n.knhj n.inhk) = n h .(n. kn ij n.jnik).
Such equations hold for all pairs h, i and j, k. Summing over all values of k,
including k = j, gives
nt .(Nnh1 n. Jn h .) = n h .(Nnti n.jni.);
[Ch. 4
then, summing over all values of h,

ni .(Nn. j - n.,N) = 0 = (Nn i f
so that Nn, f n. 5n4 . = 0, or n, f = n { .n. i/N for each i and each j.
The converse, that this condition implies the orthogonality of the contraets, is
easily shown. We call this condition the condition of proportional frequencies.
Suppose now that A and B are orthogonal by Definition 2. We show that the
condition of proportional frequencies is again a necessary and sufficient condition. There are now two models under consideration:
E(Y) = 1 + X1 a + X2 13,
(4.1)
E(Y) = 1 + X l a,
(4.2)
where X = (1, X 1 , X 2), and a and P are the vectors of main effects.
Under model (4.2) we have a one-way layout and the least squares estimate of
ah ai is y,,. y,.. Consider the estimable function 0 = n h .n { .(a h a,). Its
least squares estimate is = n,.Th . n h .T,.. Since the least squares estimates are
the unique minimum variance unbiased estimates, and since any linear combination of the observations which is an unbiased estimate of 0 under model
(4.1) will also be unbiased under model (4.2), it follows that 1 is the least squares
estimate of 0 under (4.1) if, and only if, it is unbiased. Under model (4.1)
E(^) = nh.ni.(ah aj) +
(n,.nhj nh.ntr)Nr,
and is unbiased if, and only if, n{.nh f n h .n, f = 0 for each j. Summing over all
values of h leads to the condition of proportional frequencies.
Conversely, if the condition of proportional frequencies is satisfied and the
side conditions 1, n,.a{ = 0 and > n., 5 = 0 are used in the standard least
squares procedure, it is easily seen that the least squares estimates of and a i are
y and y,. y under both models.
Under model (4.2) we may subdivide the sum of squares for regression into
two components S() and S(aI), where S() = G 2 IN. Under model (4.1)
we can have three components S(), and S(aI), as before, and the extra portion
S( , a). If A and B are orthogonal, S(I ) = S(Sl , a).
4.3 The Two-Way Layout with Interaction

An important and restrictive implication of the simple additive model used in
the previous sections for the two-factor experiment is that the expected value of
the difference between the responses at two levels of A is the same at each level
of B. In terms of the gasoline experiment, this amounts to saying that apart from
the random error if the Plymouth rates gasoline 1 to have an octane number
Sec. 4.3]
The Two-Way Layout with Interaction 71
0.5 ON higher than gasoline 2, so will every other car. Similarly, if the Plymouth
gives a rating to gasoline 1 that is one ON higher than the rating given by the
Ford, then the Plymouth will rate each gasoline (apart from the random error)
at one ON higher than the Ford rates them. This simple additive model may not
be adequate. The failure of the differences between the responses at the different
levels of A to remain constant over the different levels of B is attributed to
interaction between the two factors. To take this into account, we add an extra
term to the model.
We now consider the situation in which there are n observations at each
combination of levels of A and B, i.e., n observations in each cell. The new
model is
Yuk = + a t + P1 + (aP) {f + eilk , 1 < i < a, 1 < j < b, 1 < k < n,
where yifk is the kth observation at the ith level of A and the jth level of B and
ei1k is the error term; , ai , P1 are as we defined them before, and (as) ti is the
interaction term. There are ab such interaction terms. The design matrix X has
abn rows and ab + a + b + 1 columns. The rank of X is ab.
We can approach the subdivision of the total sum of squares into component
sums of squares in two equivalent ways. One approach is to argue that basically
we have a one-way layout with n observations on each of ab treatments, a
treatment being in this context a combination of a level of A and a level of B.
Analyzing the experiment in this way gives a sum of squares for treatments
St = l t ^ f T1 ./n C where T11. = I k Y41k and C = G 2 /N. (We shall now
write Ti .., T.,. for the sums l f I k Ytik, ^t Ik Yuk) Differences between levels of
A account for an amount S A ; differences between the levels of B account for
another SB ; the remainder St SA SB must result from interaction.
The other approach is to consider three regression models:
Ytlk = + at + jJ + (af1)u + eifk (4.3)
(If we impose the side conditions a, = 0 for each i and P, = 0 for each j, we
obtain the oneway layout, and S R = St + C):
Ytrk = FJ + a s + S ! + e tik ;
Yuk = i& + a i
(4.4)
eijk.
(4.5)
The reduction in Se at each stage gives us the desired sum of squares, and, with
the appropriate divisor (aa), the conditions of Cochran's Theorem are satisfied
when normality is assumed so that the forms have independent x 2 distributions.
In solving the normal equations under model (4.3), we impose the side conditions = 0, I f Si = 0, Ii (af1) 1J = 0 for each i, and 1t (af1), f = 0 for each j.
This actually amounts to only a + b + 1 conditions because the conditions
imposed upon the interaction terms are not independent; each of the two sets
[Ch. 4
implies li I j (aa) i5 = 0. With these side conditions, the estimates of the parameters are =9,& = yt .. 9 where y,.. = T{ ../bn, ,^, =y.. -7, and ( 1 =
yi1. y t .. y.,. + y where yif. = T,1 ./n.
TABLE 4.3
EMS
d.f.
SS
Source
Total
li G1 Gk Y
4 1k C
abn 1
l i T,?./bn C = SA
a1
a2 + bna 2
J, T../an C = SB
1, J, T!./n C S A SB = SAa
b1
(a 1)(b 1)
a2 + anaB
aa + na8
lt I1 Ik Ys1k lt St T 5./n
ab(n 1)
Interaction
Error
aa
The sum of squares for interaction may be derived in two ways. It is the
difference between the sum of squares for regression with the model of a one-way
layout with ab treatments and the sums of squares attributable to A and B, i.e.,
n -^^, ^1 C S A SB . Alternatively, the equation SR = IPX!'IY derived in
Chapter 2 may be rewritten as SR = l i Pt (l1 x{1y,) where 1 xy1 is the single
term on the left side of the normal equation corresponding to P { . With all the x, i
being zero or one, X'Y is the vector of totals, and so we have
SA =
Ti .. tl SB = L T.5., SAS =
Tt1( ^)t1
The expression for SAB can also be written

SAB =n
(yi1.y1..y.1.+9)2=n
or else
SAB =n
( (yu.
an (y.1. 9) 2 ,
which expresses SAB in terms of the variante of the estimates of a{ over the
several levels of B.
The expectations of the mean squares require a word of explanation. Our,side
conditions simplify the computations. Thus we have
G = Np + bn a4 + an 1 P 1 + nl 1(afl)t1 + 1 E es>k,
1
and this simplifies to G = N +
1 f
ik
Lk euk so that E(G 2 /N) = Np 2
+ aa
Sec. 4.4]
The Subdivision of the Sum of Squares for Interaction 73
Similarly, yt .. y reduces from + at + fi. + (af) t + et .. a.

^. (as).. e..., where e t .. = aN -1 Ii Ik etik and e... = N -1 lt Ii Ik etik to
.
yt
Y=at+et..
e...;
thus, E(SA ) = bn l t E(y, 9) 2 = bn I t aa + (a 1)02 , and we shall, in an

obvious notation, write I t at as (a 1)4
Finally, we can proceed in the same way and obtain
E(SAB) = n: E(yu. y,.. y.j. + 9) 2
(a,6) i + (a 1)(b 1)a 2 .
We then write I t Ii (aa) ti as (a 1)(b 1)B.
4.4 The Subdivision of the Sum of Squares for Interaction

We discussed in the previous chapter the subdivision of the sum of squares for
treatments in the one-way layout. This process can be extended to the factortal
experiment, and the sums of squares for A, B, and AB interaction can all be
subdivided into single degrees of freedom. We can take ab 1 orthogonal
contrasts in the cell totals, of which (a 1) are contrasts in the A totals and
(b 1) are contrasts in the B totals. The squares of the remaining contrasts
with the appropriate divisors each correspond to one degree of freedom, and
their total is S B . We mention two uses of this procedure, one of which, Tukey's
one degree of freedom for interaction, will be discussed in the next section.
We have already seen that when A is a quantitative factor, preferably with its
levels equally spaced, SA can be divided into such components as linear, quadratic, and cubic. If both A and B are quantitative, this approach can be extended
to fit the model.

E(y) = Po +
3toxi
t=^
a b
boix +
i1
Ptixix,
1=15=1
where x l and x 2 are coordinates for the levels of A and B respectively. Replacing
xi by the ith degree orthogonal polynomial for a levels, u t , and x by the corresponding polynomial vi , we obtain
E(y) = Yo + 1 Ytout + 1 Yoivi + 1 Ytiutvi
The estimates of the yti are orthogonal contrasts. The contrasts corresponding to
yto and yoi are the linear, quadratic, cubic, and so on contrasts for A and B and
[Ch. 4
give the subdivision of SA and SB into single degrees of freedom as before.

The cross-product contrasts, yjj (i > 0,] > 0), give a subdivision of the interaction sum of squares into single degrees of freedom. The contrasts corresponding to yll, Y122 Y21. and yaa are usually called the lin A x lin B, lin A x quad B,
quad A x lin B, quad A x quad B contrasts respectively. All (a 1)(b 1)
degrees of freedom for interaction can be accounted for in this way by associating
them with estimates of binomial coefficients. The physical interpretation of the
quad A x quad B and higher order interactions is by no means straightforward. A large higher order interaction can sometimes be the result of bizarre
behavior in a single ceil.
Let T be the matrix of cell totals and let LA , LB , Q,, QB be the vectors of the
coefficients for the linear and quadratic contrasts for A and B. The lin A x lin B,
lin A x quad B, quad A x lin B, and quad A x quad B contrasts are LATLB ,
L ITQ B , QTL B , and QTQ B . The divisors for the sums of squares are
nLLALLB, nLALAQBQB, nQ.sQALBLB, and nQQAQQB.
Yates (1937, p. 36) gives the data for an experiment carried out at Rothamsted
with two factors, varieties of oats (A) and quantity of manure (B). There were
three varieties of oats; four levels of manure were used, 0, 0.2, 0.4, 0.6 cwt per
acre. The figures given are the celi totals for six plots per cell; the recorded
responses are yields measured in quarter pounds per plot. Although the oats
factor (A) was not quantitative, it is nevertheless instructive to compute the
sums of squares for the several components of interaction.
B1
B2
Al
Az
A3
429
480
520
538
591
651
Total
1429
1780
B3
B4
665
711
688
749
703
761
205 6 2221
Total
2343
2508
2635
7486
It is important to remember that these are cell totals, and the factor n = 6 must
appear in the denominator of each sum of squares; C, for example, is 7486 1 /72 =
778336.06.
The eleven degrees of freedom between cells are partitioned in the ordinary
analysis of variance table into
TABLE 4.4
ANALYSIS OF VAIUANCE
Source
SS
d.f.
Total
22128.61
1786.36
20020.50
321.75
11
2
3
6
A
B
AxB
Interaction with One Observation per Cell 75
Sec. 4.5]
To obtain the interaction contrasts, we first obtain the B contrasts at each

level of A by using L = (-3, 1, 1, 3), Q = (1, 1, 1, 1), C =
(1, 3, 3, 1). Then we multiply by L = (-1,0, 1), Q = (-1, 2, 1). Thus
lin A x lin B = 775 -973 = 198, and quad A x quad B = 63 2(50) +
73 = 36. The B contrasts are
Al
Aa
A3
lin A
quad A
B divisors
'
lin B
quad B
cub B
973
904
63
50
73
10
36
4
99
22
85
184
30
20
775
198
60
20
A divisor
2
6
The individual sums of squares are

lin A x lin B
lin A x quad B
lin A x cub B
quad A x lin B
quad A x quad B
quad A x cub B
(-198) 2 /(2 x 20 x 6) = 163.35

(-10) 2 /(2 x 4 x 6) = 2.08
(-184) 2 /(2 x 20 x 6) = 141.07
(60) 2 /(6 x 20 x 6) = 5.00
(36) 2 /(6 x 4 x 6) = 9.00
(30) 2 /(6 x 20 x 6) = 1.25
321.75
The value of the lin A x lin B term is large, and we note that the lin B
contrast decreases as we change levels of A: 973, 904, 775. (If we had interchanged varieties 2 and 3, this inconsistency of the lin B contrast would have
been manifested as a large quad A x lin B interaction; however, A is not really
a quantitative factor.) It would appear that the increase in the response caused by
using extra manure is greater for A l than for A 3
-
4.5 Interaction with One Observation per Cell

Our earlier discussion of the two-way layout with one observation per cell
came before we introduced the idea of interaction. If there is interaction present,
the sum of squares called Se is actually S AB with n = 1. We have, therefore, two
problems:
(i) Our estimate of the error variance is biased inasmuch as E(M.)
a2 + B;
(ii) We have lost our F test for the hypothesis a = 0.
[Ch. 4
There is not much that we can do about problem (i) except to assume that
vAB = 0 and to comfort ourselves with the thought that if our assumption were
incorrect we should be erring on the conservative side because s 2 would tend to
be larger than a2 . Thus, it would take a larger value of 1 a? or of fl than
otherwise to provide a significant F ratio.
In regard to problem (ii), Tukey (1949) has suggested a method of obtaining a
sum of squares with one degree of freedom to test the hypothesis a = 0.
He takes the sum of squares for one of the interaction contrasts, a'Tb where
a' = (... , y .. y, ...) and b' _ (... , y. j . y, ... ). He gives an example in
his paper in which he replaces the response y ijk in a layout where c S is small
by zirk = He thereby obtains a situation in which there is a large interaction
sum of squares, much of which is absorbed in this single degree of freedom.
In practical applications we can more easily compute Tukey's statistic as
{
G(SA + SE + C)}2 /abSASB .
T* _
The distribution of T* is discussed at lome length by Graybill (1961).

The model that we considered at the beginning of the chapter had parameters
a, and $, which were assumed to be unknown constants corresponding to fixed
effects. We discussed in the previous chapter the situation in which the levels of
the single factor were random. We proceed now to consider the two-way layout
with both factors random. The procedure that we shall adopt is to take the same
breakdown into sums of squares that we used in the fixed effects case, and then
to investigate the expectations of the mean squares under the changed model.
The new model is
ytfk =
+ at + bf + (ab)i1 +
where y, ik , , and e, ik are defined as they were for the fixed effects case; ai , bj , and
(ab), f are uncorrelated random variables (as usual, we require normality, and,
hence, independence, when we test hypotheses) with zero means and variances
4, o, and 4B respectively.
We write
a. = a
l at ,

b. =
1i b lb ,
f
(ab). ; = a -1 (ab)ii, (ab).. = a'

c
e{ f . = n
-1
1k e,rk,
e,.. =
b '
ei f .,
(ab),.
= b'
1
(ab)ii,
(ab) t . = b' > (ab). i ,

f
e... =
N1
1 t 1 k
e.
The Mixed Model 77
Sec. 4.7]
Then we also have yti . = T 1 ./n, yi .. _ (bn) - 'Ti .., y. f . = ( an) - 'T. 1 . and y = GEN.
In computing E(SA), we again note that SA = nb (y f .. y) 2 ;
y{ ..y=
+ a,+b.+(ab),.+e,..a.b.(ab)..e...
= (a i a.) + ((ab). (ab)..) + (e i .. e...).

When we square these expressions and take the expected values, the crossproduct terms drop out because the random variables concerned are uncorrelated, and we have
E(SA) = (a 1)bn{V(a 1 ) + V(ab) 1 . + V(et ..)}

= (a 1)bn{a + b -l aAB + b - 1 n -1 a2 }
= (a 1)(6 2
+ natB + bna).
Similarly, we obtain
E(SB) = (b 1)(a 2 + naA2B + ana),

E(SAB ) = (a 1)(b 1)(a 2 + noB)
This points out the crucial difference in hypothesis testing with the two models.
In the fixed effects case, the hypothesis Ho : a l = = = 0 was tested by
the ratio F = M4 /Me where MA = SA /(a 1). In the random effects case, the
hypothesis H0 : oA B = 0 is tested by the ratio .F = MAB I M,. If this hypothesis is
rejected and the experimenter decides to act as if aA B > 0, then the correct ratio
for testing the hypothesis a = 0 is the ratio .W = MA /MAB .
If the hypothesis o B = 0 is not rejected, the experimenter has now three
choices for a test statistic for the hypothesis Q = 0.
(i) He can argue that the failure to reject v B = 0 is not the same as proving
aA, = 0 and that MAB is still his best estimate of o2 + naAB.
(ii) He can act as if v B = 0, in which case E(MA ) = 0 2 + bnaA2, and use the
test statistic JF = MA /M .
(iii) He can pool SAB and Se to obtain a joint estimate of a 2 , and use the ratio
= (abn a b + 1)M A /(Se + SAB)
4.7 The Mixed Model

We proceed now to the situation in which one factor is fixed and the other is
random. If there is no interaction between the factors, this situation poses no
problems; the experiment is similar to the randomized block experiment of the
previous chapter. If, however, interaction is present, difficulties arise. The
question of a suitable model for this situation has been widely discussed,
[Ch. 4
particularly by Wilk and Kempthorne (1955) and (1956), by Scheff (1956a and b),
and by Cornfield and Tukey (1956). The approaches of Cornfield and Tukey and
of Wilk and Kempthorne involve deriving the expected values on the assumption
that the levels of the fixed factor A are a random sample of size a from a finite
population of size p, and then putting p equal to a. We shall follow Scheff's
derivation. Again, we take the breakdown into sums of squares derived for the
fixed effects case, and concern ourselves with the expected values of those sums
of squares under the new conditions.
The new model is
Yifk=
+ ai
+b!
+Cij
+eifk;
Yuk, , b1 , and e, i ,, are defined in the same way as in the random effects model,
and the bi and the eifk are all uncorrelated; the effects at are unknown constants
subject to the side condition 1 a, = 0; c, i is the interaction term.
The interactions c15 are random variables: they involve a random factor B.
They are not correlated with any of the bi or the etfk . However, bearing in mind
that for the fixed effects a, we have 1 ai = 0, we impose a similar condition
c, 5 = 0 for each j upon the interaction terms. Consequently, the c{j are, for
any given j, correlated with each other. It is convenient to write V(c 1) as
(a 1 )aS/a.
,. . . ,
.
Lemma. Let x l , x2
x, be a set of random variables having the same variance
a2 and the same covariances pat, and being subject to the condition x, = const;
then p = (n 1) -1
PR00F. We have V(I x,) = na 2 + n(n l)pQa. But 1 x, = const implies
that V(I xi) = 0. Hence, n(1 + (n l)p)a 2 0, and so p = (n 1) -1
Applying this lemma to the c,i , with n = a and a2 = (a l)o S /a, we have
for i 96 i', j 0 j'
cov (cij , c1 5) = a -1 QAB,

.
cov (cjj , ct5 ) = cov (c tf , c,.1 .) = 0,
V(ci .) = b - 'V(cu) = b-l a -1(a 1)QA B .

We proceed now to derive the expectations of the sums of squares.
E(Se ) = ab(n 1)a 2 ,
as before.
SA = bn
Is (Yt.. y) .
2
(Y..y)=a,a.+b.b.+ c 1 . c.. + e t .. e...

=
at+ci.+ei..e...,
4.6)
Sec. 4.7]
The Mixed Model 79
since a. = 0 and c.. = 0. Thus, E(SA) = bn 1 a? + n(a 1)a B + (a 1)a 2 ,

and writing 1 a? = (a 1)o,
E(MA ) = a 2 + naL + bna.
(4.7)
Ss = an (Y.. y) a .
(y.,.
Y)=a.
a.+b5
b.+ c. 1 c.. + e. 5 . e...
=b f b.+e. 1 .e...,
since c.1 = c.. = 0. Then

E(SB ) = an(b 1){V(b1) + V(e. J .)} = (b 1)(a 2 + anuB),
and
E(MB) = aa + an4B.
Sas = n
^ r
(4.8)
(Y,1. y{ .. Y.J. + 7) 2 ,
where
(Yii. y,.. y. i . + y) = c, f c{ . + (et1 . e,.. e. j . + e...) .
The error term is the same term that appears in the earlier models; its contribution to E(SAB ) is (a 1)(b l)a 2 . E{I5 (cij c,.) 2 } = (b 1)V(c 15) for each i.
Hence, E(SAB ) = (a 1)(b 1)(a 2 + nO B ), and
E(MAB) = a 2 + na.
(4.9)
From this, we see that if a4 B 0 0, the hypothesis Q = 0 is tested by the ratio

= MA /MAB , while the hypothesis o8 = 0 is tested by .5F = MB /Me .
If there is only one observation per cel!, S e does not exist, and we have only
the three sums of squares SA , SB , and SAB . In the random effects case, this
presents no problem in testing o = 0 or oB = 0 because MAB is the appropriate
denominator for the test statistic anyway. However, in the fixed effects case (and
in the mixed case considered in this section when we test QB = 0), there is no
test for the main effects unless it can be assumed a priori that a3 = 0.
Some people find it surprising that although A is a fixed effect, its expected
mean square contains the interaction term (4 B , which would appear at first
glance to go with A random rather than with A fixed. The approaches of Wilk
and Kempthorne and of Cornfield and Tukey throw more light on this. They
obtain E(MA) as
E(MA) = a 2 + (1
b/R) (6 + bna4,
[Ch. 4
where q denotes the number of levels of B in the population of levels from which
the experimenter chose the b levels actually used.
Then, if B is fixed, b = q and (1 b/q) = 0; but, if B is random, q = co and
(1 b/q) = 1. We may also picture ourselves, when looking at the main effects
of A, as averaging the interactions cij over all the levels of B that were used. If B
is fixed, then we have averaged over all possible levels of B, and the interaction
terms "cancel out"; on the other hand, if Bis a random factor, we are averaging
over only a fraction of the levels of B, and it would not be reasonable to assume
that the interactions average out in any given experiment.
4.8 The Two-Stage Nested or Hierarchic Design

We illustrate this design by an example. Consider b batches of gasoline
produced by some process. From each batch, s samples are taken, and then n
observations are made on the octane number of each sample. If y, fk denotes the
kth observation on the jth sample from the ith batch, we have
YiJk = + b, + si(> + e,cin,
where bi is the effect of the ith batch, s, ) is the effect of the jth sample in the ith
batch, and ek(tn is the observational error.
If we assume to begin with that , b, and s,( , ) are all fixed, we have bs samples
and 1 + b + bs parameters. We impose side conditions 1i b, = 0 and I1 sf( , ) = 0
for each i. Then, if yij . and y,.. denote the sample and batch means,
= Y,
bi = Yi.. Y,
s'j(i) = Yj. y{ ...
The sum of squares attributable to batch differences is obtained by treating the

setup as just a one-way layout with b batches and sn observations per batch,
which gives
SB = T?./ns G 2 /N,
where N = bsn. Regarding the set-up as a one-way layout with bs samples and
n observations pet sample gives a sum of squares
SR =T1. /n
Ga /N.
Then, subtracting, we have the sum of squares for samples within batches, as
Ss in a = SS(B) _
T>ln
i 1
Tt?ln s ;
Sec. 4.9]
The Diallel Cross Experiment 81
and also
Se =
r k
Yttk
Tiran.
t
Alternatively, we may argue that SB = l t Fi{ T.. and Ss(B) _ ^{ ^ i sj(, ) Tf .. The
etension to more than two stages is obvious.
If either of the two factors is random, we take the same breakdown into sums
of squares. The most common situation is that in which S is a random effect, the
sf(, ) being (normally) distributed with mean zero and variance oi. We write v
for V(e, Jk ). We shall assume that the batches are distributed with mean zero and
variance az. An example from genetics is the case of s sires being chosen and
each of them being mated to d dams (there being ds dams altogether); then n
siblings are chosen from each of the ds litters. The expectations of the mean
squares are given in Table 4.5.
TABLE 4.5
Source
SS
d.f.
Total
N-1
LL I kYlk-C
2, T2./dn C
Between sires
s 1
Dams in sires
li 2 1 Taf ./n 2, T?./dn
s(d 1)
Between siblings 1, I f l kYuk l+ 1i T./n sd(n 1)
EMS
a + na + dna 2
a + noi
ao
Under normality, the testing of the hypotheses oi = 0 and a = 0 is carried out

by the F test in the usual way. The principal interest in the experiment usually
centers, however, on the problems of estimating the components of variance
v, ai, and 1.
4.9 The Diallel Cross Experiment

This kind of experiment is commonly performed by geneticists who are
interested in selecting lines and strains of plants or animals for further breeding.
The structure of the model used is similar to that of the two-way layout with
interaction. Observations are made on the offspring of the crosses of pairs of
inbred lines. The main effect of a given line is called its general combining
ability, a term originally defined by Sprague and Tatum (1942). Corresponding
to the interaction terms in the two-way layout are the specific combining abilities.
The reader is referred for a more detailed discussion to the papers by Griffing
(1956) and Kempthorne (1956b).
We consider here a situation in which p inbred lines are chosen and some
crosses are made. Griffing discusces four different experiments. In methods 1 and
2 the p parental lines are also included; in methods 3 and 4 they are omitted.
In methods 1 and 3 the reciprocal crosses are included, i.e., both the cross in
[Ch. 4
which the ith line is the male line and the jth line is the female line and the
reciprocal cross withj as the male line and i as the female line. In methods 2 and
4 the differences between reciprocal crosses are ignored. We shall consider
method 4, which is the simplest one, in detail, and confine ourselves to only a few
remarks about the others.
In Griffing's method 4 all p(p 1)/2 crosses are made. From each cross n
offspring are taken and some response, such as the yield or the number of
bristles, is observed; Y, ik denotes the kth observation on the cross between the
ith and jth lines. The model is
y, ik =m+gi +gi +s{i +e, ik , 1 <i<j<p, 1 <k<n,
where g, and g; are the general combining abilities of the ith and jth lines,
s is the specific combining ability of the pair of lines, m is a general mean, and
e{ik is a random error term with the usual assumptions.
We consider first the fixed effects model in which the parameters g { , st , and.m
are unknown constants. Let Y; denote the sum of all the observations on all
offspring from crosses made with the ith line; YY1 denotes the sum of all the
observations on the i x j cross, i.e., Yt1 = y. Let G denote the sum of all
the observations. Since each observation y, fk appears in the two totals Yt and Yi ,
we have 2G = Y. When we write 1 1 s ii the summation will be over all i
and j with i < j; there is no term si{ .
The oormal equations are
G=Nm+ n(p 1) +
where 2N = np(p 1),
Yi =n(p 1)m+ n(p 1)gt +n gf+n>su,

where the prime denotes summation over all values 1 < j < p, except i = j,
Yi =n(p 1)m+ n(p 2)g { +n1gf +n
ti
and
=nm+ngt+ng1+nsii.
Imposing the side conditions _7, g; = 0, I f stj = 0 for each i, we obtain the
estimates
m = GIN,
81 = (YY n(p 1 )m)/n(p 2),
gu = YY/ n g^ gi m.
Sec. 4.9]
The Diallel Cross Experiment 83
The corresponding sums of squares in the analysis of variance table are

(i) The mean, G'/N, with one degree of freedom,
(ii) Betv een lines, or the sum of squares for general combining ability,
,
Sg
yi81=
Ya
4G2
n(p -2 ) np(p - 2)

with (p 1) degrees of freedom,

(iii) Specific combining ability,
Sa =
Yii^{j =
n ya2
Y12 2G2
n(p
2) + n(p 1 )(p 2)
with p(p 3)/2 degrees of freedom,

(iv) Se , obtained by subtraction with p(p 1)(n 1)/2 degrees of freedom.
In the random effects case we assume that g,, s, l , and e, ik are uncorrelated
random variables with zero means and variances a, o$, and a, respectively. The
geneticist is interested in obtaining estimates of these components of variance.
We take the same sums of squares as in the fixed effects case, and we now derive
their expectations. There is no difficulty in showing that E(Se ) = p(p 1)
x (n 1)u/2. We also have
E(G 2) = N 2m 2 + n 2(p l) 2pa + n 2p(p 1)o$/2 + No,
E(Y? ) = n 2(p 1 ) 2m 2 + n 2 (p 1 ) 2 a 2
+ n 2 (p 1)a + n 2(p 1)a$ + n(p 1)a
= n 2 (p 1 ) 2m 2 + n 2p(p 1 ) a + n 2(p 1 )a + n(p 1 ),
E(Y f) = n 2m 2 + 2n 2 a + n 2 a + na.
It follows that
np(p 2)E(S9) = p 1 E( Y?) 4E(G 2) = p 2E(Ya) 4E(G 2)
n 2p(p 1 )(p 2) 2 a + n 2p(p 1 )(p 2)a,
+ np(p 1 )(p 2)4
so that the expectation of the mean square is

E(M9) = ae + na; + n(p 2)4
We may similarly show that E(M) = o + na;.
[Ch. 4
In the general case of method 1, in which the reciprocal crosses and the
parental lines are included, the model assumed by Griffing is
Ytrk = m + gt + Sr + st) + rti + etrk,
1_<i<p, 1<j<p, 1<k<n.
Here y, jk denotes the kth observation on the cross with i as the male line and j
as the female line; the specific combining ability s, is symmetric, s,i = s5{ . The
reciprocal effect, rt ,, accounts for the difference between the i x j and j x i
crosses, and is antisymmetric, rtj = rfi .
We define the totals Yt . = I i I k Ytrk, Y> = ^t I k Ytjk, and Y 5 . = I k Yt,kThen the estimates of the parameters in the fixed effects case are given by the
equations
np'th = G, 2np 2g{ = 2p(Yi . + Y.,) 2G, 2nr ti = Yt f .

2np 2 sti = p 2 (Y5. + Y,t) p(Y1. + Y. t + Y1 . + Y. t) + 2G,
(i 0 j),
np a gtt = p 2 Y0 . p(YY . + Y. t ) + G.
The models for the other two methods may be derived from that of method 1
by dropping the unnecessary terms. The analyses follow the same lines.
Exercises
1. An experiment is carried out to investigate the deterioration in a product after

storage for different lengths of time at different temperatures. The experimental
design is a two-way crossed classification with factor A being time at a levels and
factor B being storage temperature at b levels. However, one of the levels of A is
zero, and zero storage time is the same at all temperatures. Thus there are
b(a 1) + 1 cells and n observations in each cell. How would you analyze the
data ?
2. In the two-stage nested design we assumed that there were n siblings chosen from
each litter. In practice, however, some litters may have fewer than n siblings.
Modify the analysis for the case in which the cell corresponding to the jth litter by
the ith sire contains ntj members where n ij > 0 for all i and j.
3. (Continuation). Another possibility is that one or more litters might be completely
lost. Modify the analysis for the situation in which there are only d 1 litters by
one of the sires; each nonempty cell has n observations.
4. (Continuation). In the industrial context with batches, samples in batches, and
observations or analyses on samples, the hierarchical design can result in a large
bili for chemical analysis. Suppose that we decide to make two observations on the
first sample from each batch, and only one observation on each of the subsequent
samples. Modify the analysis accordingly.
5. Find the expected values of the mean squares in Griffing's method 1 for the diallel
cross experiment.
6. (Continuation). Derive the analysis for methods 2 and 3.
Exercises 85
7. Ina two-way layout the factors A and B are fixed and the cell frequencies satisfy the
condition of proportional frequencies. Derive the analysis with AB interaction
terms.
8. Batchelder et al. (1966) investigated the growth of tobacco. They grew tobacco in
four different soils (blocks) with three different mulches, A, B, and C, and continued the experiment for three years. The data were analyzed as a two-factor
experiment, years Y and mulches M in four randomized blocks. The data which
follow are yields in pounds per acre. A was a control (no mulch at all); blocks 1 and
II were a different type of soil from blocks III and IV. Show that the soms of
squares for mulches, years, and blocks each have one large contrast that accounts
for most of the sum. What conclusions would you draw from this?
Blocks
Year
Mulch
II
III
IV
A
1962
984
1440
1324
421
1276
1232
563
1016
1064
862
880
1016
1539
2141
1939
1039
1894
1833
914
1819
2394
1032
1675
2055
1607
1860
1846
1461
1912
1973
1593
1703
1043
1507
1378
1319
C
A
1963
C
A
1964
9. The following data are yields, in grams per plot, in diallel crosses between inbred
lines of winter beans [Bond (1966)]. He planted all crosses between six inbred lines,
and also the inbred lines themselves, a total of 36 varieties in all. The yields
reported are the mean values over two years. Analyze the data under Griffing's
model 1, both with fixed effects and with random effects.
Male Parent
Female
Parent
24
31
36
55
64
67
24
31
36
55
64
67
172.8
247.7
267.5
301.7
267.5
262.0
279.5
177.8
274.0
278.5
253.0
274.3
277.2
258.5
236.0
269.5
248.0
254.5
278.2
263.3
250.7
224.2
267.2
259.8
279.8
303.2
258.5
260.5
208.2
281.8
315.0
313.2
256.7
285.5
287.5
221.3
CHAPTER
Experiments
with
Several Factors
The results obtained in the previous chapter can be extended to larger experiments involving more than two factors. The simplest situation is the complete
factorial experiment in which all the factors are fixed and crossed. This is a
routine extension of the two-way crossed classification. The term crossed
classification is used to denote that each factor is crossed with the others:
factor A cuts across factor B in the sense that each level of factor A appears with
each level of factor B.
We shall see that in the crossed classification with some of the factors random
complications can arise about the Ftests. In some cases, no one mean square will
provide a suitable denominator for the F statistic to test a hypothesis, and a
denominator has to be put together from several mean squares. We shall also
consider designs that contain both crossed and nested factors. A set of rules will be
given for deriving the degrees of freedom, sums of squares, and the expectations
of the mean squares. The chapter ends with a short discussion, including some
examples, of split-plot and split-split-plot designs. Before proceeding further, it
is appropriate to make a few remarks about error terms, pooling, and replicates.
5.1 Error Terms and Pooling

We have seen in the previous chapter that in a two-way layout with a single
observation per cell we do not have an estimate of the error variance unless we
are prepared to assume, a priori, that there is no interaction between the two
factors. This situation persists when there are more than two factors, and it is a
general operating procedure that if there is only one observation per cell, the
EA
Sec. 5.2]
Replicates 87
highest order interaction (the interaction between all the factors) is taken to be
zero, a priori. If the experimenter does not want to assume this in any case
except that in which all the effects are random, he will have to find some way of
duplicating his observations to obtain an unbiased estimate of error. If several
of the factors are at more than two levels, this procedure of ignoring the highest
order interaction will provide an error term with a reasonable number of
degrees of freedom.
If an experiment involves four factors, each at four levels, the ABCD interaction will have 3 4 = 81 d.f. Assuming the effects to be fixed, the critical value
of F for testing the hypothesis that the main effects for A are zero, using a = 0.05,
is F *(3, 81), which is approximately 2.7. If the factors are at only three levels
each, we should have 16 d.f. for ABCD and F *(3, 16) = 3.24, which means
that we should need a 20 per cent larger F value in order to detect differences
between the levels of A, a considerable decrease in power. If all the factors are at
two levels each, there will only be a single degree of freedom for each of the mean
squares. In the Jatter circumstances, the experimenter is going to be rather
desperate for degrees of freedom in the denominators of his F statistirs, and it is
customary in such experiments to assume, in the absence of information to
the contrary, that all interactions involving more than two factors are negligible,
a priori, and to pool their sums of squares to obtain an error term. We shall
tacitly make that assumption in the later chapters when we discuss experiments
with factors at two or three levels.
The pooling of interaction sums of squares to provide an error term is a
subjective matter. On the one hand, purists can deplore the use of a possibly
biased denominator and a possibly invalid F statistic. On the other hand, the
practical man can hardly be condemned for making use of reasonable assumptions to obtain enough degrees of freedom in the denominator for his F test to
have enough power to be useful. What is important is that the experimenter
should use his common sense. There used to be a popular procedure, which has
hopefully passed into disuse, by which the experimenter would start his F tests
at the foot of the table, testing the mean square in the next to last line against the
mean square for the error. If the ratio was larger than the critical value of F,
he recorded the fact in the right-hand column and moved on to the next line.
If the ratio was not significant, he pooled the sum of squares with the error
term and went on to the next line. So he continued up the table automatically
pooling every sum of squares that was not significant, oblivious to the fact that
just because the F ratio is not quite big enough for significance it does not
necessarily follow that the hypothesis is true and that if you pool a sum of
squares that is almost large enough to be significant you will inflte your error
estimate and bias it upward.
5.2 Replicates
The word replicate (or replication) commonly occurs in the jargon of experimental design, but its meaning is not always the same. When in later chapters we
88 Experiments with Several Factors
[Ch. 5
come to discuss incomplete block designs, we shall consider a design in which

there are five treatments, each tested four times for a total of twenty plots; we
shall then say that each of the treatments is replicated four times, or has four
replications. In the context of the factorial experiment, however, a replicate or a
replication means a complete replica, repetition, or copy of the basic design with
one observation per cell.
Suppose, for example, that .we wish to test five varieties of beans using four
different fertilizers, and that there are to be three plots with each combination of
bean and fertilizer. In the completely randomized experiment there would be
sixty plots assigned at random, and it would be improper to speak of replications.
To say that the experiment has three replications implies that there are three
separate groups of twenty plots each. Each group constitutes a replicate, or
replication, inasmuch as it consists of a single plot with each bean-fertilizer
combination. If the three replicates are in the same field, they may just as well be
called blocks; it seems sometimes to be a matter of individual taste. The replicates need not, however, be in the same field. It would be possible to have them
on different farms or, in an industrial experiment, to run them on different days.
The following operating rule is commonly accepted, often without being
stated: replicates and blocks do not interact with treatments or factors. As a
result, the reader will often see in the literature analysis of variance tables that
contain a sum of squares for replicates but in which there is no mention of
interactions involving replicates because they have been automatically pooled
into the error term.
Henceforth in the expected values of sums of squares we shall use, for example,
the notation of a whether A is a fixed effect or a random effect. If A is random,
o is the usual component of variance. If A is fixed, a is what might be called a
pseudocomponent of variance and represents an expression such as (a a.) 2 /
(a 1).
_7
5.3 The Complete Three-Factor Experiment

We consider now a complete factorial experiment with three factors A at a
levels, B at b levels, and C at c levels, There are n observations in each cell, for a
total of N = abcn data points. The complete model has terms for two-factor
interactions and for a three-factor interaction:
YiJk. = !L + ai + bi + c k + (ab) 1 + (ac), k + (bc) ik + (abc)tik + e{jkm
The three-factor interaction essentially takes in all the scatter in the cell means
that is not explained by the main effects and the two-factor interactions. Its sum
of squares can also be derived by considering the variation in the sums of
squares for AB interaction at the several levels of C (or equivalently the AC
interaction at the levels of B, or the BC interaction at the levels of A).
The Complete Three-Factor Experiment 89
Sec. 5.3]
In the fixed effects case, the analysis follows the same lines as that of the twofactor experiment. The side conditions imposed are
lat =0,
b1 = 0,
(ab)t, = 0,
Ck
0;
(ab)t1 = 0,
(ac)tk = 0;
k
(ac)tk =
0,
k
(bc) 1k = 0 ,
(bc) 1k = 0 ;
(abc)11k = 0 .
(abc) t1k = 0 ,
(abc)uk = 0 ,
Typical estimates of parameters are

at = Yt... y,
F^ =7,
(abC) t5k
(ab)11 = Y{1.. yt ... y. 1 .. + y,
= Yt5k Yt1.. yt . k . y. 1k . + y t ... + y. 1 .. + y.. k . y r
with typical sums of squares

SA = T?../bcn C
{
witha 1 d.f.,
SAB=Tu../cn C SA SB,
t
with (a 1)(b 1) d.f.,

SABC=111 Tfk./n C SA SB SC SAB SAC SBC
t 1 k
with (a 1)(b 1)(c 1) d.f..

The expectations of the mean squares are E(MA ) = a' + bcnaA2, E(MAB) _
oa + cnaA2B, E(MABC) = a 2 + naA BC where a = >_ aa/(a 1), and so on. The
mean square for error Me = Sa serves as the denominator for each of the F tests.
The extension to four or more factors is straightforward.
When one or more of the factors is random, matters become somewhat more
complex. In each case we take the breakdown into sums of squares for the fixed
effects case, but there are differences in the expectations of the mean squares.
We may follow the development in the two-factor case to derive these expectations. If all three factors are random, we have
E(MA) = a a +
nUABC
+ CnaA B + bnaA C + bcnaA2,
E(MAB) = a2 + naA2BC + cnaA2B,

E(MABC) = a2 + naABC,
E(M8) = aa.
[Ch. 5
For the mixed model, we recall that in the two-factor case with A fixed and B
random, the term oM B did not appear in E(MB ) because for each level of B the
interactions were averaged over the entire set of levels of the fixed factor A.
On the other hand, O B did appear in E(MA ) because for each level of A the
interactions were averaged over only those b levels of B that were (randomly)
chosen from the infinite population.
With this in mind, we can formulate a set of rules for writing down the
expectations of the mean squares in the complete factorial experiment which are
applicable to any number of factors, fixed or random.
1. Each expected mean square contains a term aa.
2. Each component of variance has some letters as subscripts. The coefficient
of a component in any expectation is the product of all the letters that are not
among its subscripts. Thus, in the three-factor case we have the four letters
a, b, c, and n. The coefficient of Q B whenever it appears is the product cn;
similarly, we have fla Bc and bcn4 The coefficient is the number of observations
contained in the corresponding total, such as Ti ,...
3. If all the factors in an experiment are random, the expectation of a mean
square consists of o 2 together with all the terms for effects and interactions
containing all the letters in the name of the mean square in question. Thus,
E(MAB) contains terms for all components with both A and B among the
subscripts, e.g., a 2 + nUBC + cnO9.
4. If some of the factors are fixed, we strike out in the expectation of any mean
square all those components which have amng their subscripts any letters,
other than those in the name of the mean square, which correspond to fixed
factors. For example, in the three-factor case we have, when all three factors are
random,
E(MA) = a2 +
noec
+ CnvOB + bnalc + bcno .
Whether A is fixed or random has no bearing on E(MA). However, if C were

fixed, we should strike out all terms with C as a subscript and obtain
E(MA )
= a 2 + cna$ + bcna.
If B were also fixed, we should have E(MA ) = a2 + bcna.

We give in Table 5.1 the expected mean squares for a three-factor experiment
in three cases:
(i) A, B, and C are all random;
(ii) A and B are random, C is fixed;
(iii) A is random, B and C are fixed.
In the right-hand column we list the mean square that is the denominator in the
F test for the hypothesis that the corresponding effects or interactions or components of variance are zero. An asterisk denotes that we have to construct a
Sec. 5.3]
The Complete Three-Factor Experiment 91
suitable denominator. That topic is discussed in the next section. If n = 1, we

have to assume that aAeC = 0 and use MARC to estimate a 2 .
TABLE 5.1
EXPECTED MEAN SQUARES
I.
A
B
C
AB
AC
BC
ABC
Error
a2 + naABC
a2 + naac
a2 + naBC
a2 + na4ac
a2 + na aicc
a2 + noBC
a2 + naec
a2
All Factors Random
+ cnaAB + bnaAc + bcna

+ cnaB + analc + acno
+ bn4 c + cn4Bc + abna2
+ cnae
+ bnac
+ anac
*
MABC
MABC
MABC
Me
II. A and B Random, C Fixed
AB
AC
BC
ABC
a2 + cnaB
a2 + cna3e
a2 + noAac
a2 + Cnca
a2 + naBC
a2 + naac
a2 + naBC
Error
a2
A
B
+ bcna
+ acna4
+ bnac
MAB
MAB
+ analc
+ abn4
Me
+ bnac
+ anac
MABC
MABC
MB
III. A Random, B and C Fixed

A
B
C
AB
AC
BC
ABC
Error
a2+bcna
aa+Cna +acna8
a2 + bnac + abnvc
a2 + cnaB
a2 + bncc
a2 + naec + analc
aa + naABC
a2
M,
MAB
Mnc
Me
Me
MABC
MB
EXAMPLE. The following data are taken from an experiment reported by

Woodman and Johnson (1946). They were investigating the effects of various
fertilizers on the growth of carrots. The three factors were nitrogen (ammonium
sulphate), phosphorus (monocalcium phosphate), and potassium (potassium
sulphate). The data are means of the weights (in grams) of the roots of five
[Ch. 5
plants in each group grown under the various conditions. We have multiplied
the data by 100.
No N1
Po
P1
P2
N2
Ko
Ki
K2
Ko
Ki
K2
Ko
ICi
K2
8876
8745
8601
9141
9827
10420
9785
9585
9009
9483
8457
8106
10049
9720
12080
9975
11230
10877
9990
9298
9472
10023
10777
11839
10451
11094
10287
Table 5.2 follows:

TABLE 5.2
ANALYSIS OF VnxiANce
Source
Nitrogen
Potassium
Phosphorus
NxK
NxP
KxP
NxKxP
SS
488.3675
1090.6564
49.1485
142.5844
32.3475
592.6238
185.7762
d.f.
2
2
2
4
4
4
8
5.4 Approximate F Tests

We consider now the problem of testing the hypothesis a = 0 when all three
factors are random. The difficulty lies in the fact that there is no mean square
that has for its expectation E(MA ) bcnc or a 2 + naA2 + cnc B + bnaC.
If either a B or a 2 C can be taken as zero, the problem is resolved. If we decide
that c = 0, MAB becomes a suitable denominator for the F test; if o B = 0,
we can use MAC . Otherwise, a large value for the ratio MA /MAB merely indicates
that the sum bna C + bcna is not zero, and, if the hypothesis a = 0 has
already been rejected, this throws little light on the size of o. We have, therefore,
to construct our own denominator from the several denominators available.
(We are assuming now that the appropriate tests have already rejected the
hypotheses ad BC = 0, ah = 0, and a = 0).
We shall take as denominator the linear combination
u= MAB +MacMARC,
Sec. 5.4]
Approximate F Tests 93
and take as our test statistic 9' = MA/u. MA (a 1)/E(MA) has a Xa(a 1)
distribution, and MA and u are independent. We now need to find the approximate distribution of u.
It will be recalled that in the simpler case of the one-way layout with n
observations per cell and random effects our test statistic was
.'F = MA/Me .
Under normality, SA /(aa + naA2) and Se /aa had independent X 2 distributions

with (a 1) = S6 A and a(n 1) = 0 e d.f. respectively. We then argue that
H
, _ b
A 1 SA l (a a
+ naA2)
aa
MA
ce 'Se/a 2 aa + naA2 Me
has the F[4 , q e ] distribution and that

a a + nu
a
F(#A, #)
This argument depends upon MA and Me being multiples of independent X2

variates. In the present case MA and u are certainly independent, but u does not
have a X 2 distribution. Indeed, it has not yet been possible to derive the distribution of u. We seek therefore to approximate the distribution of u by regarding u
as a multiple of a x 2 variate, i.e., by obtaining numbers p and 0 such that
z = u/p has the same mean and variante as a x 2 variate with 0 d.f.
For the more general case of the distribution of linear combinations of X2
variates, the reader is referred to Satterthwaite (1946) and Welch (1956). We shall
confine our discussion to the present example.
If z _ X2, then E(z) _ and V(z) = 20. The random variables 6ABMAa /
[E(MAB)], # ACMAC I [E(MAC )], and OABCMABCI [E(MABC)] where DAB = (a 1)
x (b 1), and so on, have independent X 2 distributions. Then, for example
V(MAB) =
[E(MAB )]2
0AB
[E(MAB )]2
.2
6AB = 2
6AB
If we fit p and by letting E(u) = E(pz) = p and V(u) = p 2 V(z) = 2p 20, then
E(MAB + MAC MABC) =p6 = E(u) and V(MAB)+V(MAC)+V(MAsc)=2p2C=
V(u), so that 0 = 2 [E(u)] 2 / V(u) and p = 0 - 'E(u). Unfortunately, we are
in the difficult position of not knowing any of the expectations. We therefore

replace E(MAB ) by the observed value MAB , and so on. We then obtain p6 = u,
and
pa^
__ Ma
MAC M BC
OAB + ÂC + OABC^
[Ch. 5
so that,
(a 1)(b 1)(c 1)u 2

= (c 1)M B + (b l)MAC +MaABc
With this approximation we have
MA __ SA __ E(MA) ,
SA , ^ = E(M4)
. F(a 1, ^)
u
(a 1)zp
p^
(a 1)E(MA ) z
E(u)
The other cases are handled in a similar fashion.
5.5 An Experiment with Crossed and Nested Factors

Suppose that we have a chemical process in which a feed stock is pretreated in
one of several ways to form an intermediate, and then the intermediate is
proceseed in a second stage at one of several temperatures. We suppose that there
are a methods of pretreatment. We prepare b batches by each method. From
each batch we take cn samples; n of these are then treated at each of c temperatures. If ytikm denotes the response observed on the mth of those samples which
were treated at the kth temperature and taken from the jth of the batches that
were prepared by the ith method, we assume a model
Y(lkrn = + ai + bl(i) + Ck + (aC)ik + bClk(i) + eilkm
The batches are nested within the methods (of pretreatment). The temperatures
are crossed with the batches and, therefore, with the methods.
In the formulation of the model (we assume to begin with that all the effects
are fixed), is a grand mean and at is the main effect of the ith level of the first
factor; ai represents the difference between the expected average response on all
ben runs made on stock pretreated by the ith method and the grand mean ,
i.e., ai = E(y( ...) . We have the side condition l i ai = 0. The b batches made
by the ith method may differ; bf(() denotes the main effect of the jth batch in the
ith method. It is the difference between the expected average response of all cn
runs made on stock taken from the (ij)th batch and the expected average for all
the batches made by the ith method, i.e., bf(() = E(y ii .. yj ...). It follows that
Ii bt(() = 0 for each 1. The main effects of the levels of temperature are denoted
by C k ; we have ck = E(y.. k .) , the averages being taken over every batch
from every method, and I k Ck = 0.
There are two interaction terms. The method x temperature interaction
(ac) ik represents nonadditivity between methods and temperatures. Perhaps
changing from temperature 1 to temperature 2 produces a greater change in
response on feed stock pretreated by method 1 than on feed stock pretreated by
Sec. 5.5]
An Experiment with Crossed and Nested Factors 95
method 2; we have (ac), k = E(yt . k .) at ck = E(Ytk. Yi--- Yk +

y....). The side conditions are 1, (ac) ik = 0, and l k (ac) ik = 0. The temperature x
batch within methods interaction, (bc) ik(j) , represents nonadditivity between
temperatures and different batches made by the same method, and, in essence,
accommodates all the scatter between cell means that has not been taken into
account so far. The side conditions are k (bc) 1k(j) = 0 for each batch, and
^1 (bc) fk(i) = 0 for each combination of temperature and method.
There is no AB interaction term. We cannot talk about the main effect of
batches changing from method to method because batches are nested in methods.
A run made at temperature 1 on a stock pretreated by method 1 and a run at
temperature 1 on a stock pretreated by method 2 have in common the (possibly
important) fact that they were both made at the same temperature. The first
batch made by method 1 and the first batch made by method 2 do not have a
batch factor in common; as we have formulated the experiment there are just b
batches made by each method, and being the first or second batch is just a label.
Similarly, there is no ABC interaction. That would have corresponded to an AB
interaction varying from temperature to temperature.
We could now proceed in two ways. We could go back to the linear model,
obtain the normal equations, solve them, and derive sums of squares by routine
algebraic procedures. We shall, however, take instead a more practical approach,
which can be formalized by considering a set of nested hypotheses.
Our experiment involves ab batches and c temperatures. We can analyze it as a
two-way layout with factors B (batches), C (temperatures), and the usual sums
of squares.
TABLE 5.3
SS
Source
Total
d.f.
Y'Y C
abcn 1
T,,../cnC=SB
4
ab-1
T?k,/abn C = ST c
BC
T,k./n C SB ST (c 1)(ab 1)
4 1 k
SB By subtraction
abc(n 1)
The sum of squares for batches falls into two components, just as in the ordinary
nested design.
SA = 1, Ti?../bcn C is the sum of squares between (or for) methods.
SB(A) = lt Ij T,ll ../cn l i T../bcn is the remainder, and is attributable to
differences between batches made by the same method.
A similar process of subdivision can now be applied to the interaction sum of
squares. We divide SBC into two components; SAC is the sum of squares for
[Ch. 5
interaction that we would obtain were we to ignore batches and to consider a

two-way layout with a methods, c temperatures, and bn observations in each
cell. Thus SAC =1,i E, T 2 i . k../bn-1,^ T i 2 ../bcn ET 2 . k/abn + C witli (a 1)
(b 1) d.f.; the rest of SABC with a(b 1) (c 1) d.f. is the sum of squares
for interaction between temperature and batches made by the same method (or
more concisely temperature x batches in methods).
In practice, it is more reasonable to regard the b , as random variables,
uncorrelated (or independent under the normality assumption) with mean zero
and variance oB. The interactions (bc) jk(i) will also be random variables. Suppose
that we continue to take methods (A) and temperatures (C) as fixed. The expectations of the sums of squares may be derived by following the same procedures
that we used for the mixed model in the chapter on the two-way layout. This can,
however, be a lengthy process and we shall give in the next section some rules for
writing down the expected mean squares more easily. In our example the table of
expected mean squares with A and C fixed and B random is
;( )
TABLE 5.4
EXPECTED MEAN SQUARES
EMS
d.f.
SS
+ cn4(A) + bcncA
SA
a1
a2
SB(A)
a(b 1)
a' + CnaB(A)
Sc
c1
02 + n4 c(A) + abnac
SAC
(a 1)(c 1)
SBc(A)
a(b 1)(c 1)
a 2 + naBC(A) + bnaA C
a 2 + n4C(A)
Se
abc(n 1)
02
5.6 General Rules for Sums of Squares and EMS

We present now a set of rules for writing down the form of the model,
formulae for the sums of squares, their degrees of freedom, and the expectations
of the mean squares. These rules are also given by Scheff (1959, pp. 282-289)
and by Bennett and Franklin (1954, Sec. 7.6). We shall use the example of the
last section as an illustration.
1. The model contains, in addition to th mean and the error term, all main
effects and interactions corresponding to sets of 2, 3,..., factors, except that
there are no interaction terms containing two factors, one of which is nested in
the other. Thus, in the example there was no AB interaction and no ABC
interaction. The subscripts attached to a term are those corresponding to the
factors appearing in the term. If one of the factors in a term is a nested factor,
the letters corresponding to the factors in which it is nested are added in brackets
as subscripts. Thus, we had for the BC in A interaction, (bc) ;k( , ) . The error term is
written in the form em( , jk) for our present purposes.
General Rules for Sums of Squares and EMS 97
Sec. 5.6]
2. For each term or effect in the model (we let the word effect now embrace
both main effects and interactions), we divide the subscripts into three classes:
(a) livethose subscripts that are present in the term and are not bracketed;
(b) deadthose that are present and bracketed; (c) absentthose subscripts
that appear in the model but not in the particular term. Thus, in bc;k( ), j and k
are live, i is dead, and m is absent. We then form the symbolic product for the
term, in which Bach dead subscript is represented by its letter and each live
subscript by its letter minus one. For (bc) ;k(j) the symbolic product is i(j 1)
x (k 1).
3. The corresponding product in the numbers of levels, e.g., a(b 1)(c 1),
gives the number of degrees of freedom for the sum of squares.
4. Expanding the symbolic product, we obtain (ijk ij ik + i). The
corresponding sum of squares is 1(yi;k. Yij. Yi.k + y{...) 2 or
Tj tik
J ^ n
TZ ti
J cn
TZ{"'
T2 ik
11 bn + { ""
Similarly, we have for the main effect A the product (i 1), which gives
1 T?..
S,4 = (bcn)
C.
5. To obtain the EMS we prepare the following auxiliary table. There is a

row for each mean square in the analysis of variance and a column for each
letter. At the head of each column we write the subscript letter, the number of
levels of the factor, and whether it is random (R) or fixed (F). At the beginning
of each row we write the subscripts for that term in the model. The rules for
filling in the table follow the table itself.
TABLE 5.5
AUXILIARY TABLE
i ja k' mn
i
j(i)
ik
jk(i)
m(ijk)
(i) In any row, write the number of levels in any column headed by a letter
which is an absent subscript.
(ii) Write 1 if the column is headed by a dead subscript.
[Ch. 5
(iii) In any column corresponding to a random effect, write 1 in the remaining

cells.
(iv) In any column corresponding to a fixed effect, write 0 in the remaining
cells.
To obtain, for example, E(MA), we cover the ith column. In each row for
which i is a live or dead subscript, we take the product of the visible numbers and
multiply by the corresponding component of variance. Thus, E(MA) = aa +
cnaB (A) + bcnQ.
For E(MAC ), we cover the i column and the k column and obtain
E(MAC )
= a 2 + n4C(A) + bnaAC.
5.7 Split-Plot Designs

The experiment that we have just considered is an example of a split-plot
design. In the agricultural context from which the name is derived, we picture a
piece of land divided into several plots, perhaps strips of land or fields. Each
of these plots is then given a single treatment. It is convenient to call these plots
whole plots and to designate the treatments that are applied, one to each whole
plot, as main treatments or levels of a main factor. The whole plots are then
split into subplots, and one or more subtreatments are applied to each subplot.
We might, for example, take several varieties of a erop and sow several fields
(whole plots) with each variety, one variety to a field. Then each field could be
divided into six subplots, and one subplot from each field could be harvested
each week over a six-week period. The varieties would be the main treatments,
and the hervesting times the subtreatments. In comparing the varieties, we must
take into account the field-to-field variation. In comparing harvest times, we can
compare yields within each field. It follows that the main treatments are tested
against a larger error term (the whole-plot error) than the subtreatments.
In our example the methods of preparing the feedstock are the main treatments, the batches are the whole plots, the temperatures are the subtreatments,
and the samples drawn from the batches to be run at the various temperatures
are the subplots. It will be recalled that the main effect C is tested against
MBC(A) , as is the interaction AC; however, the main effect A is tested against
MB(A) . The most common situation is to have one subplot in each whole plot for
each subtreatment, i.e., n = 1. The main treatments are sometimes said to be
confounded with the whole plots.
The experiment reported by Yates (1937) involving oats and manure
(nitrogen), to which we referred in the previous chapter, was a split-plot experiment. Six blocks of three plots each were taken, and one plot in each block was
sown with each of the three varieties of oats chosen for the experiment. Each
plot was divided into four subplots, one of which was assigned at random to
each level of the nitrogen factor. The four levels were no manure and 0.01, 0.02,
,
Split-Plot Designs 99
Sec. 5.7]
and 0.03 tons per acre. The main treatments were the varieties of oats and the
subtreatments were the levels of nitrogen. The data and the analysis of variance
table are given below. The sum of squares for whole-plot error is the same as the
sum of squares for block x varieties interaction. In practice, the necessary sums
of squares can be obtained on a computer by using a program for a three-factor
experiment with factors A (oats), N (nitrogen), C (blocks), and one observation
per cel!. Then Se = SNC + SANC
Block
nl
n2
n3
n4
Block n l
n2
n3
n4
Al
AZ
A3
111
117
105
130
114
140
157
161
118
174
141
156
II
74
64
70
89
103
89
81
132
104
122
133
117
III
AI
A2
A3
61
70
96
91
108
124
97
126
121
100
149
144
IV
62
80
63
90
82
70
100
94
109
116
126
99
Al
A2
A3
68
64
102
129
112
89
132
86
96
124
VI
53
89
97
74
82
99
118
86
119
113
104
121
60
89
TABLE 5.6
Source
SS
d.f.
Total
51985.95
71
Blocks
Varieties
Whole plot error
Between whole plots
15875.28
1786.36
6013.30
5
2
10
23674.94
17
Nitrogen
20020.50
321.75
7968.76
3
6
45
NxV
Subplot error
MS
3175.06
893.18
601.33
6673.50
53.63
177.08
The following situation occurs occasionally in industrial experimentation and

is similar to the split-plot experiment. We shall illustrate it by an example.
Consider a chemical process in which propylene is to be polymerized in a reactor
using an acid catalyst, and pilot plant experiments are to be conducted using
acid strength and temperature as the two factors with a and t levels respectively.
The response measured is the amount of propylene converted into polymer.
At each experimental run, a quantity of the product is taken from the plant, and
the amount of conversion is determined by a laboratory analysis.
100 Experiments-with Several Factors
[Ch. 5
Suppose now that it is decided to have four observations at each combination

oflevels of temperature and acid strength. There are two possibilities that come to
mind. The first is to make 4 at runs in random order, and then to make an analysis
of the product of each run. The second is to make only one run at each combination of levels, but to make four analyses on the product of each run. In either
case, we end up with four numbers for each cell in the acid-temperature table.
In the first case, the scatter among the numbers in each cell reflects both the
precision of the laboratory analysis and also the `noise' in the whole system.
It measures the reproducibility of the results and throws lome light upon the
question: if we make another run at the ith acid strength and the jth temperature
tomorrow, how well shall we be able to reproduce the results that were obtained
last week? If y ;; ,^ denotes the response on the kth run at the ith level of acid and
the jth level of temperature, we could write
Yiik = m + ai + b; + (ab) t; + ei3k + f;k,

where ei;k is a term for the general noise, such as the inability to reproduce
exactly the same temperature and perhaps the same acid strength, on different
runs; f ; ,, is the analytic error. They are independent random variables with zero
means and variances Q and of respectively, and V(y ;;k ) = (a + af), V(y 15 .) _
(a + af)/4. The ordinary two-way analysis of variance procedure produces a
mean square for error s 2 with 3 at d.f. and expectation o + o, and we have
the situation of the previous chapter.
On the other hand, in the second experimental plan, y i;k denotes the result of
the kth analysis on the (ij)th run. We write
y k =m+at +b; +(ab), ; + e, ; +f ;k ,

because for any given i and j, ei. is the same for all four values of k. We have
V(Y;rx) = (a
+ af) as before, but now V(yt; .) = a + (o/4).
The sum of squares between duplicate analyses provides an estimate of o

with 3 at degrees of freedom. It does not, however, provide any information
about o. In consequence, the experimenter who uses the mean square between
duplicates as his error term in comparing temperatures is using too small an
error estimate. He is using only the analytic variance and is tacitly assuming that
his experimental conditions are perfectly reproducible. This will result in his
making too many wrong decisions to reject null hypotheses.
How, then, do we test for temperature differences in this case? The answer
is that we use for the denominator the mean square for A x T interaction. We
essentially have a two-factor experiment with one run at each combination of
levels; nested within each run are the four analyses. The setup is the same as if
there were at whole plots, each split into four subplots to which no subtreatment
is applied. If either A or T is fixed, as such factors usually would be, we have the
same problem that we faced in the last chapter when there was one observation
per cell and interaction was present. We may sum up with a useful reminder.
If, in a factorial experiment, all the main effects and all the interactions turn out
Split-Split-Plot Designs
Sec. 5.8]
101
to be highly significant in the analysis of variance, go back and look at the

experiment again. It may actually be a split-plot situation, and you may be
testing everything against the repeatability of a chemical balance in the hands of
an expert technician!
The following example illustrates a slightly different aspect of the topic that
has just been discussed. Chanda, et al. (1952) fed carotene to goats and cows over
consecutive two-day periods and used two methods to measure the digestibility
of the carotene. The data shown are for four goats over five periods. In each cell,
the first number given is the apparent digestibility measured by method 1. The
second number is the measurement by method 2.
The goat-period combinations can be regarded as 4 x 5 = 20 whole plots.
Each whole plot is split into two subplots. Method 1 is applied to one subplot
and method 2 to the other. The goat and period main effects are tested against
the mean square for goat x period interaction. The M x G x P mean square
is used to test M, M x G and M x P. It wilt be noted that if the M x G x P
mean square had been used to test G and P, each of those effects would have been
deemed significant.
Periods
Goats
Anna
Betty
Bluebell
Diana
II
I
75.1,
65.4,
61.9,
69.4,
75.2
62.7
60.1
70.1
69.0,
63.3,
57.7,
60.7,
IV
III
63.5
61.2
59.6
57.0
74.3,
59.4,
74.4,
57.0,
80.7
65.3
75.8
54.5
51.2,
62.9,
62.9,
56.6,
46.7
62.1
55.3
49.5
V
72.6,
63.8,
54.3,
64.3,
71.0
63.1
58.3
63.3
TABLE 5.7
Source
Methods (M)
Goats (G)
Periods (P)
MxG
MxP
GxP
MxGxP
Total
SS
d.f.
11.03
328.44
760.13
10.23
67.27
1090.77
57.99
2325.86
1
3
4
3
4
12
12
39
MS F values
11.03
109.48
190.03
3.41
16.82
90.90
4.83
2.28
1.20
2.09
0.71
3.48
18.82
5.8 Split-Split-Plot Designs

In some experimental situations, further subdivision of the subplots is used. An
example is the experiment conducted by Tandon (1949) to investigate the response
of flax to various rates of application and formulations of the herbicide 2,4-D.
[Ch. 5
He chose seven varieties of flax with three formulations of 2,4-D and four
rates of application. There were four replications in the experiment. Each
replicate involved twelve blocks, each with seven plots, so that there were 336
plots altogether. One plot in each block was sown with each variety of flax.
The three formulations of 2,4-D were made up, and four blocks in each replicate
were assigned to each formulation. The 2,4-D was then applied to one of the
blocks at the lowest rate, to another at the second rate, and so on. In any block,
each of the seven plots received the same formulation at the same rate. In the
split-split-plot context there are twelve whole plots with formulations as the
whole-plot factor. Each of the whole plots is split into four subplots (the blocks),
with rates of application as the subplot factor. Each subplot is split into seven
sub-subplots with flax as the sub-subplot factor. The sources of variation and
the degrees of freedom for their sums of squares are
Replications
F = formulations
Reps x F= Error (a)
R = Rates
RxF
Error (b)
V
VxF
VxR
VxRxF
Error (c)
3
2
6
3
6
27
6
12
18
36
216
He obtained his error (b) term for testing R and R x F by pooling the sums of
squares for reps x R and reps x R x F. The sum of squares for error (c) is
obtained by subtraction and consists of everything that has not been used
already. As usual, it is assumed that replications do not interact with any of the
other factors.
5.9 Computer Programs

The analysis of variance calculations for all these designs can readily be carried
out on a computer. It is relatively easy to write a program for a complete
factorial design with one observation per cell, and most computer centers have
one available. It is not necessary to write a special program for each design when
some factors are nested or when there are split plots. A few examples will suffice
to show how the simple program can be used for the more complex designs with
the help of a few additions with pencil and paper. We have already mentioned
this in regard to the Yates experiment with oats and manure.
1. A two-factor design with r observations per cell. We have merely to pretend
to ourselves that there are three factors. If y, fk is the kth observation at the ith
Exercises 103
level of A and the jth level of B, we suppose it to be at the k th level of a third
factor, C. The computer program then gives the correct sums of squares SA , SB ,
and SAB . The error term is obtained by adding Sc, S c , SBC, and SABC . If the
experiment is blocked, Sc is the sum of squares for blocks.
2. The two-stage nested design. Suppose that we have s sires, d dams per sire,
and n offspring per dam. Take three factors A, B, and C so that y ijk , the observation on the kth offspring from the jth dam by the ith sire, becomes the observation at the i th level of A, the j th level of B, and the kth level of C. The sums of
squares appearing in the analysis table are then:
Between sires SA
Between dams in sires SAB + SB
Between offspring in the
same litter Sc + SAC + SBC + SABC.
The idea involved here is that assuming a sires and b dams per sire, there are
a(b - 1) d.f. for dams in sires. We write this in terms of (a - 1) and (b - 1)
as a(b - 1) = ab - a = (a - 1)(b - 1) + (b - 1). Hence, SB(A) = SAB + SB .
3. The example with methods, batches, and temperatures. We add a fourth
factor, D, for pretended replications and then present the data to the computer
as a four-factor experiment in A, B, C, and D. We have SB(A) = SB + SAB ,
SBC(A) = SABC + SBC and Se = SD + SAD + SBD + SCD + SABD + SACD + SBCD
+ SABCD-
Exercises
1. The following data are yields of dry herbage (cwt/acre) from an experiment on
grazed grass [Widdowson et al., (1966)]. There were three fertilizers, nitrogen at
three levels, phosphorus at two levels, and potassium at two levels. Yields are
given for four years, 1961-1964.
Years
1961
1962
1963
1964
NoPo Ko
N0 P1 K0
N0 P0 K1
N0P1 K1
N1 P0 K0
N,P 1 Ko
N,Po Ki
N1 P1 K1
N2Po Ko
N2 P1 Ko
N2 Po K1
N2P1 K1
56.6
77.5
63.3
65.5
82.9
116.8
99.4
111.7
78.7
100.0
102.6
123.8
46.7
50.8
46.8
61.8
79.0
85.4
80.0
84.3
73.3
84.6
74.8
95.2
62.4
65.6
65.4
78.5
89.8
104.4
88.0
106.2
78.2
98.8
100.4
100.7
69.1
70.8
87.8
88.7
100.6
107.8
105.8
118.7
98.6
109.4
106.0
119.5
[Ch. 5.
We denote the levels of the three factors by No , Nl , N2 Po , P l , and Ko , K.

Analyze the data, assuming all factors to be fixed.
2. Peart (1968) reports lactation studies with ewes. In one of his experiments nine ewes
were divided into groups of three. In group 1 the ewes and their lambs were fed
freely during the lactation period; in the other groups two different methods of
restricting the food intake were used. He records the yields of residual milk in
grams for each ewe during the fourth through tenth weeks of lactation.
,
Group Ewe
Lactation Week
184
284
294
4
190
220
200
5
240
370
200
6
270
390
220
7
280
400
230
8
280
390
260
9
280
360
290
10
260
410
300
243
272
236
580
230
190
580
200
110
570
200
100
560
170
100
570
240
80
490
290
110
370
270
120
266
245
280
160
140
390
110
90
130
130
210
200
110
130
130
100
130
140
100
100
120
80
120
80
Analyze the data.

