a

V43
l^iMiM
bsKt'P^fi ipipl>J:.-|

M
Jy;j i

^1 i|;i sMi
 Digitized by the Internet Archive
in 2017 with funding from
China-America Digital Academic Library (CADAL)

https://archive.org/details/fundamentalconceOOhick_0
 /

/•

■*

.4
■-#

vW

• a

4
-i*'"

•'*. ■« ,

H.

r
Fundamental
Concepts in
the Design of
Experiments
Fundamental
Concepts in
the Design of
Experiments
Third Edition

CHARLES R. HICKS
Professor of Statistics and Education

Purdue University

Saunders College Publishing

Fort Worth Chicago San Francisco
Philadelphia Montreal Toronto
London Sydney Tokyo
Library of Congress Cataloging in Publication Data

Hicks, Charles Robert, 1920-

Fundamental concepts in the design of experiments.

Bibliography: p. 377
Includes index.
1. Experimental design. I. Title.
QA279.H53 1982 001.4'34 82-12062
ISBN □-□3-Dtil70b-S

Copyright © 1982 by Saunders College Publishing.

All rights reserved. No part of this publication may
be reproduced or transmitted in any form or by any
means, electronic or mechanical, including photocopy,
recording or any information storage and retrieval
system, without permission in writing from the
publisher.

Requests for permission to make copies of any part

of the work should be mailed to; Permissions, Holt,
Rinehart and Winston, Inc.
Orlando, Florida 32887.
Printed in the United States of America
3 4 5 038 9 8

Holt, Rinehart and Winston, Inc.

The Dryden Press
Saunders College Publishing
 Contents

Preface to the Third Edition xi

Preface to the Second Edition xiii
Preface to the First Edition xv

/
The Experiment, the Design, and the Analysis 1
1.1 INTRODUCTION TO EXPERIMENTAL DESIGN 1
1.2 THE EXPERIMENT 2
1.3 THE DESIGN 4
1.4 THE ANALYSIS 6
1.5 SUMMARY IN OUTLINE 6
1.6 EXAMPLES 7
PROBLEMS 14

2
Review of Statistical Inference 15
2.1 INTRODUCTION 15
2.2 ESTIMATION 17
2.3 TESTS OF HYPOTHESES 19
2.4 THE OPERATING CHARACTERISTIC CURVE 21
2.5 HOW LARGE A SAMPLE? 25
2.6 APPLICATION TO TESTS ON VARIANCES 25
2.7 APPLICATION TO TESTS ON MEANS 27
2.8 APPLICATION TO TESTS ON PROPORTIONS 30
PROBLEMS 32

Single-Factor Experiments with No Restrictions on Randomization 36

3.1 INTRODUCTION 36
3.2 ANALYSIS OF VARIANCE RATIONALE 38
3.3 AFTER ANOVA—WHAT? 46
3.4 TESTS ON MEANS 47
3.5 CONFIDENCE LIMITS ON MEANS 54

V
vi Contents

3.6 COMPONENTS OF VARIANCE 55

3.7 SUMMARY 59
PROBLEMS 61

4
Single-Factor Experiments—Randomized Block and
Latin Square Designs 66
4.1 INTRODUCTION 66
4.2 RANDOMIZED COMPLETE BLOCK DESIGN 67
4.3 ANOVA RATIONALE 72
4.4 MISSING VALUES 73
4.5 LATIN SQUARES 76
4.6 INTERPRETATIONS 78
4.7 GRAECO-LATIN SQUARES 79
4.8 EXTENSIONS 80
4.9 SUMMARY 80
PROBLEMS 80

5
Factorial Experiments 86
5.1 INTRODUCTION 86
5.2 FACTORIAL EXPERIMENTS 91
5.3 INTERPRETATIONS 94
5.4 COMPUTER PROGRAM 96
5.5 ANOVA RATIONALE 98
5.6 REMARKS 103
5.7 SUMMARY 104
PROBLEMS 105

6
2^ Factorial Experiments 112
6.1 INTRODUCTIONf 112
6.2 2^ FACTORIAL 112
6.3 2^FACTORIAL 120
6.4 2^—REMARKS 123
6.5 SUMMARY 124
PROBLEMS 125

7
Qualitative and Quantitative Factors 129
7.1 INTRODUCTION 129
7.2 LINEAR REGRESSION 130
7.3 CURVILINEAR REGRESSION 137
7.4 ORTHOGONAL POLYNOMIALS 139
7.5 COMPUTER PROGRAM 144
7.6 QUALITATIVE AND QUANTITATIVE FACTORS 144
7.7 TWO FACTORS—ONE QUALITATIVE, ONE QUANTITATIVE 145
7.8 TWO FACTORS—BOTH QUANTITATIVE 152
 Contents

7.9 SUMMARY 159

PROBLEMS 159

a
Multiple Regression 164
8.1 INTRODUCTION 164
8.2 THE MULTIPLE REGRESSION MODEL 165
8.3 COMPUTER PROGRAMS 175
8.4 REMARKS 176
PROBLEMS 177

9
3^ Factorial Experiments 187
9.1 INTRODUCTION 187
9.2 3^ FACTORIAL 187
9.3 3^ FACTORIAL 198
9.4 SUMMARY 207
PROBLEMS 207

lO
Fixed, Random, and Mixed Models 210
10.1 INTRODUCTION 210
10.2 SINGLE-FACTOR MODELS 211
10.3 TWO-FACTOR MODELS 212
10.4 EMS RULES 214
10.5 EMS DERIVATIONS 218
10.6 THE PSEUDO-F TEST 221
10.7 REMARKS 223
PROBLEMS 224

II
Nested and Nested-Factorial Experiments 227
11.1 INTRODUCTION 227
11.2 NESTED EXPERIMENTS 228
11.3 ANOVA RATIONALE 232
11.4 NESTED-FACTORIAL EXPERIMENTS 233
11.5 COMPUTER PROGRAMS 238
11.6 REPEATED-MEASURES DESIGN AND NESTED-FACTORIAL
EXPERIMENTS 239
11.7 SUMMARY 244
PROBLEMS 244

Experiments of Two or More Factors—Restrictions on Randomization 252

12.1 INTRODUCTION 252
12.2 FACTORIAL EXPERIMENT IN A RANDOMIZED BLOCK
DESIGN 252
viii Contents

12.3 FACTORIAL EXPERIMENT IN A LATIN SQUARE DESIGN 258

12.4 REMARKS 259
12.5 SUMMARY 260
PROBLEMS 261

Factorial Experiment—Splot-Plot Design 265

13.1 INTRODUCTION 265
13.2 A SPLIT-SPLIT-PLOT DESIGN 273
13.3 SUMMARY 277
PROBLEMS 277

14
Factorial Experiment—Confounding in Blocks 281
14.1 INTRODUCTION 281
14.2 CONFOUNDING SYSTEMS 282
14.3 BLOCK CONFOUNDING—NO REPLICATION 285
14.4 BLOCK CONFOUNDING WITH REPLICATION 290
14.5 SUMMARY 299
PROBLEMS 300

15
Fractional Replication 303
15.1 INTRODUCTION 303
15.2 ALIASES 304
15.3 FRACTIONAL REPLICATIONS 306
15.4 SUMMARY 314
PROBLEMS 315

lO
Miscellaneous Topics 318
16.1 INTRODUCTION 318
16.2 COVARIANCE ANALYSIS 318
16.3 RESPONSE-SURFACE EXPERIMENTATION 328
16.4 EVOLUTIONARY OPERATION (EVOP) 341
16.5 ANALYSIS OF ATTRIBUTE DATA 351
16.6 RANDOMIZED INCOMPLETE BLOCKS—RESTRICTION ON
EXPERIMENTATION 355
16.7 YOUDEN SQUARES 360
PROBLEMS 362

17
Summary and Special Problems 366

Glossary of Terms 373

References 377
 Contents

Statistical Tables 380

TABLE A. AREAS UNDER THE NORMAL CURVE 380
TABLE B. STUDENT’S t DISTRIBUTION 382
TABLE C. CHI SQUARE 383
TABLE D. F DISTRIBUTION 384
TABLE E. STUDENTIZED RANGES 390
TABLE F. COEFFICIENTS OF ORTHOGONAL POLYNOMIALS 392

Answers to Odd-Numbered Problems 393

Index 421
 r[

' » 4

t y '»' V' ■ ‘ V
4
'■^* ..f,. -. .■/ «>•
•s V 1
. - / JV, /' * •
.
■ "'■ ■ .V-, - ■'•- ' V'-';T>;- • ■ ' ' " '» ’ ^

^ V

W‘
. :■ 4 ;' V

*>0f ,/'■' : ♦S' f. ,t:w<f.‘‘^''''- •'^''

. ' i
f •- - vy .

■* i tw--
•'!,
i

H>
Vf

‘ ‘•‘ ' ' 4 ' - ■

'■■» •- * .■ - "J ‘
■ i - » ■• ; . • ■ *,

r^- ■-'•■
iTj?.. ^ ■> --. y
-f ?»)
. • V'

. '1 .
•* •
V
i-rr "}*
4-* *- . ; '
1 ^

fS ■ -St
. V -■ ■ I’ ■
«■' :'•» T
.i •
Kt "I
1 *' t’*' 1 ■■ i.*i’ V '' ■■ -:i-*
-I *
, V'..ii ’ ** •i - ■'
iV' I , I ‘ ‘i >5i ^ .f /■-!» ■ ■'■< :.<*
m.' '>.1 ’■f ■: -
Hp< a
■'■ K . ?j#.'
“•It ■ "■

f■ - i .:«•
I

:^V ^ ^ ••■*'
I * .■ »■? ' - V': ‘
f • < f
■-v»
■t.V
f.. s*'- • .
.■

^ * 1 i
^1.
'{♦ ■ ».
»• t ^s',1 t -^^.V k

««
. ' f 1^^'* TUN'* ' ** , -::
V.... i - 1 ' .' o. •
■JU. ►
■ ■ . 4.*'> -W > • ' >t< > V• » ' ,

r-v ^. <■ ‘ .
- - •'■ i ^ '• , ~ ,j .Z S \^*V'■ ,'
■ • .'v- V'
- t ‘ ■ '
^■■V . J ?•' .'jy ... -'V .■ ^ '
/ { gt ;fi 11
'y/'*A •-A ’ * / ^ '* 1 '' '
K

■ ■> f i ; i "\i
T: ■ ■’ ■>•
. i. ■■ i *i
'.
I, -* ■ '.Y(

it d' ■"\ PK^li ,-• . i.

v>.
■ - -.1
’■^. sr
111- r,.,- ••' ('•■■ S«svV .■• .A r-' :k
’ I . - i
>■'“ i .4 '‘.Wi "f V
* ■l
i&n
•* ■•>»-' . •t « »
V-i ^

*•» 1 ^ ■.. '■ ?■■. .

.M . • -r* # . • ’ »
;^jf ■’ ' ’ V 1= -tAi. •* .V ®
,', ■ l, -
' • j -• ,V"«‘"
V li',•
’ y
W -'^1? , ri?- -:.4

f' f* >. A '

^ *
„•
*
1 •; » ■ ■/!
^Xri ’
■, ■■*' ' ■ i -* - f
* ., V*
m ^ » «?.*•

• 1' >
n,
•A-
I 4
t *
» .. I
'V .
A • ' . ■■*■ l» : •) '
V,. 1* ■■
I'. '•; ^ ■, ^ na ' t A-'*'
“a
P.'
»' ) t:r7
i Q. *
|V

. •

' 1
I “8|

.1 . ♦t'.'A''■' ,4''m -■ , •■ . i.’^..r.4u

',S» .
r * t
^4 .♦: t.-> 1 ■
. 1
*
I A. .l;.' •■' 't>,
.;'i. “ •. . _
. 4.. >1* ■’ .
 Preface to the Third Edition

One of the main reasons for a third edition has been the use of computers to
simplify the analysis of data in an experiment. Computer programs follow most of
the complex examples showing the input instructions to a computer, the printout
from the computer, and the interpretation of the results.
The introduction has been completely revised to set the stage for ex-
perimental research while still maintaining the general theme of experiment,
design, and analysis. Chapters 4 and 5 of the previous edition have been combined
as they both represent restrictions on randomization, and a new chapter on
multiple regression has been added. Some topics rarely taught in a single-semester
course have been relegated to the chapter on miscellaneous topics.
Many illustrative examples have been replaced with examples that are more
practical and better illustrate the technique. There are 60 percent more problems
than in the second edition, resulting in more than 300 practical problems. Chapter
17 presents ten somewhat challenging problems for class discussion.
Two minor points: (1) Factorials have been introduced as 2^ and 3^ instead of
as 2" and 3", where/represents the number of factors in the experiment and n is
reserved for the number of observations per treatment. (2) The use of single and
double horizontal and vertical lines to show how the randomization has been
restricted should help illustrate how data were collected.
I am indebted to McElrath and Associates, Inc. (MAI), Management Consul-
tants, Minneapolis, Minnesota, for the opportunity to be exposed to many practic-
al problems in various industries over the past 20 years. Many of the text examples
have come from MAI clients. Several examples, especially in the field of educa-
tion, have come from students’ master’s and doctoral research. I want to thank the
following students for examples taken from their theses: Roger Bollenbacher,
Robert Horn, James Lehman, Willard Lewallen, Charlotte Macy, Thomas Mar-
shall, Larry Miller, Barbara O’Conner, Sallam Sallam, Nadia Sherief, Ruth
Tanner, John Wetz, and Catherine Yeotis.

xi
xii Preface to the third edition

Last, and most significant, I wish to thank Purdue University for granting me
sabbatical leave to complete this third revised edition.

Charles R. Hicks
Lafayette, Indiana
January, 1982
 Preface to the Second Edition

At first glance the reader may not notice much change in this edition from the
original. An attempt has been made to keep chapter and section headings the same.
It cannot be considered as a major revision although there are many new features.
The new material adds approximately 25% to the size of the book and the number
of problems has been incrased by 70%.
Main features include a section on covariance analysis and analysis of
attribute data in the 16th chapter, introduction of linear and curvilinear regression
in Chapter 8, and an amplification of the concept of orthogonal contrasts and
orthogonal polynomials. Some topics have been rearranged to appear earlier in the
text and methods of testing means after analysis of variance have been changed
and extended.
Several users of the original text have sent me helpful suggestions for this
revision. Whenever a suggestion came from more than one person, I attempted to
include it in this edition. Naturally, all suggestions were not followed and the
responsibility for their adoption is mine alone. I am indebted to the following
people who made detailed suggestions:
V. L. ANDERSON Purdue University
O. A. BECKENDORF General Motors Institute
H. C. CHARBOMEAU General Motors Institute
R. I. FIELDS University of Louisville
R. J. FOPMAN University of Cincinnati
B. HARSHBARGER Virginia Polytechnic Inst.
O. B. MOAN Arizona State University
R. G. MORRIS General Motors Institute
V. V. NAMILOV MOSCOW State University
J. B. NEUHARDT The Ohio State University
C. R. PETERSON General Motors Institute
B. T. RHODES, JR. University of Houston
H. W. STOKER The Florida State University
xiv Preface to the second edition

B. J. TODD Clemson University

L. A. WEAVER University of South Florida
I am also grateful to Purdue University for granting me sabbatical leave to
complete this revised edition.

Charles R. Hicks
Lafayette, Indiana
March, 1973
 Preface to the First Edition

It is the primary purpose of this book to present the fundamental concepts in the
design of experiments using simple numerical problems, many from actual re-
search work. These problems are selected to emphasize the basic philosophy of
design. Another purpose is to present a logical sequence of designs that fit into a
consistent outline; for every type of experiment, the distinctions among the
experiment, the design, and the analysis are emphasized. Since this theme of
experiment-design-analysis is to be highlighted throughout the text, the first
chapter presents and explains this theme. The second chapter reviews statistical
inference, and the remaining chapters follow a general outline, with an experi-
ment-design-analysis summary at the end of each chapter.
The book is written for anyone engaged in experimental work who has a good
background in statistical inference. It will be most profitable reading to those with
a background in statistical methods including analysis of variance.
Some of the latest techniques have not been included. The reason for this is
again the purpose of presenting the fundamental concepts in design to provide a
background for reading journals in the field and understanding some of the more
recent techniques. Although it is not assumed that every reader of this book is
familiar with the calculus, some theory is presented where calculus is used. The
reader should be able to grasp the underlying ideas of each chapter even if he must
omit these sections.
This book evolved from a course in Design of Experiments given at Purdue
University over the past four or five years. The author is indebted to Professors
Gayle McElrath of the University of Minnesota and Clyde Kramer of Virginia
Polytechnic Institute whose lectures, along with the author’s, form the basis of this
book. Credit is also due Mr. James Dick and Mr. Roland Rockwell, who took
notes of the author’s lectures.
I shall be forever indebted to the University of Florida for granting me a
visiting professorship during the academic year 1962-1963 with time to work on

XV
xvi Preface to the first edition

this book, and to Mrs. Alvarez of their Industrial Engineering Department for
typing the manuscript.

Charles R. Hicks
Lafayette, Indiana
November, 1963
Fundamental
Concepts in
the Design of
Experiments
 1
The Experiment,
the Design,
and the Analysis

1.1 INTRODUCTION TO EXPERIMENTAL DESIGN

One of the most abused words in the English language is probably the word
“research.” Research means many things to rhany people. The high school
student who visits the library to gather information on the Crusades is doing
research. A man who checks the monthly sales of his company for the year
1980 is doing research. Many such studies do make use of bibliothecal or
statistical tools or similar research tools but are a far cry from a scientist’s
conception of the term. Let us narrow the concept a bit and define research
as “a systematic quest for undiscovered truth.” (See Leedy [16].^)
In this definition we note that the quest must be “systematic,” using
sound scientific procedures—not a haphazard, seat-of-the-pants approach—
to solving a problem. The truth sought should be something that is not
already known, which implies a thorough literature search to ascertain that
the answer is not available in previous studies.
Not all studies are necessarily research, nor is all research experimental.
A true experiment may be defined as a study in which certain independent
variables are manipulated, their effect on one or more dependent variables
is determined, and the levels of these independent variables are assigned at
random to the experimental units in the study. Thus manipulation and
randomization are essential for a true experiment from which one may be
able to infer cause and effect. It is not always physically possible to assign
some variables to the experimental units at random but it may be possible
to run a quasi-experiment in which groups of units are assigned to various
levels of the independent variable at random. For example, to study the
effects of two methods of in-plant instruction on worker output, it may not
be possible to assign workers at random to the two methods, but only to
decide by a random flip of a coin which two of four intact classes will receive
method I and which two will receive method II. In such quasi-experiments
^ Numbers in brackets refer to the references at the end of the book.

1
2 The experiment, the design, and the analysis

the experimenter is obligated to demonstrate that the four classes were

similar at the start of the study. Since they can be dissimilar in many differ-
ent ways, this is no easy task. Randomizing all individuals to the four groups
will assure a high probability of group equality on many variables.
There are other types of research that are not experimental. One of the
most common is ex-post-facto research. The variables have already acted
and the researcher only measures what has occurred. An example of some
note would be the studies showing a high incidence of cancer in people who
smoke heavily. There is no manipulation here as the researchers do not
assign X number of cigarettes per day to some people and none to another
group, and then wait to see whether or not cancer develops. Instead they
study what they can find from people’s historical records, and then attempt
to make an inference. However, such inferences are dangerous, as the re-
searchers must be able to rule out other competing hypotheses as to what
may have caused cancer in the people studied.
In this category are survey research, which is also ex post facto, histor-
ical research, and developmental research. All of these may use statistical
methods to analyze data collected but none are really experimental in nature.
An illustration familiar to statisticians is found in the techniques of
regression and correlation. In a regression problem levels of the independent
variable X are set and observations are made on the dependent variable Y,
and the given level of X is assigned at random to each of the experimental
units in the study. The purpose of such a study is to find an equation for
predicting Y from X over the range of specified X’s. In a correlation problem
each experimental unit is measured on two variables, the n pairs {X, 7) are
plotted to look for a relationship, and certain statistics are computed to
indicate the strength of the relationship. In this case a strong correlation
(high r) does not imply that X caused Y or that Y caused X. Hence regression
analysis can often be used in a true experiment whereas correlational tech-
niques are more likely to be used in ex-post-facto studies. This distinction
between regression and correlation is often misunderstood.
Assuming that true experiments are our basic concern, what steps
should be taken to make a meaningful study?
To attack this problem one may easily recognize three important
phases of every project: the experimental or planning phase, the design
phase, and the analysis phase. The theme of this book might be called:
Experiment, Design, and Analysis.

1.2 THE EXPERIMENT

The experiment includes a statement of the problem to be solved. This sounds

rather obvious, but in practice it often takes quite a while to get general
agreement as to the statement of a problem. It is important to bring out all
 1.2 The experiment 3

points of view to establish just what the experiment is intended to do. A

careful statement of the problem goes a long way toward its solution.
Too often problems are tossed about in very general terms, a sort of
“idea burping” (see [16]). As an example, note the difference in these two
statements of a problem: What can Salk vaccine do for polio? Is there a
difference in the percentage of first, second, and third grade children in the
United States who contract polio within a year after being inoculated with
Salk vaccine and those who are not inoculated? Much care must be taken to
spell out the problem in terms that are well understood and that point the
way in which the research might be conducted.
The statement of the problem must include reference to one or more
criteria to be used in assessing the results of the study. The criterion is also
called the dependent variable or the response variable Y; and there may be
more than one Y of interest in a given study. The dependent variable may
be qualitative, such as the number or percentage contracting polio, or quan-
titative, such as yield in grams of penicillin per kilogram of corn steep.
Knowledge of this dependent variable is essential because the shape of its
distribution often dictates what statistical tests can be used in the subsequent
data analysis.
One must also ask: Is the criterion measurable, and if so, how accurately
can it be measured? What instruments are necessary to measure it?
It is also necessary to define the independent variables or factors that
may affect the dependent or response variable. Are these factors to be held
constant, to be manipulated at certain specified levels, or to be averaged out
by a process of randomization? Are levels of the factors to be set at certain
fixed values, such as temperature at 70°F, 90°F, and 110°F, or are such levels
to be chosen at random from among all possible levels? Are the factors to
be varied quantitative (such as temperature) or qualitative (operators)? And
how are the various factor levels to be combined?
If one factor A is set at three levels and a second factor B is set at two
levels, can one set all six combinations as follows?

Factor A
1 2 3

1
Factor B
2

This is called a factorial experiment as all levels of A are combined with

all levels of B. Sometimes one chooses three suppliers and then randomly
selects two batches from each supplier to be analyzed chemically. Here the
two batches are unique to a given supplier and the resulting arrangement is
a nested or hierarchical arrangement.
4 The experiment, the design, and the analysis

Supplier
1 2 3

Batch Batch Batch

12 3 4 5 6

This involves six sets of data as did the above factorial, but now the factors
are in a decidedly different arrangement.
All of these points should be considered as part of the experiment phase
of any research.

1.3 THE DESIGN

Of primary importance, since the remainder of this book will be devoted to
it, is the design phase of a project. Many times an experiment is agreed upon,
data are collected, and conclusions are drawn with little or no consideration
given to how the data were collected. First, how many observations are to be
taken? Considerations of how large a difference is to be detected, how much
variation is present, and what size risks are to be tolerated are all important
in deciding on the size of the sample to be taken for a given experiment.
Without this information the best alternative is to take as large a sample as
possible. In practice this sample size is often quite arbitrary. However, as
more and more tables become available, it should be possible to determine
the sample size in a much more objective fashion (see Section 2.5).
Also of prime importance is the order in which the experiment is to be
run, which should be random order. Once a decision has been made to control
certain variables at specified levels, there are always a number of other
variables that cannot be controlled. Randomization of the order of experi-
mentation will tend to average out the effect of these uncontrolled variables.
For example, if an experimenter wishes to compare the average current
flow through two types of computers and five of each type are to be tested,
in what order are all ten to be tested? If the five of type I are tested, followed
by the five of type II, and any general “drift” in line voltage occurs during
the testing, it may appear that the current fiow is greater on the first five
(type I) than on the second five (type II), yet the real cause is the “drift” in
line voltage. A random order for testing allows any time trends to average
out. It is desirable to have the average current flow in the type I and type II
computers equal, if the computer types do not differ in this respect. Ran-
domization will help accomplish this. Randomization will also permit the
experimenter to proceed as if the errors of measurement were independent, a
common assumption in most statistical analyses.
What is meant by random order? Is the whole experiment to be com-
pletely randomized, with each observation made only after consulting a table
 1.3 The design 5

of random numbers, or tossing dice, or flipping a coin? Or, possibly, once a

temperature bath is prepared, is the randomization made only within this
particular temperature and another randomization made at another tem-
perature? In other words, what is the randomization procedure and how are
the units arranged for testing? Once this step has been agreed upon, it is
recommended that the experimenter keep a watchful eye on the experiment
to see that it is actually conducted in the order prescribed.
Having agreed upon the experiment and the randomization procedure,
a mathematical model can now be set up, which should describe the experi-
ment. This model will show the response variable as a function of all factors
to be studied and any restrictions imposed on the experiment as a result of
the method of randomization.
Since the objective of the research project is to shed light on a stated
problem, one should now express the problem in terms of a testable hypoth-
esis or hypotheses. A research hypothesis states what the experimenter
expects to find in the data. For example: “Inoculation with Salk vaccine will
lower the incidence of polio.” Or: “The four treatments will produce different
average yields.”
To test hypotheses, a statistician usually states the hypothesis in null
form, since this is the only testable hypothesis. For example: “There will be no
difference in the percentage of youngsters contracting polio between those
inoculated and those not inoculated.” Or: “The treatment means will have
the same average yield.” This null hypothesis form can usually be expressed
in terms of the mathematical model set up in the last step of the design phase.
In this discussion of variables that influence the criterion either by
being purposefully introduced into the study, representing restrictions on
randomization, or being somewhat extraneous, one notes that there are
basically three ways to handle independent variables in an experiment:

1. Rigidly controlled the variables remain fixed throughout the experiment,

which, of course, means that any inferences drawn from the experiment
are valid only for these fixed conditions. Here, for example, one might
decide to study only children from the first three grades.
2. Manipulated or set at levels of interest. These would be the variables
whose effect on Y are to be studied. They could be either qualitative or
quantitative and their levels either fixed or random. To be effective,
every effort should be made to include the extreme levels of such variables
to provide opportunity to maximize their effect, if present.
3. Randomized. The order of experimentation is randomized to average
out the effects of variables that cannot be controlled. Such averaging
does not remove their effect completely as they still increase the variance
of the observed data. Proper design of the experiment can also reduce or
minimize the experimental error when such factors are anticipated.
6 The experiment, the design, and the analysis

A basic principle in such experimentation is referred to as the max-min-

con principle (see Kerlinger [14]), where we seek to maximize the effect of
the independent variables of interest, minimize the error variance, and control
some of the variables at specified levels.

1.4 THE ANALYSIS

The final step, analysis, includes the procedure for data collection, data
reduction, and the computation of certain test statistics to be used in making
decisions about various aspects of an experiment. Analysis involves the
computation of test statistics such as t, F, and their corresponding decision
rules for testing hypotheses about the mathematical model. Once the test
statistics have been computed, decisions must be made. These decisions
should be made in terms that are meaningful to the experimenter. They
should not be couched in statistical jargon such as “the third-order A x
B X E interaction is significant at the 1 percent level,” but instead should
be expressed in graphical or tabular form, in order that they be clearly
understood by the experimenter and by those persons who are to be “sold”
by the experiment. The actual statistical tests should probably be included
only in the appendix of a report on the experiment. These results should also
be used as “feedback” to design a better experiment, once certain hypotheses
seem tenable.

1.5 SUMMARY IN OUTLINE

I. Experiment
A. Statement of problem
B. Choice of response or dependent variable
C. Selection of factors to be varied
D. Choice of levels of these factors
1. Quantitative or qualitative
2. Fixed or random
E. How factor levels are to be combined

II. Design
A. Number of observations to be taken
B. Order of experimentation
C. Method of randomization to be used
D. Mathematical model to describe the experiment
E. Hypotheses to be tested

III. Analysis
A. Data collection and processing
B. Computation of test statistics ,
C. Interpretation of results for the experimenter
 1.6 Examples 7

1.6 EXAMPLES

Example 1.1 Salk Vaccine Experiment. The experiment conducted in the

United States in 1954 concerning the use of Salk vaccine in the control of
polio has become a classic in its employment of good design principles. A
more detailed description of the study can be found in Statistics: A Guide
to the Unknown (see [23]).
Before the statement of the problem was agreed upon as given in
Section 1.2, there was much discussion as to what population was to be
sampled. It was decided that children in grades 1, 2, and 3 would be most
appropriate, as it was among children of these ages that the disease was
most prevalent.
Implied in the statement of the problem “Is there a difference in the
percentage of first, second, and third grade children in the United States
contracting polio within a year after being inoculated with Salk vaccine
and those not inoculated?” is the concept of a criterion. The stated criterion
is the percentage of children contracting polio. This dependent variable
could be either the percent contracting the disease or the number contracting
it. In discussing this variable it was noted that the overall percentage of
children contracting polio was quite small—about eight or ten cases in
10,000 children. It was therefore necessary to take a fairly large sample in
order to get any percentages large enough to discriminate between those
who were inoculated and those who were not. The basic variable here is a
qualitative one—a binomial variable—as children fall into just two cate-
gories: contracting polio or not contracting polio. One can safely assume an
approximately normal distribution of this Y if the average number in the
sample is expected to be four or more.
The main factor of interest in affecting Y is whether or not the child
is inoculated, so it was planned to have an approximately equal number of
children receive the vaccine and be given a placebo (a salt solution). This
latter group would act as a control.
Other factors that might affect the contraction of polio included socio-
economic status, grade level, doctors who make the diagnosis, and geo-
graphic area of residence. Grade level was considered and they agreed to
sample from the first, second, and third grades. The other factors, or nui-
sance variables, were handled in the design procedures.
Since there is only one qualitative, fixed factor of interest—inoculated
or not inoculated—there is no basic concern about combining factors.
However, because three grade levels were to be considered, it was decided to
stratify their sample by grades; that is, take approximately equal sample
sizes of inoculated and noninoculated children from each grade level.
Much planning went into the design phase. At first it was proposed
that children in grades 1 and 3 receive the vaccine and those in grade 2 be
given the placebo treatment. This plan was later abandoned because polio
8 The experiment, the design, and the analysis

is a contagious disease, and if children in a given grade get polio, others in

that grade might contract it as well. Another early idea was to use children
of parents who would consent to the inoculation as “experimental” and
children of parents not consenting as a “control.” This too was abandoned
because parents of a higher socioeconomic status are more likely to consent
than parents with a lower socioeconomic status and polio is a socially
related disease—it affects more children from the upper echelons of society
than from the lower!
The final decision was to do a complete randomization countrywide
involving about 1 million randomly selected children from grades 1-3 in
areas of the United States where polio had been quite prevalent in the past.
Of these 1 million, more than 400,000 parents consented to have their children
participate in the study. These children were then divided into two groups
at random with group 1 getting the Salk vaccine and group 2 getting the
placebo. To handle other concerns, the experiment was run as a “double-
blind” experiment. The children did not know whether they were receiving
the vaccine or the salt solution and the doctors who diagnosed the children
did not know which children were inoculated and which were not.
The samples of approximately 200,000 each seemed adequate to permit
the detection of differences in the percent contracting the disease, if it did
show up.
A mathematical model could be written as

Yij - Tj

where Y^j is equal to 0 or 1: 0 if no polio is diagnosed and 1 if polio is found.

The subscipt i represents the child, a number from 1 to approximately
200,000, and7 is 1 or 2:1 if treated with Salk vaccine and 2 if not so treated.
is SL constant or general average of the O’s and Ts for the whole population.
Xj is the treatment effect: 7 = 1 if treated, 7 = 2 if not treated. 6,j is a random
error associated with a child receiving treatment 7.
To see the hypothesis to be tested, note that

is the proportion of children contracting polio who received treatment 7.

tij is the number who received treatment 7.
The statistical hypothesis to be tested is then

HQ : P\ = p'2
with alternative
^ 1 • P'l < P'l
where the sign < indicates that we are interested in showing that the true
proportion contracting polio will be less for treatment 1 (vaccine) than for
 1.6 Examples 9

treatment 2 (placebo). The primes on the p values are to show that the
hypotheses are statements about population proportions whereas the sam-
ples will give Pi and P —observed sample proportions. The data are shown
2

in Table 1.1.

Table 1.1 Polio Experiment Results

Number
Sample Contracting Proportion
Treatment Size Polio Contracting

Salk vaccine /q = 200,745 56 Pi = 28 X 10“^

Placebo /i2 = 201,229 142 = 71 X 10“^

Totals 401,974 198 p = 49 X 10“^

The test statistic for testing the hypothesis stated is

Pi - Pi
^pq(l/ni + I/MJ)

where p is the proportion contracting polio in the total sample of ^i + ^2

children. Here q = i — p.
If one now assumes an a risk of 0.001—taking one chance in 1000 of
rejecting HQ when it is true—the rejection region for z is given by z < — 3.09
(from the normal distribution table). Substituting the values above:
28 X 10“^ - 71 X 10"^
749 X 10^^(1 - 49 X 10“^)[l/200,745 + 1/201,229]
z = —6.14

which gives a very, very strong indication that we should reject HQ and
conclude that the vaccine was indeed effective.
In fact, the probability of getting a z as low as or lower than —6.14
when there is really no difference in the two groups is less than one chance
in a billion. These amazing results, of course, have been strongly substan-
tiated by the almost complete eradication of polio in the United States.

Example 1.2 The following example [18] is presented to show the three
phases of the design of an experiment. It is not assumed that the reader is
familiar with the design principles or analysis techniques in this problem.
The remainder of the book is devoted to a discussion of many such principles
and techniques, including those used in this problem.
An experiment was to be designed to study the effect of several factors
on the power requirements for cutting metal with ceramic tools. The metal
10 The experiment, the design, and the analysis

was cut on a lathe and the y or vertical component of a dynamometer

reading was recorded. As this y component is proportional to the horse-
power requirements in making the cut, it was taken as the measured variable.
The y component is measured in millimeters of deflection on a recording
instrument. Some of the factors that might affect this deflection are tool
types, angle of tool edge bevel, type of cut, depth of cut, feed rate, and spindle
speed. After much discussion, it was agreed to hold depth of cut constant at
0.100 inch, feed rate constant at 0.012 in./min, and spindle speed constant at
1000 rpm. These levels were felt to represent typical operating conditions.
The main objective of the study was to determine the effect of the other three
factors (tool type, angle of edge bevel, and type of cut) on the power require-
ments. As only two ceramic tool types were available, this factor was con-
sidered at two levels. The angle of tool edge bevel was also set at two levels,
15° and 30°, representing the extremes for normal operation. The type of
cut was either continuous or interrupted—again, two levels.
There are therefore two fixed levels for each of three factors, or eight
experimental conditions (2^), which may be set and which may affect the
power requirements or y deflection on the dynamometer. This is called a 2^
factorial experiment, since both levels of each of the three factors are to be
combined with both levels of all other factors. The levels of two factors (type
of tool and type of cut) are qualitative, whereas the angle of edge bevel
(15° and 30°) is a quantitative factor.
The question of design for this experiment involves the number of tests
to be made under each of the eight experimental conditions. After some
preliminary discussion of expected variability under the same set of condi-
tions and the costs of wrong decisions, it was decided to take four observa-
tions under each of the eight conditions, making a total of 32 runs. The order
in which these 32 units were to be put in a lathe and cut was to be completely
randomized.
In order completely to randomize the 32 readings, the experimenter
decided on the order of experimentation from the results of three coin toss-
ings. A penny was used to represent the tool type T: heads for one type,
tails for the other; a nickel represented the angle of bevel B: heads 30°,
tails 15°; and a dime represented the type of cut C: heads interrupted, tails
continuous.
Thus if the first set of tosses came up THT, it would mean that the
first tool to be used in the lathe would be of tool type 1, bevel 2, and sub-
jected to the continuous type of cut. A data layout is given in Figure 1.1
that shows each of the 32 experimental conditions. The numbers 1, 2, 3, 4,
and 5 indicate the first five conditions to be run on the lathe, assuming the
coins came up THT, TTH, HHT, HHT, HTH.
In this layout it should be noted that the same set of conditions may be
repeated (for example, runs 3 and 4) before all eight conditions are run once.
 1.6 Examples 11

Tool Type T

1 2

Bevel B Bevel B

Type Cut
of C
2 2
3
4

Figure 1.1 Data layout of power consumption for ceramic tools.

The only restriction on complete randomization here is that once four

repeated measures have occurred in the same cell, no more will be run
using those same conditions.
The coin flipping continues until the order of all 32 runs has been
decided upon. This is a 2^ factorial experiment with four observations per
cell, run in a completely randomized manner. Complete randomization
ensures the averaging out of any effects that might be correlated with the
time of the experiment. If the lathe-spindle speed should vary and all of the
type 1 tools are run through first and the type 2 tools follow, this extraneous
effect of lathe-spindle speed might appear as a difference between tool types
if the speed were faster at first and slower near the end of the experiment.
The mathematical model for this experiment and design would
be Yijkm = + Bj -\- TBij + + TCj^ + + TBCij^ + Sm{ijk)^
where represents the measured variable, fi a common effect in all
observations (the true mean of the population from which all the data came),
Tj the tool type effect where i = 1,2, Bj the angle of bevel where j = 1,2,
and Cy. the type of cut where k = 1, 2. represents the random error in
the experiment where m = 1, 2, 3, 4. The other terms stand for interactions
between the main factors T, B, and C.
The analysis of this experiment consists of collecting 32 items of data in
the spaces indicated in Figure 1.1 in a completely randomized manner. The
results in millimeter deflection are given in Table 1.2.
This experiment and the mathematical model suggest a three-way
analysis of variance (ANOVA), which yields the results in Table 1.3.
In testing the hypotheses that there is no type of tool effect, no bevel
effect, no type of cut effect, and no interactions, the EMS column indicates
that all observed mean squares are to be tested against the error mean
12 The experiment, the design, and the analysis

Table 1.2 Data for Power Requirement Example

Tool Type T
1 :1

Bevel Angle B Bevel Angle B

Type of Cut C 15° 30° 15° 30°

Continuous 29.0 28.5 28.0 29.5

26.5 28.5 28.5 32.0
30.5 30.0 28.0 29.0
27.0 32.5 25.0 28.0
Interrupted 28.0 27.0 24.5 27.5
25.0 29.0 25.0 28.0
26.5 27.5 28.0 27.0
26.5 27.5 26.0 26.0

Table 1.3 ANOVA for Power Requirement Example

Degrees of Sum of Expected Mean

Freedom Squares Mean Square Square
Source of Variation (df) (SS) (MS) (EMS)

Tool types T 1 2.82 2.82 + 1 ^

6 Y'

Bevel B 1 20.32 20.32* + 16</^B

T X B interaction 1 0.20 0.20 + S(j>TB

Type of cut C 1 31.01 31.01* + \6(/>c
T X C interaction 1 O.Ol 0.01 + 8 (j^YC
B X C interaction 1 0.94 0.94 +
T X B X C interaction 1 0.19 0.19 +
Error 24 53.44 2.23

Totals 31 108.93

* One asterisk indicates significance at the 5 percent level; two, the 1 percent level; and three,
the 0.1 percent level. This notation will be used throughout this book.

square of 2.23 with 24 degrees of freedom (df). The proper test statistic is
the F statistic (Appendix, Table D) with 1 and 24 df. At the 5 percent signifi-
cance level (a = 0.05) the critical region of F is F > 4.26. Comparing each
mean square with the error mean square indicates that only two hypotheses
can be rejected: bevel has no effect on deflection and type of cut has no
effect on deffection. None of the other hypotheses can be rejected, and it is
concluded that only the angle of bevel and type of cut affect power consump-
tion as measured by the y deffection on the dynamometer. Tool type appears
to have little effect on the y deflection, and all interactions are negligible.
 1.6 Examples 13

Table 1.4 Average Deflections

Tool type T 1 28.1

2 27.5

Bevel angle B 15° 27.0

30° 28.6

Type of cut C 1 28.8

2 26.8

Calculations on the original data of Table 1.2 show the average y deflections
given in Table 1.4.
These averages seem to bear out the conclusions that bevel affects y
deflection, with a 30° bevel requiring more power than the 15° bevel (note
the difference in average y deflection of 1.6 mm), and that type of cut affects
y deflection, with a continuous cut averaging 2.0 mm more deflection than
an interrupted cut. The difference of 0.6 mm in average deflection due to
tool type is not significant at the 5 percent level. Graphing all four B-C
combinations in Figure 1.2 shows the meaning of no significant interaction.

Y
y Deflection
(mm)

Figure 1.2 B x C interaction for power requirement example.

A brief examination of this graph indicates that the y deflection is

increased by an increase in degree of bevel. The fact that the line for the
continuous cut Cj is above the line for the interrupted cut C2 shows that the
continuous cut requires more power. The fact that the lines are nearly
parallel is characteristic of no interaction between two factors. Or it can
14 The experiment, the design, and the analysis

easily be seen that an increase in the degree of bevel produced about the
same average increase in y deflection regardless of which type of cut was
made. This is another way to interpret the presence of no interaction.
The experiment here was a three-factor experiment with two levels for
each factor. The design was a completely randomized design and the analysis
was a three-way ANOVA with four observations per cell. From the results
of this experiment the experimenter not only found that two factors (angle
of bevel and type of cut) affect the power requirements, but also determined
that within the range of the experiment it makes little difference which
ceramic tool type is used and that there are no significant interactions
among the three factors.
These are but two examples that might have been used to illustrate the
three phases of the project: experiment, design, and analysis.

PROBLEMS

1.1 In your own area of interest, write out the statement of a problem that you believe
is researchable.
1.2 For the problem stated in (1), list the dependent variable or variables of interest
and how you propose to measure them.
1.3 For the problem stated in (1), list the independent variables that might affect your
dependent variable(s) and explain how you would handle each variable.
1.4 Prepare a data layout for collecting data on your problem.
 2
Review of
Statistical Inference

2.1 INTRODUCTION
In any experiment the experimenter is attempting to draw certain inferences
or make a decision about some hypothesis or “hunch” concerning the
situation being studied. Life consists of a series of decision-making situations.
As you taste your morning cup of coffee, almost unconsciously you decide
that it is better than, the same as, or worse than the coffee you have been
drinking in the past. If, based on your idea of a “standard” cup of coffee, the
new cup is sufficiently “bad,” you may pour it out and make a new batch.
If you believe it superior to your “standard,” you may try to determine
whether the brand is different than the one you are used to or the brewing
is different, and so forth. In any event, it is the extreme differences that may
cause you to decide to take action. Otherwise, you behave as if nothing has
changed. In any decision you run the risk that the decision is wrong since
it is based on a small amount of data.
In deciding whether or not to carry an umbrella on a given morning, you
“collect” certain data: you tap the barometer, look at the sky, read the
newspaper forecast, listen to the radio, and so on. After quickly assimilating
all available data—including such predictions as “a 30 percent probability
of rain today”—you make a decision. Somehow a compromise is made
between the inconvenience of carrying the umbrella and the possibility of
having to spend money for a cleaning bill for one’s clothes.
In these instances, and in most everyday events, decisions are made in
the light of uncertainty. Statistics may be defined as a tool for decision making
in the light of uncertainty. Uncertainty does not imply no knowledge, but
only that the exact outcome is not completely predictable. If ten coins are
tossed, one knows that the number of heads will be some integer between
0 and 10. However, each specific integer has a certain chance of occurring
and various results may be predicted in terms of their chance or probability
of occurrence. When the results of an experiment that are observed could
have occurred only 5 times in 100 by chance alone, most experimenters

15
16 Review of statistical inference

consider that this is a rare event and will state that the results are statistically
significant at the 5 percent significance level. In such cases the hypothesis
being tested is usually rejected as untenable. When statistical methods are
used in experimentation, one can assess the magnitude of the risks taken in
making a particular decision.
Statistical inference refers to the process of inferring something about a
population from a sample drawn from that population. The population
consists of all possible values of some random variable Y, where Y represents
the response or dependent variable being measured. The response Y may
represent tensile strength, weight, score, reaction time, or whatever criterion
is being used to evaluate the experimental results. Characteristics of the
population of this random variable are called parameters. In general 6 will
be used to designate any given population parameter. The average or ex-
pected value of the random variable is designated as E{Y) = p. If the
probability function defining the random variable is known,

E(Y) = Y. Y^PiYi)
where p(Yi) is a discrete probability function, and
E{Y) = jYf(Y)dY

where f{Y) is a continuous probability density function.

The long-range average of squared deviations from the mean of a
population is called the population variance a}. Or
E[{Y - = C7?
The square root of this parameter is called population standard deviation
<7y. Quantities computed from the sample values drawn from the population
are called sample statistics, or simply, statistics. Examples include the sample
mean ^
r = Z YJn
i

where n is the number of observations in the sample, and the sample variance

= t (Yi - Yf/in - 1)
i= 1

Italic letters will be used to designate sample statistics. The symbol u will
be used to designate a general statistic corresponding to the population
parameter 6.
Most statistical theory is based on the assumption that samples drawn
are random samples, that is, that each member of the population has an
equal chance of being included in the sample and that the pattern of variation
in the population is not changed by this deletion of the n members for the
sample.
 2.2 Estimation 17

The notion of statistical inference may be divided into two parts: (1)
estimation and (2) tests of hypotheses.

2.2 ESTIMATION

The objective of statistical estimation is to make an estimate of a population

parameter based on a sample statistic drawn from this population. Two types
of estimates are usually needed, point estimates and interval estimates.
A point estimate is a single statistic used to estimate a parameter. For
example, the sample mean F is a point estimate of the population mean p.
Point estimates are usually expected to have certain desirable characteristics.
They should be unbiased, consistent, and have minimum variance.
An unbiased statistic is one whose expected or average value taken over
an infinite number of similar samples equals the population parameter being
estimated. Symbolically, E{u) = d, where £(•) means the expected value of
the statistic in (•). The sample mean is an unbiased statistic, since it can be
proved that
E(Y) = PY (or simply,/i)

Likewise, the sample variance as defined above is unbiased, since

E{s^) = E Yf/in - 1) CTy (or cr^)

Note that the sum of squares Yj= i (F ~ F)^ must be divided by n and
not by n if s^ is to be unbiased. The sample standard deviation

s Yf/in - 1)

is not unbiased, since it can be shown that

E{s) ^ (7

This somewhat subtle point is proved in Burr (see [7], pp. 101-104).
A few basic theorems involving expected values that will be useful later
in this book include
E{k) = k (where /c is a constant)
E{kY) = kE{Y)
E{Y, + Y2} = E(YY E{Y2)
E{Y - PYY = E{Y^) - pj = al

The statistic (F — F)^ is used so frequently that it is often referred^

to as the sum of squares, or SSy. It is actually the sum of the squares of
deviations of the sample values from their sample mean.
18 Review of statistical inference

From the above it is easily seen that

E — {n — 1)<7Y

or
£(SSy) = {n - 1)(7?

where n — 1 represents the degrees of freedom associated with the sum of

squares on the left. In general, for any statistic u,

Z («i - uf (df on SSX

or
£(SS J = (df on SS„)(T„^

When a sum of squares is divided by its degrees of freedom, it indicates an

averaging of the squared deviations from the mean, and this statistic, SS/df,
is called a mean square, or MS. Hence

E(MS) = £(SS/df) = (7^

of the variable or statistic being studied.

A consistent statistic is one whose value comes closer and closer to the
parameter as the sample size is increased. Symbolically,

lim Pr(|w„ — 0| < 8) ^ 1

n-^ 00

which expresses the notion that as n increases, the probability approaches

certainty (or 1) that the statistic which depends on n, will be within a
distance e, however small, of the true parameter 6.
Minimum variance applies where two or more statistics are being com-
pared. If u^ and U2 are two estimates of the same parameter 6, the estimate
having the smaller standard deviation is called the minimum-variance
estimate. This is shown diagrammatically in Figure 2.1. In Figure 2.1, <
(j„2, and is then said to be the minimum-variance estimate as variance is the
square of the standard deviation.

Figure 2.1 Minimum-variance estimates.

 2.3 Tests of hypotheses 19

An interval estimate consists of the interval between two values of a

sample statistic that is asserted to include the parameter in question. The
band of values between these two limits is called a confidence interval for the
parameter, since its width may be determined from the degree of confidence
that is assumed when we say that the parameter lies in this band. Confidence
limits (the end points of the confidence interval) are determined from the
observed sample statistic, the sample size, and the degree of confidence
desired. A 95 percent confidence interval on is given by

Y ± 1.96

where 1.96 is taken from normal distribution values (see Appendix, Table A),
n is the sample size, and a is the population standard deviation. If 100 sample
means based on n observations each are computed, and 100 confidence
intervals are set up using the above formula, we expect that about 95 of the
100 intervals will include {i. If only one interval is set up based on one sample
of n, as is usually the case, we can state that we have 95 percent confidence
that this interval includes fi. If cr is unknown, the Student’s t distribution
(Appendix, Table B) is used, and confidence intervals are given by

Y ± tl-a/2

where s is the sample standard deviation and t has n — 1 dL 100(1 — a)

percent gives the degree of confidence desired.

2.3 TESTS OF HYPOTHESES

A statistical hypothesis is an assumption about the population being sampled.

It usually consists of assigning a value to one or more parameters of the
population. For example, it may be hypothesized that the average number
of miles per gallon obtained with a certain carburetor is 19.5. This is expressed
as HQ \ p = 19.5 mi/gal. The basis for the assignment of this value to p
usually rests on past experience with similar carburetors. Another example
would be to hypothesize that the variance in weight of filled vials for the
week is 40 grams^ or HQ : = 40 g^. When such hypotheses are to be
tested, the other parameters of the population are either assumed or esti-
mated from data taken on a random sample from this population.
A test of a hypothesis is simply a rule by which a hypothesis is either
accepted or rejected. Such a rule is usually based on sample statistics, called
test statistics, when they are used to test hypotheses. For example, the rule
might be to reject HQ: p = 19.5 mi/gal if a sample of 25 carburetors averaged
20 Review of statistical inference

18.0 mi/gal (F) or less when tested. The critical region of a test statistic con-
sists of all values of the test statistic where the decision is made to reject
HQ. In the example above, the critical region for the test statistic Y is where
F < 18.0 mi/gal.
Since hypothesis testing is based on observed sample statistics computed
on n observations, the decision is always subject to possible errors. If the
hypothesis is really true and it is rejected by the sample, a type I error is
committed. The probability of a type I error is designated as a. If the hypoth-
esis is accepted when it is not true, that is, some alternative hypothesis is
true, a type II error has been made and its probability is designated as p.
These a and P error probabilities are often referred to as the risks of making
incorrect decisions, and one of the objectives in hypothesis testing is to de-
sign a test whose a and p risks are both small. In most such test procedures
Of is set at some predetermined level, and the decision rule is then formulated
in such a way as to minimize the other risk, p. In quality control work, a
is the producer’s risk and p the consumer’s risk.
To review hypothesis testing, a series of steps can be taken that will
apply to most types of hypotheses and test statistics. To help clarify these
steps and to illustrate the procedure, a simple example is given parallel to
the steps.

Steps in Hypothesis Testing Examples

1. Set up the hypothesis and its 1. HQ \ p = 19.5 mi/gal.
alternative. H^ \ p < 19.5 mi/gal.
2. Set the significance level of the 2. a = 0.05, n = 25.
test a and the sample size n.
3. Choose a test statistic to test HQ 3. Test statistic: Y or standard-
noting any assumptions neces- ized F
sary when applying this statistic.
Y - p
Z -
(J \ n

(Assume a known and =2.)

4. Determine the sampling distri- 4. Y is normally distributed with
bution of this test statistic when mean p and standard deviation
HQ is true. (j/Jn. Or Z is N{0, 1).
5. Set up a critical region on this
test statistic where HQ will be
rejected in (lOO)a percent of the
samples when HQ is true.

-z
-1.645 0
%

Critical region: Z < —1.645

 2.4 The operating characteristic curve 21

6. Choose a random sample of n 6. If n = 25 and Y = 18.9 mi/gal.

observations, compute the test
statistic, and make a decision on
Ho. 2/V25
As —1.5 > —1.645, do not re-
ject HQ.
In the example above a one-sided or one-tailed test was used. This is
dictated by the alternative hypothesis, since we only wish to reject HQ when
low values of Y are observed. The size of the significance level a is often set
in an arbitrary fashion such as 0.05 or 0.01. It should reflect the seriousness
of rejecting many carburetors when they are really satisfactory, or when the
actual mean of the lot (population) is 19.5 mi/gal or better. In using the
normal variate Z, a is assumed known; a difiTerent test statistic would be
used if (7 were unknown, namely. Student’s t. The critical region may also
be expressed in terms of Y using the critical Z value of —1.645:

- 19.5
-1.645
2/V25
or Y^ = 18.8, and the decision rule can be expressed as: Reject HQ if Y <
18.8. Here Y = 18.9 and the hypothesis is not rejected.
The procedure outlined may be used to test many different hypotheses.
The nature of the problem will indicate what test statistic is to be used, and
proper tables can be found to set up the required critical region. Well known
are tests such as those on a single mean, two means with various assumptions
about the corresponding variances, one variance, and two variances. These
tests are reviewed in Sections 2.6-2.8.

2.4 THE OPERATING CHARACTERISTIC CURVE

In the example given above no mention was made of the type II error of
size p. P is the probability of accepting the original hypothesis HQ when it is
not true or when some alternative hypothesis, is true. Now in this
example states that ji < 19.5 mi/gal, and there are many possible means that
satisfy this alternative hypothesis. Thus ^ is a function of the fx that is less
than the hypothesized fi value. To see how p varies with ji, let us consider
several possible ji values and sketch the distributions of sample means Y
that these various jx values would generate. Also note the critical region
that already has been established as 7 = 18.8 based on HQ.
From Figure 2.2 it can be seen that as one considers fx farther and farther
to the left of 19.5, the P error decreases. The size of any p error can be deter-
mined for any given fi from Figure 2.2 and a normal distribution table. For
example, if ix is assumed to be 19.0, the critical mean 18.8 is standardized
22 Review of statistical inference

with respect to this assumed mean and

^ 18.8 - 19.0 -0.2

^ = -0.5
IjsJTs 0.4
indicating that the rejection point is one half of a standard deviation below
19.0. For a normal distribution (see Appendix, Table A), a Z = —0.5 has
an area above it of 1 — 0.3085 or 0.6915. Thus the p error for f.i = 19.0 is
0.69, or 69 percent of the samples would still accept the original hypothesis
that fi = 19.5 when it has really dropped to 19.0.
If one considers the mean /i at 18.0 as in the bottom sketch of Figure 2.2

18.8 - 18.0
Z = 2.00
04
and the P error is only 0.0228.
If P is plotted for various values of fi, the resulting curve is called the
operating characteristic (or OC) curve for this test. Table 2.1 summarizes
the computations for several assumed values of p.
 2.5 How large a sample? 23

Table 2.1 Data for Operating Charac-

teristic Curve

18.8 — p

p 2/V25 P 1 - P

18.0 -f2.00 0.02 0.98

18.3 + 1.25 0.11 0.89
18.5 + 0.75 0.23 0.77
19.0 -0.50 0.69 0.31
19.2 -1.00 0.84 0.16
19.5 -1.75 0.96 0.04*

* (approximately = a)

j3 (probability

Figure 2.3 Operating characteristic curve.

The plotted OC curve is then as shown in Figure 2.3.

Some people prefer to plot 1 — P versus /i. This curve is called the
power curve of the test as 1 — is called the power of the test against the
alternative p.
It should be noted that when p is quite a distance from the hypothesized
value of 19.5, the test does a fairly good job of detecting this shift; that is,
if p = 18.5, the probability of rejecting 19.5 is 0.77, which is fairly high. On
the other hand, if p has shifted only slightly from 19.5, say 19.2, the probability
of detection is only 0.16. The power of a test may be increased by increasing
the sample size or increasing the risk a.

2.5 HOW LARGE A SAMPLE?

The question of how large a sample to take from a population for making
a test is one often asked of a statistician. This question can be answered
provided the experimenter can answer each of the following questions.
24 Review of statistical inference

^ Reject * Accept ►

Figure 2.4 Determining sample size.

1. How large a shift (from 19.5 to 19.0) in a parameter do you wish to detect?
2. How much variability is present in the population? (Based on past expe-
rience, (7 = 2 mi/gal.)
3. What size risks are you willing to take? (a = 0.05 and p = 0.10.)

If numerical values can be at least estimated in answering the above ques-

tions, the sample size may be determined.
Set up two sampling distributions of F, one when HQ is true, /i = 19.5,
and the other when the alternative to be detected is true, fi = 19.0 (Figure 2.4).
Indicate by F value between the two /t’s that will become a critical point,
rejecting HQ for observed values of Y below it and accepting HQ for Y values
above it. Indicate the a and P risks on the diagram. Set up two simultaneous
equations, standardizing F first with respect to a of 19.5 (a equation) and
second with respect to a of 19.0 {p equation). Solve these equations for n
and F:

a equation:
F - 19.5
—— = — 1.645 (based on a = 0.05)
llsfn

„ .
P equation:
7 - 19.0
—— = +1.282 (based on p = 0.10)
2/y/n

Subtracting the second equation from the first and multiplying both sides
of each equation by 2/y/n gives

-0.5 - -2.921

5.854
11.71
0.5

n = (ll.llf = 136.9 or 137

 2.6 Application to tests on variances 25

Keeping a = 0.05 gives

19.22 mi/gal

The decision rule is then: Choose a random sample of 137 carburetors, and
if the mean mi/gal of these is less than 19.22, reject HQ ; otherwise, accept HQ.
Excellent tables are available for determining n for several tests of
hypotheses, such as those in Davies [9, pp. 606-615] or Owen [20, pp. 19,
23, 36, 41, ff.].

2.6 APPLICATION TO TESTS ON VARIANCES

To review how the concepts of confidence limits and hypothesis testing

applies to variance tests consider the following.

I. Tests on a single variance.

HQ : 0-2 = al
^ GQ [or G^ > GQ or G^ < (JQ]
Choose a and n.
Test statistic is
jn - i)s^

which follows the chi-square distribution with {n — 1) degrees of freedom,

if one can assume that Y is normally distributed.
If > xh or < x\-«i2, reject
[If x^ ^ xl, reject, or if < Xi-,, reject]
To set confidence limits on G^ one computes

(n - l)s^ 2 /
2 < (^ < 2
Xa/2 ll-01.12
using the chi-square table with {n — 1) df to give 100(1 — a) percent con-
fidence limits.
As an example consider that one might wish to determine whether a
process standard deviation is greater than a specified value of 5 ounces. If
a sample of ten weights is found to have an average of 35.2 ounces and a
standard deviation of 7 ounces, what can we conclude about the standard
deviation of this process?
We hypothesize HQ: G^ = 25 oz^ and take as alternative H^ : G^ >
25 oz^. Selecting a = 0.05 and n = 10 gives n — 1 = 9 df. One should
reject HQ if
> ZL5(9 df) = 16.919

based on Table C of the Appendix.

26 Review of statistical inference

From the sample data

2 9(49)
X = -
17.64

so we reject the hypothesis and conclude that our process standard deviation
is greater than 5 ounces.
We might wish to set 90 percent confidence limits on (or a) based on
our observed variance of 49:
9(49) 9(49)
< G^

Zo.05 Xo.95

9(49) ^ ^ 9(49)
16.919 “ “ 3.325
26.07 <<j^< 132.63
or
5.1 < cr < 11.5 ounces

and we feel 90 percent confident that the process standard deviation is

between 5.1 and 11.5. We note that this band does not include the hypoth-
esized value of 5 ounces. We also note that this is a two-sided confidence band
whereas the hypothesis test was one-sided. By making the a for confidence
limits twice the a in the test of HQ, the results are comparable.

II. Tests on two independent variances.

As another example of a test of a hypothesis, consider testing whether or

not the variances of two normal populations are equal. This example is
included here as the test statistic involved will have many applications in
later chapters.
In accordance with the steps outlined in Section 2.3, we have the
following.

Ho- = (^2 H^IG^ > GI

a = 0.05.
Test statistic F = s\ls\ (often called the variance ratio, where is based
on — 1 df and si is based on 112 — 1 df).
If the two samples are independently chosen from normal populations
and HQ is true, the F statistic follows a skewed distribution, formed as
the ratio of two independent chi-square distributions. A table giving a
few percentiles of this F distribution appears as Appendix Table D. This
table is entered with — 1 df for the numerator and U2 — 1 df for the
denominator.

The critical region is set at F > ^0.95 in this example (a == 0.05) with
the degrees of freedom dependent on the size of the two samples.
 2.7 Application to tests on means 27

If the first sample results are

= 156
and the second are
^2 = 100
then
156
F - 1.56
100

and the critical region is F > 3.29 for 7 and 9 df. Hence the hypothesis is not
rejected.
In this example only a one-sided test has been considered as one is
usually concerned with showing that one variance is greater or less than
another. It is customary then to place the larger observed variance in the
numerator of F and call it s^.

2.7 APPLICATION TO TESTS ON MEANS

I. Tests on a single mean.

HQ : H = ^0
# //o [or /I > JUQ or II < /IQ]
Choose n and a.
Test statistic when:
A. cr is known.

Z is A(0, 1)
Reject HQ if |Z S Zi-ai2 [orZ > Zi_„ orZ < ZJ.

The problem of the carburetors in Section 2.3 illustrates this test.

B. a is unknown, normal population.

slyjn

t follows a t distribution (Appendix, Table B) with n — 1 degrees of

freedom.
Reject HQ if |t| > fi _„,2
[or t > tj or t < - tj _ J.

Consider a sample of six cylinder blocks whose cope hardness values

are 70, 75, 60, 75, 65, and 80. Is there evidence here that the average cope
hardness has changed from its specified value of 75?
HQ: H = 15
HY'. H ^ 75
28 Review of statistical inference

If a = 0.05, n = 6, one should reject HQ if t ^ ^0.975 with 5 df or if

t > 2.57. From the given sample of six: 7 = 70.8, s = 7.4. So,

70.8 - 75
t = = -1.39
7.4/V6

Since \t\ = 1.39 < 2.57, one cannot reject HQ. We might have used our
data to set 95 percent confidence limits on fi. From Section 2.2 100(1 — a)
percent confidence limits are

^ + h-aiisl-Jn
70.8 + 2.51(1 A) iS
70.8 + 7.8

or from 63.0 to 78.6, which includes our hypothesized value of 75.

II. Tests on two means.

A. If independent samples:

1. If variances are known: G\ and G\.

HQ: = fi2

Hi'. ^ [or/^i > 1^2]

Choose a, n^, and ^2.

Test statistic is

Reject HQ if Z ^ '^i-a/2 [or Z > Zi_J.

2. If variances are unknown but equal and populations are normal:

Test statistic is

Y, - Y2
t =
l{n^ - l)si + (772 - l)s
77i + 772 — 2 77i ^ 772^

with 77^ + 772 — 2 df. One rejects if |t| > fi_<^/2 [or t > fi_J.

3. If variances are unknown and unequal: If a preliminary F test

shows the variances to be unequal, no test on means may be
necessary as the two samples are so heterogeneous with respect to
their variances. If, however, a test on means is desired, some tests
are given in statistics books. This situation is often referred to as the
Behrens-Fisher problem. One method is to take as the test
 2.7 Application to tests on means 29

statistic

and check its significance using a special formula for its degrees of
freedom:

B. If dependent or correlated samples: c.

Here one often has the same sample “before and after” some treatment
has been applied. The usual procedure is to take differences between
the first and second observations on the same piece, person, or
whatever, and test the hypothesis that the mean difference is zero.
This reduces the problem to a test on a single mean:

^0- I^D ~~ 0
HI: fij) ^ 0 (/rj5 > 0)
n differences, oc
Test statistic is
d
t = with n — 1 df

Reject if |t| > [or t >

Choose a random sample of n, take two sets of measurements, one on
each unit in the sample, compute the differences. Using these differ-
ences compute d and s^, then t, and render a decision on HQ.

Some examples illustrating the foregoing tests are given in the following.
Example 1 If two samples randomly selected from two independent normal
populations give

X . ^2 = 4, F = 12.0,
2 si = 2.0, =9, Fi = 16.0 s\ = 5.0

is there enough evidence to claim that the mean of population 1 is greater

than the mean of population 2?

^0- ~ 1^2
Hl'lll > fil
(jj = aj (note F = 2.5 < Tg,3(0.975))
30 Review of statistical inference

Reject if
^ ^ ^0.95,11 df ~ 1.V96
16.0 - 12.0
t = = 3.25
8(5) + 3(2) 1 1
11 9^4
Hence reject HQ and claim is greater than /i2-
Example 2 A group of ten students was pretested before instruction and
posttested after six weeks of instruction with the following achievement
scores.

Student Before After Differences = d

d.'
1 14 17 3 ‘\
2 12 16 4 i L
3 20 21 1
i
4 8 10 2
U
5 11 10 -1
\
6 15 14 -1 I
7 17 20 3
8 18 22 4 1
9 9 14 5
10 7 12 5

2^*5
Is there evidence of an improvement in achievement over this six-week
period?
^0- I^D ~ 0
HI: p,j) > 0
Above
d = 2.5, = 2.22

t = —= 3.56
2.22/yio
Since t is >^9,0.95 1.833, we reject HQ and conclude that there has
been an improvement.

2.8 APPLICATION TO TESTS ON PROPORTIONS

I. Tests on a single proportion p.

Ho-P = Po
; p # Po [or p > po or p < Po]
Choose a and n (n had best be large enough so that np > 4), then the
 2.8 Application to tests on proportions 31

test statistic is
P - Po

where ^ “ Po P is the observed proportion in a sample of n.

Here Z is iV(0, 1) and one rejects if Z > Zi-^/2 [or Z > Zi or
z < -Zi-J.
As an example consider whether or not a class of 25 students has an
excessive number of left-handed students if we find ten left-handed students
and the national average is 20 percent.
HQ] p = 0.20
H^ :p > 0.20
n = 25, OL = 0.05
Reject HQ if Z > 1.65. Here

^ ^ 10/25 - 0.20 ^
V(0.2)(0.8)/25
SOwe reject HQ and conclude that our class does have an excessive
number of left-handers.

II. Tests on two independent proportions.

Ha'. Pi = Pi
Hz-Pi # Pi [or Pi > pj]
Choose a, and ^2.
Test statistic is

where p is the proportion of defectives in both samples together:

^iPl + ^2P2
Hi -h ^2

Reject HQ if S Zi-,/2 [or Z > Zi_J

If the number of defectives found in a sample of 100 drawn from a
production process was 12 on Monday and a sample of 200 from the same
process showed 16 defectives on Tuesday, has the process improved?

^o-Pi = Pi
^1- Pi > Pi where p^ is proportion on Monday,
a = 0.05, = 100, ^2 = 200.
32 Review of statistical inference

Reject if Z > 1.64.

Here
12/100 - 16/200
Z = ^ ' ' ■
v'(28/300)(272/300)[l/100 + 1/200]
Z = 1.12

SO one cannot say that the process has improved significantly.

PROBLEMS

2.1 For the following data on tensile strength in psi (pounds per square inch), deter-
mine the mean and variance of this sample:

Tensile Strength
(psi) Frequency

18,461 2
18,466 12
18,471 15
18,476 10
18,481 8
18,486 3

Total 50

2.2 Using the results of Problem 2.1, test the hypothesis that the mean tensile strength
of the population sampled is 18,470 psi (assume that = 40, a = 0.05).
2.3 Plot the operating characteristic curve for Problem 2.2.
2.4 Determine how large a sample would be needed to detect a 10-psi increase in the
mean tensile strength of Problem 2.2 if a = 0.05, f] = 0.02, and = 40.
2.5 Repeat Problem 2.4 for the detection of a 10-psi shift in the mean in either direction.
2.6 For a sample mean of 124 g (grams) based on 16 observations from a population
whose variance is 25 g^ set up 90 percent confidence limits on the mean of the
population.
2.7 For a sample variance of 62 based on 12 observations test the hypothesis that the
population variance is 40. Use a one-sided test and a 1 percent significance level.
2.8 For Problem 2.7 set up 90 percent confidence limits (two-sided) on
2.9 Two samples are taken, one from each of two machines. For this process

Hi =8, L = 42.2 g, sf = 10 g^
^2 = 15, F2 = 44.5 g, si = 18 g^

Test these results for a significant difference in variances.

2.10 Test the results in Problem 2.9 for a significant difference in means.
 Problems 33

2.11 Reflection light box readings before and after dichromating the interior of a metal
cone were

Test
Number 1 2 3 4 5 6 7 8

Before 6.5 6.0 7.0 6.8 6.5 6.8 6.2 6.5

After 4.4 4.2 5.0 5.0 4.8 4.6 5.2 4.9

Test for a significant difference in mean light-box readings.

2.12 To test the hypothesis that the defective fraction p of a process is 0.20, a sample of
100 pieces was drawn at random.
1. Use an a = 0.05 and set up a critical region for the number of observed defectives
to test this hypothesis against a two-sided alternative.

2. If the process now shifts to a 0.10 fraction defective, find the probability of an
error of the second kind for this alternative.
3. Without repeating the work, would you expect the j? error to be the same if the
process shifted to 0.30? Explain.
2.13 Given the following sample data on a random variable, 7: 12, 8, 14, 20, 26, 26, 20,
21, 18, 24, 30, 21, 18, 16, 10, 20. Assuming cr = 7, and a = 0.05, test each of the
following hypotheses:
1. HQ: P = 12
H^ : p > 12
2. Ho : p = 16
H ^; p 7^ 16
3. HQ] P = 18
/flip > 18
2.14 For test 1 in Problem 2.13 evaluate the power of the test with respect to several
alternative hypotheses from p = 13 to p = 24. Plot the power curve.
2.15 Repeat the results of Problem 2.14 with cc — 0.01. Plot on the same graph and
compare the two power curves.
2.16 Given the following data on the current flow in amperes through a cereal forming
machine: 8.2, 8.3, 8.2, 8.6, 8.0, 8.4, 8.5, 7.3. Test the hypothesis of the standard
deviation a = 0.35 ampere versus the alternative that a > 0.35 ampere.

2.17 For the data in Problem 2.16 test the hypothesis that the true mean current flow
is p = 7.0.
2.18 The percent moisture content in a puffed cereal where samples are taken from two
different “guns” showed

Gun I: 3.6, 3.8, 3.6, 3.3, 3.7, 3.4

Gun II: 3.7, 3.9, 4.2, 4.2, 4.9, 3.6, 3.5, 4.0

Test the hypothesis of equal variances and equal means. Use any assumptions you
believe appropriate.
34 Review of statistical inference

2.19 Pretest data for experimental and control groups on course content in a special
vocational-industrial course indicated:
Experimental: = 9.333, = 4.945, = 12
Control: Y2 = 8.375, S2 = 1.187,^2 = 8

Make any statistical tests that the data might suggest. Comment on the results.
2.20 The WISC performance scores of ten boys and eight girls were recorded as follows:

Boys: 101, 86, 72, 129, 99, 118, 104, 125, 90, 107
Girls: 97, 107, 94, 101, 90, 108, 108, 86

Test whether or not there is a significant difference in the means of boys and girls.
Take a = 0.10.
2.21 An occupational therapist conducted a study to evaluate the relative merits of two
prosthetic devices designed to facilitate manual dexterity. The therapist assigned
21 patients with identical handicaps to wear one or the other of the two devices
while performing a certain task. Eleven patients wore device A and ten wore device
B. The researcher recorded the time each patient required to perform a certain
task, with the following results:
— 65 seconds = 81
yg = 75 seconds s| = 64

Do these data provide sufficient evidence to indicate that device A is more effective
than device B1
2.22 An anthropologist believes that the proportion of individuals in two populations
with double occipital hair whorls is the same. To see whether there is any reason
to doubt this hypothesis, the anthropologist takes independent random samples
from each of the two populations and determines the number in each sample with
this characteristic. Results showed:

Population n Number with Characteristic

1 100 23
2 120 32

Is there reason to doubt this hypothesis?

2.23 A psychologist randomly selected ten wives and their husbands from among the
residents of an urban area, and asked them to complete a questionnaire designed
to measure their level of satisfaction with the community in which they lived. Do
the following results indicate that the husbands are better satisfied with the
community than are their wives?

Scores

Wife 33 57 32 54 52 34 60 40 59 39
Husband 44 60 55 68 40 48 57 49 47 52
 Problems 35

2.24 Measurements were made at each of two plants on the tensile strength of samples
of the same type of steel. The data are given below in thousands of psi. Can it be
stated that the steels made in the two plants have the same mean tensile strength?

Plant 1 Plant 2

207 219
195 228
209 222
218 197
196 225
217 184
218
211

2.25 The standard medical treatment for a certain ailment has proved to be about
85 percent effective. One hundred patients were given an alternative treatment
with the result that 78 showed marked improvement. Is the new treatment actually
inferior or could the difference noted be reasonably attributed to chance?
 3
Single-Factor Experiments
with No Restrictions
on Randomization

3.1 INTRODUCTION
In this and several subsequent chapters single-factor experiments will be
considered. In this chapter no restrictions will be placed on the randomiza-
tion so that the design will be completely randomized. Many of the tech-
niques of analysis for a completely randomized single-factor experiment can
be applied with little alteration to more complex experiments.
For example, the single factor could be steel manufacturers, where the
main interest of an analyst centers on the effect of several different manu-
facturers on the hardness of steel purchased from them. It could be tem-
perature in an instance where the experimenter is concerned about the effect
of temperature on penicillin yield. Whenever only one factor is varied,
whether the levels be quantitative or qualitative, fixed or random, the experi-
ment is referred to as a single-factor experiment, and the symbol TJ will be
used to indicate the effect of the jth level of the factor. TJ suggests that the
general factor may be thought of as a “treatment” effect.
If the order of experimentation applied to the several levels of the
factor is completely random, so that any material to which the treatments
might be applied is considered approximately homogeneous, the design is
called a completely randomized design. The number of observations for each
level of the treatment or factor will be determined from cost considerations
and the power of the test. The model then becomes

fj = p Tj + Sfj

where Y^j represents the ith observation (/ = 1, 2,..., nf on the jth treat-
ment (j = 1, 2,..., /c levels). For example, T23 represents the second ob-
servation using level 3 of the factor, p is a common effect for the whole
experiment, TJ represents the effect of the jth treatment, and represents
the random error present in the ith observation on the jth treatment.
The error term 8,^ is usually considered a normally and independently
distributed (NID) random effect whose mean value is zero and whose

36
 3.1 Introduction 37

variance is the same for all treatments or levels. This is expressed as: 8j/s
are NID (0, where is the common variance within all treatments.
is always a fixed parameter, and T2,..., TJ, .. ., are considered to
be fixed parameters if the levels of treatments are fixed. It is also assumed
that
k

J=1

If the k levels of treatments are chosen at random, the T/S are assumed
NID (0, (T^). Whether the levels are fixed or random depends upon how
these levels are chosen in a given experiment.
The analysis of a single factor completely randomized experiment
usually consists of a one-way analysis of variance test where the hypothesis
HQ : TJ = 0 for all j, is tested. If this hypothesis is true, then no treatment
effects exist and each observation Yij is made up of its population mean ji
and a random error Sij. After an analysis of variance (ANOVA), many other
tests may be made, and some of these will be shown in the example that
follows.

Example 3.1 In the manufacture of clothing a wear-testing machine is

used to measure the resistance to abrasion of different fabrics. The dependent
variable Y gives the loss of weight of the material in grams after a specified
number of cycles and the problem is to determine whether or not there is a
difference in the average weight loss among four competing fabrics. Here
there is but one factor of interest: the fabric at levels A, B, C, and D. These
are qualitative and fixed levels of this single factor. It was agreed to test
four samples of each fabric and completely randomize the order of testing
of the 16 samples. The mathematical model for this example then is

Yij = Y Tj &ij

with i = 1, 2,.. ., 4 and j = 1, 2,.. ., 4 since there are four samples for
each of four treatments (fabrics). The data are shown in Table 3.1.

Table 3.1 Fabric Wear Resistance Data

Fabric
A B C D

1.93 2.55 2.40 2.33

2.38 2.72 2.68 2.40
2.20 2.75 2.31 2.28
2.25 2.70 2.28 2.25
38 Single-factor experiments with no restrictions on randomization

An analysis of variance performed on these data gave the results in

Table 3.2.

Table 3.2 Fabric Data ANOVA

Source of Variation df SS MS

Between Fabrics: TJ 3 0.5201 0.1734

Within fabrics or error: e,-,- 12 0.2438 0.0203

Totals 15 0.7639

To test HQ -. Tj 0 for all j = 1, 2, 3, and 4, the test statistic is

0.1734
8.54
0.0203
which is significant at the 1 percent significance level since Appendix Table D
shows 5.95 as the 1 percent F {FQ gg) for 3 and 12 degrees of freedom. We
can then reject the hypothesis and claim that there is a considerable differ-
ence in average wear resistance among the four fabrics.

3.2 ANALYSIS OF VARIANCE RATIONALE

To review the basis for the F test in a one-way analysis of variance, k popula-
tions, each representing one level of treatment, can be considered with
observations as shown in Table 3.3.

Table 3.3 Population Layout for One-Way ANOVA

Treatment
1 2 • • • j • • • k

Fii I'M Y,,

F21 L22 ^2J Y2,
— —
F31 L32

hi Yn Yu

Population means M.i R2 • ■ • Fj ■ ■ • iKk

Here the use of the “dot notation” indicates a summing over all observa-
tions in the population. Since each observation could be returned to the
population and measured, there could be an infinite number of observations
 3.2 Analysis of variance rationale 39

taken on each population, so the average or E{Yn) = ji and so on. ji will

represent the average over all populations, or E{Yij) = /i. In the model TJ,
the treatment effect can also be indicated by ^ — //, and then the model is
either _
Yij = E E Tj + Sij
or
Yij = fi Y ill j — fi) + {Yij — /I j)

This last expression is seen to be an identity true for all values of Y^j. Ex-
pressed another way,
Yij - fi = {fij - fi) Y {Yij - (3.1)

Since these means are unknown, random samples are drawn from each
population and estimates can be made of the treatment means and the grand
mean. If rij observations are taken for each treatment where the numbers
need not be equal, a sample layout would be as shown in Table 3.4.

Table 3.4 Sample Layout for One-Way ANOVA

Treatment
1 2 J k

Ell EI2
... Y^j ••• Eu
E21 E22 ... Y2J E2.

En E^2 ■■■ y'u E-fc

;
E«.2 Ynfji

Totals Ti T2 ••• . ••• Tk

^ -J
Number ^2 ... rj. N
Means E.i Y.2 ••• Y ‘ * Efc F.
-J

Here T j represents the total of the observations taken under treatment j,

Hj represents the number of observations taken for treatment j, and Yj is
the observed mean for treatment j. T represents the grand total of all
observations taken where
k tij k

r. = I I J = 1 I= 1
Yii = Z
j= 1
Tj

and

N= Z nj
 40 Single-factor experiments with no restrictions on randomization

and Y is the mean of all N observations. Note too that

Y.= i njYJN
J=1

If these sample statistics are substituted for their corresponding popu-

lation parameters in Equation (3.1), we get a sample equation (also an
identity) of the form
Yij = + {Yij - YJ (3.2)

This equation states that the deviation of any observation from the grand
mean can be broken into two parts: the deviation of the observation from its
own treatment mean plus the deviation of the treatment mean from the
grand mean.
If both sides of Equation (3.2) are squared and then added over both i
and j, we have

i I
j=l 1= I
(Y, - = i I
j=l 1=1
(Fj - F.)^
j=l 1=1
(Yu - Yjf

+ 2x1; {Yj- YXY,J- Y_j) (3.3)

j=l i=l

Examining the last expression on the right, we find that

y y (F, -
/= 1 ^ = 1
- F,) = y (y, - yj
7= 1
i
i= 1
(Yij - F,)
The term in brackets is seen to equal zero, as the sum of the deviations about
the mean within a given treatment equals zero. Hence

(3.4)
1
^
i= 1
= I I 1 i= 1
(^J - F..)' + Z I
7=1 i =1
(Yij- Y.jf
/0/]no\A9.
This may be referred to as “the fundamental equation of analysis of variance,”
and it expresses the idea that the total surn of squares of deviations from the
grand mean is equal to the sum of squares of deviations between treatment
means and the grand mean plus the sum of squares of deviations within
treatments. In Chapter 2 an unbiased estimate of population variance was
determined by dividing the sum of squares (1^ — Y)^ by the corre-
sponding number of degrees of freedom, — 1. If the hypothesis being
tested in analysis of variance is true, namely, that Zj = 0 for all j, or that
there is no treatment effect, then i ^ i^.2 ^ ‘ ‘ ‘= /kfc
there is in the model is the population mean fi and random error Then,
any one of the three terms in Equation (3.4) may be used to give an unbiased
estimate of this common population variance. For example, dividing the
 3.2 Analysis of variance rationale 41

left-hand term by its degrees of freedom, N — I will yield an unbiased

estimate of population variance Within the jth treatment, Yji ~
divided by Uj — 1 df would yield an unbiased estimate of the variance
within the jth treatment. If the variances within the k treatments are really
all alike, their estimates may be pooled to give YJ='^YJ=I ~
degrees of freedom

y - 1) = N - fc
j
which will give another estimate of the population variance. Still another
estimate can be made by first estimating the variance (Ty between means
drawn from a common population with cry = njGy An unbiased estimate
for (jy is given by “ Y_)^/{k — 1) so that an unbiased estimate of

= i MYj - Y.f/{k - 1)
j=i

which is found by summing the first term on the right-hand side of Equa-
tion (3.4) over i and dividing by /c — 1 df. Thus there are three unbiased
estimates of possible from the data in a one-way ANOVA if the hypothesis
is true. Now all three are not independent since the sum of squares is additive
in Equation (3.4). However, it can be shown that if each of the terms (sums
of squares) on the right of Equation (3.4) is divided by its proper degrees of
freedom, it will yield two independent chi-square distributed unbiased esti-
mates of cr^ when HQ is true. If two such independent unbiased estimates of
the same variance are compared, their ratio can be shown to be distributed
as F with k — 1, A — /c df. If, then, HQ is true, the test of the hypothesis
can be made using a critical region of the F distribution with the observed
F at k — 1 and N — k dt given by

. lUinjiYj - ff/{k- 1}
Z5=i (Yj - Yjf/iN - k)

The critical region is usually taken as the upper tail of the F distribution,
rejecting HQ if T > F^_^ where <x is the area above In this F ratio the
sum of squares between treatments is always put into the numerator, and
then a significant F will indicate that the differences between means has
something in it besides the estimate of variance. It probably indicates that
there is a real difference in treatment means (/i i, 2^ • • •) that HQ
should be rejected. These unbiased estimates of population variance, sums
of squares divided by df, are also referred to as mean squares.
The actual computing of sums of squares indicated in Equation (3.5) is
much easier if they are first expanded and rewritten in terms of treatment
totals. These computing formulas are given in Table 3.5. Applying the
42 Single-factor experiments with no restrictions on randomization

Table 3.5 One-Way ANOVA

Source df SS MS

Between treatments TJ k - 1 2 «/F, - F.)^ ^^treatment/(^ 1)

T^j

Within treatments
or error Sij
N - k it
j= l i=l
SSerror/CA" “
fc Tij h '7^2
=
m " tij

k tij

Totals N - 1 EE
j=l i=l
(F« - F/

k nj rj.2

J=1 i=l

formulas in Table 3.5 to the problem data in Table 3.1, Table 3.6 shows
treatment totals T /s and the sums of squares of the YJ/s for each treatment
along with the results added across all treatments.