3. Gross et al. (1953) tested twelve fertilizer treatments (T) and four variaties of
alfalfa (V). They took four blocks; in each block there were forty-eight plots,
one for each alfalfa-fertilizer combination. Assuming that V and T are fixed but
that the blocks are random, set up the analysis of variance table and show which
ratio would be used to test the three hypotheses
4 = 0,
4 = 0,
QVT = 0 .
4. Aircraft are now used to sow seed and to spread fertilizer. Suppose that in a large
field we set up a grid pattern in the following way. Flying east and west, a plane
sows av strips, that is, a strips with each of v varieties. Flying north and south,
the pilot spreads fertilizer, b strips for each of t fertilizer treatments. The intersection of a sowing strip and a fertilizer strip is a plot. There are avbt of these, and n
observations are made from each of them. In both directions, the varieties or the
fertilizers are assigned to the strips at random. Set up the analysis of variance
table.
5. Lessman and Nyquist (1966) investigated the growth of an oil seed erop called
crambe. They tested eleven varieties in two locations, and in both locations
they made five replicates, each of which was divided into two blocks. In each
block four plants of each variety were sown. In one block the plants were
closely spaced; in the other they were widely spaced. There were thus 880
plants in all. Set up the analysis of variance table with the expectations of
mean squares regarding locations and plants as random, and regarding replicates, spacings, and varieties as fixed.
CHAPTER
Latin Square
Designs
In Chapter 3 we discussed the experiment in which five gasolines were tested in

cars and their road octane numbers observed. Four observations were to be made
on each gasoline. We began with the completely randomized experiment in
which twenty cars were chosen and four of them were assigned by some random
procedure to each gasoline. Later in the chapter we introduced the randomized
complete block design. There we took only four cars. Each gasoline was tested
once in each car. The car differences canceled out when gasolines were compared,
and the precision of the experiment for comparing gasolines was thereby improved. In an agricultural experiment, where we have reason to think that
the land becomes more fertile as we move from north to south, suitable blocks
might be strips of land running from east to west with one plot per block for
each of the treatments or varieties being tested. In either case, the treatments
are randomized within the blocks. In the case of the gasolines, the order in
which they are to be tested is chosen randomly for each car; the agronomist
assigns the plots within each block at random to the different varieties.
Let us change the experiment slightly so that the response we observe is the
rate of consumption in miles per gallon, with the added restriction that each
car can only make one run, i.e., test one gasoline, per day. It may be that conditions vary during the working week. Perhaps early in the week the weather is
cold and damp, but as the week wears on the sun appears and the weather
becomes warmer and drier. This could conceivably have some influence upon the
performance of the cars and upon the ratings that they give the gasolines.
Unless something is done about it, this day-to-day variation may ruin the whole
experiment. In an extreme situation suppose that all the cars use only gasoline A
on Monday, which is bright and Bunny, and that all the cars make their test of
105
[Ch. 6
106 Latin Square Designs
gasoline B on Friday when the temperature is twenty degrees lower and it is

raining steadily. We might gain some insight into the relative merits of the cars
used under a variety of conditions, but we should not be able to tell whether a
higher rating in miles per gallon of gasoline A over gasoline B results from the
differences between the gasolines or from the differences in the weather. We were
able to increase the precision of the earlier experiment when we removed the car
effects by blocking. We should be able to make a further improvement in the
precision if we could balance out the day effects. This calls for an experimental
scheme in which each gasoline is tested once in each car and once on each day
(or in each position in the block). This cannot be done with five days and only
four observations on each gasoline. Suppose that we allow ourselves a fifth car.
We can then use a Latin square design for the experiment.
A Latin square of side p, or a p x p Latin square, is an arrangement of p
letters, each repeated p times, in a square array of side p in such a manner that
each letter appears exactly once in each row and in each column. Four Latin
squares of side five are
ABCDE
BCDEA
CDEAB
DEABC
EABCD
ABCDE
CDEAB
EABCD
BCDEA
DEABC
ABCDE
DEABC
BCDEA
EABCD
CDEAB
ABCDE
EABCD
DEABC
CDEAB
BCDEA.
Suppose that we were to use the first of these squares and to let the rows denote
the cars, the columns the days, and the letters the gasolines. Then, on the
second day (column 2) the first car (row 1) uses gasoline B, the second car (row 2)
uses gasoline C, and so on. As we have written it, this square has a more systematic pattern than is needed or is desirable since for every car the gasolines follow
one another in alphabetical order. In practice, the experimenten should assign
the cars at random to the rows and the days at random to the columns (i.e., order
the columns by a random procedure).
6.1 Analysis
We assume a completely additive model; more will be said about this assumption later. Let Yifk denote the observation corresponding to the kth letter in the
ith row and the jth column, if there is one, i.e., to the observation on the kth
gasoline in the ith car on the jth day where 1 < i, j, k < p. The row, column,
and letter effects are denoted by pi , yf , and Tk , respectively, and for the present
we assume them to be fixed effects; is the grand mean and e1 ik is the random
error with the usual assumptions about homoscedasticity and normality. We then
have
Yiik
= 1 + Pi + '!
t
+ Tk
eilk
Analysis
Sec. 6.1]
107
We shall denote the row, column, and letter totals by Ti.., T. 1 ., and T.. k , and the
corresponding means by y,.., y.,., Y..k; the grand mean will be denoted by y....
In addition to the side conditions 1, p t = 0 and 1f y, = 0, we impose the
condition I k Tk = 0 . The normai equations are then easily solved, and the least
squares estimates of the parameters are
^ = Y..., Pc = Y:.. Y..., Yr = Y. i . Y... 1 fk = Y.. k y....
The analysis of variance table follows. Under normality, the sums of squares
are distributed independently. The appropriate test statistics for the hypotheses
of no row effects, no column effects, or no letter effects are the ratios of the
corresponding mean squares to the mean square for error. Under the null
hypotheses, these ratios will each have the tentral F distribution with (p 1)
and (p 1)(p 2) degrees of freedom. If, for example, the row effects are
random variables with mean zero and variance 4, we use the same breakdown
into sums of squares. The expected value of the mean square for rows is then
a 2 + p4R.
TABLE 6.1
Source
SS
d.f.
EMS
Rows
1, I J I k Y4lk C
p - 1 1, T,. C
Columns
p -1 Ii T.J. C
p1
p 1
a2 + p 1, p?/(p 1 )
a2 + p Ii yf/(p 1)
Letters
Error
p-1 Ik T.?k C
By subtraction
p1
(p 1)(p 2)
oz + p Ik Tkl(p 1)
02
Total
p2 1
In the example of the gasoline experiment, the purpose of using the Latin
square design was to improve the precision of the gasoline comparisons by
eliminating the car and day effects. In agricultural experiments we could use
such designs to eliminate fertility trends in two perpendicular directions, eastwest as well as north-south; this was the context in which they were first introduced, and they are sometimes called designs for the elimination of two-way
heterogeneity. It is customary to refer to the three classifications as rows,
columns, and treatments. The requirement that there should be the same
number of rows and columns as there are treatments can present a difficulty, and
alternative designs which involve incomplete blocks can be used. That, however,
is a topic for later discussion.
The Latin square can also be used in factorial experiments. A complete
factorial experiment with three factors each at five levels, which we may call a
5 3 experiment, calls for 125 points. If we take one of the Latin squares we can
let the rows denote the levels of the first factor, the columns the levels of the
second factor, and the letters the levels of the third factor. Then we have a set of
[Ch. 6
twenty-five out of the 125 points with the property that for any two of the
factors each level of one appears exactly once at each level of the other. In the
general case we have a (l/p)th fraction of the complete factorial design. This
reduction in the number of points can be very attractive. When p = 5, it
represents an 80 per cent reduction in the number of observations that have to be
made. The additive model that we have assumed may be rewritten
Yi1k=
N+ai +Ni+Yk+etlk,
where Y, ;k is the observation, if any, made at the ith level of the first factor, the
jth level of the second factor, and the kth level of the third factor, and ai , PJ and
Yk are the corresponding main effects. The analysis follows the same lines, and
we have tests for the main effects of the factors. Contrasts may be tested by any
of the usual multiple comparison methods. The reduction in the number of
points obliges us to assume in our model that there are no interactions. This
restriction may be serious; it is discussed in the next section.
This type of fractional factorial design may become necessary because of
budget restrictions. It may also be necessary if time is one of the factors, as it is
for the dairy cattle and crop rotation experiments that we shall discuss later in
this chapter. Consider, for example, an experiment to investigate the consumption of lubricating oil in buses. Suppose that there are three brands of oil, X, Y,
and Z. We could take three groups of new buses (A, B, and C) and conduct a
test over three months (I, II, and III), in which case an observation would show
the average consumption of a test oil by a group of buses in a month. We might
have group A use oil X for the first month, followed by oil Y for the second
month, and oil Z for the third month. We cannot use two oils simultaneously in
the same bus and so, even though we might think of a factorial experiment in the
three factors (oils, groups, and months, each at three levels), we can observe only
one of the three points AIX, AI Y, and AIZ. We are obliged to use one-third
fractions of the three-factor design, and the use of one or more Latin squares
would be appropriate.
,
6.2 Limitations of the Model

In the analysis of the Latin square design we assumed a completely additive
model with no interaction terms. As long as this model is justified, there is no
problem about the validity of the analysis. However, if there is, for example,
interaction between rows and columns, this interaction biases the estimates of
the treatment effects. This difficulty can be illustrated in the gasoline experiment
that has just been discussed. Using the first square, suppose that there is indeed
a row-column interaction (py) j to be added to the model. Over the set of rows
and the set of columns we have (py) ij = 0 for each j and Lf (py) u = 0 for
each i. These side conditions ensure that the interactions cancel out of E(fc).
Limitations of the Model 109
Sec. 6.2]
Let us consider, in particular, the gasoline that had been labeled treatment A.
We shall use subscripts 1, 2, 3, 4, and 5 for gasolines A, B, C, D, and E. We then
have
5
T.. 1 = 5
+ i=1
1 pi +j=1
yj + 5,r1 + (py)t>
4 j
5e.. 1
The terms 1 f yf and 1i pi vanish by virtue of the side conditions but the interaction terms that are actually involved are only a subset of the p 2 possible (py)
interactions, and there is no reason to expect them to vanish. They remain as an
unwanted and unknown nuisance. In this case the bias is
(py)11 + (PY)25 + (py)34 + (py)43 + (py)52
We might have chosen to label the gasoline in question treatment B rather than
treatment A, in which case the bias would have been
(PY)12 + (py)21 + (py)35 + (py)44 + (py)53,
or we might have retained the label A and used the second square, in which case
the bias would have been
(Py)11 + (PY)24 + (PY)32 + (py)45 + (py)53,
and so on. Thus, summing, or averaging, over all the labels and over all four
squares, it can be seen that the bias cancels out. In any particular square and for
any particular label there is a bias, but over the whole set it averages out. This
leads to the argument that the experimenter should choose his Latin square at
random from the set of all squares of side p and, having done so, assign the
letters at random to the treatments; by this procedure he would eliminate the
unwanted bias.
It is, alas, cold comfort to the experimenter, who has for his particular experiment one particular square, to know that if he could repeat his experiment
many times using randomly selected squares, the bias would on the average
disappear and to that extent, even in the presence of row-column interactions,
treatment differences would be estimable. He is saddled with the single square that
has been chosen, and his results are conditional upon the choice of that square.
Nor is the error term free from bias. Scheff (1959) shows that in the fixed
effects case the expectation of the mean square for error contains a term involving
the interactions, which is nonnegative and may or may not be appreciable.
The problem of bias in Latin squares was first noted by Neyman et al. (1935).
Wilk and Kempthorne (1957) consider a random effects model in a similar way.
In their model, rows, columns, and treatments are all considered as random
effects; they show that in this case, too, the mean squares for rows, columns,
treatments, and error are all biased if the simple additive model does not hold.
Kempthorne (1952) and Scheff (1959) discuss randomization models as alternatives to the models presented here.
[Ch. 6
6.3 Graeco-Latin Squares

Suppose that we take a Latin square and superimpose upon it a second square
with the treatments denoted by Greek letters. If the two squares have the property that each Latin letter coincides exactly once with each Greek letter when
the squares are superimposed, then they are said to be orthogonal. The combined squares are said to form a Graeco-Latin square. Adding a third square
orthogonal to each of the others gives a hyper-Graeco-Latin square. The four
squares of side five that were given at the beginning of this chapter form a set of
mutually orthogonal squares.
Graeco-Latin squares can be used to provide designs for four factors (rows,
columns, Latin letters, Greek letters), each at p levels in p 2 runs. Thus, in our
example of the gasoline experiment, a fourth factor, drivers, might be introduced
by combining the first two squares and using the Greek letters to represent the
drivers:
Aa Bfl Cy DS Ee
By
C8
Ce
Da ES
Ay
B8
Dfl Ey AS
Be
Ca
Ba
Cp
Dy
ES
Ae
De Ea
AP
On the fourth day of the test, driver a uses gasoline E in car 2, driver P uses
gasoline C in car 5, driver y uses gasoline A in car 3, driver 8 uses gasoline D in
car 1, and driver e uses gasoline B in car 4. The analysis with the additive model
follows the same procedure as the analysis of the Latin square design. There is an
additional sum of squares for Greek letters with (p 1) degrees of freedom and
there are only (p 1) (p 3) degrees of freedom for error.
Indeed, we could go further by having five different test courses and by having
each of the cars driven over a different course each day. Courses could be added
as a fifth factor by incorporating the third Latin square and using lower case
letters a, b, c, d, and e to denote the five courses. That would leave only (p 1)
x (p 4) = 4 degrees of freedom for error. A sixth factor, perhaps predetermined average driving speeds, could be added by incorporating the fourth
square and using Arabic numerals to denote the levels. There would then be no
degrees of freedom left for error, and, without some prior estimate of the
variance, testing hypotheses would present some difliculties. However, there is a
chance that perhaps one or two of the main effects might stand out much larger
than the others, which would point out the way to further experimentation. It
may be observed that if the experiment is going to need five cars, each on all-day
trips, we are certainly going to need at least five drivers. It does not, however,
follow that we are obliged to add drivers as a fourth factor and to use a GraecoLatin square. It depends upon the nature of the interest in the cars and drivers.
Sec. 6.4]
Sets of Orthogonal Latin Squares 111
If we wish to investigate whether there are car differences and whether there are
driver differences, then we must add drivers as a fourth factor and, unless we are
prepared to increase the size of the experiment, the Graeco-Latin square would
be appropriate, with only eight degrees of freedom for error. However, if we are
interested only in comparing gasolines, and thus regard car and driver differences, if any, as nuisances to be eliminated, we should be better advised to use
the same driver with the same car throughout the experiment. Then the driver
effects and the car effects would both be absorbed in the row effects, which
would now denote car-driver combinations, and we should have twelve degrees
of freedom for error.
The Graeco-Latin squares and the hyper-Graeco-Latin squares are, like the
Latin squares, fractional factorial designs. The Graeco-Latin square is a (1/p 2)
fraction of the p 4 design, i.e., the factorial with four factors, each at p levels.
The square with no degrees of freedom left for error is a (1 /p' -1 ) fraction of the
pP+l
factorial. Such a fraction, in which every degree of freedom is used for
estimation, is said to be a saturated fraction.
This ability to add extra factors without increasing the size of the experiment
appears attractive on the surface, but the experimenter should note that not
only does the number of degrees of freedom for error diminish as more factors
are added, but also the problems with nonadditivity, which are so bothersome
in the Latin square layout, become even more troublesome. Graeco-Latin
squares exist whenever p is a prime, or a power of a prime, and for some other
values. Complete sets of p 1 mutually orthogonal squares for p = 3, 4, 5, 7, 8,
and 9 are given in the Fisher and Yates tables.
6.4 Sets of Orthogonal Latin Squares

Now we present a method, due to Bose (1938), of obtaining sets of p 1
mutually orthogonal Latin squares of side p when p is a prime or a power of a
prime. We associate each of the treatments with an element of the Galois field
of p = s n elements, GF(s n), in a one-to-one correspondence. The elements of
the field are ordered as g o = 0, g 1 = 1, g 2 = x, g3 = x 2 , ... , gp _ 1 = X - 2
where x is a primitive element of the field, i.e., x is an element such that xp -1 = 1
and there is no other power x = 1, 0 < q < p. Then the addition table forms
a Latin square, and the other squares are obtained by rotating cyclically all
the rows except the first.
EXAMPLE 1. p = 4. GF(2 2 ) has elements 0, 1, x, and 1 + x. Arithmetic is
carried out mod 2, and x is a primitive element. The irreducible polynomial of
the second degree in the field is x 2 + x + 1. Equating this polynomial to zero,
and recalling that 1 - 1 (mod 2), we have x 2 = 1 + x, and x 3 = x + x 2 = 1.
In the addition table, the first row consists of the elements 0, 1, x, and 1 + x;
[Ch. 6
the second row is 0 + 1 = 1, 1 + 1 = 0, x + 1 = 1 + x, 1 + x + 1 = x; and

so on. The complete table is
0
1
x
1
x
l+x
0
l+ x x
1
1+x 0
1+x x
1
0
Writing A, B, C, and D for the elements 0, 1, x, and 1 + x, respectively, and

rotating the rows other than the first cyclically, we obtain the following three
squares. They form a mutually orthogonal set of squares of side 4.
ABCD
ABCD
ABCD
BADC
CDAB
DCBA
CDAB
DCBA
BADC
DCBA
BADC
CDAB
The first row of each square is the same. The second row of the second square
is the same as the third row of the first square. The third row of the second square
is the fourth row of the first square. The last row of the second square is the
second row of the first square. The third square is obtained from the second
square in the same way.
We now show that Bose's method is valid in the general case. We define the ith
square of the set to be the square that has gi = x i -1 as the first element of the
second row. The first element of the (m + 1)th row is gi .xm -1 = gi gm . We write
the ith square, therefore, as
0
...
gi
g+1
... gi + gP -1
gig2
gig2
gigP-1
gigP-1
+1
+1
gP-1
+ gP-1
...
gig2
...
gigp-1 + gP-1
We need to establish two properties:

Property 1. Such a square is a Latin square.
Suppose that it is not a Latin square and that two of the elements that
appear in the (q + 1)th row are identical, i.e., for some pair t, u, (t # u), gi gq +
gt = gig, + g,. This implies gi = g, which is false because the elements of the
field are distinct.
PROOF.
Sets of Orthogonal Latin Squares 113
Sec. 6.4]
Property 2. The squares in the set are mutually orthogonal.

Suppose that the ith and jth squares are not orthogonal. This means
that when the jth square is superimposed on the ith square, one pair of elements
occurs together in two of the cells. Suppose that these cells are in the (q + 1)th
row and the (t + 1)th column and in the (r + 1)th row and the (u + l)th
column where q r, t # u. In these two cells the elements of the i th square
coincide to give
PROOF.
gjgq + gt = g{g. + g.,
and the elements of the jth square coincide to give

gjgq + gt = grg. + g
Subtracting, we have (g{ gj)g q = (g{ g f)g,. This would imply that either
g = g. or g Q = g, in contradiction. E
This method was also presented by Stevens (1939). In his paper, Stevens
points out that the orthogonal set for p = 9 given by Yates in the 1948 edition
of the Fisher and Yates tables is not isomorphic to the set that Stevens himself
obtained using Galois fields, i.e., the one set cannot be obtained from the other
by a reassignment of the letters to the treatments.
It is easily verified that if p is a prime, rather than sn where n > 1, it is not
necessary to find a primitive element of GF(p) in order to obtain an initial
square. Bose points out that we can, instead, work with the residue classes
(mod p) and write the rows of the initial square in any order we wish.
{
EXAMPLE 2. p = 5. We shall now give two possible initial squares of side 5.

The first of these is derived by the previous procedure. The elements of GF(5)
are 0, 1, 2, 3, and 4. Arithmetic is mod 5 and x = 2 is a primitive element. Its
powers are x 2 = 4, x 3 = 3, and x 4 = 1. The first row (and column) of the
addition table as we write it to form the initial square is, therefore, 0, 1, 2, 4,
and 3.
The second initial square is obtained by writing the residue classes (mod 5) in
the usual order 0, 1, 2, 3, and 4 and forming the addition table. The ith square
of this set is obtained by re-ordering the rows of the initial square so that the
first column is 0, i, 2i, 3i,.... Writing A, B, C, D, and E for 0, 1, 2, 3, and 4,
respectively, produces the set of mutually orthogonal squares of side 5 given at
the beginning of the chapter.
01243
12304
23410
40132
34021
01234
12340
23401
34012
40123
[Ch. 6
There are no similar methods of obtaining sets of p 1 mutually orthogonal

squares of side p when p is not a prime or a power of a prime. Indeed, Bruck and
Ryser (1949) have shown that if p - 1 or 2 (mod 4), and if the square free part
of p contains at least one prime factor q - 3 (mod 4), then no orthogonal set of
p 1 squares exists. It had been conjectured by Euler that if p = 2 (mod 4),
there did not exist a pair of orthogonal squares, and this conjecture was verified
for the case p = 6 by Tarry (1901). However, Bose, Shrikhande, and Parker, in a
series of papers, have shown that if p - 2 (mod 4) and p > 6, a pair of orthogonal squares always exists. A pair of orthogonal squares found by them (1960)
for p = 10 follows.
A pair of orthogonal Latin squares of side 10.
00
86
95
59
73
68
37
14
21
42
47
11
80
96
69
74
08
25
32
53
18
57
22
81
90
09
75
36
43
64
76
28
67
33
82
91
19
40
54
05
29
70
38
07
44
83
92
51
65
16
93
39
71
48
17
55
84
62
06
20
85
94
49
72
58
27
66
03
10
31
34
45
56
60
01
12
23
77
89
98
61
02
13
24
35
46
50
88
97
79
52
63
04
15
26
30
41
99
78
87
6.5 Experiments Involving Several Squares

The single Latin square contains only p observations with each treatment.
If the experimenter wishes to have more replications of each treatment, he may
find it desirable to use several squares. The following analysis is valid whether the
squares chosen are the same or different. We suppose that there are s squares.
Let Yhifk denote the observation, if there is one, made upon the kth treatment
in the ith row and the jth column of the hth square. The model taken is
Yh(Jk = 9 + 7 rh + Pi(h) + y1(h) + Tk + ( IT T)hk + ehilk,
where 7rh is the effect of the hth square and (1r r) hk represents the interaction
between squares and treatments; P, (h) and yi(h) are the row and column effects in
the hth square. The side conditions on the parameters are h 7rh = 0 , -y^ P((h) = 0
for each h, Ii y f(h) = 0 for each h, I k rk = 0 , (ITT) hk = 0 for each h, and
lh (1TT)hk = 0 for each k.
With these side conditions, the estimates of the parameters are
^L = Y....,
17h = Yh... Y....,
Pi(h) = Y,a.. Yh,
fk = Y...k Y....,
y!(h) = Yh! Yn...,
()hk = Yh..k Yh... Y...k + Y.....
Change-Over Designs 115
Sec. 6.6]
The analysis of variance table is
TABLE 6.2
SS
Source
Total
Squares
Treatments
TxS
Rows in squares
Columns in squares
Residual
d.f.
spa 1
lh Gt G1 Gk Yhjk C
lh
T,a.../p 2 C = Ss
C = St
>hIkThk /pC S8 S,
Tnt.. /p l h Th ../pa
lh 2 1 T,? f .lP I h T.../p 2
By subtraction
Ik T.?. k /(SP)
s 1
p1
(s-1)(p-1)
s(p 1)
s(p 1)
s(p l)(p 2)
6.6 Change-Over Designs

In the previous section we considered the s squares as separate squares. We
may also consider them as being arranged in a rectangle with p rows and sp
columns. This situation typically occurs when time, in the form of periods, is a
factor in the experiment. In experiments with dairy cattle, for example, we may
test p treatments (such as kinds of feed) overp periods using sp cows. Each cow has
each feed for one of the p periods, and the schedules are arranged so that during
each period each feed is given to s cows. In the rectangular arrangement, the
periods correspond to rows and the cows to columns. Alternatively, we may
have p subjects receiving p treatments over sp periods; each subject receives each
treatment for s of the periods; during each period each of the treatments is used
on one of the subjects. In either case, we may let
Ytik =
Pt
y1
k + etik,
where 1 <i<p,1 <j< sp, and 1 <k<p.

TABLE 6.3
Source
Total
Rows
Columns
Treatments
Residual
SS
21211k ytlk C
1, T./(sp) C
T. f /p C
Ik T.? k /(sp) C
By subtraction
d.f.
sp e
p1
sp 1
p1
(sp 2)(p 1)
EMS
+ sp I, pi l(p 1)
+ p 1, y2 /(sp 1)
02 + sp Ik Tk/(p 1)
a2
02
02
EXAMPLE. The following example is taken from Patterson (1950). In the

experiment eighteen cows were fed, for consecutive five week periods, three
rations: A, good hay; B, poor hay; and C, straw. The cows were divided into six
[Ch. 6
blocks of three cows each, and a Latin square design was used for each block.
The response recorded was the yield of milk in pounds for each five-week
period. The analysis will be developed further in the next section.
We take for our model an adaptation of the previous model to include the
blocking, and we shall now use Latin letters rather than Greek letters to denote
the parameters, including the mean. (This is done merely for conveniente
and there is no deeper significance to the change.) We have
Yhtjk = M + ph + C{(p
bf + tk
ehiik,
where Yh{fk is the yield during the hth period of the ith cow in the jth block;
during that time the cow was receiving the kth treatment; Ph , c{(j) , b5 , and tk are
the period, cow, block, and treatment effects.
Period
1
2
3
Total of Cow
Total of Total of
Period
Block 2
Period
Block 1
A 768
B600
C411
1779
B 662
C 515
A 506
1683
C 731
A 680
B 525
1936
2161
1795
1442
5398
A 669
C 550
B 416
1635
Block 3
1
2
3
Total of Cow
A 1091 B 1234 C 1300

B 798 C 902 A 1297
C 534 A 869 B 962
2423
3005
3559
Total of Cow
A 941
B 718
C 548
2207
B 794
C 603
A 613
2010
3625
2997
2365
8987
A 1105
C 712
B 453
2270
B 891
A 830
C 629
2350
C 779
A 718
B515
2012
2514
2039
1676
6229
A 933
C 658
B576
2167
B 724
A 649
C 496
1869
TABLE 6.4
Source
SS
d.f.
Total
Blocks
Cows in blocks
Periods
2,756,705
1,392,534
318,242
814,223
57,791
121,147
52,768
53
5
12
2
Treatments
Errors
1752
1421
1064
4237
C 859
B 617
A 597
2073
2855
2159
1679
6693
C 749
B594
A 612
1955
2406
1901
1684
5991
Block 6
PxB
C 624
B 462
A 426
1512
Block 4
Block 5
1
2
3
B 459
A 409
C 222
1090
10
2
22
Sec. 6.7]
Change-Over Designs Balanced for Residual Effects 117
6.7 Change-Over Designs Balanced for Residual Effects

The change-over design that has just been discussed in the example is balanced
for treatments. It is also a balanced design for the comparison or elimination of
residual effects, and it was as an illustration of this that Patterson (1950) presented the data. When a cow that received treatment A for the first period is
switched to treatment B for the second period, the yield in the second period may
reflect not just the main effect of treatment B, but also the residual effect, if there
is one, of treatment A.
In the design, blocks I and II together form a pair of orthogonal Latin squares,
ABC ABC
BCA CAB
CA B B CA
and the same is true for blocks III and IV and for blocks V and VI. In the pair of
squares, each treatment is used once on each of the six cows. There are no
residual effects during the first period. In the subsequent periods, the residual
effect of each treatment appears four times. In each of the last two periods, the
main (or direct) effect of each treatment is observed in two cows, and the
residual effect of each treatment is also observed in two cows. Since no treatment can follow itself, no yield reflects both the main effect and the residual
effect of the same treatment. This problem was first considered by Cochran et al.
(1941). Further developments are due to Patterson (1950) and (1952), and Lucas
(1951) and (1957). The designs are also applicable to experiments, such as those
involving the rotation of crops, in which several treatments are applied in
succession to the same experimental units.
In the general case with the dairy cattle, we make use of sp cows over p periods,
and the feeding schedule is arranged in such a way that
(i) Each cow has each treatment for one period;
(ii) In each period each treatment is used on s cows;
(iii) Each treatment follows every other treatment s times.
We shall not consider blocks in our analysis; they present only a minor complication. The discussion given here is based on that of Patterson (1950).
A set of s of the cows receive the qth treatment in the last period, and therefore never experience the residual effect of that treatment. We call the kth of
that subset of s cows the (kq)th cow. Let Y hijkq denote the yield of the (kq)th cow
during the hth period in which she received the ith treatment, preceded, if
h > 1, by the jth treatment. Then
Yhilk4
= m + dh + ti +
r1 + Ck9
ehi5k9,
[Ch. 6
where dh , ti , ri , and C kq are the period, (direct) treatment, residual treatment, and
cow effects, respectively. The side conditions on the parameters are
1 di, =01
1 t1 = 0,
r1=0,
11Ck9=
0.
k q
In the normai equations we denote the various totals by G, Dh , Ti , R5 , and Ckq ,

and we have
G = sp 2ih,
Dh = spin + spJh ,
Ti =spin +sp$i sPi ,

R1 =s(p-1)z+s(p-1)r"5 sd1 ^ckf si5 ,
k
Ckq = p)h + PCk q r y .
We now write {C} i for the quantity G I k Cki; {C} i is the sum of all the yields
on all the cows that experience the ith residual effect. We have
{C} i = sp(p 1)m p 5. cki si.
k
Let Ti and Ri denote the adjusted totals

Ti = Ti p - 1 G = spti sr"i,
Ri = Ri p -1 {C} i + p - 'D 1 p -2 G = p -1s(p 2 p 1)fi s?i .

Solving these equations we have
_ (p 2 p 1)T{ + pR{
+ 1 )(P 2) '
ti sp(p
ri
Ti + pRi .
s(P + 1 )(p 2)
We turn now to the analysis of variance table.

Whether or not there are direct or residual treatment effects included in the
model, we have m = G/(sp 2 ) and cd,, = (Dh spm)/(sp). If, however, there are
no residual effects in the model, the least squares estimate of ckq becomes
(Ckq pm)/p, which is not the same as the estimate with our present model.
We shall consider four models:
Oo: E(Yhtjk9
) = m + dh + Ckq,
1: E(Yhiika) = m + dh + C kq + t i + r5 ,
^2 : E (Yhilk9 )
= m + dh + Ckq + ti,
: E(Yh15k4)
= m +
dh + Ck q + r!.
Sec. 6.7]
Change-Over Designs Balanced for Residual Effects 119
Let S denote the sum of squares for regression under the nth model including
the correction for the mean, G 2 /(sp 2 ). Then
So
Dhh /(sP) + Ga
S 1
Ckq
/P G 2 /(sp 2),
k q
Dn/(sp) +
Ckgekq
k q
and

S
l So = Tt it + Rt rt +
t
Ti lt
Tilt +
t
Ri 'i,
t
(CkgCkq Ckj
jp) + G 2 /N
k q
Ri r"i
r"t{C}tlp =
i
Ti fi
t
Rir"i
t
_ {(p 2 p 1)T{ 2 + 2pTtRi + p 2 Ri 2}/{sp(p + 1)(p 2)}.

Under SZ 2 , we have for the least squares estimate of t i
ii = Til(sP),
so that
S2 = So + Tt ft = So + T T, l (sp) = So + T{ 2 /(sp),
and the sum of squares for direct effects ignoring residual effects is
S2 S o
= : Ti 2 /(sp)
with (p 1) d.f., and the sum of squares for residual effects eliminating direct
effects (i.e., for residual effects over and above the sum for direct effects alone)
is
Sl S2 = (Ti + PR) 2 /sp(P + 1 )(p 2) = s(p + 1 )(P 2) ri/p.

t
Under S2 3i we have s(p 2 p 1)r1 = pR{, and
PS3 =
Ckq G 2 /(sP) r"i{C} t + p rtRt^
Dh/s +
h
k q
so that the sum of squares for residual effects ignoring direct effects is
S3 So
r"i(Rt p -1 {C} i ) = p Ri 2 /s(p 2 p 1),
=
t
and the sum of squares for direct effects, eliminating residual effects is
sp(P + 1 )(P 2)y f2

Si S3 = 1i ((P 2 p 1 )Ti + PR) 2 _
p2 p 1
sp(P + 1 )(P 2)(P 2 p 1)
[Ch. 6
Thus, to test the hypothesis that there are no residual effects we take the test
statistic F = (S, S2 )/ Me , which has under the null hypothesis the F[(p 1), ]
distribution. In the particular example that we have been considering, a term
for block x period interaction has been removed. In the general case there are
(p 1) d.f. each for periods, direct effects, and residual effects, and (sp 1)
d.f. between cows; this leaves = spe (s + 3)p + 3. If the sum of squares
for block x period interaction is allo taken out there is a further reduction of
(s 1)(p 1) d.f.
In the example the sum of squares listed for treatments is the sum of squares
for direct effects neglecting residual effects. To obtain the sum of squares for
residual effects eliminating direct effects we have, where D l = 15313,
T,
A
13713
B
12000
C 11822
37535
Ti
Rt
{C}i
3R
T + 3Ri
1201.33
511.67
689.67
7385
7036
7801
22222
25297
23956
25817
340.67
46.67
387.33
860.67
558.33
302.33
Then (S 1 S2 ) = [( 860.67) 2 + ( 558.33) 2 + ( 302.33) 2 ]/72 = 15,888. The sum of

squares for error is 52,768 15,888 = 36,880, and Me = 1,844.
6.8 Designs for Experiments with Residual Effects

The design in the example just considered was a pair of orthogonal Latin
squares repeated three times. Williams (1949) has investigated designs for this
type of experiment. He has shown that designs balanced for the residual effect of
the preceding treatment and satisfying the requirements listed in the previous
section can be obtained with only p cows, or replications, if p is even, and with 2p
cows ifp is odd, as was the case in the example. His designs are based upon Latin
squares.
EXAMPLE 1. p even. We let the treatments be represented by the residue
classes, mod p, and we illustrate Williams' method by the following example with
p = 6. The rows correspond to periods and the columns to cows.
0
1
5
2
4
3
1 2 3 4
2 3 4 5
0 1 2 3
3 4 5 0
5 0 1 2
4 5 0 1
5
0
4
1
3
Exercises 121
The first step in Williams' procedure is to find a suitable initial (first) column.
Then, each subsequent column is obtained from the previous column by adding
one to each entry and reducing mod p. The elements of the initial column in the
example are, in order, 0, 1, 5, 2, 4, and 3. The differences between adjacent pairs,
reduced mod 6, are 1, 4, 3, 2, and 5. Every number other than zero appears
exactly once, and it follows that as we take the several columns each treatment
will be preceded exactly once by each of the other treatments. Thus, treatment 3
will be preceded by treatment 2 in the column where 3 appears in the second
row because the differente between the second and first entries in any column is
1 = 3 - 2.
We leek, therefore, for the initial column a permutation of the elements
0, 1, .. ., p - 1 such that in the p - 1 differences between adjacent elements
each nonzero member, mod p, appears exactly once. When p is even, such an
arrangement is given by 0, 1, p - 1, 2, p - 2, 3, p - 3.....p/2.
This is not, however, a unique solution. If we multiply each element of the
initial column by an integer that is prime to p, and reduce mod p, we obtain
another initial column. In the example, multiplying by 5 gives the new column
0, 5, 1, 4, 2, and 3. Another initial column for p = 6 is 0, 2, 1, 4, 5, and 3.
Multiplying this by 5 gives 0, 4, 5, 2, 1, and 3.
EXAMPLE 2. p odd. When p is odd, the sum p(p - 1)/2 of the elements is
divisible by p, and if each difference is to occur once in a column beginning with
zero the last entry would also be zero, which is not satisfactory. We have, however, seen an example of a scheme for p = 3 with two Latin squares in which
half the differences appeared in one square and half in the other.
Williams makes use of two Latin squares of side p. One of the squares is
derived from an initial column in which each of the odd elements appears
twice as a differente. The other initial column gives each even difference twice.
To obtain the odd differences, we take the initial column 0, 1, p - 1, 2, p - 2, ... ,
(p - 1)/2, (p + 1)/2. For the even differences, we take the same initial column,
but we write it in reverse order.
Using this procedure we obtain for p = 5 the following design using ten
cows.
01234
34012
12340
23401
40123
40123
23401
12340
01234
34012
In a subsequent paper Williams (1950) goes on to consider designs balanced
with respect to the residual effects of the previous two treatments.
Exercises
1. Derive the following missing plot formula for the Latin square design; G', R', C',
and L' denote the sum of the (p 2 - 1) actual observations taken and (p - 1)
[Ch. 6
actual observations occurring in the same row, column, and with the same letter as
the missing plot.
x=
p(R' + C' + L') 2G '

(p 1 )(P 2)
Show that if the missing plot received the ith treatment and is replaced by a value
calculated by the previous formula then
V(Th T)=Qa
\P
+ (P 1)(p - 2)/
2. Youden and Hunter (1955) suggested modifying the Latin square design in the
following way to obtain an unbiased estimate of the error. They chose a square
in which each variety appeared exactly once on the diagonal and duplicated those
plots, thus obtaining an estimate of error between duplicates with p degrees of
freedom. Because the design is no longer orthogonal the analysis calls for testing
subhypotheses: rows, columns (after rows), and treatments (after columns and
rows). Derive the analysis.
3. Suppose that one entire row is omitted from a Latin square design and modify the
analysis accordingly.
4. For the data for the dairy cattle experiment, compute
(i) the sum of squares for residual effects ignoring direct effects, and
(ii) the sum of squares for direct effects eliminating the residual effects.
CHAPTER
Factors with
Two or
Three Levels;
Confounding
In the previous chapters we have developed the analysis of variance for the
general multifactorial experiment. We finished up with algorithms for obtaining
degrees of freedom, sums of squares, and expected mean squares for fixed,
mixed, or random models. The only blocking situation considered was that of
the randomized complete block, in which case the blocks appeared in the
algorithms as the levels of another (random) factor.
In this chapter we shall take another look at the factorial experiment. The
2n and 3' factorials will be introduced. We shall also consider blocking in the
case in which the blocks are not large enough to contain all the treatments, a
situation which we saw for the first time in the split-plot design where each whole
plot contained only one level of each of the main treatments. We shall assume
throughout that all the factors have fixed effects, that the block effects are
random effects with mean zero and variance aB, uncorrelated with each other
or with the error terms, and that there is no interaction between blocks and
treatments. We mean by the last statement that if y j denotes an observation on
the ith treatment in the jth block, and y denotes an observation on the hth
treatment in the same block, then, apart from random error, the difference
Yhj y ij remains the same for all values of j.
The reader will not fail to notice the enormous contribution made by Yates
to the development of the topics covered in this chapter. In particular, numerous
references will be found to his pioneer monograph, The Design and Analysis of
Factorial Experiments, which was published in 1937 as Technical Communication
35 by the Imperial Bureau of Soil Science, Harpenden, England.
123
124 Factors with Two or Three Levels; Confounding
[Ch. 7
7.1 Factors at Two Levels

Consider an experiment with two factors A and B, each at two levels. There
are four sets of experimental conditions: 1ow A, low B; high A, low B; low A,
high B; high A, high B. Yates calls these the four treatment combinations, and
denotes them by (1), a, b, and ab; the letter a appears whenever the factor A
is at its high level. The use of the words high and low is arbitrary and has no
particular significance in the analysis. With a quantitative factor, one level is
clearly higher than the other; with a qualitative factor, we arbitrarily call one
level high and the other low. We shall also use the notation (1), a, b, and ab to
denote the mean responses at those treatment combinations. Thus, the symbol a
is used in two senses: to denote the set of conditions having A at its high level
and B at its low level, and also the average of the observations taken under those
conditions. The notation can be extended to any number of factors. With four
factors, the treatment combination cd has C and D at their high levels and A and
B at their low levels. In the earlier chapters, the response cd would have been
denoted by something like Y1122.
The usual model has
Ysik = N- + at
+ Pi + ( aP)ti + etirc.
(We shall use P later to denote something else.)

The side condition a l + a 2 = 0 gives al = a2i we write a 2
= y... and & = Y2.. Y... =(Y2.. yl..)/ 2 , so that
= a.
Then
4= ab+a+b+(1),
4= 2ab+2aabab(1)=ab+ab(1).
The corresponding sums of squares are C = G 2 /N, as usual, and SA = N& 2
where N is the total number of observations and we assume that each of the
treatment combinations is observed the same number of times. Similarly,
writing N = NZ = P1 and ( aN)11 = (a)22 = (aP)12 = ( aa)21 = (aN), we
P)
also have SB =NO 2 and SAB = N( 2 .

We call ab + a b (1) the A contrast. The B and AB contrasts are
4ft= aba+b(1),
4()= abab+(1).
Symbolically they can be written (a 1)(b + 1), (a + 1)(b 1), and (a 1)

x (b 1). The three contrasts are mutually orthogonal. Henceforth we shall
assume, in the absence of a statement to the contrary, that we have only one
observation on each treatment combination.
An alternative approach, which we shall use later, is to associate with the
factors A and B two coordinates, x l and x 2 ; x l takes the value + 1 when A is at
Factors at Two Levels 125
Sec. 7.1 ]
its high level and the value 1 when A is at its low level; similarly, x 2 = + 1
at high B and x2 = 1 at low B. The four treatment combinations are thus
(-1, 1), (+ 1, 1), (-1, + 1), and (+ 1, + 1). Since the coordinates are 1,
it is convenient to write ( ), (+ ), ( +), (+ +). We can now fit the
model
YiJ = Po + flixi + P2x2 + R12x1x2 + e1,
and we shall have Po = , P1 = a, P2 = P, P12 = ( aP).

The interaction contrast may be written in two ways, both of which provide
insights into its nature. Writing it as {ab b a + ( l)} we see it as the difference between the increase in response obtained by changing the levels of A
at high B and the corresponding increase at low B. Writing it as {ab a}
{b ( 1)} shows it as the difference between the B effects at the two levels of A.
In either case, we have the same story: the interaction measures the failure of the
main effect of A (or B) to remain constant over the levels of B (or A). An
interesting question then arises, one that is also relevant when the factors have
more than two levels. What does it mean to have a significant A effect when there
is also a significant AB interaction?
EXAMPLE. Consider the following three situations:

(i)
(ii)
(iii)

(1) 40 a 60
(1) 40 a 80
(1) 60 a 40

b 60 ab 80
b 60 ab 60
b 60 ab 80.
Situation (i) is simple. At either level of B, increasing the level of A from low to
high causes an increase of 20 units in the response. We have 4 = 40 and
4(a) = 0. In situation (ii) we stilt have 4& = 40, but now 4(aa) _ 40. On the
average, assuming that a larger response is more desirable than a smaller one,
it is a good idea to work at the high level of A. Actually, however, there is no
difference between the responses at the two levels of A at high B, and all the
forty units difference- that went into 4& was found at low B. In situation (iii)
= 0 and 4( a) = 40. It is not correct to say that changing the level of A has no
effect on the response. Changing the level of A from low to high produces a
decrease in the response at low B and an equal increase in the response at high B.
In general we must take the point of view that a significant value of , or of the
ratio M4 /Me indicates that the responses are, on the average, higher at one level
of A than they are at the other. A significant interaction indicates that looking at
the average change in response is not enough, and that the effect of changing
levels of A is not the same at all levels of B; equivalently, the effect of changing
levels of B is not the same at all levels of A.
If we add a third factor, C, also at two levels, we introduce a three-factor
interaction. The model becomes
,
Ytik = N + ai + fl + Yk + ( X)ij + ( aM)uo + (fY),k + (afY)iik + efjk,
[Ch. 7
and, proceeding as before, we have
811=abc+ab+ac+bc+a+b+c+(1),
8&=abc+ab+acbc+abc(1),
8 (a) = abc + ab ac bc a b + c + (1),
which may be written symbolically as (a + 1)(b + 1)(c + 1), (a 1)(b + 1)

x (c + l), and (a 1)(b 1)(c + 1).
The eight interaction terms (a!y), ik have the same absolute value. If they are
not all zero, four of them are positive and the other four are negative. The side
conditions
11(aPY)tik = 0,
11(ajY)>k = 0,
4 j
^1, (aSY)tlk = 0
!
lead to
(aNY)22a =
("PY)221 =
(aaY)212 =
(afY)122 = (aaY)211 = (aPY)1a1
= (aaY)112 = (afY)111 = (afY),
whence
8(afly)=abcbcacab+a+b+c(1).
Alternatively, we may write the model as
E(y) = Ro + Rlxl + 192x2 + 9 3x3 + P12x1x2
+ P13x1x3 + P23x2x3 + N123x1x2x3
Then 8P123 = (responses at those points with x 1 x 2 x3 = 1) Y (responses at

those points with x 1 x2 x3 = 1) = 8(apy). Symbolically we have 8(aay) _
8$123 = (a
1)(b 1)(c 1).
There are several notations that can be used to denote the ABC contrasts and,
similarly, other contrasts. We shall use either {ABC} or {x 1 x2x3 }. We shall also
use {x 1 x 2x3 = + 1} and {x l x2x3 = 1} to denote the sums of the responses at
the points at which x 1 x 2x3 = +1 or 1, respectively. Then {ABC} = {x 1 x2 x3} _
{x1 x2 x3 = + 1} {x 1 x2x3 = 1}. Yates has denoted ABC by A.B.C..
The ABC contrast can be written (abc ac bc + c) (ab a b + (1)),
which is the difference between the AB contrast at high C and the AB contrast
at low C. It may also be considered as the difference between the AC contrasts at
the two levels of B, or the difference between the BC contrasts at the two levels
of A.
Yates' Algorithm and the General 2n Design 127
Sec. 7.2]
The variance of the estimate of any effect, either main effect or interaction, is
N -1 0 2 . Each point contributes with the same weight to the estimate of each
effect. In particular, so far as the estimation of a is concerned, we have lost
nothing in precision by adding B and C to the experiment because the variance
of P 1 in this experiment is the same as it would have been if we had not had the
factors B and C and had made instead only four runs at low A and four runs at
high A. Alternatively, we may take the point of view that the addition of the
extra factor C to the experiment also adds two more points at high A and at low
A, and two more at high B and low B for estimating A, B, and AB. This is an
example of what is sometimes referred to as the hidden replication property of
the factorial experiment.
We have in the 2 3 experiment four direct comparisons to estimate the main
effect of A, namely, a (1), ab b, ac c, and abc bc. We obtain a vivid
pictorial representation when the experimental points are portrayed as the
vertices of a cube with coordinates ( 1, 1, 1).
bc 24
abc 40
c 20
30
(1) 12
a 22
The four constituent A contrasts are 10, 12, 10, and 16. The overall A contrast
is the sum, 48, with an average of 12. At low B we have A contrasts 10 and 10;
at high B we have 12 and 16, giving an interaction contrast 28 20 = 8. At low
C the AB contrast is 12 10 = 2; at high C the AB contrast is 16 10 = 6;
the ABC contrast is 6 2 = 4.
7.2 Yates' Algorithm and the General 2n Design

The 2' design is the complete factorial for n factors, each at two levels. We
have now seen enough to be able to tell how the derivation of the contrasts will
proceed. If 0 is some effect (and we emphasize that we use the term effect to
embrace both main effects and interactions), the contrast will be the symbolic
product given by
{O} = 2n0 _ (a 1)(b 1)(c 1) ...,
where the sign of 1 is negative for all those letters appearing in

elsewhere. Thus, in a 2 6 experiment we should have
[Ch. 7
0 and positive
64135 = 64(ACE) = ( a 1)(b + 1)(c 1)(d + 1)(e 1)(f + 1).

We shall, henceforth, use upper case letters such as ACE to denote either the
effects or their estimates in a particular experiment. We shall no longer write the
hat over ACE unless its omission would cause confusion. Some statisticians
prefer to take as the effects twice the corresponding regression coefficients, e.g.,
A = 2 1 rather than A = as we do. Their procedure corresponds to a change
from x l = 1 to x l = +1 rather than a change from x l = 0 to x l = + 1.
Based on the symbolic product rule, Yates developed an algorithm for
computing the contrasts in a 2n design. We now illustrate the algorithm with the
data for the 2 design. There are several steps.
(i) The data are listed in the standard order (1), a; then b, ab; then c, ac, bc,
abc, and so on; they are entered in this order in column 1.
(ii) The first 2n -1 entries in any of the subsequent columns are obtained by
addition. The entry in the ith row of the (k + l)th column is the sum of
the (2i 1)th and the (2i)th entries in the kth column.
(iii) The next 2'' entries are differences. The (2'' + i)th entry in the
(k + 1)th column is obtained from the kth column by subtracting the
(2i l)th entry from the (2i)th entry. It is easy to make a mistake in
the sign at this stage.
(iv) The entries in the (n + 1)th column are the grand total M, followed by
the contrasts in standard order.
(v) W, X, Y, and Z are part of the checking procedure. W is the sum of the
entries in rows 1, 3, ... , 2n -1 1, and X is the sum of the entries in rows
2, 4, ... , 2" -1 . Y and Z are the corresponding sums for the second half of
the table. The check in the calculations is that in any column X + W is
equal to X + W + Z + Yin the previous column, and Z + Yis equal to
X W + Z Y in the previous column.
(vi) We may add an extra column, the (n + 2)th, for the sum of the squares of
the contrasts. The entries in this column are obtained by squaring the
entries in the (n + 1)th column and dividing by 2. The sum of the
entries in this column is equal to the sum of the squares of the original
observations.
The calculations in this algorithm are notoriously error prone, especially in
respect to wrong signs used in recording the differences. Three other methods of
checking the calculations are given in the literature. Good (1960), in an addendum
to his earlier paper (1958) on the topic, suggests reversing the order of the
elements of the (n + 1)th column so as to begin with ABC and end with M.
The Yates procedure is now carried out on these figures, and the new (n + 1)th
column will have for its entries the original data, multiplied by 2n, in the reverse
Yates' Algorithm and the General 2" Design 129
Sec. 7.2]
order. For the 2 2 design with (1) = 12, a = 22, b = 18, ab = 30, we should
have
(1)
12 34 82 M
16
120 = 4ab
22 48 22 A
14
104
72 = 4b
18
10
14 B
22
12
88 = 4a
ab
30
12
2 AB 82
60
48 = 4(1)
It is easily shown that the sum of the entries in the (n + l)th column should
be 2' times the last entry in the data column, e.g., 82 + 22 + 14 + 2 = 4 x
30 = 120. Quenouille (1955) points out that this "check" does not assure that
the calculations are correct. Consider again a 2 2 design and replace the first and
third entries in the second column by any numbers x and y. The entries in the
third column will be ab + b + x, ab b + y, ab + b x, and ab b y.
No matter what values are given to x and y, the entries in the third column sum
to 4ab.
Quenouille (1953) suggests using the fact that the sum of the squares of the
entries in the (n + 1)th column is 2' times the sum of the squares of the data.
It is obvious that this check is vulnerable to errors of sign in the (n + 1)th
column. However, Rayner (1967) has shown that this "check" holds "true" in
spite of mistakes in sign at any stage of the algorithm. To illustrate this we return
to the 2 2 example and consider what happens if an error is made in the sign of
the last entry in the second column. We should then have
(1)
12
34
82
22
48
18
10
14
ab
30
12
22
and 82 2 + (_2)2 + (14) 2 + (*22)2 = 7408 = 4 x 1852. Yates' Algorithm for

a 2 3 design follows.
(1)
(2)
b
ab
12
22
18
30
34 82 196 M
48 114 48 A
50 22 28 B
64 26
8 AB
20
10
14
ac
30
24
40
12
10
14
2
4
0
16
(1)
a
bc
abc
(3)
(5)
(4)
32 C
4802
288
98
8
128
AC
BC
ABC
0
2

(1)
[Ch. 7
(2) (3) (4)
W
30 84
X
52 112
Y
44 20
Z
70 28
X+W
82 196
XW
22 28
Z+Y
114 48
Z Y
26
8
X + W+Z+ Y 196 244
X W+Z Y 48 36
104
140
16
20
244 280
36
36 40
4
280
40
7.3 The 3n Series of Factorials

The 3n factorial experiment has n factors, each at three levels. An example of a
3 3 experiment appeared in Chapter 5. We denote the levels of the factors by
0, 1, and 2. Then the treatment combination with A and C at level 1 and B at
level 2 may be denoted by 121 or a 1 b 2 c 1 or ab 2 c. We shall use either of the first
two notations as convenient, and we shall sometimes not mention factors at their
low levels, writing b 1 c2 for the point with A at level 0, B at level 1, and C at
level 2.
To illustrate the 3 2 experiment we take some of the data from the 3 3 experiment
of Chapter 5. Subtracting 80 and rounding off we obtain
.
(1) 9,
a l 15,
a2
20,
bl
7,
a l b e 5,
a2 b 1
13,
b2
6,
a l b e 1,
a2 b 2
15.
TABLE 7.1
Total
8 d.f.
SA
SE
2
1
1
2
lin B
quad B
1
1
SAS
lin A
quad A
1211 920.1 = 290.9

1076.3 920.1 = 156.2
112.7
43.5
1015 920.1 = 94.9
80.7
14.2
39.8
The interaction sum of squares is usually subdivided in one of two ways. The
first method involves the use of orthogonal Latin squares. As was pointed out in
Sec. 7.3]
The 3" Series of Factorials 131
Chapter 6, a (p x p) Latin square is a square array of side p in which p letters

are arranged in such a way that each letter appears exactly once in each row and
each column. Two particular Latin squares of side 3 are
FQR P' Q R
QRF and RFQ
RPQ
Q R P.
These two squares have an additional property. If the second square is superimposed on the first square, each letter in the first square will appear exactly
once with each letter in the second square. Such a pair of squares is said to be
orthogonal. If we denote the coordinates (levels) for A and B by x l and x2 ,
respectively, P occupies in the first square those places for which x l + x2 - 0
(mod 3), while for Q we have x l + x 2 = 1 (mod 3), and for R, x l + x2 = 2
(mod 3). The totals for the letters are P 23, Q 37, R 31. Any contrast between
the letter totals is orthogonal to any contrast in the A totals or in the B totals.
We can, therefore, calculate a sum of squares between the letters in square 1
with 2 d.f. as (23 2 + 37 2 + 3 1 2 )/3 920.1 = 32.9. Alternatively, we could take
two orthogonal contrasts in the totals such as P Q and P + Q 2R and add
their individual sums of squares to obtain (-14) 2 /6 + (-2) 2 /18 = 32.9. For
the second square, the letter totals are P 29, Q 34, R 28. Any contrast in these
totals is orthogonal to any of the previous contrasts, and we have another sum
of squares, also with 2 d.f.: (29 2 + 34 2 + 28 2 )/3 920.1 = 6.9. The sum of
the two components is 32.9 + 6.9 = 39.8 = SAB . Because their derivations
involve the sets of points defined by x l + x 2 - 0, 1, 2 (mod 3) and by
x l + 2x 2 = 0, 1, 2 (mod 3), respectively, they are sometimes called the sums of
squares for AB and AB 2 interactions respectively.
For the 3 3 design we may subdivide SAB , SAc , and SBC in the same way.
SABc with 8 d.f. can be divided into four constituent sums of squares, each with 2
d.f., by breaking up the twenty-seven points into three sets of nine points each in
four orthogonal ways. We then have
+ x3 - 0, 1, 2 (mod 3),
+ 2x 2 + x3 =_ 0, 1, 2 (mod 3),
+ x2 + 2x3 = 0, 1, 2 (mod 3),
+ 2x 2 + 2x3 = 0, 1, 2 (mod 3).
ABC: x l + x 2
AB 2 C: x l
ABC 2 : x l
AB 2 C 2 : x l
Yates calls the sums of squares for AB and AB 2 in the 3 2 design the sums of
squares for the J diagonals and the I diagonals, respectively. He calls the
classifications that we have used for the 3 3 design Z, X, Y, and W, respectively.
The second approach is to split up the sum SAB into single degrees of freedom
for lin A lin B, and so on. We should then have the following four component
sums of squares: lin A lin B 1.0; lin A quad B 5.3; quad A lin B 33.3; quad A
[Ch. 7
quad B 0.1. There is no connection between these sums of squares and the sums
AB and AB 2 . The four components that we have just derived can also be obtained
by fitting the model
Yti = Yo + Yioui + Y2ou2 + Yoivi + Y02v2 + Y 11 u 1 v 1 + Y 12 U 1 V 2
+ Y21U2V1 + Y22U2V2 + ejj,
where u l = x i 1, u 2 = 3u1 2, v l = x 2 1, and v 2 = 3v1 2.

Yates' algorithm for the 2n designs may be adapted to the analysis of the 3Ty
designs. We illustrate the procedure by applying it to the 3 2 design. The procedure leads to the lin A, lin A quad B contrasts, and so on. The steps are similar
to those of the original algorithm. We begin by arranging the data in standard
order: (1), a,, a2, b1, albl, a2b1, b2, a1b2, a2b2, cl, alcl, a2c2, ..., and so on. The
entries in the (k + 1)th column are obtained from those in the preceding
column. Denote the entries in the kth column by ki . The entries in the (k + 1)th
column fall into three groups. In the first group the ith entry is k3i + k31 1 +
k3i_2; in the second group the ith entry is k, k_2; in the last group the ith
entry is k31 2k3i _ 1 + k 31 _ 2 . The sums used in the check are S l = k l + k 4 +
k7 +, S2= k 2 +k 5 +k 8 +, S3=k 3 +k 6 +k 9 + , X=S 1 +
S2 + S3 , and Y = S l S2 + 3S3 . The check is that X in any column should
equal Y in the preceding column.
(1)
44
91
al
15
20
25
22
26
28
lin A
quad A
7
5
11
a l be
a2b1
13
22
2
20
1
10
19
16
8
2
54
41
50
145
163
85
32
46
163
a2
bl
b2
a2b2
1
15
S1
22
S2
S3
X
Y
21
48
91
145
aj b2
lin B
lin A lin B
quad A lin B
quad B
lin A quad B
quad A quad B
7.4 Confounding
When we first discussed blocking we considered only the situation in which the
blocks were complete, which meant that each block contained as many points or
Confounding 133
Sec. 7.4]
plots as there were treatments or treatment combinations. We turn our attention

now to the situation which arises in a factorial experiment when the block size is
too small to accomodate one plot with each of the treatment combinations, but
is large enough for some fraction of them. We begin by discussing a 2 2 experiment in two blocks of two points each. We shall refer to the experimental points
as points or runs; the Jatter is a common usage among chemical engineers who
speak of making a run on their pilot plant at some set of conditions.
We have three choices for the division into blocks:
(i) (1), a;
b, ab;
a, ab;
(ii) (1), b;
(iii) (1), ab; a, b.
The revised model, including the block effects, may be written

Yijg = PO + N1x1 +
P2X2 + P12X1X2
+ sg + eij,
where Y ijg is the response at the ith level of A and the jth level of B in the gth
block, and 8 g is the gth block effect, which we now treat as fixed. The usual side
condition l g S g = 0 becomes 8 + 6 2 = 0, and so we write 8 1 = 6 and
S2=+ 8.
Consider blocking scheme (i). The experimental points in block I are both at
low B. The experimental points in block II are both at high B. Hence, the
expectation of the B contrast is (using the suffix 1 to denote the low level of a
factor and 2 for the high level)
E{B}
= E(y22 +
Y12
Y21 Y11) = 4P 2 + 48.
On the other hand, E{A} = 4P 1 and E{AB} = 4P 12 .

Another way of looking at the situation is to say that, since both points in one
block are at high B and both points in the other block are at low B, we cannot,
when we look at the B contrast, tell whether the observed difference results from
the B effect, from block differences, or from a combination of both. On the
other hand, each block contains one treatment combination at high A and one at
low A, and so the block effects 8 cancel out; they also cancel out for the AB
contrast. We say, therefore, that the B effect is completely confounded, i.e.,
confused, with blocks. In scheme (ii), A is confounded but B and AB are not.
In practice most experimenters would choose scheme (iii) because here it is the
two-factor interaction AB that is confounded, while the main effects A and B
are clear of the block effects.
When we consider dividing the 2 design into two blocks of four runs each,
there are more options; indeed, there are thirty-five possible ways in which to
choose three treatment combinations out of leven to be put in the same block as
(1). One possibility is to take as the first block (1), a, b, c, which is the traditional
one-at-a-time experiment (inasmuch as we start at the base point and vary
[Ch. 7
A, B, and C singly in turn). With this blocking the expected value of the contrast
between block totals is
E{abc + ab + ac + bc a b c (1)}
= 88 + 441+ 4P2 + 453
4123.
The expectation of the A contrast is 8P 1 + 48. All three main effects and the
three-factor interaction are confounded with blocks.
We can, however, arrange the blocking so that one and only one of the
effects is confounded. This we achieve by taking any one of the leven contrasts
for estimating effects, and placing in one block all the treatment combinations
that appear with a positive sign in the contrast and in the other block those
treatment combinations that have a negative sign. The chosen contrast will then
be identical with the difference between block totals, and the chosen effect will be
completely confounded with block effects. However, the estimating contrasts are
mutually orthogonal, and so the other contrasts are orthogonal to the contrast in
the block totals; the other effects are clear of the block biases.
If we decided to confound ABC with blocks, the division would be
block I abc, a, b, c;
block II ab, ac, bc, (1).
We note that in each block there are exactly two points at the high level of A
and two at the low level; the same is true for B and C. Block I consists of all the
treatment combinations with positive signs in the ABC contrast; we shall call
them the set of points defined by I = +ABC. They are also the set of points for
which x 1 x 2 x 3 = + 1. The other block contains the set of points defined by
I = ABC or by x 1 x 2 x 3 = 1. In general, let P be any effect. We shall call the
set of points that have positive (or negative) signs in the P contrast the set (or
fraction of the 2n points) defined by 1 = +P (or by I = P).
The chosen contrast is said to be the defining contrast of the confounding
scheme. The analysis proceeds as before, using Yates' algorithm, except that the
sum of squares for the defining contrast is discarded, since it is biased by the
block effects. The extension to blocking the general 2n design into two blocks of
2n -1 points each is obvious. When there are more than two blocks the situation
is more interesting; this will be discussed later.
7.5 Partial Confounding

1f the experiment contains several replications, i.e., if the set of treatment
combinations is to be repeated several times instead of only once, we have two
choices. Either we can repeat the same basic design each time, which would
amount in our example to dividing each replicate into two blocks always
confounding ABC, or else we can confound one or more of the other effects in
the subsequent replications. Thus, if the experiment were to call for four
Partial Confounding 135
Sec. 7.5]
replicates we might elect to confound ABC in the first replicate, AB in the second
replicate, AC in the third, and BC in the fourth. In that case the four interactions
are said to be partially confounded with blocks. Each of them is completely
confounded with blocks in one of the replicates, but it is unconfounded in the
other three replicates. We thus lose 25 per cent of the information about each of
the interactions by dividing the thirty-two observations into eight blocks of four
points each.
The experimenter has, therefore, two choices. He can either elect to learn
nothing about the three-factor interaction while at the same time retaining 100
per cent information about each of the two-factor interactions, or he can obtain
estimates of all four interactions at the colt of an increase in variance for the
estimates of AB, AC, and BC. This is a choice which he himself must make in
designing his experiment.
AN EXAMPLE OF PARTIAL CONFOUNDING. The following example,
given by Yates (1937), deals with an experiment on potatoes. The factors are
A, sulphate of ammonia; B, sulphate of potash; and C, nitrogen (manure). The
responses are the yields in pounds per plot.
ac 373
bc 398
ab 291
1163
la
c 312
a
106
265
b
abc 450
1133
Ib
(1) 106
ab 306
c
324
abc 449
1185
IIa
b 272
a
89
bc 407
ac 338
1106
IIb
(1)
(1) 131
(1) 101
87
ac 324
b 279
abc 471
1161
IIIa
323
128
a
bc 423
ab 334
1208
IIIb
a
103
bc 445
abc 437
1116
IVa
324
ac 361
b 302
ab 272
1259
IVb
Using Yates' algorithm for each of the four replicates separately, we obtain the
following totals and contrasts. The entries in the last column are the row totals.
The underlined contrasts are those that are confounded with blocks, and they
are not included in the corresponding row totals.
The estimates of the unconfounded effects are obtained by dividing the
corresponding totals by 32; for the partially confounded interactions the
divisor is 24. In Table 7.2 the rums of squares for the effects are obtained by
squaring the row totals and dividing by 32 or 24, as appropriate, the divisor
being the number of observations that went into the total.
The results of applying Yates' algorithm to the replicates follow.
[Ch. 7
II
III
IV
B
AB
2296
144
512
12
2291
73
577
79
2369
145
645
61
2375
29
537
47
9331
333
2271
26
C
AC
770
82
745
39
713
47
759
87
2987
208
BC
186
30
189
23
151
33
143
43
526
33.
M
A
ABC
TABLE
Total
7.2
Source
Total
Blocks
Main effects
Interactions
Error
d.f.
SS
ry2
C
= 466779.7
(11632 + 1133 2 +) C
= 4499.0
= 443453.1
(333 2 + 2271 2 + 2987 2 )/32
(26 2 + 208 2 + 526 2 + 33 2)/24 = 13404.4
5423.2
31
7
3
4
17
7.6 Confounding 2 Designs in 2k Blocks of 2
k Points
MS
642.7
147817.7
3351.1
319.0
Each
Consider the set of all effects with the addition of a unit element I. They are
I, A, B, ..., AB, ..., ABC, and, with the condition A 2 = B 2 = = I, they
form a multiplicative Abelian group of order 211. The product of the effects
ABC, ABD is A 2 B 2 CD, or CD. The product CD is said to be the generalized
interaction of ABC and ABD. We note that we could also have said (x 1 x 2 x3 )
x (x1x2x4) = 4xx3x4 = x3x4 (since xi = x = 1 at each experimental point).
A similar group is formed by the treatment combinations with (1) as the
identity and a 2 = b 2 = c 2 = = ( 1). Here, multiplication by a corresponds to
multiplying by x l or to changing the level of factor A. The reduction of the 2 11
experiment to an experiment on the vertices of a hypercube with coordinates 1,
and the corresponding equivalence of A 2 = I and xi = 1, appears to be due to
Box. The group property for the sets of effects and of treatment combinations is
due to Fisher (1942).
Suppose now that we elect to divide a 2n experiment into four blocks, confounding ABC and ABD. We can place in the first block all points with
x1x2x3 = + 1 and x 1 x2 x 4 = + 1, and in the second block all the points with
x1x2x3 = + 1 and x1x2x4 = 1, and so on. But whenever we have x1x2x3 =
x1x2x4 = + 1, we also have x 3x4 = + 1. The division into blocks is thus
The Allocation of Treatment Combinations to Blocks 137
Sec. 7.7]
Block I Block II Block III Block IV

X1X2X3
X 1 X2 X 4
X3 X 4
+1
1
1
+1
+1
+1
1
+1
1
1
1
+1
Each of the blocks contains 2n -2 points. The contrast in the block totals
I + II III IV is the ABC contrast; l II + III IV is the ABD contrast.
The third degree of freedom between blocks corresponds to I II III + IV,
which is the CD contrast. No other effects are confounded.
In the general case, let P and Q be two effects. We may divide the 2n design
into four blocks, each of size 2n 2 , by assigning to block I all points which have
positive signs in both the P and Q contrasts; block II contains the points which
have positive signs in the P contrast and negative signs in the Q contrast, and so
on. Then the generalized interaction PQ will also be confounded with blocks,
and these will be the only three effects confounded.
If we wish to make eight blocks of size 2i -3 , we may take a third effect R
(other than PQ). One block will contain all the 2' points that have positive
signs in the P, Q, and R contrasts. The Beven effects confounded will be P, Q,
PQ, R, PR, QR, and PQR. In the general case of 2" blocks, we take for the
defining contrasts a set of k independent generators P, Q, R, S,. , by which
we mean k effects such that no one of them is the generalized interaction of any
set of the others. The blocks are then the 2k sets obtained by taking the interThe
sections of the sets defined by I = P, I = Q, I = R, I = S,
complete set of confounded effects consists of the generators and all possible
generalized interactions that can be derived from them.
If we add I we obtain a set that includes the identity and is closed under the
group operation. The set of defining contrasts together with I is thus a subgroup of the group of effects. This subgroup is called the defining contrast
subgroup. It is usually desirable to choose the defining contrasts, if possible, so
that only higher order interactions are confounded. For example, in dividing a 2 5
experiment into four blocks of eight runs each, we might consider confounding
the five-factor interaction and one of the four-factor interactions. However,
confounding ABCDE and ABCD would be unwise because we should then
automatically confound their generalized interaction, which is the main effect E.
Similarly, choosing ABCDE and ABC as defining contrasts also confounds the
two-factor interaction DE. A preferable choice of defining contrasts would be
two three-factor interactions and one four-factor interaction such as ABD, ACE,
and BCDE. When the block effects are random, the design in which there are
one or more main effects confounded is similar to a split-plot design.
-
..
....
7.7 The Allocation of Treatment Combinations to Blocks

We have already mentioned that the blocks are the intersections of the sets
defined by 1 = P and 1 = Q,.. .. The actual assignment of points to blocks
[Ch. 7
is simplified in practice by taking advantage of the fact that the treatment

combinations form a group. The procedure used is to obtain the principal block,
and then to derive the other blocks from it. The principal block is defined to be
the block that contains the base point, or identity element, of the group, (1).
Theorem. The treatment combinations in the principal block form a subgroup of
the group of all treatment combinations.
PROOF. Let P be a defining contrast and let S. be the set of all treatment
combinations having an even number of letters (or none) in common with P.
Thus, if P were ABC, the set Sp would contain (1), ab, ac, bc, abd, and so on.
The set S, is closed and contains (1), for, if ab and ac are both in S P , so is their
product bc. Hence, S, is a subgroup. If P contains an even number of letters, S
is the set defined by I = +P. If P contains an odd number of letters, Sp is defined
by I = P. The principal block is the intersection of the sets St SQ , ..., and is
thus a subgroup. The principal block is sometimes called the intrablock subgroup.
3
If a treatment combination has an even number of letters (or none) in common
with both P and Q, it also has an even number (or none) in common with PQ.
Hence, it is enough to find n k independent members of the principal block,
each having an even number (or none) of letters in common with a set of k
generators of the defining contrast subgroup, and then to complete the block by
taking products.
The other blocks are obtained by changing the levels of one or more factors.
Changing the level of A is equivalent to multiplying each member of the principal
block by a. To change the levels of A and B simultaneously, we multiply each
member by ab. Thus the other blocks are the cosets of the principal block.
,,
EXAMPLE. Confounding a 2 5 design in four blocks of eight runs. The defining

contrasts are ABD, ACE, and BCDE.
Principal
Block
(1)
ade
bd
abe
ce
acd
bcde
abc
Block II
d
ae
b
abde
cde
ac
bce
abcd
Block III
e
ad
bde
ab
c
acde
bcd
abce
X 4 = X 1 X 2 +X 1 X2 X1X2
X5 = - X1X3 -X1X3
+X1X3
Block IV
de
a
be
abd
cd
ace
bc
abcde
+X1X2,
+XJX3
Sec. 7.81
The Composition of the Defining-Contrast Subgroup 139
The second block is derived by multiplying each member of the principal block
by d, the third by multiplying by e, and the fourth by multiplying by de.
We could also approach the problem of allocation of points in the following
way. The principal block consists of the points for which x 1 x 2 x 4 = 1 and
x,x 3x 5 = 1. Multiplying these equations by x4 and x5 , respectively, and noting
that 4 = 4 = 1, we have x 4 = x 1 x 2 , x 5 = x 1 x3 . We can thus obtain the
principal block by writing down a basic 2 3 design in the factors A, B, and C, and
then adding the other two factors D and E by putting x 4 = x l x2 and x 5 =
x l x3 . Similarly, the second block could be obtained by letting x 4 = + x 1 x2 and
x5 = x 1 x3 .
7.8 The Composition of the Defining-Contrast Subgroup

In this section we shall temporarily refer to the effects in the defining contrast
subgroup as words. The number of letters in an effect is the length of the word.
A word with an odd (even) number of letters is called an odd (even) word. The
product of two words is the word for the corresponding generalized interaction.
The patterns of word lengths in confounding a 2" design in 2k blocks have been
investigated by Brownlee et al. (1948), Burton and Connor (1957), and John
(1966a). We shall assume in the remainder of this section that each of the n
letters appears in at least one of the words in the subgroup. This eliminates from
our consideration, for example, the design for the 2 3 factorial in two blocks of
four runs confounding the main effect A.
The product of two even words, or of two odd words, is an even word. The
product of an odd word and an even word is an odd word. The following
necessary (but not sufficient) conditions can be deduced for the set of 2 1 1
words in the defining contrast subgroup:
(i) Each letter appears in exactly 2k -1 words. Hence, if w, is the length of the
ith word, 1, w, = 2 1 n;
(ii) Either all the words are even, or else exactly 2k -1 of them are odd. In
either case, the even words together with the identity form a subgroup;
(iii) If one of the words contains all n letters, the remaining words occur in
pairs such that the sum of the lengths of the words in any pair is n;
(iv) If one of the words consists of a single letter, the remaining words fall
into pairs, such that the lengths of the words in any pair differ by one.
Suppose that we take any set of k generators. We define a set of parameters:
ti denotes the number of letters appearing only in the ith generator; t, ; is the
number of letters that appears only in the ith and jth generators; and so on.
The sum of these parameters is t = n. We shall let w(i) denote the length of the
ith generator and w(i, j) the length of the product of the ith and jth generators.
Then
w(1)= t1+112 +...+tras +...,
w(1, 2) = w(1) + w(2) 2(t 12 + t123 + - - ).
[Ch. 7
In general let E O(s) t denote the sum of all the t which contain in their subscripts an odd number of members from a set S of integers i 1 , i2 ,..., 1 < i < k.
Then, if w(S) denotes the length of the generalized interaction of the i l th, i2 th, .. .
generators we have
w(S) = 1 t.
0(S)
Burton and Connor (1957) derived the formula
1w
2 = 2k - 2(1 t2
+ n2 ^
and showed that properties (i) and (ii) were not sufficient for the existence of a
confounding scheme.
Their example involved trying to block a 2 9 design into 2 4 blocks, using seven
words of length 4, six words of length 5, and two words of length 7. We have
w = 72 and so property (i) is satisfied. There is no difficulty in finding a set of
seven four-letter words satisfying property (ii); ABCD, ABEF, and ADFG
generate such a set.
We note next that 1 w 2 = 360 so that t 2 = t = n = 9. It follows that
nine of the t must be unity and the others zero. If such a scheme existed we could
take as generators the two seven-letter words and two of the five-letter words.
We should then have
w(1) + w(2) + w(3) + w(4) = 24.
But
1 W(i) =
4 t1234 + 3 1
tht7
+ 2 1 t{f + 1 t{.
A total of twenty-four can be obtained only if t1234, all the th1j , and all but two of
the t;5 are unity.
We can now try to reconstruct a set of generators with these values of t. One
letter, say A, must be in all four words; another letter, say B, must be in the
first three words but not the fourth, and so on. There are only two cases to be
considered: t12 = t34 = 0 and t12 = t13 = 0. In the first case each of the
generators must have six letters; in the second case w(2) = w(3) = 6. Thus an
arrangement with w(1) = w(2) = 7 and w(3) = w(4) = 5 is impossible. Burton
and Connor demonstrated the nonexistence differently, using some other results
from their paper.
7.9 Double Confounding and Quasi-Latin Squares

In a Latin square design p treatments or treatment combinations are applied to
p 2 plots arranged in a square array. The plots are assigned to the treatments in
Double Confounding and Quasi-Latin Squares 141
Sec. 7.9]
such a way that each treatment or treatment combination appears exactly once in
each row and in each column. In consequence, the treatment differences are not
confounded with either the row or the column effects. Yates (1937) introduced
the term quasi-Latin square to denote a design in which more than p treatment
combinations are used in a p x p square array. There is double confounding.
One set of effects is confounded with rows and another set is confounded with
columns.
In a 4 x 4 square we may have one complete replication of a 2 design or two
replicates of a 2 3 factorial. In dividing a 2 4 design into four blocks of four runs
each, we may confound the four-factor interaction and two of the two-factor
interactions, or else two of the three-factor interactions and their product, which
would be a two-factor interaction. A design in which ABCD, AB, and CD
are confounded with columns, and ABC, ACD, and BD are confounded with
rows is
(1) bcd abd ac

ab acd d bc
cd b abc ad
abcd a c bd
With this design we can estimate all four main effects and three of the six
two-factor interactions, but we have only 2 d.f. that can be used for error. These
are the 2 d.f. corresponding to ABD and BCD.
When two replicates of the 2 3 factorial are incorporated in one 4 x 4 square,
we might choose to have ABC and AB confounded with rows and AC and BC
confounded with columns. There are two ways of achieving this:
ABC
abc c
(1) ab abc c
a
bc ac (1) ab
AB
BC
AC
BC
ac bc
or
AC
abc c
(1) ab ac bc
abc c
(1) ab
bc ac b
In either case the first two rows and the last two rows are separate replicates.
The first two columns and the last two columns also constitute replicates. If,
however, the rows and columns are randomized, the design will not in practice
have this nice pattern with runs in the same replicate being in adjacent rows and
columns. The analysis is similar to that in the example of partial confounding
given earlier. Each of the partially confounded effects (the four interactions) is
estimated from the two rows (or columns) in which it is not confounded.
Yates gives a design for 2 6 in an 8 x 8 square confounding ACE, ADF, BDE,
BCF, ABCD, ABEF, and CDEF between the rows and ABF, ADE, BCD, CEF,
[Ch. 7
ABCE, ACDF, and BDEFbetween the columns. A problem with this design was
noted by Yates (1948) and was discussed in detail by Grundy and Healy (1950).
The difficulty lies in the fact that, when the columns and rows are randomized,
there is a considerable probability of obtaining an arrangement in which the
contrast for one of the main effects turns out to be the contrast between diagonally opposite quarters of the design. This is illustrated in the arrangement that
we wrote down for the 2 4 factorial in the 4 x 4 square; the ABD contrast pits
the four points in the top left quadrant and the four in the bottom right quadrant
against the points in the other two quadrants.
Yates presents other designs involving the use of several squares. He also
gives quasi-Latin squares for 3n designs and for two replicates of a design with
two factors at three levels and one at two levels in a 6 x 6 square. He discusses the
use of quasi-Latin squares in split-plot designs, choosing his terminology from
the weaving of tartans. A design in which main effects are confounded with both
rows and columns is called a plaid square. If main effects are confounded in only
one direction, we have a half-plaid square.
An example of a half-plaid square for a 2 3 factorial repeated twice, with A
confounded with rows and BC and ABC confounded with columns is
BC
ABC
abc ac
a ab
(1) b
a ab
abc ac
bc c
bc
(1).
Although we have introduced double confounding with the quasi-Latin

square, it is not essential that the designs have the same number of rows and
columns. The following Latin rectangle designs are given by Healy (1951)
using four rows and eight columns. The first design consists of four replicates of a
2 3 factorial. Each row consists of a complete replicate, and the treatment
combinations are arranged within the rows so that only ABC is confounded with columns. As the design is written, columns 1 and 2 form a replicate
confounding ABC. So do columns 3 and 4, columns 5 and 6, and columns
7 and 8.
(1) a
ab
ac
bc
abc
ab b
(1) abc bc
ac
ac
abc bc
(1)
ab
bc
abc c
ab
(1) a
ac
Confounding in 3 Designs 143
Sec. 7.10]
In the second design there are two replicates of a 2 4 factorial. Rows 1 and 2,
and rows 3 and 4 each comprise a complete replicate of the 2 4 . The first four
columns are a complete replicate confounding AB, ACD, and BCD. In the last
four columns, CD, ABC, and ABD are confounded.
(1) bd ac abcd ab ad bc cd
abc a
bcd d
cd bc ad ab
abd acd b
acd c
abd b
(1) bd ac abcd
bcd abc d
He also gives some designs for the 2 5 factorial in a 4 x 8 rectangle. One of

these confounds ABC, A DE, and BCDE with rows and A C, DE, ABD, and so
on, with columns. The other makes use of fractional factorials, which will be
discussed later.
7.10 Confounding in 3' Designs

We have already mentioned how in the 3 3 experiment the sums of squares for
each of the two-factor interactions and the three-factor interaction can be subdivided into components with two degrees of freedom. The same procedure leads
to a method of confounding for dividing the twenty-seven points into three
blocks of nine points each. If, for example, AB2 C 2 is to be confounded with
blocks, we place in the first block the nine points for which x l + 2x 2 + 2x3 =_ 0
(mod 3); the second block contains the points with x l + 2x 2 + 2x3 = 1
(mod 3); the third block has x l + 2x2 + 2x 3 - 2 (mod 3). The three blocks are
(I) 000, 012, 021, 101, 110, 122, 202, 211, 220;
(II) 002, 011, 020, 100, 112, 121, 201, 210, 222;
(III) 001, 010, 022, 102, 111, 120, 200, 212, 221,
where, for example, a2 b 1 c 1 is written as 211.
Confounding two of the components of the three-factor interaction to obtain
nine blocks of three points each also confounds a main effect. If we take, for
example, the three points for which x l + x 2 + x3 = i ( mod 3) and x l + 2x2 +
2x3 = j (mod 3), we also have 2x 1 - i + j (mod 3) or x l - 2(i + j) (mod 3).
We can avoid confounding main effects by confounding components of twofactor interactions such as AB, AC, and BC 2 .
Partial confounding can also be used for 3n designs with several replicates.
We may take a 3 3 with four replicates and confound each of the four pairs of
degrees of freedom for the three-factor interaction in one of the replicates. The
analysis follows the same general pattern as the analysis of the 2 3 example that
was discussed earlier.
[Ch. 7
7.11 Confounding with Factors at More Than Three Levels

The procedure that we have used in the previous section for the 3n series is a
special case of a more general procedure, due to Bose and Kishen (1940) and
Bose (1947a). They consider confounding in an experiment with n factors each at
s levels where s is a prime or a power of a prime.
They take a geometric approach to the problem. The treatment combinations
become points in n- dimensional Euclidean space, and the blocks are sets, or
intersections of sets, of parallel hyperplanes, or flats. We shall not go into the
details of this work. Reference to it will be made again in a later chapter on the
construction of incomplete block designs.
Let s = pm where p is a prime. We denote each of the treatment combinations
by a set of n coordinates (z 1 , z 2f ... , za ). These coordinates are elements of the
Galois field ofpn elements, GF(pn). If n = 1, as is the case when s = 3 or 5, the
elements can be taken as the integers 0, 1, ..., s 1, and arithmetic is carried
out mod s. If s = 4 = 2 2 , the elements are 0, 1, x, 1 + x where x 2 = x + 1,
and arithmetic is mod 2.
Consider the set of points for which 1 a,z = c where at and c are elements of
the field. The equation defines a hyperplane, or a "flat," which contains sm-1
points. Letting c take in turn each value in the field gives us a set of s blocks, each
with sn -1 points. If 1 b; z, = c denotes another set of hyperplanes, the intersections of pairs of hyperplanes, one from each set, form a set of s 2 blocks, each
of size sti -2 . In the first case, the s 1 d.f. corresponding to ai yi , or the
Aa1Ba2... interaction, are confounded with blocks. In the second case, we also
confound Ab1Bb2... and its generalized interactions with Aa1Ba2.
{
7.12 Confounding in the 4 2 Design

We illustrate now two approaches to the problem of dividing a 4 2 design with
factors A and B into four blocks of four runs each.
(i) We denote each treatment combination by a pair of elements from GF(2 2).
For convenience we write y for x + 1. Then x 2 = y and xy = 1. The
sixteen points in the design are
00 10 xO y0
01 11 xl yl
Ox lx xx yx
Oy ly xy yy.
Sec. 7.13] Confounding in Other Factorial Experiments 145

There are 9 degrees of freedom for AB interaction. We elect to confound
what might be called the ABx component by taking as blocks the flats
z l + xz 2 = c. The blocks are
z l +xz2 = 0
00 Oy 01
ly
10
Ox
lx 11
xl xx xO xy
yx yl
yy y0.
(ii) We may replace each four-level factor by two pseudofactors, each at two
levels, A by P and Q, B by R and S. Let the coordinate one denote a
pseudofactor at its high level and zero denote it at its low level. The
points are
0000
1000 0100
1100
0010
1010 0110
1101
0001
1001
0101
1110
0011
1011
0111
1111
This really amounts to replacing 0, 1, x, and y in the previous representation by 00, 10, 01, and 11, respectively.
The three degrees of freedom for A correspond to the main effects P and Q
and the interaction PQ. The main effects of B are R, S, and RS. The other nine
interactions between the pseudofactors correspond to the interaction AB. We
can now elect to split the 2 design into four blocks confounding PQR, QRS, and
PS. This method will give the same blocks as the previous method.
7.13 Confounding in Other Factorial Experiments

The methods described in this chapter can be applied, although not so easily,
to factorial experiments in which the factors do not all have the same number of
levels. If some of the factors have two levels and the others have four levels, the
confounding can be achieved by the use of pseudofactors, which has just been
described. Yates (1937) discusses confounding in designs with some three-level
factors and some four-level factors. Binet et al. (1955) consider dividing a single
replicate of a design with factors at several levels into blocks.
146 Factors with Two or Three Levels; Counfounding
[Ch. 7
Exercises
1. Apply Yates' algorithm to the following set of data for a 2 4 experiment:
(1)
a
b
ab
3.96
4.05
2.71
2.94
c
ac
bc
abc
3.71
3.87
2.67
3.02
d
ad
bd
abd
4.53
4.69
3.01
3.37
cd
acd
bcd
abcd
4.52
4.72
3.32
3.53
2. Use the algorithm to evaluate the contrasts and obtain the analysis of variance
table for the 3 3 design given as an example in Chapter 5.
3. Devise an algorithm in the style of Yates for calculating the contrasts in a
2 2 x 3 design. Use it to carry out the analysis of variance for the following
design which has two levels of P, two levels of N, and three levels of K.
Ko
No N1
K1
No N1
KZ
No N1
Po
23 36 25 46 25 31
P1
34 28 19 54 19 38
4. The data in exercise 1 were actually taken from a 2 2 x 4 experiment [Raese and
Decker (1966)]. Factors C and D correspond to the factor with four levels. Use
the results of exercise 1 to obtain the analysis of variance table for the 2 2 x 4
design.
5. In the section on partial confounding we considered the example of a 2 3 design
with four replicates, each divided into two blocks of four points each. Prove that
the estimates we obtained are actually the least squares estimates.
6. Prove the necessity of the four conditions on the words in the defining contrast
subgroup given at the beginning of the section on the composition of that group.
7. Find the principal block when the 2 8 design is split into blocks of sixteen runs
each, using the defining contrast subgroup generated by ABCD, ABEF, ABGH,
ACEG.
8. Derive the quasifactorial design for a 2 6 design in an 8 x 8 square by Yates,
which was mentioned in the text.
9. Derive a design for a 2 5 factorial in a 4 x 8 rectangle, confounding ABC, ADE,
BCDE with rows and AC, DE, ABD with columns.
10. Obtain the partially confounded design for a 3 3 factorial with four replicates,
each in three blocks of nine points each. Confound ABC with blocks in the first
replicate, ABC 2 in the second, AB 2 C in the third, and AB 2 C 2 in the fourth.
11. A reasonable procedure when a plot is missing in a 2n design is to replace it by
a value which makes the highest order interaction zero. Apply this technique to
the data of exercise 1, assuming bc to be missing. What is the variance of the
estimate of an effect when a missing plot is included?
12. (Continuation). In the 2n design the highest order interaction has only one
contrast. However, in a 3" design the highest order interaction has 2n degrees
Exercises 147
of freedom; thus, there are 2" orthogonal contrasts, each of which might be
equated to zero to calculate a missing value. A less ambiguous procedure
would be to omit all the highest order interaction terms from the model and use
the procedure of the earlier chapters. Investigate this for the 3 3 experiment of
exercise 2, assuming that the point 012 is missing.
CHAPTER
Fractions of
2n Factorial
Designs
As the number of factors in a 2n design increases, the number of points soon

outgrows the facilities and budgets of most experimenters. The full 28 experiment calls for 256 runs. Only eight of the degrees of freedom correspond to main
effects, and twenty-eight to two-factor interactions. The remaining 219 degrees of
freedom correspond to interactions involving three or more factors. If these
higher order interactions are all negligible, the experimenter has 219 degrees of
freedom for error, which is more than he really needs for an adequate estimate.
On the other hand, he hasto make 256 runs to estimate only 8 + 28 effects, and
only 14 per cent of the degrees of freedom are used for estimation.
We have already seen how Latin squares can be used to estimate the main
effects in a p3 factorial experiment with only p2 runs. Graeco- Latin squares and
hyper-Graeco-Latin squares result in even greater savings, but at the cost of
assuming that some of the interactions do not exist.
This chapter is devoted to an investigation of fractions of 2n designs. One of
the fractions that will be discussed can be derived from a Latin square of side 2,
either
I 2
2 I
or
2 1
1 2.
If we let rows represent the levels of A, columns the levels of B, and I and 2 the
levels of C, we have two four-point fractions of the 23 design: (1), be, ab, ae, and
b, e, a, abc. These are half replicates of the 23 design, or 23 - 1 designs, from which
the main effects A, B, and C are estimable, if we neglect their interactions. The
2n - " fractions, or 1{2" fractions of Z" designs were introduced by Finney (1945).
148
Sec. 8.1 ]
The 2n - k Fractional Factorials
149
They will be the topic of the first part of the chapter, after which some other
fractions will be discuseed.
Fractional factorials are commonly used in industrial experimentation where
runs are expensive and where an acceptable prior estimate of the variance is
available. They may also be used for screening experiments. These are experiments performed in the early stages of an investigation in which many factors
are considered, most of which may turn out to have little or no effect upon the
response under investigation. The purpose of the screening experiment is to
screen several factors with the purpose of spotting those factors, if any, that have
appreciable effects. The factors that seem to be interesting are then investigated
more closely in subsequent experiments. The emphasis is not so much upon
estimating the effects with small variance, but upon finding out which of the
factors merit further study. We shall assume throughout most of this chapter
that the experimenter is interested in estimating the main effects, and perhaps the
two-factor interactions; that he is prepared to assume before carrying out the
experiment that the higher order interactions are negligible; and that he is not
particularly concerned with using the actual experiment to provide himself with
an estimate of the variance. The emphasis will therefore be on finding small
experiments in which a high percentage of the degrees of freedom are used for
estimation. In these small experiments there will be few degrees of freedom for
error even if higher order interactions are used, and it will be assumed that the
experimenter has some prior estimate of the error variance. Daniel (1959) has
introduced a graphical method for deciding which effects in a 2n factorial are
significant when there is no estimate of error. If there is no prior estimate
available such graphical procedures seem to be the only recourse.
8.1 The 2' Fractional

Fractional Factorials
When in the previous chapter we divided the 2 5 design into four blocks of
eight runs, each block constituted a quarter replicate, or a 2 5 2 fraction. The
term 2n - k fraction or (1/2k)th replicate is used to describe only those sets of
2n-k treatment combinations that could occur in the same block when the
complete factorial is divided into 2 k blocks confounding exactly 2k 1 effects
in the manner derived in the previous chapter. Thus we may define a 2' -k
fraction as a set of 2n -k different points which constitute a subgroup of the
group of treatment combinations, or a coset of such a subgroup. This definition
throws very little light upon the nature of fractional factorials and their importance. Henceforth we shall regard a subgroup as a trivial coset of itself, and use
the term coset to embrace both the subgroup itself and the other cosets.
The first fraction listed (the principal block) was the fraction defined by
I = ABD = ACE = + BCDE. The other fractions in order were defined by
I= +ABD=ACE=BCDE,I=ABD=+ACE=BCDE,and I=
+ABD = +ACE = +BCDE. The four fractions are said to belong to the same
family. The term principal fraction has been used in the literature in two different
-
150 Fractions of 2n Factorial Designs
[Ch. 8
senses and we shall refrain from employing it here. Some authors use it to
denote the fraction containing (1) by analogy with the term principal block.
Box and Hunter (1961) use it to denote the fraction for which the defining
contrasts all have positive signs; in our example that would be the fraction
containing de, a, be, abd, cd, ace, bc, and abcde. We now consider that fraction in
more detail.
For each treatment combination in the fraction the relationship 1 = x 1 x 2 x 4 =
x 1 x 3 x 5 = x 2 x3 x4 x5 holds. Multiplying by x 1 and recalling that xi = 1, we have
xl =
x2x4 = x3x5 = x1x2x3x4x5
at each point so that A, BD, CE, and ABCDE are indistinguishable. The four
effects A, BD, CE, and ABCDE are said to constitute an alias set and to be
aliased (or confounded) with each other. The alias sets are the cosets of the
defining contrast subgroup (which is sometimes called the alias subgroup in
this context). There are 2n k alias sets (including the subgroup itself) and each
set has 2k members. Each effect is aliased with the effects in the same alias set
but with no others.
If the elements of the vector of effects F3 are written by order of alias sets so
that P' = (PO, P124, P1359 P2345, P13 P24, P35, P12345, ...) the cross-product matrix
X'X in the normai equations consists of square matrices 8J 4 along the main
diagonal and zeros elsewhere. The rank of X'X is 8. A matrix of full rank is
obtained by choosing one effect from each of the alias sets and striking out the
remaining rows and columns. If, therefore, we were to assume a restricted model
which included the mean, the five main effects, either BC or DE, and either CD
or BE, but no other terms, the corresponding least squares equations would have
a unique set of solutions.
The A contrast from this fraction is
-
(a + abd + ace + abcde de be cd bc).

Its expectation is 8(fl 1 + P 24 + f35 + P 12345), so that when we use the contrast
to estimate A we are actually estimating A + BD + CE + ABCDE. The other
fractions give chains A BD CE ABCDE, with the signs corresponding
to the signs of the defining contrasts. The full set of alias chains is
I +ABD+ ACE +BCDE
B + AD + ABCE + CDE
C + ABCD + AE + BDE
BC + ACD + ABE + DE
A + BD + CE + ABCDE
AB + D + BCE + ACDE
AC + BCD + E + ABDE
ABC + CD + BE + ADE
For A to be estimated we must assume that BD, CE, and ABCDE are all zero.
If we plan to use this fraction with the intention of estimating A we must decide
a priori that we are going to ignore the other three interactions. The prior
Sec. 8.2]
The Classification of Fractions
151
decision to ignore the five-factor interaction will cause little concern, but it may
not be reasonable to assume without further investigation that the two-factor
interactions BD and CE are negligible. If these two particular interactions are
the only ones that concern the experimenter, he can avoid the difficulty by
choosing a different fraction such as one of the 2 5 2 defined by I = ABC =
A DE = BCDE. The two interactions will then be aliased with each other,
but not with any of the main effects. There is not a quarter replicate from which
the mean, the five main effects, BD and CE can all be estimated. If, on the
other hand, the experimenter is not willing to declare prior to carrying out his
experiment that any of the two-factor interactions can be ignored, he cannot
estimate without bias the five main effects from any 2 5-2 fraction. In that case he
is going to have to perform a larger experiment or to omit one of the factors.
-
8.2 The Classification of Fractions

We have seen that in the use of fractional replicates it is necessary to assume
a priori that certain effects are zero. It is a common practice in industrial
experimentation to make the blanket assumption in the absence of prior
information to the contrary that all interactions that involve three or more
factors can be ignored. We shall make that assumption in this chapter. An
interaction which is taken to be zero in this way will be said to be suppressed.
We shall sometimes suppress the two-factor interactions also, and consider
estimating only the main effects. It will be convenient to use the abbreviation
2 f.i. for the expression two factor interaction.
An effect will be said to be estimable if, and only if, there is a contrast in the
data which has for its expectation the particular effect biased only by other
effects which we choose to suppress. The effect being estimated is said to be
aliased with the effects that were suppressed. Designs in which main effects are
aliased with each other are of no interest. The remaining designs are divided into
several classes.
One method of classification has three classes:
(i) Main effects only designs. These are fractions in which some of the main
effects are aliased with 2 f.i. and are estimable only when those 2 f.i. are
suppressed;
(ii) Main effects clear plans. These are fractions in which al] the main effects
are estimable whether the 2 f.i. are suppressed or not, but some of the 2
f.i. are aliased with each other;
(iii) 2 f.i. clear plans. These are fractions in which all the main effects and all
the 2 f.i. are estimable.
These classes were originally called types A, B, and B' by Box and Wilson
(1951). Box and Hunter used the terras resolution III, resolution IV, and
resolution V. Daniel (1962) called them three-, four-, and five-letter plans. The
[Ch. 8
last two terminologies are in common use. The justification for Daniel's nomenclature becomes clear when the alias sets are considered. If one of the defining
contrasts is a three-letter word, say ABC, then A is aliased with BC, B with AC,
and C with AB; the three main effects are estimable only if the three 2 f.i. are
suppressed. With ABCD as a defining contrast, A is aliased with BCD, but AB
is aliased with CD. With a five-letter word ABCDE, the main effects and 2 f.i.
are all aliased only with higher order interactions.
We shall use the following definition of the resolution of a fractional factorial,
which is applicable to other fractions as well as the 2n ' series. An interaction
involving t factors will be called a tth order effect; main effects are first-order
effects. Let t be any integer.
A design is said to be of resolution (2t + 1) if it satisfies the condition that all
effects of order t or less are estimable whenever all effects of order higher than t
are suppressed.
A design is said to be of resolution 2t if all effects of order (t 1) or less are
estimable whenever all effects of order (t + 1) or higher are suppressed.
-
8.3 Some
2n - k Fractions
The 2 3 -1 Design
The smallest usable 2n - k fraction is the 2 3-1 consisting of four points. Two
such designs are useful: one, defined by I = ABC, contains the four points,
a, b, c, and abc; the other, defined by I = ABC, consists of (1), ab, ac, and bc.
Both these designs are three-letter plans. Each main effect is aliased with a
two-factor interaction. In the first of these half replicates the estimates of the
main effects are
= (abc + a b c)/4 with expectation A + BC;

= (abc + b a c)/4 with expectation B + AC;
= (abc + c a b)/4 with expectation C + AB.
The variance of each estimate is a2 /4.

In the traditional one-at-a-time experiment, which appeals to some experimenters, each factor is varied in turn while the other factors are maintained at
their low levels. With three factors this design consists of the four points (1),
a, b, and c. It is interesting to compare this design with the half replicates. Each
uses four points to estimate the mean and the three main effects. However, the
estimate of the main effect A, for example, from the one-at-a-time experiment is
p 1 = (a (1))/2 with expectation A AB AC + ABC and variance o 2 /2.
The 2 -1 fraction thus provides an estimate which has a smaller variance and
which is biased by fewer interactions.
The one-at-a-time experiment should not, however, be dismissed completely
because of its relative inefficiency. It is attractive to some experimenters in
Sec. 8.4]
Designs with Eight Points
153
laboratories and pilot plants, especially when they are not quite sure at the
beginning of a series of experiments how far apart to take the levels of the
quantitative factors; if a ten-degree change of temperature is tried and does not
appear to produce an adequate change in response, perhaps the high and low
levels of the temperature factor should be twenty degrees apart. A useful compromise is to begin with the one-at-a-time experiment and then to add a fifth
point, abc. The five points then comprise the half replicate defined by I = + ABC
plus the base point (1). With the fifth point included in the experiment, one can
retain the same model with mean and main effects only, or else add to the model
a term for one of the interactions. It will be found that when (1) is added to the
half replicate and P123 is added to the model, the least squares estimates of the
main effects A, B, and C remain the same as the estimates from the half replicate
alone. On the other hand, if P 12 were the interaction added to the model, the
estimates of the mean, A and B, would be unchanged but that of C would be
altered.
8.4 Designs with Eight Points

The two half replicates of the 2 4 factorial defined by I = ABCD are fourletter plans with eight points. Each f the four main effects is aliased with a
three-factor interaction. The 2 f.i. are aliased in pairs; AB is aliased with CD,
AC with BD, and AD with BC. If it is known beforehand that one of the four
factors does not interact with any of the others, the interactions involving this
factor can be suppressed and then the other three 2 f.i. will be estimable.
To illustrate these designs we consider a complete 2 design which formed
part of a gasoline experiment. Each of the data points is the octane requirement
of a car (from which eighty has been subtracted). The four factors were all
qualitative.
(1) 7, a 7, b 9, ab 1, c 11, ac 12, bc 14, abc 7,

d 10, ad 8, bd
12, abd 5, cd 6, acd 6, bcd 10, abcd 6.
The corresponding contrasts, each measuring 16p, are

T 131,
A -27, B -3, AB -25,

C +13,
AC +7, BC +7, ABC +1,
D -5, AD +1, BD +9, ABD -7,

CD-27, ACD + 3, BCD + 3, ABCD + 1.
The procedure for the use of Yates' algorithm is to regard each fraction as a
complete 2 design in A, B, and C, to which the fourth factor D has been added
by putting x 4 = + x 1 x 2 x 3i or x 4 = - x 1 x 2 x 3 . We give the calculations for the
[Ch. 8
two half replicates. The fourth column of figures in each section is the column of
contrasts; these figures, which estimate 8fl, should be compared with one half of
the values of the contrasts for the complete design.
I = + ABCD
(1)
ad
bd
ab
cd
ac
bc
abcd
7
15
28
66
8
13
38 12 A
12
18 10
0 B
1
20 2 26 AB + CD
1 2
10 C
6
12 11
2
8 AC + BD
14
6 12
4 BC + AD
6 8 14 2 D
I = ABCD
d
a
abd
c
acd
bcd
abc
10
7
9
5
11
6
10
7
17
14
17
17
3
4
5
3
31
65
34 15 A
7 3 B
8
1 AB CD
3
3 C
0 1 ACBD
1
3 BC AD
2
3 D
The last contrast in each table is the ABC contrast. In the first fraction this
contrast estimates 8(ABC + D); suppressing ABC we have 8$ 4 = 2. In the
second half replicate the contrast estimates 8(ABC D), so that we have
8 4 = 3, or 8 4 = 3.
It should also be noted that adding the values of the contrasts in the two half
replicates gives the values for the complete design. Thus the A contrast for the
complete factorial is 27 = (-12) + (-15) ; the AB contrast is 26 + 1 =
25, and the CD contrast is 26 1 = 27. We are reminded in the second
half replicate that it is not wise to conclude, just because the estimate of AB
CD is 1, that both AB and CD are small. Although there is no mathematical
justification for it, some scientists would as a practical matter be prepared, as
Daniel (1962) has pointed out, to assume for example in the first half replicate
that, since the A and C effects are large and the B and D effects are small, the
relatively large size of the AC + BD contrast can be attributed to the AC
interaction rather'than the BD interaction. This procedure, lacking as it does a
firm mathematical basis, can lead to pitfalls unless the scientist has a considerable amount of prior information about the probable relationships between the
factors.
If we add to the first of these half replicates a fifth factor E by putting
x 5 = x 1 x 2 we have the 25-2 design defined by 1 = ABCD = ABE = CDE.
,
Sec. 8.5]
Designs with Sixteen Points
155
This resolution III design consists of the points e, ad, bd, abe, cde, ac, bc, abcde,
and is similar to the 2 5 - 2 designs that were mentioned earlier.
Adding F by putting x s = x 1 x 3 gives the 26-3 design defined by I = ABCD =
ABE = CDE = ACF = BDF = BCEF = ADEF, which consists of the points
ef, ad, bdf, abe, cde, acf, bc, abcdef.
Finally, we may add a seventh factor by letting x 7 = x 2 x 3 and obtain efg, adg,
bdf, abe, cde, acf, bcg, abcdefg. The defining contrast subgroup for this fraction
is generated by ABCD, ABE, ACF, and BCG. The alias chains for main effects
and 2 f.i. are
A+BE+CF+DG
B+AE+CG+DF
C+AF+BG+DE
D+AG+BF+CE
E+AB+CD+FG
F+AC+BD+EG
G+AD+BC+EF
A three-letter plan for 2t t -1 factors in 2n runs can always be constructed in this
way by starting with a basic 2n design and then adding new factors by equating
them to interactions.
8.5 Designs with Sixteen Points

The half replicates of the 2 5 design defined by I = + ABCDE or by I=
ABCDE are designs of resolution V. Each main effect is aliased with a fourfactor interaction; each 2 f.i. is aliased with a three-factor interaction. There are
five main effects and ten 2 f.i. Thus there are as many effects to be estimated as
there are degrees of freedom. A fraction in which each degree of freedom is used
to estimate an effect is said to be saturated. Daniel (1956) introduced the terms
degree -of-freedom efficiency and variance efficiency for a fractional factorial.
The variance efficiency of a fraction is obtained by comparing the average of the
variances of the estimates of the effects to the smallest variance that could be
hoped for with factors at the levels 1. For a main effect such as A, the minimum
variance of the estimate $ is obtained when half the points are at x, = +1 and
half are at x l = 1; the variance is then V(P 1 ) = a 2 IN. The 2' fractions
all have 100 per cent variance efficiency. We shall discuss later some nonorthogonal fractions which have lower efficiencies. The degree-of-freedom
efficiency is the ratio of the number of effects estimated to the number of degrees
of freedom in the design. The 2 5 -1 fractions have 100 per cent degree-of-freedom
efficiency.
156 Fractions of 2' Factorial Designs
[Ch. 8
The half replicate defined by I = ABCDE consists of the sixteen points
(1), ae, be, ab, ce, ac, bc, abce, de, ad, bd, abde, cd, acde, bcde, abcd.
The complementary half replicate, defined by I = + ABCDE, contains all
treatment combinations with an odd number of letters.
Four-letter plans for six, seven, or eight factors can be obtained by starting
with a basic 2 4 design and adding the other factors by putting E = ABC,
F = ABD, G = ACD, and H = BCD. The design for all eight factors consists
of the points
(1), aefg, befh, abgh, cegh, acfh, bcfg, abce,

dfgh, adeh, bdeg, abdf, cdef, acdg, bcdh, abcdefgh.
The defining contrast subgroup is generated by ABCE, ABDF, ACDG, and
BCDH; each defining contrast contains an even number of letters.
Designs like this can also be obtained by the important method of folding
over. Suppose that we begin with a resolution III design, the 2 7 - 4 design with
eight points presented in the previous section. We can regard it as a 2 6 - 5 design
in which all the runs were made at the low level of the factor H, and add the
generator H to the defining contrast subgroup. The folding-over process
consists of repeating each run with the levels of every factor changed. Thus,
folding over efg gives abcdh. The design is
efg, adg, bdf, abe, cde, acf, bcg, abcdefg,
abcdh, bcefh, acegh, cdfgh, abfgh, bdegh, adefh, h.
For the first eight points we have x1x2x5 = x1x3x6 = x2x3x7 = x8 = + 1; for
the second eight points we have x1x2x5 = x1x3x6 = x2X3x7 = X8 = 1. For
the whole set of sixteen points we have (x 1 x 2 x 5 )(x 1 x 3 x 6 ) = ( x 1 x 2 x 5 )(x 2 x 3 x 7) =
(x1x2x5)( x8) ` (x1x3x6)(x2x3x7) = (x1x3x6)( x8) = (x2x3x7)( x8) = + 1, and
also x 1 X 2 x 3 x 4 = 1.
Thus x 2 x 3 x 5 x 6 = x 1 x 3 x 5 x 7 = x 1 x 2 x 5 x 6 = _ + 1 and the generators of
the fraction are BCFG, ABEG, ABEH, ABCD.
In general, folding over will convert any design of resolution III into a design
of resolution IV or higher. The defining contrasts for the combined design are
those defining contrasts of the original fraction that contain an even number of
letters, and, since there are no two-factor interactions in that set, they all contain
four letters or more. The original 2" fraction
fraction has 2k -1 1 defining contrasts
with an even number of letters and 2k -1 with an odd number. In the complementary fraction obtained by folding over, the signs of each of the odd defining
contrasts are changed, and the signs of each of the even contrasts are unaltered.
It also follows that folding over a resolution V design produces a design of
resolution VI or higher.
Other 2" ' Fractions of Resolution V 157
Sec. 8.6]
Folding over can also be considered from the point of view of breaking chains
of aliases. Since the three-letter defining contrasts change signs in the second
fraction, the 2 f.i. change signs in those alias chains which include main effects.
Thus, in our example, the alias chain for A in the first eight points (considering
only main effects and 2 f.i.) is A + BE + CF + DG AH, while the corresponding alias chain in the second set is A BE CF DG + AH. The A
contrast from the first eight points gives an estimate of ft + P25 + P36 +
P47 P12. From the second eight points we have an estimate of fl p
P3s P47 + . Adding them gives an estimate of 22 1 .
Another possibility is to take a design of resolution III and repeat it, changing
only the level of A without adding the extra factor H. This gives a design in
which A and all the 2 f.i. which contain A are estimable, biased only by interactions involving three or more factors. In our example the design would be
efg, adg, bdf, abe, cde, acf, bcg, abcdefg,
aefg, dg, abdf, be, acde, cf, abcg, bcdefg.
The set of defining contrasts for the combined design consists of all the defining
contrasts of the original design which do not contain A. The alias chains
containing A are A + BE + CF + DG and A BE CF DG. The alias
chains containing AB are AB + E + FG + CD and AB E FG CD.
8.6 Other 2n - k Fractions of Resolution V

For six factors the smallest 2n - k fraction of resolution V is a half replicate.
This can be defined either by the six-factor interaction, I = ABCDEF, or by
any one of the five-factor interactions. We would usually prefer one of the halves
defined by the six-factor interaction, because the main effects would be aliased
with five-factor interactions and the 2 f.i. with four-factor interactions.
For seven or eight factors, sixty-four points are needed. For seven factors this
requires a half replicate. With eight factors we can obtain a suitable quarter
replicate with two five-letter words and a six-letter word in the defining contrasts, for example, I = ABCDE = ABFGH = CDEFGH.
2n - k fractions of resolution V for nine, ten, or eleven factors call for 128 points
each. Examples of suitable sets of defining contrasts are
29-2
210-3
1= ABCDEF = ABCGHJ = DEFGHJ;
1= ABCDEF = ABCGHJ = ABDEGHK;
2 11-4 : 1= ABCDEF = ABFJK = AEFGKL = ACEHL.

The first is a design of resolution VI. In the last two fractions only sets of
generators for the defining contrast subgroups are given. For twelve or more
factors at least 256 runs are needed. A set of generators for a 212 -4 design of
resolution V is ABCDJK, ABEFJL, ADEGLM, ABGHKL.
[Ch. 8
8.7 Blocking Fractional Factorials

Blocking in fractional factorial designs is facilitated by the use of blocking
factors. These are dummy factors which are supposed to interact with each
other but not with the real factors. An example will illustrate the procedure.
Suppose that we wish to divide the 2 4 -1 design of resolution IV into four blocks
of two runs each. We introduce two blocking factors, X = AC and Y = AB.
This gives us eight treatment combinations: (1), adxy, bdy, abx, cdx, acy, bcxy,
and abcd. We place into one block the points at the high level of X and the high
level of Y, into the second block the points at low X and high Y, and so on.
The blocks in the design are
(1), abcd; ab, cd; ac, bd; ad, bc.
This particular example illustrates another point. The design is a fold-over

design, and any fold-over design of resolution IV can be divided into blocks of
size two without losing its resolution IV property. Each block consists of a pair
of complementary treatment combinations. This blocking procedure corresponds to taking the two-factor interactions as blocking factors.
We have already discussed the saturated designs of resolution III for 2n 1
factors in 2n runs in which only the main effects are estimated. If we use any one
of the factors as a blocking factor, we have a design of resolution III for 2n 2
factors in two blocks of size 2'' -2 . Choosing two of the factors and their interaction, which will correspond to a third factor, gives a design for 2n 4 factors
in four blocks, and so on.
As an example we may construct a design of resolution III for twelve factors in
four blocks of four.
We take the basic 2 4 design in the factors A, B, C, and D, and assign X = AD
and Y = BC as blocking factors, giving the sixteen points xy, ay, bx, ab, cx, ac,
bcxy, abcy, d, axy, bd, abdxy, cd, acdx, bcdy, abcdxy. Then, upon adding new
factors E = AB, F = AC, G = BD, H = CD, J = ABC, K = ABD, L = ACD,
and M = BCD, the blocks are the columns of the array
4
abehlm
acfgkm
bdfgjl
cdehjk
bfhjkm
cegjlm
abdegk
acdfhl
aghjkl
abcefj
defklm
bcdghm
efgh
bckl
adjm
abcdefghjklm.
An interesting example is found in the National Bureau of Standards list

(p. 8). A 2 -1 design of resolution V is laid out in an 8 x 8 square. All main
effects and 2 f.i. are estimable with both row and column confounding. The
Sec. 8.8]
Analyzing 2n -' Fractions 159
fraction is defined by I = + ABCDEFG. The blocking factors for the columns

are X = ACF, Y = ABE, Z = EFG. For the row confounding we have
U = ACD, V = ABF, and W = BCE.
(1) acef adfg

adef cd
eg
bcef ab
abcdeg
abcd bdef
bcfg
cdfg adeg
ac
aceg fg
cdef
bdeg abcdfg abef
abfg bceg
bd
cdeg
acfg
bdfg
abeg
ef
ad
bc
abcdef
befg
abdg
cg
acdefg
bcde
abcf
df
ae
abcg
bcdefg
aefg
dg
abdf
be
acde
cf
abde bcdf
bf
abce
acdf
de
ce
af
abcefg bg
bcdg
abdefg
ag
cefg
defg
acdg
8.8 Analyzing 2 - k Fractions

In the example of the 2 4 -1 design defined by I = + ABCD, we proceeded as if
the design were a 2 3 design in A, B, and C to which D had been added by
putting x4 = x 1 x 2 x 3 . We applied Yates' algorithm to the basic 2 3 design and
ignored the factor D until we wrote down the effects in the last column of the
table. We could just as well have applied the algorithm to the 2 3 in A, B, and D,
or in A, C, and D, or in B, C, and D. The general technique is to consider the
2n - k fraction as a complete factorial in a set of (n k) of the factors, apply
Yates' algorithm, and then incorporate the other factors in the last column.
It is not always the case that any subset of (n k) factors will provide a basic
set. Consider for example the 2 8 - 2 defined by I = ABCD = ABEF = CDEF.
We cannot analyze this as a 2 in A, B, C, and D. Nor can we take as the basic
set of four factors A, B, E, and F, or C, D, E, and F. We can take A, B, C, and E
and put x4 = x 1 x 2 x3 , x 6 = x 1 x2 x5 , or else A, B, C, and F and put x 5 = x 1 x 2 x6 ,
but we cannot have both 1E and F out of the basic set. There is no defining
contrast that contains E but not F, or F but not E, and it is impossible to express
E as a product of factors that does not involve F. Indeed, any basic set of four
factors in this example will have to contain at least one member from each of the
pairs A, B; C, D; E, F.
In theory the problem of assigning the remaining effects and interactions to
the alias sets with the correct signs presents no difficulties; all one has to do is to
write down all the 2n - k cosets of the defining contrast subgroup and cross out
those interactions that are to be suppressed. Unfortunately, that procedure
rapidly becomes a burden as the number of factors increases. We have considered in detail a 2 - 4 design; even in a 215 -11 there are only sixteen points but
there are 32,767 effects, including fifteen main effects and 105 two-factor
interactions. Before continuing we digress to consider another problem.
Given a set of 2" - k points which form a fractional factorial, how do we find
the set of defining contraets? This question has been known to arise in practice
160 Fractions of 2" Factorial Designs
[Ch. 8
when, by the time the experiment has been completed, the information on how
the fraction was constructed has been lost. It has been discuseed by Nelder
(1963). The method of solving the problems presented in this section is due to
Margolin (1967).
We assume that we have already found by inspection a basic set of factors or
Ietters. Margolin gives a simple method of doing this and calls the letters in the
basic set live letters and the others dead. It will be helpful to consider an example.
Consider the eight-point fraction bc, acde, abdf, ef, ag, bdeg, cdfg, abcefg. E, F,
and G are a set of live letters. The first step is to find a defining contrast that
contains A and some of the live letters, but no other dead letters. We shall then
do the same for B, C, and D, and the four contrasts will be a set of generators for
the defining-contrast subgroup. These contrasts are called the special defining
contrasts for the dead letters. Margolin gives two rules.
To determine the sign of the special contrast for any dead letter in the defining
relationship of the fraction, we examine the treatment combination in which all
(n k) of the live factors are at their high levels. If, at this point, the particular
dead factor is also at its high level, then the special contrast for that factor has a
positive sign. If the factor is at its low level, the contrast has a negative sign. In
the example the special contrasts for A, B, and C have positive signs; that for D
has a negative sign.
We now find which live factors appear in the special contrast of the dead
factor being considered. We consider each live factor in turn and look at two
points, the point at which all the live factors are at their low levels, and the
point at which the particular live factor is at its high level and the others are at
their low levels. If the dead factor is at the same level in both points, then the live
factor does not appear in the special contrast; if the levels are different, the live
factor is present.
In the example we consider first the special contrast for A. We have already
established that it has a positive sign. The point at which E, F, and G are all
at their low levels is bc. In each of the points acde, abdf, ag, A is at its high
level and therefore all three letters, E, F, and G, appear in the special contrast
for A, which is thus AEFG. The other special contrasts are BEG, CFG, and
DEF.
We return now to the problem of the aliases. We have A = EFG, B = EG,
C = FG, and D = EF. We next write down the Beven sets of three effects in the
live 2 3 that have the identity for their generalized interaction: (E, F, EF D),
(E, G, EG = B), (E, FG = C, EFG = A), (F, FG, G), (F, EG, EFG), (G, EF,
EFG), (EF, EG, FG). When the interactions are expressed as dead factors, the
products give the set of three letter defining contrasts,
DEF, BEG, ACE, CFG, ABF, ADG, BCD,

and the alias chains for the main effects and 2 f.i. follow immediately, A + CE +
BF DG, B + EG + AF CD, and so on.
Designs Obtained by Omitting 2n k Fractions 161
Sec. 8.9]
8.9 Designs Obtained by Omitting 2n k Fractions

-
We begin by considering an example. In the 2 4- k series there is nothing

between the half replicate of resolution IV, from which the main effects can be
estimated, and the complete factorial with sixteen points. The main effects and
the six 2 f.i. call for only ten degrees of freedom. Can we estimate all of them
with only twelve runs? John (1961a) showed that there were two ways in which
this could be done by omitting a quarter replicate from the complete design.
The problem arose in a practical situation: an experiment had to be planned
with four factors at two levels each and there was only enough raw material
available to make twelve runs. How might those twelve runs best be chosen?
Consider the family of quarter replicates defined by
(i)
I = +AB = +ACD = +BCD
d, ab, c, abcd;
(ii)
I = AB = +ACD = BCD
bd, a, bc, acd;
(iii)
I = +AB = ACD = BCD
(1), abd, cd, abc;
(iv)
I = AB = ACD = +BCD
b, ad, bcd, ac;
and suppose that we omit the first of them. The remaining fractions form three
overlapping half replicates
(ii) + (iii) I = BCD,
(ii) + (iv) I = AB,
(iii) + (iv) I = ACD.
We can estimate A from (ii) + (iii) sine there A is aliased with ABCD; A is
not, however, estimable from either of the other two half replicates. B is estimable from (iii) + (iv); C, D, and CD from (ii) + (iv); AC and AD from (ii) +
(iii); BC and BD from (iii) + (iv). We shall see later that these are the least
squares estimates. AB is estimable from both (ii) + (iii) and (iii) + (iv); the
least squares estimate is the average of these two estimates.
We recall the model
E(y) = Po + $i x= +
t
$1xx1 + .. .
s
that we presented before, and again consider the normal equations X'Y =
X'XO. The alias sets for the quarter replicates are I, AB, ACD, BCD; A, B, CD,
ABCD; C, ABC, AD, BD; and AC, BC, D, ABD. We write the elements of the
parameter vector f3 in that order: P = (9p, P12, $134, $234, Nl, N2, P349 $1234, )
'
The cross-product matrix, X'X, for the complete desigr. is 2 4I4 . The crossproduct matrix for the omitted fraction (i) consists of submatrices 4J 4 along the
main diagonal and zeros elsewhere. The cross-product matrix for the twelvepoint design is the differente between these two matrices and consists of submatrices 16I 4 4J 4 along the diagonal. Each of these submatrices is of rank 3.
When we suppress from the model the higher order interactions and strike out
[Ch. 8
the corresponding rows and columns of X'X, the first submatrix is reduced to
16I2 4J2 with inverse (2I 2 + J2 )/32 and the other three submatrices become
16I 3 4J 3 with inverses (I 3 + J 3 )/16.
The inverse of the cross-product matrix is the variance-covariance matrix of
the least squares estimates of the parameters. The estimate of A, for example,
has variance V(P 1 ) = (1 + 1)x 2 /16 = a 2 /8. This is the variance of the estimate
from the half replicate (ii) + (iii), and thus, since the least squares estimates are
the unique minimum variance estimates, the estimate from the half replicate is
indeed the least squares estimate of A. Similarly, the variance of the average of
the estimates of AB from the two intersetting half replicates (ii) + (iii) and
(iii) + (iv) is 3a 2 /32, which is the same as (2 + 1)a 2 /32, the variance of the least
squares estimate.
The calculations in the previous example were simplified because we chose
to omit the fraction for which the defining contrasts each had positive signs.
The general case is only slightly more complicated. Suppose that we have omitted
from a 2n factorial any 2n - k fraction. We can again write the elements of the
parameter vector in order of alias sets. Then the cross-product matrix for
the omitted fraction will consist of identical square submatrices 2n - k P along the
diagonal and zeros elsewhere. The elements of P may be positive or negative,
depending on the signs of the defining contrasts. They do, however, satisfy the
following three conditions:
R^ = P>i; p = 1 ; Ph>Pi> = Pni.
For any matrix P m of order m with elements satisfying these conditions it
follows that P, = mP m ; J m is a special case of the class of P matrices. This is
true in particular of the matrix obtained by striking out any (2k m) columns of
P and the corresponding rows.
Proceeding as before we see that the cross-product matrix for the actual
design consists of singular matrices 2nI 2n - kP along the main diagonal and
zeros elsewhere. If we can find at least one effect in each alias set that can be
suppressed, we can strike out that row and column and obtain a nonsingular
matrix. We can readily show that when aI m + bP m is not singular, its inverse is
cI m + dP m (Pm being the same in both matrices), where ac = 1 and d = bc/
(a + mb). In this instante we need the inverse of 2'I m 2n - kP m , which is
cI m + dP m where c = 2 - n and d = 2 - n(2k m) 1
We have now shown that if none of the effects in an alias set is suppressed,
none of them is estimable, but, if at least one effect is suppressed, the others are
estimable. Furthermore, if m of the effects in an alias set are to be estimated,
their least squares estimates will all have the same variance (c + d)a 2 or
.
V(P) = 2 - n(2k m + 1 )a 2 /( 2k m).

The covariance of estimates of parameters in the same alias set is doe or
2 - n02 /(2k m). The covariance of estimates of effects in different alias sets is
zero
Three-Quarter Replicates 163
Sec. 8.101
We now turn to the problem of finding estimates that have these variances.
If we can do so, they will be the unique least squares estimates, and the covariances will look after themselves. Now that we have established the variancecovariance matrix of the estimates, there will be no loss of generality if we
simplify matters by assuming that each of the defining contrasts in the omitted
fraction has a positive sign. Suppose that it is defined by I = P = Q =
PQ = . The actual design then consists of 2k 1 overlapping half replicates
defined by I = P, I = Q, I = PQ, and so on. Each pair of these half
replicates has a quarter in common: for example, the half replicates defined by
I = P and by I = Q both share the points for which I = P = Q = +PQ.
Let S be an effect that is to be estimated. In the half replicate defined by
I = P, S is aliased with PS, and if PS is suppressed, S can be estimated from
that half replicate. If 2" m of the effects in the same alias set as S are suppressed, S is estimable from 2k m overlapping half replicates. Each of these
estimates has variance 2 -( n -1) a 2 ; their covariance is 2 - no 2 . Let P denote the
average of these estimates. Then
V(P) = ( 2k m) -2{(2k m)2
-(n-1) + (
2k m)(2k m 1)2 - n }a2
= 2 n(2k m + 1 )a 2 /( 2k m),
-
which is the same as the variance of the least squares estimate.

We have thus established the following method of calculating the least
squares estimates of effects. To obtain the least squares estimate of an effect S,
take the estimates of S from each of the half replicates in which it is estimable
and average them.
8.10 Three-Quarter Replicates

The simplest designs of this type are the three-quarter replicates or 3(2n -2 )
designs [John (1962)]. These fractions are obtained by omitting one of the
quarters from a full factorial and are said to be defined by the missing quarter.
Thus the twelve-point design just presented is the 3(2 4-2 ) design defined by
I = AB = ACD = BCD. Had the set of points b, ad, bcd, ac been omitted
instead, we should have had the 3(2 4 2 ) defined by I = AB = ACD =
BCD. We can add other factors to 3(2n -2 ) designs and obtain 3(2n - k) designs.
The smallest of these designs has six points. We can omit from the 2 3 factorial
any one of the quarters defined by I = AB = AC = BC. These are (1),
abc; a, bc; b, ac; ab, c. The alias sets are I, AB, AC, BC and A, B, C, ABC. Each
of the main effects is estimated from the half replicate in which it is aliased with
ABC. The 2 f.i. and the mean are hopelessly confounded with each other. The
design is a resolution IV design. The variance efficiency of the estimates of the
main effects is 8/12 = 2/3.
This design can be made into a 3(2 5 - 3 ) fraction of resolution 111 by adding
two more factors, D and E. They are equated to any two of the interactions
164 Fractions of 2 Factorial Designs
[Ch. 8
AB, AC, BC. Suppose that we let D = AB and E = BC. Then D and
E are estimated from the half replicates in which they are aliased with A C. If we
omit the quarter defined by I = AB = AC = +BC, the six points are
(1), abc, bde, acde, abe, ce.
A and D are estimated from the half replicate defined by I = BC. The A
contrast is abe + acde ce bde. The D contrast is acde + bde abe ce.
B and the general mean are estimated from the half replicate defined by I =
+AC. The B contrast is abc + bde (1) acde; 4P 0 = ( 1) + abc + bde +
acde. C and E are estimated from the half replicate defined by I = + AB. The C
contrast is abc + ce (1) abe; the E contrast is abe + ce (1) abc.
In general, any 3(2' -2 ) fraction can be made into a design of resolution III
for 3(2n -2 ) 1 factors. New factors are assigned to interactions in such a way
that three of the effects in each alias set correspond to main effects and one is
left over. Each main effect is then estimated from the half replicate in which it is
aliased with the suppressed effect that is left over. In this way we can convert the
3(2 4-2 ) fraction mentioned earlier into a design of resolution III for eleven
factors. We can let E = AB, F = ACD, G = CD, H = ABC, J = AD,
K = AC, M = BC with BCD, ABCD, BD, and ABD left over. The design would
then be
(ii)
bdfhk afghm bcfjm acdfgjk
(iii)
egjkm abdej cdegh abcehkm
(iv) bghjk adhjm bcdgm ack,

where (ii), (iii), and (iv) refer to the original quarters. The main effects B, E, H,
and M are estimated from (iii) + (iv); D, F, G, and J are estimated from (ii) and
(iv); the mean, A, C, and K are estimated from (ii) + (iii).
The design that we gave earlier is not the only 3(2 4-2 ) of resolution III. We
could also have omitted one of the quarters defined by I = A = BCD =
ABCD. Some experimenters might prefer to use this fraction because in it the
one effect that is estimated from two half replicates is the main effect A rather
than a 2 f.i.
The 3(24-2) defined by I = AB = AC = BC and the 3(2 4 - 2 ) defined by
I = AB = CD = ABCD are also of interest. Neither is of resolution V, the
latter having an alias set, AC, BC, AD, BD, which contains only two-factor
interactions. Both, however, are of resolution IV and can be made to accommodate six factors. Consider in particular the first design. The alias sets are I, AB,
AC, BC; AD, BD, CD, ABCD; A, B, C, ABC; and b, ABD, ACD, BCD. We
can add two more factors, E and F, by setting them equal to two of the interactions ABD, ACD, BCD. The main effects will be clear of 2 f.i.; the
2 f.i. will be hopelessly confounded. If we let E = +ABD and F = +ACD,
the twelve points will be
aef bcef be ace abf cf
bcd
ad
acdf bdf
cde
abde.
Sec. 8.11 ]
The 3(2 5 - 2 ) Designs
165
It should be noted that this is a fold-over design and that it can be divided into
six blocks of two runs each without losing its resolution IV property. The blocks
are the columns of the array in which we have written the points.
This illustrates a general procedure for obtaining a resolution IV design for
3(2' 3 ) factors in 3(2 11 - 2 ) runs. We take a 3(2 11 - 2 ) fraction defined by omitting
a quarter replicate in which every defining contrast contains an even number of
letters. The alias sets then consist either of four effects each with an even
number of letters, or else of four effects each with an odd number of letters.
The new factors are added by equating them to odd interactions, leaving one
over in each alias set. The 2 f.i. will then all have an even number of letters, and
so they will not appear in the same alias sets as the main effects.
-
8.11 The 3(25_2) Designs

We have already seen that the 2 -1 fractions defined by I = ABCDE are
saturated fractions of resolution V with sixteen points. It is interesting to
investigate the gains, if any, that can be made by using twenty-four points.
To what extent can we, by a suitable choice of twenty-four points, obtin better
estimates, i.e., estimates with smaller variances? It is not immediately obvious
that any improvement can be made, apart from the fact that we shall have an
estimate of error with 8 d.f. from the actual data; after all, in a three-quarter
replicate some of the effects are only estimable from a single half replicate,
which is what we have in the 2 5 -1 designs.
Apart from permutations of the letters and choices of sign in the defining
contrasts of the omitted fraction, there are only six 3(25 _2) fractions of resolution
V. The defining contrasts and alias sets for them are given in the table on page
166.
Designs 4 and 5 have the property that each alias set contains exactly two of
the effects that are to be estimated. In these designs we have V(ftn ) = V (Ph) =
3a/64; the half replicate gives V(h) = V(Pht) = c 2 / 16. Designs 1, 2, and 3 each
have three alias sets with only one suppressed effect. In each case these alias sets
contain main effects, which means that in these designs several of the main
effects are only estimated from single half replicates. Taking the extra points has
not improved the precision of the estimates of those main effects.
When we consider blocking, the differences between the designs become more
apparent. The choice of the blocking variables in one of these designs affects
not only the number of blocks but also their sizes [John (1964b)]. When one of
the defining contrasts is used as a blocking factor in a 3(2n -2 ) design, we have
two blocks; one of these is a half replicate and the other is the remaining
quarter. If one of the other effects is used, we have two blocks of equal size.
If two effects in the same alias set are used for blocking, their product, which is a
defining contrast, is also a blocking factor. Thus, if two of the defining contrasts are blocking factors, so is the third, and the blocks are the quarters; if one
of the other alias sets contains two blocking factors, there are four blocks, two
I-'
A
AB
AC
C
BC ABC
E
AE
BE ABE
CE ACE
BCE ABCE
AB
B
ABC
C
AC BC
E
ABE
AE BE
CE ABCE
ACE BCE
R= ACD
S= BCE
RS = ABDE
CDE
BCDE
DE
BDE
CE
BCE
E
BE
AB
A
C
ABC
BC AC
D
ABD
BD AD
CD ABCD
BCD ACD
BCDE
CDE
BDE
DE
BCE
CE
BE
E
ABCDE
ACDE
ABDE
ADE
ABCE
ACE
ABE
AE
ACD
CD
AD
D
ACDE
CDE
ADE
DE
ABCD
ACD
ABD
AD
ABCDE
ACDE
ABDE
ADE
BCD
CD
BD
D
BCDE
CDE
BDE
DE
R = ABE
S = ABCD
RS = CDE
A
AB
C
AC
BC ABC
AD
D
BD ABD
CD ACD
BCD ABCD
ABCD.E
ACDE
ABDE
ADE
ABCE
ACE
ABE
AE
BCD
ABCD
BD
ABD
BCDE
ABCDE
BDE
ABDE
A
B
AB
D
AD
BD
ABD
ACDE
CDE
ADE
DE
ACD
CD
AD
D
R = ACE
S= BCD
RS = ABDE
ADE
DE
ABDE
BDE
C
ABCD AE
BCD E
ACD ABE
BE
CD
ABC
BC
AC
AB
B
C
ABC
AC BC
E
ABE
AE
BE
CE ABCE
ACE BCE
BCDE
ABCDE
CDE
ACDE
BCE
ABCE
CE
ACE
BCDE
ABCDE
BDE
ABDE
BCD
ABCD
BD
ABD
Sec. 8.13]
Adding Fractions 167
with 2n 2 points each and two with 2n 3 points each. Finally, if two effects in
different alias sets, neither of which is a defining contrast, are used for blocking,
their product is in a third alias set and we have four blocks of equal size. In all
cases, for an effect to be estimable there must be a suppressed effect in the alias set
in addition to any blocking factors.
Designs 4, 5, and 6 can each be blocked into four blocks of six points each by
using as blocking factors the three factors R, S, and RS, which are given in each
case. Design 6 can be run in separate quarters. Designs 1, 2, and 3 may be split
into blocks of sizes 8, 8, 4, and 4. A particularly interesting property of designs 4
and 5 is that they may be blocked, using ABCDE, into the half replicate defined
by I = ABCDE and the remaining quarter. An implication of this is that we
may elect to complete either of these designs if, after the half replicate has been
run, we find ourselves able to make an additional eight runs. Alternatively,
we may choose to run the half replicate first anyway, as a precaution against
finding ourselves in a situation in which we are not able to complete the twentyfour point sequence.
-
8.12 Some 3(2' 1- k) Designs of Resolution V

The smallest 2 8 - k design of resolution V is a quarter replicate with sixty-four
points. The following 3(2 8-4 ) fraction has resolution V in forty-eight points.
Omitting one of the factors gives a design of resolution V for Beven factors in
forty-eight points. We omit from the complete 2 6 factorial the quarter defined by
I = ABE = CDF = ABCDEF, and add new factors by putting G = ABCD
and H = A CEF. The three quarters of the design are
(i) fg, aef, befh, abfgh, c, aceg, bcegh, abch, dh, adegh, bdeg, abd, cdfgh,
acdefh, bcdef, and abcdfg;
(ii) eg, a, bh, abegh, cef, acfg, bcfgh, abcefh, defh, adfgh, bdfg, abdef, cdegh,
acdh, bcd, and abcdeg;
(iii) gh, aeh, be, abg, cfh, acefgh, bcefg, abcf, df, adefg, bdefgh, abdfh, cdg,
acde, bcdeh, and abcdgh.
The three quarters may be used as blocks.
With ninety-six points we can obtain a design of resolution V for eleven
factors. This is achieved by omitting from the 2 factorial the quarter defined by
I = AB = CDEFG = ABCDEFG and adding new factors H = ABDEF,
J = ABCEG, K = AEFG, and L = BCDE. The alias sets are presented in
John (1969).
8.13 Adding Fractions

The three-quarter replicate can fust as well be obtained by adding a quarter
replicate to a half replicate. The addition of other fractions to a half replicate
168 Fractions of 2 Factorial Designs
[Ch. 8
does not, however, give such neat results. There are some useful designs that can
be obtained in this way and in them, too, the least squares estimates of the
estimable effects can be pieced together from the estimates from the constituent
fractions. We shall illustrate the procedure by an example and state the results
for the general case.
Suppose that we take a 2n 1 design defined by I = P where P is some effect,
and add to it the 2n 3 fraction defined by I = P = Q = PQ = R = PR =
QR = PQR. Let S be some other effect. The alias set, which contains S, consists
of S, PS, QS, PQS, RS, PRS, QRS, and PQRS. This alias set is composed of
four pairs of effects, such as S, PS, which are themselves aliased in the half
replicate. If one of the effects in each pair is suppressed, the other four effects
can be estimated from the half replicate without the need for any additional
points. The extra points are needed only to break chains of aliases, such as
S PS, when neither effect is suppressed.
As an example, suppose that in a 2 4 factorial we take the half replicate defined
by I = ABCD and add the 2 3 defined by 1= ABCD = A = BCD = B =
ACD = AB = CD. There are two alias sets:
-
M, ABCD; A, BCD; B, ACD; AB, CD;

C, ABD; AC, BD; BC, AD; D, ABC.
As usual, we suppress the interactions with three or four factors.
We shall call a pair of effects such as S and PS a nonnegligible pair if neither
S nor PS is suppressed. The first alias set contains one nonnegligible pair
AB, CD; the second set has two such pairs A C, BD and AD, BC. The points in
the design are
abc, abd, a, acd, b, bcd, c, d, ab, abcd.
We can take the extra eighth replicate (ab, abcd) and combine it in turn with
each of the eighths in the original half replicate that have the same defining
contrasts. This will give us four overlapping 2 2 fractions:
-
1= A = B = AB,
abc, abd, ab, abcd;
I = A = CD = ACD,
a, acd, ab, abcd;
I = B = CD = BCD,
b, bcd, ab, abcd;
1= AB = ACD = BCD,
c, d, ab, abcd.
From the first of these quarters we have an estimate of CD: CD =

(ab + abcd abc abd)/4. In the half replicate the CD contrast gives an
estimate of CD AB: (a + acd + b + bcd abc abd c d)/8. If we
subtract this from the estimate CD, we obtain an estimate of AB:
B=(2ab+2abcdabc abd a acd b bcd +c+d)/8.
Adding Fractions 169
Sec. 8.13]
Both CD and AB have variances a 2 /4. Also the mean, A, and B are estimable
from the original half replicate. Thus all the effects in the first alias set that are
not suppressed are estimable.
We find in considering the submatrix of the cross-product matrix that if we
delete the suppressed effects and write the remaining effects in the order M, A, B,
AB, and CD the submatrix is
U1 = 2
4I 3 + J 3E

3,2 l
8I 2 3J 2 )
E2,3
where E m is a matrix of m rows and n columns with every element unity. The
corresponding covariance submatrix is
,,,
= 2 _ 4 2I3 E
( E I2 + 3J 2 )
Ui
This shows that V(P 0 ) = V(P1 ) = V(j2 ) = a 2 /8 and V(P12) = V(P34) =

and so the estimates that we have given for M, A, B, AB, and CD are the least
squares estimates.
On the other hand, the cross-product submatrix for the other alias set is
4I 2 + J 2 J2
U 2 = 2
J2
J2
8I2 3J 2 J2
J 2 J2
812-3J2
where the first two columns correspond to C and D, the next two to the nonnegligible pair AC and BD, and the last two to the nonnegligible pair AD and
BC. The matrix U 2 is singular. The main effects C and D are estimable from the
original half replicate, but the addition of the two extra points does not break
either of the chains AC BD or AD BC.
The results in the general case where a 2' - k fraction is added to a disjoint half
replicate are as follows:
(i) Any effects that are estimable in the half replicate are estimable in the
augmented design (the worst that can happen is that the estimates will
be the same);
(ii) If an alias set contains more than one nonnegligible pair, the only
estimable effects that can be estimated are those that are estimated in the
half replicate, and the estimates from the half replicate are the least
squares estimates of the effects from all 2n -1 + 2n k points;
(iii) If an alias set contains only one nonnegligible pair and, also, m pairs
with one effect suppressed and one to be estimated (m < 2k -1 1),
all the effects are estimable.
-
[Ch. 8
In the last case, when m < 21c -1 1, and in the case in which an alias set has
no nonnegligible pairs, the least squares estimates are not simply the estimates
from the half replicate. They may be found by following the procedures derived
in John (1966b) or by using the results obtained by Plackett (1950).
This technique of adding fractions may be used to obtain a design of resolution
V for seven factors in thirty-six runs. The smallest such 2 7 k fraction has
sixty-four runs; any of the half replicates having as the defining contrast an
interaction involving five, six, or seven factors will sufface. A design with fortyeight runs is obtained by using a 3(2 7 - 3 ) fraction; one can, for example, omit the
quarter defined by I = ABC = DEF = ABCDEF from the complete 2 6 design
and add a new factor by putting G = ABCD. The thirty-six point design is a subset of the forty-eight point fraction. It is obtained as follows. We add to the
26-1 defined by I = ABCDEF the four-point fraction generated by I =
+ABCDEF = +ABDEF = ABC = +BCDEF = ...; the seventh factor is
again added by putting G = ABCD.
Also of interest is an eighty-point design of resolution V for nine factors.
This is obtained by adding to the 2 -1 defined by I = AB the 2 7 - 3 defined
by I = AB = CDEFG = ABCDEFG = ACF = BCF = ADEG = BDEG and
then putting H = AEFG and J = BCDE.
-
8.14 Nested Fractions

We have already mentioned that one of the differences between industrial and
agricultural experimentation is that in the former the runs can often be made in
sequence. Advantage can be taken of this in planning an experimental strategy.
There is also a disadvantage because experiments may by misfortune be terminated before all the runs have been made. The designer would be prudent to
anticipate such risks.
We expect the size of an experiment to be determined in the planning stage
by budgetary considerations or, perhaps on some occasions, by calculations
involving the power of the t or F tests. Very often, however, time plays an
important part because the scientist is allotted a pilot plant for only a limited
time. For example, he might have the use of a pilot plant for two weeks, or for
ten working days. If he can make two runs a day, and if he allows the first day
to set up his equipment, that gives him eighteen runs. A 2 4 design followed by
two check points on the last day might be appropriate.
But what are the consequences if something goes wrong with the machinery
and three days are lost while a pump is being replaced? Such misfortunes are
not uncommon. The loss of three days in the experiment mentioned in the
previous paragraph would mean that instead of eighteen runs the experimenter
could only carry out the first twelve of the sixteen points in the 2 4 design.
Suppose that the last four runs that had been scheduled (but never made) had
been (1), abd, cd, abc. Then he would have been lucky because the twelve points
that were actually carried out would have formed a 3(2 4-2 ) of resolution V.
Sec. 8.14]
Nested Fractions
171
On the other hand, if he had lost (1), bd, bcd, ab he would have been left with a
twelve-point fraction that can, at best, be called messy. The only half replicate
that it contains is the one defined by I = + ABD.
This suggests that the experimenter should modify complete randomization
by arranging his runs so that if the experiment ends prematurely he still has a
useful fraction to analyze. The experimental strategy should be directed toward
designs in which fractions are nested within the overall design so that a partially
completed design has a high salvage content.
An obvious possibility is to choose as the first eight points one of the fractions
defined by I = + ABCD or by I = ABCD so that even if he finishes only half
his assigned runs he does at least have a resolution IV design. He might then
follow that by a set of four points at the same level of one of the factors, and
thereby obtain one of the 3(2n -2 ) replicates of the type I = A = BCD = ABCD.
We have already noted that the 3(25 _2) fraction defined by I = E = ABCD =
ABCDE can be split into a half replicate and a quarter replicate by using any of
the defining contrasts as a blocking factor. This indicates two strategies. An
experimenter who starts on the twenty-four point sequence but doubts his
ability to finish the program can make use of the fact that the 2 5 -1 design of
resolution V defined by I = ABCDE is included in his set of points and he may
decide to run that half replicate first. On the other hand, an experimenter who
had carried out a complete 2 design in A, B, C, and D (holding E fixed) could,
if he found himself with time for another eight runs, proceed to carry out one of
the 2 -1 fractions defined by I = ABCD at the other level of E, thus converting
his 2 design into a 3(25 _2) design of resolution V.
Indeed, if after he has carried out the 3(2 5- z) fraction defined by I = E _
ABCDE = ABCD which consists of all points except the omitted fraction
e, ade, bde, abe, cde, ace, bce, abcde,
he wishes to go further and add the set
e, ade, bde, abe,
he would then have a twenty-eight point 7(2 5 -3 ) design with alias sets
I, E, ABCDE, ABCD, ABD, ABDE, CE, C;
A,AE, BCDE, BCD, BD, BDE, ACE, AC;

B,BE, ACDE, ACD, AD, ADE, BCE, BC;
AB, ABE, CDE, CD, D, DE, ABCE, ABC.
Each alias set contains at least one four-factor interaction, and so the design is of
resolution VII.
[Ch. 8
8.15 Resolution III Designs

A resolution III design is defined as a design in which all the main effects are
estimable when all the interactions are suppressed. It can be shown, by following
the arguments that will be used in the next section for resolution IV designs,
that in a resolution III design the mean Po must also be estimable. Hence, a resolution III design for n factors must contain at least n + 1 points. We shall
use, henceforth, the notation 2n/ /N to denote an N-point fraction of a 2'^
factorial.
We have already seen examples of designs in which this lower bound is
attained. These are the saturated designs for 2n 1 factors in 2n runs, and for
3(2n - 2) 1 factors in 3(2n 2 ) runs. The former series of designs have variance
efficiency one; in the 3(2n - k) series of saturated designs the efficiency is only
two-thirds and, if n + 1 is divisible by four, it may be preferable to use the
Plackett and Burman designs (1946). These are main effects designs for 2' 1 //N
with N = 4k where k is an integer. They have variance efficiency one. The
Plackett and Burman designs will be discussed with the main effects plans of
Addelman in the next chapter.
Other main effects plans have been considered under the heading of weighing
designs by Hotelling (1944a), Kishen (1945), Mood (1946), Raghavarao (1959),
and Yang (1966) and (1968). A discussion of them is found in the paper by
Margolin (1969a). We shall merely point out here that a 2n/ /n + 1 design of
resolution III can always be obtained in the following way. The ith run,
1 < i < n, has the i th factor at its low level, and all the other factors at their
high levels; the (n + 1)th point has all the factors at their low levels.
Taking the model
E(Y) = Po +
we have normal equations X'Y = X'X where X is the design matrix and includes
a vector 1 for P o . The moment matrix X'X and its inverse, the covariance matrix,
are
X'X =
+ (n 3 )J71+1,
(X'X) -1 = 4
)J
4(n 3
so that V($t ) = [(n 2 3n + 4)/4(n 1) 2]a 2 and 1im 71 . V(f) = a 2 /4.
We can do considerably better by using Yang's results. We can also do

considerably worse by replacing the last run with the point at which all the
factors take their high levels. We should then have the classical one-at-a-time
design. This is a resolution III 2n/ In + 1 design with V(Pt ) = a 2 /2 for each i and
for all n.
Sec. 8.16]
Designs of Resolution IV 173
8.16 Designs of Resolution IV

Webb (1968), noting that folding over converted a 2n - k design of resolution III
into a 2n+1- k fraction of resolution IV, showed that folding over will convert
any resolution III design for n factors into a resolution IV design for n + 1
factors. He also showed that any resolution IV design for n factors must contain
at least 2n points. He called such a design with only 2n points a minimal design
and conjectured that any minimal resolution IV design must be a fold-over
design. This conjecture was proved by Margolin (1969e); Margolin had proved
it earlier (1969b) in the special case in which each factor appeared exactly n times
at its high level and n times at its low level.
A design is of resolution IV if, and only if, the main effects are estimable when
all interactions with three or more factors are suppressed. The mean may, or
may not, be estimable. We have already mentioned the resolution IV designs
for 2 3 / / 6, 3(2a_2). In these designs the two alias sets are I, AB, AC, BC, and
A, B, C, ABC, and the mean is not estimable unless at least one of the 2 f.i. is
suppressed.
We consider a model that contains only the mean, the main effects, and the 2 f.i.:
E(y) =
Phixhxi,
P O + 1h 8hxh + 11
h i
1 < h < i < n,
and a design with N points.

The normai equations are
X'Y = X'X,
where X is a matrix of 1 + n + n(n 1)/2 columns and N rows.
We denote by X h the vector with elements xh; where x hf is the value of X h for
the jth datum point; similarly Xhi is the vector with elements xhjxif ; X hi is said to
be the Hademard product of X h and X. The vector X 0 = 1 corresponds to the
mean; X 0
X,. . , Xhi , ... are the columns of X.
If all the main effects P h , h > 0, are to be estimable, then the vectors X h ,
h > 0, and Xo must be linearly independent. Furthermore, the main effects
vectors Xh must be linearly independent of the interaction vectors X hi . Suppose
that this were not so and that, for example, X 1 = Xc where c is a vector having
e l = 0 and at least one element other than c 1 not zero. We are given that P 1 is
estimable. Let 1 = a'Y. Then
, ... ,
E(N1) = Nl
= a ' E(Y) = a' {X0P0
+1X +11
hNh
Xhifht},
and, since this is an identity, a'X 1 = 1 and a'X 0 = X h = a'X hi = 0, h > 1,

which contradicts the assumption that X l is a linear combination of the other
vectors. The converse, that if all the main effects vectors are linearly independent
of the mean and interaction vectors, then the main effects are estimable, is
easily shown.
174 Fractions of 2' Factorial Designs
[Ch. 8
Consider now the set of 2n vectors X l , X2 ,.. ., X0, X12, X13, .... The elements
of X l ,, are obtained by multiplying each element xh1 of X h by x 11 , which is equal
to 1; the elements of X o are the squares of the elements of X 1 . The vectors X h
are linearly independent, i.e., there exists no set of scalars Ch such that l h Ch xhj =
0 for all j. In particular, it is not true that either + 1(2 xh f) or 1 (1 xh f) = 0
for each j. Hence, the set of vectors X 0 , X 12 , ... are Iinearly independent, and
it has already been shown that X l , X 2 , ... are independent of them. It follows
that X l , X 2 ,. . ., X 0 , X 12 ,..., is a linearly independent set of 2n vectors. Thus
r(X) = r(X'X) > 2n. But r(X'X) < N and so N >_ 2n, which proves that every
resolution IV design for n factors must contain at least 2n points. Since we have
already seen several examples of resolution IV 2n/ /2n designs, we know that
this lower bound is attainable.
We now show that when a resolution III design for n factors is folded over, the
resulting design is of resolution IV and can indeed be made into a resolution IV
design for n + 1 factors. We consider the (n + l)th factor to be held at its low
level throughout the resolution III design (if we choose to ignore this factor
there will be no difficulty). We write the X matrix for the resolution III design as
X=(1 U V),
where 1 corresponds to the mean, U = (X 1 , X 2 ,. . ., Xttt1 and V consists of the

vectors X,,,. The rank of U is n + 1. For the folded-over portion, with the
(n + 1)th factor now at its high level,
)
X=(1 U V).
For the combined design

(1 U
X'X =
U V)
2N
1
(1
U V)
21'V
0 2U'U 0
2V'V
2V'1
0
The main effect submatrix 2U'U is nonsingular. The estimable main effects are
orthogonal to the mean and the interactions.
We have already seen that there are elementary weighing designs for n 1
factors in n runs with resolution III. Folding them over gives minimal resolution
IV designs 2/ /2n for any n. These are not necessarily the most efficient 2n/ /2n
designs, and some more efficient designs are given in the papers by Webb (1968)
and Margolin (1969a) that have already been cited. We shall not give Margolin's
proofs of Webb's conjecture that all minimal resolution IV designs are foldover designs. The reader is referred to Margolin's original papers.
Exercises
1. Analyze the fraction obtained by omitting the quarter defined by I = A =

BCDE = ABCDE .
Exercises 175
2. (Continuation) Analyze the 3(2 5-1) obtained by omitting the quarter defined by
I = AB= CDE = ABCDE.
3. Find the defining contrasts for the following 2 6-3 fraction: c, ae, bd, adf, bef,
abcf, cdef, and abcde.
4. The point (1) is added to the 2 3-1 fraction, a, b, c, and abc. Find the least squares
estimates of the main effects in the following cases:
(i) All interactions are suppressed;
(ii) All interactions except AB are suppressed;
(iii) All interactions except ABC are suppressed.
5. Obtain a 7(2 7 e) fraction of resolution IV and divide it into seven blocks of two
runs each without losing the resolution IV property.
6. An experimenter runs two half replicates of a 2 4 factorial. They are defined by
I = ABC and I = ABD respectively. There are sixteen observations altogether
since the four overlapping points are duplicated. How would you analyze the
data, assuming that all higher order interactions are suppressed? How is the
analysis modified if the two half replicates are in separate blocks?
7. Healy (1951) gives a plan for a single replicate of a 2 factorial in a Latin rectangle,
i.e., a rectangular array of thirty-two plots arranged in four rows and eight
columns. The interactions ABC, ABDE, and CDE are confounded with the rows.
The first four columns consist of the 2 5 -1 defined by I = BCE, and AB, CD,
ABCD are confounded with columns. The second four columns contain the
remaining sixteen points with AC, DE, ACDE confounded. Reconstruct his
design. Assuming that all higher order interactions are suppressed, confirm that
all the main effects and two-factor interactions are estimable and show how they
are estimated.
8. Finney (1946) describes a potato growing experiment in which four organic
fertilizers were applied in a Latin square design. Each of the sixteen plots was
split into two subplots to which treatment combinations from a 2 2 factorial in
N (nitrogen) and K (potash) were applied. Each whole plot either had (1) on one
subplot and nk on the other, or n on one and k on the other. The following data
are yields of potatoes.
-
D
61
52
n
k
B
(1) 51
A
(1) 48
nk 67
nk 57
A
(1) 32
nk 61
n
k
C
n
k
85
69
80
(1) 85
68
nk 87
68
D
(1) 72
50
nk 94
n
k
93
C
(1) 84
nk 104
n
k
C
n
k
B
n
k
D
70
B
(1) 96
nk 91
A
n
k
91
66
A
75
62
B
85
85
D
(1) 89
nk 83
[Ch. 8
Analyze the data. Is it necessary to suppress the NK interaction? How would the
analysis be modified if A, B, C, and D were the four-treatment combinations in a
2 2 factorial? How would it be modified if they were a 2 3 1 of resolution III?
Finney always confounded NK between whole plots. Could he perhaps have
confounded the main effects N and K sometimes? Would he have gained anything
by it?
'
CHAPTER
Fractional
Factorials
with More
Than Two Levels
The fractions of the 2n designs that were discussed in the previous chapter are
of considerable practical use as well as of academic interest. They enable us to
handle many factors without having to make an unreasonable number of
observations, and it seems only natural that experimenters should wish to seek
the advantages of fractionation with factors having more than two levels. New
complications then arise; not only do factors at more than two levels call for
more data points in the full design, but they also have more effects to be
measured.
In a 2,1 experiment with all interactions involving three or more factors
suppressed, there are 1 + n + n(n 1)/2 = ( n 2 + n + 2)/2 parameters to be
estimated. In a 3' experiment each main effect has two degrees of freedom,
corresponding to linear and quadratic effects; each two-factor interaction has
four degrees of freedom, giving a total of 2n 2 + 1 degrees of freedom. In a 5n
design the estimation of two-factor interactions soon gets out of hand; with
only three factors there are already 12 d.f. for main effects and forty-eight for
two-factor interactions.
The interest in fractionation of 3n, 4n, and 5n designs is not merely for mathematical amusement. The model that was used for 2fz designs contained no terms
in xh, and so, if the experimenter has cause to believe that for some factor there is
a quadratic component to be estimated, he must take three or more levels. (It
would be wrong to think that we always pay attention to estimating the quadratic
component, even when we suspect that it is there. In response-surface fitting,
which will be discussed in the next chapter, we often elect to ignore the quadratic
effects in the earlier stages of experimentation and to make decisions about
future experiments on the basis of first-order models.) The consideration of
curvature thus leads us to 3n factorials, and we shall begin this chapter by
recalling some of the properties of 3" designs described in Chapter 7.
177
178 Fractional Factorials with More Than Two Levels
[Ch. 9
We shall continue to use the concept of the resolution of a design that was
introduced in the previous chapter. In view of the fact that the number of twofactor interactions increases so quickly with n, resolution V designs are of little
interest and we shall not say much about them. The major use of these fractional
factorials is in screening experiments, and so most of this chapter will be concerned with resolution III designs, which will often be called main effects plans.
In the next section we shall recall the simplest of the main effects plans, the
Graeco-Latin squares, and follow them with a discussion of 3n ' factorials.
We shall then turn to the series of orthogonal resolution III designs, which are
of considerable practical interest, developed by Plackett and Burman (1946) and
by Addelman (1962). In that section we shall find some fractions of 2m3n designs,
i.e., designs with m + n factors, of which m appear at two levels and the other n
at three levels. The discussion of 2m3n designs will be developed further in later
sections, and the chapter will conclude with a few examples of nonorthogonal
fractions.
We have purposely restricted our discussion in this chapter to these topics.
The reader who is particularly interested may wish to read some of the earlier
papers on the subject of fractionation of sn designs where s = pm and p is a
prime. Among the papers of interest are some on orthogonal arrays. We note in
particular papers by Bose (1947a), Rao (1946) and (1947b), Bose and Bush
(1952), and Finney (1945).
An orthogonal array (N, k, s, d) is a set of N treatment combinations with k
factors each at s levels, which has the property that for any subset of d out of the
k factors all s' of the treatment combinations occur the same number of times;
d is called the strength of the array. An orthogonal array of strength 2 is an
orthogonal main effects plan. The levels of the factors will usually be denoted by
0, 1, 2, ..., s 1. If s is prime, these symbols will actually be elements of the
Galois field GF(s). We shall sometimes write the levels of a two-level factor
symbolically as zero and one; this will be purely symbolic and they will be
treated as the two elements of GF(2) with 1 + 1 = 0. In practice, the levels will
still be 1. The 2e ' defined by I = ABC may be written 000, 110, 101, and
011. Consider the subset of factors A, B. We see that the set of treatment
combinations 00, 01, 10, and 11 occurs once each. The same thing is true for the
treatment combinations of the subsets A, C and B, C. The design is thus an
orthogonal array (4, 3, 2, 2).
-
9.1 3n Designs
The 3 2 and 3 3 experiments were discuseed in Chapter 7. At that time we took
a 3 2 example and divided the four degrees of freedom for A x B interaction into
four components: lin A lin B, lin A quad B, quad A lin B, and quad A quad B.
We shall no longer use that method of subdivision of the interaction sum of
squares because it does not lead to reasonable procedures for finding fractions.
Instead we shall adopt a procedure used previously for confounding.
Sec. 9.2]
Latin Squares as Main Effects Plans 179
The three levels of each factor are represented by 0, 1, and 2, and arithmetic is
carried out mod 3. The twenty-seven points of the 3 3 design form a finite
geometry in three-dimensional space. The sum of squares for A may be divided
into two single degrees of freedom, corresponding to linear and quadratic
effects. The linear contrast is the difference between the sum of the observations
at x l = 2 and the sum at x l = 0, which we write as {x 1 = 2} {x 1 = 0}. The
quadratic contrast is {x, = 0} 2{x 1 = 1} + {x 1 = 2}. The equations x l = 0,
x l = 1, and x l = 2 determine hyperplanes (in three dimensions they are just
planes). The degrees of freedom correspond to the two orthogonal contrasts
between the totals of observations on the parallel planes.
For the A x B interaction we take the two sets of parallel planes x l + x 2
0, 1, 2 (mod 3) and x l + 2x 2 = 0, 1, 2 (mod 3). For each set of three planes
there are two orthogonal contrasts, and hence 2 degrees of freedom. We call the
first pair of degrees of freedom the AB interaction and the second pair the AB 2
interaction. Nothing further is gained by considering the planes 2x 1 + x 2
0, 1, 2 because the set of points with 2x 1 + x 2 - c has 2(2x 1 + x 2 ) - 2c, or,
reducing mod 3, x l + 2x 2 = 2c, and this again gives the AB 2 interaction.
The three-factor interaction A x B x C has 8 degrees of freedom and four
components, ABC, AB 2 C, ABC 2 , and AB'C 2 , corresponding to sets of hyperplanes, x l + x 2 + x 3 - 0, 1, 2, x l + 2x 2 + x 3 - 0, 1, 2, x l + x 2 + 2x 3
0, 1, 2, and x l + 2x 2 + 2x 3 - 0, 1, 2, respectively. It is conventional to write
the components of interactions with the exponent of the first factor unity. We
gave an example of confounding the 3 3 factorial in three blocks of nine points
each. Each of the blocks is a one-third replicate of the Pull factorial, which we
may write as a 33-1 fraction or, extending the notation that we introduced in the
last sections of the previous chapter, as a 3 1 / / 9 design. Indeed, each of the blocks
is a resolution III design, which leads us into the next section.
9.2 Latin Squares as Main Effects Plans