Table 3.6 Fabric Wear Resistance Data

Fabric
A B C D

1.93 2.55 2.40 2.33

2.38 2.72 2.68 2.40
2.20 2.75 2.31 2.28
2.25 2.70 2.28 2.25

Ty- 8.76 10.72 9.67 9.26 T. = 38.41

ny. 4 4 4 4 N - 16
k rij
y2.
I
i= 1
ij-
19.2918 28.7534 23.4769 21.4498 HYh = 92.9719
j '

The sums of squares can be computed quite easily from this table. The
total sum of squares states, “Square each observation, add over all observa-
tions, and subtract the correction term.” The last is the grand total squared
 3.2 Analysis of variance rationale 43

and divided by the total number of observations

fc /30 AI)2

SS,<„a, = I. 1 Yfj-^ = 92.9719 - = 0.7639

The sum of squares between treatments is found by totaling Uj observations

for each treatment, squaring this total, dividing by the number of observa-
tions, adding for all treatments, and then subtracting the correction term;
k ’j-'l
SS treatment S
1=1
"

(8.76)^ (10.72)2 (9.67)2 ^ (9 26)2 (38.41)'

= 0.5201
16
The sum of squares for error is then determined by subtraction:

S^error “ ^^total ~ ^^treatment ~ 0.7639 — 0.5201 = 0.2438

These results are displayed as in Table 3.2, and the F test is run on the
HQ : Tj = 0 as shown before.
The computation of these sums of squares can be simplified in two
ways. One could code the data by subtracting a convenient constant such as
2.00 and then multiplying all T’s by 100 to eliminate the decimals. The
subtraction of a constant will not affect the sums of squares but multiplying
by 100 will multiply all SS terms by (100)2 or 10,000. However, since the
test statistic F is the ratio of two mean squares, the F will be unaffected by
this coding scheme.
Since the advent of computers, the original data can be fed into a
computer program for a one-way ANOVA and the computer will print
out the ANOVA table. To illustrate this procedure the programs used in
this text will be SPSS (Statistical Package for the Social Sciences) programs,
although the input and output formats are quite similar for most available
commercial programs. Table 3.7 shows the input for Example 3.1 in an
SPSS one-way ANOVA.
The problem is given an identifying number PROB 7. Line 1 is a code
to identify the programmer. Lines 2 and 3 indicate that an SPSS program
is desired. Line 5 identifies the problem again. Line 6 shows the variables
Y and X herein labeled WL, FAB for wear loss and fabric. Line 7 shows
the number of spaces necessary for the data: 4.0 for WL or Y—including
the decimal point, and 1.0 for the fabric identified by level 1, 2, 3, or 4. Lines
9 through 14 indicate what statistics are to be computed on the data. The
ONEWAY ANOVA is 7 by X or WL by FAB with levels 1 through 4 of
the independent variable. Lines 10-13 will be described later as statistics of
interest to the experimenter. After indicating a desire for all statistics, line 14,
the input data are typed in as shown. (If card input is used, each step, 1-32,
44 Single-factor experiments with no restrictions on randomization

Table 3.7 SPSS Input for Example 3.1

+ + +CREATE PROB7
1. 000 = 77509,113
2. 000 = COMMON (SPSS)
3. 000 = SPSS
4. 000 = #EOR
5. 000 = RUN NAME STAT 502 PROB7
6. 000 = VARIABLE LIST WE,FAB
7. 000 INPUT FORMAT (F4.0, IX, FLO)
8. 000 = N OF CASES 16
9. 000 = ONEWAY WE BY FAB (1,4)/
10. 000 = CONTRAST =10 0 -1/
11. 000 = CONTRAST = 01-10/
12. 000 CONTRAST =1-1-11/
13. 000 = RANGES = SNK
14. 000 STATISTICS AEL
15. 000 = READ INPUT DATA
16. 000 = 1.93 1
17. 000 = 2.38 1
18. 000 = 2.20 1
19. 000 = 2.25 1
20. 000 = 2.55 2
21. 000 = 2.72 2
22. 000 = 2.75 2
23. 000 = 2.70 2
24. 000 2.40 3
25. 000 = 2.68 3
26. 000 = 2.31 3
27. 000 2.28 3
28. 000 = 2.33 4
29. 000 = 2.40 4
30. 000 = 2.28 4
31. 000 = 2.25 4
32. 000 = FINISH
33. 000 = #s

is punched on a separate card.) The program is terminated on line 32 and

it is sent to the computer. In a very short time, a printout is received as in
Table 3.8.
These results are the same as shown in Table 3.2 with an additional
item F PROB giving the probability of this F (8.534) or higher occurring
by chance. So one need not look up a significant F value in the F table.
In addition to the ANOVA table, the information given in Table 3.9
is also printed.
 H
z r- in
■r—1 1—H r-
m
ON

CQ [JH
z ON CN o
oq
T-H

o <N z <!
W
r4 (N (N (N
osi
cu
O
o
o
u
o
HoHo
HoH
[JU ON m m r-
OO oo r- o
OO m (N 1—H
oq in <N
m i-H c4 r4 r4
O ON
l—H
H m
< in
eej od
[JH P o o o o

2.7500
o o o o
l-H oo m oo o
rn r- NO
Z ni (N (N (N
00 <
W
cd
<
p m
a m o P o o O

1.9300
r- <N o
oo T-H o o O o
m m oo m
W Z HH ON m rsi <N
U < z
HH
<N (N r4
Z W
<
»-H s
osi Q
<
^ 00
OP OP
< NO in oo
O (N
Q OP ON ON m
O ^ Z OP o o o O
< w
“
00 S. oo o^ H
00
oo o m m
(N
m (N r--
00
00
<
1^
o
Q z
oP o
DH < < H
HH
(N m NO
00
ON ON <N in
Q < OO OO OO NO
p
Xi 00 rn Z t-H o o
< >
rn "a H w
oo Q
a
'B. P ro <N m oj
X
Z o o in o
2.4006

S Q w
< o o r- in
X c ON oo 1—H H
PQ o w NO rn
U4
u < (N (N r4 (N
-t-j
i2 cij
< >-' Q H
> CQ
OH z
16

O HJ w =S p
o
z u
pcS
Ui
o
o
< w
u
HJ
P
pin
TOTAL

oc O o\
PQ
00 ro ^ <N m rj-
fO <
H-: P QH OH PH PH
2 pi^ 2
2
O OP OP OP OP
SQ < OP
H O O O O O
H >

AS.
46 Single-factor experiments with no restrictions on randomization

This table shows the mean and standard deviation of each group—
or treatment level—1 through 4 and its corresponding standard error (sy =
SyZ-s/n) and 95 percent confidence limits on its true group mean.

3.3 AFTER ANOVA—WHAT?

After concluding, as in the problem above, that there is a difference in
treatment means, we must ask more questions such as: Which treatment is
the best? Does the mean wear resistance of fabric A differ from that of C?
Does the mean of A and B together differ from that of C and D? To answer
questions following an analysis of variance, one must consider the assump-
tions made when the experiment was planned originally. The decision tree
of Figure 3.1 may be helpful in outlining some tests after ANOVA.
In the next sections we examine some of the tests indicated in Figure 3.1.

Figure 3.1 After ANOVA, What?

 3.4 Tests on means \ ,47

3.4 TESTS ON MEANS

In considering qualitative levels of treatments as in Example 3.1, any com-
\
parison of means will depend upon whether the decision on what means to
compare was made before the experiment was run or after the results were
examined.

Tests on Means Set Prior to Experimentation—^Orthogonal Contrasts j

If the above decision is made prior to the running of an experiment, such

comparisons can usually be set without disturbing the risk a of the original
ANOVA. This means that the contrasts must be chosen with care, and the
number of such contrasts should not exceed the number of degrees of free-
dom between the treatment means. The method usually used here is called
the method of orthogonal contrasts.
In comparing treatments 1 and 2 it should seem quite logical to examine
the difference in their means or their totals since all treatment totals are
based on the same sample size n. One such comparison might be T ^ — T2.
If, on the other hand, one wishes to compare the first and second treatment
with the third, the third treatment total should be weighted by a factor of 2
since it is based on only four observations and the treatment totals of 1 and
2 together represent eight observations. Such a comparison would then be
T j -I- T 2 — 27 3. Note that for each of the two comparisons or contrasts,
the sum of the coefficients of the treatment totals always adds to zero:

T.i — T 2 has coefficients +1 — 1=0

Tj + T 2 — 2T 3 has coefficients +1 + 1—2 = 0

Hence a contrast is defined for any linear combination of treatment totals

as follows:

C'm — i + C2rf^,2 +‘ ‘+ Cj^T j + •'■ + <^kmT,k

where
= 0

or, more compactly

c = y c T
j=l
where

E Cjn, = 0
7=1
In dealing with several contrasts it is highly desirable to have the
contrasts independent of each other. When the contrasts are independent—
one having no projection on any of the others—they are said to be orthogonal
contrasts, and independent tests of hypotheses can be made by comparing
48 Single-factor experiments with no restrictions on randomization

the mean square of each such contrast with the mean square of the error term
in the experiment. Each contrast carries one degree of freedom.
It may be recalled that when two straight lines, say
CiiX + ^217 =

^12^ + ^22^ — ^2

are perpendicular, or orthogonal, to each other, the slope of one line is the
negative reciprocal of the slope of the other line. Changing the two lines
above into slope-intercept form.

Y = -C11/C21X + 61/C21
7 = ~ ^12/^22^ + ^2/^22

To be orthogonal,
~^lll^2l — ~{~ ^22/^12)
or
^11^12 — “^ 21^^22
or
^11^12 + <^21^22 ~ 0
The sum of the products of the corresponding coefficients (on X and
on 7) must therefore add to zero. This idea may be extended to a general
definition.
Two contrasts and are said to be orthogonal contrasts, provided
k

= 0
7=1

for equal u’s.

The sums of squares for a contrast are given by

SS

To apply this procedure to the problem above, three orthogonal

contrasts may be set up since there are 3 df between treatments. One such
set of three might be

= - 7.4
C2 = 7.2 - 7.3

C3 = 7.1 - 7.2 - 7.3 + 7.4

Cl is a contrast to compare the mean of the first treatment with the fourth,
C2 compares the second treatment with the third, and C3 compares the
average of treatments 1 and 4 with the average of 2 and 3. The coefficients
of the 7/s for the three contrasts are given in Table 3.10.
 3.4 Tests on means 49

Table 3.10 Orthogonal Coefficients

T.i T.2 T3 T4

Cl +1 0 0 -1
C2 0 +1 -1 0
C3 +1 -1 -1 +1

It can be seen from Table 3.10 that the sum of coefficients adds up
to zero for each contrast, and that the sum of products of coefficients of
each pair of contrasts is also zero.
For the example given in Section 3.2, using the data in Table 3.6, the
contrasts are

Cl = +1(8.76) + 0(10.72) + 0(9.67) - 1(9.26) = -0.50

C2 = 0(8.76) + 1(10.72) - 1(9.67) + 0(9.26) - 1.05
C3 = +1(8.76) - 1(10.72) - 1(9.67) + 1(9.26) = -2.37

The corresponding sums of squares are

(-0.50)^
0.0312
4(2)
(1.05)2
0.1378
4(2)
(-2.37)2
0.3511
^(4F
Total = 0.5201

Note that these three orthogonal sums of squares add up to the treatment
sums of squares. Each contrast has 1 df, so one may test hypotheses implied
by the contrasts as follows:

0.0312/1
= ^^4 or -^1,12 1.54
0.0203
t
0.1378/1
Ho,- ^2 = ^^3 or /^2 = 1^3 ^1,12 6.76
0.0203

0.3511/1
Ho,-Ti + T4. = '^2 + '^3 ■^1,12 17.30
0.0203
or
+4 — 1^2 + 3
50 Single-factor experiments with no restrictions on randomization

Since the F for 1 and 12 df is 4.75 at the 5 percent significance level and 9.33
at the 1 percent significance level, one can reject HQ^ at the 5 percent level
and conclude that the mean wear resistance of fabric 2 (or B) differs from that
of fabric 3 (or C). At the 1 percent level one rejects and concludes that the
mean wear resistances of fabrics 1 and 4 {A and D) differ from that of 2 and 3
{B and C).
If a computer program is used to run the ANOVA, it can be programmed
to compute any desired contrasts. In Table 3.7 lines 10, 11, and 12 indicate
the three contrasts desired. The printout is as shown in Table 3.11.

Table 3.11 Orthogonal Contrasts for Example 3.1

CONTRAST COEFFICIENT MATRIX

GRP 1 GRP 2 GRP 3 GRP 4

CONTRAST 1 1.0 0 0 -1.0

CONTRAST 2 0 1.0 -1.0 0
CONTRAST 3 1.0 -1.0 -1.0 1.0

POOLED VARIANCE ESTIMATE

VALUE S. ERROR T VALUE D.F. T PROS.

CONTRAST 1 -.1250 .1008 -1.2403 12.0 .239

CONTRAST 2 .2625 .1008 2.6046 12.0 .023
CONTRAST 3 -.5925 .1425 -4.1570 12.0 .001

At first glance these results seem different from the calculated values
above; however, the “value” of the contrast given in Table 3.11 is based on
means instead of totals, so multiplying each by rij = 4 yields the contrasts
of —0.5, 1.05, and —2.37. Then the tests are run as t tests instead of F tests.
In the case of 1 df it can be shown that tjf = ^f. Here, for example, for the
second contrast t = 2.6, which squares into 6.76 as seen above.
As this method of orthogonal contrasts is used quite often in experi-
mental design work, the definitions and formulas are given below for the
case of unequal numbers of observations per treatment.
Cjn is a contrast if
k
Z "jCjm = 0
J=1

and and Cq are orthogonal contrasts if

k
I = 0
j= 1
 3.4 Tests on means 51

The sum of squares for such contrasts is given by

Tests on Means After Experimentation

If the decision on what comparisons to make is withheld until after the data
are examined, comparisons may still be made, but the cc level is altered
because such decisions are not taken at random but are based on observed
results. Several methods have been introduced to handle such situations
[17], but only the |^ewimn-I<reuls] range test [15] and Scheffe’s test [21]
are described here.

Range Test After the data have been compiled the following steps are taken.
1. Arrange the k means in order from low to high.
2. Enter the ANOVA table and take the error mean square with its degrees
of freedom.
3. Obtain the standard error of the mean for each treatment

where the error mean square is the one used as the denominator in the
F test on means Yj.
4. Enter a Studentized range table (Appendix, Table E) of significant ranges
at the a level desired, using ^2 = degrees of freedom for error mean
square and p = 2, 3,. .., /c, and list these k — 1 ranges.
5. Multiply these ranges by Sy . to form a group of /c — 1 least significant
ranges.
6. Test the observed ranges between means, beginning with largest versus
smallest, which is compared with the least significant range for p = k;
then test largest versus second smallest with the least significant range
for p = k — 1; and so on. Continue this for second largest versus
sm.allest, and so forth, until all k{k — l)/2 possible pairs have been tested.
The sole exception to this rule is that no difference between two means
can be declared significant if the two means concerned are both contained
in a subset with a nonsignificant range.

To see how this works, consider the data of Example 3.1 as given in
Table 3.6. Following the steps given above:
1. /c = 4 means are 2.19 2.32 2.42 2.68
for treatments: A D C B
52 Single-factor experiments with no restrictions on randomization

2. From Table 3.2 the error mean square is 0.0203 with 12 df.
3. Standard error of a mean is

/0.0203
0.0712

4. From the Appendix, Table E, at the 5 percent level the significant

ranges are, for ^2 = 12,

p: 2 3 4
Ranges: 3.08 3.77 4.20

5. Multiplying by the standard error of 0.0712, the least significant ranges

(LSR) are
p: 2 3 4
LSR: 0.22 0.27 0.30

6. Largest versus smallest: B versus A, 0.49 > 0.30.

Largest versus next smallest: B versus D, 0.36 > 0.27.
Largest versus next largest: B versus C, 0.26 > 0.22.
Second largest versus smallest: C versus A, 0.23 < 0.27.
Second largest versus next largest: C versus D, 0.10 < 0.22.
Third largest versus smallest: D versus A, 0.13 < 0.22.

From the results in step 6 one sees that B differs significantly from
A, D, and C, but A, Z), and C do not differ significantly from each other.
These results are shown in Figure 3.2, where any means underscored by
the same line are not significantly different.

Treatment: D C B

Mean: 2.19 2.32 2.42 2.68

Figure 3.2 Newman-Keuls test means.

Al^
i
Schejfe’s Test Since the Newman-Keuls test is restricted to comparing pairs
of means and it is often desirable to examine other contrasts which represent
combinations of treatments, many experimenters prefer a test devised by
Scheffe [21]. Scheffe’s method uses the concept of contrasts presented earlier
but the contrasts need not be orthogonal. In fact, any and all conceivable
contrasts may be tested for significance. Since comparing means in pairs is a
special case of contrasts, Scheffe’s scheme is more general than the Newman-
Keuls. However, since the Scheffe method must be valid for such a large set
 3.4 Tests on means 53

of possible contrasts, it requires larger observed differences to be significant

than some of the other schemes. To see how Scheffe’s method applies, the
following steps are taken after the data have been compiled.

1. Set up all contrasts of interest to the experimenter and compute their

numerical values.
2. Determine the significant F statistic for the ANOVA just performed
based on a and degrees of freedom k — 1, N — k.
3. Compute A = yj{k — 1)F, using the F from step 2.
4. Compute the standard error of each contrast to be tested. This standard
error is given by
2
Sc = ^(error mean square) jm

5. If a contrast is numerically larger than A times it is declared

significant. Or, if |C^| > Asc^, reject the hypothesis that the true contrast
among means is zero.

Applying this technique to the coded data of Tables 3.1, 3.2, and 3.6,
proceed as follows.

1. Consider two contrasts:

Ci = Ti-r2 =8.76-10.72 =-1.96

C2 = 3 T1 - T 2 - T 3 - T 4 = 26.28 -10.72 - 9.67 - 9.26 = - 3.37

(Note that these are not orthogonal. The first compares treatment A
with B and the second compares the mean of treatment A with the mean
of the other three treatments.)
2. Since a = 0.05, /c - 1 = 3, iV - /c = 12, F = 3.49.
3. A = V3(3.49) = 3.24.
4. Sc, = V0.0203[4(l)^ + 4(-l)^] = 0.40 and Asc, = 1.30.
Sc, = V0.0203 [4(3)^ + 4( - 1 )^ + 4( -1)^ + 4( - 1 )^] = 0.99 and dsc, = 3.21,
5. Since |Ci| = 1.96 is > 1.30, this contrast is significant.
Since C2 = 3.37 is >3.21, this contrast is also significant.

If the computer is used, it can be instructed to carry out the Newman-

Keuls test, called SNK in the SPSS program. Item. 13 in Table 3.7 asks for
this test. The printout is given in Table 3.12.
This table determines subsets among the means and it is seen that
groups 1,4, and 3 (fabrics A, D, and C) form one group and group 2 (fabric B)
forms another, and that the subgroups do not overlap.
54 Single-factor experiments with no restrictions on randomization

Table 3.12 Newman-Keuls Printout for Example 3.1

MULTIPLE RANGE TEST

STUDENT-NEWMAN-KEULS PROCEDURE
RANGES FOR THE .050 LEVEL—
3.08 3.77 4.20
THE RANGES ABOVE ARE TABULAR VALUES.
THE VALUE ACTUALLY COMPARED WITH MEAN(J)-MEAN(I) IS ..
.1008 —RANGE —SQRT(1/N(I) + 1/N(J))
HOMOGENEOUS SUBSETS (SUBSETS OF GROUPS, WHOSE HIGHEST
AND LOWEST MEANS DO NOT DIFFER BY MORE THAN THE
SHORTEST SIGNFICANT RANGE FOR A SUBSET OF THAT SIZE)

SUBSET 1
GROUP GRP 1 GRP 4 GRP 3
MEAN 2.1900 2.3150 2.4175

SUBSET 2
GROUP GRP 2
MEAN 2.6800

3.5 CONFIDENCE LIMITS ON MEANS

After an analysis of variance it is often desirable to set confidence limits on

a treatment mean. The 100(1 — a) percent confidence limits on JLI J are given
by

mean square used to test treatment mean square

where the mean square used to test treatment mean square is the error mean
square in a one-way ANOVA but may be a different mean square in more
complex analyses. The degrees of freedom used with the Student’s t statistic
are the degrees of freedom that correspond to the mean square used for
testing the treatment mean.
In Example 3.1,95 percent confidence limits on the mean wear resistance
for fabric B, /i2 would be

with 12 df on t. By substituting and taking the t value from Appendix Table

B, the limits are 2.68 + 2.18(0.0712) or 2.68 + 0.16 or from 2.52 to 2.84.
 3.6 Components of variance 55

3.6 COMPONENTS OF VARIANCE

In Example 3.1 the levels of the factor were considered as fixed, since only
four fabrics were available and a decision was desired on the effect of these
four fabrics only. If, however, the levels of the factor are random (such as
operators, days, or samples where the levels in the experiment might have
been chosen at random from a large number of possible levels), the model is
called a random model, and inferences are to be extended to all levels of the
population (of which the observed four levels are random samples). In a
random model, the experimenter is not usually interested in testing hypoth-
eses, setting confidence limits, or making contrasts in means, but in estimating
components of variance. How much of the variance in the experiment might
be considered as due to true differences in treatment means, and how much
might be due to random error about these means?

Example 3.2 To see how to analyze and interpret a component of variance

or random model, consider the following problem.
A company supplies a customer with several hundred batches of a raw
material every year. The customer is interested in a high yield from the raw
material in terms of percent usable chemical. He usually makes three sample
determinations of yield from each batch in order to control the quality of the
incoming material. He expects and gets some variation between determina-
tions on a given batch, but he suspects that there may be significant batch-to-
batch variation as well.
To check this, he selects five batches at random from several batches
available and runs three yield determinations per batch. His 15 yield deter-
minations are completely randomized. The mathematical model is again

Yij = fi -\- Tj + Si^j)

except that in this experiment the k levels of the treatment (batches) are
chosen at random, rather than being fixed levels. The data are shown in
Table 3.13.

Table 3.13 Chemical Yield by Batch Data

Batch
1 2 3 4 5

74 68 75 72 79
76 71 77 74 81
75 72 77 73 79

An ANOVA performed on these data gave the results in Table 3.14.

The F test gives F4 jo = 36.94/1.80 = 20.5, which is highly significant.
Since these batches are but a random sample of batches, we may be interested
56 Single-factor experiments with no restrictions on randomization

Table 3.14 Chemical Yield ANOVA

Source df SS MS EMS

Between batches Xj 4 147.74 36.94 cr/ + nG\

Error 10 17.99 1.80

Totals 14 165.73

in how much of the variance in the experiment might be attributed to batch

differences and how much to random error. To help answer these questions,
another column has been added to the ANOVA table (Table 3.14). This is
the expected mean square (EMS) column. It can be derived by inserting the
mathematical model into the operational formulas for mean squares in
Table 3.5 and then computing the expected values of these mean squares.
The results will depend upon whether a fixed or random model is used as
will be discussed in greater length in the next part of this section.
The interpretation of Table 3.14 then is that the error variance GI is best
estimated as 1.80, its corresponding mean square. Also, the mean square
36.94 is an estimate of ol + 3cr^. If the numerical mean squares are set
equal to the variance components that they are estimating,
Sg = 1.80 Sg + 3s^ = 36.94
where is used since these are estimates of the corresponding a^’s.
Solving these expressions.

The total variance can be estimated as

hotal - s? + s? = 11.71 + 1.80 = 13.51

and 11.71/13.51 or 86.7 percent of the total variance is attributable to batch
differences and only 1.80/13.51 or 13.3 percent is attributable to errors within
batches.
It is interesting to note that the above estimate of total variance of 13.51
would give a standard deviation of 3.68. One might therefore expect all the
data in such a small experiment {N — 15) to fall within four standard
deviations or within 4(3.68) = 14.72. The actual range is 13 (high, 81—low,
68). This breakdown of total variance into its two components then seems
quite reasonable.
Some readers may wonder why the total sum of squares is never divided
by its degrees of freedom to estimate the total variance. If the methods of the
next section are used to determine the EMS for the total, it is found to be a
biased estimate of the total variance. In fact, a derivation will show that the
 3.6 Components of variance 57

EMS for the total in a one-way ANOVA is equal to

n(k — 1) . ,
EMS,„„, = (7? + o-£

Since the coefficient of is always less than one, this EMS will be less than the
total variance given above as aj.

Expected Mean Square Derivation

For a single-factor experiment

Yij = fl + Tj £ij (3.7)

From Table 3.5 the sum of squares for treatments with equal n’s is

SS.,eatment = I n{Yj - Y_ ^
J=1
From model Equation (3.7)

= Z Yij/n = i (p + -Cj + &ij)/n

i= 1 i= 1

n
nU HTi A
= -r
n + n ~+ Z i= 1
71

= + Zi= 1
(3.8)

Also,
k ti k fj
Y. = Z Z Ey/n/c = Z Z (/i + Tj + Sij}/nk
j=l i = 1 j i

nkp. V ; I VV / )
nk + «Zj + Zj ZI
k k n

= /^ + Z + ZZ
J I
(3.9)

Subtracting Equation (3.9) from Equation (3.8) gives

Y.j - Y.. = -^j- Z A+ Z

7=1 i =1
Bij/n - Z Z EijAfc
17

Squaring gives
n k

(Y.j - r f = (Tj - z ^ ^

(+ cross products)
58 Single-factor experiments with no restrictions on randomization

Multiplying by n and summing over j gives

SS„eatme„l = L n{Yj - 7,)-

= »! Z bj - I
n k
n
+ ri Z Z e.j “ I Z + « Z (cross product)
^ j \ i i j J j

The expected value operator may now be applied to this SS treatment

EiSS treatment ) = nE IJ \
- IJ
1 k / n n k

+ -E
n Z Z s.j - Z Z s.jA
J \ I ' J

as it can be shown that the expected value of the cross-product term equals
zero.
If the treatment levels are fixed,

j=i
Z = Z (t'.j - /i) = 0
j=i

and the E{SS treatment ) becomes

T(SS treatment
7=1
since errors are random and Z7=i ^ constant. The T(MStreatment) =
T [^SStreatment/(^

k
n{k - 1)
£(MS treatment ) = " Z 1) + (J,
L 7=1
-
n{k - 1)

If, on the other hand, treatment levels are random and their variance
IS (T:

^(MSt.eatment) ==

the EMS term corresponding to the treatments is then either

(jg + n Z - 1)
L 7
or
(Jg + UG^
 3.7 Summary 59

depending upon whether treatments are fixed or random. When fixed, we

shall designate Y,j ~~ 1) as (j)^, so the fixed treatment EMS is
For the error mean square,

SSerror = X I(
7=1 «=1
Subtracting Equation (3.8) from Equation (3.7) gives

^ij ^ij
i= 1
X ^ij/^

Squaring and adding gives

k n k n /

ij 'n
j I j I \ I

Taking the expected value, we have

k n / \2'

-^(S^error) ~ ^ S I (S/J - I eij/n

j i \ i

= 1E
J
I -I n

= Z - 1)'^^
j

= k{n -
and
£(MS„,J = E[SS,„Jk{n - 1)] = af
as shown in Table 3.14.

3.7 SUMMARY
In this chapter consideration has been given to

Experiment Design Analysis

I. Single factor Completely randomized One-way ANOVA

In applying the ANOVA techniques, certain assumptions should be

kept in mind:
1. The process is in control, that is, it is repeatable.
2. The population distribution being sampled is normal.
3. The variance of the errors within all k levels of the factor are homoge-
neous.
 60 Single-factor experiments with no restrictions on randomization

Many texts [4], [9] discuss these assumptions and what may be done if
they are not met in practice.
Several studies have shown that lack of normality in the dependent
variable Y does not seriously affect the analysis when the number of observa-
tions per treatment is the same for all treatments. To check for homogeneous
variances within treatments, several tests have been suggested. One quick
check is to examine the ranges of the observations within each treatment.
For Example 3.1 these are 0.45, 0.20, 0.40, and 0.15. If the average range
R is multiplied by a factor D4, which is found in most quality control texts,
and all ranges are less than D4_R, it is quite safe to assume homogeneous
variances. Here R = 0.30 and D^^R = 0.68 based on samples of four. Since
all four ranges are well below 0.68, homogeneity of variance may be reason-
ably assured. Values of for a few selected sample sizes are given in
Table 3.15.

Table 3.15 D4 Values for Sample Size n

U n = 2 3 4 5 6 7 8 9 10
D4 = 3.267 2.575 2.282 2.115 2.004 1.924 1.864 1.816 1.777

If a more precise test is desired, our computer program includes three

tests: Cochran’s C test, Bartlett-Box F test, and a maximum/minimum
variance test. Two of these show probabilities of obtaining the observed
statistic or a larger value when the variances are homogeneous. If the prob-
abilities are well above 0.05, homogeneity is a safe assumption. Table 3.16
shows the printout for Example 3.1.

Table 3.16 Homogeneity of Variances Tests

COCHRANS C = MAX. VARIANCE/SUM(VARIANCES) = .4406, P = .557 (APPROX.)

BARTEETT-BOX F - 1.265, P = .287
MAXIMUM VARIANCE/MINIMUM VARIANCE = 8.326

With some measured variables it may be that the means of the treat-
ments and the variances within the treatments are related in some way. For
example, a Poisson variable has its mean and variance equal. It can be
shown mathematically that a necessary and sufficient condition for normality
of a variable is that means and variances drawn from that population are
independent of each other. If it is known or observed that sample means and
sample variances are not independent but are related to each other, it may
be possible to “break” this correlation by a suitable transformation of the
original variable Y. Table 3.17 gives several suggested transformations to
try if certain specified relationships hold.
 Problems 61

Table 3.17 Transformations in ANOVA

If is proportional to Transform Lj to

^JYij (Poisson case)

5? r/i - Y,) arc sin yj Y^j (binomial case)
Y.J log Lj or log(Lj + 1)
sj Y^j VY„

Computer programs can be instructed simply to change the original

variable to some transformed variable as shown above and carry out the
ANOVA on the transformed variable.

PROBLEMS

3.1 Assuming a completely randomized design, do a one-way analysis of variance on

the following data in order to familiarize yourself with the technique.

Factor A
Level

1 2 3 4 5

A
Measurement 8 4 1 4 10
6 -2 2 6 8
J 0 0 5 7
5 -2 -1 5 4
8 3 -3 4 9

The cathode warm-up time in seconds was determined for three different tube
types using eight observations on each type of tube. The order of experimentation
was completely randomized. The results were

Tube Type

ABC

Warm-up time 19 20 20 40 16 19
(seconds) 23 20 20 24 15 17
26 18 32 22 18 19
18 35 27 18 26 18

or ;
Do an analysis of variance on these data and test the hypothesis that the three
tube types require the same average warm-up time.
 ■n
P ^ /'-C
0-^ rr^:Xyn^/^^^ i w^

7) ’ -r - reo.. - i-r^ i'v Z . yK y.^ J'" s_ /J

62 Single-factor experiments with no restrictions on randomization

For Problem 3.2 set up orthogonal contrasts between the tube types and test your
contrasts for significance.
3.4 )Set up 95 percent confidence limits for the average warm-up time for tube type
C in Problem 3.2.
3.5 Use the Newman-Keuls range method to test for differences between tube types.
3.6 The following data are on the pressure in a torsion spring for several settings of
the angle between the legs of the spring in a free position.

Angle of Legs of Spring

(degrees)

67 71 75 79 83

Pressure (psi) 83 84 86 89 90
85 85 87 90 92
85 87 90
86 87 91
86 88
87 88
88
88
88
89
90

Assuming a completely randomized design, complete a one-way analysis of vari-

ance for this experiment and state your conclusion concerning the effect of angle
on the pressure in the spring.
3.7 Set up orthogonal contrasts for the angles in Problem 3.6.
3.8 Show that the expanded forms for the sums of squares in Table 3.5 are correct.
3.9 Assume that the levels of factor A in Problem 3.1 were chosen at random and
determine the proportion of variance attributable to differences in level means
and the proportion due to error.
3.10 Verify the results given in Table 3.14 based on the data of Table 3.13.
3.11 Set up two or more contrasts for the data of Problem 3.2 and test their significance
by the Scheffe method.
3.12 Since the Scheffe method is not restricted to equal sample sizes, set up several
contrasts for Problem 3.6 and test for significance by the Scheffe method.
3.13 It is suspected that the environmental temperature in which batteries are activated
affects their activated life. Thirty homogeneous batteries were tested, six at each
of five temperatures, and the data shown below were obtained. Analyze and inter-
pret the data.
 Problems 63

Temperature (°C)

0 25 50 75 100

Activated life 55 60 70 72 65
(seconds) 55 61 72 72 66
57 60 73 72 60
54 60 68 70 64
54 60 77 68 65
56 60 77 69 65

3.14 A highway research engineer wishes to determine the effect of four types of sub-
grade soil on the moisture content in the top soil. He takes five samples of each
type of subgrade soil and the total sum of squares is computed as 280, whereas
the sum of squares among the four types of subgrade soil is 120.
1. Set up an analysis-of-variance table for these results.
2. Set up a mathematical model to describe this problem, define each term in
the model, and state the assumptions made on each term.
3. Set up a test of the hypothesis that the four types of subgrade soil have the
same effect on moisture content in the top soil.
4. Set up a set of orthogonal contrasts for this problem.
5. Explain briefly how to set up a test on means after the analysis of variance
for these data.
6. Set up an expression for 90 percent confidence limits on the mean of type 2
subgrade soil. Insert all numerical values that are known.
3.1^ Data collected on the effect of four fixed types of television tube coating on the
conductivity of the tubes is given as

Coating

II III IV

56 64 45 42
55 61 46 39
62 50 45 45
59 55 39 43
60 56 43 41

Do an analysis of variance on these data and test the hypothesis that the four
coatings yield the same average conductivity.
3.16 For Problem 3.15 set up orthogonal contrasts between tube coatings and test
them for significance.
3.17 jUse a Newman-Keuls range test on the data of Problem 3.15 and discuss any
significant results that you find.
 64 Single-factor experiments with no restrictions on randomization

3.18 Set up contrasts and test using Scheflfe’s method to answer the following questions:
Do the means of coating I and II differ? The means of II and IV? The mean of
I and II versus the mean of III and IV ?
3.19 Explain why confidence limits on a treatment mean as given in Table 3.9 might
differ from the method suggested in Section 3.5. Which method would you recom-
mend under what conditions?
y 3.2(y To determine the effect of several teaching methods on student achievement, 30
students were assigned to five treatment groups with six students per group. The
treatments given for the semester are as follows:

Treatment Description

1 Current textbook
2 Textbook A with teacher
3 Textbook A with machine
4 Textbook B with teacher
5 Textbook B with machine

At the end of a semester of instruction, achievement scores were recorded and

some of the statistics were found to be

Treatment

1 2 3 4 5
Totals 120 600 720 240 420

Source df SS MS F F 0.95

Between treatments 340

Error

Total 465

a. Complete the ANOVA table.

b. Write the mathematical model assumed here and state the hypothesis to be
tested.
c. Test the hypothesis and state your conclusion.
3.21 Set up one set of orthogonal contrasts that might seem reasonable from the treat-
^ment descriptions in Problem 3.20.
3.22 Test whether or not the mean achievement under treatment 1 differs significantly
^=^-^from the mean achievement under treatment 5 in a Newman-Keuls sense for
Problem 3.20.
3.23 Determine the standard error of the contrast 47 ^ — T 2 - T.3 — T4 — T5
based on Problem 3.20 data.
 Problems 65

3.24 Three fertilizers are tried on 27 plots of land in a random fashion such that each
fertilizer is applied to nine plots. The total yield for each fertilizer type is given by

Type

1 2 3
Tj 240 320 180

and the ANOVA table is

Source df SS MS

Between fertilizers 2 1096.30 311

Error 24 1440 60

Total 26 2062

a. Set up one set of orthogonal contrasts that might be used on these data.
b. For your first contrast, determine the sum of squares due to this contrast.
c. For your second contrast, find its standard error.
d. If one wished to compare types 1 and 2 and also the average of 1 and 3 versus
2, which method would you recommend and why?
3.25 Birth weights in pounds of five Poland China pigs were recorded from each of six
randomly chosen litters. The ANOVA printout showed:

Source df SS MS

Between litters 5 6.0 1.20

Within litters 24 9.6 0.40

Determine what percentage of the total variance in this experiment can be attrib-
uted to litter to litter differences.
 4
Single-Factor Experiments —

Randomized Block
and Latin Square Designs

4.1 INTRODUCTION
Consider the problem of determining whether or not different brands of tires
exhibit different amounts of tread loss after 20,000 miles of driving. A fleet
manager wishes to consider four brands that are available and make some
decision about which brand might show the least amount of tread wear after
20,000 miles. The brands to be considered are A, B, C, and D, and although
driving conditions might be simulated in the laboratory, he wants to try these
four brands under actual driving conditions. The variable to be measured is
the difference in maximum tread thickness on a tire between the time it is
mounted on the wheel of a car and after it has completed 20,000 miles on
this car. The measured variable Y^j is this difference in thickness in mils
(0.001 inch), and the only factor of interest is brands, say, Zj where j = 1,2,
3, and 4.
Since the tires must be tried on cars and since some measure of error is
necessary, more than one tire of each brand must be used and a set of four
of each brand would seem quite practical. This means 16 tires, four each
of four different brands, and a reasonable experiment would involve at least
four cars. Designating the cars as I, II, III, and IV, one might put brand T’s
four tires on car I, brand B's on car II, and so on, with a design as shown
in Table 4.1.

Table 4.1 Design 1 for Tire Brand Test

Car
I II III IV

Brand A B C D
distribution A B C D
A B c D
A B c D

66
 4.2 Randomized complete block design 67

Table 4.2 Design 2 for Tire Brand Test

Car

I II III IV

Brand distribution C(12) d(14) C(10) 4(13)

and loss in thickness T(17) T(13) Dill) D{9)
DiU) B(14) B(14) 5(8)
Dili) C(12) B(13) C(9)

One look at this design shows its fallibility, since averages for brands
are also averages for cars. If the cars travel over different terrains, using
different drivers, any apparent brand differences are also car differences.
This design is called completely confounded, since we cannot distinguish
between brands and cars in the analysis.
A second attempt at design might be to try a completely randomized
design, as given in Chapter 3. Assigning the 16 tires to the four cars in a
completely random manner might give results as in Table 4.2. In Table 4.2
the loss in thickness is given for each of the 16 tires. The purpose of complete
randomization here is to average out any car differences that might affect
the results. The model would be

Yij = fi -\- Tj + Sij

with
7 = 1, 2, 3, 4 i = 1, 2, 3, 4

An analysis of variance on these data gives the results in Table 4.3.

The F test shows F3 ^2 = 2.43, and the 5 percent critical region for F
is T3 12 ^ 3.49 (Appendix, Table D), so there is no reason to reject the
hypothesis of equal average tread loss among the four brands.

Table 4.3 ANOVA for Design 2

Source df SS MS

Brands 3 30.69 10.2

Error 12 50.25 4.2

Totals 15 80.94

4.2 RANDOMIZED COMPLETE BLOCK DESIGN

A more careful examination of design 2 in Table 4.2 will reveal some glaring
disadvantages of the completely randomized design in this problem. One
thing to be noted is that brand A is never used on car III or brand B on
68 Single-factor experiments—randomized block and latin square designs

car I. Also, any variation within brand A may reflect variation between cars
I, II, and IV. Thus the random error may not be merely an experimental
error but may include variation between cars. Since the chief objective of
experimental design is to reduce the experimental error, a better design
might be one in which car variation is removed from error variation. Al-
though the completely randomized design averaged out the car effects, it
did not eliminate the variance among cars. A design that requires that each
brand be used once on each car is a randomized complete block design, given
in Table 4.4.

Table 4.4 Design 3: Randomized Block Design for Tire Brand

Test

Car

I II III IV

Brand distribution 5(14) 0(11) 4(13) C(9)

and loss in thickness C(12) C(12) 5(13) 0(9)
4(17) 5(14) 0(11) 5(8)
0(13) 4(14) C(10) 4(13)

In this design the order in which the four brands are placed on a car is
random and each car gets one tire of each brand. In this way better compari-
sons can be made between brands since they are all driven over approximately
the same terrain, and so on. This provides a more homogeneous environment
in which to test the four brands. In general, these groupings for homogeneity
are called blocks and randomization is now restricted within blocks. This
design also allows the car (block) variation to be independently assessed and
removed from the error term. The model for this design is

Y- = p + Pi -\- Tj &ij (4.1)

where pi now represents the block effect (car effect) in the example above.
The analysis of this model is a two-way analysis of variance, since the
block effect may now also be isolated. A slight rearrangement of the data in
Table 4.4 gives Table 4.5.
The total sum of squares is computed as in Chapter 3

4 4

SS total
j I
(193)2
= 2409 - = 80.94
16
 4.2 Randomized complete block design 69

Table 4.5 Randomized Block Design Data for Tire Brand

Test

Brand

Car A B C D T,

I 17 14 12 13 56

II 14 14 12 11 51

III 13 13 10 11 47

IV 13 8 9 9 39

Tj 57 49 43 44 = 193
T.

in
i
823 625 469 492 izn
j '
= 2409

The brand (treatment) sum of squares is computed as usual

T
SS brands
J ^ N

{5lf + (49)^ + (43)2 (44)2 (^93):

= 30.69
16

Since the car (block) effect is similar to the brand effect but totaled across
the rows of Table 4.5, the car sum of squares is computed exactly like the
brand sum of squares, using row totals instead of column totals. Calling
the number of treatments (brands) in general /c, then
n '-pi

SS...
car = T ^
i^
=1k 77"

(56)2 (51)2 (47)2 4_ (39)2 (^93):

= 38.69
16

The error sum of squares is now the remainder after subtracting both brand
and car sum of squares from the total sum of squares

S^error ~ SS^^and ^^car

= 80.94 - 30.69 - 38.69 = 11.56

Table 4.6 is an ANOVA table for these data.

To test the hypothesis, HQ: ^ ^ = ju 2 = 1^.3 = ia.4, the ratio is
10.2
= 7.8
1.3
70 Single-factor experiments—randomized block and latin square designs

Table 4.6 ANOVA for Randomized Block

Design of Tire Brand Test

Source df SS MS EMS

Brands 3 30.69 10.2 + 40,

Cars 3 38.69 12.9
Error 9 11.56 1.3

Totals 15 80.94

which is signihcantly larger than the corresponding critical F even at the

1 percent level (Appendix, Table D). The hypothesis of equal brand means
is thus rejected. It is to be noted that this hypothesis could not be rejected
using a completely randomized design. The randomized block design that
allows for removal of the block (car) effect definitely reduced the error
variance estimate from 4.2 to 1.3.
It is also possible, if desired, to test the hypothesis that the average
tread loss of all four cars is the same. : = 1^3. = and
F3 9= 12.9/1.3 = 9.9, which is also significant at the 1 percent level
(Appendix, Table D). Here this hypothesis is rejected and a car-to-car
variation is detected.
Even though an effect due to cars (blocks) has been isolated, the main
objective is still to test brand differences. Thus it is still a single-factor
experiment, the blocks representing only a restriction on complete ran-
domization due to the environment in which the experiment was conducted.
Other examples include testing differences in materials that are fed into
several different machines, testing differences in fertilizers that must be
spread on several different plots of ground, testing the effect of different
teaching methods on several pupils. In these examples the blocks are
machines, plots, and pupils, respectively, and the levels of the factors of
interest can be randomized within each block.
When data are presented in tabular form there is usually no way to
determine how the data were collected. Was the randomization complete
over all N observations or was the experiment run in blocks with random-
ization restricted to within the blocks? To help in signifying the design of
the experiment, it is suggested that in the case of a completely randomized
design, no horizontal or vertical lines be drawn on the data, as in Table 4.2.
When randomization has been restricted, either vertical or horizontal lines
as shown in Tables 4.4 and 4.5, respectively, can be used to indicate this
restriction on the randomization. As more complex designs are presented,
this scheme may require some double lines for one restriction and single
lines for a second restriction, and so on. It is hoped, however, that such a
 4.2 Randomized complete block design 71

scheme will help the observer in noting just how the randomization has been
restricted.
Since blocking an experiment is a very useful procedure to reduce the
experimental error, a second example may be helpful as it illustrates the
situation in which the same experimental units (animals, people, parts, etc.)
are measured before the experiment and then again at the conclusion of the
experimental treatment. These before-after, pretest-posttest designs can be
treated very well as randomized block designs where the experimental units
are the blocks. Interest is primarily in the effect of the treatment and the
block effects can be removed to reduce the experimental error. Some refer
to these designs as repeated-measures designs because data are repeated on
the same units a second time. The following example will illustrate the
procedure.

Example 4.1 A study on a physical strength measurement in pounds on

seven subjects before and after a specified training period gave the results
shown in Table 4.7.

Table 4.7 Pretest and Posttest Strength Measures

Subjects Pretest Posttest

1 100 115

2 no 125

3 90 105

4 no 130

5 125 140

6 130 140

7 105 125

This problem could be handled by the methods of Chapter 2 using a

t test on differences where d — posttest measure — pretest measure. Here
d = 15.71, = 3.45, and t = 12.05, which is highly significant (p = 0.00002)
with 6 df.
If it is considered as a single-factor experiment at two levels (pre- and
post-) with sev^n blocks making up a randomized block design, the data and
ANOVA appear as shown in Table 4.8.
Here the large F indicates a strong treatment effect. In fact, the F
reported is equal to P given by the method of differences as t = 12.05
when squared (12.05)^ = 145.2.
72 Single-factor experiments—randomized block and latin square designs

Table 4.8 Strength Measure Data and Analysis

Subjects Pretest Posttest T,

1 100 115 215

2 no 125 235

3 90 105 195

4 no 130 240

5 125 140 265

6 130 140 270

7 105 125 230

Totals Tj 770 880 T = 1650

lYfr 85.,850 11 1,600 Y.lYfj= 197,450

i j i

Source df SS MS F Probability

Tests (treatments) 1 864.29 864.29 145 0.00002

Subjects (blocks) 6 2085.71 347.62
Error 6 35.71 5.95

Totals 13 2985.71

4.3 ANOVA RATIONALE

For this randomized complete block design, the model is

== F + ft + (4-2)
or
Yij = M + (/^i. - M) + iji.j - tA + (Yij - Hi, - H.j + A<) (4-3)

where represents the true mean of block 1 The last term can be obtained
by subtracting the treatment and block deviations from the overall deviation
as follows:
[Yij - H) - {Hi. - H) - (H.j - H) = Yij - Hi. - H.J + H

Best estimates of the parameters in Equation (4.3) give the sample model
(after moving Y to the left of the equation):

Yij - F. = {Yi. - f .) + {fj - F.) + {Yij - Yi. - Y.j + FJ

 4.4 Missing values 73

Squaring both sides and adding with i = 2,.. ., nj = 1, 2,.. ., /c,

i i
I= 1 7= 1
{Yu - ry = ii(Yu - Yf + tiifj-
i j i j
F.p
+ iiiYu- Yu - Yj+ Yf
I J

+ 3 cross products (4.4)

A little algebraic work on the sums of the cross products will show that they
all reduce to zero and the remaining equation becomes the fundamental
equation of a two-way analysis of variance. The equation states that

^^total ^^block T T SSgj-j-Qj.

one sum of squares for each variable term in the model [Equation (4.2)].
Each sum of squares has associated with it its degrees of freedom, and
dividing any sum of squares by its degrees of freedom will yield an unbiased
estimate of population variance cr^ if the hypotheses under test are true.
The breakdown of degrees of freedom here is

total blocks treatments error

(nk - 1) = (n - 1) -h (k - 1) + (n - l)(k - 1)
The error degrees are derived from the remainder

(nk — 1) — (n — 1) — (k — 1) = nk — n — k -h 1 = (n — l)(k — 1)

It can be shown that each sum of squares on the right of Equation (4.4)
when divided by its degrees of freedom provides mean squares that are
independently chi-square distributed, so that the ratio of any two of them
is distributed as F.
The sums of squares formulas given in Equation (4.4) are usually
expanded and rewritten to give formulas that are easier to apply. These are
shown in Table 4.9. They are the formulas applied to the data of Table 4.5
where nk = N.

4.4 MISSING VALUES

Occasionally in a randomized block design an observation is lost. A vial

may break, an animal may die, or a tire may disintegrate, so that there
occurs one or more missing observations in the data. For a single-factor
completely randomized design this presents no problem, since the analysis
of variance can be run with unequal n/s. But for a two-way analysis this
means a loss of orthogonality, since for some blocks the Yjj U longer
equals zero, and for some treatment the Pi longer equals zero. When
74 Single-factor experiments—randomized block and latin square designs

Table 4.9 ANOVA for Randomized Block Design

Source df SS MS
n T' 2 T2
Between n — \ S^block/(^ 1)
blocks Pi nk
k ’’p 2 T2
Between k - 1
E — - SStreatment/(^ 1)
treatments TJ j nk
n k ^ 2
Error 8ij {n - m- 1)
Ei E -
j i=l ^
k
-E^
r 2
Li + L.
^ 2
SS„„,/(n - 1)(A: - 1)

J=l n nk

n k
Totals nk — \
Ei Ej nk

blocks and treatments are orthogonal, the block totals are added over all
treatments, and vice versa. If one or more observations are missing, the usual
procedure is to replace the value with one that makes the sum of the squares
of the errors a minimum.
In the tire brand test example suppose that the brand C tire on car III
blew out and was ruined before completing the 20,000 miles. The resulting
data after coding by subtracting 13 mils appear in Table 4.10, where y is
inserted in place of this missing value.

Table 4.10 Missing Value Example

Brand

Car A B C D T,

I 4 1 -1 0 4
II 1 1 -1 -2 -1
III 0 0 y -2 .V - 2
IV 0 -5 -4 -4 -13

T., 5 -3 V — 6 -8 y - 12 = T

Now,
SS error = SS total SS treatment - SS block

Tl
+
n k nk
 4.4 Missing values 75

For this example,

S^error — 4^ + 1 ^ + ' ’ ’ + ‘ • ' + (— 4)^

_ (5)^ + (-3)^ + (y - 6f + (-8)^
4
(4)2 + (_1)2 + (3, _ 2)^ + (_13)2 ^ (^ _ 12)2
4 +~16
To find the y value that will minimize this expression, it is differentiated with
respect to y and set equal to zero. As all constant terms have their derivatives
zero.

2{y - 6) 2(y - 2) 2(y - 12)

- = 2v ^ ^ = 0
dy 16
Solving gives

16_y — 4y + 24 — 4y + S -y y — 12 = 0

9y = -20

20
y —2.2 or 10.8 mils
y

If this value is now used in the y position, the resulting ANOVA table is
as shown in Table 4.11. This is an approximate ANOVA as the sums of
squares are slightly biased. The resulting ANOVA is not too different from
before, but the degrees of freedom for the error term are reduced by one,
since there are only 15 actual observations, and y is determined from these
15 readings.
This procedure can be used on any reasonable number of missing values
by differentiating the error sum of squares partially with respect to each
such missing value and setting it equal to zero, giving as many equations as
unknown values to be solved for these missing values.

Table 4.11 ANOVA for Tire Brand Test

Example Adjusted for a
Missing Value (approximate)

Source df SS MS

Brands TJ 3 28.7 9.5

Cars fii 3 38.3 12.7
Error 8 11.2 1.4

Totals 14 78.2
76 Single-factor experiments—randomized block and latin square designs

In general, for one missing value y^j it can be shown that

nTl + kVj - r.
(4.5)
" (n - m - 1)

where the primed totals , T'p and T' are the totals indicated without the
missing value y. In our example,
4(-2) + 4(-6)-(-12)
yij — y 33 —
(3)(3)

y33 = = —2.2 (as before)

y
f •

' '\ ^) 0 - ‘*

4.5 LATIN SQUARES ^

The reader may have wondered about a possible position effect in the prob-
lem on testing tire brands. Experience shows that rear tires get different wear
than front tires and even different sides of the same car may show different
amounts of tread wear. In the randomized block design, the four brands
were randomized onto the four wheels of each car with no regard for posi-
tion. The effect of position on wear could be balanced out by rotating the
tires every 5000 miles giving each brand 5000 miles on each wheel. However,
if this is not feasible, the positions can impose another restriction on the
randomization in such a way that each brand is not only used once on each
car but also only once in each of the four possible positions: left front, left
rear, right front, and right rear.
A design in which each treatment appears once and only once in each
row (position) and once and only once in each column (cars) is called a
Latin square design. Interest is still centered on one factor, treatments, but
two restrictions are placed on the randomization. An example of one such
4x4 Latin square is shown in Table 4.12.
Such a design is only possible when the number of levels of both restric-
tions equals the number of treatment levels. In other words, it must be a

Table 4.12 4x4 Latin Square Design

Car

Position I II III IV

1 C D A B
2 B C D A
3 A B C D
4 D A B C
 4.5 Latin squares 77

square. It is not true that all randomization is lost in this design, as the
particular Latin square to be used on a given problem may be chosen at
random from several possible Latin squares of the required size. Tables of
such squares are found in Fisher and Yates [11].
The analysis of the data in a Latin square design is a simple extension
of previous analyses where the data are now added in a third direction—
positions. If the data of Table 4.5 were imposed on the Latin square of
Table 4.12, the results could be those shown in Table 4.13.

Table 4.13 Latin Square Design Data on Tire Wear

Car

Position I II III IV T.k

1 C 12 D 11 A 13 B 8 44

2 B 14 C 12 D 11 A 13 50

3 A 17 B 14 C 10 D 9 50

4 D 13 A 14 B 13 C 9 49

T
^ ..1
56 51 47 39 193

Treatment totals for A, B, C, and D brands are 57, 49, 43, and 44 as
before, where the model is now

Yijk — M + ft + L' + + ^ijk

and represents the positions effect. Since the only new totals are for
positions, a position sum of squares can be computed as

(44)^ + (50)2 ^ (49^2 (^93):

SS position = 6.19
16
and
^^error ^^total ^^brand ^^car
SS.position
= 80.94 - 30.69 - 38.69 - 6.19 = 5.37

Table 4.14 is the ANOVA table for these data.

Once again another restriction placed on the randomization has further
reduced the experimental error, although the position effect is not significant
at the 5 percent level. This further reduction of error variance is attained at
the expense of degrees of freedom, since now the estimate of is based on
only 6 df instead of 9 df as in the randomized block design. This means less
precision in estimating this error variance. But the added restrictions should
be made if the environmental conditions suggest them. After discovering
78 Single-factor experiments—randomized block and latin square designs

Table 4.14 Latin Square ANOVA

Source df SS MS EMS

Brands TJ 3 30.69 10.2 aj + 4(/),

Cars 3 38.69 12.9 + 4(7^
Positions 3 6.19 2.1 (jg + 40
Error 8ij„ 6 5.37 0.9

Totals 15 80.94

that position had no significant effect, some investigators might “pool” the
position sum of squares with the error sum of squares and obtain a more
precise estimate of namely, 1.3, as given in Table 4.6. However, there
is a danger in “pooling,” as it means “accepting” a hypothesis of no position
effect, and the investigator has no idea about the possible error involved in
“accepting” a hypothesis. Naturally, if the degrees of freedom on the error
term are reduced much below that of Table 4.14, there will have to be some
pooling to get a reasonable yardstick for assessing other effects.

4.6 INTERPRETATIONS
After any ANOVA one usually wishes to investigate the single factor further
for a more detailed interpretation of the effect of the treatments. Here, after
the Latin square ANOVA of Table 4.14, we are probably concerned about
which brand to buy. Using a Newman-Keuls approach, the results are as
follows:
1. Means: 10.75 11.00 12.25 14.25
For brands : C D B A
2. = 0.9 with 6 df.
/0.9
3. Sy
J-
— = 0.47

4. For p = 2 3 4
Tabled ranges: 3.46 4.34 4.90
5. LSRs: 1.63, 2.04, 2.30.
6. Testing averages:
A-C - 3.50 > 2.30*
A-D : 3.25* > 2.04*
A-B : 2.00 > 1.63*
B-C - 1.50 < 2.04
 4.7 Graeco-latin squares 79

Thus only A differs from the other brands in tread wear. A has the largest
tread wear and so is the poorest brand to buy. One could recommend
brand B, C, or D, whichever is the cheapest or whichever is “best” according
to some criterion other than tread wear.

4.7 GRAECO-LATIN SQUARES

In some experiments still another restriction may be imposed on the ran-

domization. The design may be a Graeco-Latin square such as Table 4.15
exhibits.

Table 4.15 Graeco-Latin Square Design

Car

Position I II III IV

1 A(x Cy DS
2 By Ad Dec Cfi
3 C3 Dy Al^ B<x
4 DjS Ca BS Ay

In this design the third restriction is at levels a, y, d, and not only do

these each appear once and only once in each row and each column, but
they appear once and only once with each level of treatment T, B, C, or D.
The model for this would be

Yijkm = At + ft + L + + ^ijkm

where co^ is the effect of the latest restriction with levels a, y, and S. An
outline of the analysis appears in Table 4.16.

Table 4.16 Graeco-Latin

ANOVA Outline

Source df

l^i 3
L- 3
yk 3
COfti 3
^ijkm 3

Total 15
80 Single-factor experiments—randomized block and latin square designs

Such a design may not be very practical, as only 3 df are left for the
error variance.

4.8 EXTENSIONS
In some situations it is not possible to place all treatments in every block
of an experiment. This leads to a design known as an incomplete bloek
design. Sometimes a whole row or column of a Latin square is missing and
the resulting incomplete Latin square is called a Youden square. These two
special designs are discussed in Chapter 16.

4.9 SUMMARY

Experiment Design Analysis

I. Single factor Completely randomized 1. One-way ANOVA

Yij ^ fi + Tj + £ij
Randomized block 2.
Yij = pi Tj di + ^ij

a. Complete a. Two-way ANOVA

b. Incomplete, balanced b. Special ANOVA
c. Incomplete, general c. Regression method
Latin square 3.
Yjk = p + di + c
+ Tfc + ^ijk
a. Complete a. Three-way ANOVA
b. Incomplete, b. Special ANOVA
Youden square (like 2b)
4. Graeco-Latin square 4. Four-way ANOVA
yi jkm
+ E’
+ S ijkm

It might be emphasized once again that, so far, interest has been centered
on a single factor (treatments) for the discussion in Chapters 3 and 4. The
special designs simply represent restrictions on the randomization. In the
next ehapter, two or more factors are considered.

PROBLEMS

4.1 The effects of four types of graphite coaters on light box readings are to be studied.
As these readings might differ from day to day, observations are to be taken on
each of the four types every day for three days. The order of testing of the four
 Problems 81

types on any given day can be randomized. The results are

Graphite Coater Type

Day M A K L

1 4.0 4.8 5.0 4.6

2 4.8 5.0 4.6

3 4.0 4.8 5.6 5.0

Analyze these data as a randomized block design and state your conclusions.
4.2 Set up orthogonal contrasts among coater types and analyze for Problem 4.1.
4.3 Use the Newman-Keuls range test to compare four coater-type means for Problem
4.1.
4.4 If the reading on type K for the second day was missing, what missing value
should be inserted and what is the analysis now? Jud-iT 'lOioet..
4.5 Set up at least four contrasts in Problem 4.1 and use the Scheffe method to test
the significance of these.
4.6 In a research study at Purdue University on metal-removal rate, five electrode
shapes A, B, C, D, and E were studied. The removal was accomplished by an
electric discharge between the electrode and the material being cut. For this
experiment five holes were cut in five workpieces, and the order of electrodes
was arranged so that only one electrode shape was used in the same position on
each of the five workpieces. Thus, the design was a Latin square design, with
workpieces (strips) and positions on the strip as restrictions on the randomization.
Several variables were studied, one of which was the Rockwell hardness of metal
where each hole was to be cut. The results were as follows:

Position

Strip 1 2 3 4 5

I A{64) 5(61) C(62) 5(62) £(62)

II 5(62) C(62) 5(63) £(62) A{63)

III C(61) D(62) £(63) T(63) 5(63)

IV D(63) £(64) A{63) 5(63) C(63)

V £(62) A(61) 5(63) C(63) 5(62)

Analyze these data and test for an electrode effect, position effect, and strip effect
on Rockwell hardness.
82 Single-factor experiments—randomized block and latin square designs

4.7 The times in hours necessary to cut the holes in Problem 4.6 were recorded as
follows:

} . Position

Strip 1 2 3 4 5

I A{3.5) 5(2.1) C(2.5) 5(3.5) £(2.4)

II £(2.6) A{33) 5(2.1) C(2.5) 5(2.7)

III D{2.9) £(2.6) A(3.5) 5(2.7) C(2.9)

IV C(2.5) 5(2.9) £(3.0) A{3.3) 5(2.3)

V 5(2.1) C(2.3) 5(3.7) £(3.2) A{3.5)

Analyze these data for the effect of electrodes, strips, and positions on time.
4.8 Analyze the electrode effect further in Problem 4.7 and make some statement as
to which electrodes are best if the shortest cutting time is the most desirable factor.
4.9 For an m x m Latin square, prove that a missing value may be estimated by

^ ^ m(r:,. + r,, + rj - 2r..

(m — l)(m — 2)

where the primes indicate totals for row, treatment, and column containing the
missing value and T'.. is the grand total of all actual observations.
4.10 A composite measure of screen quality was made on screens using four lacquer
concentrations, four standing times, four acryloid concentrations {A, B, C, D), and
four acetone concentrations (a, P, F, A). A Graeco-Latin square design was used
with data recorded as follows:

Lacquer Concentration

Standing Time ^ 1 li 2

30 C)5(16) 5F(12) 5A(17) Ta(ll)

20 5a(15) CA(14) AF(15) 5)9(14)

10 AA(12) 5a(6) sm) CF(13)

5 5F(9) Am Ca(8) 5A(9)

Do a complete analysis of these data.

4.11 Explain why a three-level Graeco-Latin square is not a feasible design.
 Problems 83

4.12 In a chemical plant, five experimental treatments are to be used on a basic raw
material in an effort to increase the chemical yield. Since batches of raw material
may differ, five batches are chosen at random. Also, the order of the experiments
may affect yield as well as may the operator who performs the experiment. Con-
sidering order and operators as further restrictions on randomization, set up a
suitable design for this experiment that will be least expensive to run. Outline its
analysis.
4.13 Three groups of students are to be tested for the percentage of high-level questions
asked by each group. As questions can be on various types of material, six lessons
are taught to each group and a record is made of the percentage of high-level
questions asked by each group on all six lessons. Show a data layout for this situa-
tion and outline its ANOVA table.
N/ 4.14 Data from the results of Problem 4.13 were as follows:

Group

Lesson A B C

1 13 18 7

2 16 25 17

3 28 24 14

4 26 13 15

5 27 16 12

6 23 19 9

Do an analysis of these data and state your conclusions.

4.15 For Problem 4.14 analyze further in an attempt to determine which group asks
the highest percentage of high-level questions.
4.16 Thirty students were pretested for science achievement and were subsequently
posttested using the same test after 15 weeks of special instruction. A t test on the
difference scores of those students was found to be 2.3. Consider this problem as
a randomized block design and show a data layout and outline its ANOVA table
with the proper degrees of freedom. Also indicate what F value should be exceeded
if one were to conclude that the training made a difference at the 5 percent signif-
icance level.
4.17 A student decides to investigate the accuracy of five scales in various local com-
mercial places. She decides to account for the variations in her weight at different
times of day and also on different days. She will only collect data on Saturdays,
which she has free. The design will be a Latin square so she chooses the five times
of day as 9:00 a.m., 11:00 a.m., L:00 p.m., 3:00 p.m., and 5:00 p.m. One possible
84 Single-factor experiments—randomized block and latin square designs

Latin square design is as follows:

Time

Day 9 11 1 3 5

1 A B C D E
2 B C D E A
3 C D E A B
4 D E A B C
5 E A B C D

a. Give the model and write out the ANOVA table.

b. Suppose she does a Scheffe test on the times of day and finds that the 3 and 5
are not significantly different. If she is going to do the experiment again on some
other scales, should she do a simpler randomized block design with five days
as blocks and make observations at 3 p.m. and 5 p.m.? Give reasons for your
decision.
4.18 Workers at an agricultural experiment station conducted an experiment to com-
pare the effect of four different fertilizers on the yield of a cane crop. They divided
four blocks of soil into four plots of equal size and shape and assigned the fertilizers
to the plots at random such that each fertilizer was applied once in each block.
Data were collected on yield in hundred weights per plot. The analysis gave a
total sum of squares of 540 with fertilizer sum of squares at 210 and block sum of
squares at 258. Set up an ANOVA table for these results.
4.19 From the data of Problem 4.18 state the hypothesis to be tested in terms of the
mathematical model, run the test, and state your conclusions.
4.20 In Problem 4.18 if the observed mean yield of cane for fertilizer B was 41.25 hun-
dred weight, what is its standard error?
4.21 In Problem 4.18 the original data show that all readings are divisible by 5. If one
codes the data by subtracting 40 and dividing by 5, set up an expression for 95
percent confidence limits on the mean of fertilizer B in terms of the coded data.

4.22 Write a contrast to be used in comparing the mean of fertilizers A and B with the
mean of fertilizer D and also show its standard error for Problem 4.18.
4.23 Set up three contrasts where one contrast is orthogonal to the other two but
these two are not orthogonal to each other.
4.24 a. What is the main advantage of a randomized block design over a completely
randomized design?
b. Under what conditions would the completely randomized design have been
better?
4.25 Material is analyzed for weight in grams from three vendors—A, B, C—by three
different inspectors—I, II, III—and using three different scales—1, 2, 3. The
experiment is set up on a Latin square plan with results as follows:
 Problems 85

Scale

Inspector 1 2 3

I A = 16 B = 10 C = 11

II B = 15 C = 9 A = 14

III C = 13 A = n B = 13

Analyze the experiment to test for a difference in weight between vendors, between
inspectors, and between scales.
 5
Factorial Experiments

5.1 INTRODUCTION
In the preceding three chapters all of the experiments involved only one
factor and its effect on a measured variable. Several different designs were
considered, but all of these represented restrictions on the randomization
where interest was still centered on the effect of a single factor.
Suppose there are now two factors of interest to the experimenter, for
example, the effect of both temperature and altitude on the current flow in a
small computer. One traditional method is to hold altitude constant and vary
the temperature and then hold temperature constant and change the altitude,
or, in general, hold all factors constant except one and take current flow
readings for several levels of this one factor, then choose another factor to
vary, holding all others constant, and so forth. To examine this type of
experimentation, consider a very simple example in which temperature is
to be set at 25°C and 55°C only and altitudes of 0 K (K = 10,000 feet) and
3 K (30,000 feet) are to be used. If one factor is to be varied at a time, the
altitude may be set for sea level or 0 K and the temperature varied from 25°C
to 55°C. Suppose the current flow changed from 210 mA to 240 mA with
this temperature increase. Now there is no way to assess whether or not this
30-mA increase is real or due to chance. Unless there is available some pre-
vious estimate of the error variability, the experiment must be repeated in
order to obtain an estimate of the error or chance variability within the
experiment. If the experiment is now repeated and the readings are 205 mA
and 230 mA as the temperature is varied from 25°C to 55°C, it seems obvious,
without any formal statistical analysis, that there is a real increase in current
flow, since for each repetition the increase is large compared with the varia-
tion in current flow within a given temperature. Graphically these results
appear as in Figure 5.1.
Four experiments have now been run to determine the effect of tem-
perature at 0 K altitude only. To check the effect of altitude, the temperature
can be held at 25°C and the altitude varied to 3 K by adjustment of pressure

86
 5.1 Introduction 87

t
t yy'

Altitude
' .V/

Temperature (°C)

Figure 5.1 Temperature effect on current flow.

in the laboratory. Using the results already obtained at 0 K (assuming they

are representative), two observations of current flow are now taken at 3 K
with the results shown in Figure 5.2.

Altitude

Temperature (°C)

Figure 5.2 Temperature and altitude effect on current flow.

From these experiments the temperature increase is seen to increase the

current flow an average of
30 mA + 25 inA
27.5 mA
2
and the increase in altitude decreases the current flow on the average of
30 mA + 20 mA
— = 25 mA
2
This information is gained after six experiments have been performed and
no information is available on what would happen at a temperature of 55°C
and an altitude of 3 K.
An alternative experimental arrangement would be a factorial arrange-
ment where each temperature level is combined with each altitude and only
88 Factorial experiments

four experiments are run. Results of four such experiments might appear as
in Figure 5.3.

Figure 5.3 Factorial arrangement of temperature and altitude effect on current flow.

With this experiment, one estimate of temperature effect on current flow

is 240 — 210 mA = 30 mA at 0 K, and another estimate is 200 — 180mA =
20 mA at 3 K. Hence two estimates can be made of temperature effect
[average (30 + 20)/2 = 25 mA] using all four observations without neces-
sarily repeating any observation at the same point. Using the same four
observations, two estimates of altitude effect can be determined: 180 — 210 =
— 30 mA at 25°C, and 200 — 240 = — 40 mA at 55°C, an average decrease
of 35 mA for a 3-K increase in altitude. Here, with just four observations
instead of six, valid comparisons have been made on both temperature and
altitude, and in addition some information has been obtained as to what
happens at 55°C and altitude 3 K.
From this simple example, some of the advantages of a factorial experi-
^^nt can be seen:

'ff.^ore efficiency is possible than with one-factor-at-a-time experiments

(here four sixths or two thirds the amount of experimentation).
2. All data are used in computing both effects. (Note that all four observa-
tions are used in determining the average effect of temperature and the
average effect of altitude.)
3. Some information is gleaned on possible interaction between the two
factors. (In the example, the increase in current flow of 20 mA at 3 K
was about the same order of magnitude as the 30 mA increase at 0 K.
If these increases had differed considerably, interaction might be said to
be present.)
These advantages are even more pronounced as the number of levels of
the two factors is increased. A factorial experiment is one in which all levels
 5.1 Introduction 89

of a given factor are combined with all levels of every other factor in the
experiment. Thus, if four temperatures are considered at three altitudes, a
4x3 factorial experiment would be run requiring 12 different experimental
conditions. In the example above, if the current flow were 160 mA for 55°C
and 3 K, the results could be shown as in Figure 5.4.

0-A^

3-K - •ISO •160

Altitude

0- K - #210 #240

25 55
Temperature (°C)

(a)

Figure 5.4 Interaction in a factorial experiment.

From Figure 5.4(a) note that as the temperature is increased from

25 to 55°C at 0 K, the current flow increases by 30 mA, but at 3 K for the
same temperature increase the current flow decreases by 20 mA. When a
change in one factor produces a different change in the response variable at
one level of another factor than at other levels of this factor, there is an
interaction between the two factors. This is also observable in Figure 5.4(b)
as the two altitude lines are not parallel. If the data of Figure 5.3 are plotted,
we get the relations pictured in Figure 5.5.

Figure 5.5 No-interaction temperature diagram in a temperature-altitude study.

90 Factorial experiments

It is seen that the lines are much more nearly parallel and no interaction
is said to be present. An increase in temperature at 0 K produces about the
same increase in current flow (30 mA) as at 3 K (20 mA). Now a word of
warning is necessary. Lines can be made to look nearly parallel or quite
diverse depending on the scale chosen; therefore, it is necessary to run
statistical tests in order to determine whether or not the interaction is statis-
tically significant. It is only possible to test the significance of such interac-
tion if more than one observation is taken for each experimental condition.
The graphic procedure shown above is given merely to provide some insight
into what interaction is and how it might be displayed for explaining the
factorial experiment.
In addition to the advantages of a factorial experiment, it may be that
when the response variable 7 is a yield and one is concerned with which
combination of two independent variables and X2 will produce a maxi-
mum yield, the factorial will give a combination near the maximum whereas
the one-factor-at-a-time procedure will not do so. Figure 5.6 shows two
cases in which the contours indicate the yields.

^2 V2

Figure 5.6 Yield contours using one-factor-at-a-time method.

In case (a), X2 is set and X^ is varied until the maximum yield is found,
then X^ is held there while X2 is varied and a maximum is found near the
top of the mound where the yield is around 60 units. But in case (b), the
same procedure misses the maximum by a long way, giving about a 37-unit
yield.
If, now, a factorial arrangement is used, one first notes the extremes of
each independent variable and may also choose one or more points between
these extremes. If this is done for the diagrams in Figure 5.6, Figure 5.7
shows that in both cases the maximum is reached. In some cases the maximum
may not have been reached but the factorial will indicate the direction to be
followed in the next experiment to get closer to such a maximum. This
concept is presented in Chapter 16.
 5.2 Factorial experiments 91

A', ^2

(a) Mound (b) Rising ridge

Figure 5.7 Yield contours using a factorial method.

5.2 FACTORIAL EXPERIMENTS

To see how a factorial experiment is set up and analyzed, consider Example

5.1.

Example 5.1 To determine the effect of exhaust index (in seconds) and
pump heater voltage (in volts) on the pressure inside a vacuum tube (in
microns of mercury), three exhaust indexes and two voltages are chosen at
fixed levels. It was decided to run two experiments at each of these six
treatment conditions (three exhaust indexes x two voltages). The order for
running the 12 experiments was completely randomized. This can be accom-
plished easily by labeling the six treatment conditions numbers 1 through 6
and then tossing a die to decide the order in which the experiments are to
be run.
Table 5.1 shows the results of one such complete randomization for the
12 experiments. In such a procedure some treatment combinations may be
run twice before others are run once, but this is the nature of complete

Table 5.1 Data Layout for Vacuum Tube Pressure

Experiment

Pump Heater Exhaust Index (seconds)

Voltage
(volts) 60 90 150

127 JJ 4 2J 1 iJ 6
9 11 12
220 4J 2 3 u 7
8 5 10
92 Factorial experiments

randomization. If one were to randomize six runs for each combination and
then randomize six more, the randomization will have been restricted and
no longer a completely randomized design. In such a procedure it is also
true that some numbers may come up again on the die even though that
combination already has been run twice. In these cases the third time is
ignored and the flipping continues until all treatment combinations are run
twice. This is, of course, a slight restriction to keep the number of observations
per treatment constant. The resulting experiment may be labeled a 3 x 2
factorial experiment with two observations per treatment with all N = 12
observations run in a completely randomized order. Of course, this random
order can be determined before the experiment is actually run and it is
important to supervise the experiment to make sure that it is actually run
in the order prescribed.
The mathematical model for this experiment can be written as

^ijk T Tjj -f- ^k{ij) (^•^)

where

i = 1, 2, 3 for the three exhaust indexes

j = 1, 2 for the two voltages
k = 1, 2 for the two observations in each ij treatment combination
Tij = the six treatment effects
8^ij) = the error within each of the six treatments

The notation k{ij) indicates that the errors are unique to each i,j combination
or are nested within each ij. Following the analysis procedures of Chapter 3,
Table 5.2 shows the resulting data in a between-treatments, within-treatments
format after multiplying all results by 1000 to eliminate decimal points.
If the data of Table 5.2 are subjected to a one-way ANOVA, the results
are as given in Table 5.3.

Table 5.2 Vacuum Tube Pressure as a One-Way Experiment

Fleo’ FI27 ElgO’ FI27 F/I5O, FI27 T/^Q, F22O Elgo-' F220 El 1 50’ k220

48 28 7 62 14 6
58 33 15 54 10 9

8r 106 61 22 116 24 15
lyfj. 5668 1873 274 6760 296 117
k
T. == 344

LIZ
i j k
V. == 14,988
 5.2 Factorial experiments 93

Table 5.3 One-Way ANOVA for Vacuum

Tube Pressure Experiment

Source df SS

Between treatments 5 4987.67

Within treatments 6 139.00

Totals 11 5126.67

In Table 5.3 the treatment effects are really combinations of exhaust

indexes and voltages, and if rearranged to show these main effects, the data
might appear as in Table 5.4.

Table 5.4 Two-Way Layout for Vacuum Tube

Pressure Experiment

Exhaust Index
Pump Heater
Voltage 60 90 150 T-,

127 48 28 7
189
58 33 15
220 62 14 9
155
54 10 6

222 85 37 T = 344

If the sums of squares for the two main effects are computed from the
marginal totals of Table 5.4, we find
(222)2 ^ (35)2 ^ (37)2 (344)2
— = 4608.17
12
(189)2 (^55)2 (344):
SS^ = = 96.33
12
These two add to 4704.50, which is less than the sum of squares between
treatments of 4987.67. It is this difference between the treatment SS and the
SS of the two main effects that shows the interaction between the two main
effects. Here

SS El y-V interaction ^^treatments ^^£7 SSp/

= 4987.67 - 4608.17 - 96.33 - 283.17

Since the treatments carried 5 df, the interaction has 5 — 2 — 1 = 2 df.

These results may now be displayed in the ANOVA of Table 5.5. In this
94 Factorial experiments

Table 5.5 Two-Way ANOVA for Vacuum Tube Pressure Experiment

Source df SS MS F Probability

Between EEs 2 4608.17 2304.08 99.5 0.001

Between F’s 1 96.33 96.33 4.2 0.088
El by V interaction 2 283.17 141.58 6.1 0.036
Error (within treatments) 6 139.00 23.17

Totals 11 5126.67

final table each main effect {El and V) can be tested for significance as well as
can the interaction by comparing each mean square with the error mean
square. Exhaust index is seen to be highly significant. Voltage is not significant
at the 5 percent significance level but the interaction is significant at the
5 percent level.
The mathematical model of Equation (5.1) can now be expanded to read

^ijk T T ^k{ij)
A

^ijk = Ei Vj EVij (5.2)

where Ei represents the exhaust index effect in this problem, Vj the voltage
effect, and EVij the interaction. It should be noted here that whereas in
Chapter 4 randomization restrictions came from the error term, these effects
come from the treatment term. In Equation (5.2) one tests the following
hypotheses:
/foi '■ El = 0 for all i
H02 • ^ for all j
HQ^ : EVij = 0 for all i and j
as both main effects are fixed effects.

5.3 INTERPRETATIONS
Since the interaction is significant in this example, one should be very
cautious in interpreting the main effects. A significant interaction here means
that the effect of exhaust index on vacuum tube pressure at one voltage is
different from its effect at the other voltage. This can be seen graphically by
plotting the six treatment means as shown in Figure 5.8. Note that the lines
in the figure are not parallel. One reasonable procedure for interpreting such
an interaction is to run a Newman-Keuls test on the six means. Had the
main effects not interacted one could treat each set of main effect means
separately. Here testing main effects is not recommended because the results
 5.3 Interpretations 95

ya.
Pressure
(X 1000)

Figure 5.8 Vacuum tube current flow interaction plot.

depend on how these main effects combine. Using the six means, Newman-
Keuls gives

1. Means are 7.5 11 12 30.5 53 58

for treatments: £3^2 £3^1 £^2^2 ^2^1 ^1^2

2. s? = 23.17 with 6 df from Table 5.5.

'23.17
3. Sy..
^ ij-
= 3.40.