We originally presented the s x s Latin squares as designs for eliminating
plot differences in two directions: rows and columns. There is nothing in our
model or in the derivation of the analysis to prevent us from letting the rows
represent the s levels of one factor and the columns represent the s levels of a
second factor. Thus we see the Latin square of side s as an orthogonal main
effects plan for three factors, each at s levels in s 2 runs. The Graeco-Latin square
is a main effects plan for four factors, and so on.
When we divided the 3 3 design into three blocks of nine points, the first block
consisted of the points for which x l + 2x 2 + 2x 3 = 0. If we let x, denote the
row and x 2 the column, this gives the Latin square
000 012 021
101 110 122
202
211
220.
[Ch. 9
There are eight degrees of freedom in a 3 x 3 Latin square, and so there is

room for a fourth factor. We put x 4 = x l + x 2 and obtain a saturated main
effects plan or 3 4 / /9:
0000 0121 0212
1011 1102 1220
2022 2110 2201.
In the more general case, where s = pm, the s x s Latin square can be made to
accomodate s + 1 factors, each at s levels, in an orthogonal main effects plan.
An interesting pioneering example of a 5 5-3 , due to L. H. C. Tippett, is reported
by Fisher (1947). The factors involved were four different components of a
spindle used in a cotton mill, each at five levels, and time, observations having
been made at five different periods.
The case s = 4 = 2 2 was also considered in Chapter 7. Here we have two
possible methods of attack. The finite geometry approach that we have adopted
so far implies the use of finite fields. In the case of s = 3 or 5, the elements of
the field are the residue classes mod s. For s = 4 we can denote the levels of each
factor by 0, 1, x, and y (= 1 + x) with arithmetic mod 2. When we blocked the
4 2 experiment into four blocks of four, we took as the blocking factor C:
x 3 = yx l + yx 2 . This also gives us a Latin square 4 3 / /16 main effects plan:
000
l0y
x01
y0x
Oly
110
xlx
yll
Oxl
lxx
xx0
yxy
Oyx
lyl
xyy
yy0.
We also mentioned at that time the second method of attack, which involves
replacing each four-level factor by two two-level factors.
The remarks that were made earlier about the bias resulting from interaction
in Latin square designs are again relevant. In the last design A is aliased with
B x C, B with A x C, and C with A x B.
In addition to the Graeco-Latin squares, there are other larger 3' - k fractions
which are main effects plans. The principle of their constructiop is the same as
that for 2' - k designs of resolution III; we take a basic 3n design and add other
factors by equating them to interactions, or rather to components of interactions. The following example will illustrate the procedure. We shall obtain a
twenty-seven point resolution III 313-10 design and show how it becomes a
3 9-6 in nine blocks of three points each.
3n -k Fractions with Interactions
Sec. 9.3]
181
We begin by taking a 3 3 design in the three factors A, B, and C and then add
new factors by setting
x4
x7
+ 2x3,
x5
= 2X1 + x2 + x3,
= 2x 1 + x 2 + 2x 3 ,
x8
= x l + x2
x1
X10 = Xl
X13 =
+ X3,
x2+
x11 = Xl
x2
x3 ,
+ 2x3,
x6
x9 =
x12
x2
+ 2x3,
xl +
x2,
= xl + 2x2,
X3.
This is equivalent to taking as defining contrasts AC 2 D 2 , AB 2 C 2E, BC 2 F2 ,

AB 2 CG, ABCH 2 , ABJ 2 , ACK 2 , ABC 2L 2 , AB 2 M 2 , and BCN 2 . The twenty-seven
points in the design follow:
0000000000000
1122221120000
2211112210000
0221100122011
1010021212011
2102212002011
0112200211022
1201121001022
2020012121022
1001202111110
2120120201110
0212011021110
1222002200121
2011220020121
0100111110121
1110102022102
2202020112102
0021211202102
2002101222220
0121022012220
1210210102220
2220201011201
0012122101201
1101010221201
2111001100212
0200222220212
1022110010212.
Suppose now that we decide to block the design by using the last two factors,
M and N, as blocking factors, i.e., by assigning to the same block those points
that have the same value for x 12 and the same value for x 13 . There will be nine
blocks corresponding to the pairs of coordinates 00, 01, 02, ..., and in the
arrangement in which we have written the points they are the first three points,
the middle three points, and the last three points of each column. There are
eight degrees of freedom between blocks. Four of these are the 2 degrees of
freedom each for M and N. The remaining 4 degrees of freedom are for the
generalized interactions of M and N, namely K and L, which also become
blocking factors since we have
X12
X13 = xl + x3 =
x10,
x12
+ 2x13 = xl +
x2
+ 2X3 = x11.
The blocked design is thus a resolution III 3 9-6 design in the first nine factors.
9.3
3n
Fractions with Interactions
There is an extensive list of 3 fractions in the pamphlet prepared by

Connor and Zelen (1959) for the National Bureau of Standards. In it they
consider designs of resolution IV and resolution V with blocking for 4 < n < 10.
[Ch. 9
The smallest resolution IV design for n = 4 is the one-third replicate, of which

an example will be given in the next section. For 5 < n < 9, they have resolution
IV designs with eighty-one points. Among the resolution V designs, eighty-one
points are necessary for n = 4 or 5. For 6 < n < 9, resolution V designs are
listed with 243 points.
The 3 5 / / 81 design can be obtained by using any of the five-factor interactions
as a defining contrast. The resolution IV 3 9 / / 81 design is obtained by setting
x5 = x1 + 2x3 + x4,
x6 = 2x1 + x2 + x3 + 2x4,
x8 =2x 1 +x2 +2x 3 ,
x7 = x2
+ 2x3 + x4,
x 9 =x 1 +x 2 +x3 +x 4 .
They obtain their 3 9 / /243 design by adding new factors to the basic 3 5 design as
follows :
x6 = xl + x2 + 2x3 + X5,
x8 =
x2
+ x3 + 2x4 + 2x5,
X7=
2x1 + X2 + 2x4 + x5,
x9 = xl + x2 +
x3
+ x4.
EXAMPLE OF A 34-1 DESIGN. Vance (1962) reports on a series of experiments in the course of which a 341 factorial was used. There were four major
operating variables in the treatment of lube oil at a refinery, and these were
taken as the factors in the experiment. The objective of the experimental
program was to find a set of operating conditions that would optimize a measure
of quality in the lube oil, which will be referred to as the response. Their previous
experience in the use of the plant had convinced Vance and his tolleagues that
the quadratic effects should be estimated, and so at least three levels of each
factor should be included.
In this preliminary experiment they were prepared to allow two-factor interactions to be aliased with each other. Their main concern was to obtain preliminary estimates of the linear and quadratic effects, aliased only by higher
order interactions, and to use this information to provide a basis for the design
of subsequent experiments. This approach is similar to the procedure used in
fitting response surfaces, which will be discuseed in Chapter 10.
They had a choice of several possible fractions. In the 2 4 factorial there are
only two half replicates that have ABCD as the defining contrast. There are,
however, no fewer than twenty-four 3 4-1 fractions in which the defining contrast is some form of the four-factor interaction. We can take as the points of the
fraction the set of points on any of the hyperplans defined by
xl + a 2 x2 + a3 x3 + a4 x4 = b (mod 3),
where a, = 1 or 2 and b = 0, 1, or 2.
There is usually no particular reason for choosing any one of these fractions
over the others. Vance elected to use the fraction defined by
X1 + X2 +
X3 + X4 = 0.
3"- k Fractions with Interactions
Sec. 9.3]
183
The experimental points and the reported responses are

0000 4.2
0102 13.1
0201
9.5
1002 7.7
1101 11.0
1200 8.5
2001 11.4
2100 13.5
2202 31.0
0012
0111
0210
1011
1110
1212
2010
2112
2211
5.9
16.4
22.2
16.5
29.0
37.4
21.1
51.6
74.5
0021
0120
0222
1020
1122
1221
2022
2121
2220
8.2
30.7
31.0
14.3
55.0
66.3
57.9
76.5
85.1.
We can now carry out the analysis of variance procedure for the basic 3 3
factorial in the usual way and obtain sums of squares for the main effects
A, B, C, and the interactions. These sums of squares are given in Table 9.1. The
sum of squares for D is calculated just as the sum of squares for the other main
effects, by adding the squares of the totals at the several levels of x 4 , dividing by
nine, and subtracting the correction for the mean.
The two-factor interactions are hopelessly jumbled. If we work in terms of the
symbolic components of interaction we can untangle the web to a certain
extent, but it is hardly worthwhile. Consider for example the A x D interaction,
which has four degrees of freedom. The AD interaction involves contrasts
between sums of points with x l + x 4 - const (mod 3). Since we also have the
defining relationship x l + x 2 + x 3 + x 4 = 0, it follows that for these points
(x 1 + x 2 + x 3 + x 4 ) + 2(x 1 + x4 ) - const, i.e., x 2 + x 3 = const, and so the
AD contrasts are also BC contrasts. Similarly, we have x l + 2x 2 + 2x 3 + x4
const, and so the AD contrasts are also AB 2 C 2 D contrasts.
An equivalent approach is to argue that the defining contrasts are I, ABCD,
A 2 B 2 C 2 D 2 . Multiplying by AD and letting A 3 = B 3 = C 3 = D 3 = I gives an
alias set AD, A 2 BCD 2 , B 2 C 2 . Finally, squaring the last two members to make
the exponent of the first letter unity gives the set AD, AB 2 C 2 D, BC. We can
proceed in this way to obtain the aliases of each component (2 d.f.) in the 3 3
experiment. If we ignore interactions involving three or more factors, the alias
sets are
A; B; C; AB, CD; AB 2 ; AC, BD; AC 2 ; BC, AD;

BC 2 ; (ABC), D; (AB 2 C), BD 2 ; (ABC 2 ), CD 2 ; (AB 2 C 2 ), AD 2 .
The crucial difference between this and the 2 4 -1 fractional factorial defined by
I = ABCD is that in the Jatter the A x D interaction is completely aliased, or
confounded, with the B x C interaction. In the present case two of the degrees of
freedom of the A x D interaction are aliased with BC, and the other two
degrees of freedom come from the A x B x C interaction sum of squares. A
[Ch. 9
more meaningful aliasing would have been one where, for example, lin A x lin B
was aliased with quad C x quad D, but we cannot obtain such relationships
using this method of obtaining fractions.
TABLE 9.1
Source
Total
A
(lin A
(quad A
B
(lin B
(quad B
C
(lin C
(quad C
D
(lin D
(quad D
(Total for main effects
AxB
AxC
BxC
A x B x C (excluding D)
SS
d.f.
26
2
1)
1)
2
1)
1)
2
1)
1)
2
1)
1)
8)
4
4
4
6
15544.5
4496.3
4399.2
97.1
2768.7
2647.5
121.2
5519.8
5516.0
3.8
283.4
213.6
69.8
13068.2
310.8
1232.9
669.8
262.8
9.4 Orthogonal Main Effects Plans

We have already discussed saturated 2' -' and 3n k fractions, in which the
basic design is "loaded" with additional factors by equating them to interactions. This procedure has given us orthogonal main effects plans for 2n 1
two-level factors in 2n runs and for (3n 1)/2 three-level factors in 3n runs.
Plackett and Burman (1946) discovered a series of orthogonal saturated resolution III designs for N 1 factors in N runs when N = 4k where k is an integer.
Their designs were mentioned in passing in Chapter 8, and a discussion of them
has been postponed until now. They listed designs for all values of N = 4k up to
N = 100, with the exception of N = 92. The case N = 92 was solved by
Baumert, Golomb, and Hall (1962). The other major contribution to orthogonal
main effects plans to be discussed at this time is due to Addelman (1962), who
showed how new plans could be obtained by collapsing and replacing factors in
the basic series of designs by other factors which have different numbers of
levels.
-
The Plackett and Burman Designs 185
Sec. 9.5]
9.5 The Plackett and Burman Designs

These are designs for factors at two levels which we denote by + and ,
standing for x i = + 1 and x t = 1, respectively. Let X denote the design
matrix. If there are N points and n factors, X is a matrix with N rows and n
columns in which x h{ is the value (+ or ) of x i at the hth point. We have the
main effects model
E(y) _ Po +
Psxnil
or E(Y) = D13 where D = (1, X) and 13' = Po, P1, P2, ..., Pn.
We wish to find a design such that the matrix D'D is diagonal. Since
xk i = 1 for all h and all i, the diagonal elements of D'D are each equal to N,
and so we require D'D = NI, i.e., that the column vectors of X be orthogonal
to each other and to 1.
We denote by X { the ith column of X, consisting of the values taken by x ti at
each design point. If 1'X i = c { , Margolin (1967) defines X i to be a column
vector of value c i . If c i = 0, X { is called a zero-sum column vector. lf c i 0, X i is
called a nonzero-sum column vector. If D'D is to be diagonal, each X, must
be a zero-sum column vector, which means that half the entries must be plus
and half minus. Similarly, if we let (X i , X; ) denote the inner product X,A j , we
need to have (X i , X; ) = 0 for all pairs i, j(i ^ j). If n = N 1, which is the
largest number of factors at two levels that we can hope to accommodate, the
problem of finding a suitable design becomes one of finding a matrix D of N
rows and N columns with the following properties:
(i) Each element is 1;
(ii) One column of D is 1;
(iii) N -112 D is orthogonal. Such a matrix is called a Hademard matrix.
We shall now show how Plackett and Burman obtained their designs. The
reader who does not wish to follow the derivation, which involves the use of
finite fields, can safely skip to the end of this section, where we list the design
for N = 12. This method of constructing a Hademard matrix is due to Paley
(1933).
Let z be any nonzero element of the finite field GF(pm); z is said to be a

quadratic residue of the field if there is an element w in the field such that
w 2 = z; otherwise, z is a nonquadratic residue. If p >_ 3, it can be shown that
the number of nonzero elements that are quadratic residues is equal to the
number that are nonquadratic residues. We introduce the Legendre symbol, x(z),
defined as follows (it will be used in a slightly different fashion in Chapter 13):
x(0 ) = 0 ;
x(z) = + 1,
if z is a quadratic residue;
x(z) = 1,
if z is a nonquadratic residue.
[Ch. 9
It follows that 1 x(z) = 0, summing over all z in GF(pm). Three other results
about the Legendre symbols are needed, and we state them without proof.
(i) Let z' be another element of GF(pm). Then X(z)X() = x(z ), in particular
x( z) = x(- 1 )x(z);
(ii) x( -1) = +1, ifpm = 4k + 1, x( -1) _ 1, ifpm = 4k 1 where k is an
integer;
(iii) Let u and v be two elements of the field u 0 v, p > 2. Then, summing over
all zin the field x(z u)x(z v) = 1.
-
Using these results, we can construct a Hademard matrix of order N where

N = pm + 1 = 4k.
We number the rows and columns of a matrix X, h, and 1 where h and i are
elements of GF(pm), and add one extra row, which has every element minus;
X is a matrix of pm + 1 rows and pm columns. The elements of the first pm rows
are assigned as xhh = + 1, x hi = X(h i) (h 0 i).
Consider the matrix D = (1, X). For this matrix we have
1'X { =-1 + 1 1X(hi)=0,
h
(X1, Xi)
= 1 x(i j) X(j i) + x(h i)x(h j)

h
= 1 x(i j)x(Ji) 1.
But with pm = 4k 1 we have x(j i) = x(-1)x(i j) = x(i j), so that
(X i , X1) = 0. The matrix D is thus a Hademard matrix.
We also note that if D is any Hademard matrix of order N,
(D D
D* _ D D)
is a Hademard matrix of order 2N, and so a design for 2N 1 factors in 2N
runs can be constructed from a design for N 1 factors in N runs.
If pm = 4k + 1, Plackett and Burman show that the foregoing construction
procedure can be modified to obtain a Hademard matrix of order 2(pm + 1),
and the reader is referred to their paper for the details. They used this method to
obtain designs for N = 52, 76, and 100. Their design for N = 36 was obtained
by trial.
It should also be noted that if N = p + 1, i.e., m = 1, then, apart from the
last row which represents the base point (1), X is a circulant matrix. We have
= x,, and so in their listing the authors give only one line of each
matrix for those values of N.
As an example we now derive the Plackett and Burman design for N = 12,
p = 11. The elements of GF(1 1) are 0, 1, 2,..., 10 with arithmetic being carried
out mod 11. The quadratic residues are 1 = 1 2 = 10 2 , 4 = 2 2 = 9 2 , 9 = 3 2 =
Sec. 9.6]
Addelman's Designs 187
8 2 , 5 = 4 2 = 7 2 , 3 = 5 2 = 6 2 . Then X(z) = + 1 if z = 1, 3, 4, 5, 9 and X(z) = 1

if z = 2, 6, 7, 8, 10. The design matrix X follows. This is a better design for
eleven factors than the 3(2 11-9 ) because each main effect is estimated from all
twelve points with variance efficiency one. If there are only ten factors, we can
use the extra factor for blocking and obtain two blocks of six. We cannot,
however, obtain more than two blocks because in the design there are no
factors for which, writing a, P, and y for subscripts, x, = x h$ xhy for all h.
The Plackett and Burman Design for N = 12
A
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
abdefk
bcefgl
acdfgh
bdeghj
cefhjk
dfgjkl
aeghkl
abfhjl
abcgjk
bcdhkl
acdejl
(1)
9.6 Addelman's Designs

During the discussion of the two-way layout in Chapter 4 the condition of
proportional frequencies was derived. Suppose that in a two-way layout with
factors A and B there are n 11 observations at the ith level of A and the jth level
of B, and that Ij n t1 = n { ., l i ntj = n. j . We denote the sum of all the observations
at the ith level of A by Ti ., and the sum of the observations at thejth level of B
by T. ; . It was shown that a necessary and sufficient condition for the two
factors to be orthogonal, i.e., for any comparison between the averages of the
observations at two levels of A, e.g., n ; .'Ti .. n 1 .., to be orthogonal to any B
comparison, such as n.-J- 1 T 1 - n.. i 'T 1 , is that the cell frequencies be proportional
to the marginal frequencies. This condition is that
-
n if = n { .n. f /N
for all i and all j; N = I, n i . = I f n. ; is the total number of observations.
Addelman (1962) makes use of this property to develop new orthogonal main
effects plans [see allo Addelman and Kempthorne (1961)].
In an earlier section we presented a 34_2 orthogonal main effects plan: four
factors in nine points. Suppose that we replace one of the three-level factors, say
D, by a two-level factor, E. We may let E take its low level whenever x 4 = 0 or 2,
[Ch. 9
and its high level whenever x 4 = 1. Then the condition of proportional frequencies will be satisfied and we shall have an orthogonal main effects fraction
fora 2 x 3 3 design. The correspondence between the two designs is
0000
0121
0212
1011
1102
1220
2022
2110
2201
-->
0000
0121
0210
1011
1100
1220
2020
2110
2201.
The important thing is that two levels of D are assigned to one level of E, and
the third level of D is assigned to the other level of E. The condition of proportional frequencies will be satisfied whether the two levels of D go to high E or to
low E, or whichever level of D is chosen to be the odd one. The latter choice
does, however, have a bearing upon the efficiency in larger designs where the
new factor has more than two levels. We could replace two, three, or all four of
the three-level factors by two-level factors. Changing all four factors to two
levels would not have much merit in practice because with three points at one
level of a factor and six points at the other level, we have V(21) = a 2 (1/3 + 1/6)
= a 2 /2. This is the same variance we obtain with a 2 -1 , and that design is of
resolution IV and has one fewer point. The other three designs, 2'3 3 / /9, 2 2 3 2 / /9,
and 2 3 3/ /9, are all useful in practical applications.
Suppose now that we go a step further and in a 4 4 - 2 Graeco-Latin square
design replace one of the four-level factors, D, by a three-level factor, E. There
are now two possibilities of interest, both of which give us proportional frequencies and hence orthogonal main effects plans. We make the correspondence
x 4 = 3 ^ x o = 0, x 4 = 1 or 2xx 5 = 1, x 4 = 2 ^ x 5 = 2. This is a symmetric assignment. There will be four points each at x 5 = 0 and x 5 = 2, and
eight at x 5 = 1. To obtain a measure of relative efficiency in this choice, we can
compare the variances of the linear and quadratic contrasts with the variances
that would have been obtained had we been able to take one third of the points
at each level of E. The relative efficiencies are then linear 3/4 and quadratic 18/16.
Alternatively, we could assign two levels of D to x 5 = 0, and one each to x 5 = 1
and x 5 = 2. In that case the relative efficiency of the linear contrast would be
one, but the relative efficiency of the quadratic contrast would be 18/22. The
difference is hardly worth getting excited about. We might sum it up by saying
that if we are more interested in the linear contrast we should take more observations at the end points. If we want more emphasis on the quadratic term we
should take more observations at the middle value of x 5 . In either case this idea of
going from four levels to three is a useful one. This procedure of replacing a factor
at s levels by another factor at t levels (t < s) is called collapsing the factor.
Sec. 9.6]
Addelman's Designs
189
Collapsing is not confined to replacing a factor by one other factor with fewer
levels. We have already mentioned while discussing confounding that a fourlevel factor can be considered as a pair of two-level factors. If, however, we are
interested only in main effects, we can go further and replace a single four-level
factor by three two-level factors. This is achieved by the correspondence
x = 0 -+ 000, x = 1 -+ 011, x = 2 -+ 101, x = 3 -+ 110. Indeed, if a factor has
s levels, where s = t'", t being a prime, or a power of a prime, it may be replaced
by (s - 1)/(t - 1) factors each at t levels. The 45 - 3 hyper-Graeco-Latin square
design can be converted in this way into 4n2l5-3n/ /16, n = 0, 1,2,3,4.
Another technique that Addelman uses is the addition of fractions to increase
the number of levels of a factor. Suppose, for example, that we wish to construct
a 6 x 33 design. We may take a 34 - 2 fraction and run it twice, letting Xl take
the values 0, 1, and 2 in the first replicate and 3, 4, and 5 in the second replicate.
If a 5 x 33 were required, we could then collapse the six-level factor, and have
the desired design.
Addelman uses as basic designs from which to build his plans several of the
standard saturated fractional factorials 27-\ 34 - 2, 215- 11 , 56 - \ 313-1,231-27.
In addition, he uses the following interesting basic plan for 37 / /18. The derivation of this design is given by Addelman and Kempthorne (1961). The eighteen
points fall into two nine-point groups:
00000oo
0112111
0221222
1011120
1120201
1202012
2022102
2101210
2210021
0021011
0100122
0212200
1002221
1111002
1220110
2010212
2122020
2201101.
In the first half the defining relationships are

X5
X3
= xi +
= Xl +
X3
Xl
+ X 2,
X6 = xi + Xl + X 2, X7 = xi + 2Xl + X2'
+ 2, X4 = Xl + 2X2 + 1, X5 = 2xi + X 2,
= 2xi + Xl + X2 + 1.
X2,
X2
X4
Xl
+ 2X2'
In the second half

X6
2xi
+ 2Xl +
I, X7
With these basic plans the tools are now available to generate a large number
of orthogonal main effects plans, especially in the 2m3 n series.
The Plackett and Burman designs do not lend themselves to modification,
either by collapsing or by replacing factors. The only choice they allow is to
leave out one or more factors, and doing that does not improve the precision of
the estimates of the main effects of the factors that remain. With the 2n - k
fractions it is also possible under suitable circumstances to reverse the replacement procedure by exchanging three two-level factors for a four-level factor.
This requires, however, that the triple of coordinates of the three factors replaced
takes only the set 111, 100, 010, 00I, or else the set 011, 101, 110, 000, giving
C = AB.
X2
[Ch. 9
Consider changing the 2' - 4 design, (1), adef, befg, abdg, cdeg, acfg, bcdf, abce
into a 3 x 2 orthogonal main effects plan. The generators of the defining
contrast subgroups are 1, ACD, ABCE, ABF, BCG. At each point
x l + x 3 + x4 - 0 (mod 2). We can therefore replace A, C, and D by a fourlevel factor, letting 000 * 0, 011 -> 1, 101 2, 110 3. We then introduc the
three-level factor by collapsing the four-level factor that we have just acquired;
we collapse levels 2 and 3 to x 8 = 2. The final design follows, together with the
original eight points:
BEFGH
(1)
00000
adef
befg
abdg
cdeg
01102
01011
acfg
00112
11110
10012
bcdf
10101
abce
11002.
9.7 Fractions of 2m3n Designs of Resolution V

A listing of fractions of 2m3 designs was prepared by Connor and Young
(1961) in the National Bureau of Standards series. More recently Margolin
(1967) has investigated such fractions. In this section we shall give a brief
sketch of Margolin's results. We shall use A to denote a factor with two levels
and R, S, T, U to denote factors with three levels. When we talk of a resolution
V design in this context we shall mean a design in which all such interactions as
ARL , AR Q , R L SL , R Q SL , R L,SQ , R Q SQ are estimable. Interactions with three or
more factors will be suppressed.
In a complete 3m factorial the linear contrast for R is given by {R = 2}
{R = 0}, i.e., by the difference between the totals at the highest level of R and
the lowest level of R. The RL SL contrast is {R = 2, S = 2} + {R = 0, S = 0}
{R = 0, S = 2} {R = 2, S = 0}. (This is not the kind of contrast between
hyperplanes that we considered earlier.) In a 3n p fraction of resolution IV the
main effects will be clear, but there will be some linear dependencies between the
interaction terms. As an example, Margolin points out that if we take the 34-1
fraction defined by x, + x, + xt + x = 1 (mod 3), there is a linear dependency
between R L SL , RQ SQ , TL
and T0 U0 . Suppose now that we collapse the factor
U into a two-level factor A by letting x u = 0, x,, = 2 x,, = 0, and x,, = 1 -^
x a = 1. Then the linear U contrast and the interactions involving UL disappear.
Contrasts involving levels of A correspond to U Q . The disappearance of UL
removes the dependency between the other three interactions. The ATe interaction, for example, corresponds to the former T0U0.
7
Sec. 9.8]
Two Nonorthogonal Fractions of the 4 3 Design 191
Using this idea Margolin obtains several fractions of 2m3n designs by collapsing factors in resolution IV 3n P designs. The example that we have just mentioned is a design of resolution V for 2 1 3 3 / /27. Collapsing a second factor would
give 2 2 3 2 / /27. He also gives a 2 2 3 4 / /81 fraction.
A second method of obtaining fractions is to run different 3n - p fractions at
each level of a two-level factor. An obvious example of a 2 1 3 5 / /2 1 3 4 would be to
run a resolution V 3 5 -1 fraction at low A and another 35_1 (or the same one) at
high A. Similarly, a 2 5 3'/ /96 fraction may be formed by running one of the
resolution V 2 5 -1 fractions at each level of the three-level factor. Margolin gives
two designs formed in this way. A 2 4 3'/ /24 design for four factors A, B, C, and
D at two levels each and E at three levels is obtained by running the half replicate
defined by I = ABCD at x 5 = 0 and x 5 = 1 and running the half replicate
defined by I = + ABCD at x 5 = 2.
-
9.8 Two Nonorthogonal Fractions of the 4 3 Design

We mentioned earlier that a four-level factor may be replaced by a pair of
two-level factors. The 4 3 factorial with factors P, Q, and R can thus be considered as a 2 6 factorial where A and B correspond to P, C and D to Q, and E
and F to R. The three degrees of freedom for the main effect of P correspond to
A, B, and AB. The nine degrees of freedom for PQ correspond to AC, AD,
ACD, BC, BD, BCD, ABC, ABD, and ABCD.
The first fraction that we shall consider is a main effects only plan with twelve
points. This will be a 3(2 6 - 4 ) fraction with all six main effects and the interactions AB, CD, and EF estimable. All other interactions are suppressed.
Consider the family of 2 4 - 2 fractions of the factorial in A, B, C, and D defined
by I = AD = BC = ABCD and add new factors by putting E = + ACD,
F = +ABC. The four quarters are
(1), adf, bce, abcdef;
cef, acde, bf, abd;
de, aef, bcd, abcf;
cdf, ac, bdef, abe.
Any three of them constitute a 3(26_4) fraction for which the alias sets are
I, AD, BC, ABCD; C, ACD(=E), B, ABD; D, A, BCD, ABC(=F); CD, AC,
BD(=EF), AB. This is a main effects only plan. Furthermore, because the
defining contrasts AD, BC, and ABCD are all suppressed, the design may be
run in three blocks of four runs each, the blocks being the three "quarters"
chosen.
This design is actually a three-quarter fraction of the Latin square. Suppose
that we let columns correspond to the levels (1), a, b, ab of P and rows to the
levels (1), c, d, cd of Q, and that we denote the levels (1), e, f, and ef of R by
0, 1, 2, and 3, respectively. Furthermore, let us denote points in the four
[Ch. 9
"quarters" by a, P, y, 8, respectively. When this is done we have a Graeco-Latin

square:
(1) aef
cef ac
de adf
cdf acde
bf
bce
bdef
bed
abe
abcf
abd
abcdef
Oa
3fl
ly
28
3y 2fl 18
08 la 2y
2a 38 0p
1 fl Oy 3a.
Thus our design actually amounts to using the fourth factor (Greek letters) as
a blocking factor and taking any three of the four blocks.
The second design is a resolution V fraction with forty-eight points. This
design is the 3(2 8-2 ) fraction defined by I = ACE = BDF = ABCDEF. The
alias sets follow. The effects that correspond to degrees of freedom in the
P x Q x R interaction, and are therefore suppressed, have been underlined.
This design may be run in three blocks of sixteen runs each. An example is given
by John (1970).
I, ACE, BDF, ABCDEF ;
A, CE, ABDF, BCDEF ;
B,ABCE, DF, ACDEF ;

C,AE, BCDF, ABDEF;
AB, BCE, ADF, CDEF;
BC, ABE, CDF, ADEF;
ABC, BE, ACDF, DEF;
D, ACDE, BF, ABCEF ;
AD, CDE, ABF, B CEF ;
BD, ABCDE, F, ACEF;
ABD, BCDE, AF, CEF;
CD, ADE, BCF, ABEF;
ACD, DE, ABCF, BEF;
BCD, ARDE, CF, AEF;
ABCD, BDE, ACF, EF.
AC, E, ABCDF, BDEF;
CHAPTER
10
Response
Surfaces
Most of the experimental designs that have been considered so far have had
their origins in agricultural experimentation. The development of 2n k fractional
factorials has taken place in industrial experimentation, but they were introduced
in an agricultural context by Finney (1945). The concept of response surfaces and
designs for their exploration began in the chemical industry. The early work was
done by statisticians and chemical engineers in the Imperial Chemicals Industries
in Great Britain. The first major paper in the field was published by Box and
Wilson (1951). Its title, "On the Experimental Attainment of Optimum Conditions," is indicative of its philosophy.
Suppose for simplicity that we wish to investigate the production of lome
chemical product in a pilot plant, and that there are two factors, temperature
and pressure, that are to be considered. The experimental procedure will be to
make several runs on the pilot plant at different temperatures and pressures and
to observe on each run a response, y, such as the yield of the product, or the
purity of the product, or the amount of raw material converted, and so on.
We have a two-dimensional factor space with two quantitative factors. We can
conceive, therefore, of the expected response E(y) as a function of the levels of
the factors. If x l denotes the temperature and x 2 the pressure we have
-
Y = f(xj, x2) + e,
f
with the usual assumptions about the random error e. The function f(x 1 , x 2 ) is
single valued and, writing 77 for E(y), the surface represented by 17 = f(x 1 , x 2 ) is
called the response surface. If the x l and x 2 axes are taken in the usual way
and the y axis is taken perpendicular to the plan of the paper, the response
193
194 Response Surfaces
[Ch. 10
surface can be represented on the paper, just as altitude is represented on a map,

by drawing the lines of equal response as contours.
It may be that prior knowledge of the chemical process involved will enable
us to know beforehand the form of f(x 1 , x 2 ). The discussion of such models is
outside the scope of this book. We shall consider the situation in which
f(x 1 , x 2 is approximated by a polynomial of degree d, in which case the model is
called a dth-order model. The set of points (x 1 , x 2 ) at which runs can be made is
called the factor space.
One possibility would be for us to take several levels of temperature and
several levels of pressure to straddle the factor space, carry out a complete, or
fractional, factorial experiment, and fit the model to the data as we have done in
earlier chapters. This ignores three important features of this kind of experiment.
1. The usual object of the experiment is not to investigate f(x 1 , x 2 ) over the
whole factor space but to locate the region in the factor space where the response
is at its highest and to map it. If we regard the response surface as a mountain,
the object is to find the peak and to explore the area in which the peak is found.
The industrialist is rarely interested in regions where the yield of his process will
be relatively low.
2. A polynomial model will usually be quite unrealistic over the whole space,
but its use in a relatively small region around the peak is acceptable as a Taylor
series expansion approximating a more complicated function.
3. Experimental runs on pilot plants, or actual commercial reactors, usually
have relatively high precision and can be conducted sequentially.
Our problem can thus be summarized: How do we find the region of the
factor space that interests us and, having found it, how do we design experiments
to map f(x 1 , x2 ) over that region? We shall discuss the second question first.
The remainder of the chapter can be divided into two areas: the design and
analysis of experiments for fitting first- and second-order models, and the
method of steepest ascent and EVOP.
The two most important papers on the topic of response surfaces are the one
by Box and Wilson (1951) mentioned before, and the paper by Box and J. S.
Hunter (1957). Box and a group of workers at Imperial Chemical Industries
prepared a book, Design and Analysis of Industrial Experiments, edited by
Davies (1954). The chapter on the determination of optimum conditions
contains several examples and is an excellent practical exposition. The reader
who wishes to pursue this subject further will find numerous references in the review paper on response surface methodology by Hill and W. G. Hunter (1966).
)
10.1 First-Order Designs

These are designs for fitting a first-degree equation:
y=Po +1Pix, +ei,
i= 1,2,...,k.
First-Order Designs 195
Sec. 10.1]
A simple example of such a design is the main effects only 2' - k fractional
factorial where the coordinates of x j take the values 1. Before considering
optimal designs it should be recalled that the coefficients S, are affected by the
choice of the scale of x { . If x 1 denotes the length of some physical component
measured in feet, we can triple the value of J by measuring x, in units of yards
instead. It is desirable, therefore, to scale the factors uniformly, and we shall do
this by imposing the, following restrictions upon the coordinates x, where x j' is
the value of x j at the jth experimental point:
x1 =
0 ,
x41 = b,
where b is a constant. The first of these requirements is a shift in the origin so that
x, = 0 for all i; the second amounts to a change of scale to provide uniform

scatter among the x,.
Let X denote the design matrix (x f,). The normal equations may be written
()
X Y = S,
where
S = (1, X)'(1, X)= [0
and F3' = ( Po, P1, ..., Pk). In comparing designs we shall take as a criterion the
average variance of the regression coefficients, or
v = ( k + 1 ) - 1 {V($o) + > V(i )} = (k + 1) -1 tr(S -1 ).

t
For an optimal design we seek, therefore, to minimize

k
tr(S
) _
h =0
Ah 1 ,
where X h are the latent roots of S and A o = N. The trace is minimized subject to
h = kb, in which case,
the condition l A h = tr S = N + kb or
Ah =b
for all h where h = 1, 2, ..., k. This requires that S be a diagonal matrix, or,
equivalently, that the columns of X be mutually orthogonal. This condition for
X is both necessary and sufficient for a design to be optimal and for such a
design we have v($t) = b - laa, i > 0.
Some examples of optimal designs are the 2n ' fractions and the Plackett and
Burman designs for n factors. For k = 3 we can use the vertices of the regular
-
[Ch. 10
tetrahedron with the center at the origin and any orientation. Also, for k = 2
the vertices of the regular polygon inscribed in a circle with the center at the
origin form an optimal design.
10.2 Second-Order Designs

These are designs for fitting the second-degree surface (for k = 3, the general
quadric surface) :
,8x5 +
E(Y1) = R 0 +
t
Nhixihxlt
h<t
It is convenient to write the elements of E3 in the order

p , = (PO, P11, P22, , Pkk, P19 P2, . 3 Pk, P12, .. )
Writing the normai equations as TY = Sp, we have

(TY)' _
x1Y^ ...91 x1Y, ..., > x1x2Y,....
Y^
We shall confine our investigation to designs that satisfy conditions of

symmetry and scaling, which are similar to those for first-order designs. The
following conditions, with h i, are imposed upon the set of points in the
design:
xph =
i
xih =
1
xlhxlt x'hX =
x hXji = 0,
4
x1i
= c + d.
22
^ xlnxlt = c,
The matrix S now takes a convenient form. In its top left-hand corner is the
principal minor, U, with (k + 1) rows and columns which will now be discussed. The remaining rows and columns consist of zero elements everywhere
except on the main diagonal. The diagonal element corresponding to P i is b,
so that P = x1iy1/b and V(Pt ) = a 2 /b. Similarly, /hi = 1 xlhxliyi/c and
v(h2) = a2/C.
The submatrix U may be written, with a = N, as
=[
a blk
bi k cJk + dlkJ
The inverse, when it exists, is a similarly patterned matrix,

_ p
U 1
qlk
[1 k SJ k + tIk]
Taking the product UU -1 = Ik+ 1 gives a set of five consistent equations in the
four unknowns, p, q, s, and t.
Sec. 10.2]
Second-Order Designs 197

1
(10.1)
aq + b(ks + t) = 0
(10.2)
bp + q(kc + d) = 0
(10.3)
bq+ds+ct+kcs=0
(10.4)
dt
(10.5)
ap
kbq
1.
The last equation gives t = 1/d. Eliminating t from Equations 10.2 and 10.4
gives s = (b 2 ac)q/bd, and solving Equations 10.1 and 10.3 gives bp =
(kc + d)q and q = b/[a(kc + d) kb 2 ].
If we now write [d{a(kc + d) kb 2 }] -1 = A, these solutions become
p = d(kc + d)A, q = bdA, s = (b 2 ac)A, and t = 1/d. If kb 2 = a(kc + d)
the design is singular.
It would be convenient if we could choose the design points in such a way that
cov (Po, Ahh) = 0 for all h, and cov (,hh, Pit) = 0 for all h, i, h i. The first
condition cannot be satisfied, for we have
COV (Po, Phh)
= qa2 = bdA
and b > 0, d> 0, A 0 0. Hence, q 0 0 and the covariance cannot be made to

vanish. On the other hand, cov (P hh , P; ,) = sa 2 , and a necessary and sufficient
condition for this to vanish is that b 2 ac = 0. It is customary to call the
property cov (Phh , Pt; ) = 0 the orthogonality property for response surface
designs, and b 2 = ac, the orthogonality condition.
Box and Hunter (1957) write
xfh = Na 2 ; ^; xjh xjt = NA 4 . The orthogonality condition then becomes a = d 4 .
Another interesting property for a response surface design is rotatability
[Box and Hunter (1957)]. A design is defined to be rotatable if its variance contours are concentric circles about the origin. Equivalently, if x o = (x01, x02, ,
X Ok ) denotes a point in the factor space and if y o is the estimated response at x 0 ,
the design is said to be rotatable if, and only if, V(y) is a function only of the
distante p = (x 1 x 2 + Xo k )li 2 of x o from the origin.
Box and Hunter (1957) derive the necessary and sufficient conditions for a
design of any order to be rotatable. We shall confine ourselves to considering
second-order designs with the restrictions that were listed at the beginning of
this section. Then
2 V (90) = V(0)
+
V(, h)xoh + 1 V (,hh)xh +

h
V (Pht)xhx{
h<{
+ 2 1 COV (N0, &h)xOh + 2

h
h t
COV (Nhh, BI{)XOhX01.

h 4
h<i
This is to be a function of p 2 . The first term, V($0 ), is a constant. The seconddegree terms are p 2{V(jh ) + 2 cov (Po, $hh)}. The fourth-degree terms must
[Ch. 10
reduce to a constant multiple of p 4 so that the coefficient of x h has to be half the

coefficient of x h xo i , or
2 V(, hh) = V(h) + 2 COV (, hh, nu).
In terms of the elements of the matrix U 1 , this gives
2(s+t)=c - '+ 2s,
or
2ct = 1.
But t = 1/d, and it follows that a necessary and sufficient condition for
rotatability is d = 2c, or since d = J J xh4 ^J xhJx,1,
xh1 = 3 1 xhj j.
J
For a rotatable design, A -1 = 2c{a(k + 2)c kb 2 } and a -2 V(j 0 ) = p +

(b - ' + 2q)p 2 + (s + t)p 4 . For a design that is both orthogonal and rotatable,
b 2 = ac and A -1 = 4ac 2 , so that in the notation of Box and Hunter p =

(k + 2)/(2N), q = 1/(2NA 2 ), s = 0, and t = 1/(2N\).
10.3 31c Designs and Some Fractions

In this and the following sections we shall consider a few examples of designs
for fitting second-order models. The 3k factorials are obvious candidates for
consideration because each factor is at three levels and the model can certainly
be fitted. For the complete factorial we have, when x h takes the levels 1, 0, + 1,
a=3 k ,
b = 2.3k -1 ,
4.3k-2, d
c=
= 2.3k -2 .
We see that b 2 ac = 0 and c = 2d. It follows that the complete 3k factorials

are orthogonal designs for k factors, but they are not rotatable. We have
A - ' = 4.3 3 k -4 so that
p-
2k3 1 ^
q= 3k1 1, s=0,
t=2.3k
-a'
and the variance of an estimated response at the point x o is given by
= (2k + 1) 2 xh
-I-
xh + 4
xnxi
h<f
The smallest of these factorials is the 3 2 design which we considered in Chapter

4. For this design the adjusted variance NVa -2 is given by
NVa - 2 = 5 9(x041 + xa)/2 + 9(4 i + x042)/ 2 + 9x1xz/4 .
At the origin we have Vo -2 = 5/9. At ( 1, 0) and (0, 1), Vo -2 = 5/9 again

and at the corner points ( 1, 1), Va -2 = 7.25/9. There is a minimum of
Sec. 10.3]
3k Designs and Some Fractions 199
3.2/9 at ( 0.4, 0.4). Box and Hunter (1957) give diagrams for the variance
contours for three designs for k = 2. The first of these is the 3 2 design, in which
they have scaled the factors so that x1 = 4 = 9, i.e., the levels of Xh are
\/(3/2), 0, +-,/(3/2). The third design is the 3 2 factorial rotated through 45 0
and the second is one of the polygonal designs that will be considered in the next
section.
For the 3 3 design we shall write v = NVa -2 . At the origin we have v = 7;
at the center points of the faces of the cube, v = 7; at the midpoints of the
edges, v = 9.25; and at the corners, v = 13.75. The minimum value of v occurs
when 1 = 4 = 4 = 1/3 and is 4.75. For any radius p the minimum variance
occurs on the main diagonals of the cube.
We are again faced with the fact that there are twenty-seven points in the 3 3
design and only ten coefficients in the second-order equation. Debaun (1959)
presents lome fractions of the 3 3 design. These are quite different from the
fractions based upon hyperplanes that we considered in the previous chapter;
those fractions are of little help to us in the present situation. Other designs
based on fractions of 3'1 factorials are given by Box and Behnken (1960).
Debaun considers five subsets of the complete factorial, each of which is
symmetrical about the origin. They are
,
(i)
(ii)
(iii)
(iv)
The complete factorial;

The cube with points ( 1, 1, 1);
The center point (0, 0, 0);
The octahedron with points ( 1, 0, 0), (0, 1, 0), and (0, 0, 1) at the
centers of the faces of the cube;
(v) The cuboctahedron, whose twelve points are the midpoints of the edges
of the cube given by ( 1, 1,0),(1,0, 1), and (0, 1, 1).
He goes on to consider and compare designs made by combining these fractions.

The designs that he considers are
(i)
(ii)
(iii)
(iv)
(v)
(vi)
The complete factorial;

Cube plus octahedron plus n center points;
Cube plus 2 octahedra plus n center points;
Cuboctahedron plus n center points;
Cube plus cuboctahedron plus n center points;
Cuboctahedron plus octahedron plus n center points.
In comparing the designs he uses as his criterion the information per point,
which is v -1 . He shows that the complete factorial is a relatively inefficient design
by this criterion and recommends designs (iii) and (iv). We shall now consider
design (iv) and give an example.
The Cuboctahedron
For the cuboctahedron with n center points we have N = a + 12 + n,
b=8,c=4,d= 4,andA-'=64n.
[Ch. 10
If there are no center points the design is singular. If, however, n > 0, we
have
1
_ 4 n
1
_1
S=

p n'
q __ 2n'
t 4'
16n '
and the design is orthogonal if, and only if, n = 4.
In that case we have
v=4
27xh +4 >xh +4(xlx2
+4
3+44).
Schneider and Stockett (1963) give an example of a cuboctahedron with three

center points. The experiment was designed to determine the effects of column
temperature (z 1 ), gas/liquid ratio (z2), and packing height (z 3) in reducing the
unpleasant odor of a chemical product that was being sold for household use.
The coded factors are
-80 z2 -0.5 z3 -4
x3 = 2
= 0 '
x2 _ 0.2 '
xl 4
1
The second-order model

y = NO + N1x1 + N2x2 + P3x3 + Nll4 + P224 + jS33x32
+ P12x1x2 + P13x1x3 + P23x2x3
was fitted to the data. Afterwards the authors replaced x 2 by x 4 , which is the
ratio of liquid rate to packing height, and obtained a somewhat better fit
(although it was still far from ideal). The three center points provide an unbiased
estimate of error with two degrees of freedom. The data and the analysis of
variance table follow.
y
66
39
43
49
58
17
5
40
65
7
43
22
31
35
26
x1
X2
X3
1
+1
1
+1
1
+1
1
+1
0
0
0
0
0
0
0
1
1
+1
+1
0
0
0
0
1
+1
1
+1
0
0
0
0
0
0
0
1
1
+1
+1
1
1
+1
+1
0
0
0
X4
11.75
11.75
3.75
3.75
11.00
11.00
3.67
3.67
23.50
7.50
7.80
2.50
5.50
5.50
5.50
Polygonal Designs for k = 2 201
Sec. 10.4]
TABLE 10.1
Source
SS
d.f.
Raw Total
Mean
Corrected total
First-order terms
Second-order terms
Lack of fit
Error
24,874.0
129.1
24,744.9
17,940.2
5,253.3
1,510.7
40.7
15
MS
5,980
876
504
20
295
44
25
1
14
3
6
3
2
10.4 Polygonal Designs for k = 2

A useful first-order design for two factors consists of the vertices of a regular
polygon with n sides inscribed in the unit circle. Each point in the design may be
expressed by x l = cos 6, x 2 = sin 0, where 0 is the angle between the radius
vector and the x l axis. Taking the first point to be x l = 1, x 2 = 0, the other
points correspond to 0 = 2jir/n, j = 1, 2, ... , n 1;x 1 + ix 2 = cos 6 + i sin 0
are the roots of the equation xn 1 = 0 with (1, 0) corresponding to j = n.
The sum of the roots of x' 1 = 0 is zero. Hence, 1?n 1 cos 0 + i sin 0 = 0
so that 1 cos 0 = 1 sin 6 = 0 and x fl = 1 xj2 = 0. Furthermore, the sum of
the squares of the roots is also zero. Hence,
(cos 0 + i sin 9) 2
cos 2 8 sine 0 + 2i
1 sin 0 cos 0 = 0.
Also, 7, cos 2 0 + 1 sine 0 = n. We thus have
xil = 4 = cos 2 0 =
1x
jl f2
1 sine 6 = n/2;
= 1 sin 0 cos 6 = 0.
The first-order design is orthogonal and satisfies the conditions that we imposed
on first-order designs in the previous sections. It is easily shown that the design
retains these properties if it is rotated through any angle a, i.e., if xfl = cos
(a + 2jwln) and x 52 = sin (a + 2jrr/n). A larger design may be made up by using
several polygons and letting the hth polygon have n h vertices and radius Ph; in
particular, we may add n o center points at the origin.
When we turn to second-order designs we note that for the single polygon
4 + xiz - 1 at each point, and the design is singular. Suppose that we add n o
center points and let N = n + n o . It is necessary that n > 5.
[Ch. 10
The sum of the cubes of the roots of xn = 1 is zero. Hence,
1 (cos
0 3 sin' 0 cos 6) + i 1. (3 sin 0 cos 2 0 cos 3 0) = 0.
But (cos 3 0 3 sine 0 cos B) = 4 cos 3 0 3 cos 0, and we have
1 cos
1 sin
cos 2 6 sin
6=
= 1 cos 0 sine 0 = 0.
The sum of the fourth powers also vanishes, giving cos 3 0 sin 0 = cos 0
sin3 0 = 0, and also 1(cos 4 6 + sin 4 6 6 cos 2 6 sine 6) = 0. But
(cos 2 6 + sin 2 6) 2 =
(cos 4 6 + sin4 6 + 2 cos 2 6 sine 6) = n;
hence,
1 sine 6 cos
6 = n/8.
Furthermore,
cos 4 0 =
cos 2 6(1 sine 6) = n/2 n/8 = 3n/8 =
sin 4 6.
In the notation of the previous section, we now have

a = N = n + n o ,
b = n/2,
c = n/8,
d = n/4.
It is not surprising that the rotatability condition d = 2c is satisfied. The

orthogonality condition b 2 = ac leads to
n 2 /4 = (n + n o )n/8,
or
n = no,
which calls for more center points than many experimenters might regard as
practicable. The elements of the covariance matrix are
p=n5
q=no 1 ,
s=
noln -1 ,
=4n -1 .
The variance of the estimate of the response at a point a distance p from the
origin is given by
o_2 V(9) = n' + (2n -1
2n0')p 2 + (n 1 + 3n 1 )p 4 .
If the design is orthogonal, this becomes

a -2 V()
= n -1 (1 + 4p4),
and the minimum variance occurs at the origin.
Polygonal Designs for k = 2 203
Sec. 10.4]
It is interesting to consider the case in which the variance is the same at the
center of the circle (p = 0) and on the circumference (p = 1). For this to occur
we must have
(2n -1 - 2no') + (no' + 3n -1) = 0,
or n = 5n 0 , a particular example of which is given by the pentagon with a single
center point. The minimum variance now occurs at p 2 = 1/2, and is given by
3a 2 /n; the variance at p = 0 and p = 1 is 5o 2 /n.
EXAMPLE. The following example is given by J. S. Hunter (1960). There are
two factors, reaction time and temperature, for a batch process, and the levels
are given by
time -90 temp - 205
x l = 10 , x 2 = 10
This experiment is actually part of a larger experiment and we shall refer to it
again later. The design was an octagon with four center points. The fitted
second-order equation was
y = 87.3750 - 1.3838x 1 + 0.3625x 2 - 2.14384 - 3.09384 - 4.8750x 1 x 2 .
(Hunter retained four decimal places in the calculated coefficients for use later in
reducing this equation to canonical form.) The data and the analysis of variance
table follow.
Y
Xl
X2
88.0
86.8
78.8
84.5
91.2
77.4
0
0
-1
+1
-1
+1
0
0
-1
-1
+1
+1
xi
X2
89.7
85.0
81.2
83.3
79.5
81.2
0
0
/2
-\/2
0
0
0
0
0
0
/2
-\/2
TABLE 10.2
Source
SS
d.f.
Raw total
Mean
Corrected total
First-order terms
Second-order terras
Lack of fit
84,653.24
84,436.96
216.28
16.37
171.96
16.18
11.77
27.95
12
Error
Combined residual
MS
1
11
2
3
3
3
6
8.18
57.32
5.39
3.92
4.66
[Ch. 10
10.5 Centra! Composite Designs

These second-order designs for k factors are composed of three sets of points.
(i) A 2' factorial design, or a suitable fraction, with x { = 1. These are
called the cube points, and there are n, of them. This portion of the
design is sometimes called a measure polytope;
(ii) A set of star or axial points (cross polytope). There are n a = 2k such
points, two on each axis at a distance a from the origin. Their coordinates
are ( a, 0, 0, ... ), (0, a,
(iii) n o center points.

For such a design we have
a=N=n c +n a +n o ,
b=n,+2a 2 ,
c=n c ,
d=2a 4 .
The rotatability condition becomes
a 4 = n,.
n, + 2a4 = 3n^,
For orthogonality N n, = (n, + 2a 2 ) 2 , or,
n o +n a =
4 a 2 (n + a 2
nc
For both orthogonality and rotatability, a 2 = -s/(n), and so n o = 4('./(n,) + 1)

2k. If -,/(n,) is an integer, we can obtain both orthogonality and rotatability
by proper choice of n o .
If k = 2 we have n c = 4, a = -,/2. The cube points and the star points form a
regular octagon with vertices on a circle of radius -,/2. From the results of the
previous section we recall that eight center points are needed for orthogonality.
If k = 4, we have a = 2, n o = 12.
If k = 5, we may take n c = 16 (the half replicate). Then n o = 10, a = 2.
If -,/(n c ) is irrational, we must sacrifice either rotatability or orthogonality. One
possibility is to choose the points so as to attain orthogonality while coming as
close as we can to rotatability. For k = 3, rotatability requires a 4 = 8. Then, for
orthogonality as well, we should need n o + 6 = 15.5. We should, therefore,
take n o = 9 and choose a to give orthogonality, i.e., so that (8 + 6 + 9) 8 =
(8 + 2a 2 ) 2 or a = 1.662.
It will already be apparent to the reader that in these designs, just as in the
polygon of the previous section, the price of orthogonality with rotatability is
taking several center points. Some experimenters will not be happy about
devoting such a large portion of their experimental effort to duplicating a
single point.
It is not necessary that the cube portion of the design consist of a complete 2'
Central Composite Designs 205
Sec. 10.5]
design; the use of the half replicate has already been mentioned. In seeking to
reduce the number of points in a composite design, we may note that the
response at a center point estimates Po ; the difference between the average of the
responses at the two star points on the ith axis and the average at the center
points gives an estimate of fl, and the difference between the responses at the
two axial points gives an estimate of 2afl,. Thus the center points and the star
points give estimates of all but the interaction terms fl,, and the Jatter have to be
estimated from the cube points. We may therefore use a fraction, if there is one
that will allow all two-factor interactions to be estimated, assuming the main
effects to be known. If we are prepared to suppress all interactions with three or
more factors, we may thus use any 2k ' fraction that contains no more than one
2 f.i. in any alias set. (If there is a single 2 f.i. among the defining contrasts, that
will contradict our simplifying assumption that 1 Xjh Xji = 0 for all pairs and
complicate the calculations, but the parameters will still be estimable.) Hartley
(1959) suggests using, when k = 6, the quarter replicate defined by I = ABC =
DEF = ABCDEF. Consider, for example, the AB contrast. If we suppress
higher order interactions, its expectation is 16(fl 12 + fl), and we already have an
estimate of f 3 from the star points.
Aia, et al. (1961) used a central composite design with k = 3, a = 1.6818, and
six center points in an experiment on the precipitation of stoichiometric dihydrate of calcium hydrogen orthophosphate (CaHPO 4 .2H 2 O). Among the
responses that they considered was the percentage yield. The three factors used
were the mole ratio of NH 3 to CaC1 2 in the calcium chloride solution, the
addition time in minutes of the NH 3 -CaCl 2 mix, and the starting pH of the
NH 4 H 2 PO 4 solution used.
The scaled factors were
-
x l = [(NH 3 ) /( CaC1 2 ) 0.85]/0.09,
x 2 = (t 50)/24,
x 3 = ( pH 3.5)/0.9.
The fitted second-order equation was
y = 76.03 + 5.49x 1 0.71x 2 + 10.18x 3 0.88x 1 x 2 1.46x 1 x 3 0.29x 2 x 3
+0.694 + 0.432 7.18x3.
They later decided to drop the x 2 terms from the model. The new fitted equation
was then
= 76.39 + 5.49x 1 + 10.18x 3 1.46x 1 x 3 + 0.644 7.22x3.
The changes in the quadratic coefficients and the constant term when x 2 is
omitted are small. There is no change in the coefficients of x l and x 3 because
these terms are orthogonal to the x 2 term and the other terms.
[Ch. 10
The data and the analysis of variance table follow.
X1
-1
1
-1
1
-1
1
-1
1
-1.6818
+1.6818
X2
X3
X1
-1
-1
1
1
-1
-1
1
1
0
0
-1
-1
-1
-1
1
1
1
1
0
0
52.8
67.9
55.4
64.2
75.1
81.6
73.8
79.5
68.1
91.2
0
0
0
0
0
0
0
0
0
0
X2
X3
-1.6818
0
80.6
+1.6818
0
77.5
0
-1.6818 36.8
0
+ 1.6818 78.0
0
0
74.6
0
0
75.9
0
0
76.9
0
0
72.3
0
0
75.9
0
0
79.8
TABLE 10.3
Source
SS
d.f.
Total
First-order terms
Second-order terms
Lack of fit
Error
770.96
610.93
135.15
18.71
6.17
19
3
6
5
5
10.6 Orthogonal Blocking

We now discuss breaking up the response surface design into m blocks;
the wth block contains n w points. The block effects are additive constants, and
there is no interaction between blocks and factors. We may modify our model
accordingly and write
E(Y5) = PO + Po.(ziw - Zw) +

w=1
Nhxih
h
h
h
+1
PhiXn XJI
Nhhxjh,
t
h
where P o + P O is the expected response at the origin in the wth block; S ow is the
wth block effect, and we impose the side condition Lw flownw = 0; zfw is a dummy
variable which takes the value + 1 if the jth point is in the wth block, and is
zero otherwise. We denote by zw the average n,,/N.
It is desirable that the block effects be orthogonal to the parameters of the
response surface. This is achieved if, and only if, the corresponding design
vectors are perpendicular, i.e., if for each w
(Zlw - zW)xlh =
(10.6)
Orthogonal Blocking 207
Sec. 10.6]
for each h,
(z5, ZW)xlhxfi =
for each pair h, i, h
(10.7)
i,
(Z!w - i.)XJh
(10.8)
= 0
for each h. Since we are confining our attention to designs for which
! xh = I! xlhxlt = 0 and I 4,, = b, the first two conditions become
!
0,
(10.6)'
= 0.
(10.7)'
Zlwxlh =
!
Zlwxlhxlt
The third condition is, for all the points in the wth block,
xh = n,,b/N.
The set of points in each block must thus comprise an orthogonal first-order
design with 1 xih proportional to the block size.
An example for k = 2 is provided by the regular hexagon with an even number
of center points. This design may be split into two blocks of size N/2 by dividing
the hexagon into two equilateral triangles, and letting each block consist of one
of the triangles together with half the center points.
Thus, with two center points we might have
Block 1
xl
1.000
xl
x2
0.000
0.500
0.866
0.500 0.866
0.0001
0.000
Block 2
1.000
0.500
0.500
0.000
x2
0.000
0.866
0.866
0.000
Indeed, if we take six center points and put three of them in each block, we can
simultaneously obtain orthogonal blocking, orthogonality (of the estimates of
quadratic coefficients), and rotatability.
A central composite design can be divided into two rotatable first-order designs,
namely the cube points and the axial points. Suppose that this is done and that
n oc of the center points are included in the block with the cube points and
n oa = n o n oc are included with the axial points.
For the blocking to be orthogonal we must have
n c = (n c
n oc )b/N, 2a 2 = ( 2k + n oa)b/N.
Eliminating b/N,
(2k + n oa)n c = 2_ 2 (n 0 + noc)
[Ch. 10
A convenient example is the case k = 4 and n, = 16. We have already seen that
for orthogonality and rotatability n o = 12 and a 2 = 4. Orthogonal blocking is
then obtained by putting n o , = 8 and n oa = 4.
The cube points may be divided into two half replicates confounding the fourfactor interaction. Each half replicate is an orthogonal first-order design, and so
we can add four center points to each 2 -1 and thereby have a design for n =
16 + 8 + 12 = 36 divided into three orthogonal blocks of twelve points each.
The octagonal design with four center points given as an example in the
section on polygonal designs can be run in two orthogonal blocks. The blocks
correspond to the columns in the listing of the data. If we consider the octagonal
design as a central composite design for k = 2 with a = 4 and n, = 4, the first
block consists of the "cube" points together with two center points. The second
block contains the star points plus the two remaining center points.
10.7 Steepest Ascent and Canonical Equations

So far we have talked about designs for fitting first- and second-order models
without concerning ourselves about the problem of where or when to fit them.
It will be convenient for us to confine ourselves for the moment to two factors,
xl and x 2 i the generalization to k factors will be clear. How do we decide about
the levels of the factors? There are two points to be remembered.
1. We are not really concerned about investigating the response over the
entire sample spare of points (x 1 , x 2 ) at which the plant will operate. We are
interested only in operating in the subset of the sample space in which the
response is optimized. The term optimizing the response is not as simple as it
sounds. In a practical situation there may be confusion over what response is to
be optimized. The engineer wishes to optimize the yield of some product. The
comptroller's department employees in the head office have in mind optimizing
profit, and they may calculate it by a formula unknown to the researchera.
There are, for example, some situations in which it is better to maximize
the total yield than the percentage yield. It might, perhaps, be more desirable to pass 200 gallons of feed stock per hour through a plant converting only
60 per cent, than to pass 100 gallons per hour with 80 per cent conversion. For
convenience we shall think in terms of maximizing conversion or yield, recognizing, however, that in some cases other responses may be of greater interest.
If we are interested in running the plant in the area of maximum yield, then
we are interested only in investigating closely the area of the sample space in
which the yields are high. We do not care to devote any more time than is
necessary to exploring regions of no economic interest. We envision a twodimensional factor space, with the yield as the third dimension rising out of the
x 1 , x2 plane. Hopefully, this three-dimensional figure y = f(x l , x2 has a single
peak, and we would like to find it. Having found it we shall operate the plant at
those conditions, and the stockholders will be content. Perhaps the peak is
unattainableit might require so high a temperature that we should burst a
boilerbut we shall do the best we can.
)
The Method of Steepest Ascent 209
Sec. 10.8]
2. Even if we can find a set of optimum conditions for operating the plant,
we may not be able to maintain them. Suppose, for example, that we have a
polymerization process using an acid catalyst. A simplified description of the
process might be that the feed stock is passed through a reactor containing acid
of strength x l at a temperature x 2 , and a fraction y of it is converted.
The first-order model gives an equation
Y = No + Nlxl +
N2x2.
If /3 and P 2 are both positive, the message seems to be clear. We should

increase the acid strength and the temperature as much as we can without
wrecking the equipment. As the plant runs, water and contaminants leak into
the reactor and the overall acid strength drops. All we can do is to continue to
run the plant at as high a temperature as possible until it is decided that the acid
strength has fallen far enough. Then the old acid is removed, the reactor is
charged with new acid, and we start again.
Consider, however, a process such as the polymerization of propylene. The
monomer is fed into the reactor. The dimer is unstable, and the product, or
effluent, consists of unconverted monomer together with a mixture of trimer,
tetramer, and pentamer (i.e., polymers in which the molecules consist of three,
four, or five molecules of propylene, respectively). Suppose that we want
to optimize the yield of tetramer. For any given acid strength we can vary the
operating temperature. If the temperature is too low we get mostly trimer and
little of the heavier polymers. On the other hand, if the reactor is too hot the
polymerization is so active that we have a lot of pentamer. The ideal, therefore,
for maximizing the yield of tetramer is to find an operating temperature that is
neither too hot nor too cold. For any given acid strength, an approximation of
the response curve y = f(x 2 ) by a parabola with a maximum at some intermediate temperature seems appropriate. Similarly, for a fixed operating temperature we find that when the acid is too dilute there is too much trimer, and when
the acid is too strong there is too much pentamer. Again a quadratic approximation for y = f(x 1 is appropriate. The second-order model y = f(x 1 x 2 ) is
justified, at least in the immediate area of the optimal value of y, or the peak of
the mountain. It is not enough, however, for us to know where the peak is.
As the acid grows weaker, we shall have to adjust the temperature upward to
compensate for the loss in strength. We should like to avoid overcompensation
or undercompensation.
In short, our objective should be to find the peak and map the area around it
so that we can find a path that will minimize loss, i.e., a path to follow by
adjusting x 2 so that as x l decreases we shall lose height as slowly as possible.
)
10.8 The Method of Steepest Ascent

In seeking the neighborhood of the peak we begin by fitting a first-order
design in the region where the plant is currently operating. On the basis of this
[Ch. 10
experiment we estimate the direction of steepest ascent. We march in this

direction until we find that we are no longer climbing. At that stage we perform
another first-order experiment and perhaps arrive at a new direction in which to
resume our climb toward the peak. On the other hand, we may find from this
latest first-order experiment that we are already either near the peak or near a
saddle point, in which case we augment the first-order design by further points to
make it into a second-order design.
We illustrate this by referring back to the experiment that led to Hunter's
octagonal design, which we mentioned earlier. There were two factors, reaction
time and reaction temperature.
Hunter began his experiment at a reaction time of sixty minutes and a reaction
temperature of 160 . He elected to make changes of 10 minutes in the time and
10 in the temperature, and ran a duplicated 2 2 factorial with a duplicated
center point,
_ time 60
xl 10
x2
_ temp 160
10
The data were
xi
+1
+1
1
1
0
x2
+1
1
+1
1
0
68.0, 66.8
60.3,61.7
62.2, 64.6
54.3, 56.5
60.3, 62.3
The fitted first-order model was
5'
=61.7 +2.4x 1 +3.6x 2 .
The contours of y are a series of parallel lines. In general we have a series of

parallel hyperplanes y = /3 + 1, Pi x . The direction of steepest ascent, i.e.,
the direction in which we cut across.the contours most rapidly, is parallel to the
normal to the hyperplane. This is along the live through the center of the
first-order model with direction cosines proportional to Pt . In the example this
procedure calls for moving 2.4 units in the x l direction for every 3.6 units in the
x 2 direction. The length of the steps has to be decided by the experimenter on the
basis of his experience.
A second 2 2 factorial with two center points was carried out, centered at a
reaction time of ninety minutes and a temperature of 205 . The coded levels
were
time
90
x1 = 10 ,
temp 205
x2 = 10
Sec. 10.9]
The Canonical Form of the Equations 211
This design is the first column of the octagonal design given earlier. The fitted
model including the x 1 x2 term is
y = 78.0 + 2.0x 1
1.4x 2 5.1x 1 x 2 ,
and we see that the interaction coefficient is considerably larger than either of the
main effect terms.
The design was therefore augmented to the octagon with four center points by
adding the second column of points. In this instance the experimenter did not
divide the design into two blocks. The analysis of variance table and the fitted
equation have already been given.
10.9 The Canonical Form of the Equations

We are going to confine our discussion to second-order equations. Thirdorder models may be appropriate in lome special cases, but the additional
complications that cubic curves bring make a discussion of them infeasible here.
Corresponding to each value of y is the second-degree function f(x l , x2 ) =
const in k variables. When k = 2 we have the conic sections; for k = 3 we have
the quadric surfaces.
Confining ourselves to the case k = 2, we have for the contour y = y o the
general conic
F(x1, x2) = N11x1 + ,8 12x1x2 + 8 22x2 + fl 1x1 + P2x2 + y = 0,
where y = P o Yo
For various values of y this equation gives a family of concentric conics.
Their center is given by the solution to the equations
aF/8x 1 = 0,
aF/8x 2 = 0
or 211x1 + P12x2 + Nl = 0,12x1 + 2P22x2 + N2 = 0

In the simplest case the contours form a family of concentric ellipses with the
center being the peak of the mountain. A conic is an ellipse if the discriminant
4 11 f22 P2 is positive, and a hyperbola if the discriminant is negative.
Box and Youle (1955) give an example in which two factors, both concentrations of reactants, are considered. The fitted second-order equation is
= 78.56 + 0.50x 1 0.21x 2 2.312 2.15x2 + 4.08x 1 x 2 .
The concentrations of the reactants were measured in grams per liter and
c 1 -55
x l = 5 ,
x2 =
c2-35
5
[Ch. 10
where c l is the concentration of reactant A, and c 2 the concentration of reactant

B. Experiments were conducted in the region 1 < x l < +1, 1 < x 2 < + 1.
The discriminant, 4(2.31 x 2.15) 4.08 2 , is positive, and the contours of 9
are a family of concentric ellipses with the center given by
0.50 = 4.624' 4.084,
0.21 = 4.084' + 4.304,
whence
xi = 0.41,
4 = 0.34.
The value of the predicted (maximum) response at the center is 9 = 78.63. To

derive the canonical form of the equation, we find the latent roots of the matrix
of second-order coefficients.
This is the matrix B = (b;j ) where bi; = fl and b i; = fl/2 (i j). In this case
[-2.31

2.04
B
2.04 2.15 ]'
and the roots are A l = 4.27, ,1 2 = 0.19.
The corresponding latent vectors of unit length are
Zl
= 1-0.69 Z, '
0.
Z
=[
and the canonical form of the equation is

y = 78.63 4.27zi 0.194
where z l = 0.72(x 1 xi) 0.69(x 2 4') = 0.72x 1 0.69x 2 0.06 and
z2 = 0.69x 1 + 0.72x 2 0.53.
In the new coordinate system the major axis of the ellipses is the live z l = 0.
We have already seen that the optimum operating conditions are at the center
x l = x = 0.41 and x 2 = 0.34. If, however, we are obliged to move away from
the center because of a change in concentration of one of the two reactants, we
shall lose yield most slowly if we can adjust x l and x 2 so that z l is kept at zero.
This means that if, for example, a change 8x 1 occurs in x 1 , we should adjust for
this by making a change Sx 2 in x 2 , where
0.728x 1 0.698x 2 = 0.
In the particular example this is almost equivalent to 8x 2 = 8x 1 , which suggests
that here the important factor might be the difference in concentrations rather
than the actual concentrations themselves.
Sec. 10.11]
Evolutionary Operation 213
10.10 Other Surfaces

In the example given by Hunter the second-order equation was
9 = 87.3750 1.3838x
+ 0.3625x 2 2.1438x1 3.0938x2 4.8750x 1 x 2 .
This is allo a family of ellipses, but in this case the center is at x l = 3.7369,
x 2 = + 3.0028, which is well outside the area of experimentation. It may even
be outside the range in which the plant can be operated. In the area where the
experiments took place the response surface is a rising ridge, and further experimentation might take place in the direction of a line toward the center.
The canonical form is
y = 90.50 5.10z1 0.14z,
where z l = 0.64x 1 + 0.77x 2 + 0.06 and z 2 = 0.77x 1 + 0.64x 2 4.79. Increases in the value of zi bring about sharp decreases in the yield.
When the contours are hyperbolas, the response surface is a saddle surface
with the center of the family at the center of the saddle. Moving away from the
center along one of the axes will produce an increase in the yield. Moving away
along the other axis will result in a drop in the yield. The experiment by Aia et al.
(1961) gave a saddle surface.
As the discriminant approaches zero, the family of curven approaches a
family of degenerate parabolas, and the contours become a series of pairs of
parallel lines. This is a ridge surface, and the optimum procedure would be to
find the highest of the parallel ridges and to remain on it.
With three factors the geometrical representation becomes more difficult.
There are several worked examples in the literature, particularly in the paper
by Box and Youle (1955) and the book by Davies (1956). Further examples
appear from time to time in chemical and agronomic journals. Examples of
third- and fourth-order models are not found as often because the designs and
the interpretation of the results are, on the whole, too complicated for general
use by engineers. Third- and fourth-order rotatable designs are discuseed by
Gardner, et al. (1959), and in a series of papers by Draper (1960a, b; 1961;
1962).
10.11 Evolutionary Operation

The discussion so far has been concerned with the design of specific experiments, particularly with experiments that might be carried out in a pilot plant by
a research and development group. Once an actual plant is in operation, little
experimentation usually takes place. If the plant can be persuaded to operate
[Ch. 10
under conditions that produce profitable yields, the attention of the operators is
directed toward keeping the plant working at those conditions and maintaining
the status quo. Perhaps it would be possible to improve the yield somewhat, but,
on the other hand, tinkering around with the operating variables might cause a
drop in production. This defensive approach is bolstered by the conviction that
the research personnel will have to make large changes in the operating variables
in order to obtain meaningful results.
A new plant may produce at a lower yield than had been expected because of
problems of scale up. If a plant to produce a thousand barrels a day is to be
designed on the basis of a pilot plant that produces two barrels a day, some
dimensions will have to be increased five hundred fold, some by \/5, and some
by the cube root. When the actual plant is built, it may well happen that the
actual optimum operating conditions are not those that were obtained on the
pilot plant, and the yield will fall short of the projections. Little will be gained by
asking the research staff to generate more data on the pilot plant. If they did a
reasonable job the first time, further data will only serve to confirm their
earlier results. What is needed at this stage is not more information about how to
run the pilot plant but information about optimizing production on the full-size
plant that has already been built.
Evolutionary operation was proposed by Box (1957) as an operating procedure
for nudging a plant toward optimum conditions without causing dramatic
disturbances and catastrophic cutbacks in production. It is designed as a method
of operation that can be carried out by the plant personnel themselves with a
minimum of supervision, or interference, from the research people. Indeed, it is
recommended as a standard operating procedure, whether or not there is concern
about the performance of the plant.
An evolutionary operation program for a plant contains two basic steps:
(i) A routine of systematic small changes is introduced in the levels of the
operating variables being considered;
(ii) The information generated is fed back to the plant foreman or other
supervisor.
Box and Hunter (1959) show how the latter objective is facilitated by using
simplified calculations and emphasizing graphical presentation. They give an
example of an application to a batch process in which the two operating variables
being monitored are temperature and reaction time. We shall use part of their
example. There is nothing in the theory to prevent us from monitoring several
variables but, for practical purposes, two or three variables are enough to
handle under normal circumstances. Box and Hunter also discuss the case in
which three variables are monitored.
The operating cycle is a cycle of five points that form a 2 2 factorial with a center
point. If the plant is currently operating at a temperature of 120 with a reaction
time of sixty minutes, and if fluctuations of 10 in the temperature and 10
Sec. 10.11]
Evolutionary Operation 215
minutes in the time are contemplated, the cycle would consist of the five points
120, 60 min;
110, 50 min;
130, 70 min;
110, 70 min;
130 , 50 min.
0
The cycle is repeated several times. Each time the responses Y1, Y2, Ys, Y4, Y5
are recorded and their averages are posted. The main effects and the interaction
of the two factors are calculated, and when one or more of these three effects
becomes significant, decisions about changes in operating conditions are made
by the plant foreman.
The most important feature in step (ii) is the information board. Table 10.4
TABLE 10.4
YIELDS DURING THE FIRST FOUR CYCLES
Conditions
Cycle
1
2
3
4
63.7
62.1
59.6
63.5
62.8
65.8
62.1
62.8
63.2
65.5
62.0
67.9
67.2
67.6
65.3
62.6
60.5
61.3
64.1
61.7
TABLE 10.5
INFORMATION BOARD-CYCLE 4
Requirement
Per cent yield

Other Responses
Maximize
61.9
64.6
T,
E 62.2
M
P 63.4
-TIME-
Error Limits for Averages

Effects with
95% error
Limits
65.7
2.3
time +2.5 2.3

temperature -1.3 2.3
txT
0.2 2.3
1.4 2.0
Change in mean
Standard deviation
Prior estimate of o
2.3
1.8
[Ch. 10
contains the data given by Box and Hunter for the yields during the first four
cycles of their process. Table 10.5 is the format of an information board at the
end of the fourth cycle. The information board is easily kept up to date, and the
data are clearly and simply displayed on it.
The calculation procedures recommended by Box and Hunter make use of
the range to provide an estimate of the standard deviation, a. We discuss this in
the next section.
10.12 The Method of Calculation

Table 10.6 shows a calculation sheet at the end of the fourth cycle.
Similar sheets for the first three cycles are included in the paper by Box
and Hunter. The calculation of the averages and of the estimates of the
effects is straightforward. The calculation of o is based upon the use of the
range. This is easier to compute than the usual quadratic estimate, and for
the range of live random variables the efficiency relative to the quadratic
estimate is 0.955.
Let yt , 1 denote the value of y i at thejth cycle, and let y,,, denote the average of
the values of y i for the first ] cycles. The entries in row (iv) of the table for the nth
cycle are the observed values of y j , - with variance a 2 {1 + (n - 1) -1 } =
nat /(n - 1). Their range, R, is independent of any blocking that may have
occurred inadvertently between cycles. The estimate of the standard deviation is
obtained by multiplying the range by a factor, c k , which depends upon the
number of observations included. In the example the range is 4.1 - (-4.3) =
8.4, and it is the range of five observations. The corresponding factor c, may be
found in the standard tables such as the Biometrika tables of Pearson and
Hartley (1954). For k = 5 we have c 5 = 0.430; for a 2 design with one center
point we need c 9 = 0.337, and with two center points, c lo = 0.325; c 4 = 0.486.
Then c k R estimates a{n/(n - 1) } 112 , and so the desired unbiased estimate of a is
as
= RCk, \n 1 ) = R x fk , _.
Values of fk , fl for 2 < n < 10 and k = 4, 5, 9, 10 follow.
n=
k
=4
5
9
10
10
0.34
0.30
0.24
0.23
0.40
0.35
0.27
0.26
0.42
0.37
0.29
0.28
0.43
0.38
0.30
0.29
0.44
0.39
0.31
0.30
0.45
0.40
0.31
0.30
0.45
0.40
0.31
0.30
0.46
0.40
0.32
0.31
0.46
0.41
0.32
0.31
--.l
Operating Conditions
Previous cycle average