4. For jj.
1 p: 2 z. ^3 -r ^

5 percent ranges: 3.46 4.34 4.90 5.31 5.63

5. LSR: 11.76, 14.76, 16.66, 18.05, 19.14.

6. Checking means:
58 - 7.5 = 50.5 > 19.14*
58 - 11 47 > 18.05*
58 - 12 46 > 16.66*
58 -- 30.5 —
27.5 > 14.76*
58 - 53 5 < 11.76
53 - 7.5 = 45.5 > 18.05*
53 - 11 —
42 > 16.66*
53 - 12 41 > 14.76*
53 -- 30.5 22.5 > 11.76*
30.5 - 7.5 = 23 > 16.66*
30.5 - 11 19.5 > 14.76*
30.5 - 12 18.5 > 11.76*
12 - 7.5 4.5 < 14.76
96 Factorial experiments

Hence there are three groups of means:

7.5, 11, 12 30.5 53, 58

If one is looking for the lowest pressure, any one of the three combinations
in the first group will minimize the pressure in the vacuum tube. Thus one
can recommend a 150-second exhaust index at either voltage or a 90-second
exhaust index at 220 volts. From a practical point of view this really gives
two choices: Pump at 127 volts for 150 seconds or at 220 volts for 90 seconds,
whichever is cheaper.
It might be noted that in the model for this problem. Equation (5.2), it
is assumed that the errors are NID (0, Ug). This means that the variances
within each of the six cells that make up the treatments are assumed to have
come from normal populations with equal variances. In the data of Table 5.4
the six ranges are 10, 5, 8, 8, 4, and 3, which average 6.33. From Table 3.15
in Chapter 3, = 20.7 and all six ranges are well within this maximum
range.
It may be of interest to note that whenever there are but two observations
per treatment, the sum of squares for the error may be computed directly.
When n = 2, the sum of squares of the ranges divided by 2 equals the SS^rror-
Here Y, = 278 and Y ^^/2 = 139, which equals the error sum of squares
of Table 5.5.

5.4 COMPUTER PROGRAM

As shown in Chapter 3, a computer program can be used to simplify the
arithmetic when analyzing data in a factorial experiment. For the data in
Example 5.1, Table 5.6 shows the input for an ANOVA on the data of Table
5.4. The entries should be self-explanatory. On line 11 one requests that the
ANOVA be run on Y by EI{1, 3), PV(\, 2), which indicates the effect of
exhaust index (El) at three levels and pump heater voltage (PF) at two levels
on y, which is the pressure in the vacuum tube. The printout of the ANOVA
is shown in Table 5.7, which is essentially the same as Table 5.5. (The T’s
were first multiplied by 1000 here too.) In addition to the ANOVA table,
since all statistics were requested, another printout (Table 5.8) shows the
means of the data—the grand mean of all 12 observations, the means for
each level of exhaust index, the means for each level of pump heater voltage,
and the means for each combination of main effects. Because the interaction
was significant in this problem, these last six means are ready for plotting
as in Figure 5.8.
 5.4 Computer program 97

Table 5.6 Computer Input for Data of Table 5.4

+ + + CREATE PROB3
1. 000 = 77509,113
2. 000 = COMMON (SPSS)
3. 000 = SPSS
4. 000 = #EOR
5. 000 = RUN NAME STAT 502 EX 3
6. 000 = VARIABLE LIST Y, El, PV
7. 000 = INPUT FORMAT (F2.0, IX, FEO, IX, FLO)
8. 000 = N OF CASES 12
9. 000 = VAR LABELS FI, EXHAUST INDEX/
10. 000 = PV, PUMP HEATER VOLTAGE/
11. 000 = ANOVA Y BY EI(1, 3), PV(1, 2)
12. 000 = STATISTICS ALL
13. 000 = READ INPUT DATA
14. 000 = 48 1 1
15. 000 = 58 1 1
16. 000 = 28 2 1
17. 000 - 33 2 1
18. 000 = 7 3 1
19. 000 = 15 3 1
20. 000 = 62 1 2
21. 000 = 54 1 2
22. 000 = 14 2 2
23. 000 = 10 2 2
24. 000 = 6 3 2
25. 000 = 9 3 2
26. 000 = FINISH

Table 5.7 ANOVA Printout for Example 5.1

SUM OF MEAN SIGNIF

SOURCE OF VARIATION SQUARES DF SQUARE F OF F

MAIN EFFECTS 4704.500 3 1568.167 67.691 .001

El 4608.167 2 2304.083 99.457 .001
PV 96,333 1 96.333 4.158 .088
2-WAY INTERACTIONS 283.167 2 141.583 6.112 .036
El PV 283.167 2 141.583 6.112 .036
EXPLAINED 4987.667 5 997.533 43.059 .001
RESIDUAL 139.000 6 23.167
TOTAL 5126.667 11 466.061
98 Factorial experiments

Table 5.8 Printout of Means for Example 5.1

TOTAL POPULATION

28.67
( 12)
LI
1 2 3
55.50 21.25 9.25
( 4) ( 4) ( 4)
PV
1 2
31.50 25.83
( 6) ( 6)
LI PV
1 2
1 53.00 58.00
( 2) ( 2)
2 30.50 12.00
( 2) ( 2)
3 11.00 7.50
( 2) ( 2)

5.5 ANOVA RATIONALE

For a two-factor factorial experiment with n observations per cell, run as a
completely randomized design, a general model would be

Yijk = Ai + Bj ABij -1- s^ij) (5.3)

where A and B represept the two factors, i = 1,2,a levels of factor A,
j = 1,2,..., b levels of factor B, and k = 1,2,n observations per cell.
In terms of population means this becomes

^ijk b'... ~ (A^I. . /^. ..) F (p.j. P...)

+ (Pij. ~ Pi.. ~ P.j. + P...) + {Yijk — i^ij) (5.4)
where fiij represents the true mean of the i, j cell or treatment combination.
Justification for the interaction term in the model comes from subtracting
A and B main effects from the cell effect as follows:

(Pp. - P...) - (Pi.. - P...) - (P.j. - P...) = l^ij. - Pi.. - P.j. + P...
If each mean is now replaced by its sample estimate, the resulting sample
model is
- y... = {Yi.. - F...) + (F,-. - Fj
+ (% - F.. - F^. + FJ + - %)
 5.5 ANOVA rationale 99

If this expression is now squared and summed over i, j, and k, all cross
products vanish, and the results give

I j k i j k

i j k

+
i j k
Y,.- F,. + f,f
+ tiiiYu^-
I j k
Y^if
which again expresses the idea that the total sum of squares can be broken
down into the sum of squares between means of factor A, plus the sum of
squares between means of factor B, plus the sum of squares oi A x B
interaction, plus the error sum of squares (or within cell sum of squares).
Each sum of squares is seen to be independent of the others. Hence, if any
such sum of squares is divided by its associated degrees of freedom, the
results are independently chi-square distributed, and F tests may be run.
The degree-of-freedom breakdown would be

(abn - 1) = {a - 1) + {b - 1) + {a - l){b - 1) + ab{n - 1)

the interaction being cell df = {ab — 1) minus the main effect df, (a — 1)
and {b — 1) or {ab — 1) — {a — 1) — {b — 1) = ab — a — b 1 = {a — 1) x
{b — 1), and within each cell the degrees of freedom are n — \ and there
are ab such cells givjng ab{n — 1) df for error. An ANOVA table can now
be set up expanding'and simplifying the sum of squares expressions using
totals (Table 5.9).
The formulas for sum of squares in Table 5.9 provide good computa-
tional formulas for a two-way ANOVA with replication. The error sum of
squares might be rewritten as
’ n 'T'2

SS error = IZ
I J
I Yl, - ^
k n

which points up the fact that the sum of squares within each of the a x b
cells is being pooled or added for all such cells. This depends on the assump-
tion that the variance within all cells came from populations with equal
variance. The interaction sum of squares can also be rewritten as
T2
nab
100 Factorial experiments

Table 5.9 General ANOVA for Two-Factor Factorial with n Replications per Cell

Source df SS MS
a'T' 2 T' 2.
^ i.. _ ^ Each SS
Factor Ai a - 1
^ nb nab divided by
b'J'2 rp 2 its df
^ -j- ^ ...
Factor Bj b - 1 ^ na nab
d b -y-t 2 ^ 2
^ O'- _ ^ ^■
A X B interaction {a - !)(/) - 1)
ij n 1 nb^
b rp2 7^ 2
y ^ T...
j na nab
abn a b rj-,2
Error ab{n — 1)
i j k i j ^
abn rp2
Totals abn — 1
? ? ? -...

which shows again that interaction is calculated by subtracting the main

effect sum of squares from the cell sum of squares.

Example 5.2 To extend the factorial idea a bit further, consider a problem
with three factors. Such a problem was presented in Chapter 1 on the effect
of tool type, angle of bevel, and type of cut on power consumption for ceramic
tool cutting. Reference to this problem will point out the phases of experi-
ment, design, and analysis as followed in Example 5.1.
It is a 2 X 2 X 2 factorial experiment with four observations per cell
run in a completely randomized manner. The mathematical model is

^ijkm — + Bj + TBIJ -f Cfc + TCif. + + ^m(ijk)

where represents a three-way interaction.

The data for this example are given in Table 1.2 and the ANOVA table
in Table 1.3. That this analysis is a simple extension of the methods used on
Example 5.1 will be shown with the coded data from Table 1.2 (see Table 5.10).
Table 5.10 shows the total for each small cell and the sum of the squares
of the cell observations. These results will be useful in doing the ANOVA.
By this time we should be able to set up the steps in the analysis without
recourse to formulas in dot notation.
First the total sum of squares: Add the squares of all readings (the
circled numbers) and subtract a correction term. The grand total is —13
 5.5 ANOVA rationale 101

Table 5.10 Coded Ceramic Tool Data of Table 1.2, Code: 2(X-28.0)

Tool Type
1 2

Bevel Angle Bevel Angle

Type of Cut 15° 30° 15° 30°

Continuous 2 10 3

5 (4^ 4 0 2 25
-2 9 -6 0
2 15 -5 13

Interrupted 0 —2 —7 —1
-6 2 -6 0
-3 (^54j -1 nm 0 -2 (21) -38
-3 -1 -4 -4
-12 -2 -17 -7

Totals -10 13 -22 +6 -13

squared, divided by the number of observations (32):

SS,otal = 441 - 435.72

32

Tool type sum of squares: Add for each tool type. The totals are +3
and —16; square these and divide by the number of observations per type
(16), add these results for both types, and subtract the correction term. Thus,

3^ + (-16)^ (-13)^
SS tool type 11.28
16

Bevel angle sum of squares: Same procedure on the totals for each bevel
angle, — 32 and 19:

(-32)" + (19)" (-13)^

SS bevel angle 81.28
16 32

Type of cut sum of squares: Same procedure, with cut totals of 25 and — 38

(25)" + (-38)" (-13)^

SS type of cut - 124.03
16 32
102 Factorial experiments

For the T x B interaction, ignore type of cut and use cell totals for the
T X B cells. These are —10, 13, —22, +6:

(-10)2 + (13)2 + (-22)2 13)2

SS T X B interaction 11.28-81.28
8 32

- 0.78

For T X C interaction, ignore bevel angle and the cell totals become 17,
-14, 8, -24:
(17)2+ (-14)2+ (8)2+ (-24)2 (-13)^
SS T X C interaction - 11.28 - 124.03
8 32

= 0.03
For B X C interaction, ignore tool type and the cell totals become —3, 28,
-29, -9:
(- 3)2 + (28)2 (_ 29)2 + (_ 9)2 (_ 13):
SS B X C interaction -81.28- 124.03
8 32
= 3.78

For the three-way interaction T x B x C, consider the totals of the smallest

cells, 2, 15, —5, 13, —12, —2, —17, and —7. From this cell sum of squares
subtract not only the main effect sum of squares but also the three two-way
interaction sums of squares. Thus,

X B X C interaction

(2)2 + (15)2 + (-5)2 + (13)2 + (-12)2 (_2)2 + (_17)2 ^ (_7)2

4

(-13)2
- ^ - 11.28 - 81.28 - 124.03 - 0.78 - 0.03 - 3.78 = 0.79

By subtraction
SS error 213.75

These results are displayed in Table 5.11. If they are compared with those
in Table 1.3, they appear to differ considerably. Actually they give the same
F test results, but in Table 5.11 the data were coded involving multiplication
by 2; the data of Table 1.3 are uncoded. Multiplication by 2 will multiply
the variance or mean square by 4; thus if all mean squares in Table 5.11 are
divided by 4, the results are the same as in Table 1.3. For example, on tool
types 11.28/4 = 2.82, and error 213.75/4 = 53.44. It is worth noting that
any decoding is unnecessary for determining the F ratios. However, if one
wishes confidence limits on the original data or components of variance on
the original data, it may be necessary to decode the results.
 5.6 Remarks 103

Table 5.11 ANOVA for Ceramic Tool Problem

Source df SS MS

Tool type Tj 1 11.28 11.28

Bevel angle Bj 1 81.28 81.28
T X B interaction TBij 1 0.78 0.78
Type of cut 1 124.03 124.03
T X C interaction TCn^ 1 0.03 0.03
B X C interaction BCj,, 1 3.78 3.78
T X B X C interaction TBCij,^ 1 0.79 0.79
Error 24 213.75 8.91

Totals 31 435.72

The interpretation of the results of this example is given in Chapter 1.

The purpose of presenting it again in this chapter is to show that factorial
experiments with three or more factors can easily be analyzed by simple
extension of the methods of this chapter.
This problem, of course, could be fed into the computer to determine
Table 5.11.

5.6 REMARKS

Since the examples in this chapter have contained several replications within
a cell, it would be well to examine a situation involving only one observation
per cell. In this case k = 1, and the model is written as
Yij = fd Ai + Bj + ABij + 8ij

A glance at the last two terms indicates that we cannot distinguish between
the interaction and the error—they are hopelessly confounded. Then the
only reasonable situation for running one observation per cell is one in
which past experience generally assures us that there is no interaction. In
such a case, the model is written as
Yij = jd + Ai Bj Y Sij

It may also be noted that this model looks very much like the model for a
randomized block design for a single-factor experiment (Chapter 4). In
Chapter 4 that model was written as
Yij = + Pi Y Sij

Even though the models do look alike and an analysis would be run in
the same way, this latter is a single-factor experiment—treatments are the
factor—and pi represents a restriction on the randomization. In the factorial
104 Factorial experiments

model, there are two factors of interest, Ai and Bj, and the design is com-
pletely randomized. It is, however, assumed in the randomized-block situa-
tion that there is no interaction between treatments and blocks. This is often
a more reasonable assumption for blocks and treatments since blocks are
often chosen at random. For a two-factor experiment an interaction between
A and B may very well be present, and some external information must be
available in order to assume that no such interaction exists. An experi-
menter who is not sure about interaction must take more than one observa-
tion per cell and test the hypotheses of no interaction.
As the number of factors increases, however, the presence of higher order
interactions is much more unlikely, so it is fairly safe to assume no four-way,
five-way,..., interactions. Even if these were present, they would be difficult
to explain in practical terms.

5.7 SUMMARY

Experiment Design Analysis

I. Single factor
1. Completely randomized 1. One-way ANOVA
Y,j = n + Tj + Cij
2. Randomized block 2.
Tj = + Tj + Pi + &ij
a. Complete a. Two-way ANOVA
b. Incomplete, balanced b. Special ANOVA
c. Incomplete, general c. Regression method
3. Latin square
Tjfc = g + + Tj
+ 7fc + fitjTc
a. Complete a. Three-way ANOVA
b. Incomplete, b. Special ANOVA
Youden square (like 2b)
4. Graeco-Latin square 4. Four-way ANOVA
Tjfcm = M + ^( + L
+ Tic + + ^ijkm
11. Two or more
factors
A. Factorial
(crossed)
1. Completely randomized 1.
Tjfc — ^ + Ai + Bj
+ ABij + •••
for more factors
a. General case a. ANOVA with
interactions
 Problems 105

PROBLEMS
5.1 To determine the effect of two glass types and three phosphor types on the light
output of a television tube, light output is measured by the current required in
series with the tube to produce 30 foot-lamberts of light output. Thus the higher
the current is in microamperes, the poorer the tube is in light output. Three ob-
servations were taken under each of the six treatment conditions and the experi-
ment was completely randomized. The following data were recorded.

Phosphor Type

Glass Type A B C

1 280 300 270

290 310 285
285 295 290
2 230 260 220
235 240 225
240 235 230

Do an analysis of variance on these data and test the effect of glass type, phosphor
types, and interaction on the current flow.
5.2 Plot the results of Problem 5.1 to show that your conclusions are reasonable.
5.3 For any significant effects in Problem 5.1 test further between the levels of the
significant factors.
5.4 Based on the results in Problems 5.1-5.3, what glass type and phosphor type
(or types) would you recommend if a low current is most desirable?
5.5 Adhesive force on gummed material was determined under three fixed humidity
and three fixed temperature conditions. Four readings were made under each set
of conditions. The experiment was completely randomized and the results set out
in an ANOVA table as follows:

Source df SS MS

Humidity 9.07
Temperature 8.66
H X T interaction 6.07
Error

Total 52.30

Complete this table.

5.6 For the data in Problem 5.5 test all indicated hypotheses and state your conclusions.
5.7 Set up a mathematical model for the experiment in Problem 5.5 and indicate the
hypotheses to be tested in terms of your model.
5.8 The object of an experiment is to determine thrust forces in drilling at different
speeds and feeds, and in different materials. Five speeds and three feeds are used.
 106 Factorial experiments

and two materials with two samples tested under each set of conditions. The order
of the experiment is completely randomized and the levels of all factors are fixed.
The following data are recorded on thrust forces after subtracting 200 from all
readings.

Speed

Material Feed 100 220 475 715 870

^10 0.004 122 108 108 66 80

no 85 60 50 60
0.008 332 276 248 248 276
330 310 295 275 310
0.014 640 612 543 612 696
500 500 450 610 610

Vio 0.004 192 136 122 108 136

170 130 85 75 75
0.008 386 333 318 472 499
365 330 330 350 390
0.014 810 779 810 893 1820
725 670 750 890 890

Do a complete analysis of this experiment and state your conclusions.

5.9 Plot any results in Problem 5.8 that are significant.
5.10 Set up tests on means where suitable and draw conclusions from Problem 5.8.
v/ 5.11 In an experiment for testing rubber materials interest centered on the effect of the
mix {A, B, or C), the laboratory involved (1, 2, 3, or 4), and the temperature (145,
155, 165°C) on the time in minutes to a 2-iri.-lb rise above the minimum time.
Assuming a completely randomized design, do an ANOVA on the following data.

Temperature (°C)
145 155 165

Mix Mix Mix

Laboratory A B c A B C A B C
1 1 11.2 11.2 11.5 6.7 6.8 7.0 4.8 4.8 5.0
11.1 11.5 11.4 6.8 6.7 7.0 4.8 4.9 4.9
7 'Ll . \
2 11.8 12.3 ‘ 12.3 7.3 7.5 7.5 5.3 5.4 5.3
11.8 12.3 11.9 7.2 7.7 7.3 5.3 5.2 5.3
" ?
3 11.5 12.3 12.7 6.6 7.1 7.8 5.0 5.3 5.2
p

11.6 12.0 12.5 6.9 7.2 7.3 5.0 5.0

\

11.5 11.8 12.7 7.2 6.7 7.1 4.5 4.5

.
A
f
1 11.3 11.7 12.7 6.9 7.0 7.0 4.6 4.5 4.5
i—1 <
 Problems 107

5.12 What is there in the results of Problem 5.11 that would lead you to question
some assumption about the experiment?
5.13 Tomato plants were grown in a greenhouse under treatments consisting of com-
binations of soil type (factor A) and fertilizer type (factor B). A completely ran-
domized two-factor design was used with two replications per cell. The following
data on the yield Y (in kilograms) of tomatoes were obtained for the 30 plants
under study.

Fertilizer Type (B)

Soil
Type{A) 1 2 3 T,

I 5,7 5,5 3,5 30

II 5,9 1,3 2,2 22
III 6,8 4,8 2,4 32
IV 7,11 7,9 3,7 44
V 6,9 4,6 3,5 33
Tj 73 52 36 161

Note: y y y Yfj, = 1043.

k j i

Complete the following ANOVA table.

Source df Sum of Squares Mean Square

Soil type {A)

Fertilizer type (B)
Interaction (AB)
Error

5.14 Carry out tests on main effects and interactions in an appropriate order using
levels of significance for each test that seem appropriate to you. For each test
state HQ and show the rejection region, and give your conclusion. After
doing all tests, summarize your results in words.
5.15 On the basis of your results in Problem 5.14, state how you would carry out
Newman-Keuls comparisons to find out the best combination (or combinations)
of soil type and fertilizer type to achieve maximum mean yield of tomatoes. Then
carry out these comparisons (at an a of 0.05) and state your conclusions.
5.16 Each of the following tables represents the cell means (7^^ ) for a two-factor com-
pletely randomized balanced experiment. For each table tell which (if any) of the
following SS’s would be zero for that table: SS^, SS^, SS^g. (There may be none;
there may be more than one.)

a. A

1 3

5 3
108 Factorial experiments

b. A

= 0

c.

= 0

d.

= 0

5.17 In each two-factor design table below the numbers in the cells of the table are the
population means of the observations for those cells. For each table, indicate
whether or not there is interaction between the factors and justify your assertion.

a. Factor B
1 2 3 4

Factor A 1 7 6 5 2

2 9 8 7 4

Is there interaction? YesjNo (circle one) Explain:

b. Factor B
1 2 3

Factor A 1 7 6 5

2 5 6 7

Is there interaction? Yes/No (circle one) Explain:

c. Factor B

Is there interaction? Yes/No (circle one) Explain:

 Problems 109

5.18 An industrial engineer presented two types of stimuli (two-dimensional and three-
dimensional films) of two diflferent jobs (1 and 2) to each of hve analysts. Each
analyst was presented each job-stimulus film twice and the order of the whole
experiment was considered completely randomized. The engineer was interested
in the consistency of analyst ratings of four sequences within each job stimulus
presentation. The variable recorded is the log variance of the four sequences
since log variance is more likely to be normally distributed than the variance.
Data showed:

Stimulus
T wo-Dimensional Three-Dimensional

Analyst Job 1 Job 2 Job 1 Job 2

1 1.42 1.40 1.00 0.92

1.25 1.44 0.90 0.93
2 1.59 1.83 1.46 1.43
2.09 2.02 1.68 1.02
3 1.26 1.80 0.85 1.33
1.48 1.75 1.43 1.48
4 1.76 0.98 1.21 0.73
1.47 1.23 1.25 1.22
5 1.43 1.17 1.60 1.31
1.72 1.18 2.03 1.12

Do a complete analysis and summarize your results.

5.19 Make further tests on Problem 5.18 as suggested by the ANOVA results.
5.20 The following results were reported on a study of “factors that affect the salary of
high-school teachers of commercial subjects”;

Source df SS

Sex 1 239,763
Size of school 2 423,056
Years in position 5 564,689
Sex by size interaction 2 18,459
Sex X years interaction 5 85,901
Size X years interaction 10 240,115
Sex X size x years interaction 10 151,394
Within classes 153 1,501,642

Indicate the mathematical model for the above study. Show a possible data layout.
Complete the ANOVA table and comment on any significant results.
5.21 Four factors are studied for their effect on the luster of plastic film. These factors
are (1) film thickness (1 or 2 mils); (2) drying conditions (regular or special); (3)
110 Factorial experiments

length of wash (20, 30, 40, or 60 minutes); and (4) temperature of wash (92°C or
100°C). Two observations of film luster are taken under each set of conditions.
Assuming complete randomization, analyze the following data.

Regular Dry Special Dry

Minutes 92°C 100°C 92°C 100°C

1-mil Thickness
20 3.4 3.4 19.6 14.5 2.1 3.8 17.2 13.4
30 4.1 4.1 17.5 17.0 4.0 4.6 13.5 14.3
40 4.9 4.2 17.6 15.2 5.1 3.3 16.0 17.8
60 5.0 4.9 20.9 17.1 8.1 4.3 17.5 13.9
2-mil Thickness
20 5.5 3.7 26.6 29.5 4.5 4.5 25.6 22.5
30 5.7 6.1 31.6 30.2 5.9 5.9 29.2 29.8
40 5.6 5.6 30.5 30.2 5.5 5.8 32.6 27.4
60 7.2 6.0 31.4 29.6 8.0 9.9 33.5 29.5

5.22 Plot and discuss any significant interactions found in Problem 5.21. What condi-
tions would you recommend for maximum luster and why?
5.23 To study the effect of four types of wax applied to floors for three different lengths
of polishing times it was decided to try each combination twice and completely to
randomize the order in which ffoors received which treatments. The criterion was
a gloss index.
a. Outline an ANOVA table for this problem with proper degrees of freedom.
b. Explain, in words, what a significant interaction would mean in this situation.
5.24 A manufacturer of combustion engines is concerned about the percentage of smoke
emitted by the engine in order to meet EPA standards. An experiment was con-
ducted involving engines with three timing levels (T), three throat diameters (D),
two volume ratios in the combustion chamber (K), and two injection systems (/).
Thirty-six engines were designed to represent all possible combinations of these
four factors and the engines were then operated in a completely random order
and percent smoke recorded to the nearest 0.1 percent. Results showed:

1 2 3
D D D

1 2 3 1 2 3 1 2 3

1 0.4 3.8 7.0 1.8 4.8 8.1 1.7 4.9 8.2

1 /
2 0.5 1.6 4.0 0.7 3.0 5.6 1.2 3.2 10.0
V
1 0.7 1.8 6.0 1.4 1.0 1.5 7.0 3.1 3.4
2 /
2 0.1 0.1 0.2 0.6 0.4 1.4 4.0 1.8 2.0
 Problems 111

Do a complete analysis of these data and make recommendations as to which

combination or combinations of factor levels should give a minimum percent of
smoke emitted.

5.25 Students’ scores on a psychomotor test were recorded for various sized targets at
which the student must aim, different machines used, and the level of background
illumination. Results are shown in the following ANOVA table.

Source df SS MS

Target size {A) 3 235.200 78.400

Machine (B) 2 86.467 43.233
Level of illumination (C) 1 76.800 76.800
A X B 6 104.200 17.367
A X C 3 93.867 31.289
B X C 2 12.600 6.300
A X B X C 6 174.333 29.056
Within sets 96 1198.000 12.478

Totals 119 1981.467

a. Using a 5 percent significance level indicate which effects are statistically

significant.
b. Assuming that you had the original data at your disposal, explain briefiy the
next steps you would take in analyzing the data from this experiment.
c. Devise a data layout sheet that could be used to collect each item of data for
this experiment.
d. Assuming the mean score for target size A3, machine B2, and level of illumina-
tion Cj is 9.6, set 90 percent confidence limits on the true mean for this treatment
combination.
 2^ Factorial Experiments
e

6.1 INTRODUCTION
In the last chapter factorial experiments were considered and a general
method for their analysis was given. There are a few special cases that are
of considerable interest in future designs. One of these is the case of / factors
where each factor is at just two levels. These levels might be two extremes of
temperature, two extremes of pressure, two time values, two machines, and
so on. Although this may seem like a rather trivial case since only two
levels are involved, it is, nevertheless, very useful for at least two reasons:
to introduce notation and concepts useful when more involved designs are
discussed and to illustrate what main effects and interactions there really
are in this simple case. It is also true that in practice many experiments are
run at just two levels of each factor. The ceramic tool-cutting example of
Chapters 1 and 5 is a 2 x 2 x 2, or 2^ factorial, with four observations per
cell. Throughout this chapter the two levels will be considered as fixed levels.
This is quite reasonable because the two levels are chosen at points near the
extremes, rather than at random.

6.2 2^ FACTORIAL
The simplest case to consider is one in which two factors are of interest and
each factor is set at just two levels. This is a 2 x 2 = 2^ factorial, and the
design will be considered as completely randomized. The example in Sec-
tion 5.1 is of this type. The factors are temperature and altitude, and each
is set at two levels: temperature at 25°C and 55°C and altitude at 0 K and
3 K. This gives four treatment combinations, displayed in Figure 5.3 where
the response variable is the current flow in milliamperes. To generalize a bit,
consider temperature as factor A and altitude as factor B. The model for
this completely randomized design would be

Yij = jU + AI B j -\- ABij + Sjj (6-1)

112
 6.2 2^ Factorial 113

where i = 1, 2 and j = 1, 2 in this case. Unless there is some replication, of

course, no assessment of interaction can be made independent of error.
From the data of Figure 5.3 when both temperature A and altitude B are
at their low levels, the response is 210 mA. This may be designated by
subscripts on AB as follows: AQBQ = 210 mA. Following this notation, A^BQ
means A at its high level and B at its low level, or temperature 55°C and
altitude 0 K. The response is A^BQ = 240 mA. Likewise, AQB^ = 180 mA
and A^B^ = 200 mA. Since a 2^ experiment is encountered so often in the
literature, most authors have adopted another notation for these treatment
combinations. For this new notation, just the subscripts on AB are used as
exponents on the small letters ab. If both factors are at their low levels,
= (1), and (1) represents the response of both factors at their low level.
a^b^ = b represents B at its high level and A at its low level, a^b^ = a
represents the high level of A and low level of B, and a^b^ = ab represents
the response when both factors are at their high levels. This notation can
easily be extended to more factors, provided only two levels of each factor
are involved.
For the example of Figure 5.3 the treatment combinations can be
represented by the vertices of a square as in Figure 6.1.
In Figure 6.1 the low and high levels of factors A and B are represented
by 0 and 1, respectively, on the A and B axes. The intersection of these
levels in the plane of the figure shows the four treatment combinations.
For example: 00 = (1) represents both factors at their low levels, 10 = a
represents A high, B low, and so on, 01 = fi, 11 = ab. These expressions
are only symbolic and are to be considered as merely a mnemonic device
to simplify the design and its analysis. The normal order for writing these
treatment combinations is (1), a, b, ab. Note that (1) is written first, then the

/? = 180 ab = 200

Factor B
(altitude)
114 2^ Factorial experiments

high level of each factor with the low level of the other {a, b), and the fourth
term is the algebraic product of the second and third (ab). When a third
factor is introduced, it is placed at the end of this sequence and then mul-
tiplied by all of its predecessors. For example, if factor C is also present at
two levels, the treatment combinations are

(1), fl, b, ab, c, ac, be, abc

which can be represented as the vertices of a cube.

Returning to Figure 6.1 we find that the ejfect of a factor is defined as
the change of response produced by a change in the level of that factor. At
the low level of B, the effect of A is then 240 — 210 or u — (1), whereas the
effect of A at the high level of B is 200 — 180 or nh — b. The average effect
of A is then
A = — (1) + uF — 6]
or
A = i[ —(1) + a — b + <26]

Note that the coefficients on this A effect are all -I-1 when A is at its high
level in the treatment combinations ( -\-a, -\-ab) and the coefficients are all
— 1 when A is at its low level as in (1) and b. Note also that

2A = —(l)-\-a — b + ab

is a contrast as defined in Section 3.4 (the sum of its coefficients are

—1 + 1 — 1-1-1= 0). This concept will be useful, as the sum of squares
due to this contrast or effect can easily be determined.
The average effect of B, based on low level ofT[180 — 210 = b — (1)]
and high level of A (200 — 240 = ab — a), is

B = j\^b — (1) + ab — a^
or
B — — — a + b + uh]

and again it is seen that the same four responses are used, but + 1 coefficients
are on the treatment combinations for the high level of B and — 1 coefficients
for the low level of B. Also, 2B is a contrast.
To determine the effect of the interaction between A and B, note that at
the high level of B the A effect is ab — b, and at the low level of B the A
effect is a — (1). If these two effects differ, there is an interaction between
+ and B. Thus the interaction is the average difference between these two
differences
AB = i{(uh - b) - [a - (1)]}
= j\^ab — b — a -\- (1)]
or
= i[(l) “ ^
 6.2 2^ Factorial 115

Here again the same four treatment combinations are used with a
different combination of coefficients. Note that 2AB is also a contrast, and
all contrasts 2A, 2B, 2AB are orthogonal to each other. Summarizing in
normal order gives
2A = —(1) -\- a — b + ab
2B = —{1) — a b ab
2AB = — a — b + ab

Note that the interaction effect takes the responses on one diagonal of the
square with +1 coefficients and the responses on the other diagonal with
— 1 coefficients. Note also that the coefficients for the interaction effect can
be found by multiplying the corresponding coefficients of the two main
effects. As the only coefficients used are + I’s and — I’s, the proper coefficients
on the treatment combinations for each main effect and interaction can be
determined from Table 6.1.

Table 6.1 Coefficients for Effects in a 2^

Factorial Experiment

Treatment Effect
Combination A B AB

(1) — — +
a + — —

b — + —

ab + + +

From Table 6.1 the orthogonality of the effects is easily seen, as well as
the generation of the interaction coefficients from the main effect coefficients.
Another approach to the interaction between A and B would be to
consider that at the high level of A the effect of B is ab — u, and at the low
level of A the effect of 5 is 6 — (1), so that the average difference is

AB = j{{ab — u) — [6 — (1)]}
= ^\_ab — a — b + (1)]
^ i[ + (l) ~ ^ ~ ^

which is the same expression as given before.

For the response data of Figure 6.1 the effects are
A = i(-210 + 240 - 180 + 200) = 25 mA
B = i(-210 - 240 + 180 + 200) = -35 mA
AB = i( + 210 - 240 - 180 + 200) - -5 mA
116 2^ Factorial experiments

Since 2A, 2B, and 2AB are contrasts, the sum of squares due to a contrast is

(contrast)^
SS

where n is the number of observations in each total (here n = 1). Also

^ 1 + 1 + 1 + 1 = 4 (or 2^). From this definition

SS, . . as
4

r2(-35)P
SSfi = = 1225
4

SS,. . . 25

= 1875

Since each effect and the interaction has but 1 df and there is no measure of
error as only one observation was taken in each cell, this ANOVA is quite
trivial, but it does show another approach to analysis based on effects. The
simple ANOVA table would be that shown by Table 6.2.

Table 6.2 ANOVA for a 2^ Factorial

with No Replication

Source df SS MS

At 1 625 625
fi; 1 1225 1225
Error or AB^j 1 25 25

Totals 3 1875

If the general methods of Chapter 5 are used on the data in Figure 6.1
(coded by subtracting 200), the results are those in Table 6.3.

(30)^
SS.otal = (10)" + (-20)2 + (40)2 _ = 1875

(-10)2 + (40)2 (30)'

SS 850 - 225 - 625
2 4

(50)2 (_20)2 (30)2

ss^ = = 1450 - 225 = 1225

(by subtraction) = 1875 — 625 — 1225 = 25

 6.2 2^ Factorial 117

Table 6.3 Coded Data of Figure 6.1

Factor A

Factor B 0 1 Totals

0 + 10 + 40 + 50
1 -20 0 -20

Totals -10 + 40 + 30

which are the same sums of squares as given in Table 6.2. Even though no
separate measure of error is available, a glance at the mean squares of
Table 6.2 shows that both main effects are large compared with the inter-
action effect.
The results of this section can easily be extended to cases where there
are n replications in the cells, by using cell totals for the responses at (1), a,
b, and ab and adjusting the sum of squares for the effects accordingly (see
Example 6.1).

Example 6.1 An example of a 2^-factorial experiment with two replications

per cell is considered here, using hypothetical responses to illustrate the
principles of the previous section. Consider the data of Table 6.4.
Using the general methods of Chapter 5, the sums of squares are

SStotal ~ 4^ + 6^ + 3^ + 7^ + 2^ + (-If + (-Af + (-6f

(10)"
= 170 - 12.5 = 157.5
8

(20)" + (-10)^
SS 12.5 = 112.5

(10)" + (0)'
SS^ = 12.5 = 12.5

(lOf + (lOf + (0)^ + (-10)'

SS Ax B interaction 12.5 - 112.5 - 12.5

= 12.5

SSerror = 157.5 - 112.5 - 12.5 - 12.5 - 20.0

These results could be displayed in an ANOVA table. Using the methods

of Section 6.2, we get, with the treatment combinations, the total response

(1) = 10 a = 0 b = 10 ab = —10
118 2^ Factorial experiments

Table 6.4 2^ Factorial with Two Replications

Factor A

Factor B 0 1 Totals

0 4 2
6 -2
10 0 + 10

1 3 -4
7 -6
10 -10 0

Totals + 20 -10 + 10

and the contrasts

4A ^ -10 + 0 - 10 + (-10) - -30

4B = -10 - 0 + 10 + (-10) = -10
4AB = +10 - 0 - 10 + (-10) = -10

The coefficient 4 used with each response represents the two individual
responses at each level. In general the coefficient of these effects is n • 2^“^
where n is the number of replications and / the number of factors. Here
/ = 2 and n = 2. The sum of squares for these three contrasts are

{4Af _ (-30)2 _ 900

SS^ 112.5
n • 22 “ 2-4 ” "

(45)2 ^ (-10)2 _ too

SS^ 12.5
n • 22 ~ 2-4 “ "

(4^5)2 _ (-10)2 _ 100

= 12.5
n • 22 “ 2-4 “ ^
since
4

1 + 1 + 1 + 1 = 4
i= 1

These results are the same as given by the general method. The total
sum of squares must be calculated as usual in order to get the error sum of
squares by subtraction.
For this special case of a 2^ factorial experiment, Yates [26] developed
a rather simple scheme for computing these contrasts. The method can best
be illustrated on Example 6.1, using Table 6.5.
 6.2 1? Factorial 119

Table 6.5 Yates’ Method on a 1? Factorial

Treatment
Combination Response (1) (2) SS

(1) 10 10 10 = total 12.5

a 0 0 - 30 = 4+ 112.5
b 10 -10 -10 - 45 12.5
ab -10 -20 -10 = 4AB 12.5

In Table 6.5 list all treatment combinations in the first column. Place
the total response to each of these treatment combinations in the second
column. For the third column, labeled (1), add the responses in pairs, for
example, 10 + 0 = 10 for the first two and 10 — 10 = 0 for the next two;
this completes half of column (1). For the second half, subtract the responses
in pairs, always subtracting the first from the second, for example, 0 — 10 =
— 10; —10 — (10) = —20. This completes column (1). Column (2) is deter-
mined in the same manner as column (1) using the column (1) results:
10 + 0-10; -10 - 20 - -30; 0 - 10 - -10; -20 - (-10) - -10.
Proceed in this same manner until the/th column is reached: (1), (2), (3),...,
(/). In this case / — 2, so there are just two columns. The values in column
/ are the contrasts, where the first entry is the grand total of all readings; the
one corresponding to u is n • 2^~ ^ ‘ A,b is n ’ 2^~ ^ • B, and ab is n • 2^~ ^ • AB.
When the results of the last column [column (/)] are squared and
divided by n • 2^, the results are the sums of squares as shown above. The
first sum of squares (total)^/8 is the correction term for the grand mean
given in Chapter 5. The Yates method reduces the analysis to simply adding
and subtracting numbers. It is very useful provided the experiment is a 2^
factorial with n observations per treatment.
As proof for the Yates method on a 2^ factorial go through the steps
above using the treatment combination symbols for the responses as in
Table 6.6. It is obvious that the last column does give the proper treatment
contrasts.

Table 6.6 Yates’ Method on a 2^ in General; One

Observation per Cell

Treatment
Combination (1) (2)

(1) (l) + t? (l) + fl + /? + cib = total

a b + ab —{l) + a — b + ab = 2A
b a — (1) -(1) — a + b + ab = 2B
ab ab — b (i) — a — b + ab = 2AB
120 2^ Factorial experiments

6.3 V FACTORIAL
Considering a third factor C, also at two levels, the experiment will be a
2 X 2 X 2 or 2^ factorial, again run in a completely randomized manner.
The treatment combinations are now (1), u, b, ab, c, ac, be, abc, and they may
be represented as vertices of a cube as in Figure 6.2.

Factor B

Factor A

Figure 6.2 2^ factorial arrangements.

With these 8 = 2^ observations, or 8n if there are replications, the main

effects and each interaction may be expressed by using the proper coefficients
( — 1 or + 1) on these eight responses. For the effect of factor A, consider all
responses in the right-hand plane {a, ab, ac, abc) with plus signs and all in
the left-hand plane [(1), b, be, c] with minus signs as these will show the
effect of increasing the level of A. Or,

4A = ~{1) a — b ab — c ac — be abc

For factor B consider responses in the lower and higher planes of the cube

4B — —{!) — a b A ab — c — ac + be + abc

The AB interaction is determined by the difference in the A effect from level

0 of 5 to level 1 of B, regardless of C

at BQI 2A effect a A ac — c — [I]

at B^ : 2A effect abc A ab — b — be
 6.3 2^ Factorial 121

The difference in these is the interaction

4AB = [flhc + ab — b — bc\ — \^a + ac — c — (1)]

4AB = abc + ab — b — be — a — ac + c + {1)
or
4AB = +{\) — a — b + ab A c — ac — be + abc

which gives the same signs as the products of corresponding signs on 4A

and 4B.
For factor C compare the responses in the back plane of the cube with
those in the front plane of the cube

4C = c A be A ac A abc — {1) — b — a — ab
or
4C = —{1) — a — b — ab A c A ac A be A abc
The AC and BC interactions can be determined as AB was, and it can be
seen that the resulting interaction effects are

4.4C = A{1) — a A b — ab — c A ac — be A abc

4BC = A{1) A a — b — ab — c — ac A be A abc

To determine the ABC interaction, consider the BC interaction at level 0 of

A versus the BC interaction at level 1 of A. Any difference in these is an ABC
interaction

2BC interaction at A^: A{1) — b — c A be

2BC interaction at A^: Aa — ab — ac A abc

Their difference is

4ABC = (a — ab — ac A abc) — [(1) — b — c A bc~\

or
4ABC = —{!) A a A b — ab A c — ac — be A abc

which can also be obtained from multiplying the coefficients of A and BC,
or B and AC, or C and AB. These others can also be used to get the ABC
interaction, but the results are the same. Summarizing gives us Table 6.7.
Table 6.7 illustrates once again the orthogonality of the effects and can
easily be extended to four, five, and more factors if each factor is at two
levels only. In a 2^-factorial experiment, the sum of squares is given by
(contrast)^ (contrast)^
^^contrast ^3 o
n • 2 8n

and again the Yates method leads to an easy computation of the contrasts.
122 2^ Factorial experiments

Table 6.7 Coefficients for Elfects in a 2^ Factorial Experiment

Effect
Treatment
Combination Total A B AB C AC BC ABC

(1) + — — + — + + —

a + + — — — — + +
b + — + — —
+ — +
ab + + + + — — — —

c + — —
+ + — — +
ac + + — — + + — —

be + — + — + — + —

abc + + + + + + + +

Example 6.2 In Chapters 1 and 5 the problem on power requirements for

cutting with ceramic tools was analyzed in detail. It is readily seen that this
isa2x2x2 = 2^ factorial with four replications per cell. This problem
could be analyzed by the special methods of this chapter. From Table 5.10
on coded data the treatment combinations might be summarized as in
Table 6.8.

Table 6.8 Ceramic Tool Data in 2^ Form

Tool Type
1 >

Bevel Angle Bevel Angle

Type of Cut 15° 30° 15° 30°

Continuous (1) b a ab
2 15 -5 13

Interrupted c be ae abe
-12 -2 -17 -7

Tool type = factor A, bevel = factor B, cut = factor C.

In Table 6.8 the totals of the four observations for each treatment
combination have been entered in their corresponding cells. By the Yates
method we get the entries in Table 6.9.
The sum of squares is seen to be in substantial agreement with those
of Table 5.10. We must resort to the individual readings, however, to deter-
mine the total sum of squares and then the error sum of squares. This was
done in Table 5.10 and the interpretation of this problem is given in Chapter 5.
The purpose of repeating the problem here was merely to show the use of
the Yates method on a 2^-factorial experiment.
 6.4 2^—Remarks 123

Table 6.9 Yates’ Method on Ceramic Tool Data

Treatment
Combination Response (1) (2) (3) SS

(1) 2 -3 25 — 13 = total 5.28

a -5 28 -38 -19 = \6A 11.28
b 15 -29 -9 51 = 16.6 81.28
ab 13 -9 -10 5 = \6AB 0.78
c -12 -7 31 -63 = 16C 124.03
ac -17 -2 20 -1 = 16TC 0.03
be -2 -5 5 -11 = \6BC 3.78
abc -7 -5 0 -5 = \6ABC 0.78

6.4 2^—REMARKS

The methods shown for 2^ and 2^ factorials may easily be extended to 2^

factorials where / factors are each considered at two levels. The contrasts
are determined from the responses to the treatment combinations by asso-
ciating plus signs with high levels of a factor and minus signs with low levels;
the contrasts for interactions are found by multiplication of corresponding
coefficients.

Table 6.10 ANOVA for a 2^ Factorial with n Replications

Source df

Main effects A 1
B 1
if
C 1
• * ^

Two-factor interactions AB 1
AC 1 /(/-I)
C(/, 2) = i
BC 1
♦ • >

Three-factor interactions ABC 1

ABD 1 ^ 3) = /(/ - !)(/ - 2)
C(/,
BCD 1 6

Four-factor interactions, and so on

Sum of all treatment combinations 2^ - 1

Residual or error 2^ ■ {n — 1)

Total n-2^ - i
124 2^ Factorial experiments

The general relationships for 2^ factorials with n observations per

treatment are
contrast = n • 2^'^ (effect)
or

effect = — (contrast)
nz-'
and
(contrast)^
SS contrast
n2-^

A general ANOVA would be as in Table 6.10, but not in “normal” order.

6.5 SUMMARY

Experiment Design Analysis

I. Single factor
1. Completely randomized 1. One-way ANOVA
Yij = + Tj + Sij

2. Randomized block 2.
Tj = /^ + +C+ ^ij
a. Complete a. Two-way ANOVA
b. Incomplete, balanced b. Special ANOVA
c. Incomplete, general c. Regression method
3. Latin square
Yijk = /^ + A + C
+ Tfc + ^ijk
a. Complete a. Three-way ANOVA
b. Incomplete, b. Special ANOVA
Youden square (like 2b)
4. Graeco-Latin square 4. Four-way ANOVA
Yijkm = M + C
+ + ^ijkm
II. Two or more
factors
A. Factorial
(crossed)
1. Completely randomized
^ijk = + Ai Bj
+ ABij + e;t(U) ■ ■ ■
for more factors
a. General case a. ANOVA with
interactions
b. 2^ case b. Yates’ method or
general ANOVA;
use (1), a, b, ah, . . .
 Problems 125

PROBLEMS

6.1 For the 2^ factorial with three observations per cell given below (hypothetical
data), do an analysis by the method of Chapter 5.

Factor A

Factor B ^1 ^2

0 4
2 6
1 2
B2 -1 -1
-3 -3
1 -7

6.2 Redo Problem 6.1 by the Yates method and compare results.
6.3 In an experiment on chemical yield three factors were studied, each at two levels.
The experiment was completely randomized and the factors were known only
as A, B, and C. The results were

A1 A2

B, B2 B, B'l

c, C2 Cl C2 Cl C2 c, C2
1595 1745 1835 1838 1573 2184 1700 1717
1578 1689 1823 1614 1592 1538 1815 1806
5! ''3 3434 365^ 3463 3^^ 2 3 '■ 6 '9-.; ,
Analyze by the methods of Chapter 5. •n' . .
6.4 Analyze Problem 6.3 by the Yates method and compare your results.
6.5 Plot any results from Problem 6.3 that might be meaningful from a management
point of view.
6.6 The results of Problem 6.3 lead to another experiment with four factors each at
two levels, with the following data.

^1

Bi B2

Cl C2 Cl C2

^1 D2 D, Di D, D2 D2
1985 2156 1694 2184 1765 1923 1806 1957
1592 2032 1712 1921 1700 2007 1758 1717

'i -( ■ - -
126 2^ Factorial experiments

Ai

B2

Cl C2 Cl C2

D, D2 D, D2 D, D2 D2

1595 1578 2243 1745 1835 1863 1614 1917

2067 1733 1745 1818 1823 1910 1838 1922

. € 3- 2%3^
Analyze these data by the general methods of Chapter 5.
6.7 Analyze Problem 6.6 using the Yates method.
6.8 Plot from Problem 6.6 any results you think are meaningful.
6.9 Consider a four-factor experiment with each factor at two levels.
a. Write out the treatment combinations in this experiment in normal order.
b. Assuming there is only one observation per treatment, outline an ANOVA for
this example and argue what you would use for an error term to test for signifi-
cance of various factors and interactions.
c. Explain how you would determine the AD interaction from the responses to
the treatment combinations given in part (a).
6.10 An experimenter is interested in the effects of five factors:—A, B, C, D, and E—
each at two levels on some response variable Y. Assuming that data are available
for each of the possible treatment combinations in this experiment, answer the
following.
a. How would you determine the effect of factor C and its sum of squares based
on data taken at each of the treatment combinations?
b. Set up an ANOVA table for this problem assuming n observations for each
treatment combination.
c. If n = 2 for part (b), at what numerical value of F would you reject the hypoth-
esis of no ACE interaction.
d. How would you propose testing for ACE interaction and what F would be
necessary for rejection in this problem if n = 1 ?
6.11 Consider an experiment with five factors A, B, C, D, and £, each at two levels.
a. How many treatment combinations are there?
b. Outline an ANOVA table to test all main effects, all two-way and all three-way
interactions, if there is only one observation per treatment combination.
c. Explain how you would find the proper signs on the treatment combinations
for the ACE interaction.
d. Some experimenter comments to you that “since all factors are at only two
levels, it is unnecessary to run a Newman-Keuls test as one only has to look
and see which level gives the greater mean response.” Explain situations in
which this comment would be true for your results and in which situations it
would not be true.
6.12 Three factors are studied for their effect on the horsepower necessary to remove a
cubic inch of metal per minute. The factors are feed rate at two levels, tool condition
 Problems 127

at two levels, and tool type at two levels. For each treatment combination three
observations are taken and all 24 observations of horsepower are made in a
completely randomized order. The data appear below.

Feed Rate

0.011 0.015

Tool Tool Condition Tool Condition

Type Same New Same New

Utility 0.576 0.548 0.514 0.498

0.576 0.555 0.515 0.519
0.565 0.540 0.518 0.504
Precision 0.526 0.547 0.494 0.521
0.542 0.524 0.504 0.480
0.548 0.525 0.530 0.494

Do a complete ANOVA on these data as a 2^ experiment and state your con-

clusions.

6.13 If minimum horsepower is desired for the experiment given in Problem 6.12,
what are your recommendations as to the best set of factor conditions on the basis
of your results in Problem 6.12.

6.14 Any factor whose levels are powers of 2 such as 4, 8,16,..., may be considered as
pseudo factors in the form 2^, 2^, 2"^,.... Yates’ method can then be used if all
factor levels are 2 or powers of 2 and the resulting sums of squares for the pseudo
factors can be added to give the sums of squares for the actual factors and appro-
priate interactions. Try out this scheme on the data of Problem 5.21.

6.15 A systematic test was made to determine the effects on the coil breakdown voltage
of the following six variables, each at two levels as indicated.
1. Firing furnace: Number 1 or 3
2. Firing temperature: 1650°C or 1700°C
3. Gas humidification: Yes or no
4. Coil outside diameter: Large or small
5. Artificial chipping: Yes or no
6. Sleeve: Number 1 or 2
Assuming complete randomization, devise a data sheet for this experiment, outline
its ANOVA, and explain your error term if only one coil is to be used for each of
the treatment combinations.

6.16 If seven factors are to be studied at two levels each, outline the ANOVA for this
experiment in a table similar to Table 6.10, and, assuming n = 1, suggest a reason-
able error term.
128 2^ Factorial experiments

6.17 In a steel mill an experiment was designed to study the effect of drop-out tempera-
ture, back-zone temperature, and type of atmosphere on heat slivers in samples of
steel. The response variable is a mean surface factor obtained for each type of
sliver. Results are

Drop-out Temperature

2145° 2165°

Back-Zone Temperature

Atmosphere 2040-2140° 2060-2160° 2040-2140° 2060-2160°

Oxidizing 7.8 4.3 4.5 8.2

Reducing 6.8 5.9 11.0 6.0

Assuming complete randomization, analyze these results. To get an error term

for testing, three of the data points were repeated to give a SS^rror of 1-11 based
on 3 degrees of freedom.
6.18 Further analyze the results of Problem 6.17 and make any recommendations you
can for minimizing heat slivers.
6.19 At the same time that data were collected on heat slivers in Problem 6.17, data
were also collected on grinder slivers. The results were found to be as follows;

Drop-out Temperature

2145° 2165°

Back-Zone Temperature

Atmosphere 2040-2140° 2060-2160° 2040-2140° 2060-2160°

Oxidizing 0.0 4.3 9.8 4.0

Reducing 0.0 4.4 10.1 3.9

For these data the SS^rror = 13.22 based on three repeated measurements. Com-
pletely analyze these data.
6.20 Sketch rough graphs for the following situations where each factor is at two
levels only:
1. Main effects A and B both significant, AB interaction not significant.
2. Main effect A significant, B not significant, and AB significant.
3. Both main effects not significant but interaction significant.
6.21 For a 2^ factorial can you sketch a graph (or graphs) to show no significant two-
factor interactions, but a significant three-factor interaction?
 7
Qualitative and Quantitative Factors

7.1 INTRODUCTION
In Chapter 3, Figure 3.1, we noted several possible procedures to follow
after an analysis of variance. In that chapter fixed levels of a single factor
were considered when such levels were qualitative levels. If the levels are
quantitative, such as temperature, dosage, humidity, or time, a different
route is suggested after an ANOVA yields significant differences between
response variable averages. It is also noted that the path to follow depends
upon whether or not these quantitative levels are equispaced. It will be seen
in this chapter that there are definite advantages in designing the experiment
so that the levels of any quantitative variables are equispaced.
It is assumed that the reader has had some introduction to linear re-
gression in which one tries to write a mathematical model of a straight line
that will predict some response variable Y from one independent variable
X. After reviewing this linear “fit,” the concept will be extended to fitting
a curve when a straight line is not an adequate model. These notions will
then be extended using the method of orthogonal polynomials to test for the
highest order polynomial necessary to do a decent prediction of Y from X.
The analysis will be made only if the analysis of variance shows a significant
difference in the Y averages due to levels of X and we wish to go further
and see how Y may be related to X.
After building some methodology to pull out various trend lines, the
notions will be extended to more than one factor where the factors in the
experiment may be all at quantitative levels or where some factors may be
at qualitative levels and some at quantitative levels. The next chapter will
extend these ideas to a multiple regression model where prediction equations
are sought and the experiment may not necessarily have all quantitative
levels equispaced.

129
130 Qualitative and quantitative factors

7.2 LINEAR REGRESSION

To review linear regression and to become familiar with the notation used,
consider an example where an experiment was designed to determine the
effect of the amount of drug dosage on a person’s reaction time to a given
stimulus.

Example 7.1 To study the effect of drug dosage in milligrams on a person’s

reaction time in milliseconds, 15 subjects were randomly assigned one of
five dosages of a drug—0.5, 1.0, 1.5, 2.0, or 2.5 mg—with three subjects
assigned to each level of drug dosage. The reaction times were recorded as
shown in Table 7.1. An analysis of variance of these data yields Table 7.2.

Table 7.1 Reaction Time Data

Dosage (mg)
0.5 1.0 1.5 2.0 2.5

Reaction 26 28 28 32 38
Time (ms) 28 26 30 33 39
yu 29 30 31 31 38

n 3 3 3 3 3 N = 15
83 84 89 96 115 T . = 467
Tj
lyfj 2301 2360 2645 3074 4409 XX yfj = 14,789
i i j

Table 7.2 Reaction Time ANOVA

Source df SS MS F Significance

Between dosages 4 229.73 57.42 28.72 0.000

Error 10 20.00 2.00

Totals 14 249.73

Obviously dosage has a highly significant effect on reaction time. But

since reaction time is a quantitative factor, the next question might well be:
How does reaction time vary with drug dosage? Can one find a functional
relationship between these two variables that might enable one to predict
reaction time from drug dosage?
As a first step one might plot a scattergram of all 15 X,Y pairs of points.
This is shown in Figure 7.1.
In Figure 7.1 the small x’s denote the average Y for each Xj or the T^’s.
The plot shows that Y increases with increases in X and one may try a
straight-line “fit” as a first approximation for predicting Y from X. Such
 7.2 Linear regression 131

Figure 7.1 Reaction time versus drug dosage.

a line is shown in Figure 7.1. One can see how close the line comes to the
average Y values for each X. Note that for a given Xj (say 2.5 mg) the observed
Yij can be partitioned into three parts: Y'x is the predicted value of Y^j for a
given Xj as found on the straight line, Yj — Y'x is the amount by which the
mean Y for a given Xj departs from its predicted value Y'x (referred to as
the departure from linear regression in this model), and Y^j — fj is the
amount by which an observed Y^j varies from its treatment mean.
In the analysis of variance model

Yij = fi + iHj - ^l) + {Yij - nj) (7.1)

and now one tries to predict fij from some function based on Xj. If this
predicted mean is labeled /ty/x, the model can be expanded to read

Yij = + {flyIX - /j) + (Pj - PY/X) + {Yij - fij) (7.2)

where jiy/x is the true predicted mean based on X and fij — fiy/x is the amount
by which the true mean juj departs from its predicted value.
For the sample model Equation (7.2) becomes

YIJ = Y, + {Y'^ - Y) + (Yj - yy + (Yij - YJ (7.3)

If the second and third terms on the right of this equation are combined, we
have the ANOVA sample model. Thus, we attempt here to partition the
132 Qualitative and quantitative factors

mean of a given treatment into two parts: that which can be predicted by
regression on X and the amount by which the mean shows a departure from
such a regression model The purpose, of course, is to find a model for pre-
dicting the response means such that the departures from these means are
very small By choosing a polynomial of high enough degree one can find
a model that actually goes through all the Y means. However, it is hoped
that an adequate lower degree model can be found whose departures are
small and insignificant when compared with the error of individual T’s
around their means.
In Equation (7.3) Y'x represents the predicted Y for a given X for any
assumed model. The straight-line model is only one possible attempt to “fit”
a curve to the data. The straight-line model is given by

= feo + (7.4)

where is the Y intercept and is its slope. The usual method for deter-
mining the values of BQ and is the method of least squares. The values of
bg and b^ are determined from the data in such a way as to minimize the
sum of squares of the deviations of each from its predicted value Y'x.
[One could also find bg and b^ such that the sum of squares of the departures
from regression {Yj — T^) is minimized, but this leads to the same results.]
Here
Yij - n = Y,j - foo - b,Xj
and
= y y {Y,j - yy^ = yX (Y^J -to- b,x/
i j i j

Since the summations of X’s and T’s can be found from a given set of
data, SSdeviations is a function of bg and b^.To minimize SS^eviations then, one
differentiates SS^eviations with respect to bg and b^

<)(SS, ,3.,ons) = 2yy(yy - bo - h,Xj)(-l) = 0

i j

^ 2y y (1^. - bo - b,Xj){-Xj) = 0
OOi i j

Dividing through by (— 2) and simplifying

Y.I.Y,J = boN +
i j j

llXjY,j = hon^Xj + b,n]:Xf\

I j j j

These equations are often called the least squares normal equations, which
^an now be solved for bg and
 7.2 Linear regression 133

From Table 7.1 Y.J \ = T. . = 467, N = 15, n = 3.

X Xj = 0.5 + 1.0 + 1.5 + 2.0 + 2.5 = 7.5

j

X Xj = (0.5)^ + (1.0)^ + (1.5)2 (2.0)2 + (2.5)2 = 13.75

j

To find the cross product Yj Zi ^ L ^jTj

gives

(0.5)(83) + (1.0)(84) + (1.5)(89) + (2.0)(96) + (2.5)(115) - 738.5

Substituting in Equations (7.5)

467 = 156o + 22.56i

738.5 - 22.5bo + 41.25fii
To solve one might multiply the first equation by 1.5, giving

700.5 = 22.5bo + 33.75fii

738.5 = 22.5^0 + 41.256i
Subtracting and solving for

700.5 - 738.5 _ -38.0

b1
33.75 - 41.25 “ -7.5

and from the original first equation

^ 467 - (22.5)(5.07)
bo = = 23.53

and the linear model for this set of sample data is

Y'x = 23.53 + 5.01Xj

Some previous work on a straight-line fit may have used the formula for
slope as
SP XY
SS X

or the ratio of the sum of cross products of X and Y to the sum of the squares
of X. And 6o as

bo = Y - b^X
Here

I- Y VVj
Zi I; 2^1 li YJ
SP XY N
hi =
SS X Hi 1J XJ}'-
Y IJ Xj - N
134 Qualitative and quantitative factors

In terms of Table 7.1 notation

nljXjT 3(7.5)(467)
^JTJ - 738.5 -
N 15
= 5.07
(n XJY 3(7.5)-
n 3(13.75) -
N 15

as before and
467
= Y - = (5.07) = 23.53
U
as before.

Table 7.3 Departures from Linear Regression

y.i 1'.; - (7j - 7';,P

0.5 27.667 26.067 1.600 2.560

1.0 28.000 28.600 -0.600 0.360
1.5 29.667 31.134 -1.467 2.152
2.0 32.000 33.667 -1.667 2.779
2.5 38.333 36.200 2.133 4.550

Totals 0 12.401

To see how well this linear model predicts the T’s from the X’s, we
construct Table 7.3. With the predicted values taken from the model
= 23.53 + 5S)lXj

as, for example, when Xj = 1.0,

7^ - 23.53 + 5.07(1.0) = 28.600

The table shows the departures from linear regression, for example,
7j- - 7^ = 28.000 - 28.600 = -0.600

Note that the sum of the departures adds to zero and the sum of squares
of departures from linear regression is

SS<,epar.ures = = 3(12.40) = 37.20

j
as each departure of a mean must be weighted with three observed values
for each Xj. Knowing the sum of squares of departures from linear regres-
sion, the ANOVA Table 12 can be expanded to give Table 7.4.
Table 7.4 shows a highly significant linear effect but also a significant
departure from the linear model at the 5 percent significance level. With
 7.2 Linear regression 135

Table 7.4 ANOVA on Reaction Time with Linear Regression

Source df SS MS F Significance

Between dosages 4 229.73

Linear 1 192.53 192.53 96.3*** 0.000
Departure from linear 3 37.20 12.40 6.2* 0.012
Error 10 20.00 2.00

Totals 14 249.73

such a departure it would be appropriate to try a second-degree equation

or a quadratic of the form

= f-o + b,Xj + b2Xj

Since there is a strong linear effect in this example and, in some prob-
lems, the linear model is adequate to explain most of the variation in the
Y variable, some useful statistics can be determined from Table 7.4.
The proportion of the total sum of squares that can be accounted for
by linear regression is sometimes called the coefficient of determination r^,
and its positive square root r the Pearson product-moment correlation co-
efficient. Here
r^ = 192.53/249.73 - 0.7712
and
r = 0.88

Thus linear regression will account for about 77 percent of the variation
seen in the reaction time Y.
By taking the ratio of the sum of squares between means to the total
sum of squares, one finds the maximum amount of the total variation that
could be accounted for by a curve or model that passes through all the
mean T’s for each X. This is called eta squared {q^). Here

= 229.73/249.73 = 0.9199

or approximately 92 percent of the variation in reaction time could be

accounted for by a model through all the means for each dosage Xj.
Another statistic used with a linear model is the standard error of
estimate (sy j^). It is the standard deviation of the deviations of the 1^/s from
their predicted values

Yij -Y'„ = {Y,j- F,) + {Yj- yy

This expression adds the error about the mean and the departure of the
mean from the predicted curve, which is appropriate only if the departure is
136 Qualitative and quantitative factors

nonsignificant. Then

^^departure T SS error
>y.x N - 2

'37.20 + 20.00
^Y.x — = 2.10
13

This statistic is often used to set confidence limits around a line of best fit
when linear regression is appropriate.
In this discussion of Example 7.1 no use has been made of the fact that
the dosages {XjS) are equispaced. When this is true, one can code the XjS
by considering their mean X and the width of the interval between them
c. Let
Xj-X, Xj - 1-5
Uj =
0.5
and our model becomes
n = ^0 + b\u (7.6)

whose least squares normal equations are given by

11 Y,j = b’oN + b\n y Uj

I J
(7.7)
ZZ Z «; + Z
I j
By equispacing the X'/s, the u'/s are —2, —1, 0, 1, 2, and = 0. Since
this is always the case, Equations (7.7) become

b'0 I. Z. Yi = Y
N
(7.8)
u, ^ ^ Zj "jTj
2
n ZJ “j ” Z; “ j J

and they are relatively easy to solve. Note, too, that since Yj
expression Yj ^j^.j ^ contrast and its sum of squares is (Yj UjT j)^/n Y
For our data

^0 Y = 467/15 = 31.133

-2(83) - 1(84) + 0(89) + 1(96) + 2(115) 76

b\ - 2.533
3(10) 30
and
r„ - 31.13 + 2.53u^-
 7.3 Curvilinear regression 137

That this is the same model, substitute for Uj and

Y'^ = 31.133 + 2.533{Xj - 1.5)/0.5
= 23.53 + S.OlXj

as before. Note also that the sum of squares due to linear regression can be
found from the linear contrast
(76)^
SS due to linear - 192.53
regression 3(10)
as shown in Table 7.4.

7.3 CURVILINEAR REGRESSION

When a linear regression model is not sufficient to explain all the significant
variation in the means, the next logical step is to consider a second-degree (or
quadratic) model:
Y'^ = bo + b,Xj + b2Xj (7.9)
By the methods of least squares, the normal equations are

YY,Y,J = boN + b,n Y^Xj + b^n X Xj

1 J

XX = Kn z Xj + b,n Z xj + b^n Z X] (7.10)

i j j j j

Z Z XjYij = hn Z V' + *1” Z

I J J J
+ bin Z
J

These equations can be solved for ho, b^, and 62, but it is difficult. Again
coded by w/s,
Y'„ = b'o + b\uj + b'^uj (7.11)

Z Z Yij =
I J
b’oN + b'lii Z Uj + fc'jn Z
Z Z njYij
i j
= b'on Z Uj +
j
b\n'^ u] + b'^ri
j
Z u]
j
(7.12)

Z Z njYij
1 J
= b'oii Z uj + b\n Z uj + b'^n Z uj
Because of the choice of the u/s, the sums of all odd powers of the w/s are
zero (^ Uj = ~ equations become

ZZ
I J
= b'oN + b'^n Z uj
Z Z njYij
i j
= b\nY, uj
J
(7.13)

ZZ = bon z "j + b'ln Z uj

138 Qualitative and quantitative factors

For the data of Example 7.1, the middle equation of (7.13) gives

b\ = E "ija 2.53
n Uj 30
The other two equations of (7.13) give:

467 = \5Wo + 30b'2

972 = 30h'o + 102h'2
since

Ii Z ^Ytj = I "]T,j = 4(83)

j j
+ 1(84) + 0(89) + 1(96) + 4(115) = 972

and
X uf = (-2)^ + (-1)^ + (0)^ + (1)^ + (2)^ - 34
j
and

Multiplying the first equation by 2 and subtracting it from the second:

972 = 30b'o + 102h'2

934 - 30b'o + 60b'2
38 = 42h;
and
38
b'2=-- = 0.90
^ 42
and
467 - (30)(0.90)
h'
UQ —
— = 29.33
15
and the quadratic model in terms of Uj is

y; = 29.33 + 2.53u,. + 0.90uj

Table 7.5 Departures from Quadratic

y.j
Y' Yj - Y', {y.j - yx
' )"

0.5 -2 27.67 27.87 -0.20 0.040

1.0 -1 28.00 27.70 0.30 0.090
1.5 0 29.67 29.33 0.34 0.116
2.0 1 32.00 32.76 -0.76 0.578
2.5 2 38.33 37.99 0.34 0.116

Totals 0.02 0 0.940

 7.4 Orthogonal polynomials 139

One need not write this in terms of Xj to see how well it predicts the y’s
from the X’s. Table 7.5 shows the predicted T’s and the departure of the means
from the quadratic model.
From Table 7.5 the sum of the squares of the departure from the qua-
dratic is

SSd.par.ure = 3(0.94) = 2.82

and this step can be added to refine Table 7.4 further and give Table 7.6.
Now the departure from the quadratic is not significant at the 5 percent level
so it is appropriate to stop with the quadratic for predicting reaction time
from dosage. One may also note from Table 7.6 that linear and quadratic
terms together account for 192.53 + 34.38 = 226.91 of the total sum of
squares in Y of 249.73. This is 226.91/249.73 = 0.9086 or approximately
91 percent of the variability accounted for. The square root of this statistic
is sometimes referred to as a correlation ratio R and = 0.91. The error
and nonsignificant departure term could be ‘pooled’ to give a standard error
of estimate for the quadratic

'20.00 + 2.82
^Y.X, 1.38
10 + 2
which might be used to set confidence limits on the T’s around the quadratic
curve.

Table 7.6 ANOVA with Quadratic Regression

Source df SS MS F Significance

Between dosages 4 229.73

Linear 1 192.53 192.53 96.3*** 0.000
Quadratic 1 34.38 34.38 17.2** 0.002
Departure 2 2.82 1.41 <1 0.517
Error 10 20.00 2.00

Totals 14 249.73

7.4 ORTHOGONAL POLYNOMIALS

In the above discussions on linear and quadratic regression it was noted that
the problem could be solved much more easily if the Xj values were equi-
spaced and coded such that — Xj ^j ^ advantage of this type
of curve fitting is even greater as the degree of the polynomial increases. If

y; = f{x, x\ x\..., x'‘)

140 Qualitative and quantitative factors

is a polynomial in X, it has been shown that it may be rewritten in the form

+ ^1^1 + + ■ * ■ + (7.14)
where each is a polynomial of degree j and all polynomials, such as and
are orthogonal to each other. The advantage of writing the model in this
form is that additional polynomials of higher degree may be added which are
independent of the ones already considered. The values of these polynomials
are given for the first five powers of u as

«0 = 1
=

= A-

/3k^ - T
= A. U' u{ - (7.15)
20
.2

= A. u

n
= A, U' (k^ - 7) + (15/c* - 230k^ + 407)
18 1008

where k is the number of levels of the factor X, and the A’s are chosen so that
the are integers for all u. Because of the complications of these formulas
and the practical need to test for significant high-order regression models,
the values of the and A’s have been tabled by Fisher and Yates [11], and
a section is reproduced as Appendix Table F.
The A/s are given by
L.J Yjjij (7.16)
' L.J a?"
To see how these formulas and tables apply, consider a quadratic model
with just three points on the X scale. The u/s would be — 1, 0, + 1. As these
are already integers, A^ = 1 and takes on values —1, 0, +1 for each j.
For {'2, {k^ — 1)/12 = (9 — 1)/12 = 2/3 and the values are -l-l, 0, +1,
which makes [t/^ — (/c^ — 1)/12] ^ 1/3, —2/3, 1/3 so A2 is taken as equal
to 3 in order to make the ^'2 integers. Then the {'2 are +1, — 2, +1. Appendix
Table F gives these values for /c = 3. Also given in the table are ^ (fj)^ for
each set of and the Aj. In the expression for Aj, each Y^j is multiplied by
the corresponding {j, and since — 0. this expression is a contrast. Since
the coefficients — 1, 0, +1 are orthogonal to 1, —2, 1, these provide two
orthogonal contrasts on the }//s (or their totals, T j) and the sum of squares
 7.4 Orthogonal polynomials 141

for such contrasts is given by

cc (Zi S; Yu ir
^ ^

" ZJ iQ
One needs only to apply the coefficients given in Appendix Table F to the
treatment totals of a single-factor problem to determine the sum of squares
for linear, quadratic, cubic, quartic, and so on, effects up to the (k — l)th-
degree equation. Equations (7.14), (7.15), and (7.16) are only needed if the
prediction model = /(u, u^, . . ., u^~^) is desired.
To illustrate this, consider the example of Sections 7.2 and 7.3. Since
k = 5 dosage levels. Table F gives the following coefficients for the linear,
quadratic, cubic, and quartic effects.

T, 83 84 89 96 115 I
j
ii’/ Q Contrast SS

Linear -2 -1 0 1 2 10 1 76 192.53
Quadratic 2 -1 -2 -1 2 14 1 38 34.38
Cubic -1 2 0 -2 1 10 5/6 8 2.13
Quartic 1 -4 6 -4 1 70 35/12 12 0.69

Total 229.73

Applying each set of coefficients to the treatment totals gives the contrasts
as listed and their corresponding sum of squares. For example, the cubic
contrast is
-1(83) 2(84) + 0(89) - 2(96) + 1(115) = 8

and its sum of squares

SS,,,i, = (8)V3(10) = 2.13

as shown. The complete ANOVA can now be written as in Table 7.7. The
F column again shows that there are no significant effects after the quadratic
and the quadratic should be adequate for prediction purposes.
If the equation is desired, one needs to find the values of and Aj for
each degree of the polynomial desired. For our data note that

. _ Zi ZJ _ Y.j T.ji’j _ Contrast

' LZt(CZ

This expression is very similar to the expression for the sum of squares of a
contrast, except that the contrast is not squared and it will retain its sign.
142 Qualitative and quantitative factors

Table 7.7 ANOVA for Polynomial Model

Source df SS MS F Significance

Between dosages 4 229.73

Linear 1 192.53 192.53 96.3 0.000
Quadratic 1 34.38 34.38 17.2 0.002
Cubic 1 2.13 2.13 1.1 0.325
Quartic 1 0.69 0.69 <1 0.570
Error 10 20.00 2.00

Totals 14 249.73

The A'jS and in terms of u are then

0 AQ = y. = 31.13 4'o = 1

76
1 2.53 = iw
“ 3(10) ■
38
2 A2 0.90 i'2 = - 2)
“ 3(14) “
8
3 A3 0.27 <^3 = 5/6(M^ — 3Au)
“ 3(10) "
12
4 •44 0.06 i'4 = 35/12
“ 3(70) “

If the A'jS and are multiplied together, they give a polynomial in u to as

high a degree as desired up to the {k — l)th degree. Hence the best fitting
straight line is

TL = Aoi'o + A,S',
= 31.13 + 2.53u
as given in Section 12.
For a quadratic

FL = A,S'o + + ^2C2
- 31.13 + 2.53u + 0.90(u^ - 2)

which simplifies to

7' = 29.33 + 2.53u + 0.90^^

as given in Section 7.3.
 7.4 Orthogonal polynomials 143

One has only to add more terms as each polynomial in u, is orthogonal

to those terms already in the model. Continuing to the quartic, even though
unnecessary, we find

FL = 31.13 + 2.53M + 0.90(M^ - 2) + (0.27)5/6(M^ - 3.4M)

+ 0.06(35/12)[M‘‘ - 62/14M" + (3/560)(24)(16)]

which simplifies to

y; = 29.69 + 1.76M + 0.12M^ + 0.22M=’ + 0.18M^

“Fitting” this quartic polynomial gives Table 7.8. This table shows the
quartic to go through the five means except for slight rounding errors.

Table 7.8 Departures from Quartic Model

Uj Y'

0.5 -2 27.67 27.77

1.0 -1 28.00 28.01
1.5 0 29.67 29.69
2.0 1 32.00 31.97
2.5 2 38.33 38.33

This method of orthogonal polynomials is then very useful in finding

the degree of the prediction model needed to “fit” a set of data, provided the
Xys are equispaced. One usually stops after the highest degree polynomial
that shows significance and then writes its equation. In Example 7.1 it would
be the quadratic or second-degree equation.
Some readers may wonder how the coefficients in Table F were arrived
at. Without attempting any derivation, let us consider three equispaced X’s
and the corresponding three totals: T/s as shown in Figure 7.2. From
Figure 7.2, as one increases X from to X3 the increase in Yj is reflected

T3

T.2

TA

X
Xi X3

Figure 7.2 Linear and quadratic contrasts.

144 Qualitative and quantitative factors

in the increase in T j or T 3 —T which can be considered a linear contrast.

The linear effect from to X2 is T2 — T 1 and from X2 to X3 is T 3 — T 2;
so the cumulative or average linear effect is
T3 - T2 + T2 - T.i = T3 - T.i
If, however, the response curve does not have the same slope from 1 to 2
as from 2 to 3, the slope changes and there is a quadratic effect present.
This is measured by a difference in the two linear effects, if they do differ.
Hence the quadratic effect is
(T3 - T2) - {T2 - T,)
or
7.3 - 27.2 + 7.1

which is a contrast and which is also orthogonal to the linear contrast

T2 — 7 i. We note that these are the same coefficients as given in Table F
for k = 3.
If there are four points X^, X2, X^, and X4, a cubic effect is present if
the quadratic from 1 to 3 is different from the quadratic from 2 to 4. Thus
(7.4 - 27.3 + Ti) - {T3 - 27.2 + Ti)
or
7.4 - 37.3 + 37.2 - Ti
whose coefficients of —1, 3, —3, and 1 agree with Table F for /c = 4. This
reasoning may give some notion as to the reasonableness of the orthogonal
polynomial coefficients.

7.5 COMPUTER PROGRAM

As in previous chapters a computer program could be used to analyze the

results of a problem involving quantitative levels of a factor. For Example 7.1,
in addition to asking for a ONEWAY ANOVA of RT (reaction time) by DOS
(dosage) in Table 7.9, on the next line one asks for POLYNOMIAL = 4
for the four possible curves that might fit the data. After entering the 15 data
points (steps 13 through 27) the printout is given as in Table 7.10.
These results are seen to be the same as in Table 7.7 if one notes that
the deviation from cubic is the quartic effect.

7.6 QUANTITATIVE AND QUALITATIVE FACTORS

In Sections 7.1 to 7.5 a single factor was considered at quantitative levels

only. Qualitative levels of a single factor were discussed in Chapter 3 and for
two or more factors in Chapter 5. When two or more factors are involved in
 7.7 Two factors—one qualitative, one quantitative 145

an experiment, some may be at qualitative levels and some at quantitative

levels. Two factors are considered in the sections that follow.

7.7 TWO FACTORS—ONE QUALITATIVE,

ONE QUANTITATIVE

A problem will now be considered where one factor is set at qualitative levels
and the other at quantitative (and equispaced) levels.

Example 7.2 We wish to detenuine the effect of both depth and position
in a tank (Figure 7.3) on the concentration of a cleaning solution in ounces
per gallon. Concentrations are measured at three depths from the surface
of the tank, 0 inch, 15 inches, and 30 inches. At each depth measurements

Table 7.9 Computer Input for Data of Table 7.1

+ + +CREATE PROB8
1. 000 = 77509, 113
2. 000 = COMMON (SPSS)
3. 000 = SPSS
4. 000 = #EOR
5. 000 = RUN NAME STAT 502 PROB8
6. 000 = VARIABLE LIST RT, DOS
7. 000 = INPUT FORMAT (F2.0, IX, Fl.O)
8. 000 = N OF CASES 15
9. 000 = ONEWAY RT BY DOS (1, 5)/
10. 000 = POLYNOMIAL - 4/
11. 000 = STATISTICS ALL
12. 000 = READ INPUT DATA
13. 000 = 26 1
14. 000 = 28 1
15. 000 = 29 1
16. 000 = 28 2
17. 000 = 26 2
18. 000 = 30 2
19. 000 = 28 3
20. 000 = 30 3
21. 000 = 31 3
22. 000 = 32 4
23. 000 = 33 4
24. 000 = 31 4
25. 000 = 38 5
26. 000 = 39 5
27. 000 = 38 5
28. 000 = FINISH
 O o
o
o ON
o v-H
o O'!
r^i
o <N
NO •»-H
o o O i-H <N
DH o o o o ITN m
[JH

o
HH r- o o GN m
H i-H NO o ON o NO
< r- (N (N ■1 1 P ro
od NO NO r-'
CNI ON
[JU

GO
W
cti
< rn m o o tT) m O
m m o 1—t ON fO GN o
:D
m m o oo O m oo o
a Nn m NO p
GO r-’ c4 (N r4 r4
Ln ON fO
z
pj
u “
Z S
<
M—I
Cei
< w
> CcJ
UH <
O Z) m
m
m o
m o
O
1—1
O
ON
m
ro GN
O
o
m
m
GO C/ ro m o OO <■ H m OO o m
GO
GO GN <N m p NO o p
IJH
ON <>i r4 (N o ON
<N ON ro m (N

<
O <N (N

z
0-)
< D
(y:i
’a
Sa
x tZ m <N
W
Q
,o
=5
O
< Q u
HH
*n
DH GO
w
z < 03
HH D
< P
P a U GO
> PJ D p P
O U O w p P S Z)
z Od P H o w
H
o p o o
< D O P p p H p p
O
GO z <
U
p
Q
p U
HH
p
o
w
w
h-^
z >■ < > m > ;z; <
r- w D p D p HH
H
2 H
p Q a Q U Q z
H
O
HH H
H

146
 7.7 Two factors—one qualitative, one quantitative 147

Figure 7.3 Positions in tank at each depth.

are taken at five dififerent lateral positions in the tank. These are considered
as five qualitative positions, although probably some orientation measure
might be made on them. At each depth and position, two observations are
taken. This is then a 5 x 3 factorial with two replications per cell (total of
30 observations). The design is a completely randomized design and the
model is
^ijk = + Di Pj -{■ DPij +
with
z = 1, 2, 3 7 = 1, 2,..., 5 k = 1,2
where D, represents the depth and Pj the position. The data collected are
shown in Table 7.11.
In this case where one factor is set at quantitative levels (Dj) and the
other at qualitative levels (Pj), the first step is to run a two-factor analysis
of variance just as if both were qualitative factors. Coding the data of Table
7.11 by subtracting 5.90 and multiplying by 100 gives Table 7.12.

Table 7.11 Cleaning-Solution Concentration Data

Depth from Top of Tank D,

Position Pj 0 in. 15 in. 30 in.

5.90 5.90 5.94

5.91 5.89 5.80
5.90 5.89 5.75
5.91 5.89 5.83
5.94 5.91 5.86
5.90 5.91 5.83
5.93 5.94 5.83
5.91 5.90 5.89
5 5.90 5.94 5.83
5.87 5.90 5.86
148 Qualitative and quantitative factors

From these coded data,

(-44)-
SS,.,.,, = 644 - = 579.47
30

33^ ^ (7)^.(Z)-^ (r^ _ (Z^ . 281.67

10 30
(-6)" + (-23)" + (-5)" + (0)" + (-10)"
SSp =

(-44)-
= 50.47
30
(1)" + (1)" + (4)" + ■■• + (-11)" (-44)-
SS„ X P interaction
30
- 281.67 - 50.47 = 58.33
SSerror = 579.47 - 281.67 - 50.47 - 58.33 = 189.00
giving the ANOVA table of Table 7.13.
Table 7.12 Coded Cleaning-Solution Concentration Data with Totals

Depth from Top of Tank Di

Position Pj 0 in. 15 in. 30 in. T.j.

1 0 0 4
1 -1 -10
1 -1 -6 -6

2 0 -1 -15
1 -1 -7
1 -2 -22 -23

3 4 1 -4
0 1 -7
4 2 -11 -5

4 3 4 -7
1 0 -1
4 4 -8 0

5 0 4 -7
-3 0 -4
-3 4 -11 -10

Ti.. 7 7 -58
II
1

i4
k=l j=l
37 37 570 L
i
44
J k
= 644
 7.7 Two factors—one qualitative, one quantitative 149

Table 7.13 ANOVA for Cleaning-Solution

Concentration Problem

Source df SS MS

Depths D, 2 281.67 140.83

Positions Pj 4 50.47 12.62
D X P interaction DPij 8 58.33 7.29
Error 15 189.00 12.60

Totals 29 579.47

From Table 7.13 it is seen that only depth produced a significant effect
on concentration of solution as

140.83
11.24
12.60

which is significant at the 1 percent level of significance. Since the depth

effect is significant and since the three depths are equispaced, it may be
worthwhile to extract a linear and quadratic depth effect to learn how
concentration varies with depth. The coefficients for /c = 3 levels are shown
in Table 7.14.
Applying these coefficients to the depth totals Tj gives

linear effect of depth = —1(7) + 0(7) + 1( —58) = — 65

quadratic effect of depth = +1(7) — 2(7) + 1( —58) = — 65

Since these are orthogonal contrasts, their sums of squares are

(-65)^
SS linear 211.25
10(2)
(-65)^ 70.42
SS quadratic
10(6) 281.67

and their total is the depth sum of squares.

Even though the D x P interaction is not significant, there may be an
interaction between the linear effect of depth and positions or between the

Table 7.14 Orthogonal Coefficients

Z (& A

Linear -1 0 +1 2 1
Quadratic +1 -2 +1 6 3

T,. 7 ' 7 -58

150 Qualitative and quantitative factors

quadratic effect of depth and positions. To compute these interactions, the

linear effect of depth is determined at each position and these effects are then
compared to see whether or not they differ. The same procedure is followed
for the quadratic effect of depth at each position.
Applying the linear coefficients of — 1, 0, +1 at each position gives
Pi:-l(l) + 0(-l)+ l(-6) = -7
P2: -1(1) + 0(-2) + l(-22) = -23
P3: -1(4)+ 0(2)+ l(-ll) = -15
P^: -1(4) + 0(4) + l(-8)
= -12
P^:-l(-3) + 0(4)+l(-ll)^ -8
-65
To compare these five effects, determine the sum of squares between
them
(-7)2 + (_2y + (-15)^ + (-12)^ + (-8)^ _ (-65)^ ^
2(2) 10(2)
Similarly, for the quadratic effect of depth at the five positions, we have

P,: +1(1) - 2(-l) + l(-6) = -3

P2: +1(1) - 2(-2) + l(-22) = -17
P3: +1(4) - 2(2) + l(-ll) = -11
P4: +1(4) - 2(4) + l(-8) = -12
P5: +l(-3) - 2(4) + l(-ll) = -22

Comparing these five quadratic effects gives

(-3)2 + (_I7)2 (_22)2 (_65)2

2(6^ 1^ =
Note that the sum of these two sums of squares (41.50 + 16.83) equals
the D X P interaction sum of squares (58.33) as it should. Summarizing for
this problem with one quantitative and one qualitative factor, the complete
ANOVA breakdown is as shown in Table 7.15.
The results in Table 7.15 indicate a strong depth effect, with the linear
depth effect significant at the 1 percent level (**) and the quadratic depth
effect significant at the 5 percent level (*). There is no position nor interaction
between depth and position. These results seem reasonable from a graph of
cell totals versus depth and positions (Figure 7.4).
This graph shows little interaction, as the curves are quite parallel
(statistically speaking) and there is little difference between the five position
curves. The depth effect is quite obvious, and concentration is seen to drop
 7.7 Two factors—one qualitative, one quantitative 151

Table 7.15 Complete ANOVA for Cleaning-Solution

Concentration Problem

Source df SS MS

Depths 2 281.67
Linear 1 211.25 211.25**
Quadratic 1 70.42 70.42*
Positions 4 50.47 12.62
D X P interaction 8 58.33
l>linear X P 4 41.50 10.38
n
^quadratic
Y p
^ 4 16.83 4.21
Error 15 189.00 12.60

Totals 29 579.47

+5

Cell
Totals
(Coded)
-10

-15

-20

0 15 30
Depth (inches)

Figure 7.4 Graph of cleaning-solution concentration problem.

off with increasing depth but on more of a curve than a straight line. Further
investigation is suggested at depths between those already studied. The lack
of any position effect or interaction should mean that this new experiment
could be run at only one position; the results should then be the same at all
five positions.
If one wishes the equation for predicting concentration (coded) from
depth,
X - 15
u =
15
and
Ao= Y.= -44/30 - -1.47
A, = -65/10(2) - -3.25
A2 = -65/10(6) - -1.08
152 Qualitative and quantitative factors

and
- 1
S'l =

= 3(w^ - 2/3) = 3u^ - 2

and
7' = -1.47 - 3.25w - 1.08(3w^ - 2)
= 0.69 — 3.25w — 3.24w^
giving predictions as follows:

Xj Uj Y'
^ U
Y

0 -1 0.70 0.70
15 0 0.69 0.70
30 1 -5.80 -5.80

which match exactly as they should.

7.8 TWO FACTORS—BOTH QUANTITATIVE

If both factors in a two-factor factorial are at quantitative levels, each factor

can be broken down into its linear, quadratic, or cubic effects, and all com-
binations of interaction can be determined, such as linear by linear, linear by
quadratic, and quadratic by quadratic. To illustrate the procedure for
analyzing a factorial experiment where both factors are at quantitative (and
equispaced) levels, consider Example 7.3.

Example 7.3 Data were collected in a completely randomized design on the

effect of lacquer concentrations at four equispaced levels and standing times
at three equispaced levels on screen quality as indicated on a 20-point scale.
Two observations were made for each of the 4 x 3 = 12 treatment com-
binations. The resulting data are shown in Table 7.16.
Table 7.16 Screen Quality Data

Lacquer Concentration
Standing
1
Time 2 1 li 2

30 16 12 17 13
14 11 19 11
20 15 14 15 12
15 17 18 14
10 10 7 10 9
9 6 14 13
 7.8 Two factors—both quantitative 153

The ANOVA model is

^ijk = + Li Tj LTij + e^ij)

i = 1, 2, 3, 4 j = 1, 2, 3 k= 1,2
and the analysis can be shown to be as given in Table 7.17. These results
indicate strong main effects—lacquer concentration and standing time—but
no significant interaction. One could then extract linear, quadratic, and
cubic effects for the lacquer concentrations and linear and quadratic effects
for the standing times as two independent factors. However, since some
examples may have significant quantitative by quantitative interactions, we
will go through the steps to partition the interaction as well as the main
effects into component parts. For lacquer concentrations there are four
levels, or /c = 4 in Table F. To apply these coefficients one needs the four
lacquer totals. These are seen, along with other useful totals, in Table 7.18
based on Table 7.16.

Table 7.17 Screen Quality ANOVA

Source df SS MS F Probability

Lacquer(Ld 3 63.79 21.26 7.38 0.005

Time [Tj) 2 126.58 63.29 21.98 0.000
L X T interaction 6 38.07 6.35 2.20 0.115
Error 12 34.51 2.88

Totals 23 262.95

Table 7.18 Screen Quality Data with Totals

Lacquer Concentration
Standing
Times j 1 l4 2 Tj
154 Qualitative and quantitative factors

Applying the orthogonal coefficients gives the following.

T. 79 67 93 72 ii'/ Contrast SS

Linear -3 -1 1 3 20 2 5 0.21
Quadratic 1 -1 -1 1 4 1 -9 3.37
Cubic -1 3 -3 1 20 10/3 -85 60.21
Total = 63.79

Applying appropriate orthogonal coefficients to the standing time totals

(here k = 3) gives the following.

T.J 78 120 113 d'/ Contrast SS

Linear -1 0 1 2 1 35 76.56
Quadratic 1 -2 1 6 3 -49 50.02
Total SSj = 126.58

These results partition the main effects into components with 1 df each and
they can be tested for significance. To break down the interaction term, note
that there are six single-degree-of-freedom terms
LL X LQ X TLQ X

where subscripts L, Q, and C represent linear, quadratic, and cubic respec-

tively. A simple computing scheme is to apply the product of the main effect
coefficients to the treatment totals as shown in Table 7.19 for linear by
linear, x T^, interaction.

Table 7.19 Linear by Linear Interaction

LL

-3 -1 1 3

-3 -1 1 3
1
30 23 36 24

0 0 0 0
TL 0
30 31 33 26

3 1 -1 -3
-1
19 13 24 22
 7.8 Two factors—both quantitative 155

In Table 7.19 the number in the upper-left-hand corner of each cell is

the product of the coefficients of the main effects for its row and column.
The number in the right-hand corner is the cell total of Tij. Applying the
coefficients to these ’s gives the x contrast as follows:

-3(30) - 1(23) + 1(36) -h 3(24) -h 0(30) 4- 0(31) 0(33)

+ 0(26) + 3(19) + 1(13) - 1(24) - 3(22) = -25
and
(-25)^
SS LJL X 7.81
TL
2(20)(2)
The sum of squares of coefficients in the denominator is the product of the
two {^j}^ from the two factors.

Table 7.20 Linear by Quadratic Interaction

-3-11 3

-3 -1 1 3
1
30 23 36 24

6 2 -2 -6
TQ -2
30 31 33 26

-3 -1 1 3
1
19 13 24 22

To find the x TQ use the same procedure with quadratic coefficients

on the standing times as shown in Table 7.20. The x TQ contrast is now

-3(30) - 1(23) + 1(36) -b 3(24) -b 6(30) -b 2(31) - 2(33)

- 6(26) - 3(19) - 1(13) -b 1(24) + 3(22) = 35
and
(35)^
SS LL ^ TQ
5.10
2(20)(6)
If this same procedure is now applied to LQ X T^, LQ X TQ, X T^, and
Lc X TQ, the resulting sums of squares are

^^LQXTL ~~ 5.06

^^LQ.TQ= 4.69

^^LcxTL ^ 2.81

TQ =
12.60
156 Qualitative and quantitative factors

Table 7.21 Complete ANOVA for Screen Quality Example

Source df SS MS F Probability

Lacquer 3 63.79
Linear 1 0.21 0.21 <1 0.791
Quadratic 1 3.37 3.37 1.17 0.301
Cubic 1 60.21 60.21 20.91 0.001
Time 2 126.58
Linear 1 76.56 76.56 26.58 0.000
Quadratic 1 50.02 50.02 17.37 0.001
L X T interaction 6 38.07
FL X U 1 7.81 7.81 2.71 0.125
FL TQ 1 5.10 5.10 1.77 0.208
1 5.06 5.06 1.77 0.208
LQ X TQ 1 4.69 4.69 1.59 0.231
Fc X U 1 2.81 2.81 <1 0.343
Fc X TQ 1 12.60 12.60 4.38 0.058
Error 12 34.51 2.88

Totals 23 262.95

Adding the six components gives 38.07, the interaction sum of squares
reported in Table 7.17. The complete breakdown and tests of significance
can be summarized as in Table 7.21. This table shows three significant effects:
a cubic effect of lacquer concentration on screen quality, and a linear and
quadratic effect of standing time on screen quality. Since there is no signifi-
cant interaction one might plot the mean responses to each main effect as
in Figure 7.5.
Because the two main effects do not interact one might find an equation
to predict the lacquer means (a cubic) and another equation to predict
standing time means (a quadratic). These two equations could be combined
to predict the cell or treatment means. Use Equation 7.15 and the contrasts
found in the data to give
For lacquer (/c = 4):

^/s
= F = 311/24 = 12.9583
A, = 5/6(20) = 0.0417
.4 2 = -9/6(4) = -0.3750
A, = -85/6(20) = -0.7083

and
7' = 12.9583 + 0.0834U - 0.375(u^ - 1.25) - 0.7083(10/3)(u=' - 2.05M)
 7.8 Two factors—both quantitative 157

Figure 7.5 Screen quality averages versus lacquer concentration and standing times.

which simplifies to

= 13.4270 + 4.9234W - 0.375w^ - 2.361t^^

For standing times (/c = 3):

^0 =: Y 12.9583 io = 1
•4i —
35/8(2) = 2.1875 = iM
A, -49/8(6) = -1.0208 ^2 = 3(M" - 2/3)
and
Y' = 12.9583 + 2.1875U -- 1.0208(3u^ -- 2)
or
Y'
^ u - 14.9999 + 2.1875t/ -- 3.0624M^

Now the lacquer concentration equation will have u values as coded for the
lacquer data and the standing time equation for coded standing times.
Calling the lacquer variable

- 1.25
0.5
158 Qualitative and quantitative factors

and Uj for standing time

- 20
To
As no interaction is assumed the two equations could be combined by addi-
tion, remembering, however, that the grand mean is in both equations. The
equation for predicting the mean screen quality for any combination of L
and T is then
r = 13.4270 + 4.9234wi^ - 0.375wi - 2.361ui + 14.9999 2.1875^7^
- 3.0624u| - 12.9583
or
r = 15.4686 + 4.9234ui^ - 0.375wi - 2.361ui + 2.1875wr - 3.0624u?-

Table 7.22 Screen Quality Based on Prediction Equation

Standing Lacquer Concentration

Times
Az. = i 1 li 2
XY 1 1 3
IAY 2 2 2

14.33 12.33 16.67 13.17

30 1
15.0 11.5 18.0 12.0

15.21 13.21 17.54 14.04

20 0
15.0 15.5 16.5 13.0

9.96 7.96 12.29 8.79

10 -1
9.5 6.5 12.0 11.0

To see how good a predictor this is. Table 7.22 gives the predicted
value at the upper left of each cell using this equation and the observed
mean of the cell at the lower right of that cell. If one now examines the de-
partures of the cell mean in Table 7.22 from their predicted values, the sum
of squares of the departures is
(15.0 - 14.33)2 + (11.5 - 12.33)2 + • • • + (11.0 - 8.79)2 _
Since each mean represents two observations

SSdeparture = 2(19.04) = 38.08

which compares well with the interaction sum of squares of 38.07 reported
in Table 7.17 or Table 7.21.
 Problems 159

7.9 SUMMARY

The examples of this chapter may easily be extended to higher order fac-
torials whenever one or more factors are considered at quantitative levels.
The use of orthogonal polynomials makes the analysis rather simple, pro-
vided the experiment is designed with equispaced quantitative levels. The
summary of designs at the end of Chapter 6 has not been changed by this
chapter, as these methods can be used in all experiments where quantitative
levels are involved.
Chapter 8 presents a more general method for handling many inde-
pendent variables in determining their effect on some dependent variable 7.

PROBLEMS

7.1 For the following data on X and Y plot a scattergram and determine the least
squares straight line of best fit.

A
3 4 5 6 7

7 8 10 11 10
7 8 8 9 9 10
9 9 9 10 9

7.2 For Problem 7.1 present your results in an ANOVA table and comment on these
results.
7.3 From your table in Problem 7.2 find and the standard error of estimate.
7.4 Use the method of orthogonal polynomials on the data of Problem 7.1 and find
the best fitting polynomial.
7.5 An experiment to determine the effect of planting rate in thousands of plants per
acre on yield in bushels per acre of corn gave the following results.

Planting Rate Yield

A 7

12 130.5, 129.6, 129.9

16 142.5, 140.3, 143.4
20 145.1, 144.8, 144.1
24 147.8, 146.6, 148.4
28 134.8, 135.1, 136.7
160 Qualitative and quantitative factors

Plot a scattergram of these data and comment on what degree polynomial might
prove appropriate for predicting yield from planting rate.
7.6 From the data of Problem 7.5 use orthogonal polynomials and determine what
degree polynomial is appropriate here.
7.7 Find the equation of the polynomial in Problem 7.6.
7.8 From an ANOVA table of the results of Problem 7.6 find R^, and comment
on what these statistics tell you.
7.9 Seasonal indexes of hog prices taken in two different years for hve months gave
the following results.

Month

Year 1 2 3 4 5

1 93.9 94.3 101.2 96.7 107.5

2 95.4 96.5 94.7 97.1 107.2

a. Sketch a scatterplot of these data and comment on what type of curve might
be used to predict hog prices when the month is known.
b. Do a complete analysis of these data and find the equation of the best fitting
curve for predicting hog price from month.

7.10 A bakery is interested in the effect of six baking temperatures on the quality of its
cakes. The bakers agree on a measure of cake quality, Y. A large batch of cake mix
is prepared and 36 cakes poured. The baking temperature is assigned at random
to the cakes such that six cakes are baked at each temperature. Results are as
follows:

Temperature
o

0
oc

175° 195° 215° 225°

22 4 10 22 32 20
10 21 22 14 27 36
18 18 18 21 22 33
26 28 32 25 37 33
21 21 28 26 27 20
21 28 25 25 31 25

Do a complete analysis of these data and write the ‘best’ fitting model for predicting
cake quality from temperature.
7.11 Compute statistics that will indicate how good your fit is in Problem 7.10 and
comment.
 Problems 161

7.12 Data on the effect of age on retention of information gave

Ages: 6-14 15-23 24-32 33-41

Total score of ten
people, T j: 80 92 103 90

Source df SS MS

Between ages 3 26.675 8.892

Linear 1 8.405 8.405
Quadratic 1 15.625 15.625
Cubic 1 2.645 2.645
Error 36 72.000 2.000

Total 39 98.675

a. Determine the significance of each prediction equation and state which degree
equation (linear, quadratic, or cubic) will adequately predict retention score
from age.
b. Write the equation for part (a).
c. Find the departure from regression of the mean retention score of a 37-year-old
based on your answer to part (b).
7.13 A study of the effect of chronological age on history achievement scores gave the
following results.

History Achievement
Chronological Age
O .. _ J
T - 18
J 1- -
ocol t/ L)y
^ 15 8 9 10 11 12

5 1 6 5
4 1 7 4 5
3 6 2
2 3 2
1 6 1
0 1

Do a complete analysis justifying your curve of best fit.

7.14 In studying the effect of the distance a road sign stands from the edge of the road
on the amount by which a driver swerves from the edge of the road, hve distances
were chosen. Then the five distances were set at random, observing four cars at
each set distance. The results gave

rj^ = 0.70

= 0.61
162 Qualitative and quantitative factors

a. Test whether or not a linear function will “ht” these data. (You may assume a
total SS of 10,000 if you wish, but this is not necessary.)
b. In fitting the proper curve, what are you assuming about the four readings at
each specified distance?
7.15 Thickness of a film is studied to determine the effect of two types of resin and three
gate settings on the thickness. Results showed:

Gate Setting (mm)

Resin Type 2 4 6

1 1.5 2.5 3.6

1.3 2.5 3.8
2 1.4 2.6 3.7
1.3 2.4 3.6

Do an analysis of variance of these data following the methods of Chapter 5.

7.16 For Problem 7.15 extract a linear and quadratic effect of gate setting and test for
significance. Also partition the interaction and test its components.
7.17 Write any appropriate prediction equation or equations from the data of Prob-
lem 7.15.
7.18 If the experiment of Problem 7.15 is extended to include a third factor, weight
fraction at three levels, the data are as follows:

Resin Type

1 2
Gate Weight Fraction Weight Fraction
Setting
(mm) 0.20 0.25 0.30 0.20 0.25 0.30

2 1.6 1.5 1.5 1.5 1.4 1.6

1.5 1.3 1.3 1.4 1.3 1.4
4 2.7 2.5 2.4 2.4 2.6 2.2
2.7 2.5 2.3 2.3 2.4 2.1
6 3.9 3.6 3.5 4.0 3.7 3.4
4.0 3.8 3.4 4.0 3.6 3.3

Compile an ANOVA table for this three-factor experiment as in Chapter 5.

7.19 Outline a further breakdown of the ANOVA table in Problem 7.18 based on
how the independent variables are set. Show the proper degrees of freedom.
7.20 For the significant (a = 0.05) effects in Problem 7.18 complete the ANOVA
breakdown suggested in Problem 7.19.
7.21 Graph any significant effects found in Problem 7.20 and discuss.
 Problems 163

7.22 Data on the effect of knife edge radius R in inches and feedroll force F in pounds
per inch on the energy necessary to cut 1-inch lengths of alfalfa are given as follows:

Feedroll Knife-Edge Radius (inches)

Force Fj
(Ib/m) 0.000 0.005 0.010 0.015 T,.

5 29 98 44 84
30 128 81 100
20 67 77 63
79 293 202 247 821

10 22 35 53 103
26 80 93 90
16 29 59 98
64 144 205 291 704

15 18 49 58 80
17 68 103 91
11 61 128 77
46 178 289 248 761

20 38 68 87 86
31 74 116 113
21 47 90 81
90 189 293 280 852

T,.. 279 804 989 1066 T = 3138

Write the mathematical model for this experiment and do an ANOVA on the
two factors and their interaction.
7.23 Outline a further ANOVA based on Problem 7.22 with single-degree-of-freedom
terms.
7.24 Do an analysis for each of the terms in Problem 7.23.
7.25 Plot some results of Problem 7.22 and try to argue that they confirm some results
found in Problem 7.24.
 8
Multiple Regression

8.1 INTRODUCTION
In Chapter 7 we considered a prediction model to predict some dependent
variable Y from one independent variable X where the model may be some
polynomial in X. We also considered examples of two or more independent,
quantitative variables for cases where the levels of each such variable were
set at equispaced values. We now consider a more general situation in which
Y may be a function of several independent variables X2,..., with
no restrictions on the settings of these k independent variables. In fact, in
most such multiple regression situations the X variables have already acted
and we simply record their values along with the dependent variable Y. This
is, of course, ex-post-facto research as opposed to experimental research
where one manipulates the X’s and observes the effect on Y.
In practice there are many studies of this type. For example, one may
wish to predict the surface finish of steel from drop-out temperature and
back-zone temperature. Here 7 is a function of two recorded temperatures
X^ and X2. It may be of interest to predict college grade-point average
(GPA) in the freshman year for students whose input data include rank in
high school, high school regents’ average, SAT (Scholastic Aptitude Test)
verbal score, and SAT mathematics score. Here T, the freshman year GPA,
is to be predicted from four independent variables: X^—high school rank,
X2—high school regents’ average, X3—SAT verbal, and X^—SAT mathe-
matical.
As in the previous chapter, a mathematical model is written and the
coefficients in the model are determined from the observed sample data by
the method of least squares, making the sum of squares of deviations from
this model a minimum.

164
 8.2 The multiple regression model 165

8.2 THE MULTIPLE REGRESSION MODEL

To predict the value of a dependent variable (7) from its regression on several
independent variables (X^, X2,..., X^) the linear population model is
given as
r = /So + + ■■■ + /SA
where the ^’s are the true coefficients to be used to weight the observed X’s.
In practice a sample of N is chosen and all variables (7, X2,..., X;^)
are recorded for each of the N items in the sample and the corresponding
sample model is

Y' = Bo + + B2X2 + • • + B,X, (8.1)

where the B's are estimators of the j5’s. (Capital B's are used here because
computer programs do not usually provide for lowercase letters.) Note that
if /c = 1, Equation (8.1) gives

Y' = Bo Y B,X,

or the straight-line model of Chapter 7.

Since each observed 7 can be expressed as

7, = 7; + i= 12,..., N

by making the sum of the squares of the errors of estimate (^ ef) a minimum,
one can derive the following least squares equations.

y y = BoW + I -^1 + 1 ^2 + • ■ ■ + B. z
y x, y = Bo I + Bi I + B y
2 + ... + y x,x,
y x,y = Bo y X, + Bi y + B, y + ■ ■ ■ + B,y x^x, (8.2)
y x,Y = Bo y X* + Bi y x,x, + B, y + ■ ■ ■ + B, y 2^,^
This set of /c + 1 equations in /c + 1 unknowns {Bo, B^, . . ., 5 J can be solved
for the B's. The solution is tedious if more than two or three independent
variables are involved and so these equations are usually solved by a com-
puter program that can handle a large number of independent variables.
To see how such problems are handled consider a problem from Burr
[8],
Example 8.1 Twenty-four samples of a steel alloy are checked as to their
percent elongation under stress and for each sample a chemical analysis is
made and the percentages of five specific chemical elements are recorded.
166 Multiple regression

These are identified only as X^, X , X , X , and X^. The data are given
2 3 4

in Table 8.1.

Table 8.1 Percent Elongation Data

Item No. Y -^1 ^2 ^3 ^4 ^5

1 11.3 0.50 1.3 0.4 3.4 0.010

2 10.0 0.47 1.2 0.3 3.6 0.012
3 9.8 0.48 3.1 0.7 4.3 0.000
4 8.8 0.54 2.6 0.7 4.0 0.022
5 7.8 0.45 2.8 0.7 4.2 0.000
6 lA 0.41 3.2 0.7 4.7 0.000
7 6.7 0.62 3.0 0.6 4.7 0.026
8 6.3 0.53 4.1 0.9 4.6 0.035
9 6.3 0.57 3.7 0.8 4.6 0.000
10 6.3 0.67 2.7 0.6 4.8 0.013
11 6.0 0.54 3.1 0.7 4.2 0.000
12 6.0 0.42 3.1 0.7 4.4 0.000
13 5.8 0.33 2.6 0.6 4.7 0.008
14 5.5 0.51 3.9 0.9 4.4 0.000
15 5.5 0.54 3.1 0.7 4.2 0.000
16 4.7 0.48 4.0 1.1 3.7 0.024
17 4.1 0.38 3.3 0.8 4.1 0.000
18 4.1 0.39 3.2 0.7 4.6 0.016
19 3.9 0.60 2.9 0.7 4.3 0.025
20 3.5 0.54 3.2 0.7 4.9 0.022
21 3.1 0.33 2.9 2.9 1.0 0.063
22 1.6 0.40 3.2 3.2 1.0 0.059
23 1.1 0.64 2.5 0.7 3.8 0.018
24 0.6 0.34 5.0 1.3 3.9 0.044

To enter Equations (8.2) the following sums are computed from Table
8.1.
N = 24
yx, = 11.68 y X? = 5.9038 y XI262 = 35.553
1^2 = 73.7 'ZXl = 240.85 y 261X3 = 10.101
1363 = 22.1 yxl = 31.17 y X1X4 = 47.644
y 264 = 96.1 y = 408.09 y X1X5 = 0.18405
y X5 = 0.397 Y^Xl = 0.0145 y XiY = 67.318
y y = 136.2 ^ = 943.78 y X2X3 = 71.17
y ^2X4 = 297.57 y X3X4 = 74.58 y X4X5 = 1.2881
y XjX; = 1.2886 y 2('32C5 = 0.5945 X ^47 = 563.35
y Xjy = 390.32 y XjF = 102.53 y XjY = 1.6384
 8.2 The multiple regression model 167

and the equations to be solved would be

136.2 = 24Bo + 11.68J5i + 73.7^2 + 22.1^3 + 96.1^4 + 03913^
67.318 = 11.68^0 + 5.9038^1 + 35.553^2 + IO.IOIJB3 + 47.644^4 + 0.18405^5
390.32 = 13JBo + 35.5535i + 240.85^2 + 71.17^3 + 291.513^ + I.2886B5
102.53 = 22.IB0 + lO.lOl^i + 71.17^2 + 31.17^3 + 74.58^4 + 0.5945^5
1.2881 = 96.1Bo + 47.644^1 + 297.57^2 + 74.58^3 + 408.09^4 + 1.2881^5
1.6384 - 0.397^0 + 0.184055i + 1.2886^2 + 0.5945^3 + 1.2881^4 + 0.0145^5
Many computer programs are available to solve such a set of equations.
An SPSS BACKWARD solution will give the B values for all variables
in the equation as follows (to two decimal places):

r = 14.92 - 1.58Xi - 1.28X2 - 1.69X3 - 0.58X4 - 41.14X5

This prediction equation could be used to predict elongation Y from the
five variables X^ through X5.

The SPSS program also gives the following printout.

MULTIPLE R 0.71634
R SQUARE 0.51314
STD DEVIATION 2.14964

ANOVA df SS MS F Significanee

Regression 5 87.66795 17.53359 3.7943 0.016

Residual 18 83.17705 4.62045

The value of R SQUARE indicates that 51.314 percent of the variation (SS)
in Y is accounted for by the above equation. Note: Total SS = 170.845 and
87.66795/170.845 = 0.51314. The F statistic is significant at the 0.016 level,
indicating the equation accounts for a significant part of the total variation.
The STD DEVIATION reported is really the standard error of estimate—
the standard deviation of the errors remaining after the prediction equation
has been “fitted” to the data. The standard deviation of Y prior to prediction
is 2.73:
I - (Z ^943.78
(136.2)2
N 2^
2.73
N 23
If R2 — 0.51314 of the variation in Y is accounted for by the regression of
Y on Xi, X2, . . ., X5, then 1 — = 0.48686 is unaccounted for or residual.
So the SS of the error or residual should be
SS error SSy(l 12345)

170.845(0.48686)
83.177
168 Multiple regression

and
SS error '83.177
12345 2.15
N - k - 1 18
as given.
Sometimes the multiple R is designated as 12345 to show the joint

effect of all five X's on Y. In like manner, Sy 12345 denotes the standard
error of estimate in Y from the regression model based on five X's.
The question might now be raised as to whether or not one needs all
five independent variables to predict Y. If a Stepwise Regression program is
used, the computer will enter the X's in the order of their contribution to
the prediction of Y. At each step in the process the input can be examined
to see whether or not adding another variable will contribute significantly to
the prediction of Y. If not, one may end up with a reduced regression equa-
tion based on a subset of the X’s that will be adequate in predicting Y.
Our data were entered into an SPSS Stepwise Regression as shown in
Table 8.2.
The printout will be considered in several stages to see what one can
determine from the computer. The first part of the printout gives the mean
and standard deviation of each variable.

Variable Mean Standard Deviation Cases

Y 5.6750 2.7254 24
X, .4867 .0977 24
X2 3.0708 .7948 24
A3
.9208 .6859 24
A4
4.0042 1.0063 24
As .0165 .0186 24

Note that the standard deviation of Y is 2.73 as we computed it above. Next,

the correlations (Pearson’s r’s) between Y and each of the X variables are
given along with the intercorrelations such as between the independent
variables. This form is referred to as the intercorrelation matrix.

Ai .16884
A2 -.56054 -.17600
A3 -.53234 -.42456 .26356
A4 .28508 .38712 .13389 -.87641
As -.52747 -.21924 .20448 .78077 -.70099
Y A, A, A3 A4

In examining the correlations between Y and each X, one notes that the
highest numerical r is T with X2 {ry2 = —0.56054) and thus a Stepwise Re-
gression will enter X2 first into the regression equation and give the resulting
simple linear equation and other useful information as seen in Table 8.3.
 UJ
8.2 The multiple regression model 169

Table 8.2 Computer Input for Elongation Data

+ + + CREATE PROB4
1.000 = 77509,113
2.000 = COMMON (SPSS)
3.000 = SPSS
4.000 = #EOR
5.000 = RUN NAME STAT 502 PROB4
6.000 = VARIABLE LIST Y, XI, X2, X3, X4, X5
7.000 = INPUT FORMAT (5F4.0, F5.0)
8.000 = N OF CASES 24
9.000 = REGRESSION METHOD = STEPWISE/
10.000 = VARIABLES = Y, XI, X2, X3, X4, X5/
11.000 = REGRESSION = Y WITH XI, X2,
X3, X4, X5/
12.000 = STATISTICS ALL
13.000 = READ INPUT DATA
14.000 = 11.3 .50 1.3 .4 3.4 .010
15.000 = 10.0 .47 1.2 .3 3.6 .012
16.000 = 9.8 .48 3.1 .7 4.3 .000
17.000 = 8.8 .54 2.6 .7 4.0 .022
18.000 = 7.8 .45 2.8 .7 4.2 .000
19.000 = 7.4 .41 3.2 .7 4.7 .000
20.000 = 6.7 .62 3.0 .6 4.7 .026
21.000 = 6.3 .53 4.1 .9 4.6 .035
22.000 = 6.3 .57 3.7 .8 4.6 .000
23.000 6.3 .67 2.7 .6 4.8 .013
24.000 = 6.0 .54 3.1 .7 4.2 .000
25.000 = 6.0 .42 3.1 .7 4.4 .000
26.000 = 5.8 .33 2.6 .6 4.7 .008
27.000 = 5.5 .51 3.9 .9 4.4 .000
28.000 = 5.5 .54 3.1 .7 4.2 .000
29.000 = 4.7 .48 4.0 1.1 3.7 .024
30.000 = 4.1 .38 3.3 .8 4.1 .000
31.000 = 4.1 .39 3.2 .7 4.6 .016
32.000 = 3.9 .60 2.9 .7 4.3 .025
33.000 = 3.5 .54 3.2 .7 4.9 .022
34.000 = 3.1 .33 2.9 2.9 1.0 .063
35.000 = 1.6 .40 3.2 3.2 1.0 .059
36.000 = 1.1 .64 2.5 .7 3.8 .018
37.000 = .6 .34 5.0 1.3 3.9 .044
38.000 = FINISH

39.000 = #s
 PJ
u
z
<
u O
o
z
a
GO
>
H -H Ol
r- o
< m o
U m Tf
H o o ll-
PU H ^ —: OS
CO r-
CQ o
< 7 ' o
-J
PJ
PJ
u
z o so 00 o ITN so NO

PJ < 00
CNI
Os
so NO o o
ro
p m P
PJ DC u 00
U 00 NO OO
<
z oo 'sf ro o ON
PM NO ro
o
< o c?
Tf c5
o
o
oo
ON
NO
ITN z NO irS r-*
u ON • G^
a
CO r- <N Z o
r- ro NO m
rsi O
X z o
z •TN
►-H
H
2
o
h-^
z o 'NO
m <
PJ <
o
c3 H
< in a
PJ
u
Q a
PJ
z <N r- ON

PJ GO
PJ < O o
<N
1-^
00
c CO
PJ
X DC Os
SO
o
00
_o cC
PJ
DC DC
H PJ OS ON ON ON
-I—> w X O
Tf OO
m < z MJ

CQ H NO 1-^ 1:3 os O
OD DC OO H
c z DC
<N
<r)
ON
r-
T—<
O ON
r^
t/N
NO
H
PJ O a^
in
NO 1—1 O
W z CO
PJ Q
NO PM rO
T—<
z -J
CO ON UTi <N o
c Q. Tt O H O
< o 00 ro
W IT) (N O PQ CO PJ
o H
O Tf m
< MJ H
NO
OO 00
oo
m
OS
O
G in
>0 --I
in m
OO O
<N ro D
ca DC o m

_o fN
5 CO < <
t/D Z < ON
r^i
oo
oo 2 CP
t/D 0 > <
D r~
VH Q PJ DO
r^l
(N
r-
m H >
(=0 tu os U PJ
<u cC < pp CP MJ
pj Q r- CQ
D m
H
az <N o < Tf
X
in
Z c/0 O (N ■It
2 X X
OH UP
W 0£ Osi P H
<
>
H PJ
(>0 PJ w PJ Q j z PJ
>
HJ Oi PJ > oa < U H
oa
< PS|S
< H
CO
z
<
G
•-H 2
<
z CQ
o 01
< > <>l o 2
< <
u
> S Di < GO X > 2
<
PM

o z >
GO
o -J PM

cn in
CO
CO
< o
00 > PJ pp
MJ DC Q
-2
2i < O 2
pp

S
Z PJ PJ O
H < DC DC U

170
 8.2 The multiple regression model 171

The “Variables in the Equation” portion of this printout gives the B

values (here B2), the coefficient of X2, and BQ, the constant term. The equa-
tion at this step is then
Y' = 11.577488 - 1.9221129X2
or
Y' = 11.58 - 1.92X2

as the best fitting straight line for predicting Y from X2 alone. With each
B is recorded its standard error and a test as to whether it is significantly
greater than zero. Here both BQ and B2 are highly significant (probabilities
of 0.000 and 0.004 respectively).
This printout also shows an ANOVA for the single X’s effect of Y and
the multiple R and which, in this first step, is the simple Pearson r of
y versus X2. Thus this equation will account for 0.3142 of the variation in
y. Note from the ANOVA that the linear regression on X2 with 1 df has
an SS of 53.68 out of a total (53.68 + 117.17) of 170.85, or 53.68/170.85 -
0.3142. The STD DEVIATION is actually the standard error of estimate or
standard deviation of the errors of estimate from this first equation. Here
Sy 2 = 2.31, which is some reduction from the original Sy of 2.73.
The “Variables Not in the Equation” portion of the printout records
partial correlation coefficients. These are correlations between Y and each
X after X2 has been removed. They are sometimes designated as ry^ 2? and
so forth, indicating the correlation between y and X^ with X2 removed.
These four partials are now examined to see whether any are significantly
greater than zero. Here three of the four are (at the 0.05 level) and the largest
partial is Vys.j = —0.50903, so X5 is next entered into the regression equa-
tion as STEP 2 as printed out in Table 8.4.
The resulting equation is now
Y' = 11.69 - 1.62X2 - 63.I8X5
At this stage Ry 25 = 0.49209 and the standard error of estimate is 2.03. An
examination of the next set of partial r’s

^yl.25 = -0.01274
^y3.25 — —0.16275
r,4.25 = 0.09925

shows none significant at the 0.05 significance level. Hence one may stop in
this problem after X2 and X5 have been entered and there is no added
significance in going to an equation with the other three variables included.
One might also note that the SS for the error after X2 and X5 have
been accounted for is 86.77 whereas it is only 83.18 when all three X’s are
included. One might also note that the standard error of estimate appears to
 UJ
U
z
<
p o
o

>
H o — r- NO
NO ON —'
<
H
U m

ONI
r-
NO

oo
NO
00
o —
(DO r-
oo
UJ H ro r-
CQ GO
<
-J
UJ UJ
U 04
z OI
UJ
UJ < NO r- ON o
UJ DC u rO lO O NO ON NO
u ON ON ON 00 NO
< UH wn OO
Z r- NO ro ro O P ON ^ 1— ON
< C^ o to 1—H NO
OO o m P a
m
iT)
—
(N z OJ
OO
ON
"TD
Z
o
u
<N
00 O
ON
CO m
p
m
— Z g ro
U. to oo (N Tf g
X p ro z
H
< z od r-' <
UJ
H
<
o
■<c—<3
P p UJ
a CO o
UJ U
(N UJ
z
Q UJ
GO
UJ <
00
ON ^
ON Tt
c X CQ UJ X DC ro r-
500
(N
o CC H
DC DC H UJ ON

03
UJ O
ON
o
to
<N
1—H
<N < J
o
CJ) ca DC NO rn p
GO ON to ON r- (N H
c DC r- C^ ON a o ^ H
UJ (N 00 CO r-
-J UJ p O
p r-
s Z CQ Q
rn
(N u. Tit \6 z
c cu ON ON (N
< H O 00 00 CO 'Ct r- (N
o UJ o r-- r-"
ro <N 2
GO UJ
HJ H
04
>0
ON
ON
c
H
GO O
r- ■ TT O
ON ro < P CQ DC P o
o H
> GO < <
• Z oo Ot NO l-H f
C/D tn NO (N OH

(/) o ON 00 DC
<u ON ON <
ON r- ON
&0
Q UJ oa NO H >
UJ CC p u UJ
<u Qi
< NO 0- P
0:5 UJ m
Q
<N H P oq m Tt
OH
Z
UJ
az
00 O
CNI
CNl m
—GO

QC
P 'X
X X

W DC DC H f-H <
>
\
H UJ
CO UJ UJ UJ Q < J z UJ
p DC UJ > < >
m U
CQ p < H UJ < H H
CO CO Z
< H P Q
2 <
3 J o P Z
Qi < oa
o < p CO Q Q
> CNI tn o DC
<
H u <
> DC < GO X X > 2
iH
UM <
OH
O >
CO -J UH
h-N CO
CO
< o
00 >- UJ p u.
J DC Q UH

2 < o ^ S
es z UJ UJ O
CtC
H < I U

111
 8.2 The multiple regression model 173

increase from the two-variable equation to the five-variable equation. This

is simply due to the fact that we are dividing by a smaller number of degrees
of freedom and using three nonsignificant variables.
Even though only X and 2 are needed to predict Y quite adequately,
the stepwise program will continue and add X , X^, and finally 3 unless
programmed to stop and give the equation shown above as the BACK-
WARD solution.
To see that this is a reasonable prediction equation we might insert
X and A values into it and examine the errors of estimate, for example, if
2 5

^2 = 1.3 and X = 0.010,

5

Y' = 11.69 - 1.62(1.3) - 63.18(.010)

= 8.957 versus an observed Y = 11.3
giving an error of estimate 7 — T' = 11.3 — 8.95 = 2.343. For all 24 points
we find as in Table 8.5. Here X ^ ^ ~ 0.003 (close to zero) and = 86.77
as given in the ANOVA table after X and X^ are accounted for.
2

One might also note in this table that the residuals or errors of estimate
seem to be nonrandom as in general they are positive for large T’s and
negative for small T’s. This may indicate a need to consider some higher
order terms such as X\, X\, or X X^. These could be entered into the
2

equation by proper instructions at the start of a program. Some computer

programs will give a plot of these residuals when requested.
If one wishes only to check the significance of a given the proper
test is
ryl
^y.l2. .k, 'k
F k, N-k-l (8.3)
1 - RyA2. 'N - k - \

Table 8.5 Deviations from Regression on Elongation Data

T Y' e T Y' e

11.3 8.957 2.343 5.8 6.977 -1.177

10.0 8.993 1.007 5.5 5.377 + 0.123
9.8 6.673 3.127 5.5 6.673 -1.173
8.8 6.093 2.707 4.7 3.699 1.001
7.8 7.159 0.641 4.1 6.349 -2.149
7.4 6.511 0.889 4.1 5.500 -1.400
6.7 5.192 1.508 3.9 5.417 -1.517
6.3 2.842 3.458 3.5 5.121 -1.621
6.3 5.701 0.599 3.1 3.017 0.0831
6.3 6.500 -0.200 1.6 2.783 -1.183
6.0 6.673 -0.673 1.1 6.508 -5.408
6.0 6.673 -0.673 0.6 0.815 -0.215
 S u
O'dN'^I^On
< S (N oo ^
r- m
D 5
Q < _H ^ o O
ro O O O

a: ^

tu
csi
< O
(N o
<dN
’-^
Tj- (N O
—H o^ O —I
w a m ^ CO
CO
HJ
OQ QC
i <
H

> Qi
CeS
W
< J ^ 0^ —I Tf
D- CO O CD r<o
h-H O ^ CO
H CO O —' —'
CO 1^
D J
I CO D

w
u
z
< ro oc Oc oo
u
1—1 O ^ cO CO CO
UH
p p cq
h-^
Z
o
HH
CO

UJ
>
O

yj —H CO O CXD
OO ON
O Oc O oc
H [V
CO Tj- o oo
p ro CO (N p
UH O O K ' ' ’
03
0<
w
Q
H
c
_o Z
■t—< w
o3
tJD
d
o Q
s w

a>
3
al
o3
H <

5 Q
c3 < w
a > Cii fN| CO ro
tu
a X z z z z
d H
CO Z
yj
'^O
00
OH
w Cxi OO CO
«s H
H CO

174
 8.3 Computer programs 175

To test whether any more variables are needed after a subset r has been
found significant the test is

Rl - Rf/k - r
(8.4)
1 - Ri/N - fc - 1
where represents the correlation of all k values with Y and the correla-
tion of a subset r of the k variables with Y.
In our example = 0.51314 and R^ = 0.49209. Testing for signifi-
cance of all five variables:

0.51314/5
3.79
0.48686/18
as in ANOVA. Testing whether one needs to go beyond X2 and X^,

_ 0.51314 - 0.49203/3
“ 0.48686/18 ^
hence nonsignificant.
Many computer programs give a SUMMARY TABLE at the end of
their printout, which shows the variables entered in order of their entry and
the multiple R and R^ increase as more variables added will account for more
and more of the variability in the Y variable. In this example Table 8.6 gives
this information. One notes that after adding X5, 49 percent of the variation
in Y is accounted for and after that additional variables increase this only
slightly. The increase can be tested for significance at any stage by Equation
(8.4).

8.3 COMPUTER PROGRAMS

Many computer programs are available to solve multiple regression prob-
lems. However, there is no way to be sure that a given reduced equation is
the best predictor of a given Y unless all possible regressions are considered.
This would mean, in our example, Y as a function of Y as a function of
X2, and so forth. Then 7 as a function of the X’s taken two at a time, three
at a time, four at a time, and finally with all five X’s. Some programs will
handle all possible regressions. In our example the number of these is 2^ — 1
or 2^ — 1 = 31 possible equations.
We have used SPSS programs in our example. There are three such
programs available. The BACKWARD method enters all k independent
variables at first as seen above. Then it removes any variables that are not
making a significant contribution to the prediction equation. One can
designate the level of significance desired to remove such variables by giving
the desired or smaller F to be removed. The procedure continues until no
176 Multiple regression

more variables can be removed and all those left are contributing significantly
to the prediction equation.
A FORWARD method will enter the variables one at a time until the
equation is found such that entering any more variables will not improve
the prediction significantly. Again, one may designate the size of F necessary
for addition to the equation.
A STEPWISE method is similar to the FORWARD method except that
at each stage all variables already entered are examined to see whether they
are all needed after entering the last variable. In some cases an earlier entry
is dropped. For example, if X^, and X7 were already selected in that
order, when X2, the next most significant variable, is entered X^ might be
removed. This may be unnecessary, now that X2 has been added, and the
equation will include X4, X7, and X2 only.
In all these methods default values are in the programs that will operate
in adding or removing variables if one does not instruct the computer as to
the level desired. It is also possible to include certain specified X’s in the
equation regardless of their significance.
In the problems at the end of this chapter it is assumed that some
multiple regression program is available to the reader.

8.4 REMARKS

In examining a multiple regression equation one must be careful not to

assume that because a B value is numerically large the corresponding X has
a greater influence in predicting Y than an X with a smaller regression
coefficient. These regression coefficients (B’s) are all based on their respective
means and standard deviations as are the X’s and Y. If one desires an equation
whose coefficients do reflect the relative effect of the associated variable, the
basic measurements (7, X2,..., X;,) can be converted to standardized
values or Z values where

and Zj =

If this is done, the resulting equation for predicting the standardized 7 can
be shown to be given by

z' = B\z, + F2Z2 + • + 5;z,

and the least squares equations become

ryi — B[ + ^2^12 + ■ ■ ' + B'j^r

^Y2 = ^1^21 + ^'2 + ■ ■ ' + B'k^lk
> (8.5)

^Yk ^ ^i^ki + ^'i^ki +■■■ +

 Problems 177

Here

= or =
Sy Sj

and the rs are simple Pearson r’s. Solving these equations for the B'^s is still
difficult for very many independent variables. In the special case where /c = 1,
i3 y
B\ = Vyi and B^ = — •
(8.6)
and
Bo=Y B,X,

In Example 8.1 the first variable entered was Xj and the preliminary data
gave the following.

Y = 5.6750 Sy = 2.7254
Xj = 3.0708 S2 = 0.7948
and
^Y2 - -0.56054
Hence
Sy 2.7254
(-0.56054) = -1.92
^ S2 0.7948
and
Bo = Y - B2X2 = 5.6750 - (-1.92)(3.0708) = 11.58

giving Y' = 11.58 — 1.922^2 as before.

A word of caution seems necessary also when one plans a regression
study. Often experimenters seem inclined to include every X value that they
can think of as computers can deal with so many X’s. However, it is difficult
to believe that all X's would be deemed important in their effect on Y by
those knowledgeable in the field. Thus the use of multiple regression should
not substitute for careful “soaking” in a problem and searching out of past
studies and theoretical considerations before including a given X in a
multiple regression study. Sometimes simple scattergrams of Y versus a
given X will give some idea as to whether they may be related and the scatter-
gram may also indicate an association that is nonlinear. This latter may
indicate a need for a term such as X^. The literature in the area may also
suggest that some X is logarithmically related to Y and one should use
log X as the independent variable instead of X. Careful consideration of all
these notions before the experiment is essential to good experimentation
and increases the prospect of meaningful prediction.

PROBLEMS

8.1 The sales volume E of a product in thousands of dollars, price Xj per unit in
dollars, and advertising expense X2 in hundreds of dollars were recorded for n =
178 Multiple regression

8 cases. The results are as follows:

Case

1 2 3 4 5 6 7 8

Y 10.1 6.5 5.1 11.0 9.9 14.7 4.8 12.2

X, 1.3 1.9 1.7 1.5 1.6 1.2 1.6 1.4
X2 8.8 7.1 5.5 13.8 18.5 9.8 6.4 10.2

a. Find the means and standard deviations of each variable and the intercorre-
lation matrix.
b. On the basis of the discussion in Section 8.4 find the regression equation Y
from XI alone.
8.2 Submit the data of Problem 8.1 to a Stepwise Regression program and find the
regression equation and the value of Ry 12 decide whether or not both X'’s
are necessary.
8.3 In studying the percent conversion (7) in a chemical process as a function of time
in hours and average fusion temperature in degrees Celsius {X2) the following
data were collected.

Y ^2

62.7 3 297.5
76.2 3 322.5
80.8 3 347.5
80.8 6 297.5
89.2 6 322.5
78.6 6 347.5
90.1 9 297.5
88.0 9 322.5
76.1 9 347.5

Analyze these data using multiple linear regression. Reexamine these data and
comment on any pecularities noted.
8.4 A management consulting firm attempted to predict annual salary of executives
from the executives’ years of experience (^1), years of education (^2), sex (X3), and
number of employees supervised (^^4). A sample of 25 executives gave an average
salary of $29,700 with a standard deviation of $1300. From a computer program
the following statistics were recorded.

= 0.42

RL4 = 0.78
R? 124 = 0.90
^r.i234 = 0.95
 Problems 179

a. Explain how you would interpret these statistics.

b. Test whether one could stop after variables 2 and 4 have been entered into the
equation.
c. If variables 1, 2, and 4 were used in the prediction equation, what would the
limits on the salaries expected for a predicted salary have to be in order to be
correct about 95 percent of the time?
8.5 Consider a prediction situation where some dependent variable Y is to be predicted
from four independent variables XX2, X2, and X^ and assume you have a step-
wise regression printout on this problem involving 25 observations. Explain
briefly each of the following with respect to this problem and its printout.
a. How does the computer decide which independent variable to enter first into
a regression equation?
b. How does it decide which variable to enter next in the equation?
c. How do you decide when to stop adding variables based on the printout?
d. Do you need to add any more variables if i^y.1234 = 0-7 and Rj 2A = 0-6?
8.6 A study was made to determine the effect of four variables on the grade-point
average of 55 high school students who attended a junior college after graduation.
These variables were high school rank (26 SAT score {X2), IQ (263), and age in
months (264). From the computer printout the following statistics were recorded.

^y.i234 ~ 0.45 RYA2 ~ 0.33

^2 ^23 ^ 0.42 = 0.19

Use this information to test for the significance of each added variable and deter-
mine whether or not an ordered subset of these four will make a satisfactory
prediction. If you wish, you may assume a total of sum of squares in this study of
10,000 units. Also explain how you would predict with approximate 2 to 1 odds
the maximum GPA expected for an individual student based on his or her scores
of the variables in the appropriate equation.
8.7 In a study of several variables that might affect freshman GPA (T) a sample of
55 students reported their college entrance arithmetic test (Xj), their analogies
test score on an Ohio battery (X2), their high school average (X3), and their
interest score on an interest inventory (X4). Results of this study are given in the
following (oversimplified) ANOVA table.

Source* df SS MS

Due to X2 1 1360 1360

Due to X3/X2 1 480 480
Due to XJX2, X3 1 80 80
Due to X4/X1, X2, X3 1 80 80
Error 50 2000 40

Total 54 4000

* / means “given.”
180 Multiple regression

Assuming a stepwise procedure was used so that variables were entered in their
order of importance, answer the following.
a. Determine RY.I234 explain its meaning in this problem.
b. Determine which of the four variables are sufficient to predict GPA (F) and
show that your choice is sufficient.
c. If a regression equation based on your answer to (b) is used to predict student
A’s GPA based on test scores, how close do you think this prediction will be
to the actual GPA? Justify your answer and note any assumptions you are
making.

8.8 In an attempt to predict the average value per acre of farm land in Iowa in 1920
the following data were collected.
Y = average value in dollars per acre
Xi = average corn yield in bushels per acre for ten preceding years
Xi = percentage of farm land in small grain
X3 = percentage of farm land in corn

Y ^2 -Y3

87 40 11 14
133 36 13 30
174 34 19 30
385 41 33 39
363 39 25 33
274 42 23 34
235 40 22 37
104 31 9 20
141 36 13 27
208 34 17 40
115 30 18 19
271 40 23 31
163 37 14 25
193 41 13 28
203 38 24 31
279 38 31 35
179 24 16 26
244 45 19 34
165 34 20 30
257 40 30 38
252 41 22 35
280 42 21 41
167 35 16 23
168 33 18 24
115 36 18 21

Analyze these data using multiple linear regression and comment on the results.
 Problems 181

8.9 In a study involving handicapped students an attempt was made to predict GPA
from two demographic variables, sex and ethnicity, and two independent measured
variables, interview seore and contact hours used in counseling and/or tutoring.
A dummy variable for sex was taken as 1 = male, 2 = female. For ethnicity:
1 = black, 2 = Hispanic, and 3 = white, non-Hispanic. Do a complete analysis
of the data below and justify the regression equation that will adequately “fit”
the data.

Students Ethnic Sex Interview Hours GPA

1 1 2 11.0 4.0 5.50

2 1 2 10.0 5.0 4.10
3 1 2 12.0 73.0 5.00
4 1 2 11.5 68.0 4.22
5 1 2 10.8 82.0 5.00
6 1 1 12.5 72.5 5.00
7 1 1 9.5 64.0 4.60
8 1 1 9.5 78.0 4.25
9 1 1 8.0 64.0 4.00
10 1 1 7.5 13.0 2.00

11 2 2 9.0 37.0 4.25

12 2 2 8.2 4.0 4.00
13 2 2 10.7 38.5 4.61
14 2 2 8.5 3.0 2.93
15 2 2 12.5 10.5 5.50
16 2 1 12.0 80.0 4.77
17 2 1 12.2 6.0 5.00
18 2 1 7.0 6.5 3.25
19 2 1 8.6 22.0 2.66

20 2 1 8.3 28.5 3.37

21 3 2 10.9 12.0 5.00
22 3 2 9.0 9.0 4.00
23 3 2 10.0 5.0 5.00
24 3 2 7.2 12.0 3.87
25 3 2 8.5 4.0 3.00
26 3 1 10.0 8.0 4.77
27 3 1 8.5 8.0 5.00
28 3 1 10.0 22.0 5.08
29 3 1 11.4 61.5 5.57
30 3 1 11.9 37.0 6.00

8.10 Given the following intereorrelation matrix between measures where Y = fresh-
man mathematics grade, = Ohio State psychological examination, X2 =
English-usage examination, A = algebra examination, A = engineering apti-
3 4

tude test:
182 Multiple regression

Variable Y ^2 ^3 ^4

0.51
^2 0.51 0.70
^3 0.61 0.53 0.61
^4 0.39 0.39 0.29 0.28
Means 5.70 4.10 5.44 5.37 4.95
Standard deviation 2.42 1.92 1.84 2.26 2.14

Do a complete multiple regression on these data and summarize your conclusions.

Assume that N = 100 students.
8.11 In a study of the effect of 15 possible variables on the current annual gift of a
corporation to a college {X^) the variables are

^6 = number of students enrolled in current year

= number of families represented
^8 = number of years gifts have been given
^9 = average previous year’s gift
^10 = related company gift giving
^12 = school ranking score
^13 = previous industry average
^14 = last annual gift
^15 = continuity of giving
400 = initialized year code

401 = times Xg or X^^Xg

402 = ^15^9

X 403 = ^9^10

X 404 = ^9^12

405 = ^10^13

This means, standard deviations, and intercorrelations are given on p. 183. (Note
that Y is X^.)
a. On the basis of the above printout what is the hrst variable to be entered in a
stepwise regression? Explain how you know.
b. On the basis of just the information given hnd the best fitting prediction
equation for predicting ^5 from this first variable.
c. How good is this prediction?
d. What is the standard error of estimate?
8.12 At step 4 in Problem 8.11 the printout is as appears on p. 184.
Explain what the next step should be and discuss the results up to this point.
That is, what is the equation and how good is it at step 4?
 ON <N ON <r) 1-H
o NO (N O
OO ON
1'-' 00 rn r- {y^
m cn m ro <;
CASES

oooooooooooooooo >
rNi<NrNirNirsi<Nr^(N
o T-H m T-H o
NO i-H o
OO 00
m 1-H 1-H (N
CT'i o O o
<
STANDARD DEV

1 1 >'

00 so o NO* r- oo o o O 80 m
T—1
00 OO so so ON OS NO
m so oo ro oo ON OO CNI O
r- r- p p (N r- oo 1-H 00 (N (N
NO p (N (N (N
<
r-‘ rn oo os od
O O o >'
o(N
CN so r- O
NO •,..1 r-
OO OO r- OS oo 1
r-4 o •rs o o (N o
o rn 1-H ro ^-H ro 0^
ro o ON OO ON ro
Q <
PJ >
so Os vzs H
i—H SO (N ro •rs OO to to 00 ON ro NO ro
MEAN

rn OO OS so so OS ro NO ro 00 CD ON t-H t-H
OS p SO p so OO (N OH o o 00 NO ON 80 ON r-' O
NO 0^4 r- r- r- (N ivl
O rxi (N so o S ro o 04 ro ro oo
r--
oo
SO ro
Tf 1-^ 00
t—H so (N
o 1
<
>
u
(N m
w r- r- NO ON NO NO NO OO to (N
CQ oo t-H ro r- ON r- ON ■^H
r- Tt to ^H to ro O
H rg NO to 04 00 ro t-H 80 NO fy^
VARIABLE

rN o ro
•ri' uri o (N ro
o o o o o o
o CM to 04
<r'
O O ■rf- ■Tf z
z
< O ro OO ro ro O ro
< < < < < < < < o OS 8 00 oo SO o ro SO 00
> > > > > > > > u NO o NO ro ON SO ON r- 0
Ovj oo ON ON to OO NO o (N to 01
H CM ro (N ro ro r--
z <
>
CASES

PJ
oo OO 00 00 OO oo OO OO
<N (N CNI CN^ (N (N (N u OO 80 r- 00 80 to 80
Tf SO ro o ro ro SO 80 o
UH os to r- so OS to r- r- oo 0
Lin NO (N r- Tf ON to ON oo t-H ro 01
ro os <N o ro ON 00 ON ro
PU <
o 1 >
u
STANDARD DEV

o (N r- r- ON NO ro r- ro T—H ro oo
< r- to 80 ON 80 o 04 to oo O
00 ON ON ON OO ON
ro ro 00 os os OS OO SO 0"^ to ro (N O
pp to r- 00 80 o Tf ro r- 80 OO ro SO Cxi
00 00 O SO ro ro ro (N NO ro ON ro ro ro ro
rn OO —H Tt so <
r- ON NO r- m o
NO rn NO (N rN
Q
PJ
>
NO 80
m H T-H r- ofN o <N
VAR404

(N NO 00 80 NO
.29343

ON
1-^ z r- <N (N 80 O o ON 04 ro t-H
r- oo OO —H to NO to 00 ■t-H o t-H OO
o NO oo to 1—H oo ON r- 1-H r- 00 to ro ex:
ro r- o ro r- o ro r- NO r- (N
f~ CL <
z ^ 1 >
S ^
Qs o o ro NO ro ro —-H ro to rs)
.57640
.79807

U o
VAR403

(N o r- <N oo OO -H O O ro (N oo ro r- so m ro G^. ro 00 OO o
rn ro so o fO ON so P^ O to 00 oo SO so ro ON ro to os ro 0
to to so 00 SO r- ON 80 to
MEAN

OS oo ro (N O so pp s p (N r- o 'Tf (N r- O CN so r-
01
so •rs
o6 H
oo O
CM oo
ro
oi to
S ^ ■
1
<
<N
a-.
80 m
r-
80
o
o^ >

T—^ to r- OO o so r- so i-H ON ro ro o oo NO rg
.34442
VAR402

lo
.88608
.92208

Z o o r-- ON NO
ON to Tf NO
ON
NO
to
80
80 O NO (N
oo NO (N
ro ON ro 00 NO
OO to o ON
o
o
O PJ (N ON! o 80 80 to ro O ro 1-H oo 00 O D£;
OO OO ro ON ro ON (N O ro OS r- oo ro
H ^ <
>
VARIABLE

NO OO ON o CNJ ro
-J <
o o o
rs o o PJ >• r- OO ON o ONI ro 80 o (N ro
VAR404

o so to
VAR403

yAR405

Qi Qi CC Qi Cti ai ex: ex: o o o o O O O o O

s o 8 o o o o o O o
< < < < < < < < OS < ex: ex: ex: ex: o£ ex: ex: ex: CeJ ex: ex: ex: ex: ex: ex:
> > > > > > > > o < < < < < < < < < < < < < < <
V

u > > > > > > > > > > > > > > >

183
 w o w
o
u o u
Z Z so—'^0<N^r~-Os(Nroosro OOIOO—'■^^sO-^sO'-O
< < r-^ooOsOO^O^O'/'iOOro-^rf-osoorOooOO^OsO
r''U-iroO'^'^r~'(N—i-Hr'--™soosoo>oornrsioorN|
u u so • NO • o • oo • CNJ * NO • <N
h-H [-!-( ON <N y—t NO ■»—( (N r- r-- OO Os
UH IJH r- 00 NO NO Tf NO CN y—t
HH m wo CNj NO r4
o o p p ON.
z Z z NO r-; r4 r4 r^i r^i
o O o
So H GO
<
Z)
lO a PJ
ro w O
o^ z
ro UJ m m oo NO ro o ro NO
O X < (N
OO
<N r-
T—(
rn
(N
O OO
r-j
Ov
r-
00 CM
■—
r<-l Di O 00 O
H i-H NT) (N r^i wo ro o r^i
Csl PU o o OO r- oo wo 00 o (N OO
z HJ
O
PJ OO IT) H H
oi O <N
r-- t"- 0
< o o z HJ lO)
o ON NO 04 o- CN 04
(Jr oo
O GO < o OS 1—( wo O ON ON 00
PJ 04 ro O i-H ON ro o- ro
03 CM J H ro NO NO O- ON 00 OO NO wo
OS CQ Cci O 04 o o O o o o o o o
CNl
Z < <
•< rn CP
W 01
<
> PP
-J r-- OS o Ol m G-) o n-) G-)
GO 00 m o o i-H ^H ^H o o O o o
m ro OO o o O o o o
oo < DC DC DC DC DC DC DC DC DC DC DC
0£ r^i so
0£ <c < < < < < < < < < <
< Tj- \o < > > > > > > > > > > >
IZ) oo r-
t" so
>
5/5 so oo
OS <N
(j_ '
^ m
U so
>-
H NO NO O- 00 WO ■_<
00 ro 0- O oo ON
< NO 04 r- wo o r- WO
GO y~-l ro T-^ ro ON o o-
GIFT AMOUNT-CURRENT YEAR

r
P-^ Os p NO 04 y—t ON o
P-1 (
■»— ■»—( WO oo
PQ GO r- 04 o 1 o
H <
U _1
OH PP
so
PH m pp
<N
Q U
Z ON O wo O 04 o r^) oo NO
VAR402

PP T—1 ro o 04 ro
u >- < 04 o WO P NO
H u NO i-H
z HH
J
PH ON
O
wo G^
04
ro
^—(
r-
< PH
a< O ro
Z wo O ro y^
S 3
< o z Osl ’—1 (N
<
>
HH HH
H
a
l-H
DC 00
<
4 . .

<
o z
PH
> D
GO
o HP pp
a
pp
cri
So
00 GO <C o pp
o
VARIABLE(S) ENTERED ON STEP NUMBER

> PP w PP
HP DC Q
PH
X DC o O
PH
H l-H oo ON ON O
C a So PP
z
DC NO ON 04
z PP PP O pp
l-H
04 wo
Ovl
NO
ON
o o
< DC DC U
HH OO r- 04 ro GO
VAR005

GO
Q l-H ro 04 ro no’
H o6 04 no
PP GO
^ G-l ^ O J
SO OO (N CQ
OS O O so
GO Csi —H <
HH
os Os os no O
DC
so PP
oo < NO o ON
DEPENDENT VARIABLE. .

no > NO
ON
NO NO o
04
ro
ro Os) ro 04
r- r- r- O
PP 04 NO 04 ^H
DC O) wo wo NO
r- oo
< NO
ro
00 "=:f
WO

a Z
GO O
DC DC H H
PP pp Q < pp
Z
HP DC pp HP
CP
> QQ <
H PP < so oo CM H
GO
H D Q o
o o O GO
HP a D DC o o Z
DC DC DC DC
D c/3 Q Q
H
<
> < < < < o
S DC < c/3 > > > > u

184
 p
U
o
z
< P
w o r \ r- GO r- fN o o (N GO o m CO (N r~ 0-1 ON
o U (N oo OO (N GO o oo OS o 1 4 0-1 <N T-H
u o OO (N T—H ■ r -H (N ro O (N oo GO GO T-H
[J-N GO T—H o OI • t" p ON O
z (N 0-1 oo
< o z ON
GO
GT) so
oo
rn
GS
ON
oo
o
oo
o GO
0-1
ON
m
O
oo
u O ON
H o (N t"; p P p p
l-u < CO
D
z a
a w
p
U
00
PJ z
< v-H
r-
T-H
OO
OA
OS
yn
o
ro
(N GO
0-1
o
(N
r-
X on 1■ \ OO
H O GO T-H
(N m ON (N o ON) o
UJ p oo. r" p P P oo.
m z ■] C P
O
u. r-; H H
lA o
oo
oo z P (N ON on 0-1 ON OO oo
GO < O (N ro >o O m OO oo o
HH r- t" ON O m m
PJ H o ro lo 0^ 0-1 GO r^i P
p p p p p P p P p p
p o
p (N < l' l’
r- < p
< GO
p GO S
od <N < p
a (N > p r- O (N ro o ^—1 ro 0-1
COm O p O ^—1 o O O O O
ON ON o O O O P
Z fO <, P P
p P P P p P P
< m p < < < < < < < < <
P < > > > > > > > > >
GIFT AMOUNT-CURRENT YEAR

>

Ui 00 OO
nj H ■^r'rNir''0(N'0GD<^r<-)r<l0
O CNI O Ot^OOCNroC-'^CNl^O^OO
0^ o o^ < r<^'^^00t^r~-0^0^00O'^O
< p P U OGD'^(Nun'Oror''rnir)’^r<-i
H HH
(N H ■
D D OO P 00
^ • m ,■ ON
CTl a ^ m P c o pic C I P I
ON <
p
VAR015

o p
o\ oo
m ^n
p
U
z 0-1 O r<-) O r-
oo 0-1
-H O ON o
r~- ■^H ON ON ■^H O
1 T-H T

H < 0-1 ON ro p r-- OO P o

U u '^H r- f" GO ON ^H
(N ON 0-1
6 . .

D- p o
OO o 0-1 OO p
OS o (N GO O o o-i
U- ^ (N o
H
<d
z <N ^—1 rn m ro NO
Q o
D
VARIABLE(S) ENTERED ON STEP NUMBER

a
P- p
p
O H p
gj O p p p o o
GO X
VAR005

U 00 <
00
O
m
H o P
CNI O GO
P
0-1 Csl o
>' Z W D p p OO ON rn oo ON T—H
< oi Q p < Z
I—( p 0-1 o
<N
oo
ON
ON
CNl
i-H
o
T-H
ON
< oi O 00 p P GO ON) o oo
p p 00 0-1 OO p
z < w p O < p Q
o p
o-i
p
rd
< > os: P u > p H p od rn P
p 00
DEPENDENT VARIABLE . .

< ro ON ON NO
GO 04
o O 0-1 5 r- (N ON ON O OO
GO GO 0-1 OO 00 a\ m 0-1 m rn
<N GO 0-1 ro < rn oo r-- oo T-H GO
GO
ON
(N
ON
(N
ON
r^i
os
> P
oo
oo
r<'i
t";
T-H

p
GO 0-1
0-1
T-H

p
p
o T—H P m ON OS GO
r- OO (N P r— r- (N
m

P P
P p Q
z p H
z
o p
P P P l-H p <
P
< H P H < GO OO (N ON 0-1 H
H p 00 < HH o o O o 00
P a D<
HH oi o o O o o z
P CO Q a Q > < p p P P p p
o
H P > < < < < <
u
P < GO 00 Q > > > > > >

185
186 Multiple regression

8.13 For Problem 8.11 the printout at step 6 is as appears on p. 185.

Explain the next step and analyze the results as completely as you can.
8.14 In a study on the effect of several variables on posttest achievement scores in
biology in a rural high school involving 48 students the following table was
presented. It included variables that had a significant effect on achievement at
the 10 percent level of significance.

Multiple R^
Step Variable Entered R R^ Change

1 Figural elaboration pretest 0.46587 0.21704 0.21704

2 Achievement pretest 0.83022 0.68927 0.47223
3 Piagetian pretest 0.86536 0.74885 0.05958
4 Attitudes pretest 0.90606 0.82094 0.07209

Discuss these results and test whether or not the last two variables were necessary.
8.15 To study the effect of planting rate on yield in an agricultural problem, 13 data
points were taken with the following results.

V 12 12 16 16 16 20 20
Y 130.5 129.6 142.5 140.3 143.4 144.8 144.1

X 23 23 23 23 28 28
Y 145.1 147.8 146.6 148.4 134.8 135.1

a. Plot a scattergram and note what type of curve you might expect of Y versus X.
b. Taking = X2 and X = X^, use multiple regression to find the equation
of the best fitting quadratic.
c. Compare this quadratic with the best linear fit.
8.16 The printout of a computer program yielded the following statistics.

Bo ^ 140.27 ij, = 0.0907

1.24 = 0-3285
B2 - 2.65 Rl,23 = 0.4211
^3 = -0.30 i^.v.i234 = 0.4548
^4 = -1.27 S, = 16.0
N = 45
a. Explain the meaning of the R squares given above.
b. Test and determine which regression model is adequate for predicting Y.
c. Compute the standard error of estimate for your model in (b).
d. Explain whether or not the program was a stepwise (step-up) program.
 9
y Factorial Experiments

9.1 INTRODUCTION
As 2^ factorial experiments represent an interesting special case of factorial
experimentation, so also do 3^ factorial experiments. The 3^ factorials
consider / factors at three levels; thus there are 2 df between the levels of
each of these factors. If the three levels are quantitative and equispaced, the
methods of Chapter 7 may be used to extract linear and quadratic effects
and to test these for significance. The 3^ factorials also play an important
role in more complicated design problems, which will be discussed in
subsequent chapters. For this chapter, it will be assumed that the design is a
completely randomized design and that the levels of the factors considered
are fixed levels. Such levels may be either qualitative or quantitative.

9.2 3^ FACTORIAL
If just two factors are crossed in an experiment and each of the two is set at
three levels, there are 3x3 = 9 treatment combinations. Since each factor
is at three levels, the notation of Chapter 6 will no longer suffice. There is now
a low, intermediate, and high level for each factor, which may be designated
as 0, 1, and 2. A model for this arrangement would be

^ij = + Bj -\r ABij + Gij

where i = 1, 2, 3, j = 1, 2, 3, and the error term is confounded with the AB
interaction unless there are some replications in the nine cells, in which case,

^ijk = -\- Ai -\- Bj ABij +

and k = 1, 2, . . ., u for n replications.

To introduce some notation for treatment combinations when three
levels are involved, consider the data layout in Figure 9.1. In Figure 9.1
two digits are used to describe each of the nine treatment combinations. The

187
188 3^ Factorial experiments

2r • • •
02 12 22

• • •
Factor B 01 11 21

0 • • •
00 10 20

0 1

Factor A

Figure 9.1 3^ data layout.

first digit indicates the level of factor A, and the second digit the level of
factor B. Thus 12 means A at its intermediate level and B at its highest level.
This notation can easily be extended to more factors and as many levels as
are necessary. It could have been used for 2^ factorials as 00, 10, 01, and 11,
corresponding respectively to (1), a, b, and ab. The only reason for not using
this digital notation on 2^ factorials is that so much of the literature includes
this (1), u, h, . . . notation. By proper choice of coefficients on these treatment
combinations, the linear and quadratic effects of both A and B can be
determined, as well as their interactions, such as A^ x /l^ x BQ, AQ X
Bi^, and AQ X BQ. The methods of analysis will be illustrated in a simple
hypothetical example.
Example 9.1 Suppose the responses in Table 9.1 were recorded for the
treatment combinations indicated in the upper-left-hand corner of each cell.

Table 9.1 3^ Factorial with Responses and Totals

Factor A

Factor B 0 1 2 T.i

00 10 20
0 1 -2 3 2

01 11 21
1 0 4 1 5

02 12 22
2 2 -1 2 3

T. 3 1 6 10
 9.2 3^ Factorial 189

Table 9.2 ANOVA for 3^ Factorial of Table 9.1

Source df SS MS

Ai 2 4.22 2.11
Bj 2 1.56 0.78
AB,j 4 23.11 5.77

Totals 8 28.89

Analyzing the data of Table 9.1 by the general methods of Chapter 5 gives

(10)"
SStotal — l" + O" + 2^ + • • • + 2^ = 28.89

ss. ^ ^ 4.22

,, _ 2^ + 5^ + 3^ (10)^ _ ,
SSg — ^ ^ — 1.56

SSe„or 28.89 - 4.22 - 1.56 = 23.11

and the ANOVA is given in Table 9.2.
A further breakdown of this analysis is now possible by recalling that
coefficients of — 1, 0, +1 applied to the responses at low, intermediate, and
high levels of a factor will measure its linear effect, whereas coefficients of
+ 1, —2, +1 applied to these same responses will measure the quadratic
effect of this factor. As in the case of 2^ factorials, products of coefficients
will give the proper coefficients for various interactions. This can best be
shown by Table 9.3, which indicates the coefficients for each effect to be used
with the nine treatment combinations.

Table 9.3 Coefficients for a 3^ Factorial with Quantitative Levels

Treatment Combination

Factor 00 01 02 10 11 12 20 21 22 Zcf

AL -1 -1 -1 0 0 0 +1 +1 +1 6
AQ +1 +1 +1 -2 -2 -2 +1 +1 +1 18
BL -1 0 +1 -1 0 +1 -1 0 +1 6
BQ +1 -2 +1 +1 -2 +1 +1 ~2 +1 18
ALBL +1 0 -1 0 0 0 -1 0 +1 4
ALBQ -1 +2 -1 0 0 0 +1 -2 +1 12
AQB]^ -1 0 +1 +2 0 -2 -1 0 +1 12
AQBQ +1 -2 +1 -2 +4 -2 +1 -2 +1 36

Yu 1 0 2 -2 4 -1 3 1 2
190 3^ Factorial experiments

From Table 9.3 it can be seen that Ai^ compares all highest levels of
/4(+1) with all lowest levels of A{—1). AQ compares the extreme levels with
twice the intermediate levels. Both of these effects are taken across all levels
of B. Now compares the highest versus the lowest level of B at the 0 level
of A, then at level 1 of A, then at level 2 of A, reading from left to right across
Similarly, BQ compares the extreme levels of B with twice the intermediate
level at all three levels of A. The coefficients for interaction are found by
multiplying corresponding main-effect coefficients. An examination of these
coefficients in the light of what interactions there are should make the
coefficients seem quite plausible. The sums of squares of the coefficients are
given at the right of Table 9.3.
Applying these coefficients to the responses for each treatment com-
bination gives
A^ = -1(1) - 1(0) - 1(2) + 0(-2) + 0(4)
+ 0(-l) + 1(3) + 1(1) 1(2) - 3
AQ = +1(1) + 1(0) + 1(2) - 2(-2) - 2(4)
- 2(-l) + 1(3) + 1(1) + 1(2) = 7
B^ = -1(1) + 0(0) + 1(2) - l(-2) + 0(4)
+ 1(-1) - 1(3) + 0(1) + 1(2) = 1
BQ = +1(1) - 2(0) + 1(2) + l(-2) - 2(4)
+ 1(-1) + 1(3) - 2(1) + 1(2) = -5
A^B^ = +1(1) + 0(0) - 1(2) + 0(-2) + 0(4)
+ 0(-l) - 1(3) + 0(1) + 1(2) = -2
A^BQ = -1(1) + 2(0) - 1(2) + 0(-2) + 0(4)
+ 0(-l) + 1(3) - 2(1) + 1(2) = 0
AQB^ = -1(1) + 0(0) + 1(2) + 2(-2) + 0(4)
- 2(-l) - 1(3) + 0(1) + 1(2) ^ -2
AQBQ = +1(1) - 2(0) + 1(2) - 2(-2) + 4(4)
- 2(-l) + 1(3) - 2(1) + 1(2) = 28
The corresponding sums of squares become

SS,,,, = ^ = 0.00

SSB, = ^ = 0.17

Summarizing, we obtain Table 9.4

 9.2 3^ Factorial 191

Table 9.4 ANOVA Breakdown

for 3^ Factorial

Source df SS

A, 2 4.22
AL 1 1.50
AQ 1 2.72
2 1.56
BL 1 0.17
BQ 1 1.39
AB,^ 4 23.11
ALBL 1 1.00
ALBQ 1 0.00
AQBI^ 1 0.33
AQBQ 1 21.78

Totals 8 28.89

The results of this analysis will not be tested, as there is no separate

measure of error, and the interaction effect is obviously large compared with
other effects. As the numbers used here are purely hypothetical, the purpose
has been only to show how such data can be analyzed and how the notation
can be used.
It may be instructive to examine graphically the meaning of this high
AQBQ interaction since it dwarfs all main effects and other interactions in the
example.
If the interaction is graphed using only the extreme levels of each
factor, the results are as shown in Figure 9.2. Although the lines do cross,
this is a very small interaction. Note also that the average change in response
for the lowest level of A to the highest level of A (shown by x ) is very small,
indicating a very small effect as the data show.

Figure 9.2 interaction.

192 3^ Factorial experiments

/I

Figure 9.3 A^^BQ interaction.

Plotting AI^BQ means using all three levels of B, giving Figure 9.3, which
again shows little interaction.
For AQBI^ the results are as shown in Figure 9.4, which once again shows
little interaction. This graph does indicate that the quadratic effect of A is
more pronounced than its linear effect. This is borne out by the data.

Figure 9.4 AQBI^ interaction.

Replotting Figure 9.3 with B as abscissa gives Figure 9.5, which shows
the same lack of serious interaction as in Figure 9.3, but does show a slight
quadratic bend in factor B over the linear effect.
So far no startling results have been seen. Now plot the middle level of
B on Figure 9.4 to obtain Figure 9.6. Note the way this “curve” reverses its
trend compared with the other two. This shows that not until both factors
are considered at all three of their levels does this AQBQ interaction show up.
It can also be seen by adding the middle level of A on Figure 9.5 (see Figure
9.7).
 9.2 3^ Factorial 193

Figure 9.5 AI^BQ interaction again.

BQ

Bi
Bx

Figure 9.6 AQBQ interaction.

Figure 9.7 AQBQ interaction again.