New observations
Differences (ii) less (iii)
New sums
New averages YI
A (time) effect
= t(Ya +
B (temperature) effect = t(Ya +
A x B
= t(Y2 +
Change in mean
= t(Y2 +
(ii)
(iii)
(iv)
(v)
(vi)
(i) Previous cycle sum
2
190.7
63.6
62.8
0.8
253.5
63.4
3
190.7
63.6
67.9
-4.3
258.6
64.6
Calculation of Effects
+2.5
.Y4 - Y2 - Ys)
-1.3
.Ys - Y2 - Y4)
+0.2
Ya - Y4 - Ys)
.Ya + Y4 + Ys - 4Yl) = +1.4
1
185.4
61.8
63.5
-1.7
248.9
62.2
Calculation of Averages
4
200.1
66.7
62.6
4.1
262.7
65.7
5
185.9
62.0
61.7
0.3
247.6
61.9
CALCULATION SHEET-END OF FOUR CYCLES
TABLE 10.6
3.66
1.83
3.10
8.4
= 6.76
= 2.25
=
=
=
=
Calculation of Limits
For new average
2a/ V n = 2.25
For new effects
2a/ V n = 2.25
For change in mean 1.78a/Vn = 2.00
Previous sum s
Previous average s
New s = range x Ikon
Range
New sum s
New average s = (sum)/(n - 1)
Calculation of a
(Ch. 10

Exercises
1. Show that the polygonal design retains its properties even if rotated through an
angle a, i.e., it is still a rotatable design and n center points are still needed for
orthogonality.
2. Prove that for any rotatable design, satisfying the conditions of the early part of
the chapter, section 10.2,
a -2 V(9o) = p + (b -1 + 2q)p2 + (s + t)p4 .
3. Prove that, in the notation of Box and Hunter, a design is both orthogonal and
rotatable if, and only if,
p = (k + 2)/(2N), q = 1/(2NA2),
t = 1 /(2N)t 2).
4. Prove that for prediction based on a 3 3 factorial design with the vertices of the
cube at ( 1, 1, 1) the variance of a predicted response is a minimum at the
points where 3x2 = 3x2 = 3x3 = 1.
5. Find the variance of an estimated response at the point (x 1 , x2i x 3) for Debaun's
designs (2), (3), (5), and (6).
6. Show that for fitting a first-order model in three dimensions, with the points to
be taken on the surface of a sphere with unit radius, the vertices of a regular
tetrahedron with any orientation is an optimal design for four points.
7. Consider a 3 2 design with k = 2 that has been rotated through 45 so that the
design points are (0, 0), ( V2, 0), (0, v/2), and so on. Analyze the design
and compare the variance contours with those of a 3 2 design in the usual
orientation.
8. A regular icosahedron inscribed in a sphere of radius 3 has twelve vertices given
by (0, a, + b), ( b, 0, a), ( a, b, 0) where a = 1.473 and b = 0.911.
Consider this set of points together with some center points as a second-order
design.
9. (Continuation). Another regular figure is the dodecahedron with twenty
vertices given by (0, c -1 , c), (c, 0, c '), (c -1 , c, 0), (1, 1, 1)
where c = 1.618. Consider this set of points plus some center points as a
second-order design.
10. (Continuation). Show that the dodecahedron with 5n center points can be
divided into five orthogonal blocks of 4 + n points each. Each block is a regular
tetrahedron plus n center points.
0
CHAPTER
11
Incomplete
Block Designs
In Chapter 3 we considered the randomized complete block design. In that

design v varieties or t treatments were tested in b blocks, and each block contained
exactly one observation on each treatment, or one plot with each variety. The
advantage of this design is that because the variance between plots in the same
block is smaller than the variance between plots in different blocks, we obtain an
increase in the precision of comparisons between varieties by being able to
eliminate the block differences. An obvious problem arises when the blocks
available are not large enough to accommodate one plot with each variety. We are
then led to the consideration of incomplete block designs: incomplete inasmuch
as the blocks do not each contain a complete set of varieties.
Suppose that in the gasoline experiment there had been seven gasolines to be
tested in cars but because of a shortage of time each car could be used only for
three gasolines. Clearly, each gasoline cannot be tested in each car. Our interest is
in comparing gasolines, and we might try to allot the gasolines to the cars in such
a way as to have each pair of gasolines appear together in the same number of
cars. One way of doing this is to take 35 = ( 3) cars and to give to each car a
different combination of three gasolines. Each pair of gasolines will appear
together in five of the cars. A smaller design, in which each pair appears together in one car, is given by
ABE, CDE, ACF, BDF, ADG, BCG, EFG,
where the letters denote gasolines and each set of three gasolines is assigned to a
different car. Such designs are called balanced incomplete block designs.
This chapter will begin with a discussion of balanced incomplete block designs.
This will be followed by sections on the general incomplete block design, in
which the blocks may be of different sizes, and where there may be no semblance
219
220 Incomplete Block Designs
[Ch. 11
of balance. We shall also discuss such topics as the recovery of interblock

information. We have a choice of notations. We may talk either of t treatments
or of v varieties. In this chapter we shall follow the more modern usage of talking
about t treatments, although we shall still call the experimental units to which
they are applied, plots. We shall thus have t treatments applied to several plots in
b blocks. In the succeeding chapters we shall discuss partially balanced association schemes and the construction of incomplete block designs. In those chapters
we shall employ the older usage of v varieties because v is the usual notation in
that area.
11.1 Balanced Incomplete Block Designs

These designs were introduced by Yates (1936). They are designs for t treatments in b blocks. Each block contains k plots, k < t. Each plot contains a
single treatment and each treatment appears in r plots. We say that each treatment has r replications, or is replicated r times. No treatment appears more than
once in any block. Each pair of treatments appears in the same block ,1 times.
We shall sometimes use the abbreviations BIB design or BIBD.
The five parameters b, t, r, k, and A are not independent. They are integers
subject to the following restrictions:
(i) N = rt = bk,
where N is the total number of plots;
(ii) A(t 1) = r(k 1);
(iii) b >: t.
A design with b = t and r = k is raid to be symmetric.
To establish restriction (ii) we consider any treatment, such as A. The treatment A appears in r blocks, in each of which are also (k 1) other treatments.
There are, therefore, r(k 1) plots, which are in the same block as a plot
containing A. These plots have to contain the remaining (t 1) treatments
exactly \ times each. Hence, t(t 1) = r(k 1). The four parameter values
b = 12, k = 4, t = 8, and r = 6 satisfy restrictions (i) and (iii). However,
restriction (ii) gives A = 18/7, which is not an integer. Hence, there is no
balanced incomplete block design with those values of b, t, r, and k.
The incidence matrix of a design is the matrix N = ( n,) of t rows and b
columns; nis is the number of times that the ith variety occurs in thejth block.
For a balanced incomplete block design, n is either one or zero, and n,2f = n i; .
The matrix NN' has t rows and t columns. It is of considerable interest in the
development of the theory of incomplete block designs. If we denote the elements
of NN' by q1 , we see that q = I f n2, and q ,, = L n,,n hf , ( i 0 h). For the
balanced incomplete block design we have qi , = r and q, h = a, (i s4 h); we may
therefore write
;
{;
NN' = ( r 2 )' + AJt,
Sec. 11.2]
Some Balanced Incomplete Block Designs 221
where I t is the identity matrix of order t and Jt is a square matrix of order t with
every element unity.
To evaluate the determinant of NN' we subtract the first column from each of
the other columns, and then add all the other rows to the first row. It is now
readily seen that
INN'l =(r)t) t-1 (r+(t l)))=(rA) t-l rk>0.
Thus NN' has rank t, and so N has rank t. However, N is a matrix of t rows and
b columns so that r(N) < min(b, t). Hence, t < min(b, t), and the inequality
(iii) is proved. This inequality is due to Fisher (1940). The proof that has been
presented follows that of Bose (1949).
If the design is symmetric, b = t and N is square. In that case NN'I = ^Nl 2 ,
and so (r A)t - lr 2 is a perfect square. This implies that when t is even, (r A)
must be a square. Consider the set of parameters t = b = 22, r = k = 7, A = 2;
t is even but r A = 5, which is not a square, and so no design with these
parameters exists. This result [Shrikhande (1950)] shows that the conditions
given are necessary, but not sufficient, for the existence of a balanced incomplete
block design.
11.2 Some Balanced Incomplete Block Designs

For any pair of integers t, k(k < t), a balanced incomplete block design can
be obtained by taking all possible combinations of k out of t treatments. For
these designs
b= (k)' r= (k 1)' (k-2)
There are, however, smaller designs for some values of t and k. The design given
at the beginning of the chapter with t = b = 7, r = k = 3, and 2. = 1 is an
example.
A design with t = 6, b = 10, r = 5, k = 3, and A = 2 is
ABE, ABF, ACD, ACF, ADE, BCD, BCE, BDF, CEF, DEF.
A design for t = 9, b = 12, r = 4, k = 3, and A = 1 can be obtained by using
orthogonal Latin squares. (The design is employed in the example in the next
section.) A design with t = 16, b = 20, r = 5, k = 4, and A = 1 may be
constructed in a similar fashion. Both designs are from the orthogonal series of
Yates. Their derivation will be discussed in Chapter 13. The second design
follows. We denote the treatments by 1, 2,..., 16. The treatments are written
[Ch. 11
in five 4 x 4 arrays and each array forms a complete replicate. The rows are the
blocks.
II
I
1 2 3 4
5 6 7 8
9101112
13 14 15 16
1
2
3
4
5 913
6 10 14
7 11 15
8 12 16
III
1 61116
2 5 12 15
3 8 914
4 7 10 13
IV
1
2
3
4
71214
8 11 13
51016
6 9 15
V
1
2
3
4
81015
7 9 16
61213
5 11 14
This design has at least two applications outside the field of experimental
design. In each of the five groups of four blocks every treatment appears exactly
once. The design may, therefore, be used in the following situation to allocate
bridge players to tables. Suppose that sixteen bridge players are gathered together for an evening and that there is enough time for five games to be played at
each table. How do we arrange the players so that each player plays five games;
four tables are in continuous use and no player has to sit out for a time; and
every player has the opportunity to play one game at the same table as every
other player. The solution is to let replications correspond to games and treatments to players. Thus, during the fourth game of the evening the players at the
first table are 1, 7, 12, and 14; at the second table they are 2, 8, 11, and 13, and so
on.
The second application was found in Great Britain before the Second World
War in professional motorcycle racing on cinder tracks, a sport which is still
popular there. The tracks are about 350 yards in circumference. Four riders take
part in each four-lap race, and an evening's program calls for about twenty
races. Usually the evening is devoted to team races between various teams in a
league, but occasionally the riders race as individuals. The following situation
then occurs. There are sixteen riders; they are to be arranged, four at a time, in
such a way that each rider races five times, and, during the evening, he races once
against each of the other fifteen men. This is achieved by using the balanced
incomplete block design and letting treatments represent riders and blocks
represent races. Unfortunately, the program cannot be arranged so that no
rider is obliged to ride in two consecutive races.
A design with t = 8, b = 14, r = 7, k = 4, and \ = 3 is obtained when a 2 3
design is repeated seven times in two blocks of four runs each time, confounding
in turn the three main effects and the four interactions. Writing, 1, 2, 3, 4, 5, 6,
7, 8 for (1), a, b, ab, c, ac, bc, abc, respectively, the design is
1357 1256 1234 1458
2468 3478 5678 2367
1368 1278 1467
2358
2457
3456
Sec. 11.3]
The Analysis of Balanced Incomplete Block Designs 223
The design that we have just given for t = 16 could be derived by adding
d, ad, bd, ... = 9, 10, 11,.... Then the first replicate is obtained by dividing the
2 4 factorial into four blocks, confounding C, D, CD. The effects confounded in
the other replicates are A, B, AB; AC, BD, ABCD; ABC, AD, BCD; BC, ABD,
ACD.
The problems of the existence and construction of balanced incomplete block
designs have been of considerable interest to both statisticians and algebraists.
(They will be the topic of Chapter 13.) The major paper on the construction of
designs is by Bose (1939). The Fisher and Yates tables contain a listing of almost
all the known designs for r < 10 and k < 10. The restrictions on the parameters
have led to the investigation of other types of incomplete block design, notably
the partially balanced incomplete block designs of Bose and Nair (1939); these
designs will be discussed in the next chapter.
11.3 The Analysis of Balanced Incomplete Block Designs

We consider now the usual method of analyzing balanced incomplete block
designs. This is called the intrablock analysis because the block differences are
eliminated and the estimates of all contraets in the treatment effects can be
expressed in terms of comparisons between plots in the same block.
We have the following model:
ytf=+T{
+;3 +e L ,,
where y, f is the observation, if there is one, on the ith treatment in the jth block,
i-, is the effect of the ith treatment, fi, is the jth block effect, and ei, is the random
error with the usual assumptions. The mean p and the treatment effects -rt are
unknown constants. The block effects may be fixed or random; we treat them as
if they are fixed. It is important to note that the model assumes that there is no
interaction between treatments and blocks.
The normal equations are
G=N+r2f i +k,1$i ,
Ti = r + rfj
B1 = k +
+
nijfj + kl^,,
(11.1)
(11.2)
(11.3)
where Ti is the sum of all the observations on the ith treatment and B, is the sum
of all the observations in the jth block.
[Ch. 11
We now eliminate the block effects from Equation 11.2 and obtain
niiBf = rkf i rfj 1l'n tin hJ fh ,
kTT
(11.4)
1h
where the prime denotes summation over all values of h except h = i.

We rewrite this as
kQ t = r(k l)fi ArT h =(r(k-1)+,1) T{
Alf h .
We call Q, the adjusted treatment total for the ith treatment; it is easily seen
that 1, Q, = 0. Imposing the side condition l h fh = 0, and recalling that
r(k 1) = A(t 1), we obtain
kQ j = ^1tfi .
If 0 = 7 c,T, is any contrast in the treatments, we have

^ = k
ctQil(At).
The adjusted treatment totals are not independent. To derive the variance
we note that
V(Q j )
kQt =(k 1)T, (1n i ,Bi T,).

i
The expression 1 nij Bj T, is now the sum of r(k 1) observations on treatments other than the ith treatment. Then
k 2 V(Q { ) = r(k
1) 2 a 2 + r(k
1)a 2 ,
and
V(Q1) = r(k 1)a 2 /k.
It is clear from symmetry that cov (Q,^, Q,) is the same for all pairs h, i (h 0 i).
Then
0 = V(> Q ) = tV(Qi) + t(t 1) cov (Qh, Q,),
{
and so
cov (Qh, Q^) _
r(k 1) o2
k(t 1)
= \a2 /k.
It follows that
V(fh Tt ) = 2ka2/(At).
The Analysis of Balanced Incomplete Block Designs 225
Sec. 11.3]
Indeed, we may write the covariance matrix of the Q 1 as
cov Q = (AtI AJ)a 2 /k.

Then, writing 0 = c'T where r is the vector of treatment effects and c'1 =
1'c = 0, we have
V(1) = ,1Z t2 V(c ' Q) = ^zt2 ' k c
/( ,1t ) .
( ca) ka2
The analysis of variance table is slightly complicated by the fact that treatments are not orthogonal to blocks. We take the sum of squares for regression
G + 1, TT T1 + I, B; p; and then subtract the sums of squares for the mean
G 2 IN (which we write as C), and for blocks (ignoring treatments), 2 Balk C.
The remainder is called the sum of squares for treatments (adjusted for blocks).
We write it as St (aai) . Then
St(u) = T1T4 + B,P, 1 Bi/k + C.

But kB1P1 = Bi kBf Bi nS7 T1 , and so
st(aa1) _
(T1
n11B5/k) Tf =
Q 1T,
= k
Q?/(\t).
TABLE 11.1
SS
Source
d.f.
N 1
1 1 y Jk C
Total
b1
Blocks (ignoring treatments) I BB lk C
Treatments (adjusted for
IQ1 = k Q?/(t) t 1
blocks)
Ntb+1
By subtraction
Residual
The residual sum of squares is the sum of squares for intrablock error. In the F
test for the hypothesis that all the treatment effects are zero, the mean square for
treatments (adj) is tested against the error mean square.
The efficiency of an incomplete block design in which each treatment appears r
times is measured relative to the complete block design with the same number of
replicates. For a complete block design with r observations on each treatment,
V( Th T1 ) = 2a 2 /r. If, for an incomplete block design, the average of the variances of all contrasts Th T1 ) is 2a 2 /a, the efficiency of the design is defined to
be E = a/r.
For the balanced incomplete block designs we have a = At/k, so that E
,1t/(rk). It is clear that for any design E < 1. Some authors write T 1 = Q1 /rE and
(
[Ch. 11
St(BuJ) _ QZ/rE. That sum of squares is also called by some authors the sum of
squares for treatments eliminating blocks.
Under the normality assumption the hypotheses -rh ri = 0 may be tested by
the t test or by Tukey's method; Scheff's S method may also be used for testing
the more complex contrasts.
EXAMPLE. The following data are taken from a dishwashing experiment
[John (1961b)]. The treatments were detergents and the response observed was
the number of plates washed under standard test conditions before the foam
disappeared. In the experiment, detergent solutions are made up and plates
soiled with a standard soil are washed one at a time until they are clean. The
testing procedure calls for three basins to be used (i.e., for three treatments to be
tested simultaneously), and the three operators wash at a common speed during
a test. A block is thus a set of three operators testing three detergents simultaneously. We have t = 9, b = 12, r = 4, k = 3, and A = 1. The data and the
analysis of variance table follow. The letters denote treatments and the responses
are given in parentheses.
1. A (19) B(17)
2. D (6) E (26)
3. G (21) H(19)
4. A (20) D (7)
5. B (17) E (26)
6. C (15) F (23)
7. A (20)
8. B (16)
9. C (13)
10. A (20)
11. B (17)
12. C (14)
C(1l)
F (23)
J (28)
G (20)
H(19)
J (31)
TABLE
E(26)
F (23)
D (7)
F (24)
D (6)
E (24)
J (31)
G (21)
H(20)
H(19)
J (29)
G (21)
11.2
d.f.
Total
Blocks (unadjusted)
Treatments (adjusted)
Residual
35
11
8
16
SS
1512.75
412.75
1086.81
13.19
MS
37.52
135.85
0.82
The adjusted treatment totals (Q, = 3T ) are given by 3Q A = 3, 3Q B = 20,

3Q,=-56, 3Q D =- 116, 3Q E =53, 3Q F =32, 3Q G =15, 3Q H =- 2,
and 3Q,=91.
;
11.4 Resolvable Designs

An incomplete block design in which each treatment appears r times is said to
be resolvable if the blocks can be divided into r groups in such a way that each
group consists of a complete replication of the treatments, i.e., each group of
blocks contains each treatment exactly once. The designs for t = 8 and t = 16,
The General Incomplete Block Design 227
Sec. 11.5]
given earlier, and the design for t = 9, used in the example, are all resolvable
designs. Bose (1942c) showed that the following condition is necessary for a
balanced incomplete block design to be resolvable:
b> t +r-1.
If b = t + r 1, the design is said to be affine resolvable. Bose's condition can
be derived from the weaker assumption that t is a multiple of k. It was subsequently derived from this assumption by Roy (1952), Mikhail (1960), and
Murty (1961). We now give the proof by Murty.
Let t = nk where n is an integer, and b = nr. We have
r _ A _ r) r A
t-1
k-1
tk k(n-1)
Then
r.lr(n-1)rnrbr
t -1
t-1
t-1
But r-1=rk t,1= k(rnA), and so (r a) /k=rnA=(br)/

(t 1). All the parameters are integers, and so r nA is an integer. Furthermore, r > A, and so (r ,1)/k > 0. It follows that (b r) /(t 1) > 1, and so
b > r + t 1.
In the analysis of variance table for a resolvable design, the sum of squares for
blocks (ignoring treatments) may be divided into a sum for replicates and a sum
for blocks in replicates.
11.5 The General Incomplete Block Design

The balanced incomplete block design is a special case of the general incomplete design; in the Jatter the blocks need not be of the same size, nor need
each treatment appear the same number of times. We shall stils consider t
treatments in b blocks. Suppose that the i th treatment is tested in ri plots and that
the jth block contains k plots, (ri , k ; > 0); we do not need to assume that
kf < t for any or all j. As before, Ti denotes the total of all observations on the
ith treatment; B ; is the sum of all the observations in the jth block; T, B are the
corresponding vectors of treatment and block totals.
We define two matrices, R and K; R is a t x t matrix with diagonal elements
r 1 , r2
rt the off-diagonal elements are zero. We write this as R = diag
(r 1 , r 2i ... , r t ). The matrix K is defined as K = diag(k l , k 2
k b ). The incidente
matrix of a design has already been defined as N = ( n i ,) where n ij is the number
of times that the ith treatment appears in the jth block. If n ij = 0 or 1 for all
pairs, the design is said to be binary. If k. = k for all j, the design is called
;
...,
, ... ,
[Ch. 11
proper. If ri = r for all i, the design is called equireplicate. The most commonly
used designs are proper, binary, and equireplicate.
For all designs we have
Rl t = Nl,,,
Kl b = N'lt,
where 1,, denotes a vector of n unit elements. For a proper design, K = kI;
for an equireplicate design, R = rI. Some authors let r and k denote the vectors
of ri and k,; they then write r and k for the matrices that we have called R
and K.
Let yifm denote the mth observation on the ith treatment in the jth block; for a
binary design we omit the subscript m. The model is
Yijm = !L + Ti +
Pi + eijm ,
where , Ti , and ,S are defined in the same way as they were for the balanced
incomplete block design. We denote by T, (3, T, and the corresponding vectors
of effects and their estimates. For the present, Ti and ,8 are assumed to be
unknown constante.
The normal equations are
ri f,+
G=N11+
Ti = ri + r,f, +
Bf =ki+
kj
ni , ft> ,
n ij f,+kfftf ,
11.5)
i = 1, . .., t,
(11.6)
j=1,...,b.
(11.7)
Multiplying each equation of the set of equations (11.7) by the corresponding

fraction n i; /k, for a given value of i and subtracting them all from the ith
equation of the set (11.6) eliminates the block effects and , giving
Qi = Ti
nijBj/k> = rif,
1 (1 niJnh5Iki) Th.
(11.8)
In matrix notation the normal equations are

G
N 1tR 16K
T = Ri t R N
B
%
T
K1b N' K
Multiplying both sides of the equation on the left by
F=
0
[N '
-
It
NK -1
N'R -1 Ib
(11.9)
Sec. 11.5]
The General Incomplete Block Design 229
and recalling that Rl t = Nl b , and Kl b = N'l t , we obtain the reduced equations

G/N = + 1'R' /N + 1'K/N,
Q=TNK -1 B=(RNK B N'R -1 T = (K N'R - 'N).
If we now impose the side conditions
r.i- = 1'RT = 0
and
1 kjpj = 1'Kp = 0,
the first of these equations becomes % = G/N = y.

The second set of equations may be written AT = Q. They are the adjusted
intrablock equations. Some authors write CT = Q. The third set of equations
are the counterparts for estimating P. We note that Al = 0 and (K N'R -1 N)1
= 0, and we now confine our investigations to the case in which A has rank
t 1 and (K N'R - 'N) has rank b 1. Such a design is said to be a connected design and under these conditions all the contrasts T l T t . and P ; P; .
are estimable.
To calculate the sum of squares for regression, we obtain from the equations
K1+N'T+K = B
as
^ =K -1 BK - 'N 'T
1 b .
Then the sum of squares for regression is

G+B' +T'T=G+B'K -1 B+(T'B'K - 'N')TG
= B'K - 'B + Q'T.
The analysis of variance table for the intrablock analysis follows:
TABLE 11.3
Source
SS
1 1 1 y j m G Z IN
Blocks (ignoring treatments) B' K 1 B G 2 /N
Q'R
Residual
By subtraction
Total
d.f.
N1
b 1
t1
Nbt+1
If there are some cells with more than one observation, i.e., n t5 > 1, we may
take out of the residual a sum of squares within cells 11 t ijm (Y iim y ;5.) 2 for
[Ch. 11
the error term and attribute the remainder to interaction between treatments and
blocks. For binary designs we assume that there is no interaction.
11.6 The Adjusted Treatment and Block Totals

The adjusted treatment totals Q. are linearly dependent since Q'l t =
1 Q{ = 0. (Were this not so the set of equations At = Q would be inconsistent.)

We have
Q'1 t = T'1 t B'K - 'N'1 t = G B'K - 'Kl b = GG=0.
Similarly, (B N'R -1 T)'l b = 0.

We now calculate the covariances of the adjusted totals. The basic form of the
normal equations in the format of the general linear hypothesis (Chapter 2) is
X'Y = X'XO with cov Y = Iu 2 and cov (X'Y) = X'(Ia 2 )X = X'Xo 2 . Thus, in
the present case the square matrix on the right in the normal equations 11.9
is the covariance matrix of the totals G, T, B; call it X'X. The covariances of
G/N, Q, (B N'R -1 T) are given by cov (F(G, T', B')') = FX'XFa 2 =
0
N -1 0
R NK - 'N'
0 (K N'R -1 N)K -1 N'
(R NK - 1 N')R - 'N a 2 .
K N'R -1 N
In particular we have cov Q = ( R NK - 'N')a 2 = Ao 2 , and cov (B N'R - 'T)

= (K N'R - 1 N)a 2 .
11.7 Solving the Intrablock Equations

The reduced intrablock equations, Q = AT, may be solved by using either of
the two methods for obtaining solution matrices that were discussed in Chapter
2. The side condition H' = 0 is usually taken with H = 1' or with H = 1'R;
in any case, H'r is not a contrast, and 1'H' = 1 h, 0. For an equireplicate
design these side conditions are, of course, equivalent; for other designs the side
condition H'r = 1'R r = 0 (coupled with 1'Kr3 = 0) gives = y. A discussion of
the two methods of solving the equation follows.
(a) Graybill, Kempthorne.
Let
A H'
B12l
A* [H 0 J [B
B21 B221
Solving the Intrablock Equations 231
Sec. 11.7]
Then A*A* - I = I gives four equations
AB + H'B 21 = I,
(11.10)
ABia + H'B 22 = 0,
(11.11)
HB 11 = 0,
(11.12)
HB 12 = 1
(11.13)
Multiplying Equations 11.10 and 11.11 on the left by 1' and recalling that
1'A = 0, we have
1'H'B 21 = 1'
and
1'H'B22 = 0,
so that B 21 = 1'(l'H') 1 = 1'/(1 h,) and B 22 = 0. If H = 1'R, B 21 = 1'/N; if

H = 1', B 2 = 1'/t.
Multiplying Equation 11.10 on the left by B 11 gives B11AB11 = B11; multiplying on the right by A gives AB 11 A = A. It follows that AB and B 11 A are
idempotent. Taking the transpose of Equation 11.10 and noting that B 11 and A
are symmetric, we have, after multiplying on the right by T,
B 11 AT + B 1 HR = T;
but AR = Q and HT = 0, and so we have obtained a solution vector B11Q = T,
and B 11 is a solution matrix. Furthermore,
2
cov T = cov B11Q = B11AB11a 2 = B11a .
(b) Scheff, Plackett.
Let P l = (A + H'H)
-1
. Then T = P 1 Q is a solution vector, for we have

Q=A4=(A+H'H)T.
It was shown in Chapter 2 that P l = B 11 + J/(1 h1) 2.

Let P be any solution matrix, i.e., let T = PQ be any solution to the equations
Q = AT. Then, if 0 = c',r is a contrast, = c'PQ and V(^) = c'PAP'ca 2 . But
= c'B 11 Q so that c'PQ c'B 11 Q = 0 and
c'(P B11)Q = 0
is an identity for all c, Q subject to c'1 = 0, 1'Q = 0. We note therefore, that
the general form of P is P = B11 + u l l' + 1u where u l and u 2 are arbitrary
vectors. It follows that
V(^) = c'(B 11 + u 1 1' + 1u)A(B l1 + lui + u 2 1')ca 2
= c ' B11AB11ca 2 = c ' Bl1ca 2 = c Pca 2 .
This result implies that if P is any solution matrix we may, in computing the
variance of the estimate + of a contrast 0, act as if P were the covariance matrix
of the estimates R.
[Ch. 11
For the balanced incomplete block design we have

kA = (r(k 1) + A)I t ^J t = AtI, ^Jt .
Using the first method,
A*
1[ AtI AJ kl
kl'
0],
kI
A* -1 =
`t
kJ
1'It
1/t
so that
Bll = at (I` t J `)'
and
T = B 11 Q = kQ/(at).
In using the second method it is easier to take H = \/(A/k)1' so that H'H =
AJ/k and A + H'H = AtI/k. Then (A + H'H) -1 = k1/,1t and t = kQ/(At).
Taking the side condition as H',r = 1'T, we have A + H'H = (AtI + (k ^1)J)lk,
and (A + H'H) -1 = {kI (k ^1)J/t}/(,1t).
Side conditions other than H' = 1 or H' = Rl are sometimes useful. Consider
a design in which a single plot with a control treatment is added to each block of
a balanced incomplete block design. The first design in this chapter for seven
treatments in blocks of three would become a design for eight treatments in
blocks of four. Denoting the control treatment by 0, the new design would be
OABE, OCDE, OACF, OBDF, OADG, OBCG, OEFG.
Let t, b, r, k', and A be the parameters of the balanced design. The whole
design has b blocks of k = k' + 1 plots each, ro = b, ri = r (i > 0). Then,
using the first row and column for the control treatment,
r1
_1 bk'
A
[rl t (r + ,1t)It ^1Jt1
We choose H = (h o , hlt), in order to make A + H'H a diagonal matrix; but
h hh o l ^
H'H = of hh lt h2 Jt 1
and so we wish to have hh o = r/k and h 2 = A/k, or,
h = 1/(/k),
ho = rI1^(ka).
The Recovery of Interblock Information 233
Sec. 11.9]
With this choice of H we have
1 [bk' +r2/A
A +HM
( + At)It ']
whence
-0
kAQ0
kQ,
_ r(r + At)
, T__
i r + At
(l
> O).
11.8 Connected Designs

The idea of connected designs is due to Bose. A block and a treatment are said
to be associated if the treatment appears in the block. Two treatments, A and B,
are said to be connected if we can form a chain of treatments and blocks:
treatmentblocktreatmentblock treatment, beginning with A and
ending with B, such that every block in the chain is associated with both the
treatments adjacent to it. The relationship A connected to B is an equivalence
relation which forms disjoint equivalence classes for the treatments. A design is
said to be connected if there is only one equivalence class, i.e., if every pair of
treatments is connected.
The following design for four blocks I, II, III, IV and seven treatments is
connected:
I ABCD,
II BCE,
III DE,
IV EFG.
A and G are connected by the chain AI B IIEIVG. Denoting the

-r 7 is estimable. We follow the
we see that r 1
effects of A, B,. . . , by T l , r 2i
chain and consider the contrast
...
= Yl l - Y21 + Y22 - Y52 + Y54 - Y74 ;
E(z) = T1 + N1 - T2 - N1 + T2 + N2 - T5 - N2 + T5 + N4 - T7 - N4
= T 1 -
T7-
The following design with t = b = 6, which is the union of two disjoint

designs having no treatment in common, is disconnected:
AB, AC, BC, DE, DF, EF.
11.9 The Recovery of Interblock Information

Yates (1940) pointed out that under certain circumstances a second set of
estimates of the treatment effects, called the interblock estimates, can be obtained.
[Ch. 11
In discussing these estimates we shall confine our consideration to proper

designs and we shall write the variance of the intrablock errors e ;J as Q.
The additional assumption to be made is that the block effects ,J are random
variables distributed with zero means and variance ab; they are uncorrelated,
and are also not correlated with the intrablock errors e ;J . Under these circumstances the block totals can be regarded as a set of observations:
BJ = kp + n ij rt + (kf3 + ^ e15).
We may write the combined error term as fJ . Then E(J) = 0 and af = V(f,) _
k 2 o + ka'. The interblock estimates are obtained by minimizing
(BJ k
where ji and T I denote the interblock estimates; and Tf will be used to denote
the intrablock estimates.
Consider first the balanced incomplete block design. The normal equations
for interblock estimation are
G=N+^rTt ,
n, JB, = rk + rT { + A 7" Th ,
(h # i ).
We shall denote 1 ni5BJ by T. With the side condition 1'r = 0, the first equation
gives = G/N, and the second equation becomes
Ti' = (r a)T{ + rk,

whence
rk
T,_ TirA
In the example of the balanced incomplete block design given earlier,
r=4,k=3,rA=3.ThenTi 1212=T1 G/3=T, 233=3f1 .

The interblock and intrablock estimates of the treatment effects follow.
T,
Ti
234
221
215
194
253
247
234
233
266
0.33
4.00
6.00
13.00
6.67
4.67
0.33
0.00
11.00
0.33
2.22
6.22
12.89
5.89
3.55
1.67
0.22
10.11
It should be noted that 1 Ti = kG = trk so that T, = 0.
Combining Intrablock and Interblock Estimates 235
Sec. 11.10]
In the general case, we minimize

(B k1 6 N'T)'(B k1 6 N'T)
and obtain the normal equations
INB, Rlk NN'] [T]

Multiplying on the left by
1
R1
b
gives the equations
bk
1'R
NB G
b1 0 NN' R
J
R [:1
or
N(B G 1/b) = (NN' RJR/b)T.
If we apply the side condition 1'R'r = 0 the equations become
G/N = and
N(B G1/b) = NN',
so that
T = (NN') - 'N(B G1/b) _ (NN') - 'NB 1(G/bk);
the existence of the interblock estimates requires that NN' exist, and hence that
the incidence matrix N has rank t.
We now have two sets of estimates: T = B 11 Q (or some other function of Q),
and T, which is a function of B and G. Because, for any i and any j, cov (Q ; , B.) =
0, it follows that the two sets of estimates are orthogonal, and, under normality,
independent.
11.10 Combining Intrablock and Interblock Estimates

Consider a contrast 0 = c'T in the treatments. If P is a solution matrix for the
intrablock equations, we have
= c'PQ
and
V(^) = c'Pca.
[Ch. 11
The interblock estimate of is

= c'T = c'(NN') -1 NB;
the term G 1/b drops out because 0 is a contrast and V(^) = c'(NN') - 'ca'.
We now wish to combine these two uncorrelated estimates of 0 to obtain an
unbiased joint estimate with minimum variance. We shall use the following
well-known property of combined estimates.
Let 0 and 0 2 be unbiased estimates of a parameter 0 with variances v l and v 2
and zero covariance. Consider the class of combined estimates of the form
u 1 0 1 + u 2 0 2 . In this class the minimum variance unbiased estimate of 0 is
(w1B1 + w202)/(wi + w2) where the weights are wl = vi l and w2 = v2 1 ; i.e.,

the individual estimates 0 and 0 2 are given weights inversely proportional to
their variances.
In the present case we should take for the combined estimate of + s,
0* = (w + w2G)/(w1 + w2),
where w l = 1 /c'Pco and w 2 = 1 /c'(NN') lcaf. This simplifies to
* _ cV() + cV(^)
V(0) + V(+)
If we knew v and af, we could now obtain &*. The original method of Yates
for the balanced incomplete block design is to use this weighted estimate, but
with the weights computed by substituting for o and a estimates obtained from
the data. Rao (1947a) derived combined estimates under the assumption that the
e 5 and the P are normally distributed. For the balanced incomplete block design
the methods of Yates and of Rao give the same estimates. Rao's derivation will
be given later in this chapter.
We shall now confine our discussion to the case of balanced incomplete block
designs. In that case, P = kI/(At) and V(^) = c'cka/(At). For the interblock
estimates, NN' = (r A)I, + and

(NN , ) _ 1 = I
rd
(rA)(r+(t-1)d)
^1J
I
rA1
(rA)rk'
but c'J = 0, and so we have
ra
_ c'NB _ c i Ti '
r,1
c'ca2
()rA
The minimum variance joint estimate is thus

0* = (w14 + w2)/(w1 + w2),
Combining Intrablock and Interblock Estimates 237
Sec. 11.101
where w l = At/ka and w z = (r A)/o, and

V(b*)
= c c .
w l + w2
As the first step toward an estimate of of we obtain an estimate of ab. In the

intrablock analysis, the sum of squares for regression was separated into three
components; they were the sums of squares for the mean (G 2 /N), blocks
(ignoring treatments), and treatments (adjusted for blocks).
An alternative is to subdivide into sums for the mean, treatments (ignoring
blocks), and blocks (adjusted for treatments).
TABLE 11.4
SS
Source
1 1 y g G 2 /N
Total
Treatments (ignoring blocks) r' 1 T{ G 2 /N
Blocks (adjusted for treatBy subtraction
ments)
From intrablock
Intrablock error
analysis
d.f.
N1
t1
b1
Nbt+1
The sum of squares for blocks (adjusted for treatments) is

2
ilbl adJ
= Ijk B1 + k GAl Q? Jt rT2S .
Scheff's method of obtaining the expectation of this sum of squares is to

consider for convenience the special case in which and -r, are all zero because
they do not appear in the result. Then B; , Qi , and T, all have zero means, so that
the expected values of their squares are equal to their variances, giving
E(S au) =
bV(B,) kV(Qti) _ tV(Ti )

r
k +
a
But,
V(B5) = af = k 2 ab + ka,
V (T=) = r (ab +
V(Q) = r(k 1)o/k,
tee),
and, substituting above,

E(Sbi aaJ) = ab(bk t) + a(b 1).
[Ch. 11
Writing EQ and Eb for the mean squares for intrablock error and for blocks
(adjusted for treatments) respectively, we obtain
Q = Eei &2 =
(En Ee)(b 1)
(N t)
We may now write
+/!* _ +^ +
ww(+ )
i4 (r
+ A t f + )a ^k(r ^) A) __ + ( ktv +
) rk)te
and replace o and ab by $' and Q6. This is essentially what Yates (1940) did.
Graybill and Seshadri (1960) have shown that this estimate is unbiased; Graybill
and Weeks (1959) have shown that it is based on a set of minimal sufficient
statistics.
11.11 Some Computational Formulae for Combined Estimates

Yates (1940) also gave some formulae for calculating the adjusted yields,
Y,; Y, = r(r? + *). We shall make use of the following identity:
rk At= r(tk)/(t 1)=rA.

The intrablock estimate of r, + is
kQ 1 GTi
+T
xt
T` + At + N r (
`
T1
1 1T1 G
rl At + N
T,(rkAt)rT1 +AGTi
Art
+ Wi
At(t 1)
where Wi = (t k)Ti (t 1)T{ + (k 1)G. Similarly,

W{
_ T1 rk _ TiT1(rA)rTi+AGTi
r(rA)
r
r(tk)
Tt+ rA i7_
1
Thus, we can think of W/At(t 1) and Wi /r(t k) as the intrablock and interblock adjustments to the raw treatment means T{ /r. We now combine the two
estimates, using the weights w l = At/(ka) and w2 = (r A)/of to obtain the
estimate (T* + *) = (Ti + 9W)/r = Y,/r, where
B r (At(t 1 1)
r(t w? k) ) (w l + w 2) -1
af ka
ob
= t(k 1)a 2 + k(t k)o = t(k 1)ab + (t 1)o2
(11.14)
Sec. 11.11 ]
Some Computational Formulae for Combined Estimates 239
We may rearrange Equation 11.14 and obtain

of = k(1 + (t k)0)o/(1 t(k 1)9),
and, after some further algebraic manipulation,

V(S*) = c'c(w l + w 2 ) ' = c'c(l + (t k)B)o /r.
-
When Eb is large by comparison to E e we have, taking the limit as Eb 00,

0 = l/{t(k 1)}; then (1 + (t k)B)o/r becomes ko/(at), which gives the
variance of the intrablock estimator.
Substituting &, &b for oe, ob as before, we have
(EbEe)(b-1)
__
B t(k 1)(b 1)(Eb Ee ) + (N t)(t 1)E e
(Eb Ee)(b 1)
t(k 1)(b l)Eb + (t k)(b t)Ee
For the symmetric design with b = t, 0 takes the simpler form
9=Eb EQ /t(k-1)Eb .
If Eb < Ee , it is usual to take 0 = 0. This is equivalent to putting Eb = Ee in the
formula and to acting as if there were no block effects; thus the design is analyzed
as a completely randomized design.
Yates also considered the situation in which the blocks can be broken up into
c groups of bie blocks each, each group consisting of r/c complete replications
of the treatments; r/c is an integer. Although Equation 11.14 still holds, the
estimate of ob is changed. We return to Scheff's method of calculation and add a
term for the sum of squares between groups in the analysis of variance table.
Let G denote the group total for the gth group. Then
Gg k Q, _ T{2 G 2
Bf
Sb' 8a^ k k f N/c + At
r + N
Assuming that
r,, g
are all zero, and also that there is no group effect,
V(G9 ) = bV(Bj )/c
and
V(G) = bV(B,),
where V(B5) = of. Then

E(Sbl ") = k V(Q{) r V(Ti) +
b(b c + 1)
V(B)
= (k(b c + 1) t)ab + (b c)o f

so that
ob = (Eb EQ )(b c)/{N t k(c 1)}.
[Ch. 11
If the design is resolvable, the blocks can be arranged to form single replicates.
In this case, c = r, and the previous result simplifies to
ob = ( Eb E )r/{k(r 1)}.
Q
Substituting in Equation 11.14, we obtain for the general case

__ (Eb Ee)(b c)
B t(k 1)(bc)Eb +(tk)(bct+ 1

and, for the resolvable case,
e rt(k 1)E
(EbEe)r
b
+k(brt+ 1)Ee
A special case is the balanced lattice. This is the balanced design with t = k 2
r = k + 1, and b = k(k + 1). We have already seen examples for t = 9 and
t = 16. In this case, the formula is further simplified to
,
8 = (Eb Ee)Ik 2 (k 1)Eb .

EXAMPLE. James and Bancroft (1951) carried out a 2 3 experiment in a
balanced incomplete block design. They took clover plants, split them into two
halves, and applied a different treatment combination to each half. A block is a
pair of half plants. The response observed was the percentage of hard leed; they
considered using the arcsine transformation on the data, but decided against it.
The treatments in the design were the 2 3 treatment combinations in a factorial
experiment, with factors calcium (C), phosphorus (P), and potassium (K). The
treatments represented by A, B,. . . , H are, respectively, (1), k, p, pk, c, ck, cp,
cpk. The design is resolvable and the data are arranged in seven replications;
the parameters of the design are t = 8, b = 28, r = 7, k = 2. and A = 1.
II
III
A 37 C 27
B 37 H 50
D 90 E 89
F 28 G 71
A 15 D 23
B 47 G 64
C 35 F 39
E 22 H 18
I
A 38
C 49
E 32
G 64
B 29
D 28
F 29
H 32
V
A 23 F 39
B21 D 14
C 18 H 10
E 23 G 53
VI
A 66
B 23
C 22
D 23
IV
A 3
B 45
D 11
F 39
VII
G 68
F 46
E 28
H 39
A 28
B 10
C 32
D 18
H 30
E 40
G 33
F 23
E 13
C 36
G 24
H 37
Sec. 11.11 ]
Some Computational Formulae for Combined Estimates 241
The various totals for each treatment and the estimates of the treatment effects
are
A
T,
Tt
Q,
Y,
W,
Tt
T*
210
439
-9.5
220
118
-2.4
-7.3
-3.0
212
500
-38.0
186
-297
-9.5
2.9
-7.9
219
439
-0.5
234
172
-0.1
-7.3
-1.0
207
467
-26.5
198
96
-6.6
-2.6
-6.2
-
247
472
11.0
257
109
2.8
-1.8
2.2
243
482
2.0
244
15
0.5
-0.1
0.4
G
377
616
69.0
366
-119
17.2
22.2
17.8
H
216
447
-7.5
225
98
-1.9
-6.0
-2.3
The intrablock and interblock analysis of variance table follows:

TABLE 11.5
Source
SS (Intrablock)
SS (Interblock)
d.f.
MS
Total
Replicates
Blocks
Treatments
Intrablock error
19,123.99
4,977.11
10,428.37
1,794.72
1,923.79
19,123.99
4,977.11
8,994.25
3,228.84
1,923.79
55
6
21
7
21
428.30 Eb
91.61 EQ
The weighting factor 0 is given by the formula for the resolvable case:
_ 7(428.30 - 91.61)
56 x 428.30 + 28 x 91.61 = 0.08876.
Then W, = 6Ti - 7T1 + G and Yt = Ti + 0W. The estimates have just been
given; T* = Y1 /7 - where = G/56 = 34.48.
In the factoraal experiment, the treatment contrasts are of primary interest.
When the joint estimates are used, the C contrast is Y5 + Y6 + Y, + Y8
Y, - Y2 - Y3 - Y4 = 254. Dividing by 8r gives the estimate, 4.54, for
the main effect of C. (James and Bancroft called 2 1 the main effect and estimated it by 9.1). The hypotheses /3 = 0 may be tested by the F test under
normality. The numerator of the test statistic is 254 2 /56 = 1152.1; the denominator is (1 + ( t - k)0)EQ = 140.4. In this particular experiment, the only two
significant effects were the main effects of C and K; for the latter, Bk = - 4.0,
with mean square 896.0.
If the intrablock analysis is used, the estimate of the main effect of C is
= k 1 Q 1 /(8At) = 2(149)/64 = 4.66. The corresponding sum of squares is
(149) x k/(8)t) = 693.8. This is tested against Eef and, again, C and K are the
-
P*,
[Ch. 11
only two significant effects. The sums of squares for the several contraets add up
to the sum of squares for treatments, adjusted for blocks. They are
C P K CP CK PK CFK
1152.1 240.3 896.0 73.1 126.0 301.8 283.5
11.12 Other Combined Estimates

In previous sections the estimate of ab was obtained from Eb . This is not the
only method. Yates (1940) subdivided the sum of squares for blocks ignoring
treatments into a treatments component and a remainder. The interblock
estimates, T, were obtained by least squares analysis from the set of b observations Bi . The total sum of squares 1, B1 G/b was then divided into two
components: regression, l { TtT,, and remainder, with t 1 and b t d.f.,
respectively. The remainder has expectation (b t)af. The total is Sbl i g tr
multiplied by k. Hence, dividing each of the component sums of squares by k, we
have a subdivision of Sb l ig tr into a treatment component l i Ti T,/k and a
remainder. The mean square of the Jatter component is an unbiased estimate of
'af/k with (b t) degrees of freedom.
The ten different sums of squares that have been considered during this
discussion of balanced incomplete block designs are collected in table 11.6,
together with their expectations, under the assumptions that treatments are
fixed with the side condition 1, T i = 0 and that blocks are random. Under
normality they have, with the proper divisors, chi-square distributions, and the
F test can be used to test hypotheses, r = 0, and ab = 0.
S1 , S2 , S3 , S6 , and S 9 appear in the intrablock analysis and are distributed
independently under normality. S 4 and 55 are independent of S l , 52, S6 , and Sa
under normality. S1 , S2i S7 , S8 , and 59 appear in the interblock analysis and
are independent under normality.
Graybill and Deal (1959) use as their combined estimator (T 1 + BW)/r, with
0 given by (11.14) and Q and ab estimated by E e and S5 /(b t). By an argument
that depends heavily upon the assumption of normality, they show that when
the flf are normally distributed their estimator is unbiased and is uniformly
better than either or l in the sense that V(i*) < min (V(#), V(^)) for all
values of o and ab when b >_ t+ 10 or when b= t+9 and N b t+
1 >_ 18. Seshadri (1963a,b) also discusses alternate combined estimates of this
type.
11.13 Maximum Likelihood Estimation

We now consider the maximum likelihood method under normality, given by
Rao (1947a). We shall assume in the derivation that we have a general incomplete
block design, the only restriction being that it is proper; it does not have to be
84
85
86
87
88
89
80
81
82
83
Raw total
Mean
Total
Blocks (unadjusted)
divided into
Treatment component
Remainder
Treatments (unadjusted)
Blocks (adjusted)
Intrablock error
Source
81
k- 1 L BJ - 8 1
[L T?/k(r - A)] - [81rk/(r - A)]
83 - 84
kQ'Q/At
r : ! L T? - 8 1
S3 + 8 6 - 8 7
82 - 8 3 - 86 = 82 - 8 7 - 8 6
80
Y'y
G2/N
SS
N-b-I+l
b - 1
1 - 1
b - 1
1 - 1
1 - 1
b - 1
N
1
N - 1
d.f.
TABLE 11.6
(b (t (b (I (I (b (N -
NfL2
NfL2
(N -
+ (r - A) L -rNk + (N - k)ag
+ (r - A) L -rNk + ttk - k)ag
I)a~ + (N - Ik)a~ = (b - I)a~/k
l)a~ + At L -rNk
l)a~ + r L -r~ + (I - k)ag
l)a~ + (N - I)a~
b - 1 + l)a~
l)a~
l)a~
Na~
(N - k)a~
orr +
+ r L orr +
Na~
r L
a~ + ka~
l)a~
+
+
ESS
[Ch. 11
equireplicate. Although we are, by our notation p ij , tacitly assuming a binary

design, that assumption is not necessary. We suppose that our model is
yir= !L+T, +flr+etr,
where P; , e{; are random normal variables and and Tl are unknown constants.
We have V(y) = ob + o, cov (yt; , yt - ; ) = ab ifj = j' and zero ifj j'.
If V is the covariance matrix of the observations Y, the maximum likelihood estimates are obtained by minimizing the quadratic form [Y' E(Y')]
V -1 [Y E(Y)] of the multivariate normal distribution of the y.
If the observations in Y are arranged by blocks, V consists of a series of square
matrices a Ik + abJ k along the main diagonal and zeros elsewhere. Then V'
consists of matrices Ik /al [abJk /o(Qe + kab)] along the main diagonals and
zeros elsewhere. The quadratic form to be minimized in order to obtain the
maximum likelihood estimates may now be written
= 11(y,r *
T*) 2
oe(Q f
kab)
(B' k *
ntiT)2
The asterisks denote the estimates of the parameters; in particular, * = G/N.

In matrix notation, the second member is
( 12
\, e
12 (B k *1 N'T*)'(B k*1 N'*).

ar )
Differentiating,
Differentiating, and recalling that Rl = Nl, we have
f
R1* RT *]
(Fa
8r = Q [T
NB Rlk* NN'r*] = 0,
whence
b1 NN',r*] =0.
Q[QAs *]+Qf [NB R
The first term alone gives the intrablock estimates; the second term gives the
interblock estimates.
Writing w l = 1 /v and w2 = k /v? gives
(w 1 A
+ k NN') T* =
w,R
[w,Q
w1
k w2) NN') T*
+ k N ` B b111
Sec. 11.14]
The General Balanced Design 245
The individual equations may be written

Pi = R i 1 T* Ah{T
k
h#t k
where, for binary designs,

Pi = Wi Qi +
k (1 nijB; w(riG/ b))
Ri = ri(wi + w2/(k 1 )),
Aht = Ahi(wl w2),
and > hi = 1 n hjn ij For a binary design, A h i is the number of times that the hth
and ith treatments appear in the same block. The combined estimates can thus
be obtained by substituting P i , R, and A, for Q, ri , and 'h in the solutions to
the intrablock equations. As before, we substitute for a and ab estimates from
the intrablock and interblock analyses of variance.
For the balanced incomplete block design, the methods of Yates and of Rao
give the same combined estimates. This is not, however, true in the general case
[see Sprott (1956a)]. It should be noted that the use of Rao's method does not
require that the matrix NN' be of full rank. We shall use this method in the
next chapter for the simple lattice design in which n 2 treatments are tested in a
design with k = n, b = 2n, and r = 2, in this case, b < t.
11.14 The General Balanced Design

A general incomplete block design is said to be balanced if, with T = 0, in
the intrablock analysis, V(T h Ti ) is the same for all pairs of treatments h and i,
h 0 i. Atiqullah (1961) showed that this implied that if the covariance matrix of
the estimates is Vol, the diagonal elements of V are all equal and the offdiagonal elements are all equal.
Solving the normal equations by the first method with H = 1', we have
cov 4 = B 11 a 2 so that V = B and Vl = B 11 1 = 0. Denote the elements of V
by V ht . Then, if the design is balanced,
1J (Th Ti) = (Vhh + vtt 2 vht )a 2 = da 2
for all pairs h, i, where d is some constant. Summing over all i, 1 < i < t, we
have
(t 1)d = (t l)vhh + r
'
vii 2
r / vht,
where the prime denotes summation over all i, except i = h. However, since
Vl = 0, l i vhi = 0, and the equation becomes
(t 1)d = tv,^ h + tr (V),
[Ch. 11
where tr (V) is the trace of the matrix V. It follows that

v,, h
= {( t
1)d tr (V) }/t
for each h, and the diagonal elements of V are all equal. Furthermore, all the
off-diagonal elements of V are also equal, and, since v hh + ( t l)V hi = 0, V
must be of the form
V = c(It
Je /t),
where c is some constant.

If the side condition Hr = 1'T = 0 is imposed in solving the equations, we
have AB = I H'B 21 = I J /t. Suppose B 11 = V = c(I J /t). Then
AB = cA cAJ /t = cA,
and
A = c -1 AB 11 = c -1 (I J /t).
Conversely, if A = c(I J /t), we have B 11 A = cB ll and B 11 = c -1 (I J /t).
We have thus shown that a necessary and sufficient condition for a design to be
balanced is that A shall have its diagonal elements equal and its off diagonal
elements equal.
An equivalent condition is that the nonnegative latent roots of A should be
equal [Rao (1958)]. It follows that if a proper binary design is balanced, it must
be equireplicate and be a balanced incomplete block design as defined in the
previous sections. If, however, the requirement that the design be binary, or the
requirement that the design be proper, is relaxed, balanced designs can be
obtained that are not equireplicate [John (1964c)].
We now generalize the definition of the efficiency of a design and compare the
average variance of a comparison (Th - Ti) to the variance in a completely
randomized design with no block effects and r = r = N/t. We then have
E = 2a 2 /rv,
where
v = { t(t 1)} 1(v

-
hn
vti 2vnt)Q 2 ,
and so E = N -1 t8, where 8 = (t 1)/tr (V).

It is now appropriate to consider how E might be maximized, and how E
might depend upon the choice of the design. The problem becomes one of
minimizing tr (V), which is the sum of the latent roots of B 11 . The rank of B 11 is
the same as the rank of A, which, so long as we confine our attention to connected designs, is t 1; both B 11 and A have a single zero latent root. It can be
The General Balanced Design 247
Sec. 11.14]
shown that the nonzero latent roots of B 11 are the reciprocals of the nonzero
roots of A with the same multiplicities. Thus, if the nonzero roots of A are
denoted by 0i and each root is counted as many times as its multiplicity,
= (t 1) { or
'}
',
which is the harmonic mean of the nonzero roots. We can proceed to find an
upper bound for 8.
The nonzero latent roots of A are all positive, and so their harmonic mean is
not greater than their arithmetic mean. Hence,
(t 1)S < tr (A) _
(r, > n?/k ; ) = N
m f /k ; ,
where m f = ^; n. The elements n 15 are either zero or positive integers with

l i n, f = k ; so that m f >_ kf and m f >t kf /t. Hence, m ; > max (k ; , k2 /t).
There are two cases to be considered. If t >t k ; , then k f >t kf /t and min (m,)
occurs when k ; of the treatments appear once each in the block. If k f = pt + q
where p is a positive integer, and 0 < q < t, min (m 5) occurs when q treatments
appear p + 1 times each, and the other treatments appear p times; in this case
mi ? q(p +
1 2
)
+ (t q)p 2 = p 2 t + 2pq + q > t
-1 (pt
+ q) 2 = k; /t.
We have, therefore, m i /k >_ 1 if kf < t, and m 1 /k5 >_ k j /t if k! >_ t.

Suppose that b' of the blocks have k ; < t, and that there are N' plots in all
the blocks that contain t or more plots. Then
tr(A)<Nb't 'N'<Nb
-
and E < Eo = {t(N b') N'}/{N(t 1)}.

If we now restrict ourselves to designs with kf < t for all j, we have b' = b and
N' = 0 so that
Eo = {t(N b)}/{N(t 1)}.
For a proper design, with all ki = k, this simplifies to
E o = t(k 1)/{k(t 1)},
which is the efficiency factor for a balanced incomplete block design if one
exists. This derivation of Eo is due to J. Roy (1958), and proves a conjecture
made by Kempthorne (1956a) that in the class of proper designs of block size
k with b blocks and t treatments the most efficient design is the balanced incomplete block design, if' one exists. Nothing in the derivation restricts the
designs under consideration to being binary.
[Ch. 11
11.15 Extended Incomplete Block Designs

In some experimental situations we have t < k < 2t. A reasonable procedure
in this case is to let each block consist of a single replication of the t treatments
together with a block from an incomplete block design with block size k t.
We consider the case in which the incomplete block design is balanced [John
(1963)]. For example, with t = 4, b = 3, and k = 7, we could take the design
ABCDABC,ABCDABD,ABCDACD,ABCDBCD.
Suppose that there are t treatments in b blocks of size k, and that each treatment appears r times. The parameters of the balanced incomplete block design
are t, b, r' = r b, k' = k t, and A. The diagonal elements of NN' are
^nt = 4(r b) + (2b r) = 3r 2b;

the off-diagonal elements are
fl hj ij =4A+ 2(r b A) + 2(r b A) + (b 2(r b A) A)

=2rb+A.
Writing A' = b + A and A* = 2r b + t, we have
NN' = (r ')I t + A*J t ,
and
kA= (rk r+'\')I t ^1*Jt =(tIJ)A*,
which differs from the normai equations for the balanced incomplete block
design only in the substitution of )* for A.
Imposing the side condition H'T = (A*/k) 112 1'T = 0, we have A + H'H =
tA*It /k, and 1 = kQ i /(t.k*). Then V( Th, - T i ) = {(t 1)k/t 2 A*}o, and the
efficiency of the design is E = t,1*/(rk).
The residual sum of squares can be divided into two components: the sum
of squares between duplicates and a remainder attributable to interaction between
blocks and treatments. The analysis of variance table follows. Interblock estimates and combined estimates may be derived in the same manner as they were
for the balanced incomplete block design.
TABLE 11.7
Source
Total
Blocks (unadjusted)
Intrablock error
Interaction
d.f.
N 1
b 1
SS
1 y 2 GZIN
Bj lk G 2 f N
t 1
k 7- Q;/(t ) *)
b(k t)
(b 1)(t 1)
(Between duplicates) _ (diff) 2 /2

By subtraction
Exercises 249
Exercises
1. Is the equation
E(Sm aai) = a (bk t) + o(b 1)
true for any binary, proper, equireplicate design? Is it true if any of the three
restrictions, binary, proper, and equireplicate, are removed?
2. A design for two treatments has p pairs of blocks. Each block has three plots.
In each pair one block has two plots with A and one with B, and the other has two
plots with B and a single plot with A. Analyze the design. How would your analysis
be modified if p = mq, and the p pairs are actually m pairs at each level of a
factor with q levels?
3. Apply the methods of this chapter to the case of a randomized complete block
design in which the first treatment is missing from the first block.
4. In a resolvable balanced incomplete block design consider the first block in the
first replication. Let m, j denote the number of treatments that this block has in
common with the jth block of the !th replication (1 < i <
_ r, 1 < j < b/r).
Show that the average of the m, j is m = k 2 /v, and that their variance is given by
(m, j m) 2 =
k(t
k)(b t r + 1)/{n 2 (r 1)(t 1)},
15
where n = tik. For an affine resolvable design mt, = m for all i, j [Bose (1 942c)].
5. Find the combined estimates for the extended block design, Sec. 11.15.
CHAPTER
12
Partially Balanced
Incomplete
Block Designs
Much of our emphasis in the last chapter was upon balanced incomplete block
designs. Unfortunately, such designs do not exist for all combinations of
parameters that we might wish to use. In particular, the requirement that A be an
integer imposes a severe restriction upon the number of treatments for which
balanced designs of a reasonable size can be found. It will be recalled that if there
are t treatments each repeated r timer in blocks of size k, then A = r(k 1)1
(t 1). If (t 1) is a prime, r has to be a multiple of (t 1), and then there is
the further restriction that b = rt/k has to be an integer. For example, if t = 32,
r must be a multiple of 31; if, in addition, k is odd, and thus prime to 32, b has
to be a multiple of 31 x 32.
The characteristic property of a balanced incomplete block design is that
V( fh - fi ) has the same value for all pairs, h, i, of treatments. An alternative to
requiring balance is to find a design in which the condition of equal variances is
modified somewhat. This can be achieved by the use of partially balanced
designs. In a partially balanced design with two associate classes, V(fh fl )
may take either of two values, depending upon whether the hth and ith treatments are first or second associates of each other. When there are m associate
classes, V( Th - f, ) may take any of m values.
Partially balanced incomplete block designs (PBIB designs) were introduced
by Bose and Nair (1939). Their idea has been developed considerably since then,
largely by Bose and his students. In 1952 Bose and Shimamoto introduced the
concept of partially balanced association schemes and the definition of a
partially balanced design was rephrased in terms of this concept. We shall use
the definition given by Bose (1963) in a review paper.
The symbol v will be used to denote the number of treatments. This will
i^,j1l
Partially Balanced Incomplete Block Designs 251

bring our notation into line with that used by Bose and his students. The
treatments may be denoted either by numbers or by letters. A partially balanced
association scheme with m classes is a relationship between the treatments,
satisfying the following conditions:
(i) Any two treatments are either 1st, 2nd, ..., or mth associates, and the
relation of association is symmetrical. Thus, if a and P are two treatments
and a is an ith associate of P, then P is an ith associate of a;
(ii) Each treatment has exactly n t ith associates, and the number n { is independent of the treatment chosen;
(iii) If a and P are ith associates, the number of treatments that are both jth
associates of a and kth associates of Pis p, and is independent of the
pair of ith associates chosen. Hence, pl k = pki.
The numbers v, n i , Pk are called the parameters of the first kind. The parameters Pìk may conveniently be written in the form of m matrices; P i (i = 1, 2,...,
m) has Pik as the element in the jth row and the kth column.
A partially balanced incomplete block design is a design based upon such a
scheme. It is a proper, binary, equireplicate design, having b blocks of k plots
each with every treatment repeated r times, and the additional requirement that
if a and P are i th associates, then they appear together in exactly t. blocks where
,1 t is independent of the particular pair of ith associates chosen. It was originally
thought that the A l should all be different, but Nair and Rao (1942) pointed out
that this was not necessary. Furthermore, lome of the . may be zero. We shall
usually write PBIB(m) to denote a design with m associate classes. The numbers
b, r, k, and A, are called the parameters of the second kind.
There are several relationships between the parameters of a partially balanced
incomplete block design and the corresponding scheme. Clearly,
>n j A =r(k l),
rv=bk,
^ni =v 1.
It is also readily shown that
Pik = n 1 ,
k
if
ij;
Pk = nj 1,
if
i =j,
and also that

niPìk = nJP+fk = nkP r
To derive the last relationship, consider a treatment a. Let G { denote the set of
ith associates of a, and Gi the set of jth associates. Each treatment in G i has
exactly pfk kth associates in G 1 ; similarly, each treatment in Gf has exactlyp' k kth
associates in G. Thus, the number of pairs of kth associates that can be obtained
by taking one treatment from G i and another from G ; is, on the one hand,
n,p`1k and, on the other, njpik.
252 Partially Balanced Incomplete Block Designs
[Ch. 12
Most of the research carried out so far has been concerned with designs having
two associate classes. Chapter 14 will be devoted to the construction of designs
from association schemes. Bose and Shimamoto (1952) classified all the known
PBIB(2) designs into five types. Simple designs are those for which either A l or
A, = 0. The other designs correspond to one of the following four association
schemes: group divisible, triangular, Latin square, and cyclic. Extensive
collections of designs are given by Bose and Shimamoto (1952), and by Bose,
Clatworthy, and Shrikhande (1954). For m = 2, the relationships between the
parameters just given take the simpler form:
+n 2 =v 1,
Pij +
n1A1 + n 2 A 2 = r(k 1),
p1 2 = n l 1,
P21 + Pz2 = n2,

P21 + PL = n 2 1,
Pi + P12 = ni,
niP12 = n2Pii,
n1P22 = n2P12.
12.1 The Group Divisible Association Scheme

This is a scheme for v = mn treatments where m and n are integers. The
treatments are divided into m groups of n treatments each. Treatments in the
same group are first associates; treatments in different groups are second
associates. This corresponds to writing the v symbols for the treatments in a
rectangular array of m rows and n columns, and letting the rows be the groups.
For this scheme v = mn, n l = n 1, and n 2 = n(m 1),
Pl
n -2
0
n(m )]'
= 1[ 0
P2
0
= [n
n-1
n(m 2),
It can be shown that a necessary and sufficient condition for a PBIB(2) association scheme to be group divisible is that p1 2 (or P1 2 ) = 0. It is customary to
abbreviate group divisible to GD.
As an example of a GD design, consider the following design for v = 6:
1,2,3; 3,4,5; 2,5,6; 1,2,4; 3,4,6; 1,5,6.
The three groups are 1 and 2, 3 and 4, 5 and 6. We have A l = 2, X1 2 = 1.
We shall show in Chapter 14 that the following inequalities hold for all GD
designs:
r>_Al,
rkA2 v>-0.
Bose and Connor (1952) have divided the GD designs into three classes:
(i) Singular designs having r = A l ;
(ii) Semiregular designs having r > d l and rk = A 2 v;
(iii) Regular designs having r > A l and rk > Azv.
Sec. 12.2]
Intrablock Analysis of GD Designs 253
A trivial example of a singular design is obtained by letting each row of the

rectangular array constitute a block.
12.2 Intrablock Analysis of GD Designs

The intrablock equations for a GD design are easily solved. In Chapter 11 we
derived the intrablock equations, Q = A-?, for the general incomplete block
design. The reduced equation for .the hth treatment is
Qh
nhiniifil k.
nhfTh
Writing S1(h) = 1 s (TS) where the summation is made over all the first associates
of the hth treatment and defining S2 (fh ) similarly for the second associates,
we have
kQ h = r(k l)fh
1S1(h)
A2SI(fh)
_ {r(k 1) + 1 }Th 1{Th + S1(Th)} A2S2(rh)

Suppose now that the treatments are labeled in order so that treatments 1
through n form the first group, treatments n + 1 through 2n form the second
group, and so on. The matrix A then consists of square submatrices
{r(k 1) + k 1 }In A1J.n
along the main diagonal and submatrices 1\ 2 J n off the diagonal. We take the
side condition 1 Ti = 0 in the form HT = 0 where H = hl and h 2 = A2 /k.
Then k(A + H'H) consists of square matrices {r(k 1) + A l } I + ( a 2 ,1 1 )Jn
along the diagonal and zero elsewhere. The solution matrix (A + H'H) -1
thus consists of submatrices
k _ (,12 A l )J
r(k 1)+j[I {r(k 1) +A 1 +nA2 na l },
r(k-1)
+a LI (A2 va
k 1
Al)
JJ
along the main diagonal and zero elsewhere.

It follows that
{r(k 1) + A }f = kQh k(\ 2 'Al){Qh + Sl(Qh)}/(vA2)
where Sj (Q h ) is the sum 1 Q for all the first associates of the hth treatment.
$
[Ch. 12
The expression Q h + S(Q) is the sum of the Q. for all the treatments in the
group. We also have
2ka2
V(Th-T(
)= r(k- 1)+A l '
if h and i are first associates, and

2ka2
V(ThTi)=r(k-1)+,1
i r 1
_(A2 i) 1
mnJ J'
if they are second associates. In the example given earlier with v = 6, m = 3,

n = 2, we have
T^ = ( 7 Q1 + Q2)/ 16 ,
V( T1 - T2 ) = 3a 2 /4,
T2 = (Qi
+ 7Q 2)/16,
V(T1 - T3) = 70 2 '8.
12.3 The Triangular Scheme

There are v = n(n 1)/2 treatments. The association scheme is obtained by
writing the treatments as the elements above the diagonal of a square array of
side n, leaving the diagonal blank, and repeating the treatments below the
diagonal in order to obtain a symmetric array. Thus, if n = 5, we have
* 1 2 3 4
1 * 5 6 7
2 5 * 8 9
3 6 8 * 10
4 7 9 10 *
Treatments appearing in the same row, or column, are first associates. Thus the
first associates of 1 are 2, 3, 4, 5, 6, 7, and the second associates are 8, 9, 10. For
this scheme
nl = 2(n 2),
n 2 = (n 2)(n 3)/2,
n-3
_ n-2
Pl [n-3 (n-3)(n-4)/2,'
_
PZ
4
2(n-4)
[2(n-4) (n-4)(n-5)/2]
The following design has b = 10, k = r = 4, d l = 1, A 2 = 2.

2,10,6,7; 10,1,2,5; 7,3,8,2; 6,2,9,4; 1,9,10,8;
5,4,3,10; 8,7,4,1;
3,5,7,9; 9,6,1,3; 4,8,5,6.
Sec. 12.5]
The Cyclic Association Scheme 255
Another way of deriving this association scheme is to let one treatment correspond to each of the ordered pairs of integers (x, y), 0 < x <y < n. Then two
treatments are first associates if their representations have an integer in common.
The number pairs are equivalent to the coordinates in the upper triangle of the
array mentioned before. Thus we can associate 1 with (1, 2), 2 with (1, 3), and
so on. Then the first associates of (1, 2) would be (1, 3), (1, 4), (1, 5), (2, 3), (2, 4),
and (2, 5).
12.4 The Latin Square Scheme (L i)

There are v = n 2 treatments. For the L 2 scheme the treatments are arranged
in a square array, and two treatments are first associates if they appear in the
same row or column. For the L, scheme, i > 2, we take a set of i 2 mutually
orthogonal Latin squares and superimpose them on the array. Then two treatments are first associates if they appear in the same row or column of the array,
or if they correspond to the same letter in one of the squares.
For this scheme n l = i(n 1), n 2 = (n i + 1)(n 1),
i 2 -3i+n
(i 1)(ni+1)
Pl [(i 1)(ni+ 1) (ni)(ni+ 1)]^
_
P2
i . 1)
i(n i)
i(ni) (ni) 2 +(i-2)]^
For n = 3 we may take the array
1 2 3
4 5 6
7 8 9
and obtain the following L 2 design:
123, 456, 789, 147, 258, 369,
in which i\ 1 and A 2 = 0. This is a simple lattice design; it is further discussed

later in this chapter.
12.5 The Cyclic Association Scheme

The treatments are denoted by the integers 0, 1, ..., v 1. The first associates
of treatment i are i + dl , i + d2i ..., i + d 1 , (mod v), where dl , ..., d,, 1 are
integers satisfying the following conditions:
(i) The di are all different, and 0 < d, < v for each j;
[Ch. 12
(ii) Among the n l (n l 1) differences d5 d; - (mod v), each of the integers

dl , d2 ,. . . , d, occurs g times, and each of the other n 2 positive integers
less than v occurs h times.
Then n 1 g + n 2 h = n1(n1 1 ), g = Pil, h = P119
g
n1g-1
n 1 g-1 n 2 n 1 +g+1
P1 =
h
nl h
P2 =
n l h n 2 n 1 +h-1
As an example consider the following design for v = 5:
012, 123, 234, 340, 401.
Here n 1 = n 2 = 2, A l = 2, A 2 = 1, dl = 1, and d2 = 4. The differences d1 d5

are 4 1 = 3 and 1 4 - 2 (mod 5), so that g = 0 and h = 1. The first
associates of 2 are 2 + 1 = 3 and 2 + 4 - 1.
Much less attention has been paid to this scheme than to the three schemes
mentioned earlier. A fifth association scheme, the negative Latin square scheme,
has been discovered by Mesner (1967). (See section 15.1.)
12.6 The Intrablock Equations for PBIB(2) Designs

We have already seen, when we considered GD designs, that the hth intrablock
equation may be written
kQh = r(k 1)
S1 ( h ) A2S2(h).
(12.1)
Writing Sl (Q h) and S2 (Q h) for the sums of the Q S for the first and second associates of the hth treatment, we have
kSl(Qh) =
iln1 ,, + S1(Th ){r(k 1) A1P111 A2P12}
+ S2(){
kS2(Qh) =
i\1Pi1 ^2Pi2},
(12.2)
'2n2Th + Sl( Th){ A 1Pi2 A2P122}
+ S2 (Th ){r(k 1) >'1P12 "2P22}.
(12.3)
There are two methods of continuing the analysis. They are given by Rao
(1947a), and by Bose and Shimamoto (1952).
The Intrablock Equations for PBIB(2) Designs 257
Sec. 12.6]
Rao's Solution: We eliminate S2 (T h ) by adding to Equation 12.1 the quantity

n 2
T4 = '2{ T h +
S1(h)
2( fh)} = 0,
and adding (aspil + A2p22) to Equation 12.2.