194 3^ Factorial experiments

Table 9.5 Diagonal Computations

Factor A Factor A

Factor B 0 1 2

0 1 -2 3 1 -2 3
1 0 4 1 0 4 1
2 2 -1 2 2 -1 2

Before leaving this problem, reconsider the data of Table 9.1. Add the
data by diagonals rather than by rows or columns. First consider the diag-
onals downward from left to right where the main diagonal is 1 + 4 + 2 = 7,
the next one to the right is —2 + 1 + 2= 1, and the last +2 is found by
repeating the table again to the right of the present one as in Table 9.5.
Similarly, the next downward diagonal gives 3 + 0 + (—1) = 2. The
sum of squares between these three diagonal terms is then

(If + {If + [If (lOf ^

If the diagonals are now considered downward and to the left, their
totals are
3 + 4 + 2 = 9
1 + 1 - 1 = 1
-2 + 0 + 2 = 0
and their sum of squares is

{9f + {If + {Of (10)2

These two somewhat artificial sums of squares of 6.89 and 16.22 are seen to
add up to the interaction sum of squares

6.89 + 16.22 = 23.11

These two components of interaction have no physical significance, but sim-

ply illustrate another way to extract two orthogonal components of interac-
tion. Testing each of these separately for significance has no meaning, but this
arbitrary breakdown is very useful in more complex designs. Some authors
refer to these two components as the / and J components of interaction

I{AB) - 6.89 2 df
J{AB) = 16.22 2 df
total/I X B = 23.11 4 df
 9.2 3^ Factorial 195

Each such component carries 2 df. These are sometimes referred to as the
AB^ and AB components of A x B interaction. In this notation, effects can
be multiplied together using a modulus of 3, since this is a 3^ factorial. A
modulus of 3 means that the resultant number is equal to the remainder when
the number in the usual base of 10 is divided by 3. Thus 4 = 1 in modulus 3,
as 1 is the remainder when 4 is divided by 3. The following associations also
hold:
numbers: 0 3 = 0 6 = 0 9 = 0
1 4=1 7 = 1 10 = 1
2 5 = 2 8 = 2 11 =2

When using the form A^B"^, it is postulated that the only exponent allowed
on the first letter in the expression is a 1. To make it a 1, the expression can
be squared and reduced, modulus 3. For example,
A^B = {A^Bf = A^B^ = AB^
Hence AB and AB^ are the only components of the A x B interaction with
2 df each. Here the two types of notation are related as follows:
I{AB) = AB^
J{AB) = AB
To summarize this simple experiment, all effects can be expressed with
2 df each, as in Table 9.6.

Table 9.6 3^ Factorial by 2-df Analysis

Source df SS

At 2 4.22
Bj 2 1.56
1{AB) = AB^ 2 6.89
J{AB) = AB 2 16.22

Totals 8 28.89

It will be found very useful to break such an experiment down into 2-df
effects when more complex designs are considered. Here, this breakdown is
presented merely to show another way to partition the interaction effect.
Let us consider a practical example.
Example 9.2 The effect of two factors—prechamber volume ratio {V) and
injection timing (T)—on the parts per million of noxious gas emitted from
an engine was to be studied. The volume ratio was set at three equispaced
196 3^ Factorial experiments

Table 9.7 Gas Emission Data of

Example 9.2

Volume {V)
Time
(T) Low Medium High

Short 6.27 8.08 7.34

5.43 8.04 7.87
Medium 6.94 7.48 8.61
6.51 7.52 8.32
Long 7.22 8.65 9.02
7.05 8.97 9.07

levels and the time was also set at three equispaced levels. Two engines
were built for each of the nine treatment combinations and the amount of
gas emitted was recorded for the 18 engines as shown in Table 9.7.
The ANOVA for Table 9.7, assuming a completely randomized design,
is given in Table 9.8.

Table 9.8 ANOVA for Gas Emission Data

Source df SS MS F Prob.

Volume {Vi) 2 11.44 5.72 71.5 0.000

Time [Tj) 2 4.17 2.08 26.0 0.000
V X T interaction (ETjj) 4 1.39 0.35 4.4 0.031
Error 9 0.70 0.08

The ANOVA of Table 9.8 shows strong main effects and an interaction
significant at the 5 percent significance level. As both factors are quantitative
and equispaced we can treat these data as we did the hypothetical data of
Table 9.3 to give Table 9.9.
Upon taking each contrast and squaring and dividing by 2 (as n = 2)
and the sum of the squares of the coefficients, one gets Table 9.10.
From Table 9.10 we see that volume ratio produced both a strong
linear and quadratic effect on the amount of gas emitted. The timing shows
a strong linear effect and a significant interaction is shown as a quadratic
by quadratic interaction. One must then plot all nine treatment means to
see the significant effects as shown in Figure 9.8.
Note that without the T2 line in Figure 9.8 there would be little obvious
interaction. Careful study of this illustration should confirm the reason-
ableness of the results given in Table 9.10.
 9.2 3^ Factorial 197

Table 9.9 Quantitative Level Breakdown of Example 9.2

Factor 00 01 02 10 11 12 20 21 22 Contrast

VL -1 -1 -1 0 0 0 +1 4-1 -41 10.81

+1 +1 +1 -2 -2 -2 +1 +1 4-1 -7.83
TL -1 0 +1 -1 0 +1 -1 0 +1 6.95
Ta +1 -2 +1 +1 -2 +1 +1 -2 4-1 2.25
VLTL +1 0 -1 0 0 0 -1 0 +1 0.31
VLTQ -1 +2 -1 0 0 0 +1 -2 -41 0.37
VQTL -1 0 +1 +2 0 -2 -1 0 +1 2.45
yQTQ +1 -2 +1 -2 +4 -2 +1 -2 +1 -8.97

Tu. 11.70 13.45 14.27 16.12 15.00 17.62 15.21 16.93 18.09

Table 9.10 ANOVA Breakdown for Example 9.2

Source df SS MS F Prob.

V- 2 11.44 5.72
VL 1 9.74 9.74 121.75*** 0.000
''Q 1 1.70 1.70 21.25** 0.001
Tj 2 4.17 2.08
TL 1 4.03 4.03 50.38*** 0.000
TQ 1 0.14 0.14 1.75 0.219
4 1.38 0.35
VLTL 1 0.01 0.01 <1
VLTQ 1 0.00 0.00 <1
VQTL 1 0.25 0.25 3.12 0.111
1 1.12 1.12 14.00** 0.005
Error 9 0.70 0.08

Figure 9.8 Treatrnent means of Example 9.2.

198 3^ Factorial experiments

As an exercise let us also pull out the VT and VT^ components of

interaction from the totals on the diagonals of Table 9.7.

(44.48)^ + (47.66)^ + (46.25)^ (138.39)^^^^^

^ — U.o3
18
(44.79)2 + (47.32)2 (46.28)' (138.39)2
SSyj2 = = 0.54
18

which is approximately (1.39 versus 1.38) the total interaction.

Since the chief concern in this example is to reduce the amount of
noxious gas to a minimum one could examine all nine treatment means in
a Newman-Keuls sense and look for conditions that might result in a
minimum mean.
The means are
5.85, 6.72, 7.13, 7.50, 7.60, 8.06, 8.47, 8.81, 9.04

From Appendix Table E.l for ^2 = 9 df:

p = 2 3 4 5 6 7 8 9
3.20 3.95 4.42 4.76 5.02 5.24 5.43 5.60

and LSRs = 0.64, 0.79, 0.88, 0.95, 1.00, 1.05, 1.09, 1.12.
From these the lines can be drawn as shown above. Clearly, the mini-
mum emission is when one has low prechamber volume ratio and the
shortest injection time as the mean of 5.85 ppm is significantly lower than
any other of the nine means.

9.3 3^ FACTORIAL
If an experimenter has three factors, each at three levels, or a 3 x 3 x 3 = 3^
factorial, there are several ways to break down the effects of factors A, B,
and C and their associated interactions. If the order of experimentation is
completely randomized, the model for such an experiment is

^ijk = /2 + A,- -f- Bj + ABij ACij. + BCji, + ABCijj^ +

with the last two terms confounded unless there is replication within the
cells. In this model / = 1, 2, 3,j — 1, 2, 3, and k = 1, 2, 3, making 27 treat-
ment combinations. These 27 treatment combinations may be as shown in
Table 9.11.
 9.3 3^ Factorial 199

Table 9.11 3^ Factorial Treatment Combinations

Factor A

Factor B Factor C 0 1 2

0 0 000 100 200

1 001 101 201
2 002 102 202

1 0 010 no 210
1 on 111 211
2 012 112 212

2 0 020 120 220

1 021 121 221
2 022 122 222

Association of the proper coefficients on these 27 treatment combina-

tions would allow the Table 9.12 breakdown of an ANOVA if all effects
were set at quantitative levels.
In an actual problem these three-way interactions would be hard to
explain, and quite often the ABC interaction is left with its 8 df for use as an
error term to test the main effects A, B, C, and the two-way interactions.

Table 9.12 3^ Factorial Analysis for Linear and Quadratic Effects

Source df Source df

A 2 AQCJ^ 1
AL 1 AQCQ 1
AQ 1 BCj, 4
Bj 2 BLCL 1
BL 1 BLCQ 1
BQ 1 BQCL 1
ABij 4 BQCQ 1
ALBL 1 ABC,J, 8
ALBQ 1 ALBLCL 1
AQBJ^ 1 A^B^CQ 1
AQBQ 1 ALBQCJ^ 1
Q 2 ALBQCQ 1
Q 1 AQBJ^CJ^ 1
CQ 1 AQBJ^CQ 1
AQ, 4 AQBQC 1
ALCL 1 AQBQCQ 1
ALCQ 1
Total 26
200 3^ Factorial experiments

Table 9.13 3^ Factorial in 2-df Analyses

Source df

A 2
B 2

AB 2)ia
AB^ 2) r
C 2

AC 21la
AC^ 2| r
BC 21
la
BC^ 2r
ABC 21
ABC^ 2
^8
AB^C 2
AB^C^ 2j

Total 26

Another possible partitioning of these effects is in terms of 2-df effects

using / and J components on AB, AC, and BC interactions. These could be
designated as AB, AB^, AC, AC^, and BC, BC^, each with 2 df. However,
the three-way interaction with its 8 df may need a further breakdown.
Sometimes ABC is broken into four 2-df components called X{ABC),
Y{ABC), Z{ABC), and W{ABC), or, using the notation of the last section:
AB^C, ABC^, ABC, and AB^C^. Here again no first letter is squared, and
A^BC — (A^BC)^ = A'^B^C^ = AB^C^ modulus 3. Such a partitioning
would yield Table 9.13.

Example 9.3 A problem involving the effect of three factors, each at three
levels, was proposed by Professor Burr of Purdue University. Here the
measured variable was yield and the factors that might affect this response
were days, operators, and concentrations of solvent. Three days, three op-
erators, and three concentrations were chosen. Days and operators were
qualitative effects, concentrations were quantitative and set at 0.5, 1.0, 2.0.
Although these are not equispaced, the logarithms of these three levels are
equispaced, and the logarithms can then be used if a curve fitting is warranted.
For the purposes of this chapter, all levels of all factors will be considered
as fixed and the design will be considered as completely randomized. It was
decided to take three replications of each of the 3^ = 27 treatment com-
binations. The data, after coding by subtracting 20.0, are as presented in
Table 9.14.
 9.3 3^ Factorial 201

Table 9.14 Example Data on 3^ Factorial with Three Replications

Day Di
5/14 5/15 5/16

Operator Oj
Concentration A B C ABC A B C

0.5 1.0 0.2 0.2 1.0 1.0 1.2 1.7 0.2 0.5
1.2 0.5 0.0 0.0 0.0 0.0 1.2 0.7 1.0
1.7 0.7 -0.3 0.5 0.0 0.5 1.2 1.0 1.7

1.0 5.0 3.2 3.5 4.0 3.2 3.7 4.5 3.7 3.7
4.7 3.7 3.5 3.5 3.0 4.0 5.0 4.0 4.5
4.2 3.5 3.2 3.5 4.0 4.2 4.7 4.2 3.7

2.0 7.5 6.0 7.2 6.5 5.2 7.0 6.7 7.5 6.2
6.5 6.2 6.5 6.0 5.7 6.7 7.5 6.0 6.5
7.7 6.2 6.7 6.2 6.5 6.8 7.0 6.0 7.0

If these data are analyzed on a purely qualitative basis, the methods of

Chapter 5 can be used. The resulting ANOVA is shown in Table 9.15.
The model for this example is merely = fi plus the sum of the
terms in the Source column in Table 9.15. From this analysis the concentra-
tion effect is tremendous, and the days, operators, and day x operator
interaction are all significant at the 1 percent level of significance.
Since concentrations are at quantitative levels, the linear and quadratic
effects of concentrations may be computed, as well as the interactions be-
tween linear effect of concentration and days, quadratic effect of concentra-
tion and days, linear effect of concentration and operators, and quadratic
effect of concentration and operators. It is not usually worthwhile to extract

Table 9.15 First ANOVA for Example 9.3

Source df SS MS

Di 2 3.48 1.74**
Oj 2 6.14 3.07**
DO,j 4 4.07 1.02**
Q 2 468.99 234.49***
4 0.59 0.15
OCjk 4 0.89 0.22
8 1.09 0.14
54 9.98 0.18

Totals 80 495.23
202 3^ Factorial experiments

Table 9.16 Cell Totals for D x C and 0 x C Interactions

(a) {b)
Concentration Concentration

Day 0.5 1.0 2.0 Operator 0.5 1.0 2.0

5/14 5.2 34.5 60.5 A 9.5 39.1 61.6

5/15 4.2 33.1 56.6 B 4.3 32.5 55.3
5/16 9.2 38.0 60.4 C 4.8 34.0 60.6

Totals 18.6 105.6 177.5 301.7 Totals 18.6 105.6 177.5 301.7

three-way interaction in this way. To calculate these quantitative effects

it is usually helpful to construct some two-way tables for the interactions
that are being computed. Two of these are shown as Table 9.16(a) and Table
9.16(b).
From Table 9.16(a), applying the linear and quadratic coefficients to the
concentration totals, we have

Sums of Squares
- -1(18.6) + 0(105.6) + 1(177.5) - 158.9

SSc,
CL = 27(2) = 467.58

CQ = +1(18.6) - 2(105.6) + 1(177.5) = -15.1

SS, .^+2!. 1,41

27(6)
SSc = 468.99
For the D x C interactions, consider each level of days separately. At

5/14: = -1(5.2) + 0(34.5) + 1(60.5) = 55.3

5/15: CL - -1(4.2) + 0(33.1) + 1(56.6) = 52.4
5/16: CL = -1(9.2) + 0(38.0) + 1(60.4) = 51.2

The D X CL is then

(55.3)^ + (52.4)^ + (51.2)^ (158.9)^ _ ^

9(2) 27(2)
For quadratic effects, at

5/14: CQ = +1(5.2) - 2(34.5) + 1(60.5) = -3.3

5/15: CQ = +1(4.2) - 2(33.1) + 1(56.6) = -5.4
5/16: CQ = +1(9.2) - 2(38.0) + 1(60.4) = -6.4
 9.3 3^ Factorial 203

The D X CQ SSi„,is then

(-3.3)" + (-5.4)" + (-6.4)" (-15.1)

= 0.09
9(6) 27(6)
and
== + SSpxcQ 0.49 + 0.09 = 0.58
If the same procedure is now applied to the data of Table 9.16(b), we have

(52.1)2 + (51.0)2 + (55.8)2 (158.9)2

SS Ox CL = 0.70
9(2) 27(2)
(-7.1)2 ^ (_5.4)2 + (-2.6)2 (-15.1)2
SS OXCQ 0.19
9(6) 27(6)
and
SS,’Oxc ~
= SS,
^^OXCL + SS,
^^OXCQ 0.70 + 0.19 = 0.89
The resulting ANOVA can now be shown in Table 9.17.

Table 9.17 Second ANOVA for

Example 9.3

Source df SS MS

Di 2 3.48 1 74**

Oj 2 6.14 3.07**
DO, 4 4.07 1.02**
CL 1 467.58 467.58***
1 1.41 2 41**
CQ
D X CL 2 0.49 0.24
D X CQ 2 0.09 0.04
0 X CL 2 0.70 0.35
0 X CQ 2 0.19 0.09
DOC,j, 8 1.09 0.14
^m{ijk) 54 9.98 0.18

Totals 80 495.22

This second analysis shows that the linear effect and the quadratic effect
of concentration are extremely significant. Two plots of Figure 9.8 may help
in picturing what is really happening in this experiment.
Figure 9.9(a) shows the effect of operators, days, and D x O interaction.
Figure 9.9(b) indicates that the linear effect of concentration far outweighs
the quadratic effect and there is no significant interaction. If a straight line
or three straight lines were fit to these data, the logs of the concentrations
would be used, as the logs are equispaced.
204 3-^ Factorial experiments

Figure 9.9 Plots of 3^ example of Table 9.14.

Although this would usually conclude the analysis of this problem, each
interaction will be broken down into its diagonal, or / and J, components
in order to illustrate the technique. To compute the two diagonal compo-
nents of the two-factor interactions, the two parts of Table 9.16 can be used,
along with a similar table for the D x O cells (see Table 9.18).
From Table 9.18 the diagonal components of the D x O interaction are

I{D X O)

_ (39.5 + 28.6 + 34.8)^ + (30.2 + 34.1 + 39.5)^ + (30.5 -f 33.3 + 31.2)^

~ 27

(301.7)'
= 1.74
81

Call it DO\

(30.5 + 28.6 + 39.5)^ + (96.2)^ -f (106.9)^ (301.7)'

J{D X O) =
27 81

= 2.33

Table 9.18 Cell Totals for D x 0 Interactions

Operator

Day A B C

5/14 39.5 30.2 30.5

5/15 31.2 28.6 34.1
5/16 39.5 33.3 34.8
 9.3 3^ Factorial 205

Table 9.19 Cell Totals for D x O Interaction at Each Level of

Concentration

C, D, at Cl D , at C2 D , at C3

A 3.9 1.5 4.1 13.9 11.0 14.2 21.7 18.7 21.2

B 1.4 1.0 1.9 10.4 10.2 11.9 18.4 17.4 19.5
C -0.1 1.7 3.2 10.2 11.9 11.9 20.4 20.5 19.7

Call it DO. These total 1.74 + 2.33 = 4.07, the D x O interaction sum of
squares.
Applying the same technique to the two parts of Table 9.16 gives

(98.7)^ + (102.7)2 + (100.3)2 (301.7)2

I{DC) = DC^ - = 0.30
27 81

(102.8)2 (99 j(2 (301.7)2

J{DC) = DC = 0.29
27 81
D X C - 0.59
(102.6)2 (99 9(2 (99 2)2 (301.7)2
I{OC) = OC^ - 0.24
27 81

(98.9)2 (104.0)2 + (98.8)2 (301.7)2

J{OC) = OC - 0.65
27 81
0 X C = 0.89

To break down the 8 df of the D x O x C interaction, form an O x D

table showing each of the three levels of concentration C as in Table 9.19.
For each of these concentration levels, find the / and J effect totals, for
example, at C^,
/ components are 8.1, 3.3, 7.2
J components are 5.0, 6.1, 7.5

Now form a table with these I and J components at each level of C (see
Table 9.20).

Table 9.20 Diagonal Totals for Each Level of

Concentration

l(DO) J{DO)

c, h h C jo 7i ji

0.5 8.1 3.3 7.2 5.0 6.1 7.5

1.0 36.0 33.1 36.5 34.6 33.3 37.7
2.0 58.8 58.6 60.1 59.0 56.8 61.7
206 3^ Factorial experiments

Treat each half of Table 9.20 as a simple interaction and compute the
/ and J components. Thus,
(101.3)" + (98.6)" + (101.8)" (301.7)"
DO^C^ = /[C X /(DO)] = 0.22
27 81
(99.1)" + (99.4)" + (103.2)" (301.7)"
DO^C = J[C X /(DO)] = 0.39
27 81
(100.0)" + (102.8)" + (98.9)" (301.7)"
DOC^ = /[C X J(DO)] = - 0.30
27 81
(99.8)" + (102.4)" + (99.5)" (301.7)"
= 0.19
27 81
total D X 0 X C

compared with 1.09 in Table 9.17. This last breakdown into four parts could
also have been accomplished by considering the C x 0 interaction at three
levels of Di, or the C x D interaction at three levels of Oj.
The resulting analysis is summarized in Table 9.21. This analysis is in
substantial agreement with Tables 9.15 and 9.17. No new tests would be
performed on the data in Table 9.21, as they represent only an arbitrary
breakdown of the interactions into 2-df components. The purpose of such a
breakdown is discussed in subsequent chapters. For testing hypotheses on
interaction, these components are added together again.

Table 9.21 Third ANOVA

for 3^ Experiment

Source df SS

D. 2 3.48
Oj 2 6.14
DO 2 2.33
DO^ 2 1.74
c,
DC
2
2
468.99
0.29
DC^ 2 0.30
OC 2 0.65
OC^ 2 0.24
DOC 2 0.19
DOC^ 2 0.30
DO^C 2 0.39
DO^C^ 2 0.22
^m(i jk) 54 9.98

Totals 80 495.24
 Problems 207

9.4 SUMMARY
The summary at the end of Chapter 6 may now be extended for Part II.

Experiment Design Analysis

II. Two or more

factors
A. Factorial
(crossed)
1. Completely randomized
Yijk ^ f^i + Ai + Bj

+ ABij + £k{ij)^ ■ ■ ■

for more factors

a. General case a. ANOVA with
interactions
b. 2^ case b. Yates method or general
ANOVA; use (1), a, b,
ab,.. .
c. 3-^ case c. General ANOVA; use
00, 10, 20,01, 11, ...and
A X B = AB + AB\ . ..
for interaction

PROBLEMS

9.1 Pull-off force in pounds on glued boxes at three temperatures and three humidities
with two observations per treatment combination in a completely randomized
experiment gives:

Temperature A

Humidity B Cold Ambient Hot

50% 0.8 1.5 2.5

2.8 3.2 4.2
70% 1.0 1.6 1.8
1.6 1.8 1.0
90% 2.0 1.5 2.5
2.2 0.8 4.0

Do a complete analysis of this problem by the general methods of Chapter 5.

9.2 Assuming the temperatures in Problem 9.1 are equispaced, extract linear and
quadratic effects of both temperature and humidity as well as all components of
interaction.
9.3 From Problem 9.1 extract the AB and AB^ components of interaction.
208 3^ Factorial experiments

9.4 Develop a “Yates method” for this 3^ experiment and check the results with those
above.
9.5 When the data were collected in Example 9.2 on ppm of noxious gas, data were
also recorded on percent smoke emitted by the 18 engines. Results were as follows:

Volume (F)

Time
(T) Low Medium High

Short 0.3 0.1 0.4

0.4 0.4 0.4
Medium 0.1 0.1 0.4
0.4 0.2 1.2
Long 0.8 0.7 2.5
2.0 1.6 3.6

Do an ANOVA by the methods of Chapter 5.

9.6 From Problem 9.5 extract linear and quadratic components of both main effects
and the interaction.
9.7 From Problem 9.5 extract the VT and the VT^ components of interaction.
9.8 Plot any results that stand out in Problem 9.6.
9.9 Run further tests to recommend the best volume and time combinations for
minimizing smoke in Problem 9.5.
9.10 Compare the results obtained in Problem 9.8 with the results of Example 9.2 in
the text and make some overall recommendations about these engine designs.
9.11 A behavior variable on concrete pavements was measured for three surface
thicknesses: 3 inches, 4 inches, and 5 inches; three base thicknesses: 0 inch, 3

Surface Thickness (inches)

3 4 5

Subbase Base Thickness Base Thickness Base Thickness

Thickness (inches) (inches) (inches)
(inches) 0 3 6 0 3 6 0 3 6

4 2.8 4.3 5.7 4.1 5.4 6.7 6.0 6.3 7.1

2.6 4.5 5.3 4.4 5.5 6.9 6.2 6.5 6.9
8 4.1 5.7 6.9 5.3 6.5 7.7 6.1 7.2 8.1
4.4 5.8 7.1 5.1 6.7 7.4 5.8 7.1 8.4
12 5.5 7.0 8.1 6.5 7.7 8.8 7.0 8.0 9.1
5.3 6.8 8.3 6.7 7.5 9.1 7.2 8.3 9.0
 Problems 209

inches, and 6 inches, and three subbase thicknesses; 4 inches, 8 inches, and 12
inches. Two observations were made under each of the 27 pavement conditions
and complete randomization performed. The results were as shown on page 208.
Do a complete ANOVA of this experiment by the methods of Chapter 5.
9.12 Since all three factors are quantitative and equispaced, determine linear and
quadratie effects for each factor and all interaction breakdowns. Test for signif-
icance in Problem 9.11.
9.13 Break down the interactions of Problem 9.11 into 2-df eomponents such as AB,
AB^, ABC, AB^C, and so on.
9.14 Use a Yates method to solve Problem 9.11 and check the results.
9.15 Plot any significant results of Problem 9.11.
9.16 From Example 9.3 fit the proper degree regression equation for the quantitative
variable and express it in terms of the original variable.
9.17 The following data are on the wet-film thickness (in mils) of lacquer. The factors
studied were: type of resin (two types), gate-blade setting in mils (three settings),
and weight fraction of nonvolatile material in the lacquer (three fractions).

Resin Type
1 2

Gate Weight Fraction Weight Fraction

Setting 0.20 0.25 0.30 0.20 0.25 0.30

2 1.6 1.5 1.5 1.5 1.4 1.6

1.5 1.3 1.3 1.4 1.3 1.4
4 2.7 2.5 2.4 2.4 2.6 2.2
2.7 2.5 2.3 2.3 2.4 2.1
6 4.0 3.6 3.5 4.0 3.7 3.4
3.9 3.8 3.4 4.0 3.6 3.3

Do an ANOVA of the above data and pull out any quantitative effects in terms of
their components.
9.18 Using a = 0.01, write prediction equations where possible and check them
(Problem 9.17).
9.19 If the gate setting is 5 mils and the weight fraetion 0.20, what is the best prediction
of wet-film thickness? If interaction is ignored, what is the prediction? How much
difference does it make? (Problem 9.17.)
9.20 Plot results indicated in Problem 9.18.
 10
Fixed, Random, and Mixed Models

10.1 INTRODUCTION
In Chapter 1 it was pointed out that, in the planning stages of an experiment,
the experimenter must decide whether the levels of factors considered are
to be set at fixed values or are to be chosen at random from many possible
levels. In the intervening chapters it has always been assumed that the factor
levels were fixed. In practice it may be desirable to choose the levels of some
factors at random, depending on the objectives of the experiment. Are the
results to be judged for these levels alone or are they to be extended to more
levels of which those in the experiment are but a random sample? In the
case of some factors such as temperature, time, or pressure, it is usually
desirable to pick fixed levels, often near the extremes and at some inter-
mediate points, because a random choice might not cover the range in which
the experimenter is interested. In such cases of fixed quantitative levels, we
often feel safe in interpolating between the fixed levels chosen. Other factors
such as operators, days, or batches may often be only a small sample of all
possible operators, days, or batches. In such cases the particular operator,
day, or batch may not be very important but only whether or not operators,
days, or batches in general increase the variability of the experiment.
It is not reasonable to decide after the data have been collected whether
the levels are to be considered fixed or random. This decision must be made
prior to the running of the experiment, and if random levels are to be used,
they must be chosen from all possible levels by a random process. In the
case of random levels, it will be assumed that the levels are chosen from an
infinite population of possible levels. Bennett and Franklin [4] discuss a
case in which the levels chosen are from a finite set of possible levels.
When all levels are fixed, the mathematical model of the experiment is
called a fixed model. When all levels are chosen at random, the model is
called a random model. When several factors are involved, some at fixed levels
and others at random levels, the model is called a mixed model.

210
 10.2 Single-factor models 211

10.2 SINGLE-FACTOR MODELS

In the case of a single-factor experiment the factor may be referred to as a
treatment effect, as in Chapter 3; and if the design is completely randomized,
the model is
Yij = i^i + Tj + Sjj (lOd)

Whether the treatment levels are fixed or random, it is assumed in this model
that /r is a fixed constant and the errors are normally and independently
distributed with a zero mean and the same variance, that is, Sij are NID
(0, <jj). The decision as to whether the levels of the treatments are fixed or
random will affect the assumptions about the treatment term TJ. The different
assumptions and other differences will be compared in parallel columns.

Fixed Model Random Model

1. Assumptions: i/s are fixed con- 1. Assumptions: i/s are random
stants. variables and are
k k
NID (0, of)
I = j=i
I (^^■J -0 = 0
i=i (Here represents the variance
(These add to zero as they are among the T/S or among the
the only treatment means being true treatment means fi j. The
considered.) Tj average to zero when averaged
over all possible levels, but for
the k levels of the experiment
they usually will not average 0.)
Figure 10.1(a) shows three fixed means whose average is /a as these are
the only means of concern and Xi^ — A^) = 0. Figure 10.1(b)
shows three random means whose average is obviously not fi as these are
but three means chosen at random from many possible means. These means
and their corresponding i/s are assumed to form a normal distribution with
a standard deviation of cr^.

^3
^2
/ 1
1 1 1

M2 M Ml M3

(a)

Figure 10.1 Assumed means in (a) fixed and (b) random models.

2. Analysis: Procedures as given in 2. Analysis: Same as for fixed

Chapter 3 for computing SS. model.
212 Fixed, random, and mixed models

EMS: 3. EMS:

Source df EMS Source df EMS

A
k - 1 (7s + A k - 1 cTg -I-

^ij k{n — 1) ^ij k{n — 1) (7e

4. Hypothesis tested: 4. Hypothesis tested:

Ho: Tj = 0 (for all j) Ho: (7^ - 0
The expected mean square (EMS) column turns out to be extremely
important in more complex experiments as an aid in deciding how to set up
an F test for significance. The EMS for any term in the model is the long-
range average of the calculated mean square when the Y^j from the model
is substituted in algebraic form into the mean square computation. The
derivation of these EMS values is often complicated, but those for the single-
factor model were derived in Chapter 3 and those for two-factor models
are derived in a later section of this chapter.
For the fixed model, if the hypothesis is true that Zj = 0 for all j, that
is, all the k fixed treatment means are equal, then the EMS
for Tj and e^j are both af. Hence the observed mean squares for treatments
and error mean square are both estimates of the error variance, and they
can be compared by means of an F test. If this F test shows a significantly
high value, it must mean that n Yj ~ 1) = ^4>r is not zero and the
hypothesis is to be rejected.
For the random model, if the hypothesis is true that cr^ = 0, that is, the
variance among all treatment means is zero, then again each mean square
is an estimate of the error variance. Again, an F test between the two mean
squares is appropriate.
From the two tables in step 3 above, it is seen that for a single-factor
experiment there is no difTerence in the test to be made after the analysis,
and the only difference is in the generality of the conclusions. If HQ is re-
jected, there is probably a difference between the k fixed treatment means
for the fixed model; for the random model there is a difference between all
treatments of which the k examined are but a random sample.

10.3 TWO-FACTOR MODELS

For two factors A and B the model in the general case is

^ijk = /^ + Aj + Bj + ABij -t- Sj^oj)

with
i = 1, 2,. . ., u 7 = 1, 2,. . . , h /c = 1, 2, . . ., 7?
 10.3 Two-factor models 213

provided the design is completely randomized. In this model, it is again

assumed that ft is a fixed constant and ehiij)S are NID (0, a^). If both A and
B are at fixed levels, the model is a fixed model. If both are at random levels,
the model is a random model, and if one is at fixed levels and the other at
random levels, the model is a mixed model. Comparing each of these models
gives:

Fixed Random Mixed

1. Assumptions: 1. Assumptions: 1. Assumptions:
Aj’s are fixed A/s are NID (0, aj) A/s are fixed and
constants and
lAi^O
i A, = 0 i

i= 1

B/s are fixed B/s are NID (0, al) B/s are NID (0, ai)
constants and

i B, = 0
i=i

ABijS are fixed AB^jS are ABi/s are

constants and NID (0, NID (0, crjg) but

I AB,J = 0
i t AB<j = 0
i
Z AB,, = 0
i AB,J # 0
j

(for A fixed, B
random)
2. Analysis: 2. Analysis: Same 2. Analysis: Same
Procedures of
Chapter 5 for sums
of squares
3. EMS: 3. EMS: 3. EMS:

EMS EMS EMS

Source df (Fixed) (Random) (Mixed)

Ai a — 1 aj + nb(t)^ + n(^AB + nba^^ +

Bi 6—1 CTg + na(j)B cTg + + naal + nao\

AB, {a - 1)(6 - 1) + n(l)AB + no^AB

ab{n — 1) 'yl
214 Fixed, random, and mixed models

4. Hypotheses tested: 4. Hypotheses tested: 4. Hypotheses tested:

Ai = 0 for all i Hj - 0 Hj Ai = 0 for all i

Hi BJ = 0 for all j Hi = 0 Hi = 0
Hi AB,j = 0 for all Hi ^AB = 0 Hi <^AB = 0
i and j

In the assumptions for the mixed model the fact that summing the
interaction term over the fixed factor is zero but summing it over the
random factor is not zero affects the expected mean squares, as seen
in item 3 on page 213.
For the fixed model the mean squares for A, B, and AB are each com-
pared with the error mean square to test the respective hypotheses, as should
be clear from an examination of the EMS column when the hypotheses are
true. For the random model the third hypothesis of no interaction is tested
by comparing the mean square for interaction with the mean square for
error, but the first and second hypotheses are each tested by comparing the
mean square for the main effect {A^ or Bj) with the mean square for the
interaction as seen by their expected mean square values. For a mixed model
the interaction hypothesis is tested by comparing the interaction mean
square with the error mean square. The random effect Bj is also tested by
comparing its mean square with the error mean square. The fixed effect (Ai),
however, is tested by comparing its mean square with the interaction mean
square.
From these observations on a two-factor experiment, the importance
of the EMS column is evident, as this column can be used to see how the
tests of hypotheses should be run. It is also important to note that these
EMS expressions can be determined prior to the running of the experiment.
This will indicate whether or not a good test of a hypothesis exists. In some
cases the proper test indicated by the EMS column will have insufficient
degrees of freedom to be sufficiently sensitive, in which case the investigator
might wish to change the experiment. This would involve such changes as a
choice of more levels of some factors, or changing from random to fixed
levels of some factors.

10.4 EMS RULES

The two examples above have shown the importance of the EMS column
in determining what tests of significance are to be run after the analysis is
completed. Because of the importance of this EMS column in these and more
complex models, it is often useful to have some simple method of determining
 10.4 EMS rules 215

these values from the model for the given experiment. A set of rules can be
stated that will determine the EMS column very rapidly, without recourse
to their derivation. The rules will be illustrated on the two-factor mixed
model of Section 10.3. To determine the EMS column for any model:
1. Write the variable terms in the model as row headings in a two-way
table.

^k{ij)

2. Write the subscripts in the model as column headings; over each sub-
script write F if the factor levels are fixed, R if random. Also write the
number of observations each subscript is to cover.

a b n
F R R
i j k

At

AB,j

3. For each row (each term in the model) copy the number of observations
under each subscript, providing the subscript does not appear in the
row heading.

a b n
F R R
i j k

Ai b n
a n
AB,j n
^k{ij)

4. For any bracketed subscripts in the model, place a 1 under those sub-
scripts that are inside the brackets.
216 Fixed, random, and mixed models

a b n
F R R
i j k

A, b n
a n
AB,j n
1 1

5. Fill the remaining cells with a 0 or a 1, depending upon whether the

subscript represents a fixed F or a random R factor.

a b n
F R R
i j k

Ai 0 b n
a 1 n
AB,, 0 1 n
1 1 1

6. To find the expected mean square for any term in the model:
a. Cover the entries in the column (or columns) that contain non-
bracketed subscript letters in this term in the model (for example,
for Ai, cover column i; for cover column k).
b. Multiply the remaining numbers in each row. Each of these products
is the coefficient for its corresponding term in the model, provided
the subscript on the term is also a subscript on the term whose
expected mean square is being determined. The sum of these coeffi-
cients multiplied by the variance of their corresponding terms (0^ or
a^) is the EMS of the term being considered (for example, for T,-,
cover column i). The products of the remaining coefficients are bn, n,
n, and 1, but the first n is not used, as there is no / in its term {Bj).
The resulting EMS is then bncj)^ + -I- 1 • cr^. For all terms,
these rules give:

a b n
F R R
i J k EMS

d, 0 b n + najiB + nb(f)^
B, a 1 n + naal
/IB,, 0 1 u +
1 1 1
 10.4 EMS rules 217

These results are seen to be in agreement with the EMS values for the
mixed model in Section 10.3. Here is, of course, a fixed type of variance

liAf
—

a
Although the rules seem rather involved, they become very easy to use with
a bit of practice. Two examples will illustrate the concept.

Example 10.1 The viscosity of a slurry is to be determined by four randomly

selected laboratory technicians. Material from each of five mixing machines
is bottled and divided in such a way as to provide two samples for each
technician to test for viscosity. These are the only mixing machines of interest
and the samples can be presented to the technicians in a completely ran-
domized order.
The model here assumes four random technicians and five fixed mixing
machines, and each technician measures samples of each machine twice. The
model is shown as the vertical column of Table 10.1 and the remainder of
the table shows how the EMS column is determined.
Table 10.1 EMS for Example 10.1

4 5 2
R F R
Source df i i k EMS

Ti 3 1 5 2 Cg + i0(Tj
Mj 4 4 0 2 CTg + 2(TJ]^ +
TM,j 12 1 0 2 CTg + 2<TYM
20 1 1 1

The proper F tests are quite obvious from Table 10.1 and all tests have
adequate degrees of freedom for a reasonable test.

Example 10.2 An industrial engineering student wished to determine the

effect of five different clearances on the time required to position and assemble
mating parts. As all such experiments involve operators, it was natural to
consider a random sample of operators to perform the experiment. He also
decided the part should be assembled directly in front of the operator and at
arm’s length from the operator. He tried four different angles, from 0°
directly in front of the operator through 30°, 60°, and 90° from this position.
Thus four factors were involved, any one of which might affect the time
required to position and assemble the part. The experimenter decided to
replicate each setup six times and to randomize completely the order of
experimentation. Here operators Oj were at random levels (six being chosen),
angles Aj at four fixed levels (0°, 30°, 60°, 90°), clearances Cj, at five fixed
levels and locations fixed either in front of or at arm’s length from the
218 Fixed, random, and mixed models

operator. This isa6x4x5x2 factorial experiment with six replications,

run in a completely randomized design. The expected mean square values
can be determined from the rules given in Section 10.4 as shown in Table 10.2.

Table 10.2 EMS for Clearance Problem

6 4 5 2 6
R F F F R
Source df i j k m Q EMS

0, 5 1 4 5 2 6 + 240(TJ
-4; 3 6 0 5 2 6 + 60aoA + 3600^
0/1,; 15 1 0 5 2 6 (^e 60(7^A
C, 4 6 4 0 2 6 + 48(7OC + 2880c
20 1 4 0 2 6 + 48(7CC
12 6 0 0 2 6 + i2(7})AC + 720^c
OACij, 60 1 0 0 2 6 + 13(7’O4C

Bfn 1 6 4 5 0 6 + 120crcc + 7200 c

OLi„, 5 1 4 5 0 6 + UOa^c
A1 3 6 0 5 0 6 + 30(7 + 18O0^c
OALij„, 15 1 0 5 0 6 + 30cro^c
4 6 4 0 0 6 + 24(7CCL + 1440CL
OCLij^^ 20 1 4 0 0 6 + 34(7OCL
A^ 12 6 0 0 0 6 (^8 + ^^OACL + ACL
OACLij,,, 60 1 0 0 0 6 + ^^OACL

^q{i jkm) 1200 1 1 1 1 1

From this table it is easily seen that all interactions involving operators
and the operator main effect are tested against the error mean square at the
bottom of the table. All interactions and main effects involving fixed factors
are tested by the mean square just below them in the table.
The rules given in this section are general enough to be applied to the
most complex designs, as will be seen in later chapters.

10.5 EMS DERIVATIONS

Single-Factor Experiment

The EMS expressions for a single-factor experiment were derived in Chap-

ter 3, Section 3.6.
Two-Factor Experiment

Using the definitions of expected values given in Chapter 2 and the proce-
dures of Section 3.6 the EMS expressions may be derived for the two-factor
experiment.
For a two-factor experiment, the model is

\jk — Ai Bj ABij -b (10.2)

 10.5 EMS derivations 219

with
/ = 1, 2,. . ., a j = I, 2, b k = 1,2, . . ., n
The sum of squares for factor A is

SS^ = ^ nb{Y,, - y..)^

i= 1

From the model in Equation (10.2)

= SZj k
= ZZ j k
+ h(ij))/b^

+Z j
+Z j
+ Z Z ^kiij))/b^
j k

y... = Z Z Z yijk/abn
i j k

a b n

^ ZZZi j k
+ ^kiij))/^b^
F ,. = // + Z A,la + Z B^jb + Z Z AB.^/ab
I }
a b n

+ Z Z Z <^Hii)lnab
i j k

Subtracting gives

Yi..- ¥,,, = { A, -t AJa) +

i /
(Z AB,j/b
\ j
- f Z ABJab
i j

+ ( Zj Zk et(ij)/bn - Z Z Z enij)/abn i j k

Note that the B effect cancels out of the A sum of squares as it should, since
A and B are orthogonal effects in a factorial experiment. Squaring and
adding gives

SS^ = nb - Z AJa
i= 1 \ i

+ Z fz ABij/b - Z Z ABJab
j= 1 \ j i j
a /bn a b n \2
+ Z ZZ
i= 1 \ j k
- Z Z Z Haj)/abn \
i j k J

+ (cross-product terms)
220 Fixed, random, and mixed models

Taking the expected value for a fixed model where

X = 0 X AB,J = 0
i i or j
the result is
^ nb
T(SS^) = nh ^ + 0 H—^ (abn — bn)a^

E{MS^) = ElSSJia - 1)] = nb j] Af/{a - 1) + (T^ = nbcj)^ + GI

i= 1

which agrees with Section 10.3 for the fixed model.

If now the levels of A and B are random,
X # 0 X AB,J ^ 0
i i or j

nb nb
£(SS^) = nb{a — l)aj + -^ (ab — b)G\B 3—yrr “ bn)G^
b n b
T(MS^) = nba^ + na^^ +

as stated in Section 10.3 for a random model.

If the model is mixed with A fixed and B random,

X /li = 0
i
t '45,7 = 0
i
but

Z AB,j 7^ 0
j
then
^ nb ub
£(SS^) - nb X [nab - nb)Gj
i b n b
a

£(MS^) = nbY, - 1) + na^^ + (yl = nbcj)^ + na^s +

i

which agrees with the value stated in Section 10.3 for a mixed model.
Using these expected value methods, one can derive all EMS values
given in Section 10.3. It might be noted that if A were random in the mixed
model

I AB,j = 0
j
then the interaction term would not appear in the factor A sum of squares.
This is true of B in the mixed model of Section 10.3.
These few derivations should be sufficient to show the general method
of derivation and to demonstrate the advantages of the simple rules in
Section 10.4 in determining these EMS values.
 10.6 The pseudo-/’ test 221

10.6 THE PSEUDO-E TEST

Occasionally the EMS column for a given experiment indicates that there is
no exact F test for one or more factors in the design model. Consider the
following example.

Example 10.3 Two days in a given month were randomly selected in which
to run an experiment. Three operators were also selected at random from a
large pool of available operators. The experiment consisted of measuring the
dry-film thickness of varnish in mils for three different gate settings: 2, 4, and
6 mils. Two determinations were made by each operator each day and at
each of the three gate settings. Results are shown in Table 10.3.

Table 10.3 Dry-Film Thickness Experiment

Day
1 2

Gate Operator Operator

Setting ABC ABC

2 0.38 0.39 0.45 0.40 0.39 0.41

0.40 0.41 0.40 0.40 0.43 0.40
4 0.63 0.72 0.78 0.68 0.77 0.85
0.59 0.70 0.79 0.66 0.76 0.84
6 0.76 0.95 1.03 0.86 0.86 1.01
0.78 0.96 1.06 0.82 0.85 0.98

Assuming that days and operators are random effects, gate settings are
fixed, and the design is completely randomized, the analysis yields Table 10.4.

Table 10.4 Analysis of Dry-Film Thickness Experiment

Source df SS MS EMS

Days D 1 0.0010 0.0010 6GIO + 18o-J

Operators O 2 0.1121 0.0560 + 6GIO + 12GO
D X O interaction 2 0.0060 0.0030** +
Gate setting G 2 1.5732 0.7866** + ^^DOG +
+ + 12(J>G
D X G interaction 2 0.0113 0.0056 + ^^DOG + OGQQ

O X G interaction 4 0.0428 0.0107 + ^^DOG + ^^OG

D X O X G interaction 4 0.0099 0.0025** + ^^DOG
Error 18 0.0059 0.0003

Totals 35 1.7622

** Two asterisks indicate significance at the 1-percent level.

222 Fixed, random, and mixed models

All F tests are clear from the EMS column except for the test on gate
setting, which is probably the most important factor in the experiment. Only
the three-way interaction shows significance. It is obvious from the results
that gate setting is the most important factor, but how can it be tested? If the
D X G interaction is assumed to be zero, then the gate setting can be tested
against the 0 x G interaction term. On the other hand, if the 0 x G inter-
action is assumed to be zero, gate setting can be tested against the D x G
interaction term. Although neither of these interactions is significant at the
5 percent level, both are numerically larger than the D x O x G interaction
against which they are tested. In this case any test on G is contingent upon
these tests on interaction. One method for testing hypotheses in such situa-
tions was developed by Satterthwaite and is given in Bennett and Franklin
[4, pp. 367-368].
The scheme consists of constructing a mean square as a linear combina-
tion of the mean squares in the experiment, where the EMS for this mean
squares includes the same terms as in the EMS of the term being tested, except
for the variance of that term. For example, to test the gate-setting effect G
in Table 10.4 a mean square is to be constructed whose expected value is

+ 2<TIOG + 4a^o +

This can be found by the linear combination

MS = MS^G + MSOG - MS^JOG

as its expected value is

f:(MS) = + 2aloG + 4a^G + + laloG + 6(7^c - al - 2aloa

= (Tj + 2aj)oo + 4(7OG + SCDG

An F test can now be constructed using the mean square for gate setting as
the numerator and this mean square as the denominator. Such a test is called
a pseudo-F or F' test. The real problem here is to determine the degrees
of freedom for the denominator mean square. According to Bennett and
Franklin, if

MS = GI(MS)I + G2(MS)2 + • • •

and (MS)i is based on df, (MS )2 is based on Vj df, and so on, then the
degrees of freedom for MS are

(MS)^
+ ail{MS)l/v2] + ‘ '

In the case of testing for the gate-setting effect above, = 1, ^^d

<^2

U = — 1 and the degrees of freedom are tq =4, V2 = 2, and v^, ^ 4. Here

3
 10.7 Remarks 223

MS = 0.0107 + 0.0056 - 0.0025 = 0.0138 and its df is

(0.0138)^
“ (1)^[(0.0107)V4] + (1)2[(0.0056)V2] + (-1)"[(0.0025)"/4]

_ 1.9044 X 10“^
“ 0.4586 X 10"'‘ ^

Hence the F' test is

MSG 0.7866 _ ^
“ " 0.0138 “ ■
with 2 and 4.2 df, which is significant at the 1 percent level of significance
based on F with 2 and 4 or 2 and 5 df.

10.7 REMARKS

The examples in this chapter should be sufficient to show the importance of

the EMS column in deciding just what mean squares should be compared
in an F test of a given hypothesis. This EMS column is also useful (usually in
random models) to solve for components of variance as illustrated in Chapter
3, Section 3.6.
One special case is of interest. In a two-factor factorial when there is but
one observation per cell {n = 1), the EMS columns of Section 10.3 reduce to
those in Table 10.5.

Table 10.5 EMS for One Observation Per Cell

Source EMS (Fixed) EMS (Random) EMS (Mixed)

Ai aj + b(p^ + (^AB + b(T\ (^E+ <^AB + b(i)^

Bj ol + acpB + aol GI + aol

ABij or 6ij + <PAB ^E + (^AB + ^AB

A glance at these EMS values will show that there is no test for the main
effects A. and .6 in a fixed model, as interaction and error are hopelessly con-
founded. The only test possible is to assume that there is no interaction; then
= 0, and the main effects are tested against the error. If a no-interaction
assumption is not reasonable from information outside the experiment, the
investigator should not run one observation per cell but should replicate the
data in a fixed model.
For a random model both main effects can be tested whether interaction
is present or not. For a mixed model there is a test for the fixed effect A but
224 Fixed, random, and mixed models

no test for the random effect B. This may not be a serious drawback, since
the fixed effect is often the most important; the B effect is included chiefly for
reduction of the error term. Such a situation is seen in a randomized block
design where treatments are fixed, but blocks may be chosen at random.
In the discussion of the single-factor experiment it was assumed that
there were equal sample sizes n for each treatment. If this is not the case, it
can be shown that the expected treatment mean square is

or + no(j>.
and

ttn =
S;=i nj
(k - l)N
where

N= Z yi;
J= 1

The test for treatment effect is to compare the treatment mean square with
the error mean square; the use of HQ is primarily for computing components
of variance.

PROBLEMS

10.1 An experiment is run on the effects of three randomly selected operators and five
fixed aluminizers on the aluminum thickness of a TV tube. Two readings are
made for each operator-aluminizer combination. The following ANOVA table
is compiled.

Source df SS MS

Operators 2 107,540 53,770

Aluminizers 4 139,805 34,951
0 X A interaction 8 84,785 10,598
Error 15 230,900 15,393

Totals 29 563,030

Assuming complete randomization, determine the EMS column for this problem
and make the indicated significance tests.
10.2 Consider a three-factor experiment where factor T is at u levels, factor B at
h levels, and factor C at c levels. The experiment is to be run in a completely
randomized manner with ti observations for each treatment combination. As-
suming factor A is run at a random levels and both B and C at fixed levels, deter-
mine the EMS column and indicate what tests would be made after the analysis.
10.3 Repeat Problem 10.2 with A and B at random levels, but C at fixed levels.
 Problems 225

10.4 Repeat Problem 10.2 with all three factors at random levels.
10.5 Consider the completely randomized design of a four-factor experiment similar
to Example 10.2. Assuming factors A and B are at fixed levels and C and D are
at random levels, set up the EMS column and indicate the tests to be made.
10.6 Assuming three factors are at random levels and one is at fixed levels in Prob-
lem 10.5, work out the EMS column and the tests to be run.
10.7 A physical education experiment will be conducted to investigate the effects of
four types of exercise on heart rate. Five subjects are chosen at random from a
physical education class. A subject does each exercise twice. The 40 measurements
are done in a completely randomized order with each subject allowed at least
10 minutes’ rest between exercises.
a. Give the model. State the parameters of its random variables.
b. Work out the expected mean squares of all the effects. Also give the degrees
of freedom for the effects.
c. Tell how to form the F’s for the tests that can be made.
10.8 Four elementary schools are chosen at random in a large city system and two
methods of instruction are tried in the third, fourth, and fifth grades,
a. Outline an ANOVA layout to test the effect of schools, methods, and grades
on gain in reading scores assuming ten students per class.
b. Work out the EMS column and indicate what tests are possible in this
situation.
10.9 As part of an experiment on testing viscosity four operators were chosen at
random and four instruments were randomly selected from many instruments
available. Each operator was to test a sample of product twice on each instru-
ment. Results showed:

Source df MS

0, 3 1498
Ij 3 1816
OIu 9 218
16 67

Determine the EMS column and find the percentage of the total variance in
these readings that is attributable to each term in the model, if it is significant.
10.10 Six different formulas or mixes of concrete are to be purchased from five com-
peting suppliers. Tests of crushing strength are to be made on two blocks con-
structed according to each formula-supplier combination. One block will be
poured, hardened, and tested for each combination before the second block is
formed, thus making two replications of the whole experiment. Assuming mixes
and suppliers fixed and replications random, set up a mathematical model for
this situation, outline its ANOVA table, and show what F tests are appropriate.
10.11 Determine the EMS column for Problem 3.6 and solve for components of
variance.
226 Fixed, random, and mixed models

10.12 Derive the expression for HQ in the EMS column of a single-factor completely
randomized experiment where the ns are unequal.
10.13 Verify the numerical results of Table 10.4 using the methods of Chapter 6.
10.14 Verify the EMS column of Table 10.4 by the method of Section 10.4.
10.15 Set up F' tests for Problem 10.4 and explain how the df would be determined.
10.16 Set up F' tests for Problem 10.5 and explain how the df would be determined.
 II
Nested and Nested-Factorial Experiments

11.1 INTRODUCTION
Example 11.1 In a recent in-plant training course the members of the class
were assigned a final problem. Each class member was to go into the plant
and set up an experiment using the techniques that had been discussed in
class. One engineer wanted to study the strain readings of glass cathode
supports from five different machines. Each machine had four “heads” on
which the glass was formed, and she decided to take four samples from each
head. She treated this experiment as a 5 x 4 factorial with four replications
per cell. Complete randomization of the testing for strain readings presented
no problem. Her model was

Yijk = /i 4- Mj -t- + MH^j +

with

i = 1,2,.. .,5 j - 1,..., 4 k = 1,. . .,4

Her data and analysis appear in Table 11.1. In this model she assumed that
both machines and heads were fixed, and used the 10 percent significance
level. The results indicated no significant machine or head effect on strain
readings, but there was a significant interaction at the 10 percent level of
significance.
The question was raised as to whether the four heads were actually
removed from machine A and mounted on machine B, then on C, and so
on. Of course, the answer was “no” as each machine had its own four heads.
Thus machines and heads did not form a factorial experiment, as the heads
on each machine were unique for that particular machine. In such a case
the experiment is called a nested experiment: levels of one factor are nested
within, or are subsamples of, levels of another factor. Such experiments are
also sometimes called hierarchical experiments.

227
228 Nested and nested-factorial experiments

Table 11.1 Data and ANOVA for Strain-Reading Problem

Machine
Head A B C D E

1 6 10 0 11 1
2 9 0 0 4
0 7 5 6 7
8 12 5 4 9

2 13 2 10 5 6
3 1 11 10 7
9 1 6 8 0
8 10 7 3 3

3 1 4 8 1 3
10 1 5 8 0
0 7 0 9 2
6 9 7 4 2

4 7 0 7 0 3
4 3 2 8 7
7 4 5 6 4
9 1 4 5 0

Source df SS MS EMS F EO.90

M, 4 45.08 11.27 1.05 2.04

Hi 3 46.45 15.48 1.45 2.18
MH,j 12 236.42 19.70 + ^(1>MH 1.84 1.66
60 642.00 10.70

Totals 79 969.95

11.2 NESTED EXPERIMENTS

The preceding example can now be reanalyzed by treating it as a nested
experiment, since heads are nested within machines. Such a factor may be
represented in the model as where j covers all levels 1, 2,... within the
ith level of M, . The number of levels of the nested factor need not be the same
for all levels of the other factor. They are all equal in this problem, that is,
j = 1, 2, 3, 4 for all i. The errors, in turn, are nested within the levels of /
and j; and k = 1, 2, 3, 4 for all i and j.
To emphasize the fact that the heads on each machine are different
heads, the data layout in Table 11.2 shows heads 1, 2, 3, and 4 on machine
A; heads 5, 6, 7, and 8 on machine B; and so on.
 11.2 Nested experiments 229

Table 11.2 Data for Strain- Reading Pro' :j>lem in a Nested Arrangement
C _. A.. . .
rOA' ^ ' 1
Machine A B c 1 D E

Headci>i'>>J 1 2 3, 4 5 6 1- 8 1
9 10 11 12 13 14 15 16 17 18 19 20

6 13 1 ' 7 10 2 4 0 0 10 8 7 11 5 1 0 1 6 3 3
2 3 10 4 9 1 1
3 ( 0 11 5 2 0 10 8 8 4 7 0 7
0 9 0 7 7 1 7 4 5 6 0 5 6 8 9 6 7 0 2 4
8 8 6 9 12 10 9 1 5 7 7 4 4 3 4 5 9 3 2 0

Head
totals 16 33 17 27 38 14 21 8
1 10 34 20 18 21 26 22 19 21 16 7 14

Machine
totals 93 81 82 88 58

As the heads that are mounted on the machine can be chosen from
many possible heads, we might consider the four heads as a random sample
of heads that might be used on a given machine. If such heads were selected
at random for the machines, the model would be

^ijk = ^j(i) + ^k{ij)

with
z = 1,.. ., 5 j = 1,.. ., 4 /c = 1,..., 4

This nested model has no interaction present, as the heads are not
crossed with the five machines. If heads are considered random and ma-
chines fixed, the proper EMS values can be determined by the rules given
in Chapter 10, as shown in Table 11.3.
This breakdown shows that the head effect is to be tested against the
error, and the machine effect is to be tested against the heads-within-machines
effect.
To analyze the data for a nested design, first determine the total sum
of squares

SS,„,„ = 6^ + 2^ + • • • + 4^ + 0^ - = 969.95

Table 11.3 EMS for Nested Experiment

5 4 4
F R R
Source i j k EMS

Mi 0 4 4 + 4cr^ +
Hjw 1 1 4 af + 4a^
^kiU) 1 1 1
230 Nested and nestedrfactorial experiments

and for machines

(93)^ + (81)2 ^ (32)2 ^ (ggj2 ^ (53^2 (4Q2^:

SSjvf — = 45.08
16 80

To determine the sum of squares between heads within machines, consider

each machine separately.

Machine A:
(16)2 (33^2 (27)2 (93)2
SS„ = = 50.19
4 16

Machine B :
(38)2 (14)2 (21)2 (g)2 (81)2
SS„ = = 126.18
4 16

Machine C:
(10)2 + (34)2 (2Q)2 (lg)2 (82)2
SS„ : - 74.75
4 16

Machine D :
(21)2 ^ (26)2 + (22)2 ^ (19)2 (88)2
SS„ : = 6.50
4 16

Machine E :
GO
77i

(21)2 (16)2 (7)2 (14)2

SSH 25.25
4 16
total SS^ = 282.87

The error sum of squares by subtraction is then

969.95 - 45.08 - 282.87 = 642.00

The degrees of freedom between heads within machine A are 4 — 1 = 3; for

all five machines the degrees of freedom will be 5 x 3 = 15. The analysis
follows in Table 11.4.
From this analysis, machines appear to have no significant effect on
strain readings, but there is a slightly significant (10 percent level) effect of
heads within machines on the strain readings. Note that what the experi-
menter took as head effect (3 df) and interaction effect (12 df) is really
heads-within-machines effect (15 df). These results might suggest a more
careful adjustment between heads on the same machine. Since there are
significant differences between heads within machines, the heads-within-
machines sum of squares of 282.87 might well be analyzed further. Returning
 11.2 Nested experiments 231

Table 11.4 ANOVA for Nested Strain-Reading Problem

Source df SS MS EMS F FO.90

M, 4 45.08 11.27 + 4crJ + 16(1) M < 1 2.36

15 282.87 18.85 ol + Aol 1.76 1.60
60 642.00 10.70

Totals 79 969.95

to the method by which this term was computed, one can look at its parti-
tioning into five parts with 3 df per machine, and since all parts are or-
thogonal, a further breakdown and tests of Table 11.4 can be made as in
Table 11.5.

Table 11.5 Detailed ANOVA of Example 11.1

Source df SS MS F FO.90

M, 4 45.08 11.27 <1 2.36

15 282.87
^jiA) 3 50.19 16.73 1.56 2.18
^j(B) 3 126.18 42.06 3.93* 2.18
3 74.75 24.92 2.33* 2.18
^j{D) 3 6.50 2.17 <1 2.18
^j(E) 3 25.25 8.42 <1 2.18
60 642.00 10.70

* Significant at 10 percent level.

From Table 11.5, using a 10 percent significance level, we see a signifi-

cant diflference between heads on machine B and machine C. One might
then examine the head means within each of these machines in a Newman-
Keuls sense.
For machine 5, means are: 2.00 3.50 5.25 9.50
For heads: 8 6 7 5

Sy = = 1-64 with 60 df for ^2-

From Appendix Table E.l (although one should have a 10 percent table)
p: 2 3 4
2.83 3.40 3.74
LSR: 4.64 5.58 6.13

So groups are 2.00 3.50 5.25 9.50

232 Nested and nested-factorial experiments

One should then examine head 5 as its mean is larger than two of the
other three.
For machine C means are 2.50, 4.50, 5.00, 8.50 for heads 9, 12, 11, 10,
and since the largest gap is 8.5 — 2.5 = 6, which is less than 6.13, no signifi-
cant differences are detected. This is so because the differences are significant
at the 10 percent level and not at the 5 percent level on which Table E.l is
based. One might still examine the heads on machine C for lack of consis-
tency. In a nested model it is also seen that the 5 SS between heads within
each machine are “pooled,” or added, to give 282.87. This assumes that these
sums of squares (which are proportional to variances) within each machine
are of about the same magnitude. This might be questioned, as these sums
of squares are 50.19, 126.18, 74.75, 6.50, and 25.25. If these five are really
different, it appears as if the greatest variability is in machine B. This should
indicate the need for work on each machine with an aim toward more
homogeneous strain readings between heads on each machine. This example
shows the importance of recognizing the difference between a nested experi-
ment and a factorial experiment.

11.3 ANOVA RATIONALE

To see that the sums of squares computed in Section 11.2 were correct,
consider the nested model

Yijk = + ^kiij)
or
Yijk = fi + (/9.. - /^) + - I^L.) + {Yijk -

which is an identity.
Using the best estimates of these population means from the sample
data, the sample model is

= y... + (X. - y..) + (5^;. - X.) + (% - yu.)

with
i = I, 2,a j = 1, 2,b k = I, 2,n
Transposing F .. to the left of this expression, squaring both sides, and adding
over j, and k gives

i j k
- y.f = I
i= 1
- T,y + ti,'kyu. -
i j

' j k

as the sums of cross products equal zero. This expresses the idea that the
total sum of squares is equal to the sum of squares between levels of A, plus
 11.4 Nested-factorial experiments 233

the sum of squares between levels of B within each level of A, plus the sum
of the squares of the errors. The degrees of freedom are

{abn — 1) = (u — 1) + a{b — 1) + ab{n — 1)

Dividing each independent sum of squares by its corresponding degrees of
freedom gives estimates of population variance as usual. For computing
purposes, the sum of squares as given above should be expanded in terms
of totals, with the general results shown in Table 11.6.

Table 11.6 General ANOVA for a Nested Experiment

Source df SS MS

■A Tf SS^
Ai a — \
^ nb nab a — 1
a b rj~i 2 fl
TL SSs
^j(t) a{b - 1)
i j i nb a{b - 1)
abn Q b rj-f 2
SS,
P, ^ . .V ab{n — 1) ZSZ
i J k
-

I J ab{n — 1)

abn rp 2
1
Totals abn — 1 EEL
i j k
-
nab

This is essentially the form followed in the problem of the last section.
Note that the
a b
Tfiij- “ Tf
n
yt nb
I J

a / b ryl
" ij- Tl\
" n nb )

which shows the way in which the sum of squares was calculated in the last
section: by getting the sum of squares between levels of B for each level of
A and then pooling over all levels of A.

11.4 NESTED-FACTORIAL EXPERIMENTS

In many experiments where several factors are involved, some may be fac-
torial or crossed with others; some may be nested within levels of the others.
When both factorial and nested factors appear in the same experiment, it is
known as a nested-factorial experiment. The analysis of such an experiment
234 Nested and nested-factorial experiments

is simply an extension of the methods of Chapter 5 and this chapter. Care

must be exercised, however, in computing some of the interactions. Levels
of both factorial and nested factors may be either fixed or random. The
methods of Chapter 10 can be used to determine the EMS values and the
proper tests to be run.

Example 11.2 The nested-factorial experiment is best explained by an ex-

ample. An investigator wished to improve the number of rounds per minute
that could be fired from a Navy gun. He devised a new loading method,
which he hoped would increase the number of rounds per minute when
compared with the existing method of loading. To test this hypothesis he
needed teams of men to operate the equipment. As the general physique of
a man might affect the speed with which he could handle the loading of the
gun, he chose teams of men in three general groupings—slight, average, and
heavy or rugged men. The classification of such m.en was on the basis of an
Armed Services classification table. He chose three teams at random to
represent each of the three physique groupings. Each team was presented
with the two methods of gun loading in a random order and each team used
each method twice. The model for this experiment was

Yijkm = + Gj + MGij +

where
Mj = methods, i = 1, 2
Gj = groups,; = 1, 2, 3
= teams within groups, k = 1, 2, 3 for all j
£m{ijk) = random error, m = 1, 2 for all k

The EMS values are shown in Table 11.7, which indicates the proper F tests
to run.

Table 11.7 EMS for Gun-Loading Problem

2 3 3 2
F F R R
Source i 1 k m EMS

M, 0 3 3 2 + + 18^^
Gj 2 0 3 2 + 4a} -f 12}>Q

MG,j 0 0 3 2 +
T^kU) 2 1 1 2 + 4a}
^Gku) 0 1 1 2 +
^m(ijk) 1 1 1 1
 11.4 Nested-factorial experiments 235

Table 11.8 Data and ANOVA for Gun-Loading Problem

Group I II III

Team 1 2 3 4 5 6 7 8 9

Method I 20.2 26.2 23.8 22.0 22.6 22.9 23.1 22.9 21.8
24.1 26.9 24.9 23.5 24.6 25.0 22.9 23.7 23.5

Method II 14.2 18.0 12.5 14.1 14.0 13.7 14.1 12.2 12.7
16.2 19.1 15.4 16.1 18.1 16.0 16.1 13.8 15.1

Source df SS MS EMS

M, 1 651.95 651.95 +
Gj 2 16.05 8.02 + A-(7Y ”1“ \^(J)Q
MG,j 2 1.19 0.60 ol -f
6 39.26 6.54 ol + 4(TJ
6 10.72 1.79 ol +
18 41.59
2
2.31 0-£

Totals 35 760.76

The data and analysis of this experiment appear in Table 11.8.

It would be well for the reader to verify that the sums of squares of
this table are correct. The only term that is somewhat different in this model
than those previously handled is that is, the interaction between
methods and teams within groups. The safest way to compute this term is
to compute the M x T interaction sums of squares within each of the three
groups separately and then pool these sums of squares. (See Tables 11.9
through 11.11.)

Table 11.9 Data on Gun-Loading Problem for Group I

Team
Method
1 2 3 Totals

Method I 20.2 26.2 23.8

24.1 26.9 24.9
44.3 53.1 48.7 146.1

Method II 14.2 18.0 12.5

16.2 19.1 15.4
30.4 37.1 27.9 95.4

Team totals 74.7 90.2 76.6 241.5

236 Nested and nested-factorial experiments

Table 11.10 Data on Gun-Loading Problem for Group II

Team
Method
4 5 6 Totals

Method I 22.0 22.6 22.9

23.5 24.6 25.0
45.5 47.2 47.9 140.6

Method II 14.1 14.0 13.7

16.1 18.1 16.0
30.2 32.1 29.7 92.0

Team totals 75.7 79.3 77.6 232.6

Table 11.11 Data on Gun-Loading Problem for Group III

Team
Method
7 8 9 Totals

Method I 23.1 22.9 21.8

22.9 23.7 23.5
46.0 46.6 45.3 137.9

Method II 14.1 12.2 12.7

16.1 13.8 15.1
30.2 26.0 27.8 84.0

Team totals 76.2 72.6 73.1 221.9

To compute the M x T interaction for group I, we have

(44.3)2 + (53.1)2 ^ ^ ^3Q_4)2 ^33^)2 ^ ^7.9)'

SScell —

(241.5)'
= 5116.39 - 4860.21 = 256.18

(146.1)2 -f (95.4)'
SS method 4860.21 - 214.19

(74.7)2 + (90.2)2
SS = 4860.21 - 35.74

SS M X T interaction = 256.18 - 214.19 - 35.74 = 6.25

 11.4 Nested-factorial experiments 237

For Group 11
(45.5)2 + (47.2)2 (479)2 4. + (32.1)^ + (29.7)^
SS cell
2
(232.6)'
= 199.96
12“

(140.6)2 + (92.0)'
SS method - 4508.56 = 196.83

(75.7)2 + (79.3)2 + (77.6)'

SS = 4508.56 = 1.62

SS M X T interaction 199.96 - 196.83 - 1.62 = 1.51

For Group 111

(46.0)" + (46.6)" + (45.3)" + (30.2)" + (26.0)" + (27.8)'

SS cell
2
(221.9)'
246.96
ir~
(137.9)2 + (84.0)'
SS method - 4103.30 = 242.10

(76.2)2 + (72.6)2 + (73.1)'

SS team 4103.30 = 1.90

SS M X r interaction 246.96 - 242.10 - 1.90 = 2.96

Pooling for all three groups gives

SSMXT = 6.25 + 1.51 + 2.96 = 10.72

which is recorded in Table 11.8.

The results of this experiment show a very significant method effect (the
new method averaged 23.58 rounds per minute, and the old method averaged
only 15.08 rounds per minute). The results also show a significant difference
between teams within groups at the 5 percent significance level. This points
up individual differences in the men. No other effects or interactions were
significant.
A further analysis of the significant differences between teams within
the three groups shows the sums of squares to be 35.74,1.62, and 1.90 respec-
tively. With the error mean square in Table 11.8 of 2.31 it is obvious that the
significant difference between teams within the three groups is concentrated
in group I. The mean number of rounds per minute for group Ts three teams
 238 Nested and nested-factorial experiments

(across methods) are

Team 13 2
Mean 18.68 19.15 22.55

A Newman-Keuls will show that team 2 is exceptionally faster than the

other two teams.

11.5 COMPUTER PROGRAMS

A nested or nested-factorial experiment, of course, can be analyzed by a
computer program. Many such program packages do not have one for a
nested experiment. If this is the case, one can feed the data into a factorial
program as shown in Chapter 5 and then combine appropriate terms to
give the proper sums of squares for the nested terms.
Recall in Example 11.1 concerning machines and heads within machines
that if it were (erroneously) analyzed as a factorial, the results would be as
given in Table 11.1. However, what appears as an interaction is really part
of the between-heads, within-machines term. Combining Hj and MHi/s SS
gives 46.45 + 236.42 = 282.87 with 3 + 12 = 15 df as given in Table 11.4.
Such a procedure can be extended to the nested factorial.
For the example on gun loading (Example 11.2) the data were entered
into an SPSS program taking teams as 1, 2, and 3 in each group and the
resulting ANOVA is as shown in Table 11.12. In Table 11.12 if teams (T)

Table 11.12 Computer Printout of Example 11.2

SUM OF MEAN SIGNIFICANCE

SOURCE OF VARIATION SQUARES DF SQUARE F OFF

MAIN EFFECTS 680.774 5 136.155 58.927 .001

G 16.052 2 8.026 3.474 .053
T 12.772 2 6.386 2.764 .090
M 651.951 1 651.951 282.162 .001
2-WAY INTERACTIONS 33.234 8 4.154 1.798 .143
G T 26.487 4 6.622 2.866 .053
G M 1.187 2 .594 .257 .776
T M 5.561 2 2.780 1.203 .323
3-WAY INTERACTIONS 5.161 4 1.290 .558 .696
G T M 5.161 4 1.290 .558 .696
EXPLAINED 719.170 17 42.304 18.309 .001
RESIDUAL 41.590 18 2.311
TOTAL 760.760 35 21.736
 11.6 Repeated-measures design and nested-factorial experiments 239

and the G x T interaction SS are combined: 12.772 + 26.487 = 39.259 as

the teams within groups, which checks with Table 11.8. To get the methods
by teams within groups interaction, combine M x T with
M X T X G or SS: 5.561 + 5.161 = 10.722, which is the same as the value
in Table 11.8.
The computer printout will also give means computed in all combina-
tions of effects for further analysis if needed.

11.6 REPEATED-MEASURES DESIGN AND

NESTED-FACTORIAL EXPERIMENTS

Many statisticians who work with psychologists and educators on the design
of their experiments treat repeated-measures designs as a unique topic in
these fields of application of statistics. It is the purpose of this section to
show that these designs are but a special case of factorial and nested-factorial
experiments. Numerical examples and mathematical models are used to
illustrate the correspondence between these designs.
One of the simplest of these cases is when a pretest and posttest are
given to the same group of subjects after a certain time lapse during which
some special instruction may have been administered.

Example 11.3 A recent study at Purdue University was ma^de on a measure-

ment of physical strength (in pounds) given to seven subjects before and
after a specified training period. Results are shown in Table 11.13.

Table 11.13 Pretest and Posttest

Measures for
Example 11.3

Subjects Pretest Posttest

1 100 115
2 110 125
3 90 105
4 no 130
5 125 140
6 130 140
7 105 125

Since the same seven subjects were given each test, we have two repeated
measures on each subject. Winer [24, p. 266] would handle this as a between-
subjects, within-subjects repeated-measures design and report results as
shown in Table 11.14.
240 Nested and nested-factorial experiments

Table 11.14 ANOVA for Table 11.13

Source
of Variation df SS MS F

Between subjects (S,) 6 2084.71 347.45

Within subjects 7 901.00 —

Tests {Tj) 1 864.29 864.29 145

Residual 6 35.71 5.96

Totals 13 2985.71

This analysis implies that the subjects model is

^ ^ + ^jii)
withdf: 6 7

And then the within-subjects’ data are further broken down into tests and
residual. This residual is actually the subject by test interaction so the
model becomes
Yij = ^ + Tj + STij

with df: 6 16

This model contains no error term since only one pretest score and one post-
test score are available on each subject.
Now if this problem were considered in more general terms instead of
the so-called repeated-measures format, one would consider this data layout
as a two-factor factorial experiment with one observation per treatment.
Subjects should be chosen at random and tests should be fixed. Using the
algorithm given in Chapter 10, the resulting ANOVA layout would be as
shown in Table 11.15.
Here k = 1 and a separate error is not retrievable. A glance at the
EMS column indicates that the only proper F test is to compare the test
mean square with the S x T interaction mean square, which was the test
as shown in Table 11.14.

Table 11.15 EMS Determination for Table 11.13

7 2 1
R F R
Source df i i k EMS

5, 6 1 2 1 ol + IGI
TJ 1 7 0 1 T ^ST T
S7-,, 6 1 0 1 ol + ojj
0 1 1 1 (Tg (not retrievable)
 11.6 Repeated-measures design and nested-factorial experiments 241

Of course, another approach to this example is to take differences

between post- and pretest scores on each individual and test the hypothesis
that the true mean of the difference is zero. Using this scheme, a t of 12.05
is found and it is quite well known that this t when squared equals the F
of our ANOVA table:

= (12.05)^ = 145 = F
Example 11.4 A slightly more complex problem involves three factors
where the subjects are nested within groups. These groups could be classes
or experimental conditions and then again each subject is subjected to
repeated measures. The study cited on physical strength was extended so
that three groups of subjects were involved, two being subjected to special
experimental training and the third acting as a control with no special
training. Again each subject was given a pre- and posttest where the
measurement here was velocity of a baseball throw in meters per second.
The resulting data are given in Table 11.16. If these data are treated as a

Table 11.16 Pretest and Posttest Throwing

Velocities of Three Groups of
Subjects in Meters/Second

Group Subject Pretest Posttest

1 26.25 29.50
2 24.33 27.62
3 22.52 25.71
I
4 29.33 31.55
5 28.90 31.35
6 25.13 29.07
7 29.33 31.15

8 27.47 28.74
9 25.19 26.11
10 23.53 25.45
II 11 24.57 25.58
12 26.88 27.70
13 27.86 28.82
14 28.09 28.99

15 22.27 22.52
16 21.55 21.79
17 23.31 23.53
18 30.03 30.21
III
19 28.17 28.65
20 28.09 28.33
21 27.55 27.86
242 Nested and nested-factorial experiments

Table 11.17 ANOVA of Table 11.16

Source df SS MS EMS F

Between subjects 20 271.05

Groups (Gi) 2 28.14 14.07 + 2cr| + i4(f)Q 1.04
Subjects within
groups 18 242.91 13.50 + 2(T|
Within subjects 21 35.73
Tests (U) 1 21.26 21.26 + G^s T 2107’ 183***
G X T 2 12.38 6.19 + <^TS + 70Gr 53**
T X Skj(i) 18 2.09 0.116 GI + <^TS

repeated-measures design, the results are usually presented as in Winer

[24, p. 520], here shown as Table 11.17.
If the layout of Table 11.16 is treated as a nested-factorial experiment
where subjects are nested within groups and then tests are factorial on both
groups and subjects, the model can be written as follows:

Yijk = fi + Gi + + Tfc + GTii, +

withdf: 2 18 1 2 18 0
including a nonretrievable error term. Use of this model and the algorithm
cited above gives Table 11.18.

Table 11.18 EMS Determination for Table 11.16

F R F R
3 1 2 1
Source df i i k m EMS

G, 2 0 1 2 1 G^ + 2(7| + 140G
^j(i} 18 1 1 2 1 GI + 2cr|
u 1 3 7 0 1 + ^TS + 210^
GT,, 2 0 7 0 1 Gg + (7^5 + l<pQj
18 1 1 0 1 G^ + Gjg
0 1 1 1 1 Gg (not retrievable)

This scheme generates the EMS column that was simply reported in
Table 11.17, and that column indicates the proper tests to be made. The
numerical results are, of course, the same.
The nested-factorial experiment covers many situations in addition to
the repeated-measures design. The factor in the nest may be farms within
 11.6 Repeated-measures design and nested-factorial experiments 243

townships, classes within schools, heads within machines, samples within

batches, and so on. It treats all situations where three factors are involved
where one is nested and another is factorial. Usually in a repeated-measures
design the subjects are chosen at random and the other two factors are
considered as fixed. By setting up the model and using the EMS algorithm
one may have factors that are random as well as fixed. For example, classes
could very well be chosen at random from a given grade level and then the
students within the classes chosen at random for the experiment.
As more factors are added in an experiment, one only has to expand the
mathematical model and determine the proper tests to make based on the
EMS column. One does not need to think in terms of repeat measures. To
illustrate further, Winer [24, p. 540] considers two cases of three-factor
experiments with repeated measures. Case I has two of the factors (fixed)
crossing (or repeated on) all subjects and in case II subjects are nested
within two factors (fixed) and the third factor crosses (or is repeated on) all
subjects. Both of these cases, and more complex ones, can easily be handled
as nested-factorial experiments by writing the proper model. Handling them
in this way does not require that the three factors all be fixed. It simply
requires that the subjects be treated as another factor nested within one or
more of the other factors.
Models for case I and case II show the relationship between the two
approaches.

Case I:

between
subjects

within subjects
. ^ s
+ Bk-h ABik + BSkji^i^-\- C^-\- ACim + CSmkjii) + + ABC ikm T BCSj^kj{i)

Case II:

between subjects
, ^,
Yijkm = + AB^j + Sk(ij)

within subjects

+ + ACi^ + BCj^ -t- ABCij^ + CS^kiij)

No error terms have been included since they are not retrievable.
244 Nested and nested-factorial experiments

These examples, it is hoped, are sufficient to remove any mystery

surrounding the repeated-measures designs by showing that they fit into
more general models familiar to statisticians in all fields.

11.7 SUMMARY
The summary at the end of Chapter 6 may now be extended for Part 11.

Experiment Design Analysis

11. Two or more

factors
A. Factorial
(crossed)
1. Completely randomized 1.
^ijk =
+ ABij + ...
for more factors
a. General case a. ANOVA with
interactions
b. 2" case b. Yates’ method or
general ANOVA;
use (1), u, h, ab,. ..
c. 3" case c. General ANOVA; use
00, 10, 20,01, 11,...,
and A X B =
AB + AB\. . .
for interaction
B. Nested
(hierarchical)
1. Completely randomized 1. Nested ANOVA
Tj/c = A + d,-
+
C. Nested
factorial
1. Completely randomized 1. Nested-factorial ANOVA
^ijkm = A< + d,- +
+ Q + AC 11,
+ ^^jk(i) + ^m(ijk)

PROBLEMS

11.1 Porosity readings on condenser paper were recorded for paper from four rolls
taken at random from each of three lots. The results were as follows. Analyze
these data, assuming lots are hxed and rolls random.
 Problems 245

Lot I II III

Roll 1 2 3 4 5 6 7 8 9 10 11 12

1.5 1.5 2.7 3.0 1.9 2.3 1.8 1.9 2.5 3.2 1.4 7.8
1.7 1.6 1.9 2.4 1.5 2.4 2.9 3.5 2.9 5.5 1.5 5.2
1.6 1.7 2.0 2.6 2.1 2.4 4.7 2.8 3.3 7.1 3.4 5.0

11.2 In Problem 11.1 how would the results change (if they do) if the lots were chosen
at random?
11.3 Set up the EMS column and indicate the proper tests to make if /I is a fixed
factor at hve levels, B is nested within A at four random levels for each level
of A, C is nested within B at three random levels, and two observations are
made in each “cell.”
11.4 Repeat Problem 11.3 for A and B crossed or factorial and C nested within the
A, B cells.
11.5 Two types of machines are used to wind coils. One type is hand operated; two
machines of this type are available. The other type is power operated; two
machines of this type are available. Three coils are wound on each machine
from two different wiring stocks. Each coil is then measured for the outside
diameter of the wire at a middle position on the coil. The results were as follows.
(Units are 10“inch.)

Machine Type Hand Power

Machine Number 2 3 5 8

Stock 1 3279 3527 1904 2464

3262 3136 2166 2595
3246 3253 2058 2303
Stock 2 3294 3440 2188 2429
2974 3356 2105 2410
3157 3240 2379 2685

Set up the model for this problem and determine what tests can be run.
11.6 Do a complete analysis of Problem 11.5.
11.7 If the outside diameter readings in Problem 11.5 were taken at three fixed posi-
tions on the coil (tip, middle, and end), what model would now be appropriate
and what tests could be run?
11.8 An experiment is to be designed to compare faculty morale in three types of
junior high school organizations: grades 6-8, grades 7-9, and grades 7-8. Two
schools are randomly selected to represent each of these three types of organiza-
tions and data are to be collected from several teachers in each school. Some-
one suggests that the morale of male and female teachers might differ so it is
246 Nested and nested-factorial experiments

agreed to choose random samples of five male and five female teachers from each
school so that sex differences in morale may be checked out too.
a. Set up a mathematical model for this experiment, outline into ANOVA, and
indicate what tests can be made.
b. Make any recommendations that seem reasonable based on your answers to
part (a), assuming that the data have not been collected as yet.
11.9 In an attempt to study the effectiveness of two corn varieties {V^ and V2), two
corn-producing counties (C) in Iowa were chosen at random. Then four farms
(F) were randomly selected within each county. Seed from both varieties were
sent to these farms and planted in random plots. When harvested, the number
of bushels of corn per acre was recorded for each variety on each farm.
a. Write a mathematical model for this situation.
b. Determine the EMS for varieties in this experiment.
c. Assuming only counties (C) showed significantly different average yields, and
MSc =130 and MS^ = 10 (farms or error), find the percentage of the vari-
ance in this experiment that can be attributed to county differences.
11.10 An educator proposes a new teaching method and wishes to compare the achieve-
ment of students using his method with that of students using a traditional
method. Twenty students are randomly placed into two groups with ten students
per group. Tests are given to all 20 students at the beginning of a semester, at
the end of the semester, and ten weeks after the end of the semester. The educator
wishes to see whether there is a difference in the average achievement between
the two methods at each of the three time periods.
a. Write a mathematical model for this situation.
b. Set up an ANOVA table and show the F tests that can be made.
c. If the educator’s method is really better than the traditional method, what
would you expect if you graphed the results at each time period? Show by
sketch.
d. If method A—the new method—gave a mean achievement score across all
time periods of 60.0, explain in some detail how you would set confidence
limits on this method mean.
11.11 In Example 11.4 in the text the ANOVA of Table 11.17 suggests that further
analysis is needed. Using the data of Table 11.16, make further analyses and
state your conclusions. (A graph might also be helpful.)
11.12 In studying the intellectual attitude toward science as measured on the Bratt
attitude scale, a science educator gave a pretest and posttest to an experimental
group exposed to a new curriculum and a control group that was not so exposed.
There were 15 students in each group. Set up a mathematical model for this
situation and determine what tests are appropriate.
11.13 In Problem 11.12 a strong interaction was found between the groups and the
tests with means as follows: pretest control = 57.00, pretest experimental =
54.96, posttest control = 54.66, posttest experimental = 65.66. Assuming the
sum of squares for the error term used to test for this interaction was 492.61,
make further tests and state some conclusions based on this information.
 Problems 247

11.14 A researcher wished to test the figural fluency of three groups of Egyptian stu-
dents: those taught by the Purdue Creative Thinking Program (PCTP) in a
restricted atmosphere, those taught by PCTP in a permissive atmosphere, and
a control group taught in a traditional manner with no creativity training.
Four classes were randomly assigned to each of the three methods. Before
training began, all pupils were administered a Group Embedded Figures Test
and classified as either field-dependent of field-independent students. After six
weeks of training results on the averages of students within these classes on
this one subtest (figural fluency) were found to be

Source df SS MS F Prob.

Methods (M,) 124.60

Classes within method 35.10
Cognitive style (FJ 2.76
M X F interaction 25.76
F X C interaction {FC,,j^i)) 20.97

Complete the table and state your conclusions.

11.15 Fifty-four randomly selected students are assigned to one of three curricula (C)
of science lessons. Over a period of a semester six lessons (L) were presented to
all students in these three groups according to their prescribed curricula. The
first two curricula were experimental A—taught question-asking skills with
evaluation feedback, and experimental B—taught with question-asking skills
without the evaluation feedback. The third curriculum was used as a control
with no instruction in question-asking skill or feedback. Results on the propor-
tion of high-level questions asked on a test gave

Source df SS MS F Prob.

2 0.556 0.278 4.14 0.0211

^k(i) 51 3.417 0.067
5 0.250 0.050 2.25 0.0493
CL,j 10 0.430 0.043 1.92 0.0431
255 5.610 0.022

Discuss these results and explain how the F tests were made.
11.16 For Problem 11.15 it was known in advance that the experimenter wished to
compare the two experimental curricula results with the control, and also to
compare the two experimental curricula results with each other. Set up proper
a priori tests for this purpose and test the results based on the following means
and the table given in Problem 11.15.
248 Nested and nested-factorial experiments

Mean proportion for curriculum A: 0.2224

Mean proportion for curriculum B : 0.1929
Mean proportion for curriculum C: 0.1236

11.17 In a study made of the characteristics associated with guidance competence

versus counseling competence, 144 students were divided into nine groups of
16 each. These nine groups represented all combinations of three levels of gui-
dance ranking (high, medium, low) and three levels of counseling ranking (high,
medium, low). All subjects were then given nine subtests. Assuming the rankings
as two fixed factors, the subtests as fixed, and the subjects within the nine groups
as random, set up a data layout and mathematical model for this experiment.
11.18 Determine the ANOVA layout and the EMS column for Problem 11.17 and
indicate the proper tests that can be made.
11.19 Three days of sampling where each sample was subjected to two types of size
graders gave the following results, coded by subtracting 4 percent moisture and
multiplying by 10.

Day 1 2 3

Grader A B A B A B

Sample 1 4 11 5 11 0 6
2 6 7 17 13 -1 -2
3 6 10 8 15 2 5
4 13 11 3 14 8 2
5 7 10 14 20 8 6
6 7 11 11 19 4 10
7 14 16 6 11 5 18
8 12 10 11 17 10 13
9 9 12 16 4 16 17
10 6 9 -1 9 8 15
11 8 13 3 14 7 11

Assuming graders fixed, days random, and samples within days random, set up
a mathematical model for this experiment and determine the EMS column.
11.20 Complete the ANOVA for Problem 11.19 and comment on the results.
11.21 In a filling process two random runs are made. For each run six hoppers are
used, but they may not be the same hoppers on the second run. Assume that
hoppers are a random sample of possible hoppers. Data are taken from the
bottom, middle, and top of each hopper in order to check on a possible position
effect on filling. Two observations are made within positions and runs on all
hoppers. Set up a mathematical model for this problem and the associated
EMS column.
11.22 The data below are from the experiment described in Problem 11.21.
 Problems 249

Run 1 Run 2

Head Bottom Middle Top Bottom Middle Top

1 19 0 31 25 25 13 6 6 0 6 0 0
2 13 0 19 38 31 25 0 13 0 31 6 0
3 13 0 31 19 13 0 28 6 25 19 13 28
4 13 0 19 31 25 13 28 0 6 0 28 25
5 19 19 19 38 19 25 16 13 13 19 0 0
6 6 0 0 0 0 0 13 6 6 19 0 19

Do a complete analysis of these data.

11.23 Some research on the abrasive resistance of filled epoxy plastics was carried out
by molding blocks of plastic using five fillers: iron oxide, iron filings, copper,
alumina, and graphite. Two concentration ratios were used for the ratio of filler
to epoxy resin: ji 1 and 1:1. Three sample blocks were made up of each of the
ten combinations above. These blocks were then subjected to a reciprocating
motion where gritcloth was used as the abrasive material in all tests. After
10,000 cycles, each sample block was measured in three places at each of three
fixed positions on the plates: I, II, and III. These blocks had been measured
before cycling so that the difference in thickness was used as the measured
variable. Measurements were made to the nearest 0.0001 inch. Assuming com-
plete randomization of the order of testing of the 30 blocks, set up a mathematical
model for this situation and set up the ANOVA table with an EMS column.
We are interested in the effect of concentration ratio, fillers, and positions on
the block. The three observations at each position will be considered as error,
and there may well be differences between the average of the three sample blocks
within each treatment combination.
11.24 Data for Problem 11.23 were found to be as shown in the following tables.
Complete an ANOVA for these data and state your conclusions.
11.25 From the results of Problem 11.24 what filler-concentration combination would
you recommend if you were interested in a minimum amount of wear? The 1:1
ratio is cheaper than j:l.
 H-( (N p p m p p p NO
1—(
od cK cK P P od P P P od P
m in

10 p p 00 't; P 't

420.2
00
• HH ■>—1 P P —H m P
• l—H r4 nl P P
rn r^i

p p p p p p 1
m 00 p
r^i
HH od r-' od n 't' P p P r~i P o
D in 'it •

a
HH o p p p T—( p 00 p p OO oo
o HH cK P P P 0 P 0 p P P o\
m rn (N

tri p p 00 p p p P ON p

210.7
HH in in NO NO P P P P p o^
-.|<N (N

> oo p ON P P in p 1
p OO <N r^i
od od NO 'It P NO P 0 P P P o
(N <N fN

HH '•O Cs| (N p p 0 m p NO p O)
HH Md MD NO P P P P P P P P r-'
HH
(N

1
p p P p P p 'it p 'll; p p p

156.2
1
i '' G in in 'ij’ 't '=t P 'It P '=t 't P m’
1

o^ in in OO p 00 p NO 'It
CNI
HH 'sd NO 0 NO P 't 06 P P
CNI
c3
•
c:
B h-^ p P 00 p p 00 Csl in
3 HH in P p 't' P P P P P P od
<

P p p ■It m p
163.3

T-H 00 m
G in in in n n P 't P \6 P P

p p NO ON 00 ON <N ON m p p IT)
H—1
r-' NO 0 NO P P P P P P o
nj <N

c
_o '-H CNI m
Totals

+-> (U 0 0
UH "HH ’p ■DH
+-> c c c c
c c c c
a> d d
o '4—> 0:1 00 00
C c/5
o o
U OH

250
 h—(
t ■ p lO p p p 1 oo p p p p
h-H r4 r4 (N 'O I ( T—H T—H r4 rn r4 od

h-H lO p p p p p lO p p p p oo
h-^ T—H 1—H cn T—H ^H 1 ( <ni
lO

oq p p m p p p p p p m
h“H
r4 <N 1 (
Copper
C (N (N T—H MD <N (N (N od

HH oo OO p o p m 1 ^ p T—H p m
HH O rn m' o (N rn (N OO rn rn T—H
HH T-H

- (N oo p p p p p p
HH 1
HH r4 O (N 'vO <N c4 \6 rn

rq p p p p p P p p p
1 ^
(N (N m’ oo’ (N rd r4 rn rn rn cK

l-H p p p p p
1
p p T—H T—H

h-H O 1—H <N r4 lO T—H (N ‘n)

HH m p p p P OO p 1 OO p » ^ p p p
h-H rn O r4 T—H T—H rn
Iron Oxide

HH
oq p p p p p p p
1 ^•
r ■ "H r4 r4 (N (N 'sO T—H T—H

HH (N 1—! rq 'O p P p p p p o^
HH rn rn O r4 m' H H (N (N Md
HH

r'« H
HH oc P p p p 1 p p p p p
-H|(N HH
r4 ini mi
'O

p oc p OO p p p ^H o^ p 'O
m' m’ rc rn ro <N oo (N fd T—H Md

h-H p p p (N p p 1 ^
p p p (N
HH
HH r4 (N <N r^’ 1p rd oi 'O

HH ■vO P p p T—H f ^ o^ 1 ^ p r- p p
H o m' o rn
mi
-
Iron Filings

p oq p m p p m ■^H p
1 ^
HH m’ rn (N O (N (N ' ' 'O r4 <N r4 r-’
T-H

HH m p p r-; p p p p P p p
< !
HH 1 (N r4 lo r4 mi

HH o\ P oo oo oo in T—H p
-i|(N HH
1—H O m' o (N o o o fN T—H

P p p p m 1 oo p p p p
r4 r4 1—H T—H ^H P rn r4 H od
Concentration

Sample 2

Sample 3
Sample 1

Totals
Position

251
 12
Experiments of Two or More Factors —

Restrictions on Randomization

12.1 INTRODUCTION
In the discussion of factorial and nested experiments it was assumed that
the whole experiment was performed in a completely randomized manner.
In practice, however, it may not be feasible to run the entire experiment
several times in one day, or by one experimenter, and so forth. It then becomes
necessary to restrict this complete randomization and block the experiment
in the same manner that a single-factor experiment was blocked in Chapter 4.
Instead of running several replications of the experiment all at one time,
it may be possible to run one complete replication on one day, a second
complete replication on another day, a third replication on a third day, and
so on. In this case each replication is a block, and the design is a randomized
block design with a complete factorial or nested experiment randomized
within each block.
Occasionally a second restriction on randomization is necessary, in
which case a Latin square design might be used, with the treatments in the
square representing a complete factorial or nested experiment.

12.2 FACTORIAL EXPERIMENT IN A RANDOMIZED

BLOCK DESIGN
Just as the factorial arrangements were extracted from the treatment effect
in Equations (5.1) and (5.2), so can these arrangements be extracted from the
treatment effect in the randomized complete block design of Chapter 4
(Section 4.2). For the completely randomized design, you will recall that

kjj ^ 4“ '^ij ^k(ij)

can be subdivided into

^ijk = f^i + Ai -\- Bj -f ABij -f 8^ij^

where A,- + Bj + ABij = of the first model.

252
 12.2 Factorial experiment in a randomized block design 253

When a complete experiment is replicated several times and Rj^ repre-

sents the blocks or replications, the randomized block model is

^ijkm 1
““ Rfi '^ij 4 “ £,n(ijk)

where Rj^ represents the blocks or replications and ijj represents the treat-
ments. If the treatments are formed from a two-factor factorial experiment,
then
^ij =

and the complete model is

^ijk = + Bj + ABij + Sijj^ (12.1)

Here it is assumed that there is no interaction between replications (blocks)
and treatments, which is the usual assumption underlying a randomized
block design. Any such interaction is confounded in the error term
To get a more concrete picture of this type of design, consider three levels
of factor A, two levels of B, and four replications. The treatment combinations
may be written as A^^B^, ^1^2, AjB^, A2B2, A^B^, and A2B2. Such a design
assumes that all six of these treatment combinations can be run on a given
day, if days are the replications or blocks. A layout in which these six treat-
ment combinations are randomized within each replication might be

Replication I: A^Bi A^B^ ^3^2 A2B

Replication II: A,B, A2B2 A2B1 A1B2 A^B
Replication III: ^2^1 A2B2 A1B2 A^B^ ^2^2 A^B
Replication IV: A,B, ^3^1 A2B2 ^2^1 A1B2 A2B

An analysis breakdown would be that shown in Table 12.1.

To see how the interaction of replications and treatments is used as error,
consider the expanded model and the corresponding EMS terms exhibited

Table 12.1 Two-Factor Experiment in a

Randomized Block Design

Source df

Replications 3
Treatments 5
Ai 2
1
AB,^ 2
Error eiji, 15

Total 23
254 Experiments of two or more factors—restrictions on randomization

in Table 12.2. Here the factors are fixed and replications are considered as
random.

Table 12.2 EMS Terms for Two-Factor Experiment in a Randomized Block

3 2 4 1
F F R R
Source df i j k m EMS

3 3 2 1 1 + 6o-|
At 2 0 2 4 1 + 2(T|^ + A
RA,, 6 0 2 1 1 + 2cr|^
Bj 1 3 0 4 1 G^ + 3cr^g + 12 (t)B
RBjk 3 3 0 1 1 GI + 2G\B
AB,j 2 0 0 4 1 + ^RAB + ^(I^AB
RABijk 6 0 0 1 1 + G^^B
^m(ijk) 0 1 1 1 1 G^ (not retrievable)

Total 23

Since the degrees of freedom are quite low for the tests indicated above,
and since there is no separate estimate of error variance, it is customary to
assume that = (^RAB = 0 to pool these three terms to serve
as the error variance. The reduced model is then as shown in Table 12.3.
Here all effects are tested against the 15-df error term. Although we usually
pool the interactions of replications with treatment effects for the error term,
this is not always necessary. It will depend upon the degrees of freedom
available for testing various hypotheses in Table 12.2, and also whether or
not any repeat measurements can be taken within a replication. If the latter
is possible, a separate estimate of aj is available and all tests in Table 12.2
may be made.

Table 12.3 EMS Terms for Two-Factor Experiment,

Randomized Block Design, Reduced Model

Source df EMS

Rk 3 erf + 6o-|
At 2 + 8^^
Bj 1 + 12^5
AB,j 2 + ^^AB
^ijk 15

Total 23
 12.2 Factorial experiment in a randomized block design 255

Example 12.1 At Purdue a researcher was interested in determining the

effect of both nozzle types and operators on the rate of fluid flow in cubic
centimeters through these nozzles. The researcher considered three fixed
nozzle types and five randomly chosen operators. Each of these 15 combina-
tions was to be run in a random order on each of three days. These three days
will be considered as three replications of the complete 3x5 factorial
experiment. After subtracting 96.0 cm^ from each reading and multiplying
all readings by 10, the data layout and observed readings (in cubic centi-
meters) are as recorded in Table 12.4.

Table 12.4 Nozzle Example Data

Operator 1 2 3 4 5

Nozzle A B C A B C A B C A B C A B C

Replication
I 6 13 10 26 4 -35 11 17 11 21 -5 12 25 15 -4

II 6 6 10 12 4 0 4 10 -10 14 2 -2 18 8 10

III -15 13 -11 5 11 -14 4 17 -17 7 -5 -16 25 1 24

The model for this example is

^ijk = + ^7 + NOij Eij^

with
/c = 1, 2, 3 i = 1, 2, 3 j == 1, 2,.. ., 5

The levels are random, Oj levels are random, and levels are fixed. If m
repeat measurements are considered in the model, but m = i, the EMS
values are those appearing in Table 12.5.

Table 12.5 EMS for Nozzle Example

3 5 3 1
F R R R
Source df i j k m EMS

Rk 2 3 5 1 1 15(T|
Ni 2 0 5 3 1 G^ + 3crj5o + 15^;^^
Oj 4 3 1 3 1 G^, + 9Gi
NO,j 8 0 1 3 1 G^ + 2G^O
^majk) 28 1 1 1 1 <^1

Total 44
256 Experiments of two or more factors—restrictions on randomization

These EMS values indicate that all main effects and the interaction can
be tested with reasonable precision (df). The sums of squares are computed
in the usual manner from the subtotals in Table 12.6.

Table 12.6 Subtotals in Nozzle Example

Operator
Nozzle
Replication Total Nozzle 1 2 3 4 5 Total

I 127 A -3 43 19 42 68 169
II 92 B 32 19 44 -8 24 111
III 29 C 9 -49 -16 -6 30 -32

Totals 248 38 13 47 28 122 248

Using these subtotals, the usual sums of squares can easily be computed
and the results displayed as in Table 12.7.

Table 12.7 ANOVA for Nozzle Example

Source df SS MS EMS F FQ.OS

Rk 2 328.84 164.42 GI + \5GI 1.70 3.34

Ni 2 1426.97 713.48 3.13 4.46
Oj 4 798.79 199.70 G^ + 9GI 2.06 2.71
NO,j 8 1821.48 227.68 2.35* 2.29
^ijk 28 2709.16 96.76

Totals 44 7085.24

The results of tests of hypotheses on this model show only a significant

nozzle X operator interaction at the 5 percent level of significance. The plot
in Figure 12.1 of all totals for the right-hand data of Table 12.6 shows this
interaction graphically.
If desirable, the model of this example could be expanded to include an
estimate of replication by operator and replication by nozzle interactions,
using the three-way interaction {N x O x Replication) as the error. The
results of such a breakdown are recorded in Table 12.8.
The F test values are shown in Table 12.8 except for the test on nozzles
labeled with an asterisk. Since there is no direct test, an F' test is used as
described in Chapter 10. Here

MS = MS^K + - MS^^o
 12.2 Factorial experiment in a randomized block design 257

nozzle A

Cell
Totals

Figure 12.1 Interaction plot of nozzle x operator interaction in nozzle example.

Table 12.8 ANOVA for Nozzle Example with Replication Interactions

Source df SS MS EMS F

2 328.84 164.42 (T^ + + \5al 1.86

Ni 2 1426.97 713.48 + SGNR + 15 ^N *
NR,, 4 272.23 68.06 ^E + 5aNR <1
Oj 4 798.79 199.70 + + 9aS 2.26
ROj, 8 705.60 88.20 ^E + <1
NO,j 8 1821.48 227.68 ^E + ^GNO 2.10
^ijk or
NROij, 16 1731.33 108.21 ^E

Totals 44 7085.24

Operator Nozzle

Replication 1 2 3 4 5 Total A B C Total

I 29 -5 39 28 36 127 89 44 --6 111

II 22 16 4 14 36 92 54 30 8 92
III -13 2 4-14 50 29 26 37 - 34 29

Totals 38 13 47 28 122 248 169 111 - 32 248

Here the coefficients are 1,1, and — 1, and the corresponding mean squares
and degrees of freedom are

^^NR — 68.06 with Ti — 4 df

MSNO = 227.68 with V2 = 8 df

^^NRO — 108.21 with = 16 df

258 Experiments of two or more factors—restrictions on randomization

The F' test is then

MS^v
F
~ ^^NRO

713.48
68.06 + 227.68 - 108.21

713.48
3.80
187.53
and

^ (187.53)^
" (l)^[(68.06)V4] + (1)^[(227.68)V8] + (-1)^[(108.21)V16]

With 2 and 4 df the 5 percent F value is 6.94, and with 2 and 5 df it is 5.79;
hence the F' = 3.80 is not significant at the 5 percent level. Hence all F
tests indicate no significant effects. The N x O interaction does not show
significance as it did in Table 12.7, because the test is less sensitive due to the
reduction in degrees of freedom from 28 to 16 in the error mean square.
Since both the N x R and R x O interaction F tests are less than one, we
are probably justified in assuming that and aj^o ^re zero and in pooling
these with for a 28-df error mean square as in Table 12.7.

12.3 FACTORIAL EXPERIMENT IN A LATIN SQUARE DESIGN

If there are two restrictions on the randomization, and an equal number of
restriction levels and treatment levels are used. Chapter 4 has shown that a
Latin square design may be used. The model for this design is

^ijk = F F Ri Tj Jk F Sijj.

where Xj represents the treatments, and both R^ and y^. represent restrictions
on the randomization. If a 4 x 4 Latin square is used for a given experiment,
such as the one in Section 4.5, the four treatments A, B, C, and D could repre-
sent the four treatment combinations of a 2 x 2 factorial (Table 12.9). Here

L ^ F Bq F ABf^^q

Table 12.9 4x4 Latin Square (Same as Table 4.12)

Car
Position I II III IV

1 C D A B
2 B C D A
3 A B C D
4 D A B C
 12.4 Remarks 259

with its 3 df. Thus a factorial experiment {A x B) is run in a Latin square

design. For the problem of Section 4.5, instead of four tire brands, the four
treatments could consist of two brands {A^ and A ) and two types and 2

B ). The types might be black- and white-walled tires. The complete model
2

would then be

^ikmq = /^ + + 7;c + + ^ikmq

and the degree-of-freedom breakdown would be that shown in Table 12.10.

This idea could be extended to running a factorial in a Graeco-Latin
square, and so on.

Table 12.10 Degree-of-Freedom Breakdown

for Table 12.9

Source df

R. 3
yk 3
Am i]1
Rq 1 y 3 df for T :
ij1
c
^ikmq 6

Total 15

12.4 REMARKS

In the preceding sections factorial experiments were considered where the

design of the experiment was a randomized block or a Latin square. It is
also possible to run a nested experiment or a nested-factorial experiment in a
randomized block or Latin square design. In the case of a nested experiment,
the treatment effect might be broken down as

and when this experiment has one restriction on the randomization, the
model is
^ijk = + Fj +
^ -j
V

If a nested factorial is repeated on several different days, the model would be

^ijkm F Rj^ T zlj Rj(i) F “1“ AC^j^ “h “H ^ijkm

where the whole nested factorial is run in a randomized block design. The
analysis of such designs follows from the methods of Chapter 11.
260 Experiments of two or more factors—restrictions on randomization

12.5 SUMMARY

The summary at the end of Chapter 11 may now be extended.

Experiment Design Analysis

II. Two or more

factors
A. Factorial
(crossed)
1. Completely randomized
Tjfc = ^ + Bj
+ AB-j + . . .

for more factors

a. General case a. ANOVA with
interactions
b. 2^ case b. Yates’ method or
general ANOVA
c. 3-^ case c. General ANOVA
2. Randomized block
a. Complete a. Factorial ANOVA with
Yijk = + -Rfc + Aj replications R,
+ Bj + AB^j + Sij,^

3. Latin square
a. Complete a. Factorial ANOVA with
Yijkm = B + + replications and
+ Aj- + Bj + AB^j positions
d" jkm
B. Nested
(hierarchical)
1. Completely randomized 1. Nested ANOVA
^ijk = B + A,-
+ Bjii) + h{ij)
2. Randomized block 2. Nested ANOVA with
blocks R,
a. Complete a. Nested ANOVA with
TjVc = + ^/c + A,- blocks R,
+ + ^ijk
3. Latin square 3.
a. Complete a. Nested ANOVA with
Tj/cm ^ B + J^k + Jm blocks and positions
+ A,- + Bj^i^ +
C. Nested
factorial
1. Completely randomized 1. Nested-factorial ANOVA
bjfcm = /^ + A,■ +
+ C, + AC,,
+ ^^kj(i) + ^m{ijk)
 Problems 261

Experiment Design Analysis

2. Randomized block 2.
a. Complete a. ANOVA with blocks

j{i) T £,•
3. Latin square 3.
a. Complete a. ANOVA with blocks
and positions

+ + BCqji^i^
T £(' jkmq

PROBLEMS

12.1 Data on the glass rating of tubes taken from two fixed stations and three shifts are
reeorded each week for three weeks. Considering the weeks as blocks, the six
treatment combinations (two stations by three shifts) were tested in a random
order each week, with results as follows:

Station 1 Station 2

Shift 1 2 3 1 2 3

Week 1 3 3 3 6 3 6
6 4 6 8 9 8
6 7 7 11 11 13

Week 2 14 8 11 4 15 4
16 8 12 6 15 7
19 9 17 7 17 10

Week 3 2 2 2 2 2 10
3 3 4 5 4 12
6 4 6 7 6 13

Analyze these data as a factorial run in a randomized block design.

12.2 In the example of Section 4.5 (Table 4.13) consider tire brands A, B, C, D as four
combinations of two factors, ply of tires and type of tread, where

A = P,T, B = P,T2 C = P^T, D = P2T2

representing the four combinations of two ply values and two tread types. Analyze
the data for this revised design—a 2^ factorial run in a Latin square design.
262 Experiments of two or more factors—restrictions on randomization

12.3 If in Problem 11.5 the stock factor is replaced by “days,” considered random,
and the rest of the experiment is performed in a random manner on each of the
two days, use the same numerical results and reanalyze with this restriction.
12.4 Again, in Problem 11.5 assume that the whole experiment (24 readings) was
repeated on five randomly selected days. Set up an outline of this experiment,
including the EMS column, df, and so on.
12.5 Three complete replications were run in a study of aluminum thickness. Samples
were taken from five machines and two types of slug—rivet and staple—in each
replicate. Considering replicates random and machines and type of slug fixed,
set up a model for this problem and determine what tests can be made. Comment
on the adequacy of the various tests. Data are as follows:

Machine

Replication Slug 1 2 3 4 5

I Rivet 175 95 180 170 155

Staple 165 165 175 185 130

II Rivet 190 185 180 200 190

Staple 170 160 175 165 190

III Rivet 185 165 175 195 200

Staple 190 160 200 185 200

12.6 Analyze the data of Problem 12.5 using a computer program and state your
conclusions.
12.7 Data were collected for three consecutive years on the number of days after
planting that the first flowering of a plant occurred. Each year five varieties of
plants were planted at six different stations. Considering the repeat years random
and varieties and stations fixed, outline the ANOVA for this problem.
12.8 The data for Problem 12.7 are given below.

Variety

Station/Year Hazel Coltsfoot Anemone Blackthorn Mustard

Broadchalke
1932 57 67 95 102 123
1933 46 72 90 88 101
1934 28 66 89 109 113
Total 131 205 274 299 337
 Problems 263

Variety

Station/Year Hazel Coltsfoot Anemone Blackthorn Mustard

Bratton
1932 26 44 92 96 93
1933 38 68 89 89 no
1934 20 64 106 106 115
Total 84 176 287 291 318

Lenham
1932 48 61 78 99 113
1933 35 60 89 87 109
1934 48 75 95 113 111
Total 131 196 262 299 333

Dorstone
1932 50 68 85 117 124
1933 37 65 74 93 102
1934 19 61 80 107 118
Total 106 194 239 317 344

Coaley
1932 23 74 105 103 120
1933 36 47 85 90 101
1934 18 69 85 105 111
Total 77 190 275 298 332

Ipswich
1932 39 57 91 102 112
1933 39 61 82 93 104
1934 43 61 98 98 112
Total 121 179 271 293 328

Do a complete ANOVA on these data and discuss the results.

12.9 Puffed cereal was packaged from a special machine with six heads and the
hlling could come from any one of three positions—bottom, middle, or bottom
of a hopper. Data were collected on a measure of filling capacity from each of
these 18 possible combinations of head and position. Measurements were taken
on each of two cycles within each of the combinations above. These cycles within
treatments can be considered random. The whole experiment was repeated in a
second run at a later date. Assuming runs are random, set up an analysis of vari-
ance table for this problem and also make up a data layout table for the collection
of the data.
264 Experiments of two or more factors—restrictions on randomization

12.10 Data for Problem 12.9 are given below for two runs, three positions {B, M, T),
six heads (H), and two cycles (C) with each combination:

Run 1 Run 2

B M T B M T

H c, C2 Cl C2 Cl C2 Cl C2 Cl C2 Cl C2
1 19 0 31 25 25 13 6 6 0 6 0 0
2 13 0 19 38 31 25 0 13 0 31 6 0
3 13 0 31 19 13 0 28 6 25 19 13 28
4 13 0 19 31 25 13 28 0 6 0 28 25
5 19 19 19 38 19 25 16 13 13 19 0 0
6 6 0 0 0 0 0 13 6 6 19 0 19

Do a complete analysis of these data. Explain whether or not you could combine
terms in the table as done in the text example.
12.11 Discuss the similarities and differences between a nested experiment with two
factors and a randomized block design with one factor.
12.12 Assuming the nested experiment involving machines and heads in Section 11.2
was completely repeated on three subsequent randomly selected days, outline its
mathematical model and ANOVA table complete with the EMS column.
12.13 For Problem 12.12 show a complete model with all interactions and the reduced
model. Make suggestions regarding how this experiment might be made less
costly and still give about the same information.
12.14 For the nested-factorial example in Section 11.4 on gun loading set up the model
assuming the whole experiment is repeated on four more randomly selected days.
12.15 Examine both the full and reduced models for Problem 12.14 and comment.
 13
Factorial Experiment —

Split-Plot Design

13.1 INTRODUCTION
In many experiments where a factorial arrangement is desired, it may not be
possible to randomize completely the order of experimentation. In the last
chapter restrictions on randomization were considered, and both random-
ized block and Latin square designs were discussed. There are still many
practical situations in which it is not at all feasible to even randomize within
a block. Under certain conditions these restrictions will lead to a split-plot
design. An example will show why such designs are quite common.
Example 13.1 The data in Table 13.1 were compiled on the effect of oven
temperature T and baking time B on the life Y of an electrical component
[25].
Looking only at this table of data, we might think of this as a 4 x 3
factorial with three replications per cell and proceed to the analysis in

Table 13.1 Electrical Component Life-Test Data

Baking
Oven Temperature T (°F)
Time B
(min) 580 600 620 640

5 217 158 229 223

188 126 160 201
162 122 167 182
10 233 138 186 227
201 130 170 181
170 185 181 201
15 175 152 155 156
195 147 161 172
213 180 182 199

265
266 Factorial experiment—split-plot design

Table 13.2 ANOVA for Electrical Component

Life-Test Data as a Factorial

Source df SS MS EMS

T 3 12,494 4165 aj + 9(f) T

2 566 283 + 12(/>B

BTii 6 2601 434 (Tg -h 3(f)jB

24 13,670 570

Totals 35 29,331

Table 13.2. Here only the temperature has a significant effect on the life of
the component at the 1 percent significance level.
We might go on now and seek linear, quadratic, and cubic effects of
temperature; but before proceeding, a few questions on design should be
raised. What was the order of experimentation? The data analysis above
assumes a completely randomized design. This means that one of the four
temperatures was chosen at random and the oven was heated to this tem-
perature, then a baking time was chosen at random and an electrical com-
ponent was inserted in the oven and baked for the time selected. After this
run, the whole procedure was repeated until the data were compiled. Now,
was the experiment really conducted in this manner? Of course the answer is
“no.” Once an oven is heated to temperature, all nine components are in-
serted; three are baked for 5 minutes, three for 10 minutes, and three for 15
minutes. We would argue that this is the only practical way to run the ex-
periment. Complete randomization is too impractical as well as too expen-
sive. Fortunately, this restriction on the complete randomization can be
handled in a design called a split-plot design.
The four temperature levels are called plots. They could be called blocks,
but in the previous chapter blocks and replications were used for a complete
rerun of the whole experiment. (The word “plots” has been inherited from
agricultural applications.) In such a setup temperature—a main effect—is
confounded with these plots. If conditions change from one oven temperature
to another, these changes will show up as temperature differences. Thus, in
such a design, a main effect is confounded with plots. This is necessary be-
cause it is often the most practical way to order the experiment. Now once a
temperature has been set up by choosing one of the four temperatures at
random, three components can be placed in the oven and one component
can be baked for 5 minutes, another for 10 minutes, and the third one for
15 minutes. The specific component that is to be baked for 5, 10, and 15
minutes is again decided by a random selection. These three baking-time
levels may be thought of as a splitting of the plot into three parts, one part
for each baking time. This defines the three parts of a main plot that are called
 13.1 Introduction 267

split-plots. Note here that only three components were placed in the oven,
not nine. The temperature is then changed to another level and, three more
components are placed in the oven for 5, 10, and 15 minutes baking time.
This same procedure is followed for all four temperatures; after this the
whole experiment may be replicated. These replications may be run several
days after the initial experiment; in fact, it is often desirable to collect data
from two or three replications and then decide whether more replications
are necessary.
The experiment conducted above was actually run as a split-plot experi-
ment and laid out as shown in Table 13.3.

Table 13.3 Split-Plot Layout for Electrical Component

Life-Test Data

Baking
Oven Temperature T (°F)
Time B
Replication R (min) 580 600 620 640

I 5 217 158 229 223

10 233 138 186 227
15 175 152 155 156

II 5 188 126 160 201

10 201 130 170 181
15 195 147 161 172

III 5 162 122 167 182

10 170 185 181 201
15 213 180 182 199

In Table 13.3 the lines indicate where randomization restrictions have

occurred. First a replication or block is run (double lines), then within a rep-
lication an oven temperature is chosen at random (single vertical line), and
then within that temperature level baking time is randomized. Thus a set of
observations with no lines separating them (such as 158, 138, 152 for tem-
perature 600°, replication I) indicates random order of these three observa-
tions. This notion of using straight lines to show randomization restrictions
was suggested in Chapter 4, Section 4.2. If data are presented as in Table 13.1,
there is no way to tell how they were collected. Were they completely ran-
domized—all 36 readings? Or were there randomization restrictions as
shown in Table 13.3?
Here temperatures are confounded with plots and the R x T cells are
called whole plots. Inside a whole plot, the baking times are applied to one
third of the material. These plots associated with the three baking times are
268 Factorial experiment—split-plot design

called split-plots. Since one main effect is confounded with plots and the other
main effect is not, it is usually desirable to place in the split the main effect
we are most concerned about testing, as this factor is not confounded.
We might think that factor B (baking time) is nested with the main plots
but this is not the case, since the same levels of B are used in all plots. A
model for this experiment would be

^ijk = /^ -\- Ri + Tj RTij Bf^ RBii^ + TBjj^ + RTBijj^

^^

whole plot split plot

The first three variable terms in this model represent the whole plot, and
the RT interaction is often referred to as the whole-plot error. The usual
assumption is that this interaction does not exist, that this term is really an
estimate of the error within the main plot. The last four terms represent the
split-plot, and the RTB interaction is referred to as the split-plot error. Some-
times the RB term is also considered nonexistent and is combined with RTB
as an error term. A separate error term might be obtained if it were feasible
to repeat some observations within the split-plot. The proper EMS values can
be found by considering m repeat measurements where m = 1 (Table 13.4).

Table 13.4 EMS for Split-Plot Electrical Component Life-Test Data

3 4 3 1
R F F R
Source df i j k m EMS

Whole plot Rt 2 1 4 3 1 ol -f MGI

Tj 3 3 0 3 1 + 3(7|J 9(f)J
RTtj 6 1 0 3 1 + 3^jir

Split-plot Bu 2 3 4 0 1 ol F 4a^B -1- 12^5

RBik 4 1 4 0 1
TBjk 6 3 0 0 1 + (^RTB + 3^rB
RTBijj^ 12 1 0 0 1 ^E + ^RTB
— 1 1 1 1 GI (not retrievable)

Total 35

Note in Table 13.4 that the degrees of freedom in the whole plot add
to 11 as there are 12 whole plots (3 replications x 4 temperatures) and the
degrees of freedom in the split-plot add to 24 as there are 2 df within each
whole plot and hence 2x12 within all whole plots.
Since the error mean square cannot be isolated in this experiment,
+ ^RTB is taken as the split-plot error, and is taken as the
 13.1 Introduction 269

whole-plot error. The main effects and interaction of interest TB can be

tested, as seen from the EMS column, although no exact tests exist for the
replication effect nor for interactions involving the replications. This is not
a serious disadvantage for this design, since tests on replication effects are
not of interest but are isolated only to reduce the error variance. The analysis
of the data from Table 13.3 follows the methods given in Chapter 5. The
results are shown in Table 13.5.

Table 13.5 ANOVA for Split-Plot Electrical Component Life-Test

Data

Source df SS MS EMS

Bt 2 1963 982 + 12cri

Tj 3 12,494 4165 + ^CFRT + 9
RTj 6 1774 296 + 3cr|j.

Bu 566 283 + 4(7^B -h 12

BBif^ 4 7021 1755
TBj, 6 2601 434 + ^RTB +
12 2912 243 2 , 2
^E + ^RTB

Totals 35 29,331

Testing the hypothesis of no temperature effect gives

4165
14.1 (significant at the 1 percent level)

Testing for baking time gives

283
< 1 (not significant)
ET^

Testing for TB interaction gives

434
F 6,12 1.79 (not significant at the 5 percent level)
243

No exact tests are available for testing the replication effect or replica-
tion interaction with other factors, but these effects are present only to reduce
the experimental error in this split-plot design.
The results of this split-plot analysis are not too different from the results
using the incorrect method of Table 13.2, but this split-plot design shows the
need for careful consideration of the method of randomization before starting
270 Factorial experiment—split-plot design

an experiment. This split-plot design represents a restriction on the random-

ization over a complete randomization in a factorial experiment.
To emphasize a bit more the difference between a completely random-
ized 4x3 factorial design and a split-plot 4x3 factorial design, one might
let the numbers 1 through 4 on one die (red) represent temperature and the
numbers 1 through 3 on another die (green) represent baking times and then
use the two dice to effect the proper randomization. In the case of complete
randomization both dice are tossed and the first six results are as indicated
below.

Oven Temperature
Baking
oo
o

Time 600° 620° 640°

5 3 4
10 1 5
15 6 2

The numbers 1, 2,..., 6 indicate the order of the experiment: first set tem-
perature at 600, bake for 10 minutes; then set temperature at 620, bake for
15 minutes; and so on. One notes that the order is scattered throughout
the 12 treatment combinations. In the case of the split-plot design a tempera-
ture is randomly chosen (roll the red die) and then the baking times are
chosen at random within that temperature (roll the green die). One such
randomization is shown for the first six combinations as follows;

Oven Temperature
oo
oo
(-/I

Time 600° 620° 640°

5 3 4
10 1 6
15 2 5

In this latter case it should seem more obvious that temperature is con-
founded with the order in which the experiment is run, as three treatments
are run at 580° first, then three are run at 620° second, and so on. It is hoped
that by replicating the whole experiment several times any effect of order
confounded with temperature would average out. Nevertheless, in a split-
plot design a main effect is confounded.
Some statisticians recommend introducing a restriction error into a
model wherever such restrictions occur; see Anderson and McLean [1].
No such term has been introduced here, and so it is strongly recommended
 13.1 Introduction 271

that F tests be run only within the whole plot or within the split-plot and
that mean squares in the whole plot not be compared with mean squares in
the split-plot regardless of the EMS column.
Since this type of design is encountered often in industrial experiments,
another example will be considered.

Example 13.2 A defense-related organization was to study the pulloff force

necessary to separate boxes of chaff from a tape to which they are affixed.
These boxes are made of a cardboard material, approximately 3 by 3 by 1
inches, and are mounted on a strip of cloth tape that has an adhesive backing.
The tape is 2 inches wide, and the boxes are placed 7 inches center to center
on the strip. There are 75 boxes mounted on each strip.
The tape is pulled from the box at a 90° angle as it is wound onto a drum.
During this separation process the portion of the tape still adhering to the
box carries the box onto a platform. The box trips a microswitch, which
energizes a plunger. The plunger then kicks the box out of the machine.
After this problem was discussed with the plant engineers, several factors
were listed that might affect this pulloff force. The most important factors
were temperature and humidity. It was agreed to use three fixed temperature
levels, — 55°C, 25°C, and 55°C, and three fixed humidity levels, 50 percent,
70 percent, and 90 percent. These gave nine basic treatment combinations.
Since there might be differences in pulloff force as a result of the strip selected,
it was decided to choose five different strips at random for use in this experi-
ment. There might also be differences within a strip, so two boxes were chosen
at random and cut from each strip for the test.
The test in the laboratory was accomplished by hand-holding the pack-
age, attaching a spring scale to the strip by means of a hole previously
punched in the strip, and pulling the tape from the package in a direction
perpendicular to the package.
In discussing the design of this experiment it seemed best to set the
climatic condition (a combination of temperature and humidity) at random
from one of the 3x3 = 9 conditions, and then test two boxes from each
of the five strips while these conditions were maintained. Then another of
the nine conditions is set, and the results again determined on two boxes
from each of the five strips. This is then a restriction on the randomization,
and the resulting design is a split-plot design. It was agreed to replicate the
complete experiment four times. A layout for this experiment is shown in
Table 13.6.
In Table 13.6, each replication is a repeat of replication I, but a new
order of randomizing the nine atmospheric conditions is taken in each rep-
plication. In this design atmospheric conditions and replications form the
whole plot and strips are in the split-plot.
 272 Factorial experiment—split-plot design

Table 13.6 Split-Plot Design of Chaff Experiment

Temperature T (°C)
-55 25 55

Humidity H (percent)
Strip S Box 50 70 90 50 70 90 50 70 90

Replication 11 1
2
R 2 3
4
3 5
6
4 7
8
5 9
10

Replication II (repeat as above)

Replication III (repeat as above)
Replication IV (repeat as above)

The model for this design and its associated EMS relations are set up
in Table 13.7.
From the EMS column tests can be made on the effects of replications,
replications by temperature interaction, replications by humidity interac-
tion, replications by temperature by humidity interaction, strips, and strips
by all other factor interactions. Unfortunately no exact tests are available
for the factors of chief importance: temperature, humidity, and temperature
by humidity interaction. Of course, we could first test the hypotheses that
can be tested, and if some of these are not significant at a reasonably high
level (say 25 percent), assume they are nonexistent and then remove these
terms from the EMS column. For example, if the TS interaction can be
assumed as zero (cr< 0), the temperature mean square can be tested against
the RT interaction mean square with 2 and 6 df. If this is not a reasonable
assumption, a pseudo-F {F') test can be used as discussed in Chapter 10.
Unfortunately, data were available only for the first replication of this
experiment, so its complete analysis cannot be given. The method of analysis
is the same as given in Chapter 5, even though this experiment is rather
complicated. It is presented here only to show another actual example of a
split-plot design.
 13.2 A split-split-plot design 273

Table 13.7 EMS for Chaff Experiment

4 3 3 5 2
R F F R R
Source df i j k m Q EMS

Ei 3 1 3 3 5 2 + 18cr^5 + 90ai
Tj 2 4 0 3 5 2 + 6(7^Ts + 24(7 + 30a^Y ■{“ 120^7’
Whole REj 6 1 0 3 5 2 + ^^RTS + 30(7^7’
plot 2 4 3 0 5 2 + ^^RHs + 24cr^5 + 30a^fj -f- 120 ^77
RH,, 6 1 3 0 5 2 + ^^RHS +
THj, 4 4 0 0 5 2 + ^^RTHS + ^^THS + Idcr^JH + 40
RTH,j, 12 1 0 0 5 2 + ^^RTHS + ^^^RTH

Split- 4 4 3 3 1 2 ^E + "f 22(75

plot 12 1 3 3 1 2 + IScT^s
TSj^ 8 4 0 3 1 2 + ^^RTS + 24(7^5
RTS,j^ 24 1 0 3 1 2 + ^^RTS
8 4 3 0 1 2 + ^^RHS + 24(7^5
RHS,„„ 24 1 3 0 1 2 + ^^RHS
THSjkm 16 4 0 0 1 2 + ^^RTHS + ^^THS
RTHS.jkm 48 1 0 0 1 2 + ^^RTHS
180 1 1 1 1 1

Total 359

13.2 A SPLIT-SPLIT-PLOT DESIGN

Example 13.3 In a study of the cure rate index on some samples of rubber,
three laboratories, three temperatures, and three types of mix were involved.
Material for the three mixes was sent to one of the three laboratories where
they ran the experiment on the three mixes at the three temperatures (145°C,
155°C, and 165°C). However, once a temperature was set, all three mixes
were subjected to that temperature and then another temperature was set
and again all three mixes were involved, and finally the third temperature
was set. Material was also sent to the second and third laboratories and
similar experimental procedures were performed. There are therefore two
restrictions on randomization since the laboratory is chosen, then the tem-
perature is chosen, and then mixes can be randomized at that particular
temperature and laboratory.
By complete replication of the whole experiment to achieve four replica-
tions, the three laboratories and four replications form the whole plots. Then
the temperature levels may be randomized at each laboratory and in each
 Table 13.8 Order of First Nine Rubber Cure Rate Index Experiments

Temperature (°C)
145 155 165

Mix Mix Mix

Laboratory ABC ABC ABC

1 3 4 (a)
2 8 1 5 7 9 Completely randomized
3 2 6

1 (b)
2 4 6 1 7 9 3 2 8 5 Split-plot
3

1 (c)
2 8 9 7 3 1 2 4 6 5 Split-split-plot
3

Table 13.9 Rubber Cure Rate Index Data

Temperature (°C)
145° 155 165

Mix Mix Mix

Replication Laboratory A B C A B C A B C

I 1 18.6 14.5 21.1 9.5 7.8 11.2 5.4 5.2 6.3

2 20.0 18.4 22.5 11.4 10.8 13.3 6.8 6.0 7.7

3 19.7 16.3 22.7 9.3 9.1 11.3 6.7 5.7 6.6

II 1 17.0 15.8 20.8 9.4 8.3 10.0 5.3 4.9 6.4

2 20.1 18.1 22.7 11.5 11.1 14.0 6.9 6.1 8.0

3 18.3 16.7 21.9 10.2 9.2 11.0 6.0 5.5 6.5

III 1 18.7 16.5 21.8 9.5 8.9 11.5 5.7 4.3 5.8

2 19.4 16.5 21.5 11.4 9.5 12.0 6.0 5.0 6.6

3 16.8 14.4 19.3 9.8 8.0 10.9 5.0 4.6 5.9

IV 1 18.7 17.6 21.0 10.0 9.1 11.1 5.3 5.2 5.6

2 20.0 16.7 21.3 11.5 9.7 11.5 5.7 5.2 6.3

3 17.1 15.2 19.3 9.5 9.0 11.4 4.8 5.4 5.8

274
 13.2 A split-split-plot design 275

replication to form a split-plot. Then at each temperature-laboratory-

replication combination, the three mixes are randomly applied forming
what is called a split-split-plot indicating that two main effects (laboratory
and temperature) are confounded with blocks.
Table 13.8 shows one possible order of the first nine experiments in one
replication of (a) a completely randomized, (b) a split-plot design, and (c) a
split-split-plot design. The resultant data will usually not reveal how the
experiment was ordered, but the use of lines in the data given in Table 13.9
should help indicate the restrictive order of randomization.
The model and its EMS values are given in Table 13.10.

Table 13.10 EMS for Rubber Cure Rate Index Experiment

4 3 3 3 1
R F F F R
Source df i j k m q EMS

Whole plot Fi 3 1 3 3 3 1 Hal

Lj 2 4 0 3 3 1 + 36
RL,j 6 1 0 3 3 1 +

Split-plot Tu 2 4 3 0 3 1 -f 9(T^J + 36^2’

FF, 6 1 3 0 3 1 d- 9G^T
LTj, 4 4 0 0 3 1 + ^^RLT + 12^2,r
FLT^J, 12 1 0 0 3 1 -h ^^RLT

Split-split-plot 2 4 3 3 0 1 + ^^RM + 36
RM,„ 6 1 3 3 0 1 ^^RM
4 4 0 3 0 1 + ^^RLM +
FLMij„, 12 1 0 3 0 1 + 2(^RLM
4 4 3 0 0 1 + ^^RTM + \2(f)jM
FTMii^jn 12 1 3 0 0 1 + ^^RTM
LTMjj^fn 8 4 0 0 0 1 ^RLTM + ^</>LTM
„2
RLTMijf^^ 24 1 0 0 0 1 -h ^RLTM
^qiijkm) 0 1 1 1 1 1 (not retrievable)

Total 107

The EMS column indicates that F tests can be made on all three fixed
effects and their interactions without a separate error term. Since the effect
of replication is often considered of no great interest and the interaction of
replication with the other factors is often assumed to be nonexistent, this
appears to be a feasible experiment. If examination of the EMS column
should indicate that some F tests have too few degrees of freedom in their
276 Factorial experiment—split-plot design

denominator (some statisticians say less than 6), the experimenter might
consider increasing the number of replications to increase the precision of
the test. In the above experiment, if five replications were used, the whole
plot error RL would have 8 df instead of 6, the split-plot error RLT would
have 16 instead of 12, and so on.
Another technique that is sometimes used to increase the degrees of
freedom is to pool all interactions with replications into the error term for
that section of the table. Here one might pool RT and RLT and RM, RLM
and RTM with RLTM if additional degrees of freedom are believed neces-
sary. One way to handle experiments with several replications is to stop
after two or three replications, compute the results, and see if significance
has been achieved or that the F’s are large even if not significant. Then add
another replication, and so on, increasing the precision (and the cost) of the
experiment in the hope of detecting significant effects if they are there.
The numerical results of the rubber cure rate index are given in Table
13.11.

Table 13.11 ANOVA for Rubber Cure Rate Index

Experiment

Source df SS MS

R 3 9.41 3.14
L 2 40.66 20.33*
RL 6 16.11 2.68

T 2 3119.51 1559.76***
RT 6 2.07 0.34
LT 4 4.94 1.24
RLT 12 7.81 0.65

M 2 145.71 72.86***
RM 6 3.29 0.55
LM 4 0.35 0.09
RLM 12 2.93 0.24
TM 4 43.69 10.92***
RTM 12 2.22 0.18
LTM 8 1.07 0.14
RLTM 24 4.97 0.21

Totals 107 3404.74

Here laboratory effect is significant at the 5 percent level and tempera-

ture, mixes, and temperature-mix interaction are all highly significant
(< 0.001).
 Problems 277

13.3 SUMMARY

The summary of Chapter 12 may now be extended.

Experiment Design Analysis

II. Two or more

factors
A. Factorial
(crossed)
1. Completely randomized 1.
2. Randomized block 2.
a. Complete a. Factorial
Yijk = ^ + Rk + Ai ANOVA
+ Bj ABjj + £ijj^
b. Incomplete, confounding: b.
i. Main effect—split-plot i. Split-plot
Ljfe = + VRi + Aj RAij./ ANOVA

whole plot
+ Bk A- RBn^ + ABj,^ + RTBjjfc

split-plot

PROBLEMS

13.1 As part of the experiment discussed in Example 13.2 a chamber was set at 50
percent relative humidity and two boxes from each of three strips were inserted
and the pulloff force determined. This was repeated for 70 percent and 90 percent,
the order of humidities being randomized. Two replications of the whole experi-
ment were made and the results follow.

Humidity

Replication Strips 50% 70% 90%

1 1.12 1.75 3.50 0.50 1.00 0.75

I 2 1.13 3.50 0.75 1.00 0.50 0.50
3 2.25 3.25 1.75 1.88 1.50 0.00

1 1.75 1.88 1.75 0.75 1.50 0.75

II 2 5.25 5.25 0.75 2.25 1.50 1.50
3 1.50 3.50 1.62 2.50 0.75 0.50
278 Factorial experiment—split-plot design

Set up a split-plot model for this experiment and outline the ANOVA table,
including EMS.
13.2 Analyze Problem 13.1.
13.3 The following data layout was proposed to handle a factorial experiment de-
signed to determine the effects of orifice size and flow rate on environmental
chamber pressure in millimeters of mercury.

Orifice Size (Inches)

Flow
Rate 8 10 12

3.5
4.0
4.5

a. If this experiment is run as a completely randomized design, how might data

be collected to fill in the above table?
b. If this experiment were run as a split-plot design, how might data be collected?
c. Assuming three complete replications of the above run as a split-plot design,
outline its ANOVA table. (You need not include the EMS column.)
d. Suggest how this experiment might be improved upon.
13.4 Interest centered on the yields of alfalfa in tons per acre from three varieties
following four dates of final cutting. A large section of land was divided into
three strips, and each strip was planted with one of the three varieties of alfalfa.
Once a strip had been selected, one fourth of the plants were cut on the first
cutting date, one fourth on the next, and so on, for the four dates studied. This
whole experiment could be replicated several times as necessary by choosing
sections at random.
a. Identify the type of design implied here.
b. Set up an appropriate ANOVA table for this situation with the minimum
number of replications you feel might be adequate to detect differences, if they
exist.
13.5 In a time study two types of stimuli are used—two- and three-dimensional
film—and combinations of stimuli and job are repeated once to give eight strips
of film. The order of these eight is completely randomized on one long film and
presented to five analysts for rating at one sitting. Since this latter represents a
restriction on the randomization, the whole experiment is repeated at three sit-
tings (four in all) and the experiment is considered a split-plot design. Set up a
model for this experiment and indicate the tests to be made. (Consider analysts
as random ; see also Problem 5.18.)
13.6 Determine an F' test for any effects in Problem 13.5 that cannot be run directly.
13.7 Three replications are made of an experiment to determine the effect of days of
the week and operators on screen-color difference of a TV tube in °K. On a
given day, each of four operators measured the screen-color difference on a given
 Problems 279

tube. The order of measuring by the four operators is randomized each day for
five days of the week. The results are

Replication Operator Monday Tuesday Wednesday Thursday Friday

I A 800 950 900 740 880

B 760 900 820 740 960
C 920 920 840 900 1020
D 860 940 820 940 980

II A 780 810 880 960 920

B 820 940 900 780 820
C 740 900 880 840 880
D 800 900 800 820 900

III A 800 1030 800 840 800

B 900 920 880 920 1000
C 800 1000 920 760 920
D 900 980 900 780 880

Set up the model and outline the ANOVA table for this problem.
13.8 Considering days as fixed, operators as random, and replications as random,
analyze Problem 13.7.
13.9 Explain how a split-plot design differs from a nested design, since both have a
factor within levels of another factor.
13.10 From the data of Table 13.9 on the rubber cure rate index experiment verify
the results of Table 13.11.
13.11 Since two qualitative factors in Problem 13.10 show significance, compare their
means by a Newman-Keuls test.
13.12 Since temperature in Problem 13.10 is a highly significant quantitative effect,
fit a proper polynomial (or polynomials) to the data.
13.13 Redo Table 13.10, assuming that the three laboratories are chosen at random
from a large number of possible laboratories.
13.14 Analyze the results of Table 13.9, based on the EMS values of Problem 13.13.
13.15 In a study of the effect of baking temperature and recipe (or mix) on the quality
of cake, three mixes were used; A, B, and C. One of these three was chosen at
random and the mix made up and five cakes were poured, each placed in a differ-
ent oven and baked at temperatures of 300°, 325°, 350°, 375°, and 400°. After
these five cakes had been baked, removed, and their quality (T) evaluated, an-
other mix was chosen and again five cakes were poured and baked at the five
temperatures. Then this was repeated on the third mix. This experiment was
replicated r times.
a. Identify this design. Show its data layout.
280 Factorial experiment—split-plot design

b. Write the mathematical model for this situation and determine the ANOVA
outline indicating tests to be made.
c. Recommend how many replications r you would run to make this an accept-
able experiment and explain why.
13.16 If, in Problem 13.15, only one oven was available but cakes of all three recipes
could be inserted in the oven when set at a given temperature, how would the
problem differ from the setup in Problem 13.15? Explain which of the two designs
you would prefer and why.
13.17 Running two random replications of a problem cited in Cochran and Cox
[Experimental Designs, 2nd edition, Wiley, 1957] on the effect of temperature and
mix on cake quality similar to Problem 13.15 resulted in the following data.

o Temperatures
OO

Replication Mix 175° 195° 205° 215° 225°

1 47 29 35 47 57 45

I 2 35 46 47 39 52 61

3 43 43 43 46 47 58

1 26 28 32 25 37 33

II 2 21 21 28 26 27 20

3 21 28 25 25 31 25

Outline the ANOVA table for this problem and show what tests can be run and
discuss these tests.
13.18 Analyze the data given in Problem 13.17 and state your conclusions.
13.19 What next steps would you recommend after your conclusions in Problem 13.18?
 14
Factorial Experiment —

Confounding in Blocks

14.1 INTRODUCTION
As seen in Chapter 13, there are many situations in which randomization is
restricted. The split-plot design is an example of a restriction on randomiza-
tion where a main effect is confounded with blocks. Sometimes these restric-
tions are necessary because the complete factorial cannot be run in one day,
or in one environmental chamber. When such restrictions are imposed on the
experiment, a decision must be made as to what information may be sacrificed
and thus on what is to be confounded. A simple example may help to illustrate
this point.
A chemist was studying the disintegration of a chemical in solution. He
had a beaker of the mixture and wished to determine whether or not the depth
in the beaker would affect the amount of disintegration and whether material
from the center of the beaker would be different than material from near the
edge. He decided on a 2^ factorial experiment—one factor being depth, the
other radius. A cross section of the beaker (Figure 14.1) shows the four treat-
ment combinations to be considered.
In Figure 14.1 (1) represents radius zero, depth at lower level (near
bottom of beaker); a represents the radius at near maximum (near the edge of
the beaker) and a low depth level; b is radius zero, depth at a higher level;
and ab is both radius and depth at higher levels. In order to get a reading on

Figure 14.1

281
282 Factorial experiment—confounding in blocks

the amount of disintegration of material, samples of solution must be drawn

from these four positions. This was accomplished by dipping into the beaker
with an instrument designed to “capture” a small amount of solution. The
only problem was that the experimenter had only two hands and could get
only two samples at the same time. By the time he returned for the other two
samples conditions might have changed and the amount of disintegration
would probably be greater. Thus there is a physical restriction on this 2^
factorial experiment; only two observations can be made at one time. The
question is: Which two treatment combinations should be taken on the first
dip? Considering the two dips as blocks, we are forced to use an incomplete
block design, as the whole factorial cannot be performed in one dip.
Three possible block patterns are shown in Table 14.1.

Table 14.1 Three Possible Blockings of a

2^ Factorial

Plan
Dip I II III

1 (1) b (1) a (1) ab

2 a ab b ab a b

If plan I of Table 14.1 is used, the blocks (dips) are confounded with the
radius effect, since all zero-radius readings are in dip 1 and all maximum
radius readings are in dip 2. In plan II the depth effect is confounded with
the dips, as low-level depth readings are both in dip 1 and high-level depth
readings are in dip 2. In plan III neither main effect is confounded, but the
interaction is confounded with the dips. To see that the interaction is con-
founded, recall the expression for the AB interaction in a 2^ factorial found
in Chapter 6.
AB = i[(I) — a — b nfi]
Note that the two treatment effects with the plus sign, (1) and ah, are both in
dip I, and the two with a minus sign, a and b, are in dip 2. Hence we cannot
distinguish between a block effect (dips) and the interaction effect (AB). In
most cases it is better to confound an interaction than to confound a main
effect. The hope is, of course, that there is no interaction, and that informa-
tion on the main effects can still be found from such an experiment.

14.2 CONFOUNDING SYSTEMS

To design an experiment in which the number of treatments that can be
run in a block is less than the total number of treatment combinations, the
experimenter must first decide on what effects he is willing to confound. If,
 14.2 Confounding systems 283

as in the example above, he has only one interaction in the experiment and
decides that he can confound this interaction with blocks, the problem is
simply which treatment combinations to place in each block. One way to
accomplish this is to place in one block those treatment combinations with
a plus sign in the effect to be confounded, and those with a minus sign in the
other block. A more general method is necessary when the number of blocks
and the number of treatments increase.
First a defining contrast is set up. This is merely an expression stating
which effects are to be confounded with blocks. In the simple example
above, to confound AB, write AB as the defining contrast. Once the defining
contrast has been set up, several methods are available for determining which
treatment combinations will be placed in each block. One method was shown
above—place in one block the treatment combinations that have a plus
sign in AB and those with a minus sign in the other block. Another method
is to consider each treatment combination. Those that have an even number
of letters in common with the effect letters in the defining contrast go in one
block. Those that have an odd number of letters in common with the de-
fining contrast go into the other block. Here (1) has no letters in common
with AB or an even number. Type a has one letter in common with AB
(a and A) or an odd number, b also has one letter in common with AB. ab
has two letters in common with AB. Thus the block contents are those in
Figure 14.2.

AB
Block I Block II

Evens

Figure 14.2

One disadvantage of the two methods given above is that they are only
good on 2^ factorials. A more general method is attributed to Kempthorne
[13]. Consider the linear expression
A^X^ + A X 2 2 + ■ ■ ■ + A„X„ = L

Here Ai is the exponent on the ith factor appearing in each independent

defining contrast, and X^ is the level of the ith factor appearing in a given
treatment combination. Every treatment combination with the same L value
will be placed in the same block. For the simple example above where the
284 Factorial experiment—confounding in blocks

defining contrast is AB, = 1, A2 = 1, and all other /l/s = 0. Hence

L = 1 - Xi + 1 •X2

For each treatment combination the L values are

(1):L = 1 - 0 + 1 - 0 = 0
a: L = 1 ■ i A I-0 = 1
h:L = 1 -0 + 1 • 1 = 1
ab: L = 1-1 + 1-1 = 2

Hence 0 and 2 are both 0, since a 2^ factorial is of modulus 2 (see Section

9.2). The assignment of treatment combinations to blocks is then

L = 0 block 1 (1) ab

L = \ block 2 a

For a more complex defining contrast such as ABC^, the expression is

L = X, A X 2 A 2X3
and this can be used to decide which treatment combinations in a 3^ experi-
ment go into a common block.
The block containing the treatment combination (1) is called the princi-
pal block. The treatment combinations in this block are elements of a group
where the operation on the group is multiplication, modulus 2. The elements
of the other block or blocks may be generated by multiplying one element
in the new block by each element in the principal block. Multiplying elements
together within the principal block will generate more elements within the
principal block. This is best seen with a slightly more complex factorial, say
2^. If the highest order interaction is to be confounded and the eight treat-
ment combinations are to be placed into two blocks of four treatments each,
then confounding ABC gives
L, = X^ A X 2 + X3
Testing each of the eight treatment combinations gives

(1): L-0 + 0-f0 = 0

u:L-l+0 + 0= l
h:L = 0-fl+0=l
ab:L=iAlA0 = 2 = 0
c: L = 0A0Al = i
ac: L = 1 A 0 A 1 = 2 = 0
be: L = 0 A \ A I = 2 = 0
abc: L=l + l + l= 3 = l
 14.3 Block confounding—no replication 285

The blocks are then

block 1 block 2

(1) a
ab b
ae c
be abc

L = 0 L = 1

Group theory can simplify this procedure. Determine two elements in the
principal block, for example, ab and be. Multiply these two to get ab^c = ac,
the fourth element. Generate block 2 by finding one element, say a. Multiply
this element by each element in the principal block, thus generating all the
elements of the second block as follows:

a ’ (1) = a
a • ab = a^b = b
a ■ ac = a^c = c
a ■ be = abc

These concepts may be extended to more complex confounding examples, as

shown in later sections.

14.3 BLOCK CONFOUNDING—NO REPLICATION

In many cases an experimenter cannot afford several replications of an
experiment, and is further restricted in that the complete factorial cannot be
run in one block or at one time. Again, the experiment may be blocked
and information recovered on all but some high-order interactions, which
may be confounded. Designs for such experiments will be considered for the
special cases of 2^ and 3^ factorials because they appear often in practice.

Confounding in 2-^ Factorials

The methods given earlier in this chapter can be used to determine the block
composition for a specific confounding scheme. If only one replication is
possible, one is run, and some of the higher order interaction terms must be
used as experimental error unless some independent measure of error is
available from previous data. This type of design is used mostly when there
are several factors involved (say four or more), some high-order interactions
may be confounded with blocks, and some others are yet available for an
error estimate. For example, consider a 2"^ factorial where only eight treat-
ment combinations can be run at one time. One possible confounding
286 Factorial experiment—confounding in blocks

scheme would be

confound ABCD and L = + 2^2 -l- 2^3 + X

block 1 (1) ab be cic abed cd ad bd

block 2 a b abc c bed acd d abd

and the analysis would be as shown in Table 14.2. The three-way interactions
may be pooled here to give 4 df for error, assuming that these interactions are
actually nonexistent. All main effects and first-order interactions could then
be tested with 1 and 4 df. After examining the results it might be possible to
pool some of the two-way interactions that have small mean squares with the
three-way interactions if more degrees of freedom are desired in the error
term.
If only four treatment combinations can be run in a block, one may
confound the 2^ factorial in four blocks of four treatment combinations each.
For four blocks 3 df must be confounded with blocks. If two interactions
are confounded, the product (modulus 2) of these two is also confounded,
since the product of the signs in two effects gives the signs for the product.
Thus, if AB and CD are confounded, so also will ABCD be confounded.
This scheme would confound two first-order interactions and one third order.
It might be better to confound two second-order interactions and one first
order:
ABC BCD AD
Table 14.2 2“^ Factorial in Two Blocks

Source df

A 1
B 1
C 1
D 1
AB 1
AC 1
AD 1
BC 1
BD 1
CD 1
ABC 1]
ABD 1
> 4 df for error
ACD 1
BCD ij
Blocks or ABCD 1

Total 15 df
 14.3 Block confounding—no replication 287

Note here that {ABC){BCD) = AB^C^D = AD (modulus 2). Hence 3 df

are confounded with blocks. Before proceeding to the block compositions,
consider a problem where four factors {A, B, C, and D) are involved, each at
two levels, and only four treatment combinations can be run in one day. The
measured variable is yield. The confounding scheme given above gives two
independent equations

+ X2 + X2
L2 = X2 + X2 + x^
Each treatment combination will give one of the following four pairs when
substituted into both and L = 00, 01, 10, 11. For example,
2

(l):Li =0 L2 = 0
Cl. L—— 1 B2 ~~ 0
h: Li = 1 L2 = 1
ab: Li = 2 = 0 L2 = 1
and so on. All treatment combinations with the same pair of L values will be
placed together in one block. Thus the 16 treatment combinations will be
assigned to four blocks as in Table 14.3.
Here the order of experimentation within each block is randomized and
the resulting responses are given for each treatment combination. The proper
treatment combinations for each block can be generated from the principal
block. If be and acd are in the principal block, then their product

{bc){acd) = abc^d = abd

is also in the principal block. For another block, consider a and multiply a by
all the combinations in the principal block, giving
a{l) = a, a{bc) = abc, a{acd) = cd, a{abd) = bd

the entries in the second block. In the same manner, the other two blocks may
be generated. This means that only a few readings in the principal block need
Table 14.3 Blocking a 2"^ Factorial in Four Blocks of Four
Confounding ABC, BCD, and AD

block 1 block 2 block 3 block 4

(1) = 82 a = 76 b = 19 ab = 85
be = 55 abc = 74 c = 71 ac = 84
acd = 81 cd = 72 abed = 89 bed = 84
abd — 88 bd = 73 ad = 79 d = m

L, = 0 Li = 1 Li = 1 Li = 0
L 2 = 0 Lj = 0 ~ T2 = 1 1^2=1
288 Factorial experiment—confounding in blocks

to be determined by the Lj, L2 values, and all the rest of the design can be
generated from these few treatment combinations. Tables are also available
for designs when certain effects are confounded [9].
This experiment is then a 2"^ factorial experiment. The design is a ran-
domized incomplete block design where blocks are confounded with the
interactions ABC, BCD, and AD. An example will illustrate the method.
Example 14.1 A systematic test was to be made on the effects of four
variables, each at two levels, on coil breakdown voltage. The variables were
A—firing furnace 1 or 3; B—firing temperature 1650°C or 1700°C; C—gas
humidification, no or yes; and D—coil outside diameter small (< 0.0300 inch)
or large (>0.0305 inch). Since only four coils could be tested in a given short
time interval, it was necessary to run four blocks of four coils for such testing.
Only one complete test of 2"*’ in four blocks of four was run and the results
(after dividing the voltages through by 10) are shown in Table 14.3, which
confounds ABC, BCD, and AD with the four blocks. For the analysis one
might use Yates’ method given in Chapter 6 to give Table 14.4.
In Table 14.4 the sums of squares in the last column correspond to the
effect for the treatment combination on the left. That is,
SS^ = 225.00 SSg = 0.25 and so on

Table 14.4 Yates’ Method Analysis for 2"’’ Example

Treatment
Combination Response (1) (2) (3) (4) SS

(1) 82 158 322 606 1252 —

a 76 164 284 646 + 60 225.00

b 79 155 320 + 32 2 0.25
ab 85 129 326 28 30 56.25
c 71 159 0 -20 -32 64.00
ac 84 161 + 32 22 + 32 64.00
be 55 153 14 + 18 -14 12.25
abc 74 173 14 12 -26 42.25*
d 80 -6 6 -38 40 100.00
ad 79 +6 -26 6 -4 1.00*
bd 73 + 13 2 + 32 42 110.25
abd 88 + 19 20 0 -6 2.25
cd 72 -1 12 -32 44 121.00
acd 81 15 +6 18 -32 64.00
bed 84 9 16 -6 50 156.25*
abed 89 5 -4 -20 -14 12.25

Total 1031.00

* Confounded with blocks.

 14.3 Block confounding—no replication 289

Note that the effects ABC, BCD, and AD are confounded with blocks. Their
total sum of squares is 42.25 + 156.25 + 1.00 = 199.50. If the above block
totals are computed, they yield

block 1 — 306 blocks — 318

block 2 — 295 block 4 — 333

and the sum of squares between these blocks is

(306)^ + (295)" + (318)" + (333)" (1252)"

4

SSbiock 98,168.50 - 97,969.00 = 199.50

which is identical with the sum of the three interactions with which blocks
are confounded. If all three-way and four-way interactions that are not con-
founded are pooled for an error term, the resulting analysis is that shown in
Table 14.5.

Table 14.5 ANOVA for 2"’’ Example in Four Blocks

Source df SS MS

A 1 225.00 225.00
B 1 0.25 0.25
C 1 64.00 64.00
D 1 100.00 100.00
AB 1 56.25 56.25
AC 1 64.00 64.00
BC 1 12.25 12.25
BD 1 110.25 110.25
CD 1 121.00 121.00
Error or ABD, ACD, ABCD 3 78.50 26.17
Blocks or ABC, BCD, AD 3 199.50 66.50

Totals 15 1031.00

With only 1 and 3 df, none of the four main effects or the five two-way
interactions can be declared significant at the 5 percent significance level.
Since the B effect and the BC interaction are not significant even at the
25 percent level, we might wish to pool these with the error to increase the
error degrees of freedom. The resulting error sum of squares would then be

0.25 + 12.25 + 78.50 = 91.00

with 5 df. The error mean square then = 91.00/5 = 18.20. Using this error
mean square, A and CD are now significant at the 5 percent level. We might
conclude that the effect of factor A is important, that B is not present, and
290 Factorial experiment—confounding in blocks

that there may be a CD interaction. As 2^ experiments are often run to get

an overall picture of the important factors, each set at the extremes of its
range, another experiment might now be planned without factor B, since its
effect is negligible over the range considered here. The next experiment might
include factors A, C, and D at, perhaps, three levels each, or a 3^ experiment.
This experiment might also require blocking.

14.4 BLOCK CONFOUNDING WITH REPLICATION

Whenever an experiment is restricted so that all treatments cannot appear in
one block, some interaction is usually confounded with blocks. If several
replications of the whole experiment (all blocks) are possible, as in the case
of the split-plot, the same interaction may be confounded in all replications.
In this case, the design is said to be completely confounded. If, on the other
hand, one interaction is confounded in the first replication, a different interac-
tion is confounded in the second replication, and so on, the design is said to
be partially confounded.

Complete Confounding

Considering a 2^ factorial experiment in which only four treatment com-

binations can be finished in one day yields a 2^ factorial run in two incom-
plete blocks of four treatment combinations each. We might confound the
highest order interaction ABC as follows:

ABC
As seen in Section 14.2, the blocks would be

block 1 block 2

(1) a
ab b
ac c
be abc

If this whole experiment (2^ in two blocks of four each) can be replicated, say
three times, the layout might be

replication I replication II replication III

block 1 block 2 block 1 block 2 block 1 block 2

ac a c (1) ab c
(1) c abc ac (1) b
ab abc b be ac abc
be b a ab be a
 14.4 Block confounding with replication 291

The confounding scheme in the above (ABC) is the same for all three replica-
tions, but the order of experimentation has been randomized within each
block. Also, the decision as to which block is to be run first in each replica-
tion is made at random. An analysis layout for this experiment appears in
Table 14.6.

Table 14.6 Analysis Layout for Completely Confounded

2^ Factorial

Source df df

Replications 2
Blocks or ABC 1
Replications x block interactions 2 5 between plots

A 1
B 1
AB 1
C 1
AC 1
BC 1
Replications x all others 12 18 within plots

Totals 23 23

Here the replication interaction with all three main effects and their
interactions are usually taken as the error term for testing the important
effects. The replication effect and the block (or ABC) effect could be tested
against the replication x block interaction, but the degrees of freedom are
low and the power of such a test is poor. Such a design is quite powerful,
however, in testing the main effects A, B, and C and their first-order interac-
tions. However, it will give no clean information on the ABC interaction.

Partial Confounding

In the example just considered, we might be concerned with some test on the
ABC interaction. This could be found by confounding some interactions
other than ABC in some of the replications. We might use four replications
and confound ABC in the first one, AB in the second, AC in the third, and
BC in the fourth. Thus the four replications will yield full information on
A, B, and C but three-fourths information on AB, AC, BC, and ABC, since
we can compute an unconfounded interaction such as AB in three out of four
of the replications. The following layout shows the proper block entries for
each replication.
292 Factorial experiment—confounding in blocks

replication I replication II replication III replication IV

confound ABC AB AC BC

(1) a (1) a (1) a (1) b

ab b c b ac c be c
ac c ab ac abc be a ab
be abc abc be b ab abc ac

L = X, + X2 L = X^ A X2 L = Vi + V3 L = X2 A X2

A V3
The analysis layout would be that shown in Table 14.7.
The residual term with its 17 df can be explained as follows. In all four
replications the main effects A, B, and C can be isolated. This gives 3 x 1 df
for each replication by main-effect interaction, or a total of 9 df. The AB,
AC, BC, and ABC interactions can be isolated in only three out of the four
replications. Hence the replications by AC have 2 x 1 ^ 2 df, or a total of
8 df for all such interactions. This gives 9 + 8 = 17 df for effects by replica-
tions, which is usually taken as the error estimate. The blocks and replications
are separated out in the top of Table 14.7 only to reduce the experimental
error. The numerical analysis of problems such as this is carried out in the
same manner as given in earlier chapters.

Example 14.2 Coils were wound from each of two winding machines using
either of two wire stocks and their outside diameters were measured at two

Table 14.7 ANOVA for Partially Confounded 2^ Factorial

Source df df

Replications 3
Blocks within replications 4 7 between plots
ABC (Replication I) 1
AB (Replication II) 1
AC (Replication III) 1
BC (Replication IV) 1

A 1
B 1
C 1
AB n only from
AC 1 replications
BC 1 where not
ABC ij confounded
Replications x all effects 17 24 within plots

Totals 31 31
 14.4 Block confounding with replication 293

positions on an optical comparator. Since only four readings could be made

in a given time period, the 2^ = 8 treatment combinations were divided into
two blocks of four. However, four replications were possible. At first the
experimenter decided to confound ABC in all four replicates. Here A repre-
sents machines, B stocks, and C positions. The results are shown in Table 14.8.

Table 14.8 Layout of Example 14.2 in Four Blocks of Four Confounding ABC
in All Replicates

replication I replication II
block 1 block 2 block 1 block 2

ab = 2173 b = 2300 a = 2228 ac = 3495

(1) = 2249 c = 3538 abc — 3592 (1) = 2094
ac = 3532 abc = 3524 c = 3116 be = 2249
be = 2948 a = 2319 b = 2386 ab = 2373

replication III replication IV

block 1 block 2 block 1 block 2

(1) = 2382 c = 3528 abc = 2996 ab = 2393

ab = 2240 b = 2118 c = 1934 be = 2814
be = 3495 abc = 3350 b = 2234 ac — 3400
ac = 2995 a = 2272 a = 2215 (1) = 2297

All measurements multiplied by 10^ so 2173 is an outside diameter of 0.02173 inch.

Analyzing by computer or by direct methods gives Table 14.9.

Table 14.9 ANOVA for Example 14.2

Source df SS MS F Prob.

Replications 3 409,719.59
Blocks or ABC 1 8,482.53
R X B interaction 3 515,407.85
7

A 1 364,444.53 364,444.53 3.13 0.094

B 1 5,227.53 5,227.53 <1
AB 1 18,963.78 18,963.78 <1
C 1 6,330,571.53 6,330,571.53 54.44** 0.000
AC 1 302,058.78 302,058.78 2.60 0.124
BC 1 16,698.78 16,698.78 <1
Replications x all others 18 2,093,039.32 116,279.96
24

Totals 31 10,064,614.22
294 Factorial experiment—confounding in blocks

Table 14.9 shows only positions on the coil as having a highly significant
effect on coil outside diameter.
Example 14.3 When coils from two other winding machines were to be
tested, it was decided to partially confound the interactions. Results gave
Table 14.10.

Table 14.10 Layout of Example 14.3 in Four Blocks of Four with Partial
Confounding

replication I (confounding ABC) replication II (confounding AB)

block 1 block 2 block 3 block 4

(1) = 2208 a = 2196 (1) = 2004 a = 2179

ab = 2133 b = 2086 ab = 2112 b = 2073
ac = 2459 c = 3356 c = 3073 ac = 3474
be = 3096 abc = 2776 abc = 2631 be = 3360

replication III (confounding AC) replication IV (confounding BC)

block 5 block 6 block 7 block 8

c = 2839 (1) = 1916 a = 2056 b = 1878

a = 2189 ac = 2979 (1) = 2010 ab = 2156
be = 3522 b = 2151 be = 3209 c = 3423
ab = 2095 abc = 2500 abc = 3066 ac = 2524

Analysis gives Table 14.11. In this table position on the coil is again
highly significant, but an AC interaction is also significant at the 5 percent
significance level.
In both problems the effect of A—machines—is approaching signifi-
cance. It may be that if more replications are run, this machine effect will
emerge as significant.