Then, after some simplification,
kQh = Al2Th + B12S1(fh),
kSl(Qh) = A22Th + B22S1(Th),
where
Al2 = r(k 1) + '2,
B12 = A2 A1,
A22 = (A2 al)P125
B22 = r(k 1) + A 2 + ("2 A1)(pil Pil)
whence, writing Al2B22 A22B12 = 0
ArTh
= k{B22Qh B12S1(Qh)}
(12.4)
(We use A, for Rao's A; Bose and Shimamoto use A for a different quantity.)
This gives a solution matrix P with p i; = B 22 kAr 1 and p h; _ kB 12 Ar 1 if h
and i are first associates, Phi = 0 if h and i are second associates.
Hence, V(Th T{) = 2k(B 22 + B 12 )Or lv 2 if h and i are first associates and
V(T h T t ) = 2kB 22 AT 1 Q 2 if h and i are second associates.
EXAMPLE 1. The triangular design given in the previous section has v = b =
10, r = k = 4, A = 1, a 2 = 2, n 1 = 6, and n 2 =3,
P1
= [2 1]'
P2
= [2
0]
A l2 = 14, A 22 = 2, B 12 = 1, B22 = 13, and A, = 180. Then 45T 3 = 13Q 1

i=2 Qi.
V(T1
T2) = 28a 2 /45,
Y (T1
f,) = 26a 2 /45.
EXAMPLE 2. For the GD design, Al2 = B22 = 7, B12 = A22 = 1, and

A, = 48. Then T l = 3(7Q 1 + Q 2 )/48, V(T1 T2) = 360 2 /48, and V(T1 T3) _
420 2 /48.
The Solution of Bose and Shimamoto: It is convenient to write Equations

12.2 and 12.3 as
kSl(Qh) = 1 n 1 Th + allS1(Th) + a12S2(fh),
/
kS2(Qh) = A2n2Th + a21S2(Th) + a22S2(Th)
(12.5)
[Ch. 12
Consider the linear combination Lh = k2 Q h + c l kS1 (Qh ) + c 2 kS2 (Q h). Substituting from Equations 12.5 and 12.1,
Lh = rk(k 1)fh (c,A l n 1 + c2)2n2)fh + (a11c1 + a21c2 A^k)S^ (fh )
+ (a12 c1 + a22c2 A2k)S2 ( Th ).
(12.6)
We choose c l and c2 so that Equation 12.6 becomes

k 2 Qh + cj kS1 (Q h ) + c 2kS2 (Qh ) = rk(k 1) fh .
This requires that

A l n l c l )) 2n 2 c2 = 1\ 1 k + allcl + a 21 c2
= ) 2k +
a 12 c 1 + a22c2,
giving
)1k = (a11 + A1ni)ci + (a 21 +
aak = (a12 + aini)ci + (a22 +
A2n2)c2.
2n2)c2,
Solving by determinants, c l = D l /D, c2 = D 2 /D, where

D = (a11 + A1n1)(a22 + A2n2) (a12 +
Aln1)(a21
+ A2n2)
and D l = A 1 k(a22 + A2n2) a 2k(a21 + A2n 2). After some simplification we

obtain
D= (rk r+)l)(rkr+a 2 )+(A l A2)
x {r(k 1 )(P2 P12) + A2Pi2 aiPi2}
Bose and Shimamoto write k 2 0 for D. We write their 0 as

shown that
tb
4=D=k2Lb.
Substituting for a al and a22 in D l , we obtain
kL bc l = A l (rk r + A2 ) + (Al A2)(Q2Pi2

Similarly,
kLbc2 = A2(rk r + Al ) + (A l A2)(Q2Pi2 ^1P?2)
The solution matrix P has

k
Pu= r(k
1) '
c1
Pht= r(k 1),
Ca
or r(k-1)'
It can be
Combined Intrablock and Interblock Estimates 259
Sec. 12.7]
according to whether h, i are first or second associates, so that

V(Ti - Th) =
or c 2 )a 2
2(k c l )a 2 2(k
r(k 1)
r(k 1)
In our first example, a ll = 5, a 12 = 8, a 21 = 4, and a 22 = 10. The

equations to be solved are
=
8 = 2c 1 + 16c2 ,
11e 1 + 2e 2 ,
whence
D = 160 b = 180,
V(f l T 2
c l = 4/15,
c2 = 8/15,
28x 2 26a2
V(f 1 Ta) =
) = 45 '
45
In practice it is convenient to write the solution as

rk(k 1)T,, = (k c l )kQ h + (c 2 cl)kS2(Qh)
= (k c2)kQh + ( Cl c2)kSl(Qh)
Kapadia and Weeks (1964) give an alternate formula for computing the
adjusted sum of squares for treatments in GD designs which provides a check on
the computations.
12.7 Combined Intrablock and Interblock Estimates

In Chapter 11 we showed that the combined estimates of Yates and Rao were
identical for BIB designs. This is not, in general, true for PBIB designs. Sprott
(1956a) has shown that for the case of two associate classes the two methods give
the same combined estimates only for GD designs.
In using the Rao maximum likelihood approach, Bose and Shimamoto define
an additional constant H by
kH = (2rk 2r + A l + A2) + (Pb Pi2)(A1 a2)
Then, putting
w1
1 _ k _
= ,

wa
vf
1
+
v kab'
Qt = Ti Q G/v and Pt = w1 Q1 + w2 Q,, they obtain the combined

estimators in the form
r{w2 + w l (k 1)}i = kP,, + dl Sl (Ph) + d2S2(Ph),
[Ch. 12
where
di De(w^b(ww2)2
+ rHw2 (ww 2(ww2 ) + rz wa
The weights w l and w2 are obtained by using the estimates of

the analysis of variance tables.
a and a
from
12.8 The Simple Lattice Design

Lattice designs were first introduced by Yates (1937) and were called quasifactorial designs at that time. They have been investigated extensively by
Kempthorne, Federer, and their students. Detailed accounts of these designs and
their analyses are given in the books by Kempthorne (1952) and Federer (1955).
We shall consider here only the simple lattice, and that merely as a special case
of a PBIB(2) design with the L 2 association scheme.
In the previous chapter we mentioned Yates' orthogonal series of balanced
incomplete block designs with v = n 2 , k = n, r = n + 1, b = n(n + 1), and
A = 1. They are resolvable designs and their construction depends upon the
existence of complete sets of orthogonal Latin squares of side n. Such a balanced
incomplete block design is called a balanced lattice. Any two of the squares
constitute a simple lattice; a set of three of the squares is a triple lattice. We
should, however, phrase our definition differently in order to avoid excluding
values such as n = 6, for which the balanced lattice does not exist, but simple
and triple lattices do.
Let v = n 2 . We arrange the treatments in a n x n array. In the first replicate,
two treatments are in the jth block if, and only if, they are in the jth row of the
array. In the second replicate, the blocks are given by the columns of the array.
In the third replicate (for a triple lattice), a Latin square is superimposed on
the array, and two treatments are in the same block if, and only if, they correspond to the same letter in the Latin square.
We now confine ourselves to the simple lattice consisting of the first two
replicates. An example is the first two replicates of the dishwashing experiment
in Chapter 11.
The simple lattice is a PBIB(2) design with the L 2 association scheme. Two
treatments are first associates if they occur in the same row or column of the
array. (The triple lattice is a PBIB(2) design with the L 3 scheme.) We have
v = n 2 , b = 2n, r = 2, k = n, A = 1, and A 2 =0.
Using Rao's method for the intrablock analysis,
A l2 = 2(n 1),
B12 =-1,
A ta = 2(n 2),
B22 = n + 2,
A,.=2n 2 .
Then
2nfh = (n + 2)Q h + Sl(Qh)
The Simple Lattice Design 261
Sec. 12.8]
Furthermore, V(T h T,) = n -1 (n + 1)o2 for first associates, n -1 (n + 2)a 2 for

second associates, and the intrablock efficiency of the simple lattice is (n + 1)/
(n + 3). The estimate Ee of the intrablock error is obtained in the usual way
from the intrablock analysis of variance table:
(n -1 ) 2 Ee=
y2if
n -1 >By ^flQh-
h j
We consider now the combined analysis, incorporating the interblock information. We assume that the block effects /3 are a random sample from a normal
population with mean zero and variance ab, and use Rao's maximum likelihood
method, which was derived in the previous chapter.
To obtain the combined estimates T* we return to the intrablock equations and
substitute Ph , R and A for Q h , r, and ,\ where
Ph = w1Qh + n_ 1 w2(1 nhJBJ GIn)
i
= w1Th + (w2 w1) nhjBaln w2GIn 2 ,
2w 1 + 2w 2 /(n 1),
A = (w 1 w 2 ).
Then, solving the equations, A l2 , etc., become
A 2 = R(n 1) = 2(n 1)w l + 2w 2i A22 = 2(n 2)(w 1 w2 ),

B2=
(wl
w2),
B 2 = (n + 2)w 1 + (n 2)w 2 ,
A* = 2n 2 w 1 (w l + w 2 ).
It follows that
Th = [{n(w1 + W2) + 2(w1 W2)}Ph + (Wl W2)Sl(Ph)1I Ar * .
The weights w 1 and w 2 are obtained, as usual, by substituting for a and a their
estimates from the analysis of variance tables. We have a = 2(Eb Ee )In if
the design is considered as being run in separate replications, and a2 = (Eb E )
x (2n 1)n 2 , if no replications term is taken out.
Kempthorne and Federer (1948) have shown that the variance of a difference
between the estimates of two treatments is
V(Th T*) =
(n + 1) w1 + (n l)w2 a 2
nw1 (w1 + w2 )
for first associates, and
V(- *) _
h
(n
+ 2)w 1 + (n 2)w 2 a2
nw 1 (w l + w2)
[Ch. 12
for second associates. The average variance is
_ (n + 3)wi + ( n 1)w2 2 '

(n + 1)w 1 (w1 + w2) a
The overall efficiency of the lattice relative to randomized complete blocks is
thus vr/(2o 2) = wl v, which may be written
1+
2(w1
w2)
(n + 1)(w 1 + w2 )
whence it is argued that the lattice design is more efficient than the randomized
complete block design with the same experimental material.
12.9 Partial Diallel Cross Experiments

In the most common diallel crossing system with v lines, all v(v 1)/2 crosses
between the lines are taken, but neither the parental lines nor the reciprocal
crosses are included. This is the simple diallel cross, or method 4 of Griffing
(1956), and it was discussed in Chapter 4. Because the number of crosses required increases rapidly with v, fractions, or partial diallel crosses, have been
discussed by Kempthorne and Curnow (1961), Curnow (1963), and Fyfe and
Gilbert (1963). In this section we consider the use of partially balanced incomplete block designs with two associate classes in the partial diallel cross. The
analysis is similar to the interblock analysis.
We consider v inbred lines and associate them with the v treatments in the
partially balanced association scheme. Each line is crossed with each of its
second associates, but not with its first associates.
Let y15 denote the response of the i x j cross. As before, our model is
y, f =m+g,+g5 + si +e15 ,
where g{ and gi are the general combining abilities of the ith and jth lines, and
si , is the specific combining ability; we assume that g,, s15 , and e11 are independently (normally) distributed with zero means and variances a9, as, and ag.
The objectives of the experiment are to obtain estimates of the parameters g, and
the components of variance ag, a$, and a.
We estimate the g, by minimizing the sum of squares
m gc gg) ,
with the side condition 2 g, = 0. Then m = G/N = 2G/rv where G is the grand
S= (Yir
total of all the observations, N = rv/2 is the number of crosses and r = n a , and
Ti = rm + rr
+ S (g ),
2
where Ti is the total of the responses on the crosses containing the ith line, and
S 2 (fr) is the sum 1 g! over all lines crossed with the ith line.
Sec. 12.9]
Partial Diallel Cross Experiments 263
Writing Q { = T1 2G/v, we immediately have

Qi = rgi + S2(g1)
Summing over all the first associates of i, and setting n z = r,

S1(Q1) = (r + Pi2)S1(8i) + Pi2S2(g1)
Following the approach of Bose and Shimamoto in the intrablock analysis, we

consider the quantity pQ i + n 2 S1 (Q2 ), and choose p so that the coefficients of
S1 (g1 ) and S2 (g1 ) are equal. We have, after some simplification,
p = r 2 + r(P12 Pi2) = r(2r v + 1) + (v 1)P'2

Then, since g, = 0, pQ ; + rSl (Q,) = sgt where
s = pn 2 n j n 2 p + n2p1 = r 2 (2r v) + r(v 2)Pi2.

It follows that V(g h gt ) = 2(p r)a 2 /s if h and i are first associates, and
V(gh g{ ) = 2pa 2 /s if h and i are second associates. It is interesting to note that
comparisons between lines that are not crossed have smaller variances than
comparisons between lines that are crossed.
If we assume, as we have tacitly done, only one observation on each cross, the
variance o 2 is the sum of a and os and E(S9 ) = a 2 + ca9. To obtain the
coefficient, c, of o9 in this expression, we ignore the other terms and note that
Q^ = rgi + S2(gi) 2rg,
where g is the average of the general combining abilities of the lines used in the
experiment. It follows that g i is replaced by g i g, and
E(T11 ) = E[{rgt + S2 (gi )}{gi g}] _ {r(v 1) r}ag/v.

Then, restoring the other terras, we have
E(S9) = E( Ti gi) = (v 1)o + r(v 2)a9.

If n observations are made on each cross, the following analysis of variance table
is obtained. The rums of squares, which are readily found, are not included
here.
TABLE 12.1
Source
d.f.
EMS
Total
General combining ability
(nrv 2)12
v 1
o + no8 + nr(v 2)/(v 1)ag
Specific combining ability
v(r 2)/2
a + no;
Error
(n 1)rv/2
[Ch. 12
With a group divisible design, p1 2 = 0 and the formulae take a more

convenient form with
p=r(2rv+ l),
s=r 2 (2rv).
If, however, there are only two groups, r = v/2 so that s = 0, and the design is
singular. On the other hand, if, instead of crossing each line with its second
associates, we cross it with its first associates, we have
Q, = (n 1 )S, + S1(Si),
S1(Q1) = (n l)gi
+ ( 2n 3 )S1(Si),
so that
2(n 1)(n 2)S i = (2n 3)Q j S1 (Q) = 2(n 1)Q { [Q + S1(QM)].

Thus the latter design is nonsingular except when n = 2. The design consists of
breaking the set of lines into several subsets and making all possible crosses
within each subset.
EXAMPLE. The following data were obtained in a drosophila selection
experiment. There are ten lines and a triangular association scheme. There were
fifteen crossings and the response observed was the number of fourth segment abdominal bristles. The data given are the sums of eight observations on each cross.
(1 x 8) 175, (1 x 9) 193, (1 x 10) 169, (2 x 6) 181, (2 x 7) 169,
(2 x 10) 165, (3 x 5) 190, (3 x 7) 180, (3 x 9) 185, (4 x 5) 179,
(4 x 6) 188, (4 x 8) 171, (5 x 10) 178, (6 x 9) 189, (7 x 8) 165.
For this design we have p1 2 = 2, r = 3, and v = 10 so that p = 9 and s = 12.
We work with the totals of eight observations, and modify the equations just
derived to read
8sg1
or 3 281 = 2 Qi S2(Q1),
= (p r)Q1 rS2(Qi),
where Q i = Ti GIS = Ti 535.4. The estimates of the general combining

abilities follow:
Line
Ti
Q,
32,
k,
537
1.6
19.4
0.606
Line
Tt
Q,
32k,
g,
558
22.6
31.4
0.981
2
515
20.4
18.6
0.581
555
19.6
17.4
0.544
538
2.6
4.6
0.144
547
11.6
24.4
0.762
10
514
21.4
17.6
0.550
511
24.4
31.6
0.987
567
31.6
19.4
0.606
512
23.4
39.6
1.237
Sec. 12.10]
Choosing an Incomplete Block Design 265
In the analysis of variance the correction for the mean was GIJ120 = 59719.408.
The uncorrected sum of squares for crosses was 59870.375, and the uncorrected
sum of squares for individuals y 2 (not shown in the data summary) was 60,171.
The point estimates of the components of variance were
Q = 2.863,
&9 = 0.597.
&s = 0.036,
The analysis of variance table follows:

TABLE 12.2
Source
Total
(Between crosses)
General combining ability
Specific combining ability
Individuals within crosses
d.f.
SS
ms
EMS
119
14
9
5
105
451.592
150.967
138.112
12.855
300.625
15.346
2.571
2.863
a + 8a$ + 21.3309
ae + Bos
o
12.10 Choosing an Incomplete Block Design

The reader has seen in the last two chapters various association schemes for
incomplete block designs. In the next three chapters he will see numerous
designs and recipes for many more. An extensive listing of PBIB(2) designs is
given in the tables by Bose, Clatworthy and Shrikhande (1954). A listing of
balanced designs is found in the Fisher and Yates tables (1953). He might well
ask how in a given experimental situation he should choose a design from this
long list. It is an easier question to ask than to answer because there is no final
answer.
Let us suppose that we have fixed upon the number of varieties v and decided
that a proper, equireplicate design should be used, and that the intrablock
analysis will be used. Kempthorne's conjecture tells us that for any given set of
parameters v, k, N the most efficient incomplete block design is the balanced
design, if there is one. For that design we have an efficiency factor
_Av v(k-1) v _1
E rk k(v 1) v 1 ( 1 k) The efficiency depends only upon v and k. For a given v it is an increasing
function of k.
We should, therefore, try to take k to be as large as we reasonably can. There
are practical situations in which, although the block size could if necessary be
increased so that k = v, the tost of doing so would be to introduce so much
heterogeneity into the plots that the variance, a 2 , would be greatly increased.
The efficiency factor E is calculated by comparing the average of V(Th - T{)
[Ch. 12
to the average variance of 2a 2 Ir that would be obtained with a complete block

design. The variance a 2 does not appear in E because it is assumed that a 2 would
be the same in both cases.
Once we have chosen k, Eo becomes an upper bound upon the efficiency that
we can obtain by picking and choosing designs. There are balanced designs
available for any choice of v, k, for example, the unreduced design in which
the blocks are the collection of all subsets of k out of the v varieties. Unfortunately, especially when (v 1) is prime, there may not be a balanced design with
a small enough value of r to fall within the budget. (It will be recalled that if
v 1 is a prime and k < v, then, in order for to be an integer, (r 1) has to
be a multiple of v 1, and if, further, k and v have no common factor, r has
also to be a multiple of k.) On the other hand, if v 1 is prime, v itself is not,
and so there is at least one GD scheme for the v varieties.
In choosing between partially balanced designs one might take into account
the several efficiencies E where E denotes the efficiency of a comparison between
two ith associates, as well as the overall efficiency E. In practice, except for GD
designs, this is not often very useful. With a GD scheme the experimenter can to a
certain extent control which varieties go into each group, and to that extent he
can divide his varieties into groups of mutual first associates by some rational
procedure. In the other schemes the associate classes are so interlocked that
little such control exists. All that the experimenter can actually do is to find a
design in which the E, are more or less equal and which falls within his budgetary
restrictions. Having the E nearly equal is roughly equivalent to having the At
nearly equal.
Within those guidelines the choice of a design is not really critical. J. A. John
(1966) has published a listing of cyclic designs, i.e., designs obtained by cycling
from initial blocks. He lists the designs for various values of k and v (he uses t
for the number of treatments). It is interesting to note that for most of the
designs for a given v and k, E has about the same value. Two examples serve to
illustrate this.
In a cyclic design we start with an initial block and obtain a set of blocks from
it by adding 1, 2, 3, ..., in turn to each element of the initial block and reducing
mod v. Some designs use several initial blocks. For v = 6 and k = 4, John
lists four designs. Cycling on the initial block (0, 1, 2, 3) gives the design
{
0123, 1234, 2345, 3450, 4501, 5012.
This is a PBIB(3) design. The first associates of i are i + 1, i + 5; the second

associates are i + 2 and i + 4; the third associate is i + 3. We have ,1 1 = 3,
A, = 2, A = 2, and E = 0.8948. Cycling on the initial block (0, 1, 2, 4) gives
the GD design
0124, 1235, 2340, 3451, 4502, 5013,
with E = 0.8942. The other two designs have r = 6. They are each formed by
cycling on two initial blocks and discarding duplicate blocks. The first of these
Exercises 267
designs is formed from (0, 1, 2, 3) and (0, 1, 3, 4). It has three associate classes
and E = 0.8977. The last design is formed from (0, 1, 2, 3) and (0, 1, 3, 4) and
has E = 0.8978. The upper bound is Eo = 0.9000.
John lists nineteen designs for v = 12 and k = 4. They are all partially
balanced, although two of them have no fewer than six associate classes. Their
efficiencies range from 0.8007 to 0.8166. The value of E, is 0.8182. Two of the
designs are GD designs. The first has r = 4, E = 0.8138, and initial block
(0, 1, 3, 7). The second has r = 10, E = 0.8166, and three initial blocks
(0, 1, 2, 5), (0, 1, 3, 8), and (0, 2, 6, 8). There is no advantage in using the more
complicated association schemes.
In summary, three guidelines are suggested. Take k as large as you reasonably
can. Try to find a design with as much balance as possible. The third guideline
is to choose, subject to the other considerations, the simplest design possible.
This is sound advice in the design of any experiment. Keep it as simple as you
can.
Exercises
1. Prove that a sufficient condition for a PBIB(2) scheme to be group divisible is that
either pis = 0 or piz = 0.
2. Show that A,. = k 20 6.
3. Derive the constant d l in the BoseShimamoto form of the combined intrablock
and interblock estimates.
4. Prove that for a GD design the methods of Yates and Rao give the same combined estimates for comparing first associates.
5. Show that for the simple lattice the methods of Yates and Rao do not give the
same combined estimates.
6. Derive the formula given for V(Th - T *) in the simple lattice.
CHAPTER
13
The Existence and

Construction of
Balanced
Incomplete
Block Designs
Balanced incomplete block designs were introduced and discussed in Chapter

11 and partially balanced incomplete blocks in Chapter 12. This chapter and the
next two are devoted to the construction of designs. The problems that we shall
consider are essentially problems in combinatorial mathematics. Many of the
contributions to the construction of balanced incomplete block designs have
been made by algebraists who had no other interest in statistics. Some of their
results antedate the discovery of balanced incomplete block designs by many
years.
This chapter and the next two are quite different from the preceding chapters.
They presuppose more knowledge of abstract algebra. The reader who is not
interested in this topic can safely omit this and the next two chapters; if he wants
a balanced incomplete design for a given practical situation he can get it easily
enough, if it exists, from the Fisher and Yates tables. On the other hand, the
research student in mathematical statistics will find in these chapters that there
remain numerous questions to be answered. The algebraist who knows no
statistics might find them interesting reading, and he will need only to refer back
to Chapters 11 and 12; he does not need to know anything about the analysis of
variance, or, indeed, why the statistician cares about these designs at all. In the
interest of completeness, we shall begin by repeating some basic material from
Chapter 11.
Balanced incomplete designs were introduced by Yates (1936) for use in
agricultural experiments, and the agricultural terminology has remained in the
jargon. A balanced incomplete block design, or BIBD, is a collection of b blocks,
Construction of Balanced Incomplete Block Designs 269
each consisting of k plots; v treatments, or varieties (v > k), are assigned to the
plots (one treatment to each plot) in such a way that
(i) Each treatment appears exactly r times;
(ii) No treatment appears more than once in any block (a design with this
property is called a binary design);
(iii) Each pair of treatments appears together in exactly A blocks.
The five parameters b, v, r, k, ) are not independent. They satisfy the conditions:
(a) bk = rv = N, which is the total number of plots;
(b) A = r(k 1)/(v 1), which is an integer;
(c) b >_ v.
The inequality was proved by Fisher (1940), and is called Fisher's inequality.
It was derived in Chapter 11, and the derivation is repeated here.
The 1938 edition of the Fisher and Yates tables contained a list of all sets of
parameters satisfying conditions (a) and (b) with r < 10 for which designs were
known to exist, in which case the design was given, and those for which it was
not yet known whether or not designs existed. Bose (1939) solved all the unknown cases in the listing except for twelve. Later workers have reduced the
number of unsolved cases in his list to two. The two open cases are b = 69,
v = 46, r = 9, k = 6, A = 1, and b = 85, v = 51, r = 10, k = 6, A = l; it is
still not known whether designs exist in these cases. Rao (1961) published a list
of all sets of parameters with 11 < r < 15, satisfying conditions (a) and (b),
together with known solutions, and Sprott (1962) did the same for 16 < r < 20.
Bose's 1939 paper is the major work in the field, and we shall draw heavily
upon it in this chapter.
Associated with any incomplete block design is its incidence matrix N = (n 1 ).
N is a matrix of v rows and b columns, and n, f is the number of times that the ith
variety appears in thejth block. For a BIBD, the elements of N are either zero or
one, and NN' = (r ,1)I + AJ where J is the square matrix of order v, each
element of which is unity.
The determinant NN' = (r A)" - '(r + (v 1)A) = (r )' _'rk, which is
positive since A < r.
Thus the rank of NN' is r(NN') = r(N) = v; but r(N) < min (b, v), and so
b > v.
If b = v and r = k the design is said to be symmetric. For such a design N is a
square matrix and INN'I = 1N1 2 . This implies that (r ))v -1 is a square.
Hence, if v is even, a necessary condition for the existence of a symmetric design
is that (r )) is a perfect square. This nonexistence theorem was derived
independently by Schutzenberger (1949), Shrikhande (1950), and Chowla and
Ryser (1950). It shows that conditions (a), (b), and (c) are not sufficient for the
existence of a design. Consider, for example, the case b = v = 22 and r = k =
7; we have A = 2, and so r ^1 = 5, and no such design exists.
270 Construction of Balanced Incomplete Block Designs
[Ch. 13
The simplest series of balanced incomplete block designs are the unreduced
designs; these are the designs that consist of all possible combinations of k
out of the v varieties. For these designs
b=() r=() A (k 2).
It is not necessary to investigate cases for which 2k > v. If a new design of b
blocks and v varieties is formed by assigning to each block those varieties that
do not appear in the corresponding block of the BIBD, the new design is also
a BIBD; it is called the complementary design.
13.1 The Orthogonal Series

These two series of designs, introduced by Yates (1936), depend upon the
existence of a series of s + 1 orthogonal Latin squares of side s; such sets of
squares are known to exist whenever s = pm where p is a prime.
For the first series, b = s(s + 1), v = s 2 , r = s + 1, k = s, and \ = 1. The
s 2 + s blocks fall into s + 1 sets of s blocks each. In each set every treatment
appears exactly once, and the set forms a complete replicate of all the varieties.
A design that can be divided into complete replicates is called resolvable. Each
set can be denoted by a square array of s 2 plots with the rows of the array
corresponding to blocks. The varieties are denoted by the integers 1, 2, ..., s 2
In the first set the jth row, and hence, the jth block, consists of the varieties
(j 1)s + 1,(j 1)s+ 2,...,(j 1)s + s = js. The rows in the second set
are the columns in the first set; the jth row consists of j, j + s, j + 2s,...,
j + (s 1)s. The remaining sets make use of the orthogonal Latin squares. Consider one of the squares composed of letters A, B, C, .... The corresponding set
of blocks is obtained by superimposing the Latin square on the square array of the
first set. Then the varieties in the first row of the new set are those that coincide
with the letter A; the second row contains varieties coinciding with B, and so on.
We illustrate the procedure with an example for s = 4. A set of three orthogonal squares has already been obtained in the chapter on Latin squares, in
which Bose's method of obtaining orthogonal sets was presented. The design,
which was given in Chapter 11, is
.
1 2 3 4
5 6 7 8
9101112
13 14 15 16
15 913
261014
371115
481216
161116
251215
38 914
47 10 13
171214
281113
351016
46 915
181015
27 916
361213
451114.
To show that the design is a BIBD, we note that any pair of varieties cannot
appear together in more than one block. If they appear in the same block of the
first set, they cannot appear in the same row of the second set, and vice versa;
in neither case can they appear in the same block in one of the subsequent sets,
Sec. 13.2]
Series of Designs Based on Finite Geometries 271
for that would imply that the same letter appeared twice in the same row (or
column) of a Latin square. Similarly, if a pair of varieties appears together in a
block of the (i + 2)th set and also in a block of the (h + 2)th set, that contradicts the fact that the ith and hth squares are orthogonal. However, each variety
appears r = s + 1 times, once in each set, and thus its plots appear in the same
blocks as r(k 1) = s 2 1 plots occupied by other varieties; if no variety is
to appear more than once, these 2 1 plots must be taken up by each of the
other varieties appearing once each, and the design is a BIBD with A = 1.
Conversely, the existence of a BIBD with parameters b = s(s + 1), v = s 2 ,
r = s + 1, k = s, and A = 1 implies the existence of a set of s 1 orthogonal
squares of side s. This establishes the nonexistence of a design with b = 42,
v = 36, r = 7, k = 6, and A = 1 since there is not a complete set of Latin
squares of side six.
To each member of the first series, OSI, there corresponds a member of the
second series, 0S2. These are designs for v = S 2 + s + 1 varieties. To each
block of the ith set of the OS1 design an extra plot is added containing the
(s 2 + i)th variety. One extra block is added, consisting of one plot with each of
the new varieties. The resulting design is a BIBD with b = v = s 2 + s + 1,
r = k = s + 1, and A = 1. There is no such design for forty-three varieties.
The designs of the first series are also called balanced lattices; any two of the
replicate sets constitute a simple lattice.
13.2 Series of Designs Based on Finite Geometries

Let s = p" be a prime or a power of a prime. A point in the finite projective
geometry PG(N, s = pn) is represented by a (N + 1)-tuple, each coordinate of
which is an element of the Galois field GF(s), and not all of which are zero.
We write x = (x 0 , x 1 ,. . ., xN), xt e GF(s). Two points, x and y, are identical if
there exists an element 6 of the field such that x t = Oy, for all i.
Altogether there are sN+l 1 such (N + 1)-tuples. They may be divided into
groups of (s 1) members with each member of the same group representing the
same point. Thus the geometry PG (N, s) has (sN+ 1 1)/(s 1) distinct points.
Let x and y be two of these points. The line joining x and y is the set of all
points A 1 x + a2y where Al, A 2 E GF(s). There are (s + 1) distinct points on this
line (or one-dimensional subspce), namely y, which corresponds to A l = 0,
X1 2 = 1, and the s points obtained by putting A l = 1 and letting X1 2 take all s
values in the field. The point x is joined by lines to Bach of the other points; each
of these lines contains s points in addition to x, and none of the points lies on
more than one of the lines. Hence, [(s'+ 1 1)/(s 1) 1 ]/s = (5N 1)/(s 1)
lines pass through x. Altogether there are
1 5N_1 1 (SN+1 l)(SN 1)
-1 S -1 S +1 (s2- 1)/(\s-1)
sN+1
S
lines, and each pair of points appears on one and only one line.
[Ch. 13
If we now let the points represent varieties and the lines represent blocks, the
system of points and lines gives a BIBD with
S N+1 1
v s-1
(SN+1 1)( S N 1)
b=l (s2-1)
k=s+1, a=1.
-1) '
If N = 2, we have the orthogonal series with b = v = s 2 + s + 1, r = k =

s+1.
The simplest example occurs when s = N = 2. Then s takes the valnes 0, 1,

and the arithmetic is carried out mod 2. The seven points are A (100), B (010),
C (001), D (011), E (101), F (110), G (111). The seven lines are x o = 0, BCD;
xl = 0, ACE; x 2 = 0, ABF; x o + x l - 0, CFG; xo + x 2 - 0, BEG; x l + x 2
0, ADG; x o + x l + x2 = 0, DEF. This design may also be obtained from the
geometrical figure of the complete quadrangle, using as blocks the six sides and
the live consisting of the three points in which the pairs of opposite sides meet.
We now extend these ideas from one-dimensional subspaces to m-dimensional
subspaces. Let pi , i = 1, 2, ..., m + 1 be a set of (m + 1) points whose coordinate vectors are linearly independent; denote the vectors by p i . The set of
points whose coordinate vectors are linear combinations ,l i p{ forms an mdimensional subspace of PG (N, s). If t = i for all i, the two linear combinations
1 Aly i and 1 ,p i denote the same point. Thus the sets of coefficients {^1,} themselves represent the points of PG(m, s). Hence, the number of points in an
m-dimensional subspace is the same as the number of points in PG(m, s),
namely (sm+l 1)/(s 1).
To obtain an ordered set of (m 1) linearly independent points from
PG(N, s), we may choose the first point in (sN+' 1)/(s 1) ways, and the
second point in {(sN+1 1)/(s 1)} 1 ways; for the third point the choice is
lirnited to those points that do not belong to the one-dimensional subspace
generated by the first two points; it may be chosen in {(sN+1 1)/(s 1)}
{(s 2 1)/(s 1)} ways. Similarly, the (i + 1)th point may be chosen in
{(S +1 1)/(s 1)} {(s t 1)/(s 1)} = {s' (SN i+1-1)}/(s 1) ways. Hence
there exist rj { s t(sN-1+1 1)/(s 1)} ordered sets of (m 1) independent
points in PG (N, s). By a similar argument there are n {s{(sm-s+1 1)/(s 1)}
ordered sets of (m 1) points in each PG(m, s). Taking the quotient we have
for the number of m-dimensional subspaces in PG(N, s)
-
(s"
D (N, m, s)
1)(SN 1).
1). ..
.. (SN-m+l 1)
1) l. .. (S 1)
(Sm+1 1)(Sm
Each point belongs to

( S N+1 1)(S N 1) ... (SN - m+1 1) x (s m+l
(m+1
1)(S m 1) ... (s 1)
(SN+1
1)
1)
^(N 1, m 1, s)
subspaces and each pair of points belongs to 1(N 2, m 2, s) subspaces.
Series of Designs Based on Finite Geometries 273
Sec. 13.2]
We may now identify points with varieties and subspaces with blocks to
obtain BIB designs with parameters
/ S N+1
v = I
_1
, s-1
=(
) _ '(N, 0, s);
b= t(N, m, s);
r= J?(N 1,m 1, s);
sm+l _ 1 \
I = ^(m,0,2);
s1
/
A = O(N 2, m 2, s),
where I(N, 1, s) = 1. The subspaces, or m-flats, are obtained as the sets of

= 0.
points satisfying sets of (N m) linear restrictions of the form
EXAMPLES. PG(3, 2). All arithmetic is mod 2; N = 3, s = 2. There are
fifteen points.
1. 0001
7. 0111
13. 1101
2. 0010
8. 1000
14. 1110
3. 0011
9. 1001
15. 1111.
4. 0100
10. 1010
5. 0101
11. 1011
6. 0110
12. 1100
For m = 2, each flat is defined by a single equation and the design bas b = v = 15,
r = k = 7, and i\ = 3. The blocks are
xa - 0
123 4 5 6 7,
x 1 -0 123 8 9 10 11,
x2 = 0
145 8 91213,
x 3 =_ 0 246 8101214,
x o +x 1 =0
x o +x 2 -0
x0+xg- 0
x 1 +x 2 =0
x 1 +x 3 -0
x2 + x3 0
x o +x l +x2 -0
x0 + x1 + x3 0
xo + x2 + x3 = 0
xl + x2 + x3 - 0
x o +x 1 +x 2 +x 3 -0
1 2 3 12 13 14 15,
1 4 5 10 1114 15,
246 9 11 13 15,
167 8 91415,
257 8101315,
347 8111215,
1 6 7 10 1112 13,
257 911 1214,
347 9101314,
356 8111314,
356 9101215.
If we take the intersections of pairs of two-flats, we obtain the design for

m = 1 with b = 35 and k = 3. Equating any two linear combinations to zero
also equates their sum to zero, and so the intersections are actually intersections
of three two-flats. For example, x 2 + x 3 = 0, x o - 0, and x o + x 2 + x 3 - 0
give the block 3, 4, 7; x o =_ 0, x l = 0, and x o + x l - 0 give 1, 2, 3; x o + x 3 = 0,

x o + x l + x2 = 0, and x l + x 2 + x 3 = 0 give the block 6, 11, and 13.
[Ch. 13
The PG(2, s) designs with m = 1 are the 0S2 series with v = s 2 + s + 1 and
A = 1. Other PG designs of interest are
PG(4,2) v=31
m=1 k=3
m=2 k=7
b=155 r=15
A=1,
b=155
)=7,
m=3 k=15 b=31
r=35
r=15 )=7,
PG(3,3) v=40 m=1 k=4 b=130 r=13 A=1,

m=2 k=13 b=40 r=13 )=4,
PG(3,4) v=85 m=1 k=5 b=357 r=21 A=1,
m=2 k=21 b=85 r=21 ^t=5.
13.3 Designs Using Euclidean Geometries

If we remove from PG(N, s) all the points in the (N - 1) dimensional subspace x o = 0, we obtain the Euclidean geometry EG(N, s). Dropping the now
spurious coordinate xo = 1, the points are N-tuples, and two N-tuples x and y
represent the same point if, and only if, x, = yi for all !.
These Euclidean geometries provide BIB designs with parameters
v=D(N,0, )- b(N- 1,0, )=sN,
k = 'J?(m, 0, s) - (D(m - 1, 0, s) = s-,
b = t(N, m, s) - (D(N - 1, m, s) = sN - m<D(N - 1, m - 1, s),
r = (D(N- 1, ni- 1, s),
.1 = '(N-2,m-2, ).
When m = 1 and N = 2, we have the orthogonal series OS 1 with v = s 2 ,

k = s, and A = 1. Other EG designs of interest are
EG(3,2) v=8
m=1 k=2
m=2 k=4
EG(4,2) v=16 m=1 k=2

m=2 k=4
m=3 k=8
b=28
b=14
r=7
r=7
A=3,
b=120
r=15
.1=1,
.1=1,
b=140 r=35 A=7,

b=30
r=15
,1=7,
EG(5,2) v=32 m=4 k=16 b=62
r=31
A=15,
EG(3,3) v=27 m=1 k=3

m=2 k=9
r=13 A=4,
b=117 r=13 A=1,
b=39
EG(3,4) v=64 m=1 k=4 b=336 r=21

m=2 k=16 b=84 r=21
A=1,
A=5.
Sec. 13.4]
Designs Derived from Difference Sets
275
13.4 Designs Derived from Difference Sets

xv _ 1 ) be an Abelian group of v elements with addition
Let G: (x o = 0, x 1
as the group operation. We can associate each treatment with a different element
of the group. If v is a prime, the set of residue classes mod v is a convenient group
for our purposes; we associate the treatments with the integers 0, 1, ..., v - 1,
and reduce all sums mod v. Let D = (dl , d2 ,. . ., dk be a proper subset of G
such that every element of G, except x o , can be expressed in exactly A ways as a
difference d, - d; . Then D is called a perfect differente set, or, more conveniently, a difference set. If D i = (dl + xi , d2 + x{ , ..., dk + x i ), xj E G, then
D. is also a difference set. It will be proved in a more general context that the
collection of all v difference sets Di forms a symmetric BIBD with b = v and
r = k. These designs are called cyclic designs; their incidente matrices are
circulants. The set D is called the initial block.
For v = 7 we may take the group of residue classes mod 7. Consider the set
D = ( 0 , 1, 3); dl - d2 6, d2 d1=1, d1 - d3 4, d3 - d1=3, d2 d3
5, d3 - d2 - 2, and each difference other than zero appears exactly once. We
may, therefore, take D as an initial block. Adding 1 to each element of D gives
D l = (1, 2, 4); adding 2 gives (2, 3, 5), and so on. This gives a design for seven
treatments in seven blocks of three. The blocks are
,
...,
0, 1, 3; 1, 2, 4; 2, 3, 5; 3,4,6; 4, 5, 0; 5, 6, 1; 6, 0, 2.
The complementary design with k = 4 and A = 2 is generated in a similar way
from the initial block (0, 2, 3, 4).
A difference set for v = 15 and k = 7 can be obtained by representing each
treatment by a pair of integers (x, y), x = 0, 1, 2, 3, 4, y = 0, 1, 2. Addition is
defined by (x 1 , y l ) + (x 2 , Y 2) = (x 1 + x2 , y l + y2 ) with x l + x 2 reduced mod 5,
and y l + Y2 reduced mod 3. The design is
00, 10, 40, 01, 21, 31, 02; 01,11,41,02,22,32,00; 02, 12, 42, 00, 20, 30, 01;
10,20,00, 11, 31,41, 12; 11, 21, 01, 12, 32, 42, 10; 12, 22, 02, 10, 30, 40, 11;
20, 30, 10,21,41,01,22; 21, 31, 11,22,42,02,20; 22, 32, 12, 20, 40, 00, 21;
30, 40, 20, 31, 01, 11, 32; 31, 41, 21, 32, 02, 12, 30; 32,42,22,30,00,10,31;
40,00,30,41,11,21,42; 41, 01, 31, 42, 12, 22, 40; 42, 02, 32, 40, 10, 20, 41.
If v is a power of a prime, we may use the Galois field GF(v) in a manner
similar to the double modulus system just mentioned. For example, if v = 16
and k = 6 the treatments are represented by elements of the form a o + a l x +
a2x 2 + a 3 x 3 where the coefficients a; are either zero or one. Arithmetic is carried
out mod 2. The representations of the treatments can be abbreviated to the
4-tuples a 1 a2 a3 a4 . The initial block is 0000, 0001, 1000, 1111, 0101, 0011.
We shall see later that the use of difference sets may be extended to include
designs involving several initial blocks. If there are t initial blocks, the design
[Ch. 13
will have b = tv, and r = tk. We shall discuss Bose's two methods of using
differente sets and his two fundamental theorems. After that we shall present
some of the series of difference sets that have been discovered. Before proceeding
to Bose's theorems, we shall mention two theorems about the symmetrie
designs.
13.5 Two Nonexistence Theorems

The following two theorems are important in establishing the nonexistence of
symmetrie balanced incomplete block designs. We state them without proof
because the proofs depend upon results in the theory of numbers that fall outside the scope of this book. The first gives a necessary condition for the existence
of a design with a particular set of parameters v, k, and A which satisfy the
equality A(v 1) = k(k 1). The second theorem gives a condition for the
existence of a difference set, and hence a cyclic design, with a given set of
parameters, when the treatments are represented residue classes mod v.
The first theorem uses the Legendre symbol (alb), which is defined as follows:
(alb) = 1 if a is a quadratic residue of b, i.e., if there is an integer x such that
x 2 = a (mod b); if a is not a quadratic residue of b, (alb) = 1.
Theorem 1. Chowla and Ryser (1950). Let v, k, and A be integers for which a
symmetric BIBD exists. The following conditions hold:
(i) If v is even, k A is a square;
(ii) If v - 1 (mod 4), and if there exists an odd prime p such that p divides the
square free part of k , then (Al p) = 1;
(iii) If v = 3(mod 4), and if there exists an odd prime p, such that p divides the
square free part of k A, then ( Ai p) = 1.
We have already discussed condition (i). Conditions (ii) and (iii) may be
combined into the single statement that a necessary condition for the design to
exist when v is odd is that the Diophantine equation
x2 = (k A)y2 + (- 1)cv_ 2Az2
shall have a solution in integers other than the trivial solution x = y = z = 0.
The following examples illustrate the use of the theorem. The square-free
part of k A is the part remaining when all square factors have been removed.
Thus the square free part of 12 is 12/2 2 = 3.
If v = 29, k = 8, and A = 2, we have v - 1 (mod 4), k ,1 = 6. Taking
p = 3, we note that 1 2 = 1, 2 2 = 1 (mod 3). Hence, A = 2 is not a square, so
that (Alp) = 1, and the design does not exist.
If v = 67, k = 12, and ,1 = 2, we have v= 3 (mod 4), k 1 = 10, and
p) = (315) = 1, and the design does not exist.
p = 5. Then () p)
The Fundamental Theorems of Bose 277
Sec. 13.6]
Theorem 2. Hall and Ryser (1951). Let v be odd and let e be any positive divisor
of v. A necessary condition for the existence of a symmetric design with parameters
v, k, and A, which is cyclic, is that the Diophantine equation
x 2 = (k A)y 2 + (- 1)(e- 1)12eZ2
must possess a solution in integers that are not all zero.
Smith and Hartley (1948) showed that if, for any design with v = b and r = k
(whether balanced or not), we write the design as a rectangular array of v rows
and k columns with blocks as rows, it is possible to arrange the treatments
within the rows in such a way that each treatment appears exactly once in each
column. Such an arrangement is called a Youden square; in the analysis, a sum
of squares can be taken for differences between columns if desired.
It does not follow, however, that every symmetric design is a cyclic design, or
is isomorphic to one in t1 e sense that it could be made cyclic by a suitable
relabeling of the treatments. Bhattacharya (1944b) has obtained a symmetric
BIBD for v = 25, k = 9, and A = 3.
Bhattacharya's design is the following:
ABEFKLQTW ABIJOPQUY ABMNGHQVX

CDEFOPQVX ADEHJKSVY ADILNOSTX
BCMPJKSTX BCEHNOSUW ACEGJLRUX
BDIKFHRUX DBMOJLRVW BDEGNPRTY
EIMHLPWXY GKOHLPQRS FJNHLPTUV
CDIJGHQTW CDMNKLQUY
ADMPFGSUW BCILFGSVY
ACIKNPRVW ACMOFHRTY
EIMFJNQRS EIMGKOTUV
FJNGKOWXY QRSTUVWXY.
However, the equation
x2 = 6y 2 + 5z2
has no solution in integers, other than x = y = z = 0, and so there can be no

cyclic design with these parameters.
13.6 The Fundamental Theorems of Bose

We now extend this concept of generating a design from a difference set to
Bose's methods of symmetrically repeated differences. We consider, as before,
an Abelian group G with m elements (x o = 0, x l , ... , x,n - 1 ); the group operation
is addition.
[Ch. 13
Let v = mn treatments be divided into m subsets of n treatments each. We

associate one subset with each element of the group G, and denote by vi the jth
of the treatments in the subset, which has been assigned to the element x i ;
treatments with the same second subscript, j, are said to belong to the same class.
The notation used by Bose is to denote each treatment by the numeral of its
subset, i, with the numeral, j, of its class as a subscript; thus our v 12 would be
denoted by 1 2 .
Suppose that a block contains k treatments, of which m ; belong to the jth
class. There are two kinds of differences to be considered: pure differences
[j, j] and mixed differences [j, j'], j # j'. The pure differences [j, j] are the
mi(m j - 1) differences of the type (x i - x; .) where x t and x,. correspond to
treatments v tf and v; f in the block; the mixed differences [j, j'] are the m imi .
differences of the type (x { - x{ .) where x i , x; - correspond to treatments vif and
vi , j , in the block (we may have i = i'). If n = 1 there are no mixed differences,
and we have the situation considered in the previous sections.
For example, if G consists of the residue classes, mod 5, and n = 2, there are
ten treatments; consider the block 0 1 , 1 1 , 3 1 , 3 2 , and 4 2 . The differences [1, 1] are
0-1=4, 0-3=2, 1-3=3, 1-0=1, 3-0=3,3-1=2. The
differences [2, 2] are 3 - 4 = 4 and 4 - 3 = 1. The differences [1, 2] are

0-3=2, 1-3=3, 3-3=0, 0-4=1, 1-4=2, 3-4=4. The
differences[2, 1]are3 -0 = 3,3-1 = 2,3-3 = 0,4-0 = 4,4-1 = 3,

4 - 3 = 1.
Consider now a set of t blocks satisfying the following conditions, where m1,
denotes the value of m f in the sth block:
(i) For each j the set of , m ;s (m ;s - 1) pure differences [j, j] contains each
nonzero element of G exactly 2 times;
(ii) For each pair j, j'(j j') the set of 1 s m;sm ; . s mixed differences [j, j']
contains each element of G, including zero, exactly A times.
We say that in these t blocks the differences are symmetrically repeated.
For example, if G is the set of residue classes mod 5, and n = 2, the following
set of t = 6 blocks has the property with A = 2: 0 2 , 1 2 , 2 2 ; 1 1 , 4, 0 2 ; 2, 3 1 , 0 2 ;
11, 4, 22; 2^, 3 , 22; 01, 02, 2 2.
Theorem 1. The first fundamental theorem of differences:

Let G be an Abelian group of m elements, and let v = mn treatments be divided
into m subsets of n treatments each, denoted by v i; , as described previously.
Suppose that we canfind a set of t blocks B 1 , B2 ,. . . , Bi which satisfy the following
conditions:
(i) Each block has k plots and contains k different treatments;
(ii) Among the kt varieties occurring in the set of t blocks, exactly r belong to
each class. (Some treatments may appear 'in more than one block, and each
appearance is counted; thus kt = nr);
(iii) The differences are symmetrically repeated, A times each, in the t blocks.
Sec. 13.6]
The Fundamental Theorems of Bose 279
Suppose further that these t blocks are used as initial blocks in the following
manner. From each initial block B, a set of m blocks is generated. If v h; is a
treatment in B 3 , and Xg = Xh + x t , thn v 9j appears in B P ; the block B 30 is the
initial block B.
The theorem states that the resulting set of mt blocks forms a balanced incomplete block design with v = mn, b = mt, r, k, A.
Let Vhj be a treatment in B3 . The images of Vhj in the set of blocks B8
consist of all treatments v l , belonging to the jth class. Thus each treatment vif
appears in the design exactly once for each appearance of j in the initial blocks.
Hence, by virtue of condition (ii), each treatment v t5 , i = 0, 1, ..., m - 1,
j = 1, 2, ..., n, appears exactly r times in the set of mt blocks.
Consider two treatments: v ij and vt . ; ., j j'. They appear together in a block'
B39 if, and only if, we can find two treatments v, ti5 and v h ; in B, and an element
X g in G such that X h + x 9 = x j and X h ' + X g = x i .. Subtracting, we have the
equivalent condition Xh - xh , = x, - xj . This implies that the pair v, 1 vi . ; . occurs
in exactly one of the blocks generated by B, whenever i - i' is one of the
differences [j, j'] in that block. However, the difference i - i' occurs exactly A
times among the set of differences [j, j'] in all t blocks, and so the pair of treatments vt , v n ', occurs together in exactly ) blocks of the design. A similar
argument involving the nonzero differences [j, j] shows that the pair v;5 , v l f also
occurs together exactly A times. Thus the design is a balanced incomplete block
design with the required parameters, and the theorem is proved. E
It should be noted that the requirement that each of the initial blocks should
contain exactly k plots has not been used. Even if the blocks are of different
sizes, the resulting design, though not a BIBD, is stilt equireplicate, and each
pair of treatments appears in the same block exactly ,1 times.
PROOF.
EXAMPLE 1. The six blocks with k = 3, v = 10 just given may be used as

initial blocks for a design with parameters v = 10, b = 30, r = 9, k = 3, A = 2.
The complete design is
0 2 ,12,22; 11,41,02; 21,3 1 ,0 2 ;
1 2, 2 2, 32; 2 1, 01, 1 2; 3 1, 4 1, 1 2;
2 2, 32, 42; 3 1, 1 1, 2 2; 4 1, 0 1, 2 2;
3 2, 42, 02; 41, 21, 3 2; 01, 1 1, 3 2;
4 2, 0 2, 1 2; 0 1, 3 1, 4 2; 11, 2 1, 42;
1 1, 4 1, 2 2; 21,3 1 ,2 2 ; 01,0 2 ,2 2 ;
2 1, 01, 3 2; 31, 4 1, 3 2; 11, 1 2, 3 2;
3 1, 1 1, 4 2; 4 1, 0 1, 42; 2 1, 2 2, 42;
4 1, 2 1, 0 2; 0 1, 1 1, 0 2; 31, 32, 02;
0 1, 31, 1 2; li, 2 1, 1 2; 41, 42, 1 2.
EXAMPLE 2. v = 9, b = 18, r = 8, k = 4, and A = 3. Gis the set of residue

classes mod 9, m = 9, n = 1. The initial blocks are 0, 1, 2, 4; 0, 3, 4, 7.
EXAMPLE 3. v = 25, b = 50, r = 8, k = 4, i\ 1. G is the set of residue
classes of the double modulus system mod (5, 5), or, equivalently, G can be the
Galois field GF(52). The initial blocks are 00, 01, 41, 13; 00, 32, 21, 02.
[Ch. 13
We begin the discussion of Bose's second method by considering an example.

We can obtain two symmetric designs for v = 7: a design with k = 3, ,1 = 1,
and the complementary design with k = 4, A = 2. If we take the smaller design
and add to each block a fourth plot with a new (eighth) treatment, then all
fourteen blocks comprise a balanced incomplete block design with v = 8,
b = 14,r = 7,k = 4,andA = 3.
This approach can be generalized. We need to find two BIBD, one with
parameters v, b 1 , r 1 , k, A l , and the other with parameters v, b 2 , r2 , k 1, ,1 2 .
To each block of the second BIBD we add an extra plot with a new treatment oo.
If the composite design is a BIBD, its parameters must be v + 1, b = b l + b 2
r = r l + r 2 k, = + a2. In the latter design the treatment oo appears in b 2
blocks, and each of the old treatments appears in r z of these. We must, therefore,
have r 2 = A and r = b 2 , whence b l k = b 2 (v k + 1).
,
Theorem 2. The second fundamental theorem of differences. Let G be an Abelian

group of m elements under addition and, as in Theorem 1, let there be n treatments corresponding to each element x t of G and denoted by v i . Furthermore,
let there be an additional treatment denoted by oo. Suppose that it is possible
to find a set of t + u blocks B 1 . . , Bi , B,. . . , B,* satisfying the following conditions:
,.
(i) Each block contains k plots; each block B 3 contains k of the mn treatments
vij , each block B* contains oo and k 1 of the other treatments;
(ii) Among the kt treatments in the blocks Bs , exactly (mu A) treatments
belong to each class and among the u(k 1) treatments other than o0
occurring in the blocks B*, exactly A belong to each class. (This implies
that kt = n(mu A) and (k 1)u = n.k.);
(iii) The differences obtained Erom the t + u blocks, ignoring oo, are symmetrically repeated.
From each of the initial blocks B $ we generate a set of m blocks in the same
manner as in Theorem 1. From each of the initial blocks B$ we generate a set of
m blocks by the same procedure, except that oo does not change as we go from
block to block.
The theorem states that the resulting set of m(t + u) blocks is a balanced incomplete block design with parameters v = mn + 1, r = mu, k, and A.
The proof follows the same lines as the proof of the first fundamental theorem
(Theorem 1) and will not be given.
EXAMPLE 4. Let G be the set of residue classes mod 11. The following three
initial blocks generate a BIBD with v = 12, b = 33, r = 11, k = 4, and A = 3;
0, 1,3,7; 0, 1,3,9; 00,0, 1,5.
Sec. 13.7]
Designs Derived from Symmetric Designs 281
13.7 Designs Derived from Symmetric Designs

Lemma. Any two blocks of a symmetrie balanced incomplete block design have
exactly A treatments in common.
PROOF. Consider a particular block, which we may take without loss of
generality, to be labeled as the first. Let n denote the number of treatments
common to the first block and the jth block with j 1; it will suffice to show
that n1 = A for all j.
The k treatments in the first block form k(k 1)/2 pairs. Each pair appears
again in exactly t 1 of the remaining blocks; each of the treatments in the
first block appears in exactly k 1 of the remaining blocks. We thus have the
equations
;
nj= k(k 1)=,1(v-1),

nj (nj 1)/2 = (A 1)k(k 1)/2 = A(^1 1)(v 1)/2.
Then n; = ^1, 1 n; = \ 2 (v 1), and 1(n ; A) 2 = 0, so that n ; = A for all j.
There are two methods of obtaining other balanced incomplete block designs
from a symmetric design. The first is the method of block section (sometimes
called residuation). The second is the method of block intersection (sometimes
called derivation). Both are due to Bose (1939).
The method of block section consists of cutting one of the blocks out of the
design, and dropping all the treatments contained in it from the other blocks.
This gives a new design with parameters v* = v k, b* = b 1, r, k A, A.
The simplest illustrations of this method occur in the orthogonal series of
Yates. We have seen how a symmetric design with b = v = s 2 + s + 1,
r = k = s + 1, and ,1 = 1 is obtained by augmenting the orthogonal (balanced
lattice) design with v = s 2 , b = s(s + 1), and k = s, r = s + 1. The method of
block section takes us from the symmetric design to the lattice.
For another example, we may take the symmetric design with v = b = 11,
r = k = 5, and A = 2, and by block section obtain a design for six treatments in
ten blocks of three plots each.
A partial converse result has been obtained by Hall and Connor (1953). They
show that the existence of a balanced incomplete block design with parameters
v* = v k, b* = v 1, k* = k ), r * = k, and d* = A implies the existence
of the symmetric design with parameters v, k, and A when A is 1 or 2; i.e., when
one is given the smaller design it is possible to augment it to obtain the larger
symmetric design. Their result is not true for A = 3. Connor (1952a) mentions
the following counterexample: Bhattacharya gives a design for v = 16, b = 24,
r = 9, k = 6, and ,\ = 3 which contains two blocks that have four treatments in
common. If it were possible to augment this design to a symmetric design for
v = 25, k = 9, and \ = 3, any pair of blocks of the new design would have
[Ch. 13
exactly three treatments in common; with two of the blocks already having four
treatments in common, this is clearly impossible. Bhattacharya's design for
v = 16 was first given in reference (1944a). He repeated it (1944b), together with
the symmetric design for v = 25 and k = 9 mentioned earlier and a second
design for v = 16 obtainable by block section from it.
Bhattacharya's design for v = 16, b = 24, r = 9, k = 6, and A = 3 is
ABGHNO
CEHILN
ADEMNP DFHIKO
BDHJLN ABCKLO
ABCDEF**
DEGILO
CDGJLP
CEGHKM
BEFKLP
EFHJOP
BDIJKM
AFGILM*
AEIJKN
BFGINP
ADGHKP AFHJLM*
CDFMNO CFGJKN
BCHIMP BEGJMO
ACIJOP
KLMNOP**.
The two blocks marked with a single asterisk have four treatments in common;
the two blocks marked with two asterisks have no treatments in common; the
other blocks all have two treatments in common.
The method of block intersection is of less interest. It consists of omitting one
of the blocks completely and retaining in the other blocks only those treatments
that occur in the omitted block. This leads to a balanced incomplete block design
with v* = k, b* = v 1, r* = k 1, k* _ A, and A* = 1. For example,
we may take the symmetric design for nineteen treatments in blocks of nine
plots and obtain from it by block intersection a design for v = 9, b = 18, r = 8,
k = 4, and A =3.
13.8 Designs with r <_ 10

In his 1939 paper Bose left unsolved these twelve cases of sets of parameters
with r < 10 which had been listed in the Fisher and Yates tables. The reference
numbers of the twelve cases are those given in the Fisher and Yates tables.
Reference
Number v
(8)
(10)
(12)
(14)
(17)
(20)
15
22
21
29
16
25
.1
21 7
22 7
28 8
29 8
24 9
25 9
5
7
6
8
6
9
2
2
2
2
3
3
Reference
Number v
(24)
(26)
(27)
(28)
(30)
(31)
46 69
21 30
31 31
36 45
46 46
51 85
k ,1
9
10
10
10
10
10
6 1
7 3
10 3
8 2
10 2
6
1
Since that time ten of the cases have been settled and only two, numbers (24)
and (31), remain unsolved. We have already mentioned Bhattacharya's solutions
to (17) and (20). In 1946 he obtained a solution to (27), and, hence, by block
section, to (26). The impossibility of the symmetric designs (10) and (30) follows
Sec. 13.9]
Series of Designs Based on Difference Sets 283
from the fact that v is even and r A is not a square. The symmetric design (14)
was shown to be impossible as the example following our presentation of the
theorem of Chowla and Ryser (1950). The impossibility of (8), (12), and (28)
then follows directly from the result of Halt and Connor (1953) because their
existence would imply the existence of (10), (14), and (30) respectively. The nonexistence of (8) was established by Nandi (1945). An interesting alternate
derivation of the nonexistence of (8) and (28) is given by Atiqullah (1958).
13.9 Series of Designs Based on Difference Sets

There are numerous series of designs based on difference sets. Most of them
are to be found in four papers by Bose (1939) and (1942a) and Sprott (1954) and
(1956b). Bose's series denoted by Latin letters are given in his 1939 paper, and
his series denoted by Greek letters are given in reference (1942a). Sprott's series
A, B, C, D are in his 1954 paper, and his numerical series are in his 1956 paper.
Almost all these series involve a Galois field of p" elements where p is a prime;
most often there is the requirement that v = p". We shall let x denote a primitive
element of such a field. The standard procedure is to derive a set of initial blocks
and then to generate the design by the method of one of the two fundamental
theorems.
From the equations J (v 1) = r(k 1), rv = bk, we obtain the following
necessary conditions for the existence of a BIB design:
A(v 1) = 0 mod (k 1), Av(v 1) - 0 mod k(k 1).
Hanani (1961) has shown that these two conditions are sufficient for k = 3 and
4 and all A and for k = 5 with A = 1, 4, or 20. (In his 1961 paper the design
v = 141, k = 5, and A = 1 was left open as a possible exception, but in 1965
Hanani gave a solution for that case.) Bose gave methods of obtaining many of
these designs in his 1939 paper.
In particular, for k = 3 and A = 1 we must have v = 6t + 3 or v = 6t + 1.
Bose showed how to obtain designs when v = 6t + 3 for any t, and when
v = 6t + 1 if either v = p" or t is odd. The first of these is his T l series. The
blocks in these designs are known as Steiner triplets after J. Steiner, who investigated them in 1853.
(a) Bose's Tl Series
Let v = 6t + 3, k = 3, and A = 1. We take the set of residue classes mod
2t + 1 and associate three varieties with each of the elements. Consider the
pairs 1, 2t; 2, 2t 1;..., i, 2t i + 1;..., t, t + 1. The differences from the
ith pair are 2t + 1 2i and 2i, and so the set of differences covers all
the nonzero elements of the set of residue classes.
[Ch. 13
The design is obtained from the following set of initial blocks:

{11, 2t 1 , 0 2 } . .. {il, (2t + 1 - i)1, 02} . .. {t1, (t + 1 )1, 02},
{12, 2t2, 03} .
.. {i2, (2t + 1 - l)2, 03} . .. {t2, (t + 1 )2, 03},
{13, 2t3, 01} ... {i3, (2t + 1 - i)3, 01} ... {t3, (t + 1 )3,
0 l},
{01, 02, 0 3 }.
The three smallest designs of the series are

t= 1,
v=9,
r=4,
k=3,
,1= 1,
t=2,
v=15, b =35, r=7,
k=3,
A=1,
t=3,
v=21,
k=3,
%=1.
b=12,
b =70,
r=10,
The design for t = 1 follows. The four blocks in the first row are the initial
blocks.
{11 2 1 0 2} {1 2 22 0 3 ) {132301} {00 3 03}
{2 1 0 1 1 2 } {2 2 02 1 3 } {2 3 03 1 1 } {1 1 1 2 1 3 }
{0 1 1 1 2 2 } {0 2 1 2 2 3 } {0 3 1 3 2 1 } {2 1 2 2 2 3 }.
This design is isomorphic to the OS 1 design for nine varieties.

(b) Sprott's A, B, C, D series
These four series require that v = p". The treatments are represented by the
elements of GF(v), and x is a primitive element of the field. Series D has an
additional requirement. The parameters of the designs that can be constructed
with the series are
v
,1
tk+1
tk
k-1
B
C
D
2t(2A + 1) + 1
2t(2^1 - 1) + 1
4t(4A + 1) + 1
2A+1 t(2+1) A
2A
2tA
,1
4^1 + 1 t(4^1 + 1) A.
In each of the designs there are t initial blocks, 0 < i < t - 1:

A
xi+t xi+2t
xi
it+2t
x,x
0, x i
x2
,x i+4t ,...,x i+4At.,
xi+4A-1)t.
x i+2t
x21+16At
x 2i+8t
In series A all the differences in the initial blocks may be written

xi+(s+u)t - x i+ut = x i+ut(x st - 1)
,s
Series of Designs Based on Difference Sets 285
Sec. 13.9]
where xs` 1 is an element of the field; s = 1, ..., k 1; u = 0, 1, ..., k 1.

For any s, the number of such differences is kt and they are all distinct. If two
such differences were identical, we should have
i + ut = i' + u't
(mod kt),
(u u')t
(mod kt),
so that
and
i i' - 0 (mod t).

But i, i' are both lens than t, and so i = i', u = u'. It follows that as s takes all
its values, 0 < s < k 1, the differences are symmetrically distributed with
A = k 1, and the required design is obtained by the method of the first
fundamental theorem. The validity of the other series is established in the same
way.
The additional requirement for series D is that among the 2? quantities
x 4 ts 1, which are themselves elements of the field, there should be exactly A
even powers of x and d odd powers for each s = 1, 2, ..., 2A. Sprott showed
that if this requirement is violated by any primitive element, x, it is violated by
all primitive elements, and so only one x need be checked. In the case A = 1,
this requirement simplifies to the requirement that x 41 + 1 is an odd power
of x.
The numbered series 1, 2, and 3 of Sprott (1956b) are extensions of his letter
series which also have restrictions in addition to the requirement that GF(v)
exists.
Bose's T2 series for v = 6t + 1 = p", k = 3, and A = 1 is a special case of
Sprott's B series with initial blocks (x`, x '+ 2t , x1+4c) The smallest example occurs
with t = 1 and v = 7. The elements of the field are the residue classes mod 7;
x = 3 is a primitive element, and a difference set is (x, x 2 , x 4 ), or (1, 2, 4). It
should be noted that Bose (1939) also gave a series for v = 12t + 7, k = 3, and
A = 1, which does not depend on v being a power of a prime.
(c) Other Series
Bose also gave series El and E2 for k = 3, A = 2, and v even, series Fl and Fa
for k = 4, A = 1, and G 1 , G 2 for k = 5, A = 1. In his paper, the existence of
designs in the E2 , F2 , and G 2 series was said to be dependent upon the solution of
auxiliary problems. In F2 and G 2 the problem was to find an integer 0 such that
(x 9 + l)/(x 1) is an odd power of x. Bose showed later (1942b) that this
can always be done. The auxiliary problem for the E2 series was solved
by Bhattacharya (1943). Some of these designs may also be obtained as
special cases of Sprott's letter series. The F l series is a particular case of Sprott's
series 3.
[Ch. 13
Bose's designs a l , a 2 , a 3 , and a 4 are special cases of Sprott's A, B, and C with

k = 4 and 5. His series P1i j2 , and y have the following parameters:
k A Requirements
+ 2) 4 2 2t + 1 = p", t >- 2,
5 2 2t+1=p", t>-2,
4(3t + 2) 4 3 12t + 7 =pa
p 1 2(3t
fla 10t + 5
y
Bose's a 3 series is obtained from Sprott's series B by putting A = 2. The other

parameters are v = 10t + 1, k = 5, r = 5t. If we put t = 1, we have Bose's S l
series. This a series of symmetric designs with v = 4X + 3, k = 2A + 1. A
single initial block consists of all the nonzero elements of GF(v) which are
squares. The values of v for which this series is of practical interest are v = 11,
19, 23, 27, and 31. The missing value in the sequence, 15, is not a power of a
prime, and this method does not apply in that case; we have, however, presented
earlier in this chapter a difference set for the symmetric design with v = 15 and
k = 7, using a double modulus system. The difference sets for the other values
of v are
v=11: 13459,
v=19: 145679111617,
v=23: 1 234689 1213 16 18,
v = 27: 001 211 020 021 022 100 111 110 120 121 102 202 221,
v = 31: 1 24578910141618 19202528.
For v = 27 we use GF(3 3) with x 3 = x + 2 and arithmetic mod 3. The treatments are represented by triples ao a 1 a2 corresponding to a o + a l x + a2 x2 .
13.10 Resolvable Designs

A design is said to be resolvable [Bose (1942c)] if it can be split up into groups
of blocks in such a way that each group contains each treatment exactly once.
It was shown in Chapter 11 that for such a design b >_ v + r - 1.
Applying the method of block section to the symmetric designs with v =
4A + 3 = p", k = 2,1 + 1, A gives Bose's Bl series of designs with v = 2a + 2,
b = 4a + 2, r = 2A + 1, and k = A + 1. Bose (1947b) showed that when
2,1 + 1 is a prime, A is odd and ,\ > 1, there are two nonisomorphic designs with
the parameters v = 2,1 + 2, k = A + 1, r = 2,1 + 1; one of these is resolvable
and the other is not. When A is even, there is a single solution which is not
resolvable.
The resolvable design for odd ,1 is obtained from the initial blocks
(cc x x 3 x211-1) (0 x xa x211-1) where x is a primitive element of
GF(2,1 + 1). The nonresolvable design comes from the initial blocks
(cc, x, x 2 , ... ), (0, x, x 3 ,.. . ). The resolvable design is the same as Sprott's
Exercises 287
series 4 design, and it has the additional property that all triples of treatments
occur the same number of times. The design with A = 3 has v = 8, b = 14,
r = 7, and k = 4. It is
00356, co460, 00501, co612, co023, co134, co245,

0124, 1235, 2346, 3450, 4561, 5602, 6013.
Sprott's series 5 designs are also resolvable designs with the property that
every triple occurs the same number of times. They have v = 2k, b = 4(2k - 1),
r = 2(2k - 1), A = 2(k - 1), and it is necessary that 2k - 1 = p". The design
is obtained from the four initial blocks
(0 , x,
cc
x11,
x{+ 3
x{ +2k-3)
= 0 1.
The doubly balanced designs of Calvin (1954) are special cases of these series.
Exercises
1. Show that the complement of a BIBD is also a BIBD.
2. Show that the existence of a design with v = n 2 , k = n, and A = 1 implies the
existence of a complete set of orthogonal Latin squares of side n. Hence there
exists no design with v = 36, k = 6, and ,1 = 1.
3. Show that the equation
x2
= 6y 2 + 5z 2
has no solution in integers other than the trivial solution x = y = z = 0.

4. If v is a prime and v = 4t + 3, show that the quadratic residues of v form a
perfect difference set.
5. (Continuation). Show that if v = 27, and if the varieties are represented by
elements of GF(3 3 ), the quadratic residues also form a difference set. That is the
example given in the text.
6. We stated in the discussion of Sprott's series D that if the necessary requirement
is violated by any primitive x, it is violated by all primitive elements. Prove this.
7. Show that the differences are symmetrically repeated in Bose's resolvable B1
designs.
8. Find counterexamples to show the following:
(i) The equality b = v + r - 1 does not ensure that v = nk where n is an
integer;
(ii) v = nk is not sufficient to ensure resolvability.
9. Show that for an affine resolvable design (i.e., a design with b = v + r - 1) the
parameters can be written as
v = n 2 ((n - 1)t + 1),
k = n((n - 1)t + 1),
b = n(n 2 t + n + 1),
r=n 2 t+n+1,
a =nt+1.
CHAPTER
14
The Existence and

Construction of
Partially Balanced
Designs
In Chapter 12 we initiated our discussion of partially balanced association

schemes and partially balanced designs. These were introduced in 1939 by Bose
and Nair because of the restrictions on the existence of balanced incomplete
block designs, especially in the case where v 1 is a prime. We have postponed
until now the discussion of the existence and construction of designs with the
several association schemes. This chapter, like Chapter 13, is of more mathematical interest than the others. The reader who does not wish to work his way
through it may find a design to fit his requirements by using the tables of designs
assembled by Bose, Clatworthy, and Shrikhande (1954), and by Clatworthy
(1956).
We shall present in the next section some elementary designs that can be
obtained from any association scheme. In the subsequent sections we shall,
for the most part, confine our attentions to partially balanced designs with only
two associate classes, and present methods of obtaining designs with the group
divisible, triangular, and Latin square association schemes. Even for these
schemes there are still some sets of parameters with k < 10 and r < 10 for
which it is not known whether or not a design exists. The eighteen unsolved
cases for the triangular scheme and the five unsolved cases for the L 2 scheme are
listed in the relevant sections of the chapter. More problems of interest remain
to be solved for PBIB(m) designs with m > 2.
The reader will wish to refer to Chapter 12 for the introductory remarks about
partially balanced designs. He will then recall that we begin with a partially
balanced association scheme. A PBIB design with that scheme is then an
equireplicate, proper, binary design, i.e., a design in which the treatments each
appear r times in blocks of k plots each, with no treatment appearing more than
288
The Existence and Construction of Partially Balanced Designs 289

once in any block. Each pair of ith associates appears together in exactly A i
blocks. The A { need not all be different.
We have already given a proof in Chapter 11 of Kempthorne's conjecture
that, of all proper, binary, equireplicate designs with a given set of parameters
(b, v, r, k), the balanced incomplete block design, if it exists, is the most efficient.
The experimenter wil] usually not, therefore, choose a partially balanced design
in preference to a balanced design, nor, if the blocks are large enough, will he
choose an incomplete block design over the randomized complete block design.
The efficiency of an incomplete block design has been shown (Chapter 11) to be
given by N lv8 where S is the harmonie mean of the nonzero latent roots of the
matrix A of the reduced intrablock equations.
Let N be the incidence matrix of a partially balanced design. We shall show in
Chapter 15 that if 0 is a latent root of NN' with multiplicity a, then
-
O j = (rk 9 t )^k
is a root of A with the same multiplicity a { . For every connected design NN' has a
root 0 = rk with a o = 1. It is only meaningful to talk about efficiency for
connected designs, in which case
E= (v 1)I(r ajoi 1).
The largest value of E that can be hoped for with v varieties in blocks of size k
is
E = (dv)/(rk)
where it(v 1) = r(k 1), i.e.,
v(k-1)
vk
.
-1
Emax =
k(v-1)
k(v 1)
For any design the parameters r, k, n i , and ,1, must satisfy the equation
r(k 1) _
,li nt .
Furthermore, since NN' is a nonnegative definite matrix, its roots 0 must all be
nonnegative. We shall see in Chapter 15 that for a PBIB(m) design there are,
other than the simple root 0 = rk, only m distinct roots 9 ; ; their values and the
values of the multiplicities a are readily obtained by the methods to be given in
that chapter. This provides us with a convenient method of establishing the
nonexistence of designs with some sets of parameters. In addition, for a
symmetric design the determinant INN'1 is a square and so rk0i10z2... is a
square. Thus, for example, if all the roots are positive and if is even for all
i, i > 1, but a l is odd, then 0 must be a perfect square. This is, of course, an
extension of the method that we used earlier to show that in a symmetric balanced
incomplete block design with v even, (r a) must be a square.
290 The Existence and Construction of Partially Balanced Designs
[Ch. 14
For some association schemes, other conditions for the nonexistence of

symmetric designs have been derived. These conditions make use of the Hilbert
symbol (a, b), and depend upon the HasseMinkowski invariant. Because the
calculation of these invariants is usually quite tedious, we shall content ourselves
by giving references where needed to the appropriate papers. The latent roots of
NN' are also of importance in linked block designs.
14.1 Duality and Linked Block Designs

The concept of duality comes from projective geometry in which it is possible
to obtain from one theorem a second one by interchanging in the statement such
words as line and point. The statements "There is one straight line common to
any two points" and "There is one point common to any two lines" are duals.
In the present context, the dual design is obtained by interchanging blocks and
varieties.
Consider the unreduced BIBD with v* = 4, k* = 2. The parameters are
v* = 4, b* = 6, k* = 2, r* = 3, and A* = 1; the design is AB, AC, AD, BC,
BD, CD.
In the dual the varieties are a, b, c, d, e, f, corresponding to the six original
blocks. The statement "Variety A appears in blocks 1, 2, and 3" becomes
"Block 1 contains the varieties a, b, and c." This gives the new design abc; ade;
bdf; cef with parameters v = b* = 6, b = v* = 4, k = r* = 3, and r = k* = 2.
The dual design is actually a triangular PBIB(2) design. It does not, however,
follow that the dual of a balanced incomplete block design is always a PBIB(2)
design. This will be discussed further in Chapter 14.
Youden (1951) defined linked block designs as designs with the property that
each pair of blocks has the same number (p) of varieties in common. It follows
that the dual has the property that each pair of varieties has the same number ()
of blocks in common, i.e., the dual is a balanced incomplete block design.
If = 1, we have singly linked blocks; if = 2, we have doubly linked blocks.
The main properties of linked block designs are given by J. Roy and Laha
(1957a,b).
Let NN' be the interblock matrix for a linked block design; N'N is the interblock matrix for the dual (BIBD). We recall that if A and B are matrices
such that AB and BA both exist, then AB and BA have the same nonzero latent
roots. Thus NN' has the same nonzero latent roots as N'N, namely r*k* and
r*(v* k*)/(v* 1) = r* A*. Furthermore, NN' and N'N have the same
rank and the same trace. This leads to the following result, due to J. Roy and
Laha (1957b).
A necessary and sufficient condition that a design be a linked block design is
that NN' has only one nonzero latent root other than rk, namely,
_ k(br)
e b-1
and the multiplicity of 0 is a = b 1.
The Elementary Designs 291
Sec. 14.2]
Shrikhande (1952) gave two cases in which the duals of balanced incomplete
block designs are PBIB(2) designs. The first is the dual of a balanced incomplete
block design with parameters
v*=rkk+l,
r*=k,
A*=1.
It is a PBIB(2) design with parameters
v=k(krk+ 1)/r,
k*=r,
n 1 = r(k 1),
b=rkk+ 1,
n 2 = (k r)(r 1)(k 1)/r,
r,k,
A l = 1, d 2 = 0.
The P matrices are
[ (k2)+(r-1)Z

P
l
(r
1)(k
r)
(r-1)(kr)
(r 1)(k r)(k - r 1)/r]
_r 2 r(kr 1)
P ^ [r(kr 1) n 2 r(kr 1) 1,
This is called the singly linked block (SLB) association scheme [Bose (1963)].
If we take the balanced design with v* = 9, b* = 12, r * = 4, k* = 3, and
d* = 1 used in Chapter 11 and number the blocks 1 through 12, the dual is the
design
1,4,7,10; 1,5,8,11; 1,6,9,12; 2,4,9,11; 2,5,7,12;

2, 6, 8, 10; 3,4,8,12; 3,5,9,10; 3,6,7, 11.
The second case is the triangular doubly linked block design, which is
discussed later in this chapter.
14.2 The Elementary Designs

Corresponding to each association scheme are designs that we shall call the
elementary designs. The first of these designs are the elementary symmetric
designs denoted by ES(i) and E'S(i). In the design ES(i) where i = 1, 2, ..., m,
the hth block consists of all those varieties which are ith associates of the hth
variety. For the design ES(i), we have r = k = n, and A t = p, t = 1, 2, ..., m.
In the design E'S(i), the hth block consists of the hth treatment together with
all its ith associates, and so r = k = n, + 1, ,1s = pi ; + 2, and A t = p,, (t 0 i).
For the group divisible association scheme, the designs ES(1) and E'S(l) are
both unconnected. If there are only two associate classes, the ES(l) design and
the E'S(2) design are complementary, as are the E'S(l) and ES(2) designs.
For breeding experiments, designs with k = 2 can be obtained by crossing
each variety with its ith associates. We gave an example in Chapter 12.
[Ch. 14
We can also make up elementary designs by combining a collection of v BIB

designs. The hth of these balanced designs is constructed from the set of ith
associates of the hth variety and has parameters b*, v* = n i , k*, r*, and A*. Then
we have for the combined design b = vb*, k = k*, r = n ( r*, and A t = pi,a *. If
we now add the hth variety itself to each of the blocks in the hth BIBD, we have
a design with b = vb*, k = k* + 1, r = n t r* + b*, Ai = 2r* + pA*, and
k= PILA * (t0i).
14.3 Group Divisible Designs

This was the first association scheme to be investigated extensively. Contributions have been made by Bose and Connor (1952), Connor (1952b), Bose et al.
(1953), and Freeman (1957). The matrix NN' has each diagonal element r; the
off-diagonal elements are either A l or ' 2 . For the GD design, if we order the
varieties by groups so that the first n varieties are in the first group, and so on,
NN' consists of submatrices (r a 1 )I n + ) 1 J
the main diagonal and
submatrices Mn off the diagonal. We can evaluate the determinant as
=
(r + (n 1)A1 + ( m 1)n2)(r 1 + n('\1 ,\2))m-1(r A1)m(n -1) Substituting r 0 for r we see that the latent roots and their multiplicities are
00 =r +(n 1)A l +(m 1)nd2 =rk,
a 0 =1;
B 1 = r+,1 1 (n 1),\ 2n=rkA 2 v,
a1 = m 1;
02 =rA l
a2=m(n-1)
The latent roots and their multiplicities will be obtained more easily later. For
binary designs we cannot have A l > r, and so the condition 0 2 >_ 0 is trivial.
However, the condition 0 >_ 0 can be used to establish the nonexistence of
some designs, as the following example shows.
Let m = 6 and n = 4. The condition r(k 1) = l { n i A becomes
r(k 1) = (n 1).\ + n(m 1)A 2 .
A design with A l = 0, A 2 = 11, r = 20, and k = 12 would satisfy this requirement and also the requirement that b be an integer. However, 0 = 20 44 =
24, and so no such design exists.
For symmetric designs, we note that the determinant
!NN I = r 2(r l v2) m -1 (r 1) m(7-1)
'
must be a perfect square. If neither 0 nor 0 2 is zero, the following conditions are
necessary:
(i) If m is odd and n is even, then r ,1 1 must be a square;
(ii) If m is even, then r 2 v )t2 is a square.
Sec. 14.5]
Semiregular Designs 293
Thus, for example, condition (i) rules out the design for v = b = 22, r = k = 5,
m = 11, n = 2, t = 0, and A 2 = 1; condition (ii) rules out v = b = 44,
r = k = 7, m = 22, n = 2, A, = 0, and A 2 = 1. Bose and Connor (1952) have
extended these results, using the HasseMinkowski invariant. Bose and Connor
(1952) divided the GD designs into three classes which are mutually exclusive:
(i) Singular designs, which are designs for which 0 2 = 0, i.e., r
(ii) Semiregular designs which have 0 1 = 0, 6 2 > 0;
(iii) Regular designs which have 9 1 > 0, 0 2 > 0.
= al ;
If both 0 and 0 2 are zero, we have the complete block design with b = r and
k = v.
14.4 Singular Designs

For a design with r = A l , the varieties have to occur by groups. Whenever a
given variety appears in a block, all varieties in the same group must also appear
in that block. The simplest example is the design with b = m, k = n, and r = 1,
in which each block consists of a single group of varieties. The E'S(1) design
repeats this design n timer. The ES(2) design is also singular.
Singular designs with t 2 > 0 are obtained from balanced incomplete block
designs. If we take a balanced design with parameters v* = m, b*, r*, k*, and A*,
and replace each treatment by a group of n varieties, we obtain a singular GD
design with v = mn, b = b*, r = d l = r *, k = nk*, and A 2 = A*. Indeed, every
singular GD design with A l = 0 can be obtained by this method. It is possible
to have b < v, as we saw in the preceding paragraph, because the rank of NN' is
1 + a i = m. The inequality b >_ v, which holds for balanced incomplete block
designs, is thus replaced by b >_ m.
Zelen (1954) points out that this method of replacing each variety by a group
of n new varieties can be used to derive a PBIB(m + 1) design from any
PBIB(m) design, and P. M. Roy (1953) has incorporated this idea into his
hierarchical GD designs.
14.5 Semiregular Designs

Bose and Connor (1952) prove the following theorem.
Theorem. A necessary and sufficient condition for a group divisible design to be semiregular is that k = cm where c is an integer, and that each block contain exactly c
varieties from each of the groups.
PROOF.
Al.
Suppose that we have a GD design with parameters v = mn, r, k, ,1 1 , and
[Ch. 14
Sufficiency: Consider a variety in the jth group. It appears r times, and on each
occasion there are, in the same block, c varieties from the hth group, (h 0 j).
Then
rc = Zn.
We also have k = cm so that
rk=rcm=mn
=A2 v,
and the design is semiregular.
Necessity: Suppose that the design is semiregular and that the jth block conef = nr.
tains ef varieties belonging to the first group. Then
The jth block contains (2) pairs of varieties from the first group. Each of
these pairs of varieties occurs together in A l blocks, and so
6
êj (ef -1)=Al n(n-1).