Confounding in 3^ Factorials

When a 3^-factorial experiment cannot be completely randomized, it is

usually blocked in blocks that are multiples of 3. The use of the I and J
components of interaction introduced in Section 9.2 is helpful in confounding
only part of an interaction with blocks. Kempthorne’s rule also applies, using
such interactions as AB, AB^, ABC^,. . ., and treatment combinations in
the form 00,10,01,11 instead of (1), a, h, ab as in lE Here treatment combina-
tions can be multiplied that will actually add these exponents (modulus 3).
If a 3^ factorial is restricted so that only three of the nine treatment com-
binations can be run in one block, we usually confound AB or AB^, as each
carries 2 df. The defining contrast might be AB^, which gives L = A- 2X2-
 14.4 Block confounding with replication 295

Table 14.11 ANOVA for Example 14.3

Source df SS MS F Probability

Replications 3 38,717.59
Blocks in
replications 4 401,060.13
7

A 1 224,282.53 224,282.53 3.97 0.063

B 1 52.53 52.53 <1
AB 1 737.04 737.04 <1
C 1 6,886,688.28 6,886,688.28 121** 0.000
AC 1 416,066.67 416,066.67 7.36* 0.015
BC 1 2,667.04 2,667.04 <1
ABC 1 70,742.04 70,742.04 1.25 0.279
Replications x all 17 961,283.32 56,546.08
24

Totals 31 9,002,296.97

Each of the nine treatment combinations then yields

00, L = 0 01, L = 2 02, L = 4 = 1
10, L = 1 11, L = 3 - 0 12, L = 5 = 2
20, L = 2 21, L = 4 = 1 22, L = 6 - 0

Placing treatment combinations with the same L value in common blocks

gives
block 1 block 2 block 3
L = 0 L=1 L = 2

00 10 20
11 21 01
22 02 12

Block 1 with 00 in it is the principal block. Block 2 is generated by adding one

treatment combination in block 2 to all those in block 1, giving
00 + 10 = 10 11 + 10 = 21 22 + 10 = 32 = 02

Similarly for block 3,

00 + 20 = 20 11 + 20 = 31 = 01 22 + 20 = 42 = 12

Since there are three blocks, 2 df must be confounded. Here AB^ is con-
founded with blocks. If AB with its 2 df were confounded, L = X^ A X2,
296 Factorial experiment—confounding in blocks

and the blocking would be

block 1 block 2 block 3

L = 0 L = \ L = 2

00 10 20
12 22 02
21 01 11

If the data of Table 9.1 are used as an example, and AB^ is confounded, the
responses are

block 1 block 2 block 3

00 = 1 10 = -2 20-3
11 = 4 21 - 1 01-0
22-2 02 = 2 12 - -1

The block sums of squares are determined from the three block totals of
7, 1, and 2:
7^ + + 2^ (10)^
SS block 6.89
3

which is identical with the I{AB) interaction component computed in

Section 9.2 (Table 9.6) as it should be. The remainder of the analysis proceeds
the same as in Section 9.2, and the resulting ANOVA table is that exhibited
in Table 14.12.
If these were real data, the AB part of the interaction might be considered
as error, but then neither main effect is significant. The purpose here is only
to illustrate the blocking of a 3^ experiment.
Consider now a 3^ experiment in which the 27 treatment combinations
(such as given in Table 9.7) cannot all be completely randomized. If nine can
be randomized and run on one day, nine on another day, and so on, we might
use a 3^ factorial in three blocks of nine treatment combinations each. This
requires 2 df to be confounded with blocks. Since the ABC interaction with
8 df can be partitioned into ABC, ABC^, AB^C, and AB^C^, each with 2 df.

Table 14.12 ANOVA for 3^ Example

Source df SS MS

A 2 4.22 2.11
B 2 1.56 0.78
AB 2 16.22 8.11
Blocks or AB^ 2 6.89 3.44

Totals 8 28.89
 14.4 Block confounding with replication 297

one of these four could be confounded with blocks. If AB^C is confounded,

L = 2X2 + X2

and the three blocks are

L = 0 L = 1 L = 2

000 100 200

oil 111 211
no 210 010
121 221 021
102 202 002
212 012 112
220 020 120
022 122 222
201 001 101

The analysis of data for this design would yield Table 14.13.
This is a useful design, since we can retrieve all main effects {A, B, C)
and all two-way interactions, if we are willing to pool the 6 df in ABC as
error. The determination of sums of squares, and so forth, is the same as for
a complete 3^ given in Example 9.1. Confounding other parts of the ABC
interaction would, of course, yield different blocking arrangements, although
the outline of the analysis would be the same. In practice, different blocking
arrangements might yield different responses.
More complex confounding schemes can be found in tables such as in
Chapter 9 of Davies [9].
For practice, consider confounding a 3^ in nine blocks of three treatment
combinations each. Here 8 df must be confounded with blocks. One con-
founding scheme might be
AB^C^ AB BC^ AC

Table 14.13 3^ Factorial in Three Blocks

Source df

A 2
B 2
AB 4
C 2
AC 4
BC 4
Error or ABC, ABC^, AB^C^ 6
Blocks or AB^C 2

Total 26
298 Factorial experiment—confounding in blocks

Note that
{AB^C^){AB) = A^C^ = AC
and
iAB){BC^) = AB^C^
These are not all independent. In fact, only two of the expressions are. For
this reason we need only two expressions for the L values;
L, = A 2X2 + 2X3

L2 = X,A X2
as these two will yield nine pairs of numbers—one pair for each block. First
determine the principal block, where both and L2 are zero. One treatment
combination is obviously 000. Another is 211, as
Li = 1(2) + 2(1) + 2(1) = 6 = 0
L2 = 1(2) + 1(1) = 3 = 0
A third is 122, as
Li = 1(1) + 2(2) + 2(2) = 9 = 0
L2 = 1(1) + 1(2) = 3 = 0
The other eight blocks can now be generated from this principal block,
giving
block 1 2 3 4 5 6 7 8 9

000 001 002 010 020 100 200 no 101

211 212 210 221 201 on 111 021 012
122 120 121 102 112 222 022 202 220

An analysis layout would be that shown in Table 14.14.

Table 14.14 3^ Factorial in Nine Blocks

Source df

A 2
B 2
C 2
AB^ 2
AC^ 2
BC 2
ABC 2]
AB^C 2 > 6 df for error
ABC^ 2)
Blocks or AB, BC\ AC, AB^C^ 8

Total 26
 14.5 Summary 299

This design might be reasonable if interest centered only in the main

effects A, B, and C. Such a design might be necessary where three factors are
involved, each at three levels, but only three treatment combinations can be
run in one day. These three might involve an elaborate environmental test
where at best only three sets of environmental conditions—temperature,
pressure, and humidity—can be simulated in one day.

14.5 SUMMARY

The summary of Chapter 13 may now be extended.

Experiment Design Analysis

II. Two or more

factors
A. Factorial 1. Completely randomized
(crossed) 2. Randomized block 2.

a. Complete a. Factorial
^ijk = /^ + + d,- ANOVA
+ Bj + AB^j +
b. Incomplete, confounding: b.
i. Main effect—split-plot i. Split-plot
Tjk = + Bi + Aj + RAij ANOVA
'' V ^

whole plot
+ + ^^ik + d-BjTc + V
RAB^j^
^

split-plot
ii. Interactions in 2^ and 3^ ii.
(1) Several replications (1) Factorial
Tjkq ^ y. + Ri + ANOVA
+ Rf^ij + A,^ + Bq with
replications
Rj and
blocks l^j or
confounded
interaction
(2) One replication only (2) Factorial
Tjk = B + Pi + ANOVA
+ ^k + .. . with blocks
PjOV

confounded
interaction
300 Factorial experiment—confounding in blocks

PROBLEMS

14.1 Assuming that a 2^ factorial such as the problem of Chapters 1, 5, and 6 involving
two tool types, two bevel angles, and two types of cut cannot all be run on one
lathe in one time period but must be divided into four treatment combinations
per lathe, set up a confounding scheme for confounding the AC interaction in
these two blocks of four. Assuming the results of one replication give (1) = 5,
u = 0, h = 4, uh = 2, c = —3,ac = 0,bc= —1, abc = — 2, show that the
scheme you have set up does indeed confound AC with blocks.
14.2 Assuming three replications of the 2^ experiment in Problem 14.1 can be run but
each one must be run in two blocks of four, set up a scheme for the complete
confounding of ABC in all three replicates and show its ANOVA layout.
14.3 Repeat Problem 14.2 using partial confounding of AB, AC, and BC.
14.4 Data involving four replications of the 2^ factorial described in the foregoing
problems and in Section 14.3 gave the following numerical results.

replication I replication II
confound ABC AB

(1) = 2 a — 0 (1) = -3 0=1

ab = 3 b = 1 c = —6 h = 1
ac —1 c = 0 ab = % oe = — 6
be = -2 abc = —1 abc = 0 he = 2

replication III replication IV

confound AC BC

(1) = 5 0 = 0 (1) = -2 h = 9
ac = 0 c = -3 he = —1 e = -3
o

abc = —2 he = —1 0 = —6
II

b = 4 ab = 2 abc = —4 ac — —4

Analyze these data and state your conclusions.

14.5 A 2^ factorial is to be used to study three factors and only four treatments can
be run on each of two days. On the first day the following data are collected:

(1) = 2, a = 1, be = 3, abc = 4

and on the next day:

c = 1, ac = 4, b = 6, ab = 0

By examining the variation between the two days and your knowledge of a 2-^
experiment, hnd out which term in the model (main effect or interaction) is
confounded with days.
 Problems 301

14.6 For the data of Problem 6.6 consider running the 16 treatment combinations in
four blocks of four (two observations per treatment). Confound ACD, BCD, and
AB and then determine the block compositions.
14.7 From the results of Problem 6.6 show that the blocks are confounded with the
interactions in Problem 14.6.
14.8 For Problem 9.11 consider this experiment as run in three blocks of nine with
two replications for each treatment combination. Confound ABC (2 df) where
A is surface thickness, B is base thickness, and C is subbase thickness, and deter-
mine the design. Also use the numerical results of Problem 9.13 to show that
ABC is indeed confounded in your design.
14.9 Set up a scheme for confounding a 2"^ factorial in two blocks of eight each,
confounding ABD with blocks.
14.10 Repeat Problem 14.9 and confound BCD.
14.11 Work out a confounding scheme for a 2^ factorial, confounding in four blocks
of eight each. Show the outline of an ANOVA table.
14.12 Repeat Problem 14.11 in four replications confounding different interactions in
each replication. Show the ANOVA table outline.
14.13 Work out the confounding of a 3‘^ factorial in three blocks of 27 each.
14.14 Repeat Problem 14.13 in nine blocks of nine each.
14.15 In Example 14.2, since ABC is completely confounded, the analysis could be
easily run on a computer or by using Yates’ method on the A, B, C sections and
general methods for replications and blocks. Verify the results given in Table 14.9.
14.16 Since Example 14.3 has partial confounding, it may be difficult for the computer.
However, a Yates method can be used if care is taken as to how it is applied.
Try to verify Table 14.11 results using a modified Yates procedure.
14.17 A systematic test was made to determine the effects on coil breakdown voltage
of the following six factors, each at two levels as indicated:
1. Firing furnace Number 1 or 3
2. Firing temperature 1650°C or 1700°C
3. Gas humidification No or yes
4. Coil outer diameter Below 0.0300 inch or above 0.0305 inch
5. Artificial chipping No or yes
6. Sleeve Number 1 or 2
All 64 experiments could not be performed under the same conditions so the
whole experiment was run in eight subgroups of eight experiments. Set up a
reasonable confounding scheme for this problem and outline its ANOVA.
14.18 An experiment was run on the effects of several annealing variables on the
magnetic characteristic of a metal. The following factors were to be considered.
1. Temperature 1375°, 1450°, 1525°, 1600°
2. Time in hours 2, 4
302 Factorial experiment—confounding in blocks

3. Exogas ratio 6.5 to 1, 8.5 to 1

4. Dew point 10°, 30°
5. Cooling rate in hours 4, 8
6. Core material A, B.
Since all factors are at two levels or multiples of two levels, outline a scheme for
this experiment in two blocks of 64 observations.
14.19 Outline a Yates’ method for handling the data of Problem 14.18.
14.20 If X is the surface rating of a steel sample and material is tested from eight heats,
two chemical treatments, and three positions on the ingot, the resulting factorial
is an 8 X 2 X 3 or 2^ X 2 X 3 or 2"’' X 3, which is called a mixed factorial in
the form 2"^ ■ 3”. Set up an ANOVA table for such a problem treating the eight
heats as 2^ pseudo factors.
14.21 Assuming Problem 14.20 results must be fired in furnaces that can handle only
24 steel samples at a time, devise a confounding scheme for the experiment.
14.22 Complete the analysis of Problem 14.21 for the following data.

Position
1 2 3

Chemical Treatment Chemical Treatment Chemical Treatment

Heat 1 2 1 2 1 2

1 1 0 0 0 0 2
2 -3 2 -1 -2 -2 -2
3 1 2 0 3 0 4
4 0 0 -2 1 0 1
5 -3 1 0 2 -2 0
6 -2 1 -3 2 0 0
7 3 3 3 3 1 -1
8 1 1 0 5 1 2

14.23 Consider a 2^ factorial experiment with all fixed levels to be run in four blocks
of eight treatment combinations each. The whole experiment can be replicated
in a total of three replications. The following interactions are to be confounded
in each replication.

Replication I — ABC, CDE, ABDE

Replication II — BCDE, ACD, ABE
Replication III — ABCD, BCE, ADE
Determine the block compositions for each block in each replication and outline
the ANOVA for this problem.
 15
Fractional Replication

15.1 INTRODUCTION
As the number of factors to be considered in a factorial experiment increases,
the number of treatment combinations increases very rapidly. This can be
seen with a 2^ factorial where / = 5 requires 32 experiments for one replica-
tion, / = 6 requires 64, / = 7 requires 128, and so on. Along with this
increase in the amount of experimentation comes an increase in the number
of high-order interactions. Some of these high-order interactions may be
used as error, as those above second order (three way) would be difficult to
explain if found significant. Table 15.1 gives some idea of the number of
main effects, first-order, second-order,. . ., interactions that can be recovered
if a complete 2^ factorial can be run.
Considering / = 7, there will be 7 df for the seven main effects, 21 df for
the 21 first-order interactions, 35 df for the 35 second-order interactions,
leaving
35 21 + 7 -f 1 = 64 df
for an error estimate, assuming no blocking. Even if this experiment were
confounded in blocks, there is still a large number of degrees of freedom for
the error estimate. In such cases, it may not be economical to run a whole
replicate of 128 observations. Nearly as much information can be gleaned

Table 15.1 Buildup of 2-^-Factorial Effects

Order of Interaction
Main
/ 2^ Effect 1st 2nd 3rd 4th 5th 6th 7th

5 32 5 10 10 5 1
6 64 6 15 20 15 6 1
7 128 7 21 35 35 21 7 1
8 256 8 28 56 70 56 28 8 1

303
304 Fractional replication

from half as many observations. When only a fraction of a replicate is run,

the design is called a fractional replication or a fractional factorial.
For running a fractional replication, the methods of Chapter 14 are used
to determine a confounding scheme such that the number of treatment
combinations in a block is within the economic range of the experimenter.
If, for example, the experimenter can run 60 or 70 experiments and is inter-
ested in seven factors, each at two levels, a one-half replicate of a 2^ factorial
can be used. The complete 2^ = 128 experiment is laid out in two blocks of
64 by confounding some high-order interaction. Then only one of these two
blocks is run; the decision is made as to which one by the toss of a coin.

15.2 ALIASES
To see how a fractional replication is run, consider a simple case. Three
factors are of interest, each at two levels. However, the experimenter cannot
afford 2^ = 8 experiments, but will, however, settle for four. This suggests a
one-half replicate of a 2^ factorial. Suppose ABC is confounded with blocks.
The two blocks are then
/ = ABC

block 1 (1) ab be ac

block 2 a b c abc

A coin is flipped and the decision is made to run only block 2. What informa-
tion can be gleaned from block 2 and what information is lost when only
half of the experiment is run?
Referring to Table 6.7, which shows the proper coefficients (+1 or — 1)
for the treatment combinations in a 2^ factorial to give the effects desired,
box only those treatment combinations in block 2 that are to be run (Table
15.2).
Note, in the boxed area, that ABC has all plus signs in block 2, which is
a result of confounding blocks with ABC. Note also that the effect of A is

/l = +a — h — c A- ahc
and
BC = Aa — h — c A abc
so that we cannot distinguish between A and BC in block 2. Two or more
effects that have the same numerical value are called aliases. We cannot tell
them apart. B and AC likewise are aliases, as are C and AB. Because of this
confounding when only a fraction of the experiment is run, we must check
the aliases and be reasonably sure they are not both present if such a design
is to be of value. An analysis of this one-half replication of a 2^ would be as
shown in Table 15.3.
 15.2 Aliases 305

Table 15.2 One-Half Replication of a 2^ Factorial

Effect
Treatment
Combination A B AB C AC BC

(1) —
+ —
+ -f-

a + — — — — + -f

b —
+ — — + — +

ab -f -f + _ _

c — — + + — — A

ac -f - — -I- -1- -
be - + - + - +

abc + + + -f + + -f

Table 15.3 One-Half Replication of a 2^ Factorial

Source df

A (or BC) 1
B (or AC) 1
C (or AB) 1

Total 3

This would hardly be a practical experiment unless the experimenter is

sure that no first-order interactions exist and has some external source
of error to use in testing A, B, and C. The real advantage of such fractionating
will be seen on designs with larger / values.
If block 1 were run for the experiment instead of block 2, A would have
the value — (1) ab + ac — be, and BC the value -t-(l) — ab — ac A be,
which gives the same total except for sign. The sum of squares due to A and
BC are therefore the same; again they are said to be aliases. Here,
A = -BC B = -AC C = -AB

The dehnition given above still holds where one effect is the alias of another
if they have the same numerical value, or value regardless of sign.
A quick way to find the aliases of an effect in a fractional replication of a
2^ factorial experiment is to multiply the effect by the terms in the defining
306 Fractional replication

contrast, modulus 2. The results will be aliases of the original effect. In the
example above,
/ = ABC
The alias of A is
A{ABC) = A^BC = BC
The alias of B is
BiABC) = AB^C = AC
The alias of C is
C{ABC) = ABC^ = AB
This simple rule works for any fractional replication for a 2^ factorial. It
works also with a slight modification for a 3^ factorial run as a fractional
replication.

15.3 FRACTIONAL REPLICATIONS

2-^ Factorials

As an example, consider the problem suggested in the introduction. An

experimenter wishes to study the effect of seven factors, each at two levels,
but cannot afford to run all 128 experiments and so will settle for 64, or a
one-half replicate of a 2^. Deciding to confound the highest order interaction
with blocks, the experimenter has
/= ABCDEFG
The two blocks are found by placing (1) and all pairs, quadruples, and
sextuples of the seven letters in one block and the single letters, triples,
quintuples, and one septuple of the seven letters in the other block. One of
the two blocks is chosen at random and run. Before carrying out this experi-
ment, the experimenter should check on the aliases. From the defining
contrast
/= ABCDEFG
the alias of A is A{ABCDEFG) = BCDEFG, and all main effects are likewise
aliased with fifth-order interactions.
AB is aliased with AB{ABCDEFG) = CDEFG, a fourth-order interac-
tion. So all first-order interactions are aliased with fourth-order interactions.
A second-order interaction such as ABC is aliased with a third-order inter-
action as
ABC(ABCDEFG) = DEFG
If the second- and third-order interactions are taken as error, the analysis
will be as in Table 15.4.
This is a very practical design, as there are good tests on all main effects
and first-order interactions, assuming all higher order interactions are zero.
The degrees of freedom for each test would be 1 and 35. If, for some reason
 15.3 Fractional replications 307

Table 15.4 One-Half Replication of a 2^ Factorial

Source df

Main effects A, B,..., G 1 each for 7

(or fifth order)
First-order interaction AB, AC 1 each for 21
(or fourth order)
Second-order interaction ABC, ABD 1 each for 35 > use as error
(or third order) J
Total 63

known to the experimenter, some second-order interaction were suspected,

the experimenter could leave it out of the error estimate and still have
sufficient degrees of freedom for the error estimate. The analysis of such an
experiment follows the methods given in Chapter 6.
If this same experimenter is further restricted and can only afford to run
32 experiments, a one-fourth replication of a 2^ might be tried. Here 3 df must
be confounded with blocks. If two fourth-order interactions are confounded
with blocks, one third-order is automatically confounded also, as

/ = ABCDE - CDEFG = ABEG

which confounds 3 df with the four blocks. In this design, if only one of the
four blocks of 32 observations is run, each effect has three aliases. These are

A = BCDE = ACDEEG = BEG CD =■ ABE = EFG = ABCDEG

B - ACDE = BCDEFG = AEG CE = ABD = DEG = ABCEFG
C = ABDE = DEFG = ABCFG CF = ABDEF = DEG = ABCG
D = ABCE = CEFG = ABDFG CG = ABDEG = DEE - ABCE
E = ABCD = CDFG = ABEFG DE = ABC = CFG = ABDEFG
F = ABCDEF = CDEG = ABG DF = ABCEF = CEG = ABDG
G = ABCDEG = CDEF - ABF DG = ABCEG = CEE = ABDE
AB = CDE = ABCDEFG = EG EE = ABCDE = CDG = ABEG
AC = BDE = ADEFG = BCFG EG = ABCDG = CDF = ABEF
AD = BCE = ACEFG = BDFG ACE - BDEF = ADEG = BCG
AE = BCD = ACDFG - BEFG ACG = BDEG = ADEF = BCE
AF = BCDEF = ACDEG = BG ADE - BCEF = ACEG = BDG
AG - BCDEG = ACDEF = BE ADG = BCEG = ACEF = BDE
BC = ADE = BDEFG - ACFG AEG = BCDG = ACDE = BEE
BD = ACE = BCEFG = ADFG BEG = ACDG = BCDE = AEF
BE = ACD - BCDEG = AEFG
308 Fractional replication

Table 15.5 One-Fourth Replication of a 2^ Factorial

Source df

Main effects A, B, . . ., G 1 each for 7

First-order interaction AC, AD,. .. 1 each for 15
AB (or FG), AF (or BG), AG (or BF),. . . 1 each for 3
Second-order interaction or higher (ACF,. ..) 1 each for 6

Total 31

This is quite a formidable list of aliases; but when only one block of 32 is run,
there are 31 df within the block. The above list accounts for these 31 df. If,
in this design, all second-order (three-way) and higher interactions can be
considered negligible, the main effects are all clear of first-order interactions.
Three first-order interactions {AB, AF, and AG) are each aliased with another
first-order interaction (FG, FG, and BF). If the choice of factors A, B, F, and
G can be made with the assurance that the above interactions are either
negligible or not of sufficient interest to be tested, these 3 df can be either
pooled with error or left out of the analysis. The remaining 15 first-order
interactions are all clear of other interactions except second order or higher.
There are also 6 df left over for error involving only second-order interactions
or higher. An analysis might be that shown in Table 15.5.
Here tests can be made with 1 and 6 df or 1 and 9 df if the last two lines
can be pooled for error. This is a fairly good design when it is necessary to
run only a one-fourth replication of a 2^ factorial.
Example 15.1 In an experiment on slicing yield from pork bellies, four
factors were considered as possible sources of significant variation. These
were A—press effect; B—molding temperature; C—slicing temperature;
D—side of the hog. If each of these four factors were set at two levels, it
would require 2"^ = 16 experiments to run one complete replication. If only
eight bellies could be handled at one time, it was decided to run a one-half
replicate of this 2"’’ factorial and confound ABCD with blocks. If / = ABCD,
the blocks are
block I (1) ah cd abed ae be ad bd

block II a b aed bed e abc d abd

Now if only block I is run. Table 15.6 shows the data where entries are for
the treatment combinations from block I only.
The analysis could be handled by a Yates’ method using only the one-half
replication data available as shown in Table 15.7. Note that the numerical
alias of A is A(ABCD) = BCD, of F = ACD, etc. (shown by the arrows).
The analysis follows in Table 15.8.
Table 15.6 Slicing Yields in Percent from Pork Belly Experiment (Example 15.1)

Press {A)
Normal High
Molding Temperature (B) Molding Temperature
Slicing
Temperature Side 23° 40° 23° 40°

23° E (1) = 95.29 b a ab = 86.58

R d bd = 89.38 ad = 96.45 abd
30° L c be = 86.79 ac = 88.70 abc
R cd = 90.35 bed acd abed = 89.57

Table 15.7 Yates’ Solution of Example 15.1

Treatment
Combination Y (1) (2) (3) (4) ss
(1) 95.29 95.29 181.87 357.36 723.IN
a 86.58 175.49 365.75 -0.69K\ 0.06
b 88.70 185.83 -6.80 -18.47,^\\ 42.64
ab 86.58 86.79 179.92 6.29 0.47,<X^ 0.03
e 96.45 -8.71 -10.62 -12.29 \ 18.88
ae 88.70 89.38 1.91 -7.85 2.77r\\\ \\ 0.96
be 86.79 90.35 7.07 6.38 13.09A\\\ \\ 21.42
abe 89.57 -0.78 -5.91 8.390 \ ' 8.78
d -95.29 -8.71 -6.38 8.392J j
ad 96.45 86.58 -1.91 -5.91 13.09^ J
bd 89.38 88.70 -7.07 10.62 2.71'^///1
abd -86.79 -0.78 -7.85 -12.29'V^
ed 90.35 96.45 181.87 6.80
aed -89.38 -175.49 6.29 -\ZAlW /
bed -90.35 -185.83 -357.36 -0.69'^
abed 89.57 89.57 179.92 365.75 723.1 r

Table 15.8 ANOVA for Example 15.1

Source df SS MS F Prob.

A (or BCD) 1 0.06 0.06 <1

B (or ACD) 1 42.64 42.64 5.71 0.097
C (or ABD) 1 18.88 18.88 2.53 0.210
D (or ABC) 1 8.78 8.78 1.18 0.357
AB or CD n1 0.033
AC or BD 1 ^ 3 0.96 > 22.41 7.47
AD or BC ij' 2I.42J
Total 7

309
310 Fractional replication

This experiment is not too good because the two way interactions are
hopelessly confounded and the degrees of freedom on error are so small.
Nothing shows significance but the analysis concept is seen by this example.

3^ Factorials
The simplest 3^ factorial is a 3^ requiring nine experiments for a complete
factorial. If only a fraction of these can be run, we might consider a one-third
replication of a 3^ factorial. First we confound either AB or AB^ with the
three blocks. Confounding AB^ gives
/ = AB^
L = + 23^2
and the three blocks are
L = 0 L=1 L = 2

00 10 20
11 21 01
22 02 12

If one of these three is run, the aliases are

AiAB^) = A^B^ = A^B^ = AB
since the exponent on the first element is never left greater than one. A^B^ is
squared to give A'^B'^ = AB (modulus 3). However, the alias of B is
B{AB^) = AB^ ^ A
Hence all three effects A, B, and AB are aliases and mutually confounded.
To get both aliases from one multiplication using the defining contrast, one
multiplies the effect by the square of / as well as /. This gives
A[{AB^)y = A{A^B^) - A^B = B
Hence A = AB = B; this is a poor experiment, since, within the block that
is run, both degrees of freedom confound A, B, and AB. In addition, no
degrees of freedom are left for error. This example is cited only to show the
slight modification necessary when determining the aliases in a 3^ factorial
run as a fractional replication. The value of these fractional replications is
seen when more factors are involved.
Consider a 3^ factorial in three blocks of nine treatment combinations
each. The effects can be broken down into 13 2-df effects: A, B, C, AB, AB^,
AC, AC\ BC, BC\ ABC, AB^C, ABC\ and AB^C\ Confounding ABC^
with the three blocks gives
/ = ABC^ L = + X2 + 2X2
 15.3 Fractional replications 311

and the blocks are

L = 0 000 on 022 101 112 120 210 221 202

L = 1 100 111 122 201 212 220 010 021 002

L = 2 200 211 222 001 012 020 no 121 102

If only one of these blocks is now run, the aliases are

A = A(ABC^) = A^BC^ = AB^C

and
A = A(ABC^f = A^B^C^ = B^C = B^C^ = BC^
B = B{ABC^) = AB^C^
and
B = B{ABC^f = A^B^C^ = A^C^ = AC^
C = C{ABC^) = ABC^ = AB
and
C = C{ABC^f = A^B^C^ = A^B^C^^ = ABC
AB^ = AB\ABC^) = A^B^C^ = A^B^C^ = AC
and
AB^ = AB^ABC^ = A^B^C^ = BC

The analysis would be that in Table 15.9.

Table 15.9 One-Third Replication of a 3^ Factorial

Source df

A (or BC^ or AB^C) 2

B{ox AB^C^ or AC^) 2
C (or AB or ABC) 2
AC (or AB^ or BC) 2

Total 8

This design would be somewhat practical if we could consider all

interactions negligible and be content with 2 df for error.
These methods are easily extended to higher order 3^ experiments.
Analysis proceeds as in Chapter 9.
It might be instructive to examine somewhat further this design of a
one-third replication of a 3^. A layout for 3^ as a complete factorial was
given in Table 9.7. If only one third of this were run, say block L = 1, sup-
pose the other entries in Table 9.7 were crossed out, giving the results shown
in Table 15.10.
 312 Fractional replication

Table 15.10 One-Third Replication of a 3^ Factorial

Factor A
Factor B Factor C 0 1 2

0 0 eea 100 2oa

1 -001 m 201
2 002 402 202

1 0 010 ITO 210

1 111
2 1T2 212

2 0 mo. m 220
1 021 m 221
2 122 222

The remaining nine entries in Table 15.10 can now be summarized by

levels of factor C only, to give the result shown in Table 15.11.
A second glance at Table 15.11 reveals that this design is none other
than a Latin square! In a sense our designs have now come full circle, from
a Latin square as two restrictions on a single-factor experiment (Chapter 4)
to a Latin square as a one-third replication of a 3^ factorial. The designs
come out the same, but they result from entirely different objectives. In a
single-factor experiment the general model
Yij = /t + Tj + Sij

is partitioned by refining the error term due to restrictions on the random-

ization to give Yij,^ = fi -Y Tj -\- Pi + 7/, + s'ij where the block and position
effects are taken from the original error term Sij.
In the one-third replication of a 3^, there are three factors of interest'
that comprise the treatments, giving
+ Ai Y Bj Y Cj^ Y Sij,.

Table 15.11 One-Third Replication of a 3^ in

Terms of Factor C

Factor A

Factor B 0 1 2

0 2 0 1
1 0 1 2
2 1 2 0
 15.3 Fractional replications 313

where A, B, and C are taken from the treatment effect. Both models look
alike but derive from different types of experiments. In both cases the
assumption is made that there is no interaction among the factors. This is
more reasonable when the factors represent only randomization restrictions
such as blocks and positions, rather than when each of the three factors is
of vital concern to the experimenter.
By choosing each of the other two blocks in the confounding of ABC^
(L = 0 or L = 2), two other Latin squares are obtained. If another con-
founding scheme is chosen, such as AB^C, ABC, or AB^C^, more and
different Latin squares can be generated. In fact, with four parts of the
three-way interaction and three blocks per confounding scheme, 12 distinct
Latin squares may be generated. This shows how it is possible to select a
Latin square at random for a particular design.
It cannot be overstressed that we must be quite sure that no interaction
exists before using these designs on a 3^ factorial. It is not enough to say
that there is no interest in the interactions, because, unfortunately, the
interactions are badly confounded with main effects as aliases.
Example 15.2 Torque after preheat in pounds was to be measured on some
rubber material with factors A—temperature at 145°C, 155°C, and 165°C;
B—mix I, II, and III; and C—laboratory 1, 2, and 3. The 3^ = 27 experi-
ments seemed too costly, so a one-third replicate was run. Confounding
AB^C gives L = X^ 5- 2X2 + X^ and block composition as follows:

L = 0 000 110 on 121 212 022 220 201 102

L = 1 100 210 111 221 012 122 020 001 202

L = 2 200 010 211 021 112 222 120 101 002

If, now, block L = 1 only is run, the numerical results are as given in
Table 15.12 and the analysis as in Table 15.13.

Table 15.12 Rubber Torque Data of Example 15.2

Temperature [A)

145° 155° 165°

Mix Mix Mix

Laboratory II III II III II III

1 16.8 11.2 9.9

2 15.8 14.4 17.8
3 17.1 20.5 15.7
314 Fractional replication

Table 15.13 ANOVA for Example 15.2

Source df SS MS

A (or ABC^ or BC^) 2 6.66 3.33

B (or AC or ABC) 2 38.13 19.06
C (or AB^ or AB^C^) 2 40.81 20.40
AB (or AC^ or BC) 2 0.12 0.06

Totals 8 85.72

Because of the tiny error term all main effects are highly significant
even with only 2 and 2 degrees of freedom.

15.4 SUMMARY
Continuing the summary at the end of Chapter 14 under incomplete ran-
domized blocks gives

Experiment Design Analysis

II. Two or more

factors
A. Factorial
(crossed)
1. Completely randomized
2. Randomized block
a. Incomplete, confounding
i. Main effect—split-plot
ii. Interactions in 2^ and 3-^
1. Several replications 1. Factorial ANOVA
Yijkq =
with replications
+ + Bq and blocks
+ + ^ikq
2. One replication only 2. Factorial ANOVA
^ijk = with blocks or
+ Bk + ABj,.,. . . confounded
interaction
3. Fractional replication— 3. Factorial ANOVA
aliases with aliases or
a. in 2-^ aliased interactions
b. in 3-^; /’ = 3;
Latin square
bjfc = /^ + Ai + Bj
+ 45,,, ...
 Problems 315

PROBLEMS

15.1 Assuming that not all of Problem 14.1 can be run, determine the aliases if a one-
half replication is run. Analyze the data for the principal block only.
15.2 Assuming only a one-fourth replication can be run in Problem 14.6, determine
the aliases.
15.3 For Problem 14.8 use a one-third replication and determine the aliases.
15.4 Run an analysis on the block in Problem 15.3 that contains treatment combina-
tion 022.
15.5 What would be the aliases in Problem 14.9 if a one-half replication were run?
15.6 Determine the aliases in a one-fourth replication of Problem 14.11.
15.7 Determine the aliases in Problem 14.13 where a one-third replication is run.
15.8 If a one-eighth replication of Problem 14.17 is run, what are the aliases? Comment
on its ANOVA table.
15.9 What are the aliases and the ANOVA outline if only one block of Problem 14.18
is run?
15.10 If data are taken from one furnace only in Problem 14.21, what are the aliases
and the ANOVA outline?
15.11 What are the numerical results and conclusions from just one furnace of Prob-
lem 14.22?
15.12 The experiment of Example 15.1 was later continued and the second block was
now run with results as follows:

Press {A)

Normal High

Molding Temperature {B) Molding Temperature

Slicing
Temperature Side 23° 40° 23° 40°

L b = 88.08 a = 89.26
23°
R d = 90.87 abd = 93.16
L c = 85.95 abc = 85.68
30°
R bed — 83.65 acd = 95.25

Analyze these results.

15.13 Take the results of Example 15.1 and Problem 15.12 and combine into the full
factorial in two blocks of eight and analyze the results and comment.
15.14 a. A four-factor experiment is to be run with each factor at two levels only.
However, only eight treatment combinations can be run in each block. If one
wishes to confound the ABD interaction, what are the block compositions?
316 Fractional replication

b. Outline the ANOVA for just one run of this experiment. Consider all factors
hxed and note what you might use as an error term.
c. If only one block of this experiment can be run, explain what complications
this makes in the experiment and how the ANOVA table must be altered.
15.15 Consider a 2^ factorial experiment with all hxed levels to be run in four blocks
of eight treatment combinations each. This whole experiment can be replicated
in a total of three replications. The following interactions are to be confounded
in each replication:
Replication I — ABC, CDE, ABDE
Replication II — BCDE, ACD, ABE
Replication III — ABCD, BCE, ADE
a. Determine the block composition for the principal block in replication I.
b. If only a one-fourth replication were run of replication II, what would be the
aliases of BC in this block?
c. Determine the number of degrees of freedom for the replication by three-way
interactions in this problem.
15.16 Consider an experiment designed to study the effect of six factors—A, B, C, D,
E, and E—each at two levels.
a. Determine at least six entries in the principal block if this experiment is to be
run in two blocks of 32 and the highest order interaction is to be confounded.
b. If only one of the two blocks can be run, what are the aliases of the effects
A, BC, and /45C?
c. Outline the ANOVA table for part (b) and indicate what effects might be
tested and how they would be tested.
15.17 In determining which of four factors A, B, C, or D might affect the underpuffing
of cereal in a plant, the following one-fourth replicate was run.

Bi B, B2
96.6 125.7
c,1
D, 43.5 22.4

D, 14.1 9.5
c ■)
Di 28.8 52.5

a. Determine which interaction was confounded in this design.

b. Analyze these data and comment on the results and on the design.
15.18 A company manufacturing engines was concerned about the emission charac-
teristics of these engines. To try to minimize various undesirable emission
variables, the company proposed to build and operate experimental engines
varying hve factors; throat diameter to be set at three levels, ignition system at
three levels, temperature at three levels, velocity of the jet stream at three levels,
and timing system at three levels. As this would require 243 experiments, it was
 Problems 317

decided that a one-third replicate would be run. Set up a confounding scheme

for this problem and indicate a few terms in the principal block. Also show an
ANOVA outline when running one of the three blocks, indicating some aliases.
Comment on the design.
15.19 In studying the surface finish of steel five factors were varied each at two levels;
B—slab width; C—percent sulfur; D—tundish temperature; E—casting speed;
and F—mold level. The following data were recorded on the number of longi-
tudinal face cracks observed.

B Slab width, inches

<60 >62

C Sulfur, percent

<0.015 >0.017 <0.015 >0.017

>10

2 3
>59

oo
1 0 0
>2845

>10
Tundish temperature, °F

5 0
Casting speed, 1 pm

Mold level, inches

<54

oo
1 6 0
>10

6 3
>59

F
E

oo
1 2 0
D

ly-)
<2840

>10

7 0
<54

oo
1 1 0

Determine what interaction is confounded here in order to run this one-half

replicate. Outline its ANOVA and comment on the design.
15.20 Using the data of Problem 15.19, do the analysis and state your conclusions.
 16
Miscellaneous Topics

16.1 INTRODUCTION
Several techniques have been developed, based on the general design princi-
ples presented in the first 15 chapters of this book and on other well-known
statistical methods. No attempt will be made to discuss these in detail, as
they are presented very well in the references. It is hoped that a background
in experimental design as given in the first 15 chapters will make it possible
for the experimenter to read and understand the references.
The methods to be discussed are covariance analysis, response surface
experimentation, evolutionary operation, analysis of attribute data, incom-
plete block design, and Youden squares.

16.2 COVARIANCE ANALYSIS

Philosophy

Occasionally, when a study is being made of the effect of one or more factors
on some response variable, say T, there is another variable, or variables,
that vary along with Y. It is often not possible to control this other variable
(or variables) at some constant level throughout the experiment, but the
variable can be measured along with the response variable. This variable is
referred to as a concomitant variable X as it “runs along with” the response
variable Y. In order, then, to assess the effect of the treatments on Y, one
should first attempt to remove the effects of this concomitant variable X.
This technique of removing the effect of X (or several X’s) on Y and then
analyzing the residuals for the effect of the treatments on Y is called co-
variance analysis.
Several examples of the use of this technique might be cited. If one
wishes to study the effect on student achievement of several teaching methods,
it is customary to use the difference between a pretest and a posttest score
as the response variable. It may be, however, that the gain Y is also affected

318
 16.2 Covariance analysis 319

by the student’s pretest score X as gain, and pretest scores may be correlated.
Covariance analysis will provide for an adjustment in the gains due to
differing pretest scores assuming some type of regression of gain on pretest
scores. The advantage of using this technique is that one is not limited to
running an experiment on only those pupils who have approximately the
same pretest scores or matching pupils with the same pretest scores and
randomly assigning them to control and experimental groups. Covariance
analysis has the effect of providing “handicaps” as if each student had the
same pretest scores while actually letting the pretest scores vary along with
the gains.
This technique has often been used to adjust weight gains in animals by
their original weights in order to assess the effect of certain feed treatments
on these gains. Many industrial examples might be cited such as the one
given below where original weight of a bracket may affect the weight of
plating applied to the bracket by different methods.
Snedecor [22] gives several examples of the use of covariance analysis.
Ostle [19] is an excellent reference on the extension of covariance analysis
to many different experimental designs. The Biometrics article of 1957 [5]
devotes most of its pages to a discussion of many of the fine points of covari-
ance analysis. The problem discussed below will illustrate the technique for
the case of a single-factor experiment run in a completely randomized
design. Since the covariance technique represents a marriage between re-
gression and the analysis of variance, the methods of Chapter 3 and Chapter 7
will be used.

Example 16.1 Several steel brackets were sent to three different vendors to
be zinc plated. The chief concern in this process is the thickness of the zinc
plating and whether or not there was any difference in this thickness between
the three vendors. Data on this thickness in hundred thousandths of an inch
(10“^) for four brackets plated by the three vendors are given in Table 16.1.
Assuming a single-factor experiment and a completely randomized
design, an analysis of variance model for this experiment would be

+ C + (16.1)
Table 16.1 Plating Thickness in 10”^ Inches
from Three Vendors

Vendor
A B C

40 25 27
38 32 24
30 13 20
47 35 13
320 Miscellaneous topics

Table 16.2 ANOVA on Plating Thickness

Source df SS MS

Between vendors 2 665.2 332.6

Within vendors 9 543.5 60.4

Totals 11 1208.7

where Y^j is the observed thickness of the ith bracket from the jth vendor,
ji is 3. common effect, TJ represents the vendor effect, and Sij represents the
random error. Considering vendors as a fixed factor and random errors as
normally and independently distributed with mean zero and common vari-
ance (jg, an ANOVA table was compiled (Table 16.2).
The F statistic equals 5.51 and is significant at the 5 percent significance
level with 2 and 9 df One might conclude, therefore, that there is a real
difference in average plating thickness among these three vendors, and
steps should be taken to select the most desirable vendor.
During a discussion of these results, it was pointed out that some of
this difference among vendors might be due to unequal thickness in the
brackets before plating. In fact, there might be a correlation between the
thickness of the brackets before plating and the thickness of the plating. To
see whether or not such an idea is at all reasonable, a scatterplot was made
on the thickness of the bracket before plating X and the thickness of the
zinc plating Y. Results are shown in Figure 16.1.
A glance at this scattergram would lead one to suspect that there is a
positive correlation between the thickness of the bracket and the plating

50

40

30
Y
20

10

0 30 60 90 120 150
V

Figure 16.1 Scattergram of bracket thickness X versus plating thickness Y in

10“^ inches.
 16.2 Covariance analysis 321

thickness. Since another variable, X, can be measured on each piece and

may be related to the variable of interest Y, a covariance analysis may be
run in order to remove the linear effect of X on 7 when comparing vendors.
A covariance analysis is used to remove the linear or higher order relation-
ship of one or more independent variables from a dependent variable in
order to assess the effect of a factor on the dependent variable.

Covariance

Since a positive correlation is suspected between X and Y above, a regression

model might be written as

Y,j = fiY PiX^j - X) + 8,j (16.2)

where P is the true linear regression coefficient (or slope) between Y and X
over all the data; X is the mean of the X values. In such a model it is assumed
that a measure of variable X can be made on each unit along with a corre-
sponding measure of Y. For a covariance analysis models (16.1) and (16.2)
can be combined to give the covariance model

Yjj = 11 + Pi^ij — X) + Tj + Sjj (16.3)

In this model the error term eij should be smaller than in model (16.1)
because of the removal of the effect of the covariate X^j.
To determine how to adjust the analysis of variance to provide for the
removal of this covariate, consider the regression model (16.2) in deviation
form
y = bx + e (16.4)

where x and y are now deviations from their respective means (x = X^j — X,
y = Yij — Y) and b is the sample slope. By the method of least squares it
will be recalled that

where summations are over all points. The sum of the squares of the errors
of estimate, which have been minimized, is now, from Equation (16.4),
e = y — bx

y = y (y - hx}^

= - Ib^xy +

Substituting for b
Z xy
322 Miscellaneous topics

and
Z xy
Z xy + 'Z z
z^^
= z/- (Zz xy)^ (16.5)

In this expression the term ^y)^lYu 1^® amount of reduction in the

sum of squares of the variable due to its linear regression on x. This term
is then the sum of squares due to linear regression. It involves the sum of
the cross products of the two variables xj;) and the sum of squares of the
X variable alone x^). If a term of this type is subtracted from the sum of
squares for the dependent variable y, the result will be a corrected or adjusted
sum of squares for y.
The sums of squares and cross products can be computed on all the
data (total), within each vendor or between vendors. In each case it will be
helpful to recall that the sums of squares and cross products in terms of the
original data are

z x^
(II^
N

iZll
N

(Z
N

Data on both variables appear with the appropriate totals in Table 16.3.
From Table 16.3 the sums of squares and cross products for the total
data are^

2
r,, = (110)^ + (75)^ + ••• + (59)^ -
(944)2 = 9240.7
12

- (40)2 + (38)2 ^ ^3) 2 (344)2 = 1208.7

yy
12

(944)(344)
T,, = (110)(40) + (75)(38) + • • • + (59)(13) - = 2332.7
12
The between-vendors sums of squares and cross products are computed by

' Here T, V, and E are used to denote sum of squares and cross products for totals, vendors,
and error respectively.
 16.2 Covariance analysis 323

Table 16.3 Bracket Thickness X and Plating Thickness Y in

10“^ Inches from Three Vendors

Vendor
A B C Total

X y X y X y X y

no 40 60 25 62 27
75 38 75 32 90 24
93 30 38 13 45 20
97 47 140 35 59 13

Totals 375 155 313 105 256 84 944 344

vendor totals as in an ANOVA:

(375)" + (313)2 (256)2 (944)2
V
r V -y- = - 1771.2
4 12

(155)^ + (105)2 + (84)2 (344)2

V = = 665.2
4 14~

(375)(155) + (313)(105) + (256)(84) (944)(344)

V
' xy = = 1062.2
4 ’ 12

By subtraction the error sums of squares are

= 9240.7 - 1771.2 - 7469.5

Eyy = Tyy - Vyy = .l - = 543.5

^xy = T^y - V^y = 2332.7 - 1062.2 - 1270.5

Using this information, the sums of squares of the dependent variable Y
may be adjusted for regression on X. On the totals, the adjusted sum of
squares is then
j2
adjusted ^ = Tyy - as in Equation (16.5)
^ XX

(2332.7)2
= 1208.7 619.8
9240.7

and the adjusted sum of squares within vendors is

adjusted = Eyy —
E XX

(1270.5)^
= 543.5 - = 327.4
7469.5
324 Miscellaneous topics

Then, by subtraction, the adjusted sum of squares between vendors is

619.8 - 327.4 = 292.4

These results are usually displayed as in Table 16.4.

Table 16.4 Analysis of Covariance for Bracket Data

SS and Products
Adjusted
Source df IV I^^ IV df MS

Between vendors 2 665.2 1062.2 1771.2 — — —

Within vendors 9 543.5 1270.5 7469.5 327.4 8 40.9

Totals 11 1208.7 2332.7 9240.7 619.8 10

Between vendors 292.4 2 146.2

With the adjusted sums of squares, the new F statistic is

146.2
3.57
40.9
with 2 and 8 df. This is now not significant at the 5 percent significance level.
This means that after removing the effect of bracket thickness, the vendors
no longer differ in the average plating thickness on the brackets.
Table 16.4 should have a word or two of explanation. The degrees of
freedom on the adjusted sums of squares are reduced by 2 instead of 1 as
estimates of both the mean and the slope are necessary in their computation.
Adjustments are made here on the totals and the within sums of squares
rather than on the between sum of squares as we are interested in making
the adjustment based on an overall slope of 7 on AT and a within-group
average slope of T on 2f as explained in more detail later.
In this analysis three different types of regression can be identified: the
“overah” regression of all the T’s on all the X’s, the within-vendors regres-
sion, and the regression of the three Y means on the three X means. This last
regression could be quite different from the other two and is of little interest
since we are attempting to adjust the T’s and the X’s within the vendors. An
estimate of this “average within vendor” slope is

1270.5
0.17
“ 7469.5
If this slope or regression coefficient is used to adjust the observed Y means
for the effect of X on T, one finds

adjusted Y j = Y j — b{X j — )
 16.2 Covariance analysis 325

For vendor A,

adjusted F,i = ^ - (0.17)^^ - ^

= 38.75 - (0.17)(93.75 - 78.67) = 36.19

For vendor B,

adjusted Y 2 = 26.25 - (0.17)(78.25 - 78.67) = 26.31

For vendor C,

adjusted Y ^ = 21.00 - (0.17)(64.00 - 78.67) = 23.49

A comparison of the last column above with the column of unadjusted
means just to the right of the equal signs shows that these adjusted means are
closer together than the unadjusted means which is confirmed in the pre-
ceding analysis. This can be seen graphically by plotting each vendor’s data in
a different symbol and “sliding” the means along lines parallel to this regres-
sion slope (Figure 16.2).

a = vendor A
A - mean for vendor A
b = vendor B
B = mean for vendor B
c = vendor C
C = mean for vendor C

Figure 16.2 Plating thickness Y versus bracket thickness X plotted by vendors.

At the left-hand ordinate in Figure 16.2, the three unadjusted Y means

are marked off. By moving these three means along lines parallel to the gen-
eral slope {h = 0.17), a comparison is seen between the three means (on the
X ordinate) after adjustment. This shows graphically the results of the
covariance analysis. By examining each group of observations for a given
vendor, it is seen that all three slopes within vendors are in the same direction
and of similar magnitude. By considering all 12 points, the overall slope can
be visualized. On the other hand, if the three means (indicated by the capital
letters) are considered, the slope of these T’s on their corresponding X’s is
326 Miscellaneous topics

much steeper. Hence adjustments are made on the basis of the overall slope
and the “average” or “pooled” within-groups slope.
One of the problems in covariance analysis is that there are many
assumptions made in its application. These should be examined in some
detail.
First, there are the usual assumptions of analysis of variance on the
dependent variable Y: an additive model, normally and independently dis-
tributed error, and homogeneity of variance within the groups. This last
assumption is often tested with Barlett’s or a similar test. In this problem,
since the ranges of the three vendor readings are 17, 22, and 14, the assump-
tion seems tenable.
In covariance analysis it is further assumed that the regression is linear,
that the slope is not zero (covariance was necessary), that the regression co-
efficients within each group are homogeneous so that the “average” or
“pooled” within-groups regression can be used on all groups, and finally
that the independent variable X is not affected by the treatments given to the
groups.
The linearity assumption may be tested if the experiment is planned in
such a way that there is more than one Y observation for each X. Since this
was not done here, a look at the scattergram will have to suffice for this
linearity assumption.
To test the hypothesis that the true slope p in Equation (16.3) is zero,
consider whether or not the reduction in the error sum of squares is significant
when compared with the error sum of squares. By an F test the ratio would be

adjusted Eyy/{N — k — 1)

^ _ 216.1 _ 216.1 _
“ 327.4/8 ~ ~W.9 ~ ■

As the 5 percent F is 5.32, this result is nearly significant and certainly casts
considerable doubt on the hypothesis that p = 0. Since the results after
covariance adjustment were not significant as compared with the results
before adjustment, it seems reasonable to conclude that covariance was
necessary.
To test the hypothesis that all three regression coefficients are equal, let
us compute each sample coefficient and adjust the sum of squares within each
group by its own regression coefficient.
For vendor A,

. ^ Z-xy _ y Ary - (ZA:)(Z Y)ln

z Z - (Z
 16.2 Covariance analysis 327

Within group A,

14,599 - 58,125/4 ^
35,783 - 140,625/4

adjusted ^ = YJ ^^ ^ 146.75 — 7.32 = 139.43

z X

For vendor 5,
, 9294 - 32,865/4
® ~ 30,269 - 24,492.25/4 “ '

adjusted = 286.75 - 201.72 = 85.03

For vendor C,
5501 - 21,504/4
^ 17,450 - 65,536/4
adjusted ^ = 110.00 — 14.66 == 95.34

Summarizing, we obtain Table 16.5.

Table 16.5 Adjusted Sums of Squares within Each Vendor

Within Adjusted
Vendor df b df

A 3 146.75 0.108 139.43 2

B 3 286.75 0.187 85.03 2
C 3 110.00 0.117 95.34 2

Totals 9 543.50 319.80 6

Adjusting each within-vendor sum of squares by its own regression

coefficient reduces the degrees of freedom to 2 per vendor. Adding the new
adjusted sums of squares. Table 16.5 shows a within-vendor sum of squares
of 319.80 based on 6 df. When the within-vendor sum of squares was adjusted
by a “pooled” within-vendor regression. Table 16.4 showed an adjusted sum
of squares within vendors of 327.4 based on 8 df. If there were significant
differences in regression within the three vendors, this would show as a
difference in these two figures: 327.4 — 319.8 = 7.6. An F test for this
hypothesis would then be

l, N-2k ~

adjusted Y (based on pooled within groups regression)

{k - 1)
— adjusted Y f ^ (based on regressions within each group)
[adjusted Y (within each group)]/(N — 2k)
328 Miscellaneous topics

and
(321A - 319.8)/2 3.8
^ 2,6 ~~
= <1
319.8/6 53.3

hence nonsignificant, and we conclude that the three regression coefficients

are homogeneous.
The final assumption that the vendors do not affect the covariate X is
tenable here on practical grounds as the vendor has nothing to do with the
bracket thickness before plating. However, in some applications this assump-
tion needs to be checked by an F test on the X variable. For our data from
Table 16.4 such an F test would give

_ 1771.2/2 _ 885.6
1.07
“ 7469.5/9 “ 829.9

which is obviously nonsignificant.

From this discussion of the assumptions, one notes that it is often a
problem in the use of covariance analysis to satisfy all of the assumptions.
This may be the reason many people avoid covariance although it is applic-
able whenever one suspects that another measured variable or variables affect
the dependent variable.
The above techniques may be extended to handle several covariates,
nonlinear regression, and several factors and interactions of interest. In
handling randomized blocks, Latin squares, or experiments with several
factors, Ostle [19] points out that the proper technique is to add the sum of
squares and sum of cross products of the term of interest to the sum of
squares and sum of cross products of the error, adjust this total, then adjust
the error and determine the adjusted treatment effect by subtraction.
When several covariates are involved, miany questions arise concerning
how many covariates can be handled, how many really affect the dependent
variable, how a significant subset can be found, and so on. All of these
questions are of concern in any multiple regression problem and some dis-
cussion of these and other problems can be found in [5].

16.3 RESPONSE-SURFACE EXPERIMENTATION

PhUosophy

The concept of a response surface involves a dependent variable T, called

the response variable, and several independent or controlled variables.
A 25 Xu. If all of these variables are assumed to be measurable, the
response surface can be expressed as

y = /(Ai, A^, . . ., A,)

 16.3 Response-surface experimentation 329

(a)

(d)

Figure 16.3 Some typical response surfaces in two dimensions: (a) mound;
(b) depression; (c) rising ridge; (d) saddle.

For the case of two independent variables such as temperature and time
^2, the yield 7 of a chemical process can be expressed as

Y = fix
This surface can be plotted in three dimensions, with X^ on the abscissa,
X on the ordinate, and Y perpendicular to the X^X plane. If the values
2 2

of X^ and X that yield the same Y are connected, we can picture the surface
2

with a series of equal-yield lines, or contours. These are similar to the contours
of equal height on topographic maps and the isobars on weather maps.
Some response surfaces might be of the types in Figure 16.3. Two excellent
references for understanding response-surface experimentation are [9] and
[12].
The Twofold Problem

The problem involved in the use of response-surface experimentation is two-

fold: (1) to determine, on the basis of one experiment, where to move in the
330 Miscellaneous topics

next experiment toward the optimal point on the underlying response surface;
(2) having located the optimum or near optimum of the surface, to determine
the equation of the response surface in an area near this optimum point.
One method of experimentation that seeks the optimal point of the
response surface might be the traditional one-factor-at-a-time method. As
shown in Figure 16.3(a), if X is fixed and
2 is varied, we find the X^ optimal
(or near optimal) value of response Y at the fixed value of X2. Having found
this X^ value, experiments can now be run at this fixed X^, and the X for 2

optimal response can be found. In the case of the mound in Figure 16.3(a),
this method would lead eventually to the peak of the mound or near it.
However, this same method, when applied to a surface such as the rising
ridge in Figure 16.3(c), fails to lead to the maximum point on the response
surface. In experimental work the type of response surface is usually un-
known, so that a better method is necessary if the optimum set of conditions
is to be found for any surface.
The method developed by those who have worked in this area is called
the path of steepest ascent method. The idea here is to run a simple experiment
over a small area of the response surface where, for all practical purposes,
the surface may be regarded as a plane. We then determine the equation of
this plane and from it the direction we should take from this experiment in
order to move toward the optimum of the surface. Since the next experiment
should be in a direction in which we hope to scale the height the fastest, this
is referred to as the path of steepest ascent. This technique does not determine
how far away from the original experiment succeeding sequential experiments
should be run, but it does indicate to the experimenter the direction along
which the next experiment should be performed. A simple example will
illustrate this method.
To determine the equation of the response surface, several special
experimental designs have been developed that attempt to approximate this
equation using the smallest number of experiments possible. In two dimen-
sions the simplest surface is a plane given by

Y = BQXQ + B^X^ + B X
2 2 + 8 (16.6)
where Y is the observed response, XQ is taken as unity, and estimates of the
JB’S are to be determined by the method of least squares, which minimizes the
sum of the squares of the errors e. Such an equation is referred to as a first-
order equation, since the power on each independent variable is unity.
If there is some evidence that the surface is not planar, a second-order
equation in two dimensions may be a more suitable model
Y = BQXQ + B^X^ + B X 2 2 B^^Xf + B^2^i^2 T -^22-^2 T ^ (16.7)

Here the X^X term represents an interaction between the two variables X^
2

and X -
2
 16.3 Response-surface experimentation 331

If there are three independent or controlled variables, the first-order

equation is again a plane or hyperplane

Y = BQXQ + + B X2 2 + B^X^ + £ (16.8)

and the second-order equation is

Y = BQXQ + B,X^ + B2X2 + B2X2 + B^^Xj + B22XI

+ B,,Xl + B,2X^X2 + + B23X2X2 + e (16.9)
As the complexity of the surface increases, more coefficients must be esti-
mated, and the number of experimental points must necessarily increase.
Several very clever designs have been developed that minimize the amount of
work necessary to estimate these response-surface equations.
To determine the coefficients for these more complex surfaces and to
interpret their geometric nature, both multiple regression techniques and
the methods of solid analytical geometry are used. In the example that
follows, only the simplest type of surface will be explored. The references
quoted will give many more complex examples.

Example 16.2 Consider an example in which an experimenter is seeking the

proper values for both concentration of filler to epoxy resin X^ and position
in the mold X to minimize the abrasion on a plastic die. This abrasion or
2

wear is measured as a decrease in thickness of the material after 10,000 cycles

of abrasion. Since the maximum thickness is being sought, the first experi-
ment should attempt to discover the direction in which succeeding experi-
ments should be run in order to approach this maximum by the steepest
path. Assuming the surface to be a plane in a small area, the first experiment
will be used to determine the equation of this plane. The response surface
is then
Y — BQXQ + B^X^ + B2X2 + e

As there are three parameters to be estimated, BQ, B^, and B , at least three 2

experimental points must be taken to estimate these coefficients. Such a

design might be an equilateral triangle, but since there are two factors, X ^
and X , each can be set at two levels and a 2^ factorial may be used. Two
2

concentrations were chosen, 1 and 1:1 (ratio of filler to resin), and two
positions, 1 inch and 2 inches from a reference point, with responses Y, the
thickness of the material in lO""'' inches, as shown in Figure 16.4.
To determine the equation of the best fitting plane for these four points,
consider the error in the prediction equation

e = y — BQXQ — B^X^ — B2X2

The sum of the squares for this error is

= y (7 - Bo^o - B,X, - B^X^f

332 Miscellaneous topics

5.4 6.9
+1 2 in.

Position X2

-1 1 in.

1;1
Concentration

-1 +1

Figure 16.4 2^ factorial example on response surface.

where the summation is over all points given in the design. To find the 5’s,
we differentiate this expression with respect to each parameter and set these
three expressions equal to zero. This provides three least squares normal
equations, which can be solved for the best estimates of the J5’s. These esti-
mates are designated as fi’s.
Differentiating the expression above gives

= -2 X (r - = 0

= -2X(y - BoXo - B,X, - B2X2)X, = 0

= -2 ^ (y - BoX„ - BiXi - B X )X 2 2 2 = 0

and from these results we get

X XoF = ho X ^0 + C Z 2fo2f 1 + h2 Z 2^0X2

Z x. y = ho Z 2^o2^i + C Z 2^? + C Z 2^12^2 (16.10)

Z 2^2^ = ho Z 2fo2f2 + C Z 2^I2^2 + C Z 2^2
as the least squares normal equations.
By a proper choice of experimental variables, it is possible to reduce these
equations considerably for simple solution. More complex models can be
solved best by the use of matrix algebra. For the 2“ factorial, the following
coding scheme simplifies the solution of Equation (16.10). Set
XI = 4C — 3 where C is the concentration
X 2 ^ 2P — 3 where P is the position
 16.3 Response-surface experimentation 333

Table 16.6 Orthogonal Layout for

Figure 16.4

Y Xo Xi X2
7.0 1 -1 -1
5.4 1 -1 1
7.1 1 1 -1
6.9 1 1 1

XQ is always taken as unity. For the experimental variables X-^ and X2, the
responses can be recorded as in Table 16.6. These experimental variables are
also indicated on Figure 16.4.
An examination of the data as presented in Table 16.6 shows that XQ,
X^, and X2 are all orthogonal to each other as

= X X2 = 0 and X ^1^2 = 0

Hence the least squares normal equations become

X XQY = bQU + 61 • 0 + ^2 • 0
Y,X,Y = ho-o + h.-Y^xi + b^-o (16.11)

X ^2^ = 60 • 0 4- • 0 + 62 X ^2
Solving gives
^0 = i:XoY/n
(16.12)

b2 = X^2 5"/X ^2
For this problem

1.6
61 = = 0.40
4

bi = = -0.45
4

and the response surface can be approximated as

Y = 6.60 + 0.40X1 - 0.45X2

To determine the sum of squares due to each of these terms in the model,
we can use the sum of squares due to h,’s
334 Miscellaneous topics

Here
= (6.60)(26.4) = 174.24
= (0.40)(1.6) = 0.64
SS,^ = (-0.45)(-1.8) = 0.81
each carrying 1 df. The ANOVA table is shown as Table 16.7.

Table 16.7 ANOVA for 1? Factorial Response-Surface

Example

Source df SS

bo 1 174.24
hi 1 0.64
bi 1 0.81
Residual 1 0.49

Totals 4 176.18

Here, all 4 df are shown, since the ho term represents the degree of
freedom usually associated with the mean, that is, the correction term in
many examples. The total sum of squares is ^ of the responses and the
residual is what is left over. With this 1-df residual, no good test is available
on the significance of each term in the model, nor is there any way to assess
the adequacy of the planar model to describe the surface.
To decide on the direction for the next experiment, plot contours of
equal response using the equation of the plane determined above

y = 6.60 + 0.40X1 - 0.45X2

Solve for X,2

6.60 - y + 0.40X1
045
If y =
1.10 + 0.40X1
045
when
Xi = -1 X 2 = 1.56
Xi = 1 X2 = 3.33

0.60 + 0.40X1
0.45
 16.3 Response-surface experimentation 335

when
X, = -1 X2 = 0.44
X^ = 1 ^2 = 2.22
If y -
0.10 + 0.40Xi
~ 045
when
X^ = -1 X2= -0.67
= 1 X2 = 1.11

-0.40 + 0.40X1
045
when
Xi = -1 X2 = -1.67
Xi = 1 X2 = 0
If y =
_ -0.90 + 0.40X1
^ 0.45
when
Xi = -1 X2 = -2.89
Xi = 1 X2 = -1.11
Plotting these five contours on the original diagram gives the pattern shown
in Figure 16.5.
By moving in a direction normal to these contours and “up” the surface,
we can anticipate larger values of response until the peak is reached. To

^2

Figure 16.5 Contours on 2^ factorial response-surface example.

336 Miscellaneous topics

decide on a possible set of conditions for the next experiment, consider the
equation of a normal to these contours through the point (0, 0). The con-
tours are
6.60 - T + 0.40Zi
2
045

and their slope is 0.40/0.45 = 8/9. The normal will have a slope = —9/8,
and its equation is

X2-0 = ~^{X,-0)

-9
X2 = —5— (see arrow in Figure 16.5)
8

This technique will not tell how far to go in this direction, but we might run
the next experiment with the center of the factorial at ( +1, —9/8). The four
points would be

3^2

1
0 8
0 — 7^
1
2 8
2 — 7^

These points are expressed in terms of the experimental variables, but must
be decoded to see where to set the concentration and position. Using the
coding,

X^ = 4C — 3 or C =
4
X, + 3
X, = 2P - 3 or P = —
2

the four new points would be

X, ^2 C P
1 3.1
0 8 4.1
li%
7
0 — 2-
^8 |:1 16
1
2 8 U:1 ll%
2 — 2^ li:l 7
^8 16
 16.3 Response-surface experimentation 337

if these settings are possible. Responses may now be taken at these four
points on a new 2^ factorial, and after analysis we can again decide the
direction of steepest ascent. This prodecure continues until the optimum is
obtained.
The above design is sufficient to indicate the direction for subsequent
experiments, but it does not provide a good measure of experimental error
to test the significance of b^, b^, and bj, nor is there any test of how well the
plane approximates the surface. One way to improve on this design is to take
two or more points at the center of the square. By replication at the same
point, an estimate of experimental error is obtained and the average of the
center-point responses will provide an estimate of “goodness of fit” of the
plane. If the experiment is near the maximum response, this center point
might be somewhat above the four surrounding points, which would indicate
the need for a more complex model.
To see how this design would help, consider two observations of response
at the center of the example above. Using the same responses at the vertices
of the square, the results might be those shown in Figure 16.6.

Y ^0 ^2

7.0 1 -1 -1
5.4 1 -1 1
7.1 1 1 -1
6.9 1 1 1
6.6 1 0 0
6.8 1 0 0

Figure 16.6 2^ factorial with two points in center.

338 Miscellaneous topics

The coefficients in the model are still estimated by Equation (16.12):

39.8
= 6.63
n

1.6
= 0.40
T
- 1.8
-0.45
4
and
SSb, = 6.63(39.8) = 263.87
= 0.40(1.6) = 0.64
SS,^ = -0.45(-1.8) = 0.81

For the sum of squares of the error at (0, 0),

(13.4)"
SS, = (6.6)" + (6.8)" = 89.80 - 89.78 = 0.02
2
The analysis is shown in Table 16.8.
Testing these effects against error gives
^ _ 236.87
= 11,843.5
“ ^ooT
0.64
^1- ^1,1 = 32

0.81
bi-Fi 1 = TT— = 40.5
^ 0.02
0.32
lack of fit: F2 1 = = 16
0.02
With such a small number of degrees of freedom, only bo shows signifi-
cance at the 5 percent level. However, the tests on the b terms are larger than
the test on lack of fit, which may indicate that the plane
y = 6.63 + 0.40X1 - 0.45X2

is a fair approximation to the surface where this first experiment was run.

More Complex Surfaces

For more factors—controlled variables—the model for the response variable

is more complex, but several designs have been found useful in estimating
the coefficients of these surfaces.
 16.3 Response-surface experimentation 339

Table 16.8 ANOVA for 2^ + Two

Center Points

Source df SS MS

Total 6 265.98

bo 1 263.87 236.87
bi 1 0.64 0.64
b2 1 0.81 0.81
Residual 3 0.66 —

Error 1 0.02 0.02

Lack of fit 2 0.64 0.32

When three variables are involved, a first approximation is again a

plane or hyperplane of the form

Y = BQXQ + + B2X2 + B2X2, + e

For this first-order surface, at least four points must be taken to estimate the
four^’s. Since three dimensions are involved, we might consider a 2^ factorial.
As this design has eight experimental conditions at its vertices, the design
often used is a one-half replication of a 2^ factorial with two or more points
at the center of the cube. These six points (two at the center) are sufficient to
estimate all four B’s and to test for lack of fit of this plane to the surface in
three dimensions. After this initial one-half replication of the 2^ the other
half might be run, giving more information for a better fit.
If a plane is not a good fit in two dimensions, a second-order model
might be tried. Such a response surface has the form

y = BQXQ + B^Xi + B2X2 + B^^X^ -f BI2^I^2 + ^22^2 + ^

Here six B's are to be estimated. The simplest design for this model would be
a pentagon (five points) plus center points. This would yield six or more
responses, and all six S’s could be estimated.
In developing these designs Box and others found that the calculations
can be simplified if the design can be rotated. A rotatable design is one that
has equal predictability in all directions from the center, and the points are
at a constant distance from the center. All first-order designs are rotatable,
as the square and one-half replication of the cube in the cases above. The
simplest second-order design that is rotatable is the pentagon with a point
at the center, as given above.
340 Miscellaneous topics

In three dimensions a second-order surface is given by

Y = BQXQ + + B2X2 + ^3^3 + ^11^1 + B22X2

+ B33XI + B^2^1^2 + -^13^1^3 + B23X2X2 + S (16.13)

which has ten unknown coefficients. The cube for a three-dimensional model
has only eight points, so a special design has been developed, called a central
composite design. It is a 2^ factorial with points along each axis at a distance
from the center equal to the distance to each vertex. This gives 8 + 6 = 14
points plus a point at the center makes 15, which is adequate for estimating
the B’s in Equation (16.13). This design is pictured in Figure 16.7.

In attempting to determine the equation of the response surface, some

concepts in solid analytical geometry are often helpful. If a second-order
model in two dimensions is

Y = BQXQ B^X^ + B2X2 + -^11^1 T B^2^I^2 T -^22^2 T ^

the shape of this surface can be determined by reducing this equation to what
is called canonical form. This would be

V D' V' 2 I D' V' 2

i — + £>22^2

This is accomplished by a translation of axes to remove terms in and X2,

then a rotation of axes to remove the X^X2 term. From B\ ^ and B22 we can
determine the shape of the surface, whether it is a sphere, ellipsoid, paraboloid,
hyperboloid, or other. In higher dimensions this becomes more complicated,
but it can still be useful in describing the response surface.
 16.4 Evolutionary operation (EVOP) 341

16.4 EVOLUTIONARY OPERATION (EVOP)

Philosophy

Evolutionary operation is a method of process operation that has a built-in

procedure to increase productivity. The technique was developed by Box [6].
This book and Barnett’s article [2] should be read to understand how the
method works. Many chemical companies have reported considerable
success using EVOP.
The procedure consists of running a simple experiment, usually a
factorial, within the range of operability of a process as it is currently running.
It is assumed that the variables to be controlled are measurable and can be
set within a short distance of the current settings without disturbing produc-
tion quality. The idea is to gather data on a response variable, usually yield,
at the various points of an experimental design. When one set of data has been
taken at all the points, one cycle is said to have been completed. One cycle is
usually not sufficient to detect any shift in the response, so a second cycle is
taken. This continues until the effect of one or more control variables, their
interactions, or a change in the mean shows up as significant when compared
with a measure of experimental error. This estimate of error is obtained from
the cycle data, thus making the experiment self-contained. After a significant
increase in yield has been detected, one phase is said to have been completed,
and at this point a decision is usually made to change the basic operating
conditions in a direction that should improve the yield. Several cycles may
be necessary before a shift can be detected. The objective here, as with
response surfaces, is to move in the direction of an optimum response. Re-
sponse surface experimentation is primarily a laboratory or research tech-
nique; evolutionary operation is a production-line method.
To facilitate the EVOP procedure, a simple form has been developed
to be used on the production line for each cycle of a 2^ factorial with a
point at the center. In the sections that follow an example will be run using
these forms, and later the details of the form will be developed.

Example 16.3 To illustrate the EVOP procedure, consider a chemical

process in which temperature and pressure are varied over short ranges, and
the resulting chemical yield is recorded. Since two controlled variables
affect the yield, a 2^ factorial should indicate the effect of each factor as well
as a possible interaction between them. By taking a point at the center of a
2^ factorial, we can also check on a change in the mean (CIM) by comparing
this point at the center with the four points around the vertices of the square.
If the process should be straddling a maximum, the center point should
eventually (after several cycles) be significantly above the peripheral points
at the vertices. The standard form for EVOP locates the five points in this
design as indicated in Figure 16.8.
342 Miscellaneous topics

^2 • 1

Figure 16.8

By comparing the responses (or average responses) at points 3 and 4 with

those at 2 and 5, an effect of variable may be detected. Likewise, by
comparing responses at 3 and 5 with those at 2 and 4, we may assess the X2
effect. Comparing the responses at 2 and 3 with those at 4 and 5 will indicate
an interaction effect, and comparing the responses at 2, 3, 4, and 5 with those
at 1 will indicate a change in the mean, if present. The forms shown in Table
16.9 are filled in with data to show their use, which is self-explanatory.
Not much can be learned from the first cycle unless some separate
estimate of standard deviation is available. The second cycle (Table 16.10)
will begin to show the method for testing the effects.
Since, in the second cycle, none of the effects are numerically larger than
their error limits, the true effect could easily be zero. In this case another
cycle must be run. The only items that might need explaining are under the
calculation of standard deviation. The range referred to here is the range of
the differences (iv), and 4 „ is found in a table where k = 5 for these five-point
designs, and n is the cycle number. Part of such a table shows

n = 2 3 4 5 6 7 8
0.30 0.35 0.37 0.38 0.39 0.40 0.40

The third cycle is shown in Table 16.11. Here the temperature effect is seen
to be significant, with an increase in temperature giving a higher yield. Once
a significant effect has been found, the first phase of the EVOP procedure has
been completed. The average results at this point are usually displayed on an
EVOP bulletin board as in Figure 16.9.
An EVOP committee usually reviews these data and decides whether
or not to reset the operating conditions. If they change the operating condi-
tions (point 1), then EVOP is reinstated around this new point and the
 c

aj
>
0)
Q oo s:
r-
3 V.I
C/5 O
a u,
VH
(N
c (U w (N
cS 3 4-1 3
o c3
00 X C/l
(D
II <1> O
OO 3 tJ) c/5
-*-*
s
bO _o
(N VH O
3
£ w 3 II D +-> a>
~—
o .2 3
<Oi to > dj
ON 44
3 oJ (U to
rt- ^ 3
• • 3 IT)
3 3
OD II II a ;-l
a>
>
jj
3 a>
3
CT3
■H?
CO
. •
fli .. JO
O O 00
bJD
a U
(U
3 3 o
.2 ^ 00
4-<
31 •
> > 3 u u 4H
o' 3
44 X.
a ^ ! 3 oo 00 O O o
^ ^ o^ZZ :z UH PH UH

U-»
*2
»o '3' '3‘ '3‘
o\ ON ON

in O
‘2 *2
•3' ■3- ■3‘ oo
a\ ON ON

o
o
m »n »n l>^
NO NO NO o
m ON ON ON o
2
3 o
o l>r
(/5
4—>
(U
c/5
<D 1 *2
>n »n CO
l>^
+
bO <N 1 rj- ■3' £
c3
4H
TT
03N ON ON
w l>^
a>
2 a
O c/5 > i>? +
l>^
EVOP Work Sheet - First Cycle

>> a> < l>>i V ^

U l>4
q q q .2
•3
+
3 Tj- Tj- rf 2 +
_o ON ON ON
b
4-* JO + i-i|(S l>^
2 3 i>^
3 u
JO
3 w + O
U <u
0/5 bO
'b & CO
3 oj c/5 1—i|r^
CO
tS
C/5
O S C/5 D CO 3 00
3 ^ _o
3
'O o cr
s ^
*
l>4 <u
4->
CO
cb
3 c/5 a 00
2 2 c3 '—' O

U
O 3 3 0-
Sn
<u
bJ)
cd
o
44
3
&
4H
00
4-»
a
O o 00 VH 4-< <D 3
bi) c/5 c/5 3 S
c/5 <1> cb O
_3 3 3 X) > 4H VH 00
to
O 0 a s c« lU
-4—>
1 _o 3
cb 3- c/5 3
c3
Table 16.9

’> > C/5 X

<u OO 00 00 32 00 <u a
D
O x:
Cu 1 /F'
44
H
4H
U
; PH PH :z P ^
O
,4-V
» ^ _> ^
•3

343
 c <N
O
_o
’-I-*
00
’>
C/)
-4—^ G 00
<D >r) G
o 'g
00 s:
3
X) G Ui ■>
c/2
03 O
S-i
X> XN
C ttJ (N
03 G G
■*-> C.—
in lo O G
X o <X)
II <1>
(U ^ <D c/2
<N S' bX) ,, bX) -4-4
c II G o
_o B w >^2 G
G .
o
bX)
_G !-(
<u
-<—> G > <^ G 3 > O
c/2 03 JJ (U bX)
UH G
2 (/2 V2 a> 3 G
3 G G II I > G
jj O O <U Oo c/2 G U <3 a> XG
y (1) G o
.W O) (1) 3 '> ’> c ^
G
O
^ -C ^ U g g G W O
S-H
o O
S-i
<U O
O. OH Q £ £ :z: :z; :z UH PX

lo »o O «r) *^2 <N

^ O oo ■G*
Gs Os I oo Os

»r2 •o »o o O sr2
G" Tl- ^ c4 ’-o v-2’ (N
OS OS 0\ I 0\ Os
O
o’
<^2
m
»0 U-2 O *^2 »r2 r-- l>^
m so so’ *0 ■^’ T-4 ITS «r2 G
Os Os Os Os Os
2 «o
73
I
i>r
rs| c/2 o’
D
bX) c/2
-4—>
o3 o 1/2
<L) UH
>^2 irs O *0 <^2 +
■G
c/2
c (U <N Tt rt »r2 o’ 0\ & i>^
>
EVOP Work Sheet—Second Cycle

o Os OS Os I oo OS
W l>^
a C4_l l>^
c/2
<u O o 13"
+
i>^
U P< c/2 G
C o o o o o +
rf so (N o’ »o
-4.J
l>^ +
-l-» Os Os Os I OS Os 2 l>^
2 3 w
3
3
+
l>r
3 U
U
s^|<S
+ O
c/2 bX) l>^ o
G o3 c/2 <u
O g W
C/2 ?£
C/2 <U O G <x>
S ^ o G _o G
TD p & -4—>
G
C 2 2 o3 '—' c/2 (U o (U
o <u O G
3 3 bX)
o S-I
U >2
G
S-I
?S <D
o o G <u -i-<
bX) ^ H c i-i -4-4
G
00 c/2 <U G
_c
G G > S-i <u O
G s-i
O g (U
Table 16.10

-4—> O O c/2 G G bX)

o3
Ui
. ^ ^ CX c/2 G
> > C/2
<u (U (U ^ £
<L)
^ ^ (U <U (l> X G
cx 1-1 $-1 o u -G
O 0. CU Z QZ :z: H CU u

:G > >,

344
 o <N
c o O
•
o
4-*
d C/5
> 4- >
♦
NO
to
Q
(U
ON
•^
a NO
oo
T3 «o 03
o C/5 5- 1
ccj o
T3 to c 5H
5-(
C O (N
03 ^ c w G
4—> t4H <N
00
M (1>
X 8 O
o3
O

<N
o O
60 ,,
!0
O
60
o3
c/3
4-»
a
c S S c-, C II o
• PH 5-1 O
Tj- (D
Ol O' 4-> G > rA oS
5H
^ 60
o3
> o
c/3 3 o3 60
5H
tr> ty;}" <L) G
4-> 3 G G 13 > 3 o3
o O •O <u c/3 o3 D O 40
a> b’
c/3 3 •^ PH 60 U G G O
b > >
’o’ 03
4.4
U (U O G 5H 5H 5H
5-1 -G o3 ^4 5.4 o3j o a> o o o o
0^ OH OH CLH 0:H PLH UH MH

»o fNj q fN (N
od '=1' d r4
oo ON ON oo ON
HH (N

q m q »n •T)
o
•o’ r--' I_H' NO’ to i-H
ON ON ON 1 oo ON
1 CN »o d
d
I
•o r-- O cn <N r- * IrH
1-H ud NO d r-’ tri >o
ON ON ON 1 oo ON CO
1—H 1 CN
2
13 •o i>^
c/3
<u o l>r
c/3
W) 4->
<u o3 >o •o (N O CC3 o +
VH
<U ON rn cn 1^4
C > Sp.^ oo ON ON OO (ON
<i> o l-H (N w
EVOP Work Sheet—Third Cycle

<4-(
O c/3 o +
>» O
i>r
U 0^ c/3 G
c o o »r) tn q O
•^ + l>^
o d ■^’
•^
4-»
d !0
ON ON ON oo ON JB l>^
+
■r^ CN 3 1^
in jj
3 + 1^
3 i-l|lO
3 U 1^
U
Hc^
+ O
c/3 60 1^ O
c o3 C/3 (U
o 5-1 C/3 &
<L> (/3 o G <u
S 03> C O
G
T3 o JS o3
c 0 <U ‘40 c/3
o <t)

U
o 3 3
(/3
0)
60
o3
<U
JS
o3
V4
(D
4->
a
^ ^ fl> C/5
(U
60 o a 5H
<u 03 G
G
C/3
01 P
C/3
c > 5-1 <u •1.4
o
•^ 03 o 5H
O O O 5-1 c/3 o3 G 60
Table 16.11

o3 .W GH c/3 G
VH
a>
a
o o ^ 3 g o a<u C/3
(U
5H
X
40
O £ Ok Q ^ H OH u

:=3 > >.

345
346 Miscellaneous topics

94.2 f ♦ 95.7

Pressure 94.8

94.3 4 495.5

Temperature
Figure 16.9

second phase is begun. EVOP is continued again until significant changes are
detected. In fact, it goes on continually, seeking to optimize a process.
This is a very simple example with just two independent variables. The
references should be consulted for more complex situations.

EVOP Form Rationale

Most of the steps in the EVOP form above are quite clear. It may be helpful
to see where some of the constants come from.
In estimating the standard deviation from the range of differences in
step (iv), consider a general expression for these differences

Dp - X p,n
n — 1

where Dp represents the difference at any point p of the design. represents

the observation at point p in the ith cycle, n is, of course, the number of
cycles. Since the variance of a sum equals the sum of the variances for inde-
pendent variables, and since the variance of a constant times a random
variable equals the square of the constant times the variance of the random
variable.