It follows that
1 ei = n(n 1)i\ + nr = n(rk n(m 1)A ) = n A .

2
2 2
Let = nr/b = vr/(mb) = k/m = c. Then b (e, ) 2 = n 2b 2 n 2 r 2 = 0,

and e l = e 2 = = e m = = c. But ej is an integer and so c is an integer.
The same argument applies for each group of varieties, and so each block
contains exactly c varieties from each group. u
It is easily shown that for a semiregular GD design b >_ v m + 1. From
any semiregular design with v = mn treatments, a design with parameters
v* = m i n, b, r, k* = cml, A1, A2 may be obtained by cutting out (m m l ) of
the groups of treatments.
When Bose (1942c) established the inequality
b>v+r-1
for resolvable balanced incomplete block designs, he called affine resolvable
those designs for which the equality b = v + r 1 held. He showed that for an
affine resolvable design with v = nk, each block in a given replication has exactly
k/n varieties in common with each of the blocks in every other replication; kin is
an integer.
It follows that the dual of an affine resolvable balanced incomplete block
design is a semiregular GD design. When we take the dual, blocks that were in
the same replication become varieties in the same group, and, since the blocks in
any replication are disjoint, we have A l = 0. Since blocks in different replications
Regular Designs 295
Sec. 14.6]
become varieties in different groups, it follows from the property of affine resolvable designs just mentioned that A 2 = kin. We summarize this in the following
statement. Let D be an affine resolvable balanced incomplete block design with
parameters
r* = m,
v* = nk*,
k*,
A*;
then the dual of D is a semiregular GD design with
v = mn,
a1
r = k*,
k = m (c = 1),
A 2 =kin.
=0,
The duals of the designs in Yates' orthogonal series OS1 are semiregular
designs with parameters
v=sa+s,
b=s 2 ,
m=s+1,
n=s,
r=s,
k=s+1,
A=0, A 2 =1.
If we take only the first p replications, we have a design with
v=ps,
b=s 2 ,
k=p,
m=p,
n=s,
A l =0,
A.=1.
14.6 Regular Designs

For regular designs 0 > 0 and 0 2 > 0; NN' is of full rank, and b >_ v.
Several methods of obtaining regular designs have been given by Bose et al.
(1953), and by Freeman (1957). We will present some of these methods.
(a) Section from a Balanced Design
From a balanced design with parameters v*, r*, k*, and ,1 = 1, we may delete
all the blocks containing a particular variety. The resulting design will be a GD
design with v=v* 1, m=r*, n=k* 1, r=r* 1, k=k*, X1 1 =0,
a 2 = 1. The groups of varieties correspond to the discarded blocks. The latent
roots of NN' are 0 1 = r* k* and 02 = r* 1. If the balanced design is
symmetric, we have 9 1 = 0, and the design is semiregular.
(b) Complete and Incomplete Groups
These are designs in which each block consists of u complete groups of
varieties together with h varieties from another group. An example is the E'S(2)
design, which has u = m 1, h = 1. The ES(1) design has u = 0, h = n 1.
There are several methods of obtaining designs of this type. In an earlier section
we mentioned unconnected designs in which a balanced incomplete block design
with v* = n, b*, k, A = a l is taken for each group. The combined design has
b = mb*, A, = A, ,l 2 = 0. Its complement is a design of this type with h = n k
and u =m-1.
[Ch. 14
Other designs may be obtained by using two balanced incomplete block

designs. In this context, either or both of them may be degenerate designs with
r = k = 1, A = 0. Call the designs D l , D 2 ; D l is a design for n treatments; D 2
is a design for (m 1) treatments. We begin by taking the design D l and using
the varieties in the first group as the treatments. We then substitute for the
m 1 treatments in D a the other m 1 complete groups, and combine each
block of D l with each block of D 2 . The procedure is then repeated using the
varieties of the second group for D l and so on. If the parameters of D{ are
bi , k,, r1 , A*, the overall design has
b = mb 1 b2 ,
r = r 1 b 2 + (m 1)r2b 1 ,
A l = ( m 1)b 1 r2
+ b 2 Ai,
k = k l + nk 2 ,
A2 = ( m 2)b 1 A 2
+ 2r1 r2 .
Suppose, for example, that there are twelve varieties divided into three groups
offour:A,B,C,D; E, F, G, H; I, J, K, L.
EXAMPLE 1. The E'S(2) design is
CEFGHIJKL,
AEFGHIJKL, BEFGHIJKL,
DEFGHIJKL,
ABCDEIJKL, ABCDFIJKL,
ABCDGIJKL, ABCDHIJKL,
ABCDEFGHI, ABCDEFGHJ, ABCDEFGHK, ABCDEFGHL.
r =k =9,
v =b =12,
A l =8,
,12 =6.
EXAMPLE 2. Take for the incomplete group blocks from the balanced design:
12, 13, 14, 23, 24, 34. Augment each block with the varieties in the remaining
two groups. This gives
AB EFGH IJKL, AC EFGH IJKL, AD EFGH IJKL,
BC EFGH IJKL, BD EFGH IJKL, CD EFGH IJKL,
ABCD EF IJKL, ABCD EG IJKL, ABCD FG IJKL, and so on.
b = 18,
k = 10,
r = 15,
A l = 13,
a2 = 12.
EXAMPLE 3. Taking u = h = 1 gives a design with b = 24, k = 5, r = 10,

d l = 8, and ' 2 = 2. (Each of the two balanced designs used is degenerate with
r = k = 1 and A = 0.)
A IJKL,
E ABCD,
I EFGH,
A EFGH,
E IJKL,
I ABCD,
B IJKL,
F ABCD,
J EFGH,
B EFGH,
F IJKL,
J ABCD,
C IJKL,
G ABCD,
K EFGH,
C EFGH,
G IJKL,
K ABCD,
D IJKL,
H ABCD,
L EFGH,
D EFGH,
H IJKL,
L ABCD.
It is not, however, always necessary to take the complete design. In Example 3,

either the first twelve blocks or the last twelve blocks constitute a half design
Sec. 14.6]
Regular Designs 297
which is partially balanced. Indeed, whenever m is odd such half designs are
available. They are obtained by taking, for example, those blocks in which the
hth group is complete and the (h + 1)th group is incomplete, and omitting the
corresponding blocks in which the (h + 1)th group is complete and the hth
group is incomplete (h + 1 being reduced mod m).
In analogous fashion, designs may be obtained in which each block consists
of a subset from each of two or more groups, the subsets again being chosen as
blocks of balanced designs with v = m. Thus, with six varieties in two groups
A, B, C; D, E, F, we may take all possible pairs from the first group with each
member of the second group, and vice versa, to obtain a design with b = 18.
k = 3, A 1 = 3, and A2 =4:
ABD,
BCD,
ADF,
ABE,
BCE,
BDF,
ABF,
BCF,
CDF,
ACD,
ADE,
AEF,
ACE,
BDE,
BEF,
ACF,
CDE,
CEF.
Again, fractional designs exist. The complete design for m = n = 3, with each
block containing two treatments from one group and one from another, calls
for fifty-four blocks. There is a partially balanced half design with v = 9,
b = 27, k = 3, r = 9, A l = 3, and A 2 = 2. For m = n = 4, with each block
containing two varieties from one group and one from another, 288 blocks are
needed for the complete design, but there is a one-sixth fraction with v = 16,
b = 48,k = 3,r = 9, A l = 2,andA 2 = 1.
Freeman (1957) also discusses variations of these methods in which v = 2n 2
varieties are split into m = 2n groups, and the groups are divided into two sets
of n groups each. Designs can then be found with k = 2n, b = n(3n 1),
r = ( 3n 1), A l = 2n 1, and A 2 = 2. It is necessary that there exist (n 1)
orthogonal Latin squares of side n. There are two types of block in the design.
The first n(n 1) blocks consist of all possible pairs of groups from the same
set. The remaining 2n 2 blocks each consist of a group from one set and a single
variety from each group of the other set; each group appears complete in n of
these blocks, and the assignment of the varieties from the other set uses the
orthogonal Latin squares.
The design for n = 2 has v = 8, b = 10,k=4,r=5,1 1 =3, and ,1 2 =2.
If the groups are A, B; C, D; E, F; G, H, we have
ABCD, EFGH, ABEG, ABFH, CDEH, CDFG, ACEF, BDEF,

ADGH, BCGH.
Freeman gives a design for n = 3.
(c) Adding Designs
Suppose that there exists a balanced design with parameters v = mn, k =
n, b *, r*, A*. A disconnected GD design with m blocks of size k can be obtained
by letting each of the m blocks consist of a different group of n varieties. The
[Ch. 14
design obtained by repeating the balanced design s times and then repeating the
disconnected design t times is a regular GD design with parameters
v=mn,
b=b*s+mt,
Al = A * s + t,
k=n,
2 =
r=r*s+t,
*s.
Adesignforv=9,m=3,n=3,b=27,k=3,a 1 =3,andA2 =2has

just been mentioned; each block contained two varieties from one group and
one from another. Bose, Clatworthy, and Shrikhande (1954) give a different
(nonisomorphic) design with the same parameters. They repeat the orthogonal
series design with v = k 2 = 9 and b = 12 twice, and then repeat one of the
squares a third time, which amounts to taking s = 2 and t = 1.
In the previous example of adding a disconnected design to a balanced design
it happend that the balanced design was resolvable and that the disconnected
design comprised one of the replications in it. Bose et al. (1953) use a method of
repeating a resolvable balanced incomplete block design t times and further
repeating one of the replicates an additional a times where a may be positive
(addition) or negative (subtraction).
Other designs may be obtained by adding partially balanced designs to
balanced designs. We have mentioned Freeman's design for v = 16, m = n = 4,
b = 48, k = 3, A l = 2, and A 2 = 1. A nonisomorphic design is found in the
Bose, Clatworthy, and Shrikhande tables, design R38 (1954). It may be constructed from other designs in the following way. We first take for each group
the BIBD for v = b = 4, r = k = 3, and A = 2. These sixteen blocks are a
design with A l = 2 and X1 2 = 0, and we add to them design R35 of Bose et al.
(1954), which has v = 16, m = n = 4, k = 3, A l = 0, and ' 2 = 1. Other
examples of this method are given by Bose et al. (1953).
This method of adding designs is not confined in its application to group
divisible schemes. If in the general case we can find designs D of block size k
with the property that A, = 1 and A h = 0 for all h 0 i, we can easily construct a
design of block size k for any set of parameters A l , A2 , ... by repeating each
basic design Di exactly A, times. An example will be found in the section on
Latin square designs.
{
(d) Extension of Designs

It will be recalled that the complete lattice design with v = k 2 = 9, b = 12,
r = 4, and A = 1 is resolvable and can be made into a balanced design for
v = b = 13 and r = k = 4 in the following way. A new block is added containing the four new variaties, a plot with the first new variety is added to each
block in the first replicate, the second new variety is added to each block in the
second replicate, and so on. This method of adding r additional varieties to make
a design with block size k + 1 has its parallel in the construction of GD designs.
Suppose that there exists a resolvable GD design with parameters v = mn = ks,
b = rs, r, k, A l , A2 = 1, and also that there exists another GD design with
Sec. 14.6]
Regular Designs 299
parameters v' = r = m'n, b', r' = r s, k' = k + 1, A, = A l , and X = 1

where m' is an integer; then we may allocate one new variety to each replicate
in the first design, adding it to each block in the replicate, and obtain a new GD
design with parameters
v=
rn
v+v'=v +r,
=m+m
'
b=b+b',
n = n,
P=r,
nl = A1,
kC=k+1,
"2 =
1.
There is a paucity of examples of moderate size. The following one is found in

the paper by Bose et al. (1953).
EXAMPLE. A resolvable GD design for v = 6 and k = 2 is obtained by taking
the unreduced design with b = () = 15 and repeating one of the replicates.
This gives a design with v = 6, b = 18, r = 6, k = 2, m = 3, n = 2, A l = 2, and
A 2 = 1:
AB, CD, EF; AB, CD, EF; AC, BE, DF; AD, BF, CE; AE, BD, CF;
AF, BC, DE.
A design for v'= r=6,k'=k+1=3,m=3,n =2, A 1 = 2, and A 2 =1is
UWX, VWX, WYZ, XYZ, UVY, UVZ.

The combined design with
.1 1 = 2, and ,C 2 = 1 is
v = 12, b = 24, r" = 6, k = 3, m = 6, n = 2,
ABU, CDU, EFU, ABV, CDV, EFV, ACW, BEW,

DFW, ADX, BFX, CEX, AEY, BDY, CFY, AFZ,
BCZ, DEZ, UWX, VWX, WYZ, XYZ, UVY, UVZ.
(e) Difference Sets
We quote without proof the following theorem of Bose et al. (1953), which is
similar to Bose's fundamental theorems for balanced incomplete block designs.
Theorem. Let M be a module with m elements and let n treatments correspond to
each element. Suppose that we can find a set of t initial blocks B i , B 2 ., Bt each
of size k, and an initial group G of n elements, such that
,..
(i) The n(n 1) differences (reduced mod m) arisingErom G are all different,
and
(ii) Among the k(k 1)t differences occurring in the initial blocks, each
differente occurs t 2 times, except those that arise in G, which each occur ,1 1
times.
[Ch. 14
Then a GD design is obtained by cycling (mod m) from the initial blocks B 1

B2 , ... , B.
EXAMPLE. v = b = 14, m = 7, n = 2, A l = 0, and A2 = 1. There are two

initial blocks (1 1 , 2 1 , 4 1 , 0 2 ), (1 2 , 2 2 , 42 , 0 1 ). We denote the jth variety in the ith
group by i , 0 < i <_ m - 1, 0 < j < n - 1. Arithmetic is mod 7. The initial
group is 0 1 , 0 2 . In the differences [1, 1] each value 1, 2, ..., 6 occurs once; the
same is true for the differences [2, 2]. In the differences [1, 2] and again in the
differences [2, 1], each value 0, 1, 2, ... , 6 occurs once, except for zero, which
does not appear. In the initial group there are no differences [1, 1] or [2, 2]; the
only difference [1, 2] or [2, 1] is zero. Thus, the conditions of the theorem are
satisfied and we have the following design:
;
1 1 2 1 4 1 02,
5161 1 1 42,
12 2 2 4 2 0 1,
5 2 62 12 4 1,
2 1 3 1 5 1 1 2,
6 1 0 1 2 1 52,
2 2 3 2 52 1 1,
6 2 02 22 5 1,
31 4 1 6 1 2 2,
0 1 1 1 3 1 62 ,
32 4 2 6 2 2 1,
02 1 2 3 2 6 1.
41 5 1 0 1 3 2,
4 2 5 2 02 31,
Again, it is sometimes possible to find fractional or partial designs, and the

previous theorem can be generalized.
Let M be a module with cm elements and let n jc (an integer) varieties correspond to each element of M. Suppose that we can find a set of t initial blocks of k
treatments each and an initial group G of n treatments such that
(i) The differences arising from G take n(n - 1)1c different values, each of
which is repeated c times, the values being equally spaced;
(ii) Among the kt(k - 1) differences in the initial blocks, each value occurs A 2
times, except for those occurring in G, which occur A l times each.
Then, developing the initial blocks gives a GD design with parameters
v = mn, b = met, r = kct/n, m, n, X1 1, Az.
EXAMPLE. v = b = 12, r = k = 4, m = 6, n = 2, l = 2, A 2 = 1. Initial
block 0, 1, 4, 6; initial group (0, 6); c = 2. The design is
0146,
1257,
2368,
347 9,
45810,
56911,
67100,
78111,
89 02,
91013,
101124,
11 035.
(f) Designs Based on Projective Geometries

Some designs based on finite geometries are given in the original paper by
Bose and Nair (1939). We give an illustrative example here.
In the Euclidean geometry with N = 2, consider only lines that do not pass
through the origin. Then lines are blocks, points are treatments, and we have a
Designs with the Triangular Scheme 301
Sec. 14.71
design with v = b = s 2 1, r = k = s. If points on the same line are second

associates, A2 = 1 and A l = 0. This is the design that is obtained by starting
with the orthogonal series balanced design v = k 2 = s 2 , r = k + 1, and ,\ = 1,
and cutting out all the blocks that contain a particular treatment, namely the
treatment represented by (0, 0). It is a particular case of the method (a) presented
in this section.
14.7 Designs with the Triangular Scheme

We recall that in the triangular scheme v = n(n 1)/2, n l = 2(n 2);
n a = (n 2)(n 3)/2, pi l = n 2, and pi l = 4. The varieties are arranged
above the diagonal of a square array of side n, the diagonal positions in the
array are left blank, and the varieties are repeated below the diagonal so as to
make the array symmetric. Two varieties are first associates if they appear in the
same row (or column) of the array; otherwise they are second associates. The
latent roots of NN' and their multiplicities are
B1 =r+(n-4)A,(n-3)A 2 ,
9
=r-2A 1 +
A2, a 2
a1
=n 1;
= n(n 3)/2.
Connor (1958) showed that for n >_ 9 the triangular scheme is unique in the
lense that if a PBIB(2) scheme has the values just given for the parameters of the
first kind, then it must be a triangular scheme. Shrikhande (1959a) proved
uniqueness for n < 6. Hoffman (1960) showed that the scheme was unique for
n = 7 but not for n = 8. Chang (1959) also showed that the scheme was not
unique for n = 8 and later (1960) showed that for the parameters v = 28,
n l = 12, n2 = 15, pi l = 4, and so on, there are exactly three schemes possible in
addition to the triangular scheme.
Chang et al. (1965) found that there are exactly 225 combinations of values,
all integers, of n, r, k, A l , and A 2i with r < 10 and k < 10, satisfying the following necessary conditions:
b = rv/k is an integer,
2v = n(n 1),
2(n-2)A l +(n-2)(n-3)A 2 /2=r(k 1),
0<A l <r, 0<A,<r,
and, sine the complement is also a triangular design,

2rb<
2rb<
2.
For seventy-six of these combinations, designs are given by Clatworthy (1956);

some of them are also included in the tables of Bose et al. (1954). Chang et al.
[Ch. 14
provide designs for twenty-one more combinations and show that no designs
exist in a hundred and ten other cases. Eighteen remain unsolved: their parameters are given in Table 14.1. Number 7 has been solved by K. R. Aggarwal.
See Ann. Math. Statist. (1972), 43, p. 371.
TABLE 14.1
Reference
Number
1
,11
,la
15
9
10
27
3
2
10
42
1
3
3
1
3
4
21
6
7
8
9
10
11
12
30
21
8
10
28
35
1
3
10
30
4
2
10
10
21
4
5
4
1
5
13
36
63
14
10
45
7
9
63
45
1
2
5
10
9
10
99
55
99
15
16
11
55
17
12
18
66
There are four methods of obtaining triangular designs in addition to the

general methods for all schemes which were mentioned before.
(i) The dual of a balanced incomplete block design with parameters
v*=n,
b*=n(n-1)J2,
r*=n-1,
A*,
k*=2,
i.e., the unreduced design for n treatments in blocks of two, is a triangular

design with
b=n,
r =2,
k=n-1,
a 1 =1,
I'2
=0.
This design is a special case of the first of the two dual methods given
by Shrikhande (1952). Each block is one of the rows of the array in the
association scheme. These are called the triangular singly linked blocked
Sec. 14.7]
Designs with the Triangular Scheme 303
designs because each pair of blocks has a single treatment in common.

For n = 5 the array is
* 1 2 3 4
1*567
2 5 * 8 9
3 6 8 * 0
4 7 9 0 *,
and this design has the five blocks 1234, 1567, 2589, 3680, 4790.
(ii) Shrikhande's second method gives the triangular doubly linked block
designs. These are the duals of balanced designs with parameters
v
1)
(fl
;=
b* = (n)
2,
r*
= n,
k*=n
-2,
,1*=2.
The triangular designs have parameters
=( n)
b=(n 2 1 ), k=n, r=n -2, A l =1,
)L 2
=2.
For n = 5 we may take the balanced design

ABC, BDE, AEF, CDF, CEF, BDF, ADE, ACD, ABF, BCE,
and, letting the first block correspond to treatment 1, the second to

treatment 2, and so on, obtain the triangular design
13789, 12690, 14580, 24678, 23570, 34569.
Conversely, a triangular design with r = n 2, A l = 1, and ' 2 = 2
has for the latent roots of NN' : 8 1 = 0, 0 2 = r = n 2, and a2 = b 1.
Hence, its dual has' only one latent root other than Bo = rk, namely
n 2 with multiplicity a 2 = v* 1. Thus the dual is a balanced design.
Raghavarao (1960b) proved that if for a triangular design B l = 0, or,
equivalently,
rk vA l = n(r A l ) 12,
which is true for the doubly linked block designs, then 2k is divisible by n
and every block of the design contains exactly 2k/n treatments from each
of the n rows of the scheme. This result is analogous to the similar result
for semiregular GD designs and the proof follows the same lines.
The complement of the design of method (ii) for n = 6 is a linked block
design of triangular type with v = 15, b = 10, r = 6, k = 9, n l = 8,
n 2 = 6, A l = 3, and ' 2 = 4; each block contains three treatments from
[Ch. 14
each row of the scheme. It is the dual of the balanced design with
parameters
v* = 10,
b* = 15,
r*
= 9 = n(n - 3)/2,
k* = 6 = (n - 2)(n - 3)/2, A* = 5 = (n - 1)(n - 4)/2.

It follows from the foregoing discussion that all triangular designs with
r = n - 2, A l = 1, and X1 2 = 2 are doubly linked block designs and that
the existence of such a design for any n implies the existence of the
corresponding balanced design. For n = 7, 8, 10 the corresponding
designs would be balanced designs with parameters
v= 15,k=5,A=2; v=21,k=6,A=2; v=36,k=8,A=2.
However, Hall and Connor (1953) showed that these three designs are
impossible. This implies that triangular designs with r = n - 2, A l = 1,
and a 2 = 2 do not exist for n = 7, 8, or 10.
(iii) Each of the rows in the scheme constitutes a set of (n - 1) varieties with
the property that any pair of theet are first associates. If there exists a
balanced incomplete block design with parameters v* = n - 1, b*, r*,
k* = k, A *, we can take such a design for each row of the scheme and
combine theet into a triangular design with
v = n(n - 1)/2, b = nb*, r = 2r*, k, A = \*, ,1 2 = 0.
For n = 5, v = 10, and k = 3, we note that each row of the scheme

has n - 1 = 4 varieties and use the balanced design with v = b = 4,
r = k = 3 to obtain the design
123,
124,
134,
234,
156,
157,
167,
567,
258,
259,
289,
589,
368,
360,
380,
680,
479,
470,
490,
790.
For k >- 4, this is the only way of obtaining a design with A, = 0.

Hence, a triangular design with 4 < k < n - 1, r, )t, and A Z = 0 exists
if, and only if, a balanced incomplete block design with v* = n - 1,
k* = k >_ 4, r* = r/2, and A* = A l exists.
(iv) The method of triads is due to Clatworthy. When k = 3, method (iii)
is not the only method of obtaining a design with A > 0 and A 2 = 0.
Another method is to obtain the blocks as follows. Take a pair of
varieties that appear in the same row. The first variety appears again in
another row. For the third member of the block take the variety that
Designs for the Latin Square Scheme 305
Sec. 14.9]
appears in that row and in the same column as the second member. Thus,
in the case n = 5 the blocks would be
125,
340,
136,
568,
147,
579,
238,
670,
249,
890.
In the general case, r = n - 2, 6b = n(n - 1)(n - 2).
14.8 The Nonexistence of Some Triangular Designs

The latent roots of NN' are not negative; thus it is necessary that
r +(n- 4);1 1 -(n- 3)1 2 >0,
r-21 1 +A 2 >_0.
This rules out, for example, designs with parameters

n = 5, v = 10, r = 9, k = 5, b = 18,1 1 = 2,A 2 =8,(6 1 =- 5);
n=5,v= 10,r=3,k=5,b=6, A 1 =2,1 2 =0,(0 2
=-
1).
For symmetric designs it is necessary that
I NNI = r20n-102(n-3)12
be a square. If neither 0 nor 0 2 is zero, it follows that
(i) If n = 4t, 0 must be a square;
(ii) If n = 4t + 1, 0 2 must be a square;
(iii) If n = 4t +2, 0 1 0 2 must be a square.
Examples of designs that are found to be nonexistent by these criteria are
(a) n = 8,v = b = 28,r = k = 10, d 1 =5,A 2 =2,(0 = 20);

(b)n= 9,v=b=36,r=k=8, A 1 = 1,1 2 =2,(8 2 =8);
(c) n= 6,v=b= 15,r=k=5, A 1 = 1,,1 2 =2,(9 1 6 2 =5).
Further such conditions involving the Hasse-Minkowski invariant are given
by Ogawa (1959).
14.9 Designs for the Latin Square Scheme

These designs are discussed by Clatworthy (1956) and (1967), and by Chang
and Liu (1964). They include as a special case the designs obtained by taking
some, but not all, of the replicates in the Yates orthogonal series designs for
[Ch. 14
v = k 2 , b = k(k + 1), r = k + 1, and \ = 1. Clatworthy (1956) gives a listing

of 111 designs most of which are L 2 designs. Chang and Liu have calculated that
there are 155 combinations of parameters that might lead to L 2 designs with
r < 10 and k <_ 10. Of these, seventy-two are included in Clatworthy's listing;
Chang and Liu give twenty-one others and show that fifty-seven of the designs
do not exist. The five unsolved cases are
v=36,
k=6,
r=9,
X1 1 =2,
A 2 =1,
v=36,
v=36,
k=6,
k=9,
r=10,
r=10,
a 1 =0,
,i 1 =3,
,^ 2 =2,
A2 =2,
v=49,
k=9,
r=9,
A=3,
A2 =1,
v = 100,
k = 10,
r = 9,
a 1 =0,
,12 =1.
Shrikhande (1959b) has shown that the L 2 scheme is unique except for n = 4.
In this case there are two schemes with v = 16, n, = 6, n 2 = 9, and pi l = 2.
The second scheme arises when, in one of the L 3 schemes, we interchange first
and second associates. Suppose that the square array and the Latin square are
1 2 3 4
5 6 7 8
9101112
13141516
ABCD
BCDA
CDAB
DABC.
Let two varieties be first associates if they do not appear in the same row or
column or with the same letter; this gives a set of parameters v = 16, n l = 6,
n 2 = 9, and pi l = 2 as in the L 2 scheme, but consider the first associates of
variety 1. They are 6, 7, 10, 12, 15, and 16. The first associates of 6 are 1, 4, 11, 12,
13, and 15, and so varieties 12 and 15 are first associates of both 1 and 6. If the
scheme is an L 2 scheme, it follows that 1, 6, 12, and 15 must appear in the same
column or row of a square array and, hence, 12 and 15 are first associates.
Twelve and 15 both go with the letter B, and so they are second associates, in
contradiction. This scheme was first mentioned by Mesner (1956) and (1967).
In the same paper Shrikhande gives conditions for the nonexistence of symmetric L 2 designs based on the Hasse-Minkowski invariant.
The latent roots of NN' and their multiplicities are
0 1 =r-(i-n)(A l -,la)-,12 ,
a 1 =i(n- 1),
02=r - i(Al - A2) - '12, a2=(n-1)(n-i+ 1).

For the L a case, a l is always even; a 2 is odd when n is odd, and then 0 2 = r 2a 1 + a 2 must be a square if the design is symmetric and B l > 0.
Raghavarao (1960b) gives a result for L 2 designs similar to those just
mentioned for semiregular GD designs and triangular designs. If rk - v21 1 =

n(r - A ' ), then k is divisible by n and every block of the design contains kin
varieties from each row of the association scheme. It should be noted, however,
that this condition is not equivalent either to 0 = 0 or to 0 2 = 0.
Designs for the Latin Square Scheme 307
Sec. 14.9]
Clatworthy (1967) has shown that most of the known L, designs belong to
four families. We shall discuss each of the families in turn. Our notation differs
from that of Clatworthy in that he talks of v = s 2 varieties while we have
v = n 2 . We shall alter the definitions of families A and D slightly in order to
include more designs. Clatworthy considers numerous special cases in each
family; the reader is referred to his paper for these.
Family A: These are designs with k < n and a 2 = 0. Let D be a balanced

incomplete block design with parameters v* = n, b*, k* = k, r*, d* = A l . The
first b* blocks of our design consist of the blocks of D using the varieties in the
first row of the square array of the association scheme; D is then repeated using
the varieties in the second row, and so on. After all the rows have been used, we
repeat the procedure using the columns of varieties. If i > 2, the procedure is
then repeated again for each of the i 2 orthogonal squares. We then have a
design with
b = inb*, r = ir*, k,
i\,
A 2 = 0.
If the balanced design D is the unreduced design consisting of all
(k) subsets of
k out of the n varieties, we have

n-2
b=ni(n)
k' Al=(k-2)' A2=0.
If k = n, the design consists of ni blocks forming the first i replicates of the
balanced lattice, if that design exists, with A l = 1 and A 2 = 0. We shall use this
design later.
Family B: These are designs with k = 2n. The first n(n 1)/2 blocks consist of
all possible pairs of rows from the square array, the next n(n 1)/2 blocks
consist of pairs of columns, and so on. For the L i design we have
b=in(n 1)/2,r=i(n 1),k=2n,
=n+i-2,A 2 =i.
Family C. These designs have k = n, and they exist whenever n is a prime, or a

power of a prime. We have already mentioned the design in family A with
k = n, A l = 1, and A 2 = 0. If we proceed in the same way to set up a second
design using the remaining Latin squares of the complete orthogonal set, we
shall have a second design with b = n(n i + 1), k = n, A l = 0, and A 2 = 1.
If we wish to construct a design with parameters a l , , we have now merely to
repeat the first design d l times and the second design X 2 times. The reader wilt
appreciate that if there exists a balanced design with A = 1 for a value of k <_ s,
we may use a similar. procedure.
[Ch. 14
Family D: This family consists of the ES(i) and E'S(i) designs. The parameters
for these symmetric designs are
ES(1): k = i(n 1),
,1 1 =i 2 -3i+n,
d 2 =i(i 1),
ES(2): k=(n-1)(ni+ l),
) 1 =(ni)(ni+ l),
X12 =(ni) 2 +i-2,
E'S(l): k=i(n-1)+1,
a 1 =i 2 -3i+n+2,
A 2 =i(i 1),
E'S(2): k=(n-1)(ni+1)+1,
A 1 =(ni)(ni+1),
A 2 =(ni) 2 +i.
Exercises
1. Prove that if A and B are square matrices such that AB and BA both exist, then
AB and BA both have the same nonzero latent roots.
2. Derive the P matrices for Shrikhande's two linked block designs.
3. Let U be a square matrix of order mn partitioned into m 2 square matrices U, of
order n where U 5 = al,, + bJ n if i = j, and cIn + dJn if i ^ j. Recalling that
laI. + bJ nl = an-1 (a + nb), show that
U = (a C)(n-1)(m-1)(a + nb
nd)m '(a + (m
-
1)c)n
x {a + nb + (m 1)(c + nd)}.
4. (Continuation). Order the varieties of a GD scheme so that the ith group

consists of varieties (i 1)n + 1, ..., in. Then evaluate directly the determinant
INNI
5. (Continuation). Show that if the varieties in an L 2 scheme are suitably ordered

INN'I = rk(r + (n 2),1 1 (n 1)) 2) 2( n -1) (r 2A 1 + ,12)(n -1 ) 2 .
Hence find the latent roots of NN' and their multiplicities.

6. Prove that if a triangular design has 0 1 = 0 then 2k is divisible by n and each
block contains exactly 2k/n treatments from each row of the scheme [Raghavarao
(1960b)].
7. Prove that if, in a design with the L 2 scheme,

rkva l =n(r ,1 1 ),
then k is divisible by n and every block of the design contains kin treatments
from each row of the scheme [Raghavarao (1960b)].
8. Show that for any L 2 design the following inequalities hold for b. Let
a = min(n 1, [r/1 ]), p = min(n 1, [r/,12 ]), where [x] is the integer part of x.
Then [Chang and Liu (1964)]
2b > (a + 1)(2r aa l ),
2b > (P + 1)(2r
Exercises 309
9. Consider a PBIB(2) association scheme A in which there exists a set of u,
treatments that are all i th associates of each other. Let
wt = min(u i 1, [r/,1,]),
wi = min(u 1
1, [(b r)/(b 2r + a{)]).
Show that for any design having associate scheme A
2b > (w + 1)(2r w) i ),
{
2b > (wi + 1)(2b 2r w(b 2r + ai)).
Adapt this result to obtain a nonexistence theorem for triangular designs

[Chang, et al. (1965)].
10. Show that for any PBIB(2) design which is not of the GD type [Chang et al.
(1965)],
max(2,1, A 2 , 2) 2 A l ) < r.
CHAPTER
15
Additional
Topics in
Partially Balanced
Designs
In this chapter we shall continue our investigation of partially balanced

designs, with emphasis on designs having more than two associate classes. We
begin, however, with a short introduction to the topic of partial geometries and
their application to PBIB (2) designs.
We have already emphasized the usefulness of the latent roots of NN' in
determining the nonexistence of some designs. We shall discuss the subject of the
latent roots of NN' and their multiplicities in PBIB(m) schemes at some length.
The discussion of that topic will conclude with a few words about cyclic and
pseudocyclic schemes with two associate classes. In the succeeding section, we
shall introduce several PBIB(3) schemes. The chapter will conclude with a
discussion of cyclic designs.
15.1 Geometric Designs

In the previous chapter we dealt with designs having the three principal types
of schemes with two associate classes: group divisible, triangular, and Latin
square. There was only one reference to designs obtained from finite geometries,
and that example duplicated a design given earlier. Finite geometries do, however, provide a series of simple partially balanced designs, i.e., designs having
either A l = 1, .2 = 0, or A l = 0, A 2 = 1, which do not belong to the three
principal types.
Bose and Nair (1939) presented several examples of designs obtained from
Euclidean or projective geometries. If we consider the set of all lines in a
Euclidean geometry with sN points and omit all the lines that pass through the
origin 0, we can let points denote varieties and lines denote blocks. The remaining
310
Geometric Designs 311
Sec. 15.1]
points and lines form a GD design for v = sN 1 varieties in blocks of size

k = s with A l = 0 and A 2 = 1. Two varieties are first associates if the line
joining the points passes through the origin. Then n l = s 1.
Other GD designs are obtained if we consider m-dimensional subspaces
(m > 1) and again omit those subspaces that contain the origin O. Bose and
Nair also considered omitting from a projective geometry, PG(2, s = p"), all the
points on three lines that are not concurrent, and all the lines passing through the
vertices of the triangle formed by those three lines.That procedure led them to
some symmetric L 3 designs with v = (s 1) 2 , k = (s 2), X1 2 = 1, and A l = 0.
More interesting, perhaps, are the designs obtained from partial geometries
because every partial geometry is isomorphic to a simple PBIB(2) design. A
partial geometry is a collection of v points (varieties) and b lines (blocks) with
the following axioms:
(i) There is only one line through any pair of points;
(ii) There are exactly k points on each line;
(iii) There are exactly r lines through each point;
(iv) If a point P does not lie on a line p, then there are exactly t lines through P
that intersect p.
This system constitutes a partial geometry with parameters (r, k, t). It follows
that
tv=k[(r 1)(k 1) + t],tb = r[(r 1)(k 1) + t].
To derive the latter relationship, we note that there are r lines passing through
P, each with k 1 points other than P. Thus, there are r(k 1) points connected to P. On the other hand, there are b r lines that do not pass through P.
Each of these b r lines contains t points that are connected to P; each of the
points is on r 1 of the b r lines; thus, there are t(b r)/(r 1) points
connected to P. Hence,
r(k 1) = t(b r)/(r 1) and b = r[(r 1)(k 1) + t]/t.
If we now let points denote varieties and lines denote blocks, we have a simple
partially balanced design. Two varieties are first associates if their points are
connected in the partial geometry; otherwise, they are second associates. We
have n l = r(k 1), n 2 = (r 1)(k 1)(k t)/t. To obtain pi l we consider
two first associates P and Q. There are k 2 other points on PQ. There are
r 1 lines through Q other than PQ; each of these lines contains t first associates
of P. Thus,
P1,=(t-1)(r-1)+k-2, P12 = (r 1)(k t),
Pij=rt,
P12= r(k t-1).
312 Additional Topics in Partially Balanced Designs
[Ch. 15
Bose and Clatworthy (1955) show that every PBIB(2) design with A l = 1,
sla = 0, and b < v is a partial geometry. It follows that the singly linked block
designs are partial geometries. These are one of the series of linked block
designs obtained by Shrikhande (1952), and are duals of the balanced incomplete block designs with v* = rk k + 1, r* = k, k* = r, and A* = 1. They
were mentioned at the beginning of Chapter 14. When r = 2, the singly linked
block design is a triangular design. Thus the triangular association scheme can
be obtained from a partial geometry configuration.
For a further discussion of this topic the reader is referred to the review paper
by Bose (1963) and the two monographs by Vajda (1967a,b). Further research in
this area is being carried out by students of graph theory. The representation of
association schemes as concordant graphs is mentioned by Bose and Mesner
(1959).
Mesner (1967) has extended the Li scheme to include cases where n and i are
both negative, although all the design parameters are positive. This gives the NL
family of schemes (negative Latin square). The NL(n) scheme has v = n 2 , and
the other parameters are obtained by putting n for n and i for i in the
corresponding formulae for the L i scheme. For example, the NL 3 (8) scheme has
v=64, n 1 =-3(-8-1)=27, n 2 =(-8+3+1)(-8-1)=36. The P
matrices are
P1 = [16 20,'

P2 = [1520]
Numerous designs with NL schemes can be obtained from finite geometries,

and Mesner gives some examples in the reference cited.
15.2 Latent Roots of the Intrablock and Interblock Matrices

The formula for the intrablock efficiency of an incomplete block design
obtained at the end of Chapter 11 involves the nonzero latent roots of A. We
have seen that for a linked block design A has a single nonzero latent root with
multiplicity b 1. We now investigate further the latent roots of A and NN' for
PBIB designs.
In solving the intrablock equations for the case of two associate classes
(Chapter 12) we obtained (12.2). Adding the quantity Ji= 1 )i l n l fi gives
kSl(Qh) = S l (Th){r(k 1) A1Pii A2Pi2 + it l n l }
+S2 (^ h){ a1P1
'2Pi2 + A l n l }.
In the general case of m associate classes

Si(Qh) = i 9'iSJ(fh),
j=1
where kpi! = r(k 1 )Si, ^ k=1 AkPik + Aini, and Bij is the Kronecker delta.
Sec. 15.2]
Latent Roots of the Intrablock and Interblock Matrices 313
Let C h be a matrix of m rows and v columns defined as follows: c i; = 1 if the

jth variety is an ith associate of the hth variety; otherwise c l; = 0. Then C h Q is
the vector (Si (Q h )) and C h-r is the vector (Si (f h )). Thus, writing < _ (W) we
have
C,Q = 4'C, zT.
If <b is not singular (i.e., if the design is connected),

C,T ='D -1 C,Qi
but
Th = -
Si(Th)
_ -1 ChT,
'
t=1
hence,
fh
1'b
1 C h Q
We now drop the subscript h, noting that for any choice of h,

^C = CA + Dl',
where kD' = (A 1 n 1 , A2 n 2 , i\ an 3 , ... ).
A has a latent root 0 o = 0, with the corresponding latent vector 1; if the
design is not connected, the zero latent root will be repeated.
We now show that the matrices A and 4 have the same nonzero roots.
Let 0 be a nonzero latent root of A and x a corresponding vector such that
x'1 = 0. Then
Ax = Ox and CAx = OCx,
but
CAx = (CA + D1')x = ^Cx,
and so 0 is a latent root of <D with vector Cx.
Conversely, let 0 be a nonzero root of ' with corresponding vector x. Then
x' I = Ox',
and
x'^C = Ox'C.
Then
x'CA + x'Dl' = Ox'C,
x'(C 0 'Dl')A = Ox'(C 0 -1 D1'),
and 0 is a latent root of A with [C 0 - 'D1]'x as vector.
[Ch. 15
Let II* = (p 1) where

m
= r 8i1 + 1 AkPtik niAi,

k=1
then
k> = krI II*.
Since kA = rkI NN', it follows that if B i is a latent root of NN' with

multiplicity a, then 0i = (rk 8)/k is a root of A with the same multiplicity a,.
It also follows, from the similar relationship between 4; and II*, that because the
nonzero roots of c are the nonzero roots of A, the roots of NN' are, with the
exception of rk, the same as the roots of II*. This result was first shown by
Connor and Clatworthy (1954). Bose and Mesner (1959) derived it by considering
the algebra of the association matrices; in their paper they also proved that the
multiplicities a i are independent of r, Al, A2, ..., Am. Since NN' is nonnegative
definite, none of its roots are negative.
The ith association matrix Bi is defined as follows: the element Pab in the ath
row and bth column, 0 <_ a <_ m, 0 5 b <_ m, is 1 if a and b are ith associates
and zero otherwise.
The equation 1 1I* = 0 is a polynomial of the mth degree in r. If the determinant vanishes when r takes the values s 1 , s2i . .., sm , the corresponding latent
roots of fl*, and hence of NN', are B, = r si . The zeros s,, are linear functions
of the Ai, and so we may write
9=rl+Z*?,
where 0' = ( 0 k, 02, ... , Om), X' _ (A1, A2, ... , a m).
Adding the root 0 = rk, we have
tr (NN') = rv = rk +
s=1
or, writing a' = ( al, 2, ... , am), n ' = (n1, %1 2, ... , nm),
r(vk)=a'@=ra'l+a'Z*1^.
Then, since 1 = v 1,
a'Z*a=r(k-1)=n'X,
which is, since the a i are independent of the ,\, an identity in the A t . This leads to
m equations to be solved for the m unknowns a l , ..., a m , (Z*)'a = n.
For two associate classes det 11* = 0 is a quadratic in 0. Connor and Clatworthy (1954) have shown that after some simplification it may be written
(r 0) 2 +
KAl "2)(Pi2 P112) (Al + A2)](r 0)

+ [(A1
'2)(Q2P112 A1P212) + A1A2] = 0,
Sec. 15.3]
Cyclic Association Schemes 315
whence, putting y = Pi2 Pi2,

follows that
= P12 + P12, and 0 = y 2 + 2P + 1, it
0 1= r++(/ D+Y 1 )Al f(1/O+y+ 1)A2,

92= r
j(4/ 1 y+ 1 ) )ti+ (1/ 0 y 1)a 2
Furthermore,
nl + n 2 (n l n 2 ) + Y(n1 + n2)
2/
a= 2
and
+ n 2 (n l n 2 ) + y(nl + n2)
a= nl
2 +
2\/Q
It is necessary that these multiplicities be positive integers. This imposes some
restrictions upon the parameters for the existence of designs, especially if A is
not a perfect square.
We calculate now the latent roots and their multiplicities for a group divisible
design. In this case
IH *I = Ir A l (A2
0
1)
r + l(n 1) Zn2
whence 0 1 =r+) 1 (n-1),1 2 n=rkvA 2 and 0 2 =r

y= P= n-1andA= n 2 .Hence,
=m-1,
X1 1 .
We have
a2 =m(n-1).
Bose and Connor (1952) obtained this result directly by evaluating NN'I _
rk(rk v1 2 )m -1 (r ,\ ) m(n -1 Alternatively,
).
Z* n-1
_ _ n (Z*)-i =
n
0]
n
(Z *)_1 = i 0
n [1 n 1]
[(n 1), n(m 1)]
1J
= [(m 1), m(n 1)].
15.3 Cyclic Association Schemes

We have already noted that the requirements that the latent roots Bt be nonnegative and that for a symmetric design the determinant NN'1 be a square
enable us to establish that designs with certain sets of parameters do not exist.
In all the standard association schemes for PBIB(m) designs, the formulae for a,
automatically give only integer values, and so the requirement that a, be an
integer is not very helpful. It is of interest, however, in the schemes for two
associate classes when we consider the results of Connor and Clatworthy that
have just been obtained. If 0 is not a square, VA is irrational and the term
[Ch. 15
which has -,/A in the denominator must vanish. Since y is either zero or a
positive integer, it follows that if -/A is irrational we must have n l = n 2 and
y = 0, which leads to the following result [Mesner (1965)].
If, in a PBIB(2) scheme, -/A is irrational, there is an integer t, such that
v=11 =4t+1, p12=pi2=t, n 1 =n 2 =2 = a2 =2t.
Mesner also notes that the only known cyclic association schemes as defined by
Bose and Shimamoto (1952) have parameters that satisfy these conditions. He
therefore calls any PBIB(2) scheme with %/0 irrational pseudocyclic, and proven
that a two-class association scheme with v prime must be of this type. It follows
that there are no two-class association schemes, and thus no designs, for which v
is a prime of the form 4t + 3.
We mention in passing that the only cyclic association schemes listed by Bose,
Clatworthy, and Shrikhande (1954) have v = 13, 17, 29, and 37. The varieties
can be represented by integers 1, ..., v 1, and 0. For v = 13 and 17, their
scheme gives as the first associates of the vth variety the set of nonquadratic
residues of v. For v = 29 and 37, they have the quadratic residues of v. In all
cases, either set of residues will suffice.
Mesner (1967) also uses the term pseudo-Latin square to denote any scheme
that has the same parameters as an L(n) scheme (the NL schemes are excluded).
He points out that there are such schemes which are not L, schemes, and gives as
an example the complement of an L3 (6) design. This has the same parameters as
an L 4(6) scheme would. However, no L 4 (6) scheme exists because there is no
Graeco-Latin square of side six.
15.4 Partially Balanced Association Schemes with Three

Associate Classes
Any PBIB(2) design can be made into a PBIB(3) design with v' = sv by
replacing each treatment by a set of s new treatments. These designs are said to
be singular, in the sense that the matrix NN' of the new design is singular. In
this section we shall present several partially balanced association schemes with
m = 3. In most cases the authors of the papers cited have followed the procedure
of giving the P matrices and calculating the latent roots of NN' and their
multiplicities; they have then given some examples in which designs are found to
be nonexistent because their existence would lead to a negative latent root or, in
the case of a symmetric design, to a violation of the requirement that
Bil B2O3
be a perfect square.
The intrablock and interblock equations for PBIB(3) designs can be solved by
extending Rao's method for two-associate clans designs. This is discussed in
Rao's paper (1947a) and also by Nair (1952).
Sec. 15.5]
The Hierarchic Group Divisible Scheme 317
For certain sets of parameter values, these schemes degenerate into two-class
schemes, or, more often, certain designs are two-class designs. If, in a PBIB(m)
design all the A, are equal, then the design is a balanced incomplete block design.
It is not, however, true that if in a PBIB(3) design two of the A, are equal, the
design is automatically a two-class design and that there are only two variances;
whether or not that is true depends upon the number of distinct latent roots.
We again have for each scheme the two series of elementary symmetric designs
ES(i) and E'S(i). In the ES(i) design, the jth block consists of the n i ith
associates of the jth variety; d h = p. In the E'S(i) design, the jth block also
contains the jth variety itself; then d t = p; + 2 and ,, = p 1 where h i.
;
15.5 The Hierarchic Group Divisible Scheme

This is due to P. M. Roy (1953). He considered the case of m classes; we shall
confine ourselves to the case m = 3. Suppose that v = N1 N2 N3 . We divide the
treatments into Nl groups of N2 N3 treatments each at the first stage. Treatments
in different groups are first associates; n l = (N l 1)N2 N3 . Then we divide each
group into N2 subgroups of N3 treatments each. Treatments in different subgroups of the same group are second associates; n 2 = ( N2 1)N3 . Treatments
in the same subgroup are third associates; n 3 = N3 1. Raghavarao (1960a)
considers the same scheme, but he reverses Roy's definitions. What Roy calls
first associates, Raghavarao calls third associates, and vice versa.
The P matrices are
(Ni 2)N2 N3 (N2 1)N3 N3 1
P 1 = ( N2 1)N3 0
N3 -1
0
(Ni 1)N2 N3 0
P2 =
0
0
(N2 2)N3 N3 1
N3-1
(Ni 1)N2 N 3 0
P 3 =
(N2-1)N3
N3-2
The latent roots are

B1=
r A1N2N3+\2(N2 -1 )N3+A3(N3 - 1)=rAl(n2+n3+1)

+ X 2n 2 + A3n3,
03= r A2N3+A3(N3 -1 )=r i2(n3+ 1 )+ an3,

)
03= r .13.

To find the multiplicities we solve
al
[Ch. 15
(n2 + n3 + 1)
n2
n3
0
(n3 + 1)
n3

nl
0 [al]
0 a2 = n2
1 a3
n3
and obtain
a2 = N1 (N2 1),
Nl 1,
a3 = N 1 N2 (N3 1).
Roy (1953) gives the following example of a resolvable design with parameters
v =8, b=10, r=5, k=4, N 1 =N2 =N3 =2,
n1=4, n2=2, n 3 =1, A 1 =2, A 2 =3, ,1 3 =1:

1234, 5678; 1357, 2468; 1368, 2457; 1458, 2367; 1467, 2358.
The subgroups are 12, 34; 56, 78.
If A 2 = a 3 the design is merely a GD design with Nl groups, each of which has
N2 N3 treatments. In this case, 9 2 = 0 3 = r A 3 , and there are only two distint
roots.
,
15.6 The Rectangular Scheme

Vartak (1959) discusces designs with this scheme and calls them series A.
Suppose that we arrange v = mn treatments in a rectangular array of m rows
and n columns. Treatments are first associates if they appear in the same row,
and second associates if they appear in the same column; otherwise they are
third associates. Then n l = n 1, n 2 = m 1, and n3 = (m 1)(n 1),
0
n-2
0
0
m-1
0
P 1 =
m-1 (m-1)(n-2)
0
0
n-1
0

0
m-2
0
(m-2)(n-1)
n-1
0
P2
P3 =
1
0
n-2
0
m-2
1
n-2 m-2 (m-2)(n-2)
Sec. 15.7]
The Cubic Scheme 319
0 1 =rA l +(m-1)(% 2 --fi s ),

0 2 =r +(n-1)(,1 1
B3 =rA 1
3 )-
a1= n-1=n1,
2 ,
+,1 3 ,
X1 2
2=
m-1
3=
(m-1)(n-1)
=n2,
=n3.
Vartak also proves that the scheme is unique in the sense that if a PBIB(3)
association scheme has P matrices identical with those above, then the scheme
must be a rectangular scheme.
If
m = n and also X1 1 =
we have designs with the L 2 scheme. In this case
B l = 0 2 . Note, however, that unless the array is square, putting A l = ,1 2 will not
make the two roots equal, and the number of associate classes is still three.
15.7 The Cubic Scheme

This is due to Raghavarao and Chandrasekhararao (1964). There are
v = s 3 treatments, each of which is denoted by a different set of three coordinates (x, y, z) where x, y, and z are all integers such that 1 < x < s,
1 < y < s, 1 < z < s. Two treatments are defined to be i th associates if they
have exactly (3 i) coordinates in common. For convenience we write (x, y, z)
as xyz. Thus, for s = 3, v = 27, the treatment 000 has six first associates:
001, 002, 010, 020, 100, and 200. It has eight third associates denoted by abc,
where each of the coordinates a, b, and c takes the values 1 or 2. In the general
case, n l = 3(s 1), n 2 = 3(s 1) 2 , and n 3 = ( s 1) 3 .
s-2
2(s-1)
P l = 2(s 1) 2(s 1)(s 2)
(s
0
(s 1) 2
(s 1) 2 (s 1) 2 (5 2)
2(s-2)
(s 1)
P 2 = 2(s 2) 2(s 1) + (s 2) 2 2(s 1)(s 2)
1)
2(s 1)(s 2)
0
P 3 =
(s 1)(s 2) 2
3(s-2)
6(s 2) 3(s 2) 2
3(s 2) 3(s 2) 2 (s 2) 3
3
The latent roots of NN' and their multiplicities are

B o =rk, 0 1 =r+(2s-3))) 1 +(s-1)(s-3)) 2 (s- 1) 2 A3 ,
02 =r+(s- 3)A l (2s-3)A 2 +(s 1)
ao
=1, a 1 =3(s-1) =n 1
a2
3i
03 = r-3A 1 +3A2 A l ,
=3(s-1) 2 =n 2
a3=
(s- 1)3 =n3.
[Ch. 15
The E'S(2) design for v = 64 is of particular interest. We have s = 4,

,1 1=P22=
2(s - 1)(s-2)= 12, A2=p22+2= 2(s- 1)+(s-2) 2 +2=
12, and A 3 = 6(s - 2) = 12. The design is thus a balanced incomplete block
design with v = b = 64, r = k = 28, and ^1 = 12.
15.8 The Triangular Scheme for m = 3

We defined the triangular scheme for m = 2 in two equivalent ways. The
definition involving the square array is the one most commonly used; the second
definition, in which the treatments are denoted by pairs of integers (x, y),
1 < x < y < n, may be extended to a scheme for three-associate classes. In this
scheme [John, (1966c)] we have 6v = (s + 2)(s + 3)(s + 4). Each treatment is
represented by three integers (x, y, z) where 1 < x <y < z < s + 4. Two
treatments are ith associates if there are exactly (3 - i) integers which appear in
both representations. If s = 1, there are ten treatments. The first associates of
(1, 2, 3) are (1, 2, 4), (1, 2, 5), (1, 3, 4), (1, 3, 5), (2, 3, 4), (2, 3, 5); the second
associates are (1, 4, 5), (2, 4, 5), (3, 4, 5), and the scheme reduces to a triangular
scheme for m = 2. If s = 2, treatment (1, 2, 3) has a single third associate
(4, 5, 6). In general, n l = 3(s + 1), n 2 = 3s(s + 1)/2, n 3 = (s - 1)s(s + 1)/6.
P 1 =
P z =
s+2
2s
2s
s2
s(s - 1)/2
s(s - 1)/2
s(s - 1)(s - 2)/6
2s
(s - 1)
2s
(s - 1)(s + 6)/2
(s - 1)(s - 2)
(s - 1) (s - 1)(s - 2)
P 3 =
(s - 1)(s - 2)(s - 3)/6
3(s-2)
9(s - 2)
3(s - 2)(s - 3)/2
3(s - 2) 3(s - 2)(s - 3)/2 (s - 2)(s - 3)(s - 4)/6
6 1 =r+(2s-l)\ l +-s(s -5)A 2 --s(s-1)1 3 ,

02 = r + (s - 3)A l - (2s - 3)1 2 + (s - 1)A 3 ,
63 =r-31 1 +3A 2 -2 3 ,
al=s+3,
a2 = (s + 1)(s + 4)/2,
a3=(s-1)(s+3)(s+4)/6.
This scheme should not be confused with the T(m) scheme given by Singh
and Singh (1964); they obtain PBIB(3) designs by replacing each treatment in a
design which has the triangular scheme for two associate classes, and v =
n(n - 1)/2 by a set of m new treatments. These singular designs are discussed at
lome length in their paper.
Sec. 15.9]
The LS3 Association Scheme 321
The dual of the balanced incomplete block design with v* = 6, b* = 20,

k* = 3 is a PBIB(3) design with this scheme. We have s = 2, v = 20, b = 6,
r = 3, k = 10, A l = 2, A 2 = 1, and d 3 = 0. The latent roots of NN' are
8 1 =3+31 1 -3A 2 A 3 =6,
a=5;
8 2 =3 d 1 a 2 +a 3 =0,
a
a22
93 = 3 3A 1
a 3 =5.
+ 32 2 3 = 0,
=9;
In the dual, the block that contained the old varieties x, y, and z is transformed
into the treatment (x, y, z). This method is genera]. An unreduced BIBD with
v* = n, b* = ( n ), and k = 3 has for its dual a triangular PBIB(3) design with
s = n 4,A l =2,,1 2 = 1, and A=0.
15.9 The LS3 Association Scheme

This is an extension of the two-class L 2 scheme for v = n 2 treatments [Johnson
(1967)]. The v = n 2 treatments are arranged in a square array and a Latin
square is superimposed. Two treatments are said to be first associates if they are
in the same row or in the same column of the array. They are second associates if
they correspond to the same letter in the Latin square. Otherwise they are third
associates. Then n l = 2(n 1), n 2 = (n 1), n 3 = ( n 1)(n 2).
P 1 =
n-2

1
n-2
n-2
n-2 n-2 (n-2)(n-3)

2(n-2)
n-2
2(n 2)
(n 2)(n 3)
2(n-3)
n-3
P 2 =
P 3 =
2(n-3) n-3 n 2 -6n+10

The nonzero latent roots are
0 1 =r +(n-2)A 1 A2 (n-2)A 3 ,
a 1 = 2(n-1),
62 =r-2 1 +(n 1) 2 (n-2)11 3 , a 2 =n 1,

B3 = r 2A, A, + 2A
a3 = (n 1)(n 2).
[Ch. 15
If A2 = a 3 , the design becomes an L 2 design. For n = 6, the E'S(3) design is a

balanced incomplete block design with b = v = 36, r = k = 21, and A = 12;
its complement has r = k = 15 and A = 6.
15.10 Cyclic Designs

We have seen in Chapters 13 and 14 how designs can be obtained by taking
one or more initial blocks and, in the manner of Bose, developing other blocks
by cycling. From any initial block we obtain another block by adding the same
number to each element of the initial block. It is a popular method of obtaining
balanced designs, and we have referred to its use in obtaining GD designs in
Chapter 14.
Kempthorne (1953) and David (1963) investigated cyclic designs with only
two plots per block. Kempthorne's motivation derived from breeding experiments, especially diallel crosses. Later, David and Wolock (1965) and J. A.
John (1966) discussed cyclic designs with larger blocks. All these authors
confined themselves to representing the varieties by the residue classes 0, 1, ...,
v 1 (mod v), and for the present we shall keep this restriction. These designs
are all balanced or partially balanced, but they may have considerably more than
two classes. The schemes have a cyclic structure in the sense that if x, y, and z
are integers and x is an ith associate of y, then z + x is an ith associate of
z + y. Furthermore, if x is an ith associate of zero, then v x is also an ith
associate of zero.
A consequence of this is that the determinant NN' is a circulant. If a circulant
has for its first row the elements a o , a l , ... , a n _ 1i the latent roots are given by
6r = ao + abw; + a2u,1 +... + a
,ti
_ wf
l
1,
where w1 = cos 2iij/n + i sin 21Tj/n, and j2 = 1, j = 1,2,. . . , n.

In this particular case, NN' is symmetric and n = v so that a l =
a 2 = a - 2 , ..., and the imaginary parts vanish. Hence,
Bj = ao + 2a 1 cos 2irj/v + 2a2 cos 4iij/v + :

i.e.,
(v-1)12
Bi = r + 2
ah cos 27rhj/v,
h=1
ifvisodd;
(v -2)12
B1 = r + 2
h=1
if v is even.
ah cos 27rhj/v + a 12 cos ir,
Sec. 15.10]
Cyclic Designs 323
As an example we may consider the design with the cyclic association scheme
given in Chapter 12:
012, 123, 234, 340, 401.
The first associates of zero are 1 and 4, a l = 2, a 2 = 1, and r = 3.
Applying the results of Connor and Clatworthy, we have y = 0,
0= 5, a 1 =a, =2, 0 i =(3 1/5)/2.
The matrix NN' is
3
1.
S = 2,
Hence, df = 3 + 4 cos 2ij/5 + 2 cos 47rj/5.

Recalling that cos 21T/5 = (- 1 + x/5)/4, cos 4ir/5 = (-1 - \/5)/4, we have
0 1 = 3 + 4 cos 2ir/5 + 2 cos 4r/5 = (3 + \/5)/2 = 0 4 ,
0 2 = 3 + 4 cos 47r/5 + 2 cos 87x/5 = 3( - V5)/2 = 0 3 .
Cyclic designs can be obtained with GD schemes by appropriately labeling the
treatments. We denote thejth treatment in the ith group by (i - 1) + (j - 1)m,
1 < i < m, 1 < j < n. In essence, this is what Bose et al. (1953) did in the
second example that we presented in Chapter 14. The design was for v = 12,
m = 6, and n = 2. The treatments in the first group were zero and 6, in the
second group 1 and 7, and so on. The initial block was 0 1 4 6 and the design,
consisted of twelve blocks. The differences in the initial block are
1-0=1,4-0=4,6-0=6,4-1=3,6-4=2,
and their negatives
0-1=11,8,6,9,10.
We note that the set contains each nonzero difference except 6 once, while 6 itself appears twice. We thus have a design with a l = 2 and X1 2 = 1.
J. A. John (1966) suggests using 0 13 7 as the initial block, and we may ask
whether the two designs are equivalent. It should first be pointed out that Bose
and his colleagues would have obtained the same design if they had started with
any of the other three blocks which contain zero, namely, 6 7 10 0, 8 9 0 2, or
11 0 3 5. Similarly, John could have suggested 11 0 2 6, 9 10 0 4, or 5 6 8 0. If we
multiply each of the members of an initial block by a number prime to v, in this
case 5, 7, or 11, and reduce mod v, we shall get an equivalent design [David
[Ch. 15
(1963)], inasmuch as our action will amount only to a relabeling of the treatments. In each case, 6 will remain unchanged. Similarly, if we replaced each
treatment by its negative (mod v), we should get an equivalent design. Multiplying 0 1 4 6 by 5 gives 0 5 8 6, which is the fourth of the possible initial blocks in
John's design. Thus the two designs are equivalent.
Cyclic designs with the hierarchic group divisible scheme can be obtained
by labeling the jth treatment in the ith subgroup of the hth group as
h 1 + (i 1)N1 + (j l)N1 N2 .
With this labeling, a design with Nl = N2 = N3 = 2 would have varieties 0, 2
and 4, 6 in the subgroups of the first group and 1, 3 and 5, 7 in the other group.
The E'S(l) design is now the cyclic design generated by 0 1 3 5 7. The ES(1)
design is also cyclic and consists of the two blocks 0 2 4 6 and 1 3 5 7, each
repeated four timer. The pair of blocks is called a partial cycle.
15.11 Generalized Cyclic Designs

We may generalize the concept of a cyclic design by allowing the varieties to be
elements of an Abelian group under addition. As before, it is convenient to take
the initial block to contain the identity zero. The jth block is then obtained by
adding the jth element of the group to each member of the initial block. Again,
several initial blocks may be used, and partial cycles may be included.
For group divisible designs we may use a double modulus system, representing
the jth treatment in the ith group by the ordered pair ij; addition is then defined
by gh + ij= xy where x - g + i (mod m) and y - h + j (mod n), and the
same system can be used for L 2 designs. Consider, for example, the array of nine
treatments
00 01 02
10 11 12
20 21 22,
with arithmetic for both coordinates being mod 3.
We can consider the array as a GD scheme with rows as groups. Then the
initial block 00 01 0211 generates a GD design with A l = 3 and A 2 = 1. Considering the array as an L 2 scheme, the initial block 00 01 12 22 generates the
ES(l) design with A l = 1 and A 2 = 2.
We can also use this labeling of treatments for the rectangular scheme.
Suppose we take a scheme with m = 2 and n = 3.
00 01 02
10 11 12.
Arithmetic for the first digit is mod 2, and for the second digit mod 3. Then the
initial block 00 11 12 gives the E'S(3) design with A l = 1, A a = 0, and a 3 = 2.
Sec. 15.12]
Designs for Eight Treatments in Blocks of Three 325
In the cubic scheme, the coordinate system is immediately adaptable to cyclic

designs. We give two examples for the schemes with v = 27 and v = 64.
In the scheme for v = 27, we denote the treatments by 000, 001, and so on,
with arithmetic mod 3. Consider the design generated by the initial block
000, 001, 002. The six differences are 001 and 002, each repeated three times.
The twenty-seven blocks consist of a partial cycle of nine blocks repeated three
times. Suppose now that we take the set of three initial blocks, 000, 001, 002; 000,
010, 020; 000, 100, 200, and take one partial cycle from each set of blocks. The
resulting design will have b = v = 27, r = k = 3, ,1 1 = 1, X1 2 = 0, and X1 3 = 0.
In a similar way, partial cycles from the six initial blocks 000, 012, 021;
000, 011, 022; 000, 102, 201; 000,101,202; 000, 120, 210; 000, 110, 220 give
b = 54, r = 6, k = 3, A, = 0, " 2 = 1, and A 3 = 0. The initial blocks 000, 111,
222; 000, 112, 221; 000,121,212; 000, 211, 112 likewise give a design with
b = 36, r = 4, A l = = 0, and A 3 = 1, and all three of these PBIB(3) designs
combined form a BIBD with b = 117, k = 3, r = 13, A = 1.
For v = 64 we can use for the coordinates the elements of the Galois field
GF(2 2 ), namely, 0, 1, x, and 1 + x, which we write as y. We can obtain a
similar set of designs with k = 4 and other parameters:
(i)
b =48, r=3, ,1 1 =1, 2 =0, A 3 =0;
(ii)
b = 144, r = 9, X1 1 =0, 2 =1, A3 =0;
(iii) b = 144, r = 9, A, = 0, A 2 = 0, ,1 3 = 1.
In design (iii), one of the initial blocks used is 000, 111, xxx, yyy. The differences are 111, xxx, yyy, xxx - 111 = yyy, yyy - 111 = xxx, yyy - xxx =
111 and their negatives; but sine all arithmetic is mod 2, taking the negatives
amounts to repeating each difference again. The set of twelve differences is thus
111, xxx, and yyy, each appearing four times. The actual design consists of
partial cycles (sixteen blocks) on each of the following nine initial blocks:
000, 111, xxx, yyy; 000, 11x, xxy, yyl; 000, ily, yyx, xxl;
000, lxl, xyx, yly; 000, lyl, yxy, xlx; 000, xl 1, l yy, yxx;
000, yl 1, xyy, lxx; 000, lxy, xyl, yxl ; 000, lyx, yxl, .xly.
15.12 Designs for Eight Treatments in Blocks of Three

As an illustration of the techniques that have been developed so far, we consider designs with v = 8 and k = 3. In particular, we shall look at designs with
two or three associate classes. We have the following possible schemes:
(i) GD with m = 2, n = 4, and ordering 0 24 6, 1 3 5 7;
(ii) GD with m = 4, n = 2, and ordering 04, 15, 26, 37;
(iii) Hierarchic GD with ordering 04, 26, 15, 37;
(iv) Rectangular;
(v) Cubic.
[Ch. 15
For v = 8, the last two schemes are equivalent.

We turn first to the ordinary cyclic designs with a single initial block. Each
initial block consists of zero and two other treatments. There are thus
(2)
= 21
possible initial blocks, but they occur in equivalence sets.

/
Consider the block 0, 1, 2. When we cycle on it, we obtain 0 1 7, 0 6 7, and so
these three blocks are equivalent in the sense that they generate the same design.
Furthermore, multiplying by 3 gives 0 3 6, 0 3 5, 0 2 5, and we have one equivalence set with six triples. The initial block 01 3 also gives us 0 2 7, 0 5 6, with
differences 1 - 0, 3 - 0, 3 - 1, and their negatives. We could as wel] have
considered the negative of the block, replacing 1 by 7 and 3 by 5 to give 0 5 7, and,
similarly, 01 6, 0 2 3, for another set of six. There are two more equivalence sets,
01 4,037,045,047,01 5,034,andthesetofthreetriples,024,026,046.
For designs with a single initial block, we need therefore consider only the four
cases:
(i) 01 2, with differences 1, 1, 2, 6, 7, 7 [or, as David (1963) writes, (1 2 , 2)];
this gives four associate classes, with the associates of zero being 1, 7;
2, 6; 3, 5;4,andA l =2,A2 = 1, 1 3 =A 4 =0;
(ii) 01 3, which gives a GD design with groups 0, 4; 1, 5; 2, 6; 3, 7 and
A r =0, ,12 =1;
(iii) 014, which gives a hierarchic GD design with ,1 1 = 1, A a = 0, ,1 3 = 2;
and
(iv) 024, which gives a disconnected GD design with m = 2, n = 4, and
=2,A2 =0.
J. A. John (1966) also considers combinations of several initial blocks. The
set (0 1 2), (0 1 3), (014) gives a design for four classes; a l = 4, A2 = As =
A 4 = 2. There are three other designs with three associate classes, all of which
have the hierarchic scheme. They are generated by
(0 1 3), (0 2 4)
Jol = 1, A 2 = 3, A3 = 2;
(0 1 3), (014)
Al = 2, A 2 = 1, A 3 = 2;
(0 1 3), (014), (0 2 4)
A l = 2, ,1a = 3, A 3
We can obtain three more designs by using the cubic scheme for v = 8.
Denoting the treatments by 000, 001, 010, and so on, with arithmetic mod 2, the
initial blocks and values of A, are
(000, 001, 010), (000, 010, 100), (000, 100, 001)
A 1 = 4, A 2 = 2, A3 = 0;
(000, 001, 110), (000, 010, 101), (000, 100, 011)
X1 1
= A 2 = 2, A s = 6;
(000, 011, 110)
X11
= a3 = 0, ,1
= 2.
Exercises 327
15.13 Conclusion
The construction of designs and the discovery of new schemes remains a
fertile field for research, and in this chapter we have given only a limited survey
of the field. It might, however, be appropriate to recall the last pages of Chapter
12 in which it was suggested that from a practical standpoint there is sometimes
little to choose between the various alternative designs that can be used in some
problems. We mentioned at that time the nineteen designs that J. A. John listed
for v = 12 and k = 4, with efficiencies ranging from 0.8007 to 0.8166, the
maximum being attained by a GD design. The mathematical problems discussed in the last three chapters are very interesting from a theoretical point of
view and have a fascination for pure mathematicians as well as for statisticians.
In spite of the beauty and elegance of the mathematical results, the practicing
statistician should stil) bear in mind that the more complex the design, the easier
it is for something to go wrong. We conclude with the maxim that has already
appeared at the end of Chapter 12: When you design an experiment, keep it as
simple as you can.
Exercises
1. Find the latent roots of NN' and their multiplicities for the triangular designs.
2. Find the latent roots of NN' and their multiplicities for designs with the L(n)
scheme.
3. Verify the results given for the latent roots of NN' and their multiplicities for the
several three-associate class schemes.
4. The hierarchic GD design is obtained from a GD design by replacing each
treatment by a set of N 3 new treatments. We may similarly take an L(n) scheme
and replace each treatment by a group of m new treatments. Two treatments are
said to be first associates if they belong to the same group, and second associates
if their groups correspond to first associates in the L(n) scheme. This is the
L,(n, m) scheme of Singh and Singh (1962). Find the P matrices, the latent roots
of NN', and their multiplicities for this scheme.
5. (Continuation) They also defined in the same paper the T(m) scheme, in which
they take a triangular scheme and replace each treatment by a group of m new
treatments. The three associate classes are defined as for the L,(n, m) scheme.
Derive the P matrices, the latent roots of NN', and their multiplicities for this
scheme.
6. Show that if in a PBIB(2) design 0 is not a perfect square then
P21 = P12 = Pi = t,
nt = na = al = a2 = 2t,
v = 4t + 1.
7. Show that if v = 4t + 1 and v is a prime we may obtain a cyclic association

scheme by taking as the first associates of zero either the set of quadratic
residues of v or its complement; the first associates of i are then obtained by
adding i to each of the first associates of zero.
[Ch. 15
8. (Continuation) Obtain a similar result when v = 9 or 25 by allowing the treatments to be the elements of the Galois field of v elements.
9. Show that for a PBIB(2) scheme vn 1 n 2 = 1a 1 a2 [Mesner (1965)].
10. Show that a necessary and sufficient condition for a PBIB(2) scheme to have 1
and a 2 equal in some order to n l and n 2 is that the scheme be of the pseudocyclic, the pseudo-Latin square or the NL types [Mesner (1967)].
11. Prove the following results for designs with the rectangular scheme:
(i) If 0 = 0, k is divisible by n and each block contains exactly kin treatments
from each column of the array;
(ii) If 02 = 0, k is divisible by m and each block contains kim treatments from
each row;
(iii) If 0 = B z = 0, there are k(r 1)/(b 1) treatments common to any two
blocks of the design [Vartak (1959)].
12. Show that there are only five nonisomorphic cyclic designs with v = b = 8 and
r = k = 4 and a single initial block.
13. Show that although the cyclic design for v = 8 generated by (0,. 1, 2, 4) has only
two distint values of a,, nevertheless there are four associate classes.
14. Prove that cos(27T/5) = ( 1 + x/5)/4.
Appendix
A.1 Matrices
We review in this appendix some properties of matrices, vector spaces, and
quadratic forms. The elements of all the matrices will be real numbers, and some
of the results given are not true without this restriction. Some of the results will
be proved as theorems; others will be stated without proof, and some of these
will appear later as exercises. Although it is reasonable to assume that most
students entering a graduate course in the mathematical theory of design of
experiments will have had some exposure to elementary matrix theory, many
undergraduate courses in linear algebra do not progress as far as the discussion
of quadratic forms, which are essential to the development of our subject. This
does not attempt to be a complete presentation, and the reader is referred for
further details to the standard texts. Some basic results about determinants will
be assumed.
A matrix, in the context in which we shall be considering them, is a rectangular
array of elements, which in our case will be real numbers (or, at a later stage,
smaller matrices of real numbers). We talk of a matrix A having m rows and n
columns, sometimes written (m x n), and denote by ai , the element in the ith
row and the jth column. We shall use boldface type, A, to denote matrices and we
shall write A = (a). A square matrix with m rows and m columns is said to be of
order m. If a{5 = a;; , A is called symmetrie. 0 will denote a matrix of zero elements. I, is the identity matrix, the square matrix of order n with ones along the
main diagonal and zeros elsewhere, i.e., a = 1, a, ; = 0, i j. If there is no
danger of confusion, we shall omit the subscript n and write I.
329
330 Appendix
A matrix A can be multiplied by a scalar (a real number) c. The product is
defined as cA = B where b,1 = caii .
Two matrices A and B, each with m rows and n columns, can be added.
Addition is defined by A + B = C where c sj = a,, + bi,.
The product AB of two matrices is defined only if the number of columns of A
is the same as the number of rows of B, in which case we define AB = C where
Ct, = I k a1kbkl. In general it is not true that AB = BA; indeed, one product may
exist and the other not. If AB = BA, the two matrices are said to commute.
The transpose of a matrix A with m rows and n columns will be written as A'.
Some authors write AT. It is defined as A' = C where C has n rows and m
columns and c, 1 = a51 . If A and B are two matrices such that AB exists, then
(AB)' = B'A'.
So far as we are concerned, a vector of n elements may be considered as a
matrix with n rows and a single column. Its transpose is sometimes called a
row vector. We shall use 1,, to denote a vector of n unit elements. Em will
denote a matrix of m rows and n columns with every element unity: Em,,,
1,,,i. We shall usually write J for E,a , n . The length of a vector y is defined as d
where d 2 = y'y = 1 y. Two vectors, x and y, are said to be orthogonal if
x'y = y'x = 0; this implies that x and y have the same number of elements.
Consider a set of vectors y l , ... , ym , each with n elements. They are said to be
linearly independent if there exists no set of scalars Cl, ..., C m (except the set
C 1 = = c m = 0) for which 1 c i yi = 0 where 0 is a vector of zero elements.
Otherwise, we have a linearly dependent set. The zero vector 0 is itself a linearly
dependent set. The collection of all linear combinations c {yi is called the vector
space, V, spanned by the set of vectors y i , ..., y m . If the set y l , ..., y m is a
linearly dependent set, Vis also spanned by a subset of y l , ... , Y. This leads us
to the idea of a basis for a vector space. We define a basis (or a Hamel basis) for a
vector space V to be a set of linearly independent vectors that span V. It can be
shown that every vector space has a basis and that all bases for a given vector
space contain the same number of vectors. This number is called the dimension
of the space. It is closely connected with the next concept, that of the rank of a
matrix. We may consider a matrix A, (m x n), as a collection of n columns or
vectors.
The rank of A, written r(A), is the number of linearly independent columns in
A, which is the same as the dimension of the vector space spanned by the
columns of A. It can be shown that this is the same as the number of linearly
independent rows of A, and that r(A) is equal to the order of the largest nonvanishing determinant that can be formed from the columns and the rows of A.
It follows that r(A) < min(m, n), and that r(0) = 0. We conclude these preliminaries by defining the inverse of a matrix and proving some theorems about
the ranks of matrices.
Let A be a square matrix, and let Al be the determinant with elements a is .
If Al = 0, A is said to be singular; if IAI 0 0, A is nonsingular. Let A, f denote
the cofactor of the element a, i in Al; A,j is the value of the determinant obtained
by striking out the ith row and the jth column of A, multiplied by (-1)'+'
Appendix 331
Suppose now that A is nonsingular, and consider the matrix C = (c, j) where
= AjJ1Al. It follows from elementary properties of determinants, namely
1 a 1A, f = Âl and I; akJA {f = 0 where i k, that AC = CA = I. C is called
the inverse of A and is written as A -1 . C is unique, and, if A and B are both
nonsingular, (AB) -' = B - lA -1 . Every square nonsingular matrix A has an
inverse. If r(A) is less than the order of A, Al = 0, and A has no inverse.
c{1
Theorem. Let A and B be (m x n) matrices. Then r(A + B) < r(A) + r(B).

PROOF. Let a be a set of r(A) linearly independent columns of A, and P a set of
r (B) linearly independent columns of B; let y be the union of the two sets a and /.
Every column of C = A + B is a linear combination of the columns of y.
Then r(C) is not greater than the number of columns in y, and so r(A + B) <
r(A) + r(B). 1
Theorem. Let A and B be matrices such that AB exists. Then r(AB) < r(A) and
r(AB) < r(B).
PROOF. The columns of AB are linear combinations of the columns of A. Thus
they belong to the space spanned by the columns of A, and so r(AB) < r(A).
Similarly, r(B'A') < r(B'). But r(B'A') = r(AB) and r(B') = r(B) so that
r(AB) < r(B). a
Corollary 1. If A is nonsingular and AB exists, r(AB) = r(B).
PROOF. By the theorem r(AB) < r(B); but B = (A - ')(AB), so that r(B) <_
r(AB). Hence, r(AB) < r(B) < r(AB), and so r(AB) = r(B). We may similarly
prove that, if CA exists, r(CA) = r(C). a
Corollary 2. If A and C are nonsingular r(ABC) = r(B), i.e., the rank of a matrix
is not changed by multiplying on the left and the right by nonsingular matrices.
Corollary 3. If A and B are square matrices and AB = 0, then A = 0, or B = 0,
or else A and B are both singular.
PROOF. Suppose, in contradiction, that A is nonsingular and B 0. Then
r(AB) = r(B) > 0, and AB 0. Similarly, if B is nonsingular and A = 0,
AB
0
0.
3
A.2 Orthogonality
We have mentioned that two vectors, x and y, are said to be orthogonal if the
product x'y = y'x = 0. A set of vectors of unit length which are mutually
orthogonal is called an orthonormal set. A set of n mutually orthogonal vectors of
332 Appendix
n elements each is a linearly independent set; it is not possible to find a (n + 1)th
vector other than 0 which is orthogonal to all the others. A matrix P (n x n)
which has for its columns a set of orthonormal vectors is called an orthogonal
matrix. If P is orthogonal, P'P = I and P' = P -1 .
If x is a vector of coordinates in Euclidean n-dimensional space and P is an
orthogonal matrix, the transformation into new coordinates y defined by
y = Px is called an orthogonal transformation and corresponds to a rotation of
the axes. A simple example occurs in two-dimensional analytic geometry when
the coordinate axes are rotated through an angle 0. With respect to the new axes,
the coordinates of the point (x 1 , x 2 ) become (y l , y 2 ), where y l = x l cos B x 2
sin 0, Y 2 = x l sin 0 + x 2 cos 0, or
[y 2 ] [sin B cos 9, [x 2 ,
[y 2 ] [sin B cos 9, [x 2 ,
A.3 Quadratic Forms

Let Y be a vector of n real elements y l , ..., y,,. A homogeneous seconddegree multinomial in these elements, l i b ;t yi + l i I, bi , y; yf , where i <j, is
called a quadratic form. We shall write such a form as Y'AY where A is a
symmetric matrix having a{ , = b ; , and a;1 = a;i = b ;j /2; A is called the matrix of
the form. We shall consider only forms for which A is a symmetric matrix. A
quadratic form Q is said to be positive definite if Q is positive for all values of Y
except Y = 0. The form Q = 2y1 + 2y,y2 + y2 is positive definite; it may be
written as the sum of two squares, Q = yi + (yl + y2) 2 . If Q is never negative,
but is zero for some values of Y other than 0, it is said to be positive semidefinite. Q = (y l Y 2 ) 2 is an example; it is never negative, but it takes the value
zero whenever y l = Y 2 . The term nonnegative definite is used to cover both
positive definite forms and positive semidefinite forms. The same terms are
applied to the corresponding matrices A.
Suppose that a ll = 0. Then the form ailyi = 0 for all y l , and A cannot be
positive definite. If a ll < 0, allyi < 0 for y l 0. Ifa ll = 0 and a 12 is positive,
Q = all Yi + 2a12Y1Y2 is negative whenever y 1 y2 is negative. It follows that a
positive definite matrix has all its diagonal elements, a t ,, positive. In a positive
semidefinite matrix all the diagonal elements are nonnegative, and, if a ;i = 0,
then = a; , = 0 for all j; i.e., y i does not appear in the form.
A.4 Latent Roots and Latent Vectors

If A is a square matrix of order n, and d is a root of the equation A M] = 0,
there exists a vector x other than 0 such that [A AI]x = 0, and,
Ax = ^1x;
Appendix 333
A is called a latent root (or characteristic root or proper value or eigenvalue) of
A, and x is a latent (characteristic, proper or eigen-) vector corresponding to A.
Counting multiplicities, there are exactly n roots; if A is symmetrie, the number
of nonzero roots of A is equal to the rank of A. We now confine ourselves to
symmetrie matrices A.
Theorem 1. If X11 , a ; are two latent roots of A such that ,\ 0 ,1 j , and x, x ; are two
corresponding latent vectors, then x i and x ; are orthogonal.
Axf = ,1 ; x1 , so that xj'Ax, = d i xix ; , x,Ax 5 = A.xix f .

Then, since A is symmetrie, taking the transpose of the Jatter expression gives
x;Ax = 21; x;x,. Thus A; xfx, = A jxjx t , and since A. # ,1; , x;x t = 0.
It can also be shown that if A is a latent root with multiplicity m, there is a set
of exactly m mutually orthogonal latent vectors corresponding to A. We may take
the lengths of the latent vectors to be unity, and thus obtain for A a st of n
orthonormal vectors. Let P denote the matrix whose columns are these orthonormal vectors; P is orthogonal and
PROOF. We have Ax ; = ,1 i xj ,
{
AP = PA,
where A is a diagonal matrix (i.e., a matrix with all its off-diagonal elements
zero), having as the diagonal elements the latent roots, each appearing as many
times as its multiplicity. Multiplying on the left by P' gives P'AP = P'PA = A.
We state this result as a theorem.
Theorem 2. Let A be a square symmetrie matrix, and A be the diagonal matrix
fust defined. There exists an orthogonal matrix P such that P'AP = A. Conversely, if there exists an orthogonal matrix P such that P'AP = D where D is
diagonal, the diagonal elements are the latent roots of A.
Since P is nonsingular, r(A) = r(P'AP) = r(A), and we see that the rank of A
is equal to the number of nonzero latent roots, counting their multiplicities.
Denote the ith diagonal element of A by Al . Let a diagonal matrix D be constructed as follows. The ith diagonal element d ij of D = 1 if A l = 0 and -1 / 2 if
a, 0. Let Q = PD. Q is not singular. Then the transformation Y = QZ
changes Y'AY into 1 z{ where the number of positive signs is equal to the
number of positive latent roots, and the number of negative signs is equal to the
number of negative latent roots.
If B is a nonsingular matrix we have
B -1 Ax = B 'ABB ix = B llx = AB -1 x,
-
so that A is also a latent root of B 'AB with latent vector B 'x. The two matrices
A and B 'AB are said to be similar.
-
Theorem 3. A is positive definite if, and only if, all its latent roots are positive.
334 Appendix
PROOF. Let Z = P'Y, where P is an orthogonal matrix such that P'AP = A.
Then Y'AY = Y'PP'APP'Y = Z'AZ, and, since Z = 0 if, and only if, Y = 0,
the form Y'AY is positive definite if, and only if, Z'AZ is positive definite. Thus
A is positive definite if, and only if, A is positive definite, i.e., since A is diagonal,
if, and only if, all the diagnal elements of A are positive.
3
Similarly, A is positive semidefinite if, and only if, A has no negative latent
roots and has at least one zero latent root, which implies that A is singular.
The trace of a square matrix A is defined as the sum of the diagonal elements:
tr(A) = 1 a, i . It is easily shown by performing the multiplications that, if A and
B are both square matrices, tr(AB) = tr(BA).
Theorem 4. The sum of the latent roots of A is equal to the trace of A.

PROOF.
d, = tr(A) = tr(P'AP) = tr(APP') = tr(AI) = tr(A).
Theorem 5. The latent roots of a real symmetric matrix are real.

PROOF. Suppose that we have a complex root ,A + i with a complex vector
x + iy where , , , x, and y are all real. Then
A(x + iy) = (,1 + i)(x + iy)

and, equating the real and imaginary parts,
Ax=Axy
(A.1)
Ay = Ay + x.
(A.2)
Multiply Equation A. 1 on the left by y'. Multiply Equation A.2 on the left by x'
and transpose it. Then
y'Ax = Ay'x uy'y = Ay'x + x'x,

so that (x'x + y'y) = 0. But x'x + y'y > 0, and so = 0.
(A.3)
Theorem 6. The latent roots of Aa are the squares of the latent roots of A.
PROOF.
We have P'AP = A. Then
P'APP'AP = P'A 2P = A2
so that the orthogonal matrix P also diagonalizes A 2 , and the diagonal elements
of the diagonalized matrix are d, = A. Thus the latent roots of A 2 are Aa.
Furthermore, multiplying Ax = Ax on the left. by A gives
A2x = AAx = A2x,
Appendix 335
and we see that each vector x corresponding to a root ,1 of A also corresponds to
a root A 2 of A 2 .
The following theorem is of particular use to us.
Theorem 7. Let X be a real matrix with m rows and n columns. Then X'X is
nonnegative definite and r(X'X) = r(X).
Since X'X is symmetric, there is an orthogonal P such that P'X'XP = A.
Let B = (b if) = XP. Then B'B = A, and we have, writing the columns of B as
vectors, b i , bbi = a, and babi = 0 where i # j. It follows that A, = 0 if and only
if b, = 0; otherwise, > 0. Thus the latent roots are nonnegative and X'X is
nonnegative definite.
The vectors b, are mutually orthogonal. The rank of X'X is equal to the
number of positive diagonal elements of A, which, in turn, is the number of b $
that are not 0. Thus, r(X'X) = r(B) = r(XP) = r(X). E
We state without proof the following theorem, which is to be found in many
of the standard textbooks.
PROOF.
Theorem 8. A necessary and sufficient condition for A to be positive definite is that

every principal minor of IAI shall be positive. A principal minor of Al is the
determinant of a submatrix of A obtained by striking out some rows of A and the
corresponding columns so that the diagonal elements of the submatrix are also
diagonal elements of A.
A square symmetric matrix A is said to be idempotent if Al = A. The following
properties of idempotent matrices are readily established:
(i) A necessary and sufficient condition for A to be idempotent is that the
latent roots of A are either zero or one;
(ii) If A is idempotent, r(A) = tr(A);
(iii) If A is idempotent and Q is orthogonal, Q'AQ is idempotent;
(iv) If A is idempotent and a{ , = 0, then ajj = af = 0 for all j;
(v) If A is idempotent, a,ll its diagonal elements are nonnegative.
,
A.5 Simultaneous Linear Equations

We conclude this discussion of matrices by recalling some results about the
solution of sets of simultaneous linear equations.
Consider the set of linear equations Ax = b where A has m rows and n
columns, and suppose that we wish to solve for x. If m = n, and r(A) = n, the
solution vector is x = A 'b. Otherwise, there are three possibilities to be
considered:
-
(i) There is no solution;

(ii) There is an infinite number of solution vectors;
(iii) There is a unique solution vector.
336 Appendix
For example, we may have the following four equations in two unknowns:
+ x2 = 3
(A.4)
xl
+ 2x2 = 4
(A.5)
2x 1
+ 3x2 = 7
(A.6)
2x 1
+ 3x 2 = 8.
(A.7)
x1
Equation A.4 alone admits an infinity of solutions; we may take any number t
that we wish and obtain the solution vector x' = (t, 3 t).
The pair of equations A.4 and A.5 has the unique solution vector x' = (2, 1).
The addition of equation A.6 makes no difference because it is actually just the
sum of Equations A.4 and A.5. When, however, Equation A.7 is included we have
an inconsistent set; subtracting Equation A.6 from Equation A.7 would give
0=1.
We state the following results without proof. B is the augmented matrix
(A, b), i.e., the matrix A with the column b added to it. A set of equations is raid
to be consistent if it has at least one solution.
A necessary and sufficient condition for the equations to be consistent is that
r(A) = r(B). If r(A) = r(B) = n, there is a unique solution vector; if r(A) =
r(B) < n, then there is an infinity of solutions and n r(A) of the x i may be
assigned arbitrarily.
Exercises
The exercises involve the derivation of some results stated without proof in the
text. A denotes a square symmetric matrix of order n with real elements. C is any
matrix with real elements.
1. Show directly that (AC)' = C'A'.
2. All bases for a given finite dimensional vector space have the same number of
vectors.
3. Let V be the space of all vectors in En (Euclidean n space). The dimension of
V. is n.
4. There does not exist a set of n + 1 mutually orthogonal vectors in V.

5. The rank of A is the order of the largest nonvanishing determinant in A.
6. The number of linearly independent rows of C is the same as the number of
linearly independent columns of C.
7.A -1 is unique. If B is a right inverse of A and E is a right inverse, then B = E.
If B and E are both left inverses, B = E. We have given in the text a procedure
for obtaining one matrix that is both a right and left inverse of A when A is
nonsingular.
8. We defined latent roots as solutions of IA Al] = 0. Show that any scalar ,1 for
which there exists a vector x ; 0 such that Ax = Ax is a root of JA tIj = 0.
Appendix 337
9. If A is a latent root of A with multiplicity m, show that we can find a set of m
mutually orthogonal latent vectors (but no more) corresponding to A.
10. A is positive semidefinite if, and only if, A has at least one zero latent root and no
negative latent roots.
11. tr(BC) = tr(CB).
12. If there exists an orthogonal matrix P such that P'AP is diagonal, then
P'AP = A.
13. A necessary and sufficient condition for A to be positive definite is that every
principal minor of IAI be positive.
14. If A is idempotent and I = A + B, then B is idempotent and AB = BA = 0.
15. If the elements of C are real, C'C = 0 if, and only if, C = 0. Show by a counterexample that this is not true when C has complex elements.
Bibliography
ADDELMAN, S. (1962). "Orthogonal Main-Effect Plans for Asymmetrical Factorial

Experiments." Technometrics, 4, 21-46.
ADDELMAN, S., and O. KEMPTHORNE (1961). "Some Main Effects Plans and Orthogonal Arrays of Strength Two." Ann. Math. Statist., 32, 1167-1176.
AIA, M. A., R. L. GoLDSMITH, and R. W. MOONEY (1961). "Precipitating Stoichiometric CaHPO 4 . 2H 2 0." Industrial and Engineering Chemistry, 53, 55-57.
AITKEN, A. C. (1950). "On the Statistical Independence of Quadratic Forms in
Normal Variates." Biometrika, 37, 93-96.
ALLAN, R. E., and J. WISHART (1930). "A Method of Estimating the Yield of a
Missing Plot in Field Experimental Work." J. Agr. Sci., 20, 399-406.
ATIQULLAH, M. (1958). "On Configurations and Non-isomorphism of Some
Incomplete Block Designs." Sankhya, 20, 227-248.
ATIQULLAH, M. (1961). "On a Property of Balanced Designs." Biometrika, 48,
215-218.
BANERJEE, K. S. (1964). "A Note on Idempotent Matrices." Ann. Math. Statist., 35,
880-882.
BATCHELDER, A. R., M. J. ROGERS, and J. P. WALKER (1966). "Effects of Subsoil
Management Practices on Growth of Flue-Cured Tobacco." Agron. Journal, 58,
345-347.
BAUMERT, L., S. W. GOLOMB, and M. HALL, Jr. (1962). "The Discovery of an
Hademard Matrix of Order 92." Bull. Amer. Math. Soc., 68, 237-238.
BENNETT, C. A., and N. J. FRANKLIN (1954). Statistical Analysis in Chemistry and
the Chemical Industry. John Wiley and Sans, New York.
BHATTACHARYA, K. N. (1943). "A Note on Two-Fold Triple Systems. Sankhya, 6,
313-314.
BHATTACHARYA, K. N. (1944a). "A New Balanced Incomplete Block Design."

Science and Culture, 9, 508.
339
340 Bibliography
BHATTACHARYA, K. N. (1944b). "On a New Symmetrical Balanced Incomplete

Block Design." Bull. Calcutta Math. Soc., 36, 91-96.
BHATTACHARYA, K. N. (1946). "A New Solution in Symmetrical Balanced Incomplete Block Designs (v = b = 31, r = k = 10, A = 3)." Sankhya, 7, 423-424.
BINET, F. E., R. T. LESLIE, S. WEINER, and R. L. ANDERSON (1955). "Analysis of
Confounded Factorial Experiments in Single Replications." North Carolina Agric.
Exp. Station. Techn. Bull. No. 113.
BOND, D. A. (1966). "Yield and Components of Yield in Diallel Crosses Between
Inbred Lines of Winter Beans (vicia faba)." J. Agri. Sci., 67, 325-336.
BOSE, R. C. (1938). " On the Application of the Properties of Galois Fields to the Problem of the Construction of Hyper-Graeco-Latin Squares." Sankhya, 3, 323-338.
BOSE, R. C. (1939). "On the Construction of Balanced Incomplete Block Designs."
Annals of Eugenics, 9, 353-399.
BOSE, R. C. (1 942a). "On Some New Series of Balanced Incomplete Block Designs."
Bull. Calcutta Math. Soc., 34, 17-31.
BOSE, R. C. (1942b). "A Note on Two Series of Balanced Incomplete Block Designs."
Bull. Calcutta Math. Soc., 34, 129-130.
BOSE, R. C. (1942c). "A Note on the Resolvability of Balanced Incomplete Block

Designs." Sankhya, 6, 105-110.
BOSE, R. C. (1944). "The Fundamental Theorem of Linear Estimation." Proc. 31st
Indian Sci. Congress, 2-3 (abstract).
BOSE, R. C. (1947a). " Mathematical Theory of the Symmetrical Factorial Design."
Sankhya, 8, 107-166.
BOSE, R. C. (1947b). "On a Resolvable Series of Balanced Incomplete Block Designs."

Sankhya, 8, 249-256.
BOSE, R. C. (1949). "A Note on Fisher's Inequality for Balanced Incomplete Block
Designs." Ann. Math. Statist., 20, 619-620.
BOSE, R. C. (1963). "Combinatorial Properties of Partially Balanced Designs and
Association Schemes." Sankhya, Ser. A., 25, 109-136.
BOSE, R. C., and K. A. B.uSH (1952). "Orthogonal Arrays of Strength Two and
Three." Ann. Math. Statist., 23, 508-524.
BOSE, R. C. and W. H. CLATWORTHY (1955). "Some Classes of Partially Balanced
Designs." Ann. Math. Statist., 26, 212-232.
BOSE, R. C., W. H. CLATWORTHY, and S. S. SHRIKHANDE (1954). "Tables of Partially
Balanced Designs with Two Associate Classes." North Carolina Agric. Exp.
Station. Tech. Bull. No. 107.
BOSE, R. C., and W. S. CONNOR (1952). "Combinatorial Properties of Group
Divisible Incomplete Block Designs." Ann. Math. Statist., 23, 367-383.
BOSE, R. C., and K. KISHEN (1940). "On the Problem of Confounding in the General
Symmetrie Factorial Design." Sankhya, 5, 21-36.
BosE, R. C., and D. M. MESNER (1959). "On Linear Associative Algebras Corresponding to Association Schemes of Partially Balanced Designs." Ann. Math.
Statist., 30, 21-38.
BOSE, R. C., and K. R. NAIR (1939). "Partially Balanced Incomplete Block Designs."
Sankhya, 4, 337-372.
BOSE, R. C., and T. SHIMAMOTO (1952). "Classification and Analysis of Designs with
Two Associate Classes." J. Amer. Statist. Ass., 47, 151-184.
BOSE, R. C., S. S. SHRIKHANDE, and K. N. BHATTACHARYA (1953). " On the Construction
of Group Divisible Incomplete Block Designs." Ann. Math. Statist., 24,167-195.
Bibliography 341
C., S. S. SHRIKHANDE, and E. T. PARKER (1960). "Some Further Results on
the Construction of Mutually Orthogonal Latin Squares and the Falsity of
Euler's Conjecture." Canad. J. Math., 12, 189-203.
Box, G. E. P. (1953). "Non-normality and Tests on Variances." Biometrika, 40,
318-335.
Box, G. E. P. (1957). "Evolutionary Operation: A Method for Increasing Industrial
Productivity." Applied Statist., 6, 3-23.
Box, G. E. P., and D. W. BEHNKEN (1960). "Simplex Sum Designs: A Class of
Second Order Rotatable Designs Derivable from Those of First Order." Ann.
Math. Statist., 31, 838-864.
Box, G. E. P., and J. S. HUNTER (1957). "Multifactor Experimental Designs for
Exploring Response Surfaces." Ann. Math. Statist., 28, 195-242.
Box, G. E. P., and J. S. HUNTER (1959). "Condensed Calculations for Evolutionary
Operation Programs." Technometrics, 1, 77-95.
Box, G. E. P., and J. S. HUNTER (1961). "The 2k p Fractional Factorial Designs."
Technometrics, 3, Part I, 311-352; Part II, 449-458.
Box, G. E. P., and K. J. WILSON (1951). "On the Experimental Attainment of
Optimum Conditions." J. Roy. Statist. Soc., Ser. B., 13, 1-45.
Box, G. E. P., and P. V. YOULE (1955). "The Exploration and Exploitation of
Response Surfaces: An Example of the Link Between the Fixed Surface and the
Basic Mechanism of the System." Biometrics, 11, 287-323.
BROWNLEE, K. A., B. K. KELLY, and P. K. LORAINE (1948). "Fractional Replication
Arrangements for Factorial Experiments with Factors at Two Levels." Biometrika,
35, 268-276.
BRUCK, R. H., and H. J. RYSER (1949). "The Non-existence of Certain Finite
Projective Planes." Canad. J. Math., 1, 88-93.
BULMER, M. G. (1957). "Approximate Confidence Limits for Components of
Variance." Biometrika, 44, 159-167.
BURTON, R. C., and W. S. CONNOR (1957). "On the Identity Relationship for
Fractional Replicates in the 2" Series." Ann. Math. Statist., 28, 762-767.
CALVIN, L. C. (1954). "Doubly Balanced Incomplete Block Designs for Experiments
in Which the Effects Are Correlated." Biometrics, 10, 61-88.
CHANDA, et al. (1952). "The Use of Chromium Sesquioxide to Measure the Digestibility of Carotene by Goats and Cows." J. Agr. Sci., 42, 179-185.
CHANG, L. C. (1959). "The Uniqueness and Non-uniqueness of the Triangular
Association Schemes." Science Record, 3, new series, 604-613.
CHANG, L. C. (1960)." Association Schemes of PartiallyBalanced Designswith Parameters v = 28, n l = 12, n 2 = 15, pi l = 4." Science Record, 4, new series, 12-18.
CHANG, L. C., and W. R. Liu (1964). "Incomplete Block Designs with Square
_ 10." Scientia Sinica, 13, 1493-1495.
Parameters for which k <_ 10 and r <
CHANG, L. C., C. W. Liu, and W. R. Liu (1965). "Incomplete Block Designs with
_ 10." Scientia Sinica, 14, 329-338.
Triangular Parameters for k < 10 and r <
CHOWLA, S., and H. J. RYSER (1950). "Combinatorial Problems." Canad. J. Math.,
2, 93-99.
CLATWORTHY, W. H. (1956). "Contributions on Partially Balanced Incomplete
Block Designs with Two Associate Classes." National Bureau of Standards.
BOSE, R.
Applied Math. Series, No. 47.

CLATWORTHY, W. H.
(1967). "Some New Families of Partially Balanced Designs of

the Latin Square Type and Related Designs." Technometrics, 9, 229-244.
342 Bibliography
COCHRAN, W.
G. (1934). "The Distribution of Quadratic Forms in a Normal

System, with Applications to the Analysis of Covariance." Proc. Camb. Philos.
Soc., 30, 178 -191.
COCHRAN, W. G., K. M. AUTREY, and C. Y. CANNON (1941). "A Double ChangeOver Design for Dairy Cattle Feeding Experiments." J. Dairy Sci., 24, 937-951.
COCHRAN, W. G., and G. M. Cox (1957). Experimental Designs. (Second ed.) John
Wiley and Sons, New York.
CONNOR, W. S. (1952a). "On the Structure of Balanced Incomplete Block Designs."
Ann. Math. Statist., 23, 57-71.
"Some Relations Among the Blocks of Symmetrical GroupDivisible Designs." Ann. Math. Statist., 23, 602-609.
CONNOR, W. S. (1958). "The Uniqueness of the Triangular Association Scheme."
CONNOR, W. S. (1952b).
Ann. Math. Statist., 29, 262-266.)
S., and W. H. CLATWORTHY (1954). "Some Theorems for Partially

Balanced Designs." Ann. Math. Statist., 25, 100-112.
CORNFIELD, J., and J. W. TUKEY (1956). "Average Values of Mean Squares in
Factorials." Ann. Math. Statist., 27, 907-949.
CRAIG, A. T. (1943). "Note on the Independence of Certain Quadratic Forms."
CURNOW, R. N. (1963). "Sampling the Diallel Cross." Biometrics, 19, 287-306.
DANIEL, C. (1956). "Fractional Replication in Industrial Research." Proc. Third
CONNOR, W.
Berkeley Symposium on Mathematical Statistics and Probability, 5, 87-98.
"Use of Half-Normal Plots in Interpreting Factorial Two Level

Experiments." Technometrics, 1, 311-342.
DANIEL, C. (1962). "Sequences of Fractional Replicates in the 2P -9 Series." J.
DANIEL, C. (1959).
Amer. Statist. Assoc., 57, 403-429.

A. (1963). "The Structure of Cyclic Paired-Comparison Designs."
J. Austral. Math. Soc., 3, 117-127.
DAVID, H. A., and F. W. WOLOCK (1965). "Cyclic Designs." Ann. Math. Statist., 36,
1526-1534.
DAVIES, O. L. (1954). Design and Analysis of Industrial Experiments. Oliver and
DAVID, H.
Boyd, London.
DEBAUN, R. (1959).
"Response Surface Designs for Three Factors at Three Levels."
Technometrics, 1, 1-8.
DRAPER, N. R. (1960a). "Third Order Rotatable Designs in Three Dimensions.'
DRAPER, N. R. (1960b). "A Third Order Rotatable Design in Four Dimensions.'
DRAPER, N. R. (1961). "Third Order Rotatable Designs in Three Dimensions: Some
Specific Designs." Ann. Math. Statist., 32, 910-913.
DRAPER, N. R. (1962). "Third Order Rotatable Designs in Three Factors: Analysis."
DRAPER, N. R., and H. SMLTH (1966). Applied Regression Theory. John Wiley and
Sons, New York.

DUNCAN, D. B. (1955). " Multiple Range and Multiple F-Tests." Biometrics, 11, 1-42.
FEDERER, W. T. (1955). Experimental Design: Theory and Application. The Mac-
millan Company, New York.

J. (1945). "The Fractional Replication of Factorial Arrangements."
Annals of Eugenics, 12, 291 -301.
FINNEY, D.
Bibliography 343
FINNEY, D. J. (1946). "Recent Developments in the Design of Field Experiments.
I. Split-Plot Confounding." J. Agri. Sci., 36, 56-62.
FISHER, R. A. (1925). Statistical Methods for Research Workers. Oliver and Boyd,
Edinburgh.
FISHER, R. A. (1940). "An Examination of the Different Possible Solutions of a
Problem in Incomplete Blocks." Annals of Eugenics, 10, 52-75.
FISHER, R. A. (1942). "The Theory of Confounding in Factorial Experiments in
Relation to the Theory of Groups." Annals of Eugenics, 11, 341-353.
FISHER, R. A. (1947). The Design of Experiments. (Fourth ed.) Oliver and Boyd,
Edinburgh.
FISHER, R. A., and F. YATES (1953). Statistical Tables for Biologica!, Agricultural,
and Medical Research. (Fourth ed.) Oliver and Boyd, Edinburgh.
Fox, M. (1956). "Charts of the Power of the F Test." Ann. Math. Statist., 27,
484-497.
FREEMAN, G. H. (1957). "Some Further Methods of Constructing Regular Group
Divisible Incomplete Block Designs," Ann. Math. Statist., 28, 479-487.
FYFE, J. L., and N. GILBERT (1963). "Partial Diallel Crosses." Biometrics, 19,
278-286.
GARDINER, D. A., A. H. E. GRANDAGE, and R. J. HADER (1959). "Third Order
Rotatable Designs for Exploring Response Surfaces." Ann. Math. Statist., 30,
1082-1096.
GooD, I. J. (1958). "The Interaction Algorithm and Practical Fourier Analysis."
J. Roy. Statist. Soc., Ser. B., 20, 361-372.
GooD, I. J. (1960). Addendum to "The Interaction Algorithm and Practical Fourier
Analysis," J. Roy. Statist. Soc., Ser. B., 22, 372-375.
GRAYBILL, F. A. (1961). An Introduction to Linear StatisticalModels. Vol. I. McGrawHill, New York.
GRAYBILL, F. A., and R. B. DEAL (1959). "Combining Unbiased Estimators."
Biometrics, 15, 543-550.
GRAYBILL, F. A., and G. MARSAGLIA. (1957). "Idempotent Matrices and Quadratic
Forms in the General Linear Hypothesis." Ann. Math. Statist., 28, 678-686.
GRAYBILL, F. A., and V. SESHADRI (1960). 1 On the Unbiasedness of Yates' Method
of Estimation Using Interblock Information." Ann. Math. Statist., 31, 786-787.
GRAYBILL, F. A., and D. L. WEEKS (1959). "Combining Interblock and Intrablock
Information in Balanced Incomplete Blocks." Ann. Math. Statist., 30, 799-805.
GRIFFING, B. (1956). "Concept of General and Specific Combining Ability in
Relation to Diallel Crossing Systems." Austral. J. Biol. Sci., 9, 463-493.
GRoss, H. D., E. R. Parvist, and G. A. Ahlgren (1953). "The Response of Alfalfa
Varieties to Different Soil Fertility Levels." Agron. J., 45, 118-120.
GRUNDY, P. M., and M. J. R. HEALY (1950). "Restricted Randomization and
Quasi-Latin Squares." J. Roy. Statist. Soc., Ser. B, 12, 286-291.
HALL, M., and W. S. CONNOR (1953). "An Embedding Theorem for Incomplete
Block Designs." Canad. J. Math., 6, 35-41.
HALL, M., and H. J. RYSER (1951). "Cyclic Incidence Matrices." Canad. J. Math.,
3, 495-502.
HAMMERSLEY, J. M. (1949). "The Unbiased Estimate and Standard Error of the
Interclass Variance." Me/ron, 15, 173-188.
HANANI, H. (1961). "The Existence and Construction of Balanced Incomplete
Block Designs." Ann. Math. Statist., 32, 361-386.
344 Bibliography
HANANI, H. (1965).
"A Balanced Incomplete Block Design." Ann. Math. Statist., 36,
711.
HARTLEY, H. O.
(1959). "Smallest Composite Designs for Quadratic Response

Surfaces." Biometrics, 15, 611-624.
HEALY, M. J. R. (1951). "Latin Rectangle Designs for 2n Factorial Experiments on
32 Plots." J. Agri. Sci., 41, 315-316.
HILL, W. J., W. G. HUNTER (1966). "A Review of Response Surface Methodology: A
Literature Survey." Technometrics, 8, 571-590.
HOFFMAN, A. J. (1960). "On the Uniqueness of the Triangular Association Scheme."
HOTELLING, H. (1944a). "Some Improvements in Weighing and Other Experimental
Techniques." Ann. Math. Statist., 15, 297-306.
HOTELLING, H. (1944b). "On a Matric Theorem of A. T. Craig." Ann. Math.
Statist., 15, 427-429.
HUNTER, J. S. (1960). "Sme Applications of Statistics to Experimentation."
Chemical Engineering Progress Symposium, Series 31, 56, 10-26.
JAMES, E., and T. A. BANCROFT (1951). "The Use of Half Plants in a Balanced
Incomplete Block Design in Investigating the Effect of Calcium, Phosphorus,
and Potassium, at Two Levels Each on the Production of Hard Seed in Crimson
Clover, Trifolium Incarnatum." J. Agron., 43, 96-98.
JAMES, G. S. (1952). "Notes on a Theorem of Cochran. "Proc. Camb. Philos. Soc.,
48, 443-446.
JOHN, J. A. (1966). "Cyclic Incomplete Block Designs." J. Roy. Statist. Soc., Ser. B.,
28, 345-360.
JOHN, P. W. M. (1961a). "Three-Quarter Replicates of 2 4 and 2 5 Designs." Biometrics, 17, 319-321.
JoHN, P. W. M. (1961b). "An Application of a Balanced Incomplete Block Design."
JOHN, P. W. M. (1962). "Three-quarter Replicates of 2n Designs." Biometrics, 18,
172-184.
JOHN, P. W. M. (1963). "Extended Complete Block Designs." Austral. J. Statist.,.
5, 147-152.
JOHN, P. W. M. (1964a). "Pseudo-Inverses in the Analysis of Variance." Ann. Math.
Statist., 35, 895-896.
JOHN, P. W. M. (1964b). "Blocking of 3(2n - k) Designs." Technometrics, 6, 371-376.
JOHN, P. W. M. (1 964c). "Balanced Designs with Unequal Numbers of Replicates."
Ann. Mth. Statist., 35, 897-899.
JOHN, P. W. M. (1966a). "On Identity Relationships for 2n ' Designs Having
Words of Equal Length." Ann. Math. Statist., 37, 1842-1843.
JOHN, P. W. M. (1966b). "Augmenting 2n -1 Designs." Technometrics, 8, 469480.
JOHN, P. W. M. (1966c). "An Extension of the Triangular Association Scheme to
Three Associate Classes." J. Roy. Statist. Soc., Ser. B., 28, 361-365.
JOHN, P. W. M. (1969). "Some Non-Orthogonal Fractions of 2n Designs." J. Roy.
Statist. Soc., Ser. B., 31, 270-275.
JOHN, P. W. M. (1970). "A Three-Quarter Replicate of the 4 3 Design." Austral. J. of
Statist., 12, 73-77.
JOHNSON, J. D. (1967). "Adding Partially Balanced Incomplete Block Designs."
Doctoral dissertation, University of California, Davis.
-
Bibliography 345
KAPADIA, C. H., and D. L. WEEKS (1964). "On the Analysis of Group Divisible
Designs." J. Amer. Statist. Ass., 59, 1217-1219.
KEMPTHORNE, O. (1952). The Design and Analysis of Experiments. John Wiley and
Sons, New York.
KEMPTHORNE, O. (1953). "A Class of Experimental Designs Using Blocks of Two
Plots." Ann. Math. Statist., 24, 76-84.
KEMPTHORNE, O. (1955). "The Randomization Theory of Experimental Inference."
J. Amer. Statist. Ass., 50, 946-967.
KEMPTHORNE, O. (1956a). "The Efficiency Factor of an Incomplete Block Design."

KEMPTHORNE, O. (1956b). "The Theory of the Diallel Cross." Genetics, 41,451-459.

KEMPTHORNE, 0., and R. N. CURNOW (1961). "The Partial Diallel Cross." Biometrics, 17, 229-250.
KEMPTHORNE, 0., and W. T. FEDERER (1948). "The General Theory of Prime Power
Lattice Designs." Biometrics, 4, Part 1, 54-79; Part II, 109-121.
KEULS, M. (1952). "The Use of the 'Studentized Range' in Connection with an
Analysis of Variance." Euphytica, 1, 112-122.
KISHEN, K. (1945). "On the Design of Experiments for Weighing and Making Other
Types of Measurements." Ann. Math. Statist., 16, 294-300.
KRUSKAL, W. H., and W. A. WALLIS (1952). " Use of Ranks in One Criterion Analysis
of Variance." J. Amer. Statist. Ass., 47, 583-621.
LANCASTER, H. O. (1954). "Traces and Cumulants of Quadratic Forms in Normal
Variables." J. Roy. Statist. Soc., Ser. B., 16, 247-254.
LUCAS, H. L. (1951). "Bias in Estimation of Error in Change-Over Trials with
Dairy Cattle." J. Agri. Sci., 41, 146-148.
LUCAS, H. L. (1957). "Extra-Period Latin Square Change-Over Designs." J. Dairy
Sci., 40, 225-239.
MADOW, W. G. (1940). "The Distribution of Quadratic Forms in Non-Central

Normal Random Variables." Ann. Math. Statist., 11, 100-103.
MARGOLIN, B. H. (1967). "Systematic Methods of Analyzing 2m3" Factorial Experiments with Applications." Technometrics, 9, 245-260.
MARGOLIN, B. H. (1969a). "Results on Factorial Designs of Resolution IV for the
2n and 2"3m Series." Technometrics, 11, 431-444.
MARGOLIN, B. H. (1969b). "Resolution IV Fractional Factorial Designs." J. Roy.
Statist. Soc., Ser. B., 31, 514-523.
MARGOLIN, B. H. (1969c). "Orthogonal Main Effect Plans Permitting Estimation of
All Two-Factor Interactions for the 2 1 3m Factorial Series of Designs." Technometrics 11, 747-762.
MESNER, D. M. (1956). "An Investigation of Certain Combinatorial Properties of

Partially Balanced Incomplete Block Designs and Association Schemes, with a
Detailed Study of Latin Square and Related Types." Doctoral dissertation,
Michigan State University.
MESNER, D. M. (1965). "A Note on the Parameters of PBIB Association Schemes."
MESNER, D. M. (1967). "A New Family of Partially Balanced Incomplete Block

Designs with Some Latin Square Design Properties." Ann. Math. Statist., 38,
571-582.
MIKHAIL, W. F. (1960). "An Inequality for Balanced Incomplete Block Designs."

346 Bibliography
MILLER, R. G., Jr. (1966). Simultaneous Statistical Inference. McGraw-Hill, New
York.
MooD, A. M. (1946). "On Hotelling's Weighing Problem." Ann. Math. Statist., 17,
432-446.
MURTY, V. N. (1961). "An Inequality for Balanced Incomplete Block Designs."
NAIR, K. R. (1952). "Analysis of Partially Balanced Incomplete Block Designs
Illustrated on the Simple Square and Rectangular Lattices." Biometrics, 8,122-155.
NAIR, K. R., and C. R. RAO (1942). "A Note on Partially Balanced Incomplete
Block Designs." Science and Culture, 7, 568-569.
NANDI, H. K. (1945). "On the Relation Between Certain Types of Tactical Configurations." Bull. Calcutta Math. Soc., 37, 92-94.
NATIONAL BUREAU OF STANDARDS (1957). "Fractional Factorial Experiment Designs
for Factors at Two Levels." Applied Mathematics Series, No. 48.
NATIONAL BUREAU OF STANDARDS (1959). "Fractional Factorial Experiment Designs
for Factors at Three Levels." Applied Mathematics Series, No. 54.
NATIONAL BUREAU OF STANDARDS (1961). "Fractional Factorial Designs for
Experiments with Factors at Two or Three Levels." Applied Mathematics Series,
No. 58.
NELDER, J. A. (1963). "Identification of Contrasts in Fractional Replication of 2n
Experiments." Applied Statistics, 12, 38-43.
NEWMAN, D. (1939). "The Distribution of the Range in Samples from a Normal
Population, Expressed in Terms of an Independent Estimate of the Standard
Deviation." Biometrika, 31, 20-30.
NEYMAN, J., et al. (1935). "Statistical Problems in Agricultural Experimentation."
J. Roy. Statist. Soc., Suppl., 2, 107 -154.
OGAWA, J. (1949). "On the Independence of Bilinear and Quadratic Forms of a
Random Sample from a Normal Population." Ann. Inst. Statist. Math. Tokyo, 1,
83-108.
OGAWA, J. (1959). "A Necessary Condition for Existence of Regular and Symmetrical Experimental Designs of Triangular Type, with Partially Balanced
Incomplete Blocks." Ann. Math. Statist., 30, 1063 -1071.
PALEY, R. E. A. C. (1933). "On Orthogonal Matrices." J. Math. Phys., 12, 311-320.
PAPATHANASIOU, G. A., K. J. LESSMAN, and W. E. NYQUIST (1966). "Evaluation of
Eleven Introductions of Crambe, Crambe Abyssinica Hochst." Agron. J., 58,
587-589.
PATTERSON, H. D. (1950). "The Analysis of Change-Over Trials." J. Agri. Sci., 40,
375-380.
PATTERSON, H. D. (1952). "The Construction of Balanced Designs for Experiments
Involving Sequences of Treatments." Biometrika, 39, 32-48.
PEARSON, E. S., and H. O. HARTLEY (1951). "Charts of the Power Function of the
Analysis of Variance Tests, Derived from the Non-Central F Distribution."
Biometrika, 38, 112-130.
PEARSON, E. S., and H. O. HARTLEY (1954). Biometrika Tables for Statisticians,
vol. 1, 176-177.
PEART, J. N. (1968). "Lactation Studies with Blackface Ewes and Their Lambs."
J. Agri. Sci., 70, 87-94.
PETERSEN, E. M. (1952). "Controlling Tobacco Sucker Growth with Maleic Hydrazide." Agron. J., 44, 332-334.
Bibliography 347
PLACKETT, R.
L. (1950). "Some Theorems in Least Squares." Biometrika, 37, 149-
157.
PLACKETT, R. L. (1960). Regression Analysis.
Oxford University Press, London.

L., and J. P. BURMAN (1946). "The Design of Optimum Multifactorial Experiments." Biometrika, 33, 305-325.
QUENOUILLE, M. H. (1953). The Design and Analysis of Experiment. Charles Griffin
and Company, London.
QUENOUILLE, M. H. (1955). "Checks on the Calculation of Main Effects and Interactions in a 2" Factorial Experiment." Annals of Eugenics, 19, 151-152.
RAESE, J. T., and A. M. DECKER (1966). "Yields, Stand Persistente and Carbohydrate Reserves of Perennial Grasses as Influenced by Spring Harvest Stage,
Stubbie Height and Nitrogen Fertilization." Agron. J., 58, 322-326.
RAGHAVARAO, D. (1959). "Some Optimum Weighing Designs." Ann. Math. Statist.,
PLACKETT, R.
30, 295-303.
RAGHAVARAO, D. (1960a). "A Generalization of Group Divisible Designs." Ann.
Math. Statist., 31, 756-765.
RAGHAVARAO, D. (1960b). "On the Block Structure of Certain PBIB Designs with
Two Associate Classes Having Triangular and
L2
Association Schemes." Ann.
Math. Statist., 31, 787-791.

RAGHAVARAO, D., and K. CHANDRASEKHARARAO (1964). "Cubic Designs." Ann.
Math. Statist., 35, 389-397.
RAO, C. R. (1946). "On Hypertubes of Strength d and a System of Confounding in
Factorial Experiments." Bull. Calcutta Math. Soc., 38, 67-78.

R. (1947a). "General Methods of Analysis for Incomplete Block Designs."
RAO, C.

R. (1947b). "Factorial Experiments Derivable from Combinatorial Arrangements of Arrays." J. Roy. Statist. Soc., Suppl., 9, 128-139.
RAO, C. R. (1961). "A Study of BIB Designs with Replications 11 to 15." Sankhya,
23, 117-127.
RAO, C. R. (1962). "A Note on Generalized Inverse of a Matrix with Applications to
Problems in Mathematical Statistics." J. Roy. Statist. Soc., Ser. B., 24, 152-158.
RAO, V. R. (1958). "A Note on Balanced Designs." Ann. Math. Statist., 29, 290-294.
RAYNER, A. A. (1967). "The Square Summing Check on the Main Effects and
Interactions in a 2' Experiment as Calculated by Yates's Algorithm." Biometrics,
23, 571-573.
ROLLINSON, D. H. L., K. W. HARKER, J. I. TAYLOR, and F. B. LEECH (1956). "Studies
RAO, C.
on the Habits of Zebu Cattle IV. Errors Associated with Recording Technique."
J. Agron. Sci., 46, 1 -5.
Roy, J. (1958). "On the Efficiency Factor of Block Designs." Sankhya, 19, 181-188.
Roy, J., and R. G. LAHA (1957a). "On Partially Balanced Linked Block Designs."
ROY, J., and R. G. LAHA (1957b). "Classification and Analysis of Linked Block
Designs." Sankhya 17, 115-132.
Roy, P. M. (1952). "A Note on the Resolvability of Balanced Incomplete Block
Designs." Bull. Calcutta Statist. Ass., 4, 130-132.
Roy, P. M. (1953). "Hierarchical Group Divisible Incomplete Block Designs with
m Associate Classes." Science and Culture, 19, 210-211.
SATTERTHWAITE, F. (1946). "An Approximate Distribution of Estimates of Variante
Components." Biometrics Bulletin, 2, 110 -114.
;
348 Bibliography
(1953). "A Method for Judging All Contrasts in the Analysis of
Variance." Biometrika, 40, 87-104.
SCHEFF, H. (1956a). "A `Mixed Model' for the Analysis of Variance." Ann. Math.
SCHEFF, H.
Statist., 27, 23-36.
"Alternative Models for the Analysis of Variance." Ann. Math.

Statist., 27, 251-271.
SCHEFF, H. (1959). The Analysis of Variance. John Wiley and Sons, New York.
SCHNEIDER, A. M., and A. L. STOCKETT (1963). "An Experiment to Select Optimum
Operating Conditions on the Basis of Arbitrary Preference Ratings." Chemical
Engineering Progress Symposium Series, No. 42, Vol. 59.
SCHUTZENBERGER, M. P. (1949). "A Non-Existence Theorem for an Infinite Family
of Symmetric Block Designs." Annals of Eugenics, 14, 286-287.
SESHADRI, V. (1963a). "Constructing Uniformly Better Estimators." J. Amer.
Statist. Ass., 58, 172-175.
SESHADRI, V. (1963b). "Combining Unbiased Estimators." Biometrics, 19, 163-170.
SHRIKHANDE, S. S. (1950). "The Impossibility of Certain Symmetrical Balanced
Incomplete Block Designs." Ann. Math. Statist., 21, 106-111.
SHRIKHANDE, S. S. (1952). "On the Dual of Some Balanced Incomplete Block
Designs." Biometrics, 8, 66-72.
SHRIKHANDE, S. S. (1959a). "On a Characterization of the Triangular Association
Scheme." Ann. Math. Statist., 30, 39-47.
SHRIKHANDE, S. S. (1959b). "The Uniqueness of the L 2 Association Scheme." Ann.
Math. Statist., 30, 781-798.
SINGH, N. K., and K. N. SINGH (1964). "The Non-Existence of Some Partially
Balanced Incomplete Block Designs with Three Associate Classes." Sankhya, 24,
239-250.
SMITH, C. A. B., and H. O. HARTLEY (1948). "The Construction of Youden Squares."
J. Roy. Statist. Soc., Ser. B., 10, 262-263.
SPRAGUE, G. F., and L. A. TATUM (1942). "General vs. Specific Combining Ability in
Single Crosses of Corn." J. Amer. Soc. Agron., 34, 923-932.
SPROTT, D. A. (1954). "A Note on Balanced Incomplete Block Designs." Canad.
Jour. Math., 6, 341-346.
SPROTT, D. A. (1956a). "A Note on Combined Interblock and Intrablock Estimation
in Incomplete Block Designs." Ann. Math. Statist., 27, 633-641. [Erratum vol. 27,
269 (1957)].
SPROTT, D. A. (1956b). "Some Series of Balanced Incomplete Block Designs."
Sankhya, 17, 185-192.
SPROTT, D. A. (1962). "A Listing of BIB Designs from r = 16 to 20." Sankhya, 24,
203-204.
STEEL, R. G. D. (1960). "A Rank-Sum Test for Comparing All Pairs of Treatments."
STEINER, J. (1853). "Eine Kombinatorische Aufgabe." Crelle's J. Reine und Angew.
Math., 45, 181-182.
STEVENS, W. L. (1939). "The Completely Orthogonalized Latin Square." Annals of
Eugenics, 9, 82-93.
TANDON, R. K. (1949). "The Response of Flax to Rates and Formulations of 2,4Dichlorophenoxyacetic Acid." Agron. J., 41, 213-218.
TANG, P. C. (1938). "The Power Function of the Analysis of Variance Tests with
Tables and Illustrations of Their Use." Statist. Research Memoirs, 2, 126-149.
SCHEFF, H. (1956b).
Bibliography 349
G. (1901). "Le problme de 36 officieurs." Compte Rendu de l'Assoc.
Franaise pour l'Avancement de Science Naturel, 2, 170-203.
TUKEY, J. W. (1949). "One Degree of Freedom for Non-Additivity." Biometrics, 5,
232-242.
TUKEY, J. W. (1953). "The Problem of Multiple Comparisons." Dittoed manuscript,
396 pp., Princeton University.
VAJDA, S. (1967a). "Patterns and Configurations in Finite Spaces." Griffin's
Statistical Monographs and Courses, No. 22. Charles Griffin and Company,
TARRY,
London.
(1967b). "The Mathematics of Experimental Design: Incomplete Block

Designs and Latin Squares." Griffin's Statistical Monographs and Courses, No. 23.
Charles Griffin and Company, London.
VANCE, F. P. (1962). "Optimization Study of Lube Oil Treatment by Process `X'."
VAJDA, S.
Proc. of Symp. on Application of Statistics and Computers to Fuel and Lubricant

Research Problems. Office of the Chief of Ordnance, U.S. Army, March 13-15,
1962.
VARTAK, M. N. (1959). "The Non-Existence of Certain PBIB Designs." Ann. Math.
Statist., 30, 1051-1062.
WEBB, S. R. (1968). "Non-Orthogonal Designs of Even Resolution." Technometrics, 10, 291-300.
WELCH, B. L. (1947). "The Generalisation of 'Student's' Problems When Several
Different Population Variances are Involved." Biometrika, 34, 28-35.
WELCH, B. L. (1956). "On Linear Combinations of Several Variances." J. Amer.
Statist. Ass., 51, 132-148.
WIDDOWSON, F. V., A. PENNY, and R. J. B. WILLIAMS (1966). "An Experiment
Measuring Effects of N, P, and K Fertilizers on Yield and N, P, and K Contents

of Grazed Grass." J. Agri. Sci., 67, 121-128.
WILK, M. B., and O. KEMPTHORNE (1955). "Fixed, Mixed, and Random Models."
and O. KEMPTHORNE (1956). "Some Aspects of the Analysis of

Factorial Experiments in a Completely Randomized Design." Ann. Math.
WILK, M. B.,
Statist., 27, 950-985.
B., and O. KEMPTHORNE (1957). "Non-Additivities in a Latin Square

Design." J. Amer. Statist. Ass., 52, 218-236.
WILLIAMS, E. J. (1949). "Experimental Designs Balanced for the Estimation of
Residual Effects of Treatments." Austral. J. Scientific Res., Ser. A., 2, 149-168.
WILLIAMS, E. J. (1950). "Experimental Designs Balanced for Pairs of Residual
Effects." Austral J. Scientific Res., Ser. A., 3, 351-363.
WOODMAN, R. M., and D. A. JOHNSON (1946). "The Nutrition of the Carrot. III
Grown in Gravel Soil." J. Agri. Sci., 36, 10-17.
YANG, C. H. (1966). "Some Designs for Maximal (+ 1, -1) Determinant of Order
n = 2 (mod 4)." Mathematics of Computation, 20, 147-148.
YANG, C. H. (1968). "On Designs of Maximal (+ 1, -1) Matrices of Order n = 2
(mod 4)." Mathematics of Coniputation, 22, 174-180.
YATES, F. (1933). "The Analysis of Replicated Experiments when the Field Results
are Incomplete. Emp. Jour. Exp. Agr., 1, 129-142.
YATES, F. (1936). Incomplete Randomized Blocks. Annals of Eugenics, 7, 121-140.
YATES, F. (1937). "The Design and Analysis of Factorial Experiments." Imperial
Bureau of Soil Science, Harpenden, England.
WILK, M.
350 Bibliography
YATES, F. (1940). "The Recovery of Interblock Information in Balanced Incomplete
Block Designs." Annals of Eugenics, 10, 317-325.
YATES, F. (1948). "Discussion on a paper by F. Anscombe." J. Roy. Statist. Soc.,
Ser. A., 111, 204-205.
YOUDEN, W. J. (1951). "Linked Blocks, A New Class of Incomplete Block Designs."
(Abstract.) Biometrics, 7, 124.

YOUDEN, W. J., and J. S. HUNTER (1955). "Partially Replicated Latin Squares."
Biometrics, 11, 399-405.
ZELEN, M. (1954). "A Note on Partially Balanced Designs." Ann. Math. Statist., 25,
599-602.
Index
Addelman, S., 178, 184, 187, 339

Adding fractions, 167
Adjusted block totals, 230
Adjusted treatment totals, 230
Aia, M. A., 205, 213, 339
Aitken, A. C., 31, 339
Alias set, 150
Alias subgroup, 150
Allan, R. E., 60, 339
Analysis of covariance, 60-64
Anderson, R. L., 145, 340
Approximations for distributions
F test in three-way layout, 92-94
t test with unequal variances, 6
Association matrices, 314
Association schemes, see Partially balanced association schemes
Atiqullah, M., 245, 283, 339
Autrey, K. M., 117, 342
Balanced incomplete block designs; 220,
268-87
combined estimates, 235-242
efficiency of, 225, 247
incidence matrix of, 220, 269
interblock estimates, 234
intrablock analysis, 223
orthogonal series of, 221, 270
resolvable, 226
symmetric designs, 220, 269, 276277, 281-283
unreduced designs, 270, 307
Balanced lattices, 240, 307
Bancroft, T. A., 240, 344
Banerjee, K. S., 28, 339
Baumert, L., 184, 339
Behnken, D. W., 199, 341
Bennett, C. A., 96, 339
Bhattacharya, K. N., 277, 282, 285,
292, 295, 298, 299, 323, 339, 340
Binary designs, 227
Binet, F. E., 145, 340
Block intersection, 281
Block section, 281
Blocking factors, 158
Bose, R. C., 10, 24, 111, 114, 144, 178,
221, 223, 227, 249, 250, 252, 256,
265, 269, 281, 282, 283, 285, 286,
288, 291, 292, 293, 294, 295, 298,
299, 300, 301, 310, 312, 314, 315,
316, 323, 340, 341
Box, G. E. P., 14, 46, 150, 151, 193,
194, 197, 199, 211, 213, 214, 341
Brownlee, K. A., 139, 341
Bruck, R. H., 114, 341
Bulmer, M. G., 54, 341
351
352 Index
Burman, J. P., 172, 178, 184, 347
Burton, R. C., 139, 140, 341
Bush, K. A., 178, 340
Calvin, L. C., 287, 341
Cannon, C. Y., 117, 342
Canonical equations, 208-209, 211-212
Central composite designs, 204
Chandrasekhararao, K., 319, 347
Chang, L. C., 301, 305, 308, 309, 341
Change-over designs, 115-121
Checks in calculations, 47; see also
Yates' algorithm
Chowla, S., 269, 276, 283, 341
Classification of fractional factorials,
151
Clatworthy, W. H., 252, 265, 288, 298,
301, 304, 305, 306, 307, 312, 314,
315, 316, 340, 341, 342
Cochran, W. G., 6, 33, 117, 341, 342
Cochran's theorem, 33, 71
Combined intrablock and interblock estimates, 235-245
computational formulae, 238-242
in partially balanced designs, 259260
maximum likelihood estimates (Rao),
242-245
Combining ability, see Diallel cross experiments
Complementary designs, 270
Completely randomized experiment, 4145
Components of variance, confidence intervals for, 54
Computer programs, use of, 102-103
Confounding, 12, 132-245
in fractional factorials, 158
in ,2n designs, 133-143
in 3' designs, 143
in 4n designs, 144-145
in other designs, 145
partial confounding, 134-135
with factors at more than three levels,
144-45
Connected designs, 233
Connor, W. S., 139, 140, 181, 190, 252,
281, 283, 292, 293, 301, 304, 314,
315, 340, 341, 342
Cornfield, J., 78, 342

Covariance, analysis of, 60-64
Cox, G. M., 6, 342
Craig, A. T., 31, 342
Cross polytope, 204
Cuboctahedron, 199
Curnow, R. N., 262, 342, 345
Cyclic designs, 266-267, 275-287, 322326
generalized, 324-326
Dairy cattle, experiments with, 115-121
Daniel, C., 149, 151, 154, 155, 342
David, H. A., 322, 323, 326, 342
Davies, 0. L., 194, 213, 342
Dead factors, 160
Dead subscripts, 97
Deal, R. B., 242, 343
Debaun, R., 199, 218, 342
Defining contrast, 134
Defining contrast subgroup, 137, 139140, 150
Degrees of freedom, calculation in multifactorial experiments, 97
Degree of freedom efficiency, 155
Derivation, see block intersection
Diallel cross experiments, 81-84, 262265
Difference sets, 275
Dishwashing, 226
Double confounding, 140-143
Draper, N. R., 21, 213, 342
Duality, 290-291
Duncan, D. B., 45, 342
Duncan's multiple range test, 45
Effects
main, 10
fixed, 11
random, 11
Efficiency of fractional factorials, 155
Efficiency of incomplete block designs,
10, 225, 246-247,266
Elementary symmetric designs, 291-292
Equireplicate designs, 228
Estimable functions, 24
Euler's conjecture, 114
Evolutionary operation, 213
calculations for, 214-217
Index 353
Expected mean squares,
general rules for deriving, 96-98
multi-factor crossed classification, 9091
Extended incomplete block designs, 248
F test
approximation, 92-94
bias in with missing plots, 60
power of, 7, 44, 53
Factorial experiments, 10
Federer, W. T., 45, 260, 261, 342, 345
Finite projective geometrics, 271-274,
300, 310-312
Finney, D. J., 12, 148, 175, 178, 193,
342, 343
First order designs, 194-196
Fisher, R. A., 2, 4, 136, 180, 221, 265,
268, 269, 282, 343
Fisher's inequality, 221, 269
Five-letter plans, 151
Foldover designs, 156-157, 173-174
Four-letter plans, 151
Fox, M., 8, 44, 343
Fractional factorial designs, 12, 148192
Franklin, N. J., 96, 339
Freeman, G. H., 292, 295, 297, 343
Fundamental theorems of differences,
277-280
for group divisible designs, 299-300
Fyfe, J. L., 262, 343
Gardner, D. A., 213, 343
Gauss-Markoff theorem, 34-35
Gilbert, N., 262, 343
Goldsmith, R. L., 205, 213, 339
Golomb, S. W., 184, 339
Good, I. J., 128, 343
Graeco-Latin squares, 110
as fractional factorials, 111, 148, 178,
179, 189, 191-192
Grandage, A. H. E., 213, 343
Graybill, F. A., 8, 27, 28, 76, 230, 238,
242, 343
Griffing, B., 81, 262, 343
Group divisible designs, 252-254, 292301
classification, 252, 293
intrablock analysis, 253-254

in partial diallel cross, 264
regular, 295-301
semiregular, 293-295
singular, 293
Group representation in 2n designs, 136
Grundy, P. M., 142, 343
Hademard matrices, 185
Hademard product of vectors, 173
Hader, R. J., 213, 343
Half-plaid squares, 142
Hall, M., Jr., 184, 277,, 281, 283, 304,
339, 343
Hammersley, J. M., 54, 343
Hanani, H., 283, 343, 344
Hartley, H. 0., 8, 44, 47, 205, 216, 277,
344, 346, 348
Healy, M. J. R., 142, 175, 343, 344
Hexagon design, 207
Hierarchic designs, see Nested designs
Hill, W. J., 194, 344
Hoffman, A. J., 301, 344
Hotelling, H., 31, 172, 344
Hunter, J. S., 122, 150, 151, 194, 197,
199, 203, 214, 341, 344, 350
Hunter, W. G., 194, 344
Hyper-Graeco-Latin squares, 111
as fractional factorials, 111, 148
Idempotent matrices, 15, 31, 335
Incident matrix, see Incomplete block
designs
Incomplete block designs, 9, 219-328
efficiency of, 10, 225, 247, 266, 289
incidence matrix of, 220, 269, 289
interblock analysis, 233-235
intrablock analysis, 228-233
see also balanced incomplete block
designs; Partially balanced incomplete block designs
Interaction in two-way layout, 70-80
single degree of freedom for, 76
Interblock estimates, 233-235
Intrablock analysis, 228-233
balanced incomplete block designs,
223-226
general balanced design, 228-233,
245-247
354 Index
Intrablock analysis (continued)
group divisible designs, 253-254
PBIB designs, 256-259
Intrablock subgroup, 138
James, E., 240, 344
James, G. S., 28, 344
John, J. A., 266, 322, 323, 326, 344
John, P. W. M., 28, 139, 161, 163, 165,
167, 170, 192, 226, 246, 248, 320,
344
Johnson, J. D., 321, 344
Kapadia, C. H., 259, 345
Kelly, B. K., 139, 341
Kempthorne, 0., 5, 8, 10, 27, 78, 81,
109, 187, 189, 230, 247, 260, 261,
262, 265, 289, 322, 339, 345
Kempthorne's conjecture, 10, 247, 265,
289
Keuls, M., 45, 345
Kishen, K., 144, 172, 340, 345
Kruskal, W. H., 46, 345
Lancaster, H. 0., 28, 31, 345
Laha, R. G., 290, 347
Latin rectangles, 142-143
Latin squares, 105-121
experiments using several, 114-121
orthogonal sets, 111-114, 221-222,
270
Latin square designs, 12, 105-121
analysis of, 106-107
limitations of, 108-109
use as fractional factorials, 107
Lattice designs, 240, 260-262, 271, 307
Least significant difference, 44, 226
Legendre symbol, 185, 276
Leslie, R. T., 145, 340
Linear models, 18
Linked block designs, 290-291, 302303, 312
singly linked blocks, 291, 302, 312
triangular doubly linked blocks, 303
Liu, C. W., 301, 309, 341
Liu, W. R., 301, 305, 308, 309, 341
Live factors, 160
Live subscripts, 97
Loraine, P. K., 139, 341
Lucas, H. L., 117, 345
Madow, W. G., 33, 345

Main effects, 10, 42, 67, 125
Main effects clear plans, 151
Main effects only designs, 151
Margolin, B. H., 160, 172, 173, 174,
190, 345
Marsaglia, G., 28, 343
Matrices, idempotent, 15, 31, 335
Measure polytope, 204
Mesner, D. M., 256, 306, 312, 316, 340,
345
Mikhail, W. F., 227, 345
Miller, R. G., Jr., 45, 46, 346
Minimal designs of resolution IV, 173
Missing plot formulae, 37
Latin square designs, 121
randomized complete block design, 59
2'n designs, 146
Mixed models, 10, 77-80
Mood, A. M., 172, 346
Mooney, R. W., 205, 213, 339
Motorcycle racing, 222
Multiple comparisons, 44-47
Murty, V. N., 227, 346
Nair, K. R., 10, 223, 250, 251, 288,
300, 316, 340, 346
Nandi, H. K., 283, 346
National Bureau of Standards, 158,
181, 190, 346
Nelder, J. A., 160, 346
Nested designs, 80, 103
Nested factors, 94
Nested fractions, 170-172
Newman, D., 45, 346
Newman-Keuls test, 45
Neyman, J., 105, 346
Noncentral X 2 distribution, 30-31
Octagonal designs, 203, 208
Octahedron, 199
Ogawa, J., 31, 305, 346
Omitting 2n k fractions, 161
Optimal designs, 195
Orthogonal arrays, 178
Orthogonal blocking, 206-208
Orthogonal Latin squares, sets of, 111114, 221-222, 270
construction of, 111-114
-
Index 355
Orthogonal main effects plans, 184
Orthogonal polynomials, 50
Orthogonal response surface designs,
197
Orthogonality of factors in two-way layout, 68-70
Paley, R. E. A. C., 185, 346
Parker, E. T., 114, 341
Partial confounding, 134
Partial geometry, 311
Partially balanced association schemes,
250
cubic, 319
cyclic, 255, 315
group divisible, 252, 292-301
hierarchic GD, 317
Latin square, 255, 305-308
L.(n,m), 327
LS3, 321
negative Latin square, 256, 306, 312
rectangular, 318
T.(m), 320
triangular, 254, 301-305, 320
Partially balanced incomplete block designs, 10, 223, 250-267, 288-328;
see also the various schemes
choice of, 266

combined estimates, 259
intrablock analysis, 256
Patterson, H. D., 117, 346
Pearson, E. S., 8, 44, 47, 216, 346
Peart, J. N., 346
Petersen, E. M., 346
Permutation tests, 5
Plackett, R. L., 27, 170, 172, 178, 184,
231, 347
Plackett and Burman designs, 172, 178,
184-187,195
Plaid squares, 142
Polygonal designs, 201-203
Pooling sums of squares, 86
Power calculations, 7, 44, 53
Principal block, 138
Principal fraction, 149
Proper designs, 227, 246
Proportional frequencies, condition of,
68-70, 187
Pseudo inverses, 223
Quadratic forms, 15, 17, 332

distribution of, 17, 29-31
independence of, 31
rank of, 17
Qualitative factors, 48, 124
Quantitative factors, 48, 124
Quasi-Latin squares, 140, 158, 260
Quenouille, M. H., 129, 347
Raese, J. T., 347
Raghavarao, D., 172, 303, 306, 308,
317, 319, 347
Random effects, 11, 52, 76
Randomization, 4, 5
Randomization tests, 5
Randomized complete blocks, 9, 55-60
analysis of covariance in, 63
missing plots in, 59
Rao, C. R., 23, 178, 236, 242, 245, 251,
256, 269, 316, 346, 347
Rao, V. R., 246, 347
Rayner, A. A., 129, 347
Recovery of interblock information,
see Interblock estimates
Reduced intrablock equations, 229
Replication, 4, 87
Residual effects, 117
designs for experiments with, 120
Residuation, see Block section
Resolution of a fractional factorial, 151
Resolvable designs, 226, 240, 270, 286,
299
Ridge surface, 213
Rising ridge, 213
Response surfaces, 14, 193-213
Rollingson, D. H. L., 68, 347
Rotatability, 197
Roy, J., 10, 247, 290, 347
Roy, P. M., 227, 293, 317, 318, 347
Ryser, H. J., 114, 269, 276, 277, 283,
341, 343
Saddle surface, 213
Sample size, calculation of, 7
Satterthwaite, F., 93, 347
Saturated fractions, 155, 172
Scheff, H., 2, 5, 6, 8, 27, 35, 45, 54,
65, 78, 96, 109, 231, 348
356 Index
Scheff's S statistic, 45, 226
Schneider, A. M., 200, 348
Schtzenberger, M. P., 269, 348
Screening experiments, 13, 149
Second order designs, 196
Seshadri, V., 238, 242, 343, 348
Shimamoto, T., 250, 252, 256, 316, 340
Shrikhande, S. S., 114, 252, 265, 269,
288, 291, 292, 295, 298, 299, 301,
306, 312, 316, 323, 340, 348
Singh, K. N., 320, 327, 348
Singh, N. K., 320, 327, 348
Single degrees of freedom, 48-52, 7376
Tukey's test for interaction, 76
Smith, C. A. B., 277, 348
Smith, H., 21, 342
Solution matrices, 24
general methods of obtaining, 26-28
for reduced intrablock equations,
230-233
Split plots, 11, 98, 142
Split-split plots, 101
Sprague, G. F., 81, 348
Sprott, D. A., 245, 249, 269, 283, 284,
348
Standard order, 128, 132
Steel, R. G. D., 46, 348
Steepest ascent, 208
Steiner, J., 283, 348
Stevens, W. L., 113, 348
Stockett, A. L., 200, 348
Studentized range, 45
Student's t distribution, approximation
with unequal variances, 6
Subdivision of sums of squares, 73
for 3n design, 130
Sums of squares, general rules for deriving, 96
subdivision, 73
Tanden, R. K., 101, 348
Tang, P. C., 44, 348
Tang's tables, 8, 44
Tarry, G., 114, 349
Tatum, L. A., 81, 348
Three-letter plans, 151
Three quarter replicates, 163
Transformations to stabilize variance,

7
Triads, method of, 304
Two f.i. clear plans, 151
Tukey, J. W., 45, 76, 78, 342, 349
Tukey's test for treatment differences,
45, 226
Unequal variances, 6
Uniqueness of association schemel,
-k 2 , 306
triangular, 301
Vajda, S., 312, 349
Vance, F. P., 182, 349
Variance efficiency, 155
Vartak, M. N., 318, 349
Wallis, W. A., 46, 345
Webb, S. R., 173, 174, 349
Webb's conjecture, 173
Weeks, D. L., 238, 259, 343, 345
Weighing designs, 172
Weiner, S., 145
Welch, B. L., 6, 93, 349
Whole plots, 11, 98
Widdowson, F. V., 349
Wilk, M. B., 5, 78, 109, 349
Williams, E. J., 120, 349
Wilson, K. J., 14, 151, 193, 194, 341
Wishart, J., 60, 339
Wolock, F. W., 322, 342
Woodman, R. M., 91, 349
Yang, C. H., 172, 349
Yates, F., 2, 12, 37, 74, 98, 123, 127,
135, 141, 142, 145, 220, 233, 238,
242, 260, 265, 268, 270, 282, 343,
349, 350
Yates' algorithm, 127, 153
checks on calculations, 128-130
extension to 3n designs, 132
Youden, W. J., 122, 290, 350
Youden square, 277
Youle, P. V., 211, 213, 341
Zebu cattle, 68
Zelen, M., 181, 293, 350
Zero-sum column vectors, 185

PWJohn

Uploaded by

Copyright:

Available Formats

PWJohn

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PWJohn

Uploaded by

Copyright:

Available Formats

Downloaded 12/11/12 to 129.128.207.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.

James M. Ortega, Numerical Analysis: A Second Course

F. H. Clarke, Optimization and Nonsmooth Analysis

Charles L Lawson and Richard J. Hanson, Solving Least Squares Problems

Richard E. Barlow and Frank Proschan, Mathematica) Theory of Reliability

Peter W. M. John, Statistica! Design and Analysis of Experiments

Copyright 1998 by the Society for Industrial and Applied Mathematics.

John, Peter William Meredith.

Preface to the Classics Edition

Chapter 2. Linear Models and Quadratic Forms

Chapter 3. Experiments with a Single Factor

Chapter 4. Experiments with Two Factors

The Two-Way Layout with One Observation per Cell

Chapter 5. Experiments with Several Factors

5.1 Error Terms and Pooling

An Experiment with Crossed and Nested Factors

Chapter 6. Latin Square Designs

Chapter 7. Factors with Two or Three Levels; Confounding

7.1 Factors at Two Levels

Chapter 8. Fractions of 2" Factorial Designs

The 2n k Fractional Factorials

8.8 Analyzing 2n - k Fractions

Chapter 9. Fractional Factorials with More Than Two Levels

Chapter 10. Response Surfaces

10.1 First-Order Designs

Central Composite Designs

10.6 Orthogonal Blocking

10.8 The Method of Steepest Ascent

Chapter 11. Incomplete Block Designs

Balanced Incomplete Block Designs

11.5 The General Incomplete Block Design

Chapter 12. Partially Balanced Incomplete Block Designs

The Group Divisible Association Scheme

Chapter. 13. The Existence and Construction of Balanced Incomplete Block

The Orthogonal Series

14.1 Duality and Linked Block Designs

Chapter 15. Additional Topics in Partially Balanced Designs

Appendix. Matrices and Quadratic Forms

decade to first- and second-year graduate students in the statistics department at

Preface to the Classics

When I signed my contract with Macmillan in 1966, some, but by no means

Preface to the Classics Edition

to take hours, and without the problems of making a charge to a budget.

Preface to the Classics Edition

3. The growth of experimental design in industry

Preface to the Classics Edition

the agronomists. They could think of series of small experiments: running

4. Fixed, mixed, and random models

Preface to the Classics Edition

4.2. Step 5r (the restricted rule)

4.3. Step 5u (the unrestricted rule)

an interaction term only if all its live subscripts belong to

Preface to the Classics Edition

Preface to the Classics Edition

6. Incomplete block designs

Preface to the Classics Edition

to my work at Chevron. Now I also wanted the freedom to work on any

Preface to the Classics Edition

References in the Preface

C. (1982). "Developments in Design of Experiments." Int. Statist.