Since all these x’s represent the same population, their variances are all alike,
and then
n
n — I
and
 16.4 Evolutionary operation (EVOP) 347

The standard deviation of the population can then be written in terms of the
standard deviation of these differences

(jj) can now be estimated from the range of these differences From the
quality control field

where J2 depends on the number of differences in the range, which is 5 on

this form. Here ^2 = 2.326 for samples of 5,
R,
(7 D
2.326
and the standard deviation of the population is estimated by
’n — 1 R,
=
n 2.326
The quantity
'n — 1 1
n 2.326
is called „ in the EVOP form where k = 5.
Note that

A- - J S

which tallies with the table values given.

To determine the error limits for the effects, two standard deviation
limits are used, as they represent the approximate 95 percent confidence
limits on the parameter being estimated.
For any effect such as
E = i(F3 + n - 72 - F5)

its variance would be

348 Miscellaneous topics

and two standard deviation limits on an effect would be

estimated by

For the change in mean effect

CIM = |(F, + F3 + F4 + - 4F,)

,
FC/M =

20 20 a}
— ^ (^r) ~ ^
25 25 n
[4 (7Y

and two standard deviation limits would be

s
±2 • ±1.78

as given on the EVOP form.

Example 16.4 This EVOP procedure often has three factors, each at two
levels. The 2^ = 8 experimental points are run as well as two points in the
center of the cube, giving ten experimental points per cycle. Usually this
design is run in two blocks where the ABC interaction is confounded with
blocks. The standard form for the location of the ten points is shown in
Figure 16.10.

Figure 16.10
 a\
3 r4
o o^
a
> (/5 1>5
d)
t/D II11
Q gj 3 3 CO
'3
3 0 3
u 3 1
a 3

2n
*3 X
3 *3 II 3^ 3
3
bo o^ 3
3
i->
C/3 3 (S
3 3 •
bO Vi
3
VO 3 B 5^
Vi CO 3
OO O • bO
S
05 to 3
Vi
o C/) c/5 g 3
o o 3 3 3 3 S >
3
o
lo \0
bi) O O
> •3 ^
VO ^ •^
<u c/D o C 3
3 ^ c: ^
<N .a
O' 52 a (N N U 3 3 3 3

CU OH Q
X) c
c o3 X
'3•
Vi
PH z z
03
o o
o O
»o "O
3 CO CO "Tf ON r^, 00

r4
ra 3 Tf Tt co’ o' r^* CO
a> bO
(U u* 3
3 •^
-(.j O
o3 3
o3 Vi TJ
o (U (DCX d) 00 00 O <N 00
a Ov Ov •0 3- d
O s ^ <N
g d) I
*d)

Change in mean effect = ^(^2 + F3 + Ft + F5 — 4yi) = 0.64

Vi
a> a> d)
o 3 ^ o:
n O cx
ti—I
u> N c/5 CN
3 1/5 ? cx ^ o ^ VO r- ^ CO
O) Vi
d) p o s CO i2 ^ VO 10 00 3‘
C as P CQ
I
o3 c/5
O
d> d)
'3 ^ Cq U o o p o p o 3
3 s
1/5
c/5 Vi <N 0 O 3^ 3^ d
3 s
B - AC = Ml's + Ys - ij - W = -3.20

Vi bO
O VH O c/5
EVOP Work Sheet—Block 1 in 2

a O -»-*
C - .4B = i(F4 + ^5 - y2_ - I'aJ = 5^00
A - BC = iC^s + Yi - Y2 - Xs) = 5.40

d)
1/5 o bO c/5
-t-»
d> 3 3 3
U P< Ui Vi
3 p
> ON p p r^. •'N &
< co’ CO in ON '3' w

3 3
O
I—>
-X
u 3 o 3
jj c/5
jj

3 3
s g g c/5
C/5
3 U
U c/5 33 3
3 _oc/5 C/5 l>^
O c/5
3 3^c3 3
bO
TO 0 3^
3
3 g
3
3
22 v^
O c/5
3 3 P 3
c/5 3
>
u O 3O o VH
3
Table 16.12

bO
3 ■> 3 ^
3
CQ<- ♦
g g I 3
3
Vi ,
PH P-i Z Q Z
3
CX ■•- > >.
O

349
 VO
m m
C •r5
o
•
on VO
r- CO +1
*rs *
S3 o c/5
(N
‘> -/-< OO OO <N
O
Q
(U II
(N
s
•^
+1
+1
so
3 3
ro ^ O
T3 • o P
CA
I I I
<N '-^3 c/5 Ul
03 o s; c,
T3 o^ s; u <N
C
a
fN ^ 3 (U CN
W
u

-*-» <D P P
bfl CO o3 cp o O
c/5 C o c/5 ID
«4-l
o
II
S'
i-, fO c
_o
It)
bO
o3
c/5
-*-*
s ^ ^ «P

n I I I
_o
-!->
SP >^ IT)
<u
ba
-t—> VH
(U
>
O

O 0 op ^
!/D 03 03 3 o3 (U b£)
3
c/5
D
1/5
P
II s3 sP VH
(U
> 3
p
03 1 I I
o3 O I)
O O <1^ on oo U P
P:
3
U
-r
w <i)
tj)
jrt O (U <U
UH
p
VH
o
;-i ^GG
VH tH ^ o o o ^ ^ op
PH 0^ UH 1^ UH
+ + +
c/5
U op ^
I ^ ^
O• O• 0^• ON« ^• ^ i
O O ^ -H ^
® ; II II II
-P
•W
o op U O
PP ^ ^ oq
p ON VO VO
rj- c/5
os O O »n Tt *0 (N ■/—>
<N o

&
W o

p OO p 00 r-
r4 c4 so cn ON
p
CO _o *
CN
I OO <N

(N o * r4
CJ
3
jj
(N
OO
m

O O
»o »o 3 u
Tj- ON 00 »r5
m *^5
+ 3 *rs I
s *o (N U n +
Tj- Tf- rn o* r^* (N
m’ s
m SO I
(N (N C/D i>^
<U
bO
oS
c/5
-/->
O
+ G GGS
op ^ u
<N VH
(U o o
> <N• <N• <N• o• W
l>^ I , ,
o • l>r i>:r
0) < so rj- m T-H on +
CQ
o ^ 3 2^ o
>> o O
p l>r i>^
^ C o i>r + + +
(U _o •

ID
dj s-* <N
c/:) 3
3 o
i3
3 o GGG P
o3
o
<u
bO I>^
op X X (U
o
o3 3 + + + B
O 3
U
s
B w C/5
c/5
C/5
O
U + + + £
o
oo C ^ 00 330
C s § p
cu o di ^
.2 —
c/5
ig 1^ 103 i|<s
d)
bO
O Cj ' 1 (U
i-i|fS r-ilrs (U
3 3
> T3 o o
bO
03 II II II _§
w C S <u 52

U
O c/5
P
c/5
P
/5 o
s }-i
O
> o op o
^ oq O ^
O O ■§ § s o3 op bO
bO P
'sd C
♦^
>
a>
>
(D S !C + + + o3
o
^ ^ (l> O
VH I-H
-2i
o3
i-i
<L>
PH CU ZQZ op (J
4P
u
3 a > >. -C
H

350
 16.5 Analysis of attribute data 351

The first five points (1 through 5) are run as the first block and the last
five points (6 through 10) are run as the second block. Tables 16.12 and 16.13
illustrate this EVOP technique for the second cycle of an example on the
mean surface factor for grinder slivers where the factors are drop-out
temperature A, back-zone temperature B, and atmosphere C. The work
sheets should be self-explanatory as they are based on the design principles
of confounding a 2^ factorial in two blocks of four observations. One could
terminate this process with the second cycle as it shows drop-out temperature
A, back-zone temperature B, and their AB interaction to be significant in
producing changes in the mean surface factor for grinder slivers.

16.5 ANALYSIS OF ATTRIBUTE DATA

Philosophy
Two assumptions used in the application of the analysis of variance technique
are that (1) the response variable is normally distributed and (2) the variances
of the experimental errors are equal throughout the experiment. In practice
it is often necessary to deal with attribute data where the response variable
is either 0 or 1. In such cases one often records the number of occurrences
of a particular phenomenon or the percentage of such occurrences.
It is well known that the number of occurrences per unit such as defects
per piece, errors per page, customers per unit time, often follow a Poisson
distribution where such response variables are not only nonnormal but
variances and means are equal. When proportions or percentages are used
as the response variable, the data are binomial and again variances are
related to means and the basic assumptions of ANOVA do not hold. Some
studies have shown that lack of normality is not too serious in applying
the ANOVA technique, but most statisticians recommend that a trans-
formation be made on the original data if it is known to be nonnormal.
Recommendations for proper transformations are given in Davies [9].
Another technique, called factorial chi square, has been found very
useful in treating attribute data in industrial problems. This technique is
described by Batson [3], who gives several examples of its application to
actual problems. Although this factorial chi-square technique may not be
as precise as a regular ANOVA on transformed data, its simplicity makes
it well worth consideration for many applied problems.
To illustrate the techniques mentioned above, consider the following
example from a consulting-firm study.

Example 16.5 Interest centered on undesirable marks on steel samples

from a grinding operation. The response variable was simply whether or not
these marks occurred. Four factors were believed to affect these marks: blade
size A, centering B, leveling C, and speed D. Each of these four factors was
352 Miscellaneous topics

set at two levels giving a 2"^ factorial experiment. Each of the 16 treatment
combinations was run in a completely randomized order, and 20 steel
samples were produced under each of the 16 experimental conditions. The
number of damaged samples in each group of 20 is the recorded variable
in Table 16.14.

Table 16.14 Damaged Steel Samples in 2‘^

Experiment

Factor A
0 1

Factor B Factor B
Factor C Factor D 0 1 0 1

0 0 0 0 16 20
1 0 0 10 20

1 0 0 0 10 14
1 1 0 12 20

Table 16.15 Yates’ Method on Transformed Data of Steel Sample Experiment

Transformed
Proportion Variable
Treatment P X = arc sin Vp (1) (2) (3) (4) SS

(1) 0.00 0.00 1.11 2.68 4.46 9.51 —

a 0.80 1.11 1.57 1.78 5.05 9.05 5.12

b 0.00 0.00 0.79 2.36 4.46 1.89 0.22
ab 1.00 1.57 0.99 2.69 4.59 2.35 0.35
c 0.00 0.00 0.79 2.68 0.66 -1.23 0.09
ac 0.50 0.79 1.57 1.78 1.23 -1.03 0.06
be 0.00 0.00 1.12 2.36 0.66 -0.59 0.03
abc 0.70 0.99 1.57 2.23 1.69 -0.13 0.01
d 0.00 0.00 1.11 0.46 -0.90 0.59 0.03
ad 0.50 0.79 1.57 0.20 -0.33 0.13 0.01
bd 0.00 0.00 0.79 0.78 -0.90 0.57 0.02
abd 1.00 1.57 0.99 0.45 -0.13 1.03 0.06
cd 0.05 0.23 0.79 0.46 -0.26 0.57 0.02
acd 0.60 0.89 1.57 0.20 -0.33 0.77 0.04
bed 0.00 0.00 0.66 0.78 -0.26 -0.07 0.00
abed 1.00 1.57 1.57 0.91 0.13 0.39 0.01
 16.5 Analysis of attribute data 353

If the above responses are each divided by 20, the proportions of

damaged samples can be determined. Since proportions follow a binomial
distribution, an arc sine transformation is appropriate (see [9]). Taking the
arc sin as the response variable, Yates’ method can be used to determine
the sums of squares as shown in Table 16.15.
If the three- and four-way interactions are assumed to be nonexistent,
the resulting ANOVA table is shown as Table 16.16.

Table 16.16 ANOVA for Steel Sample

Experiment

Source df SS MS

A 1 5.12 5.12***
B 1 0.22 0.22*
AB 1 0.35 0.35**
C 1 0.09 0.09
AC 1 0.06 0.06
BC 1 0.03 0.03
D 1 0.03 0.03
AD 1 0.01 0.01
BD 1 0.02 0.02
CD 1 0.02 0.02
Error, pooled 5 0.12 0.02

This analysis shows the highly significant blade size effect, which is
obvious from a glance at Table 16.14. It also shows a significant centering
effect and a blade size-centering interaction that is not as obvious from
Table 16.14.
To apply the factorial chi-square technique to this same example.
Table 16.17 is filled in. The plus and minus signs simply give the proper
weights for the main effects and two-way interactions and T is the value of
the corresponding contrast. D is the total number of samples in the total
experiment; here 20 per cell or 20 x 2"^ = 320 samples. The chi-square (x^)
value is determined by the formula

..2 _
(16.14)
~ (S)(F) ^ ~D

Here N is the total number of items in the whole experiment and S is the
number of occurrences (successes) in the N. F is the number of nonoccur-
rences (or failures). In the example N — 320 samples, 5 = 123 with un-
 * * *
(N VO fNj U-) VO Tj- (N VO VO
<S m q VO q
0 T-H OS q
fO 00 d OS d d r4
ON

•n »o ro 00 wo cn •r> 0
9. Ov (N 0 (N <N 0 0 <N r--
<s •ri d d (N d d d d d
Tj”

0 0 0 0 0 0 0 0 0 0
Q <N <N (N (N (N CN tN (N rs| <N
m m m cn ro ro m m CO

Cv Os os WO
Tj- <N 00 r4 00 00 CN
<S VO VO (N

T—H

os m r- ov cn Os WO
<N fS <N
K 1 1 1
+

Os VO 0 00 m r-
1 11 Tl- VO VO rl- VO VO VO w^ WO
1 1 1 1 1 1 1 1 1

tN m •0 VO r4 0 VO OS
+ r- VO r- W-) VO VO VO VO
w
20

+ + + + + + + + + +

0 + + + 1 + + 1 + 1 1
20

+ + 1 + + 1 + 1 + 1
0
20

0 + + 1 1 + 1 1 1 1 +

r4
+ 1 + + 1 + + 1 1 +

0 0
+ 1 + 1 1 + 1 1 + 1
0
'•>6
1—1 0
+ 1 1 + 1 1 + + 1 1
jj
0
3
cn VO
0 + 1 1 1 1 1 1 + + +
H
Ui
0 1 + + + 1 1 1 + + +
+->
(U ’^H
QJ
0 0 1 + + 1 1 1 + + 1 1
GO

Ui T—j 0 1 4- 1 + 1 + 1 1 + 1
O
0
(U 0 0 1 + 1 1 1 + + 1 1 +
d 0
:3
(T T-H
1 1 + + + 1 1 1 1 +
CO
1

0 0 1 1 + 1 + 1 + 1 + 1
u
0
TH 0 1 1 1 + + + 1 + 1 1
0
■t-> 0
0
c3 0
0 1 1 1 1 + + + + + +
UH
(/) 00
-(->
r- 0 °Q C
1-H U C)
,<1^
'O Q to 0
0 -<-> TO ■*—>
ui U 0) T3 13 <0 0
c > d) cd
(U
*T^ -(->
TJ
<D c 3 <u O- ui
d)
U q 0 q q
3 cd C > <U
(U <u D,
<N 3 U co -4-^ q q
05 c
H S U CO t-H

354
 16.6 Randomized incomplete blocks—restriction on experimentation 355

desirable marks. For factor A, then

, (320)^

= (4.23)(45.75) = 193.52

The 4.23 is constant for these data. The resulting chi-square values are
then compared with with 1 df at the significance level desired. For the
5 percent level, X[i] = 3.84 and again factors A, B, and AB are significant.
As seen in Table 16.17, the calculations for this technique are quite
simple and the results are consistent with those using a more precise method.
Extensions to problems other than 2^ factorials can be found in Batson [3].

16.6 RANDOMIZED INCOMPLETE BLOCKS—

RESTRICTION ON EXPERIMENTATION

Method for Balanced Blocks

In some randomized block designs it may not be possible to apply all treat-
ments in every block. If there were, for example, six brands of tires to test,
only four could be tried on a given car (not using the spare), and such a
block would be incomplete, having only four out of the six treatments in it.

Example 16.6 Take the problem of determining the effect on current flow
of four treatments applied to the coils of TV tube filaments. As each treat-
ment application requires some time, it is not possible to run several obser-
vations of these treatments in one day. If days are taken as blocks, all four
treatments must be run in random order on each of several days in order
to have a randomized block design. After checking it is found that even four
treatments cannot be completed in a day; three are the most that can be run.
The question then is: Which treatments are to be run on the first day, which
on the second, and so forth, if information is desired on all four treatments?
The solution to this problem is to use a balanced incomplete block
design. An incomplete block design is simply one in which there are more
treatments than can be put in a single block. A balanced incomplete block
design is an incomplete block design in which every pair of treatments occurs
the same number of times in the experiment. Tables of such designs may be
found in Fisher and Yates [11]. The number of blocks necessary for bal-
ancing will depend upon the number of treatments that can be run in a
single block.
For the example mentioned there are four treatments and only three
treatments can be run in a block. The balanced design for this problem
requires four blocks (days) as shown in Table 16.18.
356 Miscellaneous topics

Table 16.18 Balanced Incomplete Block Design

for TV Filament Example

Treatment
Block
(days) A B C D T,

1 2 — 20 1 29
2 — 32 14 3 49
3 4 13 31 — 48
4 0 23 — 11 34

T, 6 68 65 21 160 = T

In this design only treatments A, C, and D are run on the first day;
B, C, and D on the second day, and so forth. Note that each pair of treat-
ments, such as AB, occurs together twice in the experiment. A and B occur
together on days 3 and 4; C and D occur together on days 1 and 2; and so
on. As in randomized complete block designs, the order in which the three
treatments are run on a given day is completely randomized.
The analysis of such a design is easier if some new notation is intro-
duced. Let

b = number of blocks in the experiment {b = 4)

t = number of treatments in the experiment (f = 4)
k = number of treatments per block {k = 3)
r = number of replications of a given treatment throughout
the experiment (r = 3)
N = total number of observations
= bk = tr{N = 12)
X = number of times each pair of treatments appears together
throughout the experiment
= r{k - l)/{t - 1) (2 = 2)
Table 16.18 has the current readings after they have been coded by
subtracting 513 milliamperes and the block and treatment totals. The analysis
for a balanced incomplete block design proceeds as follows:

1. Calculate the total sum of squares as usual:

SS total
I j
1^
(160)^
= 3478 - 1344.67
12
 16.6 Randomized incomplete blocks—restriction on experimentation 357

2. Calculate the block sum of squares, ignoring treatments:

b nr 2
T^.
SS block
i= 1 Iv
(29)^ + (49)2 + (48)2 + (34)2
= 100.67
3 12
3. Calculate treatment effects, adjusting for blocks:

y 6)2
cc 1 ^j
(4.8)
^^treatment /c2t
where
Qj = - Z (4.9)
I

where riij = 1 if treatment j appears in block /, and riij = 0 if treatment

J does not appear in block i. Note that Yji is merely the sum of all
block totals which contain treatment j.
For the data given.

Q, = 3(6) - (29 + 48 + 34) = 18 - 111 = -93

Qi = 3(68) - 131 = 73
Q, = 3(65) - 126 = 69
<24 = 3(21) - 112 = -49
0
Note that

Z e. = 0
j=i

which is always true. Then

(-93)2 + (73)2 + (69)2 ^_49^2

SS treatment 880.83
3(2)4

4. Calculate the error sum of squares by subtraction

SS error SS total SS block SS treatment

= 1344.67 - 100.67 - 880.83 = 363.17

Table 16.19 summarizes in an ANOVA table.

An F test gives F3 5 = 293.61/72.63 = 4.04, which is not significant at
the 5 percent level (Appendix, Table D).
In Table 16.19 the error degrees of freedom are determined by subtrac-
tion rather than as the product of the block and treatment degrees of freedom.
However, this error degrees of freedom is seen to be the product of treatment
358 Miscellaneous topics

Table 16.19 ANOVA for Incomplete Block Design

Example

Source df SS MS

Blocks (days) 3 100.67 —

Treatments (adjusted) 3 880.83 293.61

Error 5 363.17 72.63

Totals 11 1344.67

and block degrees of freedom (9) if 4 is subtracted for the four missing values
in the design.
In some incomplete block designs it may be desirable to test for a block
effect. The mean square for blocks was not computed in Table 16.19 since it
had not been adjusted for treatments. In the case of a symmetrical balanced
incomplete randomized block design, where b = t, the block sum of squares
may be adjusted in the same manner as the treatment sums of squares

e; = rT, - I n,jTj
j
Q\ = 3(29) - 92 = -5
Q'2 = 3(49) - 154 = -7
2'3 = 3(48) - 139 = +5

e; - 3(34) - 95 - +7

I e; - 0
(-5)^ + (-7)^ + (5)^ + (7)-
ss block I m- h'/ib - 6.17
i= 1 3(2)4

and

6^ + 68^ + 65^ + 21^ (160)^

SS treatment (unadjusted) 975.34
3 12

The results of this adjustment and the one in treatments may now be sum-
marized for this symmetrical case as shown in Table 16.20
The terms in parentheses are inserted only to show how the error term
was computed for one adjusted effect and one unadjusted effect

SSerror = SS,,,,. - SS.treatment (adjusted) SSv

‘^‘^block

= 1344.67 - 880.83 - 100.67 = 363.17

or
SS error SS
‘^‘^total
— ‘^‘^treatment
ss — ‘^'“^block
ss (adjusted)

1344.67 - 975.34 - 6.17 - 363.17

 16.6 Randomized incomplete blocks—restriction on experimentation 359

Table 16.20 ANOVA for Incomplete Block Design Example

for Both Treatments and Blocks

Source df SS MS

Blocks (adjusted) 3 6.17 2.06

Blocks (3) (100.67) —

Treatments (adjusted) 3 880.83 293.61

Treatments (3) (975.34) —

Error 5 363.17 72.63

Totals 11 1344.67

It should be noted also that the final sum of squares values used in
Table 16.20 to get the mean square values do not add up to the total sum of
squares. This is characteristic of a nonorthogonal design. The F test for
blocks was not run because its value is obviously extremely small, which
indicates no day-to-day effect on current flow.
For nonsymmetrical or unbalanced designs, the general regression
method may be a useful alternative. If contrasts are to be computed for an
incomplete block design, it can be shown that the sum of squares for a
contrast is given by
(4.10)

where contrasts are made on the Q/s rather than T/s.

As an example consider the following orthogonal contrasts on the data
of Table 16.18:

C, = Q,-Q2 = -166
C2 = QI + Q2- 203 = -158
6^3 = Qi + 62 + 03 — 5^4 = 196
The corresponding sums of squares are
(-166)2
574.08
(2)(3)(2)(4)
(-158)2
173.36
(6)(3)(2)(4)
(196)2
133.39
(12)(24)
880.83

which checks with the adjusted sums of squares for treatments. Comparing
each of the above sums of squares with their 1 df against the error mean
360 Miscellaneous topics

square in Table 16.19 indicates that contrast is the only one of these three
that is significant at the 5 percent level of significance. This indicates a real
difference in current flow between treatments A and B, even though treat-
ments in general showed no significant difference in current flow.

16.7 YOUDEN SQUARES

When the conditions for a Latin square are met except for the fact that only
three treatments are possible (for example, because in one block only three
positions are available) and where there are four blocks altogether, the de-
sign is an incomplete Latin square. This design is called a Youden square.
One such Youden square is illustrated in Table 16.21.

Table 16.21 Youden Square Design

Position

Block 1 2 3

I A B C
II D A B
III B C D
IV C D A

Note that the addition of a column (D, C, A, B) would make this a Latin
square if another position were available. A situation calling for a Youden
square might occur if four materials were to be tested on four machines but
there were only three heads on each machine whose orientation might affect
the results.
Example 16.7 The analysis of a Youden square proceeds like the incomplete
block analysis. Assuming hypothetical values for some measured variable
where
^ijk = A Y Pi -\- Tj -\- A- Sijk
with
i=l,...,4 j = 1, 2,..., 4 k = 1, 2, 3
we might have the data of Table 16.22.
In Table 16.22
t = b ^ 4
r = k —
2 = 2
Treatment totals of A, B, C, D are
Tj/.e, 2, -1, -9
 16.7 Youden squares 361

Table 16.22 Youden Square Design Data

Position

Block 1 2 3 T,.

I A = 2 B = 1 C = 0 3
II Z) = -2 4 = 2 B = 2 2
III B = -1 C = -1 D = -3 -5
IV C = 0 D = -4 4 = 2 -2

T.k -1 -2 +1 -2 = T

From this,

SS,„,a, = Z Z Z - “ = 48 - = 47.67

Position effect may first be ignored, since every position occurs once and
only once in each block and once with each treatment, so that positions are
orthogonal to blocks and treatments:

ec _ (3)" + (2)^ + (-5)" + (-2)" (-2)^

‘^'^block (ignoring treatments) -y A ^

42 1 _ 41
13.67
T “ 3 “ T
For treatment sum of squares adjusted for blocks we get

Q, = 3( + 6) - 3 = 15
= 3( 2) - 0 = 6
Q, = 3(-l) - (-4) = 1
= 3(-9) - (-5) = -22
and
Z e,' = 0
i

(15)^ + (6)2 + (1)2 + (-22)2 _

SS treatment 31.08
(4)(6)

(-1)2 + (-2)2 + (1)2 (-2)2

SS position = 1.17
4 12

SS error = SS....,
'total - SS block
— ‘^‘^treatment
SS (adjusted) SS position
47.67 - 13.67 - 31.08 - 1.17
1.75
The analysis appears in Table 16.23.
362 Miscellaneous topics

Table 16.23 Youden Square ANOVA

Source df SS MS

Treatments TJ (adjusted) 3 31.08 10.36

Blocks Pi 3 13.67 —

Positions 2 1.17 0.58

Error &ij,^ 3 1.75 0.58

Totals 11 47.67

The position eflfect here is not significant and it might be desirable to

pool it with the error, getting

1.17 + 1.75 _ 2.92

s e2 0.58
2 + 3 “ ^

as an estimate of error variance, with 5 df. Then the treatment effect is highly
significant. The block mean square is not given, as blocks must be adjusted
by treatments if block effects are to be assessed. The procedure is the same
as shown for incomplete block designs.

PROBLEMS

16.1 One study reports on the breaking strength Y of samples of seven different starch
films where the thickness of each film X was also recorded as a covariate. Results
rounded off considerably are as follows:

Adjusted
Source df df MS

Between starches 6,000,000 50,000 500

Within starches 100

Totals 8,000,000 60,000 600

1. Complete this table and set up a test for the effect of starch film or breaking
strength both before and after adjustment for thickness.
2. Write the mathematical model for this experiment, briefly describing each
term.
3. State the assumptions implied in the tests of hypotheses made above and
describe briefly how you would test their validity.
 Problems 363

16.2 In studying the effect of two different teaching methods the gain score of each
pupil in a public school test was used as the response variable. The pretest scores
are also available. Run a covariance analysis on the data below to determine
whether or not the new teaching method did show an improvement in average
achievement as measured by this public school test.

Method I Method II

Pretest Pretest
Pupil Score Gain Pupil Score Gain

1 28 5 1 24 7
2 43 -8 2 40 9
3 36 4 3 41 4
4 35 1 4 33 4
5 31 4 5 45 5
6 34 1 6 41 -2
7 33 4 7 41 6
8 38 8 8 33 11
9 39 1 9 41 -1
10 44 -1 10 30 15
11 36 -7 11 45 4
12 21 2 12 50 2
13 34 -1 13 42 9
14 33 1 14 47 3
15 25 7 15 43 -2
16 43 1 16 34 2
17 34 3 17 44 7
18 40 5 18 43 4
19 36 4 19 28 3
20 30 9 20 46 -1
21 35 -6 21 43 2
22 38 0 22 46 -4
23 31 -14 23 21 4
24 41 4
25 35 10
26 27 10
27 45 1

16.3 Data below are for five individuals who have been subjected to four conditions
or treatments represented by the groups or lots 1-4. 7 represents some measure
of an individual supposedly affected by the variations in the treatments of the
four lots. X represents another measure of an individual that may affect the
value of 7 even in the absence of the four treatments. The problem is to deter-
mine whether or not the 7 means of the four groups differ significantly from
each other after the effects of the X variable have been removed.
364 Miscellaneous topics

Groups or Lots

1 2 3 4

X Y X y X Y X y

29 22 15 30 16 12 5 23
20 22 9 32 31 8 25 25
14 20 1 26 26 13 16 28
21 24 6 25 35 25 10 26
6 12 19 37 12 7 24 23

16.4 Consider a two-way classified design with more than one observation per cell.
1. Show the F test for one of the main effects in a covariance analysis if both
factors are fixed.
2. Show the F test for the interaction in a covariance analysis if one factor is
fixed and the other one is random.
16.5 For Example 16.4 in the text, another response variable was measured—the
mean surface factor for heat slivers. Results of the first three cycles at the ten
experimental points gave

Cycle (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

1 4.0 6.8 5.9 4.5 4.3 4.2 7.8 8.3 11.0 5.9
2 5.2 7.2 6.1 3.7 4.0 4.5 7.0 8.1 12.5 6.3
3 4.8 8.0 7.0 3.6 3.8 4.3 6.5 7.8 12.7 7.0

Set up EVOP work sheets for these data and comment on your results after
three cycles.
16.6 Data on yield in pounds for varying percent concentration at levels 29, 30, and
31 percent, and varying power at levels 400,450, and 500 watts gave the following
results in four cycles at the five points in a 2^ design.

Cycle (1) (2) (3) (4) (5)

1 477 472 411 476 372

2 469 452 430 468 453
3 465 396 375 468 292
4 451 469 363 432 460

Analyze by EVOP methods.

16.7 The response variable in the data on page 365 is the number of defective knife
handles in samples of 72 knives where the factors are machines, lumber grades,
and replications. Analyze after making a suitable transformation of the data.
 Problems 365

Machine
Lumber
Grade Replication 1 2

A I 8 4
II 4 6
B I 9 3
II 10 4

16.8 Analyze the data of Problem 16.7 using factorial chi square. Compare with the
results of Problem 16.7.
16.9 Data on screen color difference on a television tube measured in degrees Kelvin
are to be compared for four operators. On a given day only three operators can
be used in the experiment. A balanced incomplete block design gave results as
follows:

Operator

Day A B C D

Monday 780 820 800 —

Tuesday 950 — 920 940

Wednesday — 880 880 820
Thursday 840 780 — 820

Do a complete analysis of these data and discuss your findings with regard to
differences between operators.
16.10 Run orthogonal contrasts on the operators in Problem 16.9.
16.11 At the Bureau of Standards an experiment was to be run on the wear resistance
of a new experimental shoe leather as compared with the standard leather in use
by the Army. It was decided to equip several men with one shoe of the experi-
mental-type leather and the other shoe of standard leather and after many
weeks in the field compare wear on the two types. Considering each man as a
block, suggest a reasonable number of men to be used in this experiment and
outline its possible analysis.
16.12 If three experimental types of leather were to be tested along with the standard
in Problem 16.11, it is obvious that only two of the four types can be tested on
one man. Set up a balanced incomplete block design for this situation. Insert
some arbitrary numerical values and complete an ANOVA table for your data.
 17
Summary and Special Problems

Throughout this book the three phases of an experiment have been empha-
sized: the experiment, the design, and the analysis. At the end of most chap-
ters an outline of the designs considered up to that point has been presented.
It may prove useful to the reader to have a complete outline as a summary of
the designs presented in Chapters 1-15.
This is not the only way such an outline could be constructed, but it
represents an attempt to see each design as part of the overall picture of
designed experiments. A look at the outline will reveal that the same design
is often used for an entirely different experiment, which simply illustrates
the fact that much care must be taken to spell out the experiment and its
design or method of randomization before an experiment is performed.
For each experiment, the chapter reference in this book is given.
After the summary several special problems are presented. Many do
not fit any one pattern or design but may be of interest as they represent
the kinds of problems actually encountered in practice. These problems
have no “pat” solutions and the way they are set up may depend a great
deal upon assumptions the reader may make. They should serve to evoke
some good discussion about designing an experiment to solve a given prac-
tical problem.

366
 X X
<u
0
m 1)- ■G- X X o^
a> Vi Vi Vi Vi Vi Vi Vi Vi Vi
G
(*V <l-> d) d) d) d) d) d) d) d) d)
;-H ->-> ■1—> -11 -1—1 -1-1 -1-1 1-1 1-1
03 CX GH Cv Pi a PH Pi Pi Pi
JZ C-I-M G G G G cS G G a G
D
u X X X X X X X X X
U u U U U U U U U

TD
O
175
P _i • •

+-<
d)
<NI
X)
0 G
Vi
d)
• -

(U 0 G 1-H <N
G (U -,—1 cq
< Vi
C GO • • rs 0
< > d>
1-1
> _C
Vi • <I

C/D <
< •§ 0
< < 0
x' > 0 +
0 > ^ Z > > X! G
C/D >
0
Z
<
0
^
^
GO
< o o X
1-1
X
0
'3
x' Z
0 o'
(N
cq
X
0

X 03 :z 1-1 0
z d) 0"
a < o3
< < <
II G
Vi
< < o3
:>
g
♦—(
> 03
>1 :>

0
1
cG
0
Vi
^
03
03 'o
03 >
0
r-
C/D
a> 0
^^ D 0
cq +-<
_G

and
Vi A\ ^ X
1 cii X a ^ Z o3 Z 03
CO
(D
-G
d)
CO
Vi
(D H iX 0 H 00 3 < < G 0 G c2
C
. o
0 d X 0 d X lx d X cJ

(N m

CO

+
s CO

d»
Ui
3 +
CO
o3
+ D +
C cr cq
TD
d> co“ TDg c/5

C
d)
CO
N a
u< + d)
'T3
+ +
Q
I + g d)
a
d)
00
+
•“-5 G
O
>H
^
^ +
(T '
KT
o cq
T3
+ 2
G
i +3 +
^
•-» - -v
d) dT (U
C/D
G H“ P
Vi ^
G
d) ^ O
>>'"^03 CCi, ^ ^ d>
'a a
^ (U
S, CIH HJG + + G
o + .a G 3
+
G, 13 Vi
d) 03
9- S o CT
a 4—» -1 d) d)
CO CO

'a. ^ o ' I 8 CJ
CO
3.
a o oo I
(U
Oi II
■X Vi
O
G d>
G
0
G
0
§ a d)
aO H--^ "Sc3 II o £ C
U £ G
Vi
s ao -
S o "-I
<N
X
m

u lx dx' OjlX dxOlX U >- d X cj

CNI m

CO "O
Vi d)
O
1—>
CO
(75
o O
Vi
C ,G
o
d)
1>
avC Vi
o n
^

^
Cd
d)
'll
o
V11
a s
w
X O
n 03
GO
_C
O tx
H <

367
 (N 'Cj- lO
■r—1 ▼-H
^ 8 ;-! VH
'—1

^ c <U <u <u

a -^--->
Ou OH OH OH OH
3 3 ccS 3 3
X X X X X
^ a: u u U u u

cZ < < c/3

<
> c/3
co o > o > _o3
< o o o O <33 H—>

> z X3 TD(D 3 < z X(U z

< o < CJ < 3
_o3
c/3 C/D

"XX o X _o _3 X
C/D (L>
> z cd o X3 _3
—H
3 C/D

•
c/5 O
c< < CD o O3 o3 'CH _o
w
-*-H
3o 3o
H—<
'CH X3
C/D
c/3
■(—I o s Xo
1
o AJ
o CQ
1-^^
13 < a
o
o3
H—<
X
3 (L> o3 VH H—<
X
X
3O (L>
VH

o3H—>

_3 <4—»
c/5

c CO ■
"d
X G OQ 3 X CJ _3 X "3
< Ui cO
O CJ OH
CO
<N m

8
cO

<N

JS-
CQ
CO CO oT
c/5
+ C< + <D
(/5
00
S-I
•“-5
+ •it
+ •2
CQ
cd
3
cq -it C<
cq •it lo CT
'' 0 cq c/3
CQ
+
X

0
X
’OH +
+
c +
^ X
X ^ c/5
_o 3
+ '3'
ccT bX) _o + X X 3
H—> of'
_3 XJ 0
cq § cO
o
3 + X
"OH

-4—)
•■-«3
0 CQ-
3 +
c + +
CJD
c/3
3
3 +
X c<
<N
S ttf , O .-.
CC!, (U ^
V-H
3 'OH
c/5 1. + ^
Q
+
.2
3 cZ
+ _G ^
.2 +
I
cO +
X
c/5 8 cs, C “-H
3 x: 0 4-<
CQ O
’o. =t 3, (N m
_o
_o 3 3 + (D
3 _o 2 II ^ II 4-< C G
X 3" X •*«4» + H-H w =■ <u C-)
~o 3 + -HH
3
3 o > cO
H—» ccj <U
3
N
3 =t 'OH
_3
3
II
-it
l-H
lU
CO o (JL| I>-i
cO X)
X 3
S II
a
C rT m
e 0 S
0 qx! 3
0
X 3
3 u X“ 1—1
3
X cd X

<N

c
a>

s
• T*H
!-i
<U
a
X
W

368
 Chapter 12
7l- 1 1
DC T-H 'T—H
VH C/5 -r-H DC
l-l
03T—H OJ ^ !-i
CD
H—* <l> ^ - 03
-(—>
•I—I
73
OH OH O ^ c OH
^ O o 7:3
o3 o3 3 5 03 o3 !=1
HO o3 HO aq o3
U u U U U U
replications and positions

< < g
a. Factorial ANOVA with

> > o
'S DC
<
> o o
> o o <
^
<
o
OH
< Ln
:z
"d
< > > O
<
VH *rl ^
> o O a. O
H—‘ DC o ^C/5
o 'iO
—>
CD
M-H CD
< Ci? < g
< 7D c/o TD
<U
crc
JD
0
73031 3
^ 8

O DC O 1 -1—> HO
T:5 DC ■»—> Vc ^
CD Q-) O CD o (D 03 <1^ .Id
C/5 on
<V (D
z c3 o3 3 3
3.

(N m <N m

-ii

q*;s:
oi"

+ CO

-iji s +
^ I
+ Bj

oi" CQ u •"-3 CQ .c:

+ + +
+ 3 + to"
-i**
^ °q
cq 8
CQ"
+ U
+ ^ + +
+ Rk + Jm +

CO
+ S
T3 + £ r ■“
CD ^ 03 — u
N —
_N —
+ CO
+ + +^
I cq" ^ +
C<
a ^ + .i > +
o
c +•- -O
^ C< +
'§ +
c
o U
+ a ^ u +^ + o
0-) 03 3, +
. Latin square

^ ^ X! 03
a. Complete

(U (D 3. 73
>> + cq 03 H—>3, +’ ^
;-l H—1
13 3, o3 <U N 03 c3 03 II
+
Yijkm =

ID + .^
DH II 3 'DH II
;S ^ a cr OH B 3, 'OH
a 'L C/i R ^
03
’HH II o a £
•“-i
c/5 a £
^ II -g rl O 3 73 o o
s ^ U c ■§ U >< ao --s.s a u U
O ^03 3 o3 o3
U Cti d 1—1 d u d O d
m m (N rd

03
O
o a
VH
Ut
UH o
<u H—<
CD
CHH

'O 7D
<u
-(—> 03
on
CD 03
z
pp u

369
370 Summary amd special problems

SPECIAL PROBLEMS

17.1 Three different time periods are chosen in which to plant a certain vegetable:
early May, late May, and early June. Plots are selected in early May and five
different fertilizers are randomly tried in each plot. This is repeated in late May
and again in early June. The criterion is the number of bushels per acre at harvest
time. Assuming r plots for each fertilizer can be used at each planting:
a. Discuss the nature of this experiment.
b. Recommend a value for r to make this an acceptable study. (Time periods
and fertilizers are fixed.)
17.2 In the study of the factors affecting the surface finish of steel in Problem 15.19
there was a sixth factor—factor A (the flux type)—also set at two levels. It was
possible to run all 64 (2^) treatment combinations except for the fact that the
experimenter wanted a complete replication of the experiment that was to be
run. Thus, it was decided to do a one-half replicate of the 2^ and then replicate
the same block at another time. It was also learned that first a flux type had to
be chosen, and then a slab width, and then the other four factors randomized.
After this the experimenter would run the other slab width and then repeat the
experiment with the other flux type.
Outline a data format for this problem and show what its ANOVA table
might look like.
17.3 In Problem 15.18 on the 81 engines it took about a year to gather all the
data. When the data were ready they were found to be incomplete at the low-
temperature level so the experiment was really a one-third replication of a 3“^ x 2
factorial instead of the 3^ factorial. How would you analyze the data if one
whole level in your answer to this problem were missing?
17.4 In a food industry there was a strong feeling that viscosity readings on a slurry
were not consistent. Some claimed that different laboratories gave different
readings. Others said it was the fault of the instruments in these laboratories.
And still others claimed it was the fault of the operators who read the instruments.
To determine where the main sources of variability were, slurries from two prod-
ucts with vastly different viscosities were taken. Each sample of slurry was
mixed well and then divided into four jars, one of which was sent to each of
four laboratories for analysis. In each laboratory the jar was divided into eight
samples in order for each of two randomly chosen operators within that labora-
tory to test two samples on each of two randomly chosen test instruments.
Considering the laboratories as fixed, the two products as fixed, operators as
random, and instruments as random, show a data layout for this problem. Show
its model and its ANOVA outline and discuss any concerns about the tests to
be made.

17.5 A total of 256 seventh grade students with equal numbers of boys and girls and
representing both high and low socioeconomic status (SES) was subjected to a
mathematics lesson that was believed to be either interesting or uninteresting
under one of four reinforcement conditions: no knowledge of results, immediate
verbal knowledge of results, delayed verbal knowledge of results, and delayed
written knowledge of results. Task interest and type of reinforcement emerged
 Summary and special problems 371

as significant at the 1 percent significance level. The delayed verbal condition

was significantly superior to the other three, which did not dilTer from each other.
Also, low SES children learned the high-interest task as quickly as high SES
children, and more quickly than they learned the low-interest task. On the basis
of this information answer the following.
a. What is the implied criterion T?
b. List each independent variable and state the levels of each.
c. Set up an ANOVA table for this problem and indicate which results are
significant.
d. Show a data layout indicating the number of children in each treatment.
e. For the significant results set some hypothetical means to show possible
interpretation of such results.

17.6 In testing for a significant effect of treatments on some measured variable Y the
resulting F test was almost, but not quite, significant at the 5 percent level (say,
F = 4.00 with 3 and 8 df). Now the experimenter repeats the whole experiment
on another set of independent observations under the same treatments and again
the F fails to reach the 5 percent level (say, F = 4.01 with 3 and 8 df). What
conclusion—if any—would you reach with regard to the hypothesis of equal
treatment effects? Briefly explain your reasoning.
17.7 Six different formulas or mixes of concrete are to be purchased from five competing
suppliers. Tests of crushing strength are to be made on two blocks constructed
according to each formula-supplier combination. One block will be poured,
hardened, and tested for each combination before the second block is formed,
thus making two replications of the whole experiment. Assuming mixes and
suppliers fixed and replications random, set up a mathematical model for this
situation, outline its ANOVA table, and show what F tests are appropriate.
17.8 Four chemical treatments A, B, C, and D are prepared in three concentrations
(3 to 1, 2 to 1, and 1 to 1) and two samples of each combination are poured; then
three determinations of percent of pH are made from each sample. Considering
treatments and concentrations fixed and samples and determinations random,
set up the appropriate ANOVA table and comment on statistical tests to be
made. Also explain in some detail how you would proceed to set 95 percent
confidence limits on the average percent of pH for treatment B, concentration
2 to 1.
17.9 A graduate student carried out an experiment involving six experimental con-
ditions. From an available group of subjects she assigned five men and five
women at random to each of the six conditions, using 60 subjects in all. Upon
completion of the experiment, she carried out the following analysis to test the
differences among the condition means:

Source df

Between conditions 5
F = —^ with 5 and 54 df
Within conditions 54 MS^

Total 59
372 Summary and special problems

At this point a statistical consultant was called in.

a. The consultant blew his top. Why?
b. You are now the consultant. Set up the analysis the student should use, with
sources of variation, df, and F tests, and explain to her why she must use your
analysis and why hers is incorrect.
c. As consultant you must point out the possibility that one of the significance
tests in the appropriate analysis may affect the interpretation of the other.
How? What should be done in such a case?
17.10 For the problem (Problem 6.6) on the effect of four factors on chemical yield,
the original data had an unequal number of observations at each treatment
combination. The numbers in each treatment combination below are the n’s for
each particular treatment. Explain how you might handle the analysis of this
problem.

Ai

Br B2 B, B2

D, 2 2 1 5
Cl
D2 2 1 2 3

D, 4 2 5 1
Co
D2 1 1 2 2
 Glossary of Terms

Alias. An effect in a fractionally replicated design that “looks like” or

cannot be distinguished from another effect.
Alpha (a). Size of the type I error or probability of rejecting a hypothesis
when true.
Beta (^). Size of the type II error or probability of accepting a hypothesis
when some alternative hypothesis is true.
Canonical form. Form of a second-degree response-surface equation that
allows determination of the type of surface.
Completely randomized design. A design in which all treatments are assigned
to the experimental units in a completely random manner.
Confidence limits. Two values between which a parameter is said to lie with
a specified degree of confidence.
Confounding. An experimental arrangement in which certain effects cannot
be distinguished from others. One such effect is usually blocks.
Consistent estimator. An estimator of a parameter whose value comes closer
to the parameter as the sample size is increased.
Contrast. A linear combination of treatment totals or averages where the
sum of the coefficients is zero.
Correlation coefficient (Pearson r). The square root of the proportion of
total variation accounted for by linear regression.
Correlation index R. The square root of the proportion of total variation
accounted for by the regression equation of the degree being fitted to the
data.
Covariance analysis. Technique for removing the effect of one or more
variables on the response variable before assessing the effect of treatments
on the response variable.
Critical region. A set of values of a test statistic where the hypothesis under
test is rejected.

373
374 Glossary of terms

Defining contrast. An expression that indicates which effects are confounded

with blocks in a factorial design that is confounded.
Effect of a factor. The change in response produced by a change in level of
the factor.
Errors. Type I: Rejecting a hypothesis when true. Type II: Accepting a
hypothesis when false.
Eta squared (ff). Proportion of total variation accounted for by a regression
line through all average responses for each value of the independent variable.
Evolutionary operation. An experimental procedure for collecting informa-
tion to improve a process without disturbing production.
Expected value of a statistic. The average value of a statistic if it were
calculated from an infinite number of equal-sized samples from a given
population.
Ex-post-facto research. Research in which the variables have already acted
and one is seeking for some possible relationships between the variables.
Eactorial chi square. Technique for handling attribute data in an analysis
of variance.
Eactorial experiment. An experiment in which all levels of each factor in
the experiment are combined with all levels of every other factor.

Eractional replication. An experimental design in which only a fraction of a

complete factorial is run.
Graeco-Latin square. An experimental design in which four factors are so
arranged that each level of each factor is combined only once with each
level of the other three factors.
Incomplete block design. A randomized block design in which not all treat-
ment combinations can be included in one block. Such a design is called a
balanced design if each pair of treatments occurs together the same number
of times.
Interaction. An interaction between two factors means that a change in
response between levels of one factor is not the same for all levels of the
other factor.
Latin square. An experimental design in which each level of each factor is
combined only once with each level of two other factors.
Least squares method. Method of assigning estimates of parameters in a
model such that the sum of the squares of the errors is minimized.
Max-min-con principle. A design principle in which the experimenter
attempts to maximize the variance due to the manipulated variables, mini-
mize the error variance, and control other possible sources of variance.
 Glossary of terms 375

Mean. Of a sample:

Y = t YJn
i— 1
Of a population:
fi = E{Y).
Mean square. An unbiased estimate of a population variance. Determined
by dividing a sum of squares by its degree of freedom.
Minimum variance estimate. If two or more estimates u^, Uj,... are
made of the same parameter 6, the estimate with the smallest variance is
called the minimum variance estimate.
Mixed model. The model of a factorial experiment in which one or more
factors are at fixed levels and at least one factor is at random levels.
Nested experiment. An experiment in which the levels of one factor are
chosen within the levels of another factor.
Nested-factorial experiment. An experiment in which some factors are
factorial or crossed with others and some factors are nested within others.
Operating characteristic curve. A plot of the probability of acceptance of a
hypothesis under test versus the true parameter of the population.
Orthogonal contrasts. Two contrasts are said to be orthogonal if the prod-
ucts of their corresponding coefficients add to zero.
Orthogonal polynomials. Sets of polynomials used in regression analysis
such that each polynomial in the set is orthogonal to all other polynomials
in the set.
Parameter. A characteristic of a population, such as the population mean
or variance.
Point estimate. A single statistic used to estimate a parameter.
Power of a test. A plot of the probability of rejection of a hypothesis under
test versus the true parameter. It is the complement of the operating char-
acteristic curve.
The block in a confounded design that contains the treat-
Principal block.
ment combination in which all factors are at their lowest level.
Random model. A model of a factorial experiment in which the levels of
the factor are chosen at random.
Random sample. A sample in which each member of the population
sampled has an equal chance of being selected in this sample.
Randomized block design. An experimental design in which the treatment
combinations are randomized within a block and several blocks are run.
Repeated-measures design. A design in which measurements are taken on
the experimental units more than once.
376 Glossary of terms

Research. A systematic quest for undiscovered truth.

Rotatable design. An experimental design that has equal predictive power
in all directions from a center point, and in which all experimental points
are equidistant from this center point.
Split-plot design.An experimental design in which a main effect is con-
founded with blocks due to the practical necessities of the order of experi-
mentation.
Standard error of estimate. The standard deviation of errors of estimate
around a least squares fitted regression model.
Statistic u. A measure computed from a sample.
Statistical hypothesis HQ. An assumption about a population being
sampled.
Statistical inference.Inferring something about a population of measures
from a sample of that population.
Statistics. A tool for decision making in the light of uncertainty.
Steepest ascent method. A method of sequential experiments that will
direct the experimenter toward the optimum of a response surface.
Sum of squares (SS). The sum of the squares of deviations of a random
variable from its mean,

SS = t (Fi - Yf
i= 1

Test of a hypothesis. A rule by which a hypothesis is accepted or rejected.

Test statistic. A statistic used to test a hypothesis.
Treatment combination. A given combination showing the levels of all
factors to be run for that set of experimental conditions.
True experiment. A study in which certain independent variables are
manipulated, their effect on one or more dependent variables is observed,
and the levels of the independent variables are assigned at random to the
experimental units of the study.
Unbiased statistic. A statistic whose expected value equals the parameter it
is estimating.
Variance. Of a sample:

2 T,UAY,-Yf SS
f! — 1 df
Of a population:
= E(Y - nf.
Youden square. An incomplete Latin square.
 References

1. Anderson, V. L., and R. A. McLean, “Restriction Errors: Another Dimension in

Teaching Experimental Statistics,” The American Statistician (Nov. 1974).
2. Barnett, E. H., “Introduction to Evolutionary Operation,” Industrial and Engi-
neering Chemistry, vol. 52 (June 1960), p. 500.
3. Batson, H. C., “Applications of Factorial Analysis to Experiments in Chemistry,”
National Convention Transactions of the American Society for Quality Control
(1956), pp. 9-23.
4. Bennett, C. A., and N. L. Franklin, Statistical Analysis in Chemistry and the Chemical
Industry. New York: John Wiley & Sons, Inc. (1954).
5. Biometrics, vol. 13 (Sept. 1957), pp. 261-405.
6. Box, G. E. P., and N. R. Draper, Evolutionary Operation: A Statistical Method
for Process Improvement. New York: John Wiley & Sons, Inc. (1969).
7. Burr, I. W., Statistical Quality Control Methods, New York: Marcel Dekker (1976).
8. Burr, I. W., Applied Statistical Methods, New York: Academic Press (1974).
9. Davies, O. L., Design and Analysis of Industrial Experiments. New York: Hafner
Publishing Company (1954).
10. Dixon, W. J., and F. J. Massey, An Introduction to Statistical Analysis (2nd Ed.).
New York: McGraw-Hill Book Company (1957).
11. Fisher, R. A., and F. Yates, Statistical Tables for Biological, Agricultural and
Medical Research (4th Ed.). Edinburgh and London: Oliver & Boyd, Ltd. (1953).
12. Hunter, J. S., “Determination of Optimum Operating Conditions by Experimental
Methods,” Industrial Quality Control (Dec.-Feb. 1958-1959).
13. Kempthorne, O., The Design and Analysis of Experiments. New York: John Wiley &
Sons, Inc. (1952).
14. Kerlinger, F. N., Foundations of Behavioral Research (2nd Ed.) New York: Holt,
Rinehart and Winston (1973).
15. Keuls, M., “The Use of the Studentized Range in Connection with an Analysis of
Variance,” Euphytica, vol. 1 (1952), pp. 112-122.
16. Leedy, P. D., Practical Research: Planning and Design. New York: Macmillan (1974).

377
378 References

17. McCall, C. H., Jr., “Linear Contrasts, Parts I, II, and III,” Industrial Quality Control
(July-Sept., 1960).
18. Miller, L. D., “An Investigation of the Machinability of Malleable Iron Using
Ceramic Tools,” unpublished MSIE thesis, Purdue University (1959).
19. Ostle, B., Statistics in Research (2nd Ed.). Ames, Iowa: Iowa State University Press
(1963).
20. Owen, D. B., Handbook of Statistical Tables. Boston: Addison-Wesley Publishing
Company, Inc. (1962).
21. Scheffe, H., “A Method for Judging All Contrasts in the Analysis of Variance,”
Biometrics, vol. XL (June, 1953).
22. Snedecor, G. W., and W. C. Cochran, Statistical Methods (6th Ed.). Ames, Iowa:
Iowa State University Press (1967).
23. Tanur, Judith M., editor, and others. Statistics; A Guide to the Unknown (2nd Ed.)
San Francisco: Holden-Day (1978).
24. Winer, B. J., Statistical Principles in Experimental Design (2nd Ed.). New York:
McGraw-Hill Book Company (1971).
25. Wortham, A. W., and T. E. Smith, Practical Statistics in Experimental Design.
Dallas: Dallas Publishing House (1960).
26. Yates, F., Design and Analysis of E act or ial Experiments. London: Imperial Bureau
of Soil Sciences (1937).
 ‘ . -■' f. 1^ I iT T. W,- - ■• ./i.»> J. Vitt ■>■» r .4 .'-ij. *■ ' :'V • . ■ ■
.s # ;■
•’ 'J , -V*^^
'.T^-v
•

rS-' ■. jM‘ : ■ X :■ -■■ ‘/V. ‘ • “ 'iite ^

^..-^ . M-.it.4-r • *-!»■, (V ' ‘ I ' -•'V.*'-', .■' ■ .1 '■iiB
ilrji •■■- MtM^t’ ■ ^ '?■ .r. .
W«V-^ ** ■ '^ . 1* ► '^1, ^ ' '* J ‘ * • "i , * '•^1-^1,'* i. ’ ^ ’ «••* ^ ^ - •jfcu
* , .■
1%-i'

. 7,C-■■ ' <-' C-. ]-r f’^ ■'. , r’•■ • -</■:>((

.. .. . ’ll. y• ... ”' t ' I • •' 1! . .

, ».i ^r ■■• VJ.T'I • /\, •.‘•"* „'

L.^-'"f-

EG -IpIP' ifT r.
/ • • " • . *' f, **>' J't
v*i* ,'•. ^> ]

.' I ■. ';,* " •'* . • .,

fee. ■
L>i'*'/,'’■ ; ;A,. '■ ,iitf'

e .i ':i:-A ^
•TV-' AvyS*"**,
14'^-
rjH.Vi.'r

'■■ )-’A SfV ' i- '■ ‘‘‘ ■-’ iJ

t'e- •
■■■
'
^ A r.’^'-j'.i'i
' '} 1: ,^.r v. /i.^;.'yv:.''.5 .,';'^^^^l' 'y
r-' .. ‘
5' •. ''tf-V'.
 Statistical Tables

Table A Areas under the Normal Curve* (Proportion of Total Area under the
Curve from — oo to Designated Z Value)

z 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0.00

-3.5 0.00017 0.00017 0.00018 0.00019 0.00019 0.00020 0.00021 0.00022 0.00022 0.00023
-3.4 0.00024 0.00025 0.00026 0.00027 0.00028 0.00029 0.00030 0.00031 0.00033 0.00034
-3.3 0.00035 0.00036 0.00038 0.00039 0.00040 0.00042 0.00043 0.00045 0.00047 0.00048
-3.2 0.00050 0.00052 0.00054 0.00056 0.00058 0.00060 0.00062 0.00064 0.00066 0.00069
-3.1 0.00071 0.00074 0.00076 0.00079 0.00082 0.00085 0.00087 0.00090 0.00094 0.00097

-3.0 0.00100 0.00104 0.00107 0.00111 0.00114 0.00118 0.00122 0.00126 0.00131 0.00135
-2.9 0.0014 0.0014 0.0015 0.0015 0.0016 0.0016 0.0017 0.0017 0.0018 0.0019
-2.8 0.0019 0.0020 0.0021 0.0021 0.0022 0.0023 0.0023 0.0024 0.0025 0.0026
-2.7 0.0026 0.0027 0.0028 0.0029 0.0030 0.0031 0.0032 0.0033 0.0034 0.0035
-2.6 0.0036 0.0037 0.0038 0.0039 0.0040 0.0041 0.0043 0.0044 0.0045 0.0047
-2.5 0.0048 0.0049 0.0051 0.0052 0.0054 0.0055 0.0057 0.0059 0,0060 0.0062
-2.4 0.0064 0.0066 0.0068 0.0069 0.0071 0.0073 0.0075 0.0078 0.0080 0.0082
-2.3 0.0084 0.0087 0.0089 0.0091 0.0094 0.0096 0.0099 0.0102 0.0104 0.0107
-2.2 0.0110 0.0113 0.0116 0.0119 0.0122 0.0125 0.0129 0.0132 0.0136 0.0139
-2.1 0.0143 0.0146 0.0150 0.0154 0.0158 0,0162 0.0166 0.0170 0.0174 0.0179
-2.0 0.0183 0.0188 0.0192 0.0197 0.0202 0.0207 0.0212 0.0217 0.0222 0.0228
-1.9 0.0233 0.0239 0.0244 0.0250 0.0256 0.0262 0.0268 0.0274 0.0281 0.0287
-1.8 0.0294 0.0301 0.0307 0,0314 0.0322 0.0329 0.0336 0.0344 0.0351 0.0359
-1.7 0.0367 0.0375 0.0384 0.0392 0.0401 0.0409 0.0418 0.0427 0.0436 0.0446
-1.6 0.0455 0.0465 0.0475 0,0485 0.0495 0.0505 0.0516 0.0526 0.0537 0.0548
-1.5 0.0559 0.0571 0.0582 0.0594 0,0606 0.0618 0.0630 0.0643 0.0655 0.0668
-1.4 0.0681 0.0694 0.0708 0.0721 0.0735 0.0749 0.0764 0.0778 , 0.0793 0.0808
-1.3 0.0823 0.0838 0.0853 0.0869 0.0885 0.0901 0.0918 0.0934 0.0951 0.0968
-1.2 0.0985 0.1003 0.1020 0.1038 0.1057 0.1075 0.1093 0.1112 0.1131 0.1151
-1.1 0.1170 0.1190 0.1210 0.1230 0.1251 0 1271 0.1292 0.1314 0.1335 0.1357

-1.0 0.1379 0.1401 0.1423 0.1446 0.1469 0.1492 0.1515 0.1539 0,1562 0.1587
-0.9 0.1611 0.1635 0.1660 0.1685 0.1711 0.1736 0.1762 0.1788 0.1814 0.1841
-0.8 0.1867 0.1894 0.1922 0.1949 0.1977 0.2005 0.2033 0.2061 0.2090 0.2119
—0.7 0.2148 0.2177 0.2207 0.2236 0.2266 0.2297 0.2327 0.2358 0.2389 0.2420
-0.6 0.2451 0.2483 0.2514 0.2546 0.2578 0.2611 0.2643 0.2676 0.2709 0.2743
-0.5 0.2776 0.2810 0.2843 0.2877 0.2912 0.2946 0.2981 0.3015 0.3050 0.3085
-0.4 0.3121 0.3156 0.3192 0.3228 0.3264 0.3300 0.3336 0.3372 0.3409 0.3446
-0.3 0.3483 0.3520 0.3557 0.3594 0.3632 0.3669 0.3707 0.3745 0.3783 0.3821
-0.2 0.3859 0.3897 0.3936 0.3974 0.4013 0.4052 0.4090 0.4129 0.4168 0,4207
-0.1 0.4247 0.4286 0.4325 0.4364 0.4404 0.4443 0.4483 0.4522 0.4562 0.4602
-0.0 0.4641 0.4681 0.4721 0.4761 0.4801 0.4840 0.4880 0.4920 0.4960 0.5000

* Adapted from E. L. Grant, Statistical Quality Control (2nd Ed.). New York: McGraw-
Hill Book Company, Inc. (1952), Table A, pp. 510-511. Reproduced by permission of the
publisher.

380
 Statistical tables 381

Table A {Continued)

z 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0,08 0.09

+0.0 0.5000 0.6040 0.5080 0.5120 0.5160 0.5199 0.5239 0.5279 0.5319 0.5359
+0.1 0.6398 0.5438 0.5478 0.5517 0.5557 0.5596 0.5636 0.5675 0.5714 0.5753
+0.2 0.5793 0.5832 0.5871 0.5910 0.5948 0.5987 0.6026 0.6064 0.6103 0.6141
+0.3 0.6179 0.6217 0.6255 0.6293 0.6331 0.6368 0.6406 0.6443 0.6480 0.6517
+0.4 0.6554 0.6591 0.6628 0.6664 0.6700 0.6736 0.6772 0.6808 0.6844 0.6879
+0.5 0.6915 0.6950 0.6985 0.7019 0.7054 0.7088 0.7123 0.7157 0.7190 0.7224

+0.6 0.7257 0.7291 0.7324 0.7357 0.7389 0.7422 0.7454 0.7486 0.7517 0.7549
+0.7 0.7580 0.7611 0.7642 0.7673 0.7704 0.7734 0.7764 0.7794 0.7823 0.7852
+0.8 0.7881 0.7910 0.7939 0.7967 0.7995 0.8023 0.8051 0.8079 0.8106 0.8133
+0.9 0.8159 0.8186 0.8212 0.8238 0.8264 0.8289 0.8315 0.8340 0.8365 0.8389
+ 1.0 0.8413 0.8438 0.8461 0.8485 0.8508 0.8531 0.8554 0.8577 0,8599 0.8621

+ 1.1 0.8643 0.8665 0.8686 0.8708 0.8729 0.8749 0.8770 0.8790 0.8810 0.8830,
+ 1.2 0.8849 0.8869 0.8888 0.8907 0.8925 0.8944 0.8962 0.8980 0.8997 0.9015
+ 1.3 0.9032 0.9049 0.9066 0.9082 0.9099 0.9115 0.9131 0.9147 0.9162 0.9177
+ 1.4 0.9192 0.9207 0.9222 0.9236 0.9251 0.9265 0.9279 0.9292 0.9306 0.9319
+ 1.5 0.9332 0.9345 0.9357 0.9370 0.9382 0.9394 0.9406 0.9418 0.9429 0.9441

+ 1.6 0.9452 0.9463 0.9474 0.9484 0.9495 0.9505 0.9515 0.9525 0.9535 0.9545
+ 1.7 0.9554 0.9564 0.9573 0.9582 0.9591 0.9599 0.9608 0.9616 0.9625 0.9633
+ 1.8 0.9641 0.9649 0.9656 0.9664 0.9671 0.9678 0.^68^ 0.9693 0.9699 0.9706
+ 1.9 0.9713 0.9719 0.9726 0.9732 0.9738 0.9744( 0.9750 0.9756 0.9761 0.9767
+2.0 0.9773 0.9778 0.9783 0.9788 0.9793 0.9798^ 0r0803 0.9808 0.9812 0.9817

+2.1 0.9821 0.9826 0.9830 0.9834 0.9838 0.9842 0.9846 0.9850 0.9854 0.9857
+2.2 0.9861 0.9864 0.9868 0.9871 0.9875 0.9878 0.9881 0.9884 0.9887 0.9890
+2.3 0.9893 0.9896 0.9898 0.9901 0.9904 0.9906 0.9909 0.9911 0.9913 0.9916
+2.4 0.9918 0.9920 0.9922 0.9925 0.9927 0.9929 0.9931 0.9932 0.9934 0.9936
+2.6 0.9938 0.9940 0.9941 0.9943 0.9945 0.9946 0.9948 0.9949 0.9951 0.9952

+2.6 0.9953 0.9955 0.9956 0.9957 0.9959 0.9960 0.9961 0.9962 0.9963 0.9964
+2.7 0.9965 0.9966 0.9967 0.9968 0.9969 0.9970 0.9971 0.9972 0.9973 0.9974
+2.8 0.9974 0.9975 0.9976 0.9977 0.9977 0.9978 0.9979 0.9979 0.9980 0.9981
+2.9 0.9981 0.9982 0.9983 0.9983 0.9984 0.9984 0.9985 0.9985 0.9986 0.9986
+3.0 0.99865 0.99869 0.99874 0.99878 0.99882 0.99886 0.99889 0.99893 0.99896 0.99900

+3.1 0.99903 0.99906 0.99910 0.99913 0.99915 0.99918 0.99921 0.99924 0.99926 0.99929
+3.2 0.99931 0.99934 0.99936 0.99938 0.99940 0.99942 0.99944 0.99946 0.99948 0.99950
+3.3 0.99952 0.99953 0.99955 0.99957 0.99958 0.99960 0.99961 0.99962 0.99964 0.99965
+3.4 0.99966 0.99967 0.99969 0.99970 0.99971 0.99972 0.99973 0.99974 0.99975 0.99976
+3.5 0.99977 0.99978 0.99978 0.99979 0.99980 0.99981 0.99981 0.99982 0.99983 0.99983
 B Student’s t Distribution

Percentile point

70 80 90 95 97.5 99 99.5

1 .73 1.38 3.08 6.31 12.71 31.82 63.66

2 .62 1.06 1.89 2.92 4.30 6.96 9.92
3 .58 .98 1.64 2.35 3.18 4.54 5.84
4 .57 .94 1.53 2.13 2.78 3.75 4.60

5 .56 .92 1.48 2.01 2.57 3.36 4.03

6 .55 .91 1.44 1.94 2.45 3.14 3.71
7 .55 .90 1.42 1.90 2236 3.00 3.50
8 .55 .89 1.40 1.86 2.31 2.90 3.36
'9 .54 .88 1.38 ^ 1.83 2.26 2.82 3.25

10 .54 .88 1.37 1.81 2.23 2.76 3.17

11 .54 .88 1.36 1.80 2.20 2.72 3.11
12 .54 .87 1.36 1.78 2.18 2.68 3.06
13 .54 .87 1.35 1.77 2.16 2.65 3.01
14 .54 .87 1.34 1.76 2.14 2.62 2.98

15 .54 .87 1.34 1.75 2.13 2.60 2.95

16 .54 .86 1.34 1.75 2.12 2.58 2.92
17 .53 .86 1.33 1.74 2.11 2.57 2.90
18 .53 .86 1.33 1.73 2.10 2.55 2.88
19 .53 .86 1.33 1.73 2.09 2.54 2.86

20 .53 .86 1.32 1.72 2.09 2.53 2.84

21 .53 .86 1.32 1.72 2.08 2.52 2.83
22 .53 .86 1.32 1.72 2.07 2.51 2.82
23 .53 .86 1.32 1.71 2.07 2.50 2.81
24 .53 .86 1.32 1.71 2.06 2.49 2.80

25 .53 .86 1.32 1.71 2.06 2.48 2.79

26 .53 .86 1.32 1.71 2.06 2.48 2.78
27 .53 .86 1.31 1.70 2.05 2.47 2.77
28 .53 .86 1.31 1.70 2.05 2.47 2.76
29 .53 .85 1.31 1.70 2.04 2.46 2.76

30 .53 .85 1.31 1.70 2.04 2.46 2.75

40 .53 .85 1.30 1.68 2.02 2.42 2.70
50 .53 .85 1.30 1.67 2.01 2.40 2.68
60 .53 .85 1.30 1.67 2.00 2.39 2.66
80 .53 .85 1.29 1.66 1.99 2.37 2.64

.53 .84 1.29 1.66 1.98 2.36 2.63

.52 .84 1.29 1.65 1.97 2.34 2.60
.52 .84 1.28 1.65 1.96 2.33 2.59

.52 .84 1.28 1.64 1.96 2.33 2.58

382
Table C Chi Square*

Probability
V

0.99 0.98 0.95 0.90 0.80 0.20 0.10 0.05 0.02 0.01 0.001

1 0.0n57 0.0^628 0.00393 0.0158 0.0642 1.642 2.706 3.841 5.412 6.635 10.827
2 0.0201 0.0404 0.103 0.211 0.446 3.219 4.605 5.991 7.824 9.210 13.815
3 0.115 0.185 0.352 0.584 1.005 4.642 6.251 7.815 9.837 11.341 16.268
4 0.297 0.429 0.711 1.064 1.649 5.989 7.779 9.488 11.668 13.277 18.465
5 0.554 0.752 1.145 1.610 2.343 7.289 9.236 11.070 13.388 15.086 20.517

6 0.872 1.134 1.635 2.204 3.070 8.558 10.645 12.592 15.033 16.812 22.457
7 1.239 1.564 2.167 2.833 3.822 9.803 12.017 14.067 16.622 18.475 24.322
8 1.646 2.032 2.733 3.490 4.594 11.030 13.362 15.507 18.168 20.090 26.125
9 2.088 2.532 3.325 4.168 5.380 12.242 14.684 16.919 19.679 21.666 27.877
10 2.558 3.059 3.940 4.865 6.179 13.442 15.987 18.307 21.161 23.209 29.588

11 3.053 3.609 4.575 5.578 6.989 14.631 17.275 19.675 22.618 24.725 31.264
12 3.571 4.178 5.226 6.304 7.807 15.812 18.549 21.026 24.054 26.217 32.909
13 4.107 4.765 5.892 7.042 8.634 16.985 19.812 22.362 25.472 27.688 34.528
14 4.660 5.368 6.571 7.790 9.467 18.151 21.064 23.685 26.873 29.141 36.123
15 5.229 5.985 7.261 8.547 10.307 19.311 22.307 24.996 28.259 30.578 37.697

16 5.812 6.614 7.962 9.312 11.152 20.465 23.542 26.296 29.633 32.000 39.252
17 6.408 7.255 8.672 10.085 12.002 21.615 24.769 27.587 30.995 33.409 40.790
18 7.015 7.906 9.390 10.865 12.857 22.760 25.989 28.869 32.346 34.805 42.312
19 7.633 8.567 10.117 11.651 13.716 23.900 27.204 30.144 33.687 36.191 43.820
20 8.260 9.237 10.851 12.443 14.578 25.038 28.412 31.410 35.020 37.566 45.315

21 8.897 9.915 11.591 13.240 15.445 26.171 29.615 32.671 36.343 38.932 46.797
22 9.542 10.600 12.338 14.041 16.314 27.301 30.813 33.924 37.659 40.289 48.268
23 10.196 11.293 13.091 14.848 17.187 28.429 32.007 35.172 38.968 41.638 49.728
24 10.856 11.992 13.848 15.659 18.062 29.553 33.196 36.415 40.270 42.980 51.179
25 11.524 12.697 14.611 16.473 18.940 30.675 34.382 37.652 41.566 44.314 52.620

26 12.198 13.409 15.379 17.292 19.820 31.795 35.563 38.885 42.856 45.642 54.052
27 12.879 14.125 16.151 18.114 20.703 32.912 36.741 40.113 44.140 46.963 55.476
28 13.565 14.847 16.928 18.939 21.588 34.027 37.916 41.337 45.419 48.278 56.893
29 14.256 15.574 17.708 19.768 22.475 35.139 39.087 42.557 46.693 49.588 58.302
30 14.953 16.306 18.493 20.599 23.364 36.250 40.256 43.773 47.962 50.892 59.703

* Abridged from R. A. Fisher and F. Yates, Statistical Tables for Biological, Agricultural and
Medical Research. Edinburgh and London: Oliver & Boyd, Ltd (1953), Table IV. Reproduced
by perm.ission of the authors and publishers. For larger values of v, the expression —
■sj2v — 1 may be used as a normal deviate with unit variance, remembering that the probability
for corresponds with that of a single tail of a normal curve.

383
 istribution*
0 F

df for numerator

1 2 3 4 5 6 7 8 9 10 11 12

5.83 7.50 8.20 8.58 8.82 8.98 9.10 9.19 9.26 9.32 9.36 9.41
1 39.9 49.5 53.6 55.8 57.2 58.2 58.9 59.4 59.9 60.2 60.5 60.7
161 200 216 225 230 234 237 239 241 242 243 244

2.57 3.00 3.15 3.23 3.28 3.31 3.34 3.35 3.37 3.38 3.39 3.39
2 8.53 9.00 9.16 9.24 9.29 9.33 9.35 9.37 9.38 9.39 9.40 9.41
18.5 19.0 19.2 19.2 19.3 19.3 19.4 19.4 19.4 19.4 19.4 19.4
98.5 99.0 99.2 99.2 99.3 99.3 99.4 99.4 99.4 99.4 99.4 99.4

2.02 2.28 2.36 2.39 2.41 2.42 2.43 2.44 2.44 2.44 2.45 2.45
3 5.54 5.46 5.39 5.34 5.31 5.28 5.27 5.25 5.24 5.23 5.22 5.22
10.1 9.55 9.28 9.12 9.10 8.94 8.89 8.85 8.81 8.79 8.76 8.74
34.1 30.8 29.5 28.7 28.2 27.9 27.7 27.5 27.3 27.2 27.1 27.1

1.81 2.00 2.05 2.06 2.07 2.08 2.08 2.08 2.08 2.08 2.08 2.08
4 4.54 4.32 4.19 4.11 4.05 4.01 3.98 3.95 3.94 3.92 3.91 3.90
7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96 5.94 5.91
21.2 18.0 16.7 16.0 15.5 15.2 15.0 14.8 14.7 14.5 14.4 14.4

1.69 1.85 1.88 1.89 1.89 1.89 1.89 1.89 1.89 1.89 1.89 1.89
5 4.06 3.78 3.62 3.52 3.45 3.40 3.37 3.34 3.32 3.30 3.28 3.27
6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.77 4.74 4.71 4.68
16.3 13.3 12.1 11.4 11.0 10.7 10.5 10.3 10.2 10.1 9.96 9.89

1.62 1.76 1.78 1.79 1.79 1.78 1.78 1.77 1.77 1.77 1.77 1.77
6 3.78 3.46 3.29 3.18 3.11 3.05 3.01 2.98 2.96 2.94 2.92 2.9C
5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06 4.03 4.00
13.7 10.9 9.78 9.15 8.75 8.47 8.26 8.10 7.98 7.87 7.79 7.72

1.57 1.70 1.72 1.72 1.71 1.71 1.70 1.70 1.69 1.69 1.69 1.68
7 3.59 3.26 3.07 2.96 2.88 2.83 2.78 2.75 2.72 2.70 2.68 2.67
5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.64 3.60 3.57
12.2 9.55 8.45 7.85 7.46 7.19 6.99 6.84 6.72 6.62 6.54 6.47

1.54 1.66 1.67 1.66 1.66 1.65 1.64 1.64 1.64 1.63 1.63 1.62
8 3.46 3.11 2.92 2.81 2.73 2.67 2.62 2.59 2.56 2.54 2.52 2.50
5.32 4.46 4.07 3.84 3.69 3.58 3.50 3.44 3.39 3.35 3.31 3.28
11.3 8.65 7.59 7.01 6.63 6.37 6.18 6.03 5.91 5.81 5.73 5.67

1.51 1.62 1.63 1.63 1.62 1.61 1.60 1.60 1.59 1.59 1.58 1.58
9 3.36 3.01 2.81 2.69 2.61 2.55 2.51 2.47 2.44 2.42 2.40 2.38
5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.14 3.10 3.07
10.6 8.02 6.99 6.42 6.06 5.80 5.61 5.47 5.35 5.26 5.18 5.11

1.49 1.60 1.60 1.59 1.59 1.58 1.57 1.56 1.56 1.55 1.55 1.54
10 3.28 2.92 2.73 2.61 2.52 2.46 2.41 2.38 2.35 2.32 2.30 2.28
4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.98 2.94 2.91
10.0 7.56 6.55 5.99 5.64 5.39 5.20 5.06 4.94 4.85 4.77 4.71

1.47 1.58 1.58 1.57 1.56 1.55 1.54 1.53 1.53 1.52 1.52 1;51
11 3.23 2 86 2.66 2.54 2.45 2.39 2.34 2.30 2.27 2.25 2.23 2.21
4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.90 2.85 2.82 2.79
9.65 7.21 6.22 5.67 5.32 5.07 4.89 4.74 4.63 4.54 4.46 4.40

1.46 1.56 1.56 1.55 1.54 1.53 1.52 1.51 1.51 1.50 1.50 1.49
12 3.18 2.81 2.61 2.48 2.39 2.33 2.28 2.24 2.21 2.19 2.17 2.15
4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.80 2.75 2.72 2.69
9.33 6.93 5.95 5.41 5.06 4.82 4.64 4.50 4.39 4.30 4.22 4.16

384
Table D {Continued)

df for numerator
1 -a df for
15 20 24 30 40 50 60 100 120 200 500 00 denom.

9.49 9.58 9.63 9.67 9.71 9.74 9.76 9.78 9.80 9.82 9.84 9.85 .75
61.2 61.7 62.0 62.3 62.5 62.7 62.8 63.0 63.1 63.2 63.3 63.3 .90 1
246 248 249 250 251 252 252 253 253 254 254 254 .95

3.41 3.43 3.43 3.44 3.45 3.45 3.46 3.47 3.47 3.48 3.48 3.48 .75
9.42 9.44 9.45 9.46 9.47 9.47 9.47 9.48 9.48 9.49 9.49 9.49 .90 2
19.4 19.4 19.5 19.5 19.5 19.5 19.5 19.5 19.5 19.5 19.5 19.5 .95
99.4 99.4 99.5 99.5 99.5 99.5 99.5 99.5 99.5 99.5 99.5 99.5 .99

2.46 2.46 2.46 2.47 2.47 2.47 2.47 2.47 2.47 2.47 2.47 2.47 .75
5.20 5.18 5.18 5.17 5.16 5.15 5.15 5.14 5.14 5.14 5.14 5.13 .90 3
8.70 8.66 8.64 8.62 8.59 8.58 8.57 8.55 8.55 8.54 8.53 8.53 .95
26.9 26.7 26.6 26.5 26.4 26.4 26.3 26.2 26.2 26.2 26.1 26.1 .99

2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 2.08 .75
3.87 3.84 3.83 3.82 3.80 3.80 3.79 3.78 3.78 3.77 3.76 3.76 .90
5.86 5.80 5.77 5.75 5.72 5.70 5.69 5.66 5.66 5.65. 5.64 5.63 .95 4
14.2 14.0 13.9 13.8 13.7 13.7 13.7 13.6 13.6 13.5 13.5 13.5 .99

1.89 1.88 1.88 1.88 1.88 1.88 1.87 1.87 1.87 1.87 1.87 1.87 .75
3.24 3.21 3.19 3.17 3.16 3.15 3.14 3.13 3.12 3.12 3.11 3.10 .90 5
4.62 4.56 4.53 4.50 4.46 4.44 4.43 4.41 4.40 4.39 4.37 4.36 .95
9.72 9.55 9.47 9.38 9.29 9.24 9.20 9.13 9.11 9.08 9.04 9.02 .99

1.76 1.76 1.75 1.75 1.75 1.75 1.74 1.74 1.74 1.74 1.74 1.74 .75
2.87 2.84 2.82 2.80 2.78 2.77 2.76 2.75 2.74 2.73 2.73 2.72 .90 6
3.94 3.87 3.84 3.81 3.77 3.75 3.74 3.71 3.70 3.69 3.68 3.67 .95
7.56 7.40 7.31 7.23 7.14 7.09 7.06 6.99 6.97 6.93 6,90 6.88 .99

1.68 1.67 1.67 1.66 1.66 1.66 1.65 1.65 1.65 1.65 1.65 1.65 .75
2.63 2.59 2.58 2.56 2.54 2.52 2.51 2.50 2.49 2.48 2.48 2.47 .90 7
3.51 3.44 3.41 3.38 3.34 3.32 3.30 3.27 3.27 3.25 3.24 3.23 .95
6.31 6.16 6.07 5.99 5.91 5.86 5.82 5.75 5.74 5.70 5.67 5.65 .99

1.62 1.61 1.60 1.60 1.59 1.59 1.59 1.58 1.58 1.58 1.58 1.58 .75
2.46 2.42 2.40 2.38 2.36 2.35 2.34 2.32 2.32 2.31 2.30 2.29 .90 8
3.22 3.15 3.12 3.08 3.04 3.02 3.01 2.97 2.97 2.95 2.94 2.93 .95
5.52 5.36 5.28 5.20 5.12 5.07 5.03 4.96 4.95 4.91 4.88 4.86 .99

1.57 1.56 1.56 1.55 1.55 1.54 1.54 1.53 1.53 1.53 1.53 1.53 .75
2.34 2.30 2.28 2.25 2.23 2.22 2.21 2.19 2.18 2.17 2.17 2.16 .90 9
3.01 2.94 2.90 2.86 2.83 2.80 2.79 2.76 2.75 2.73 2.72 2.71 .95
4.96 4.81 4.73 4.65 4.57 4.52 4.48 4.42 4.40 4.36 4.33 4.31 .99

1.53 1.52 1.52 1.51 1.51 1.50 1.50 1.49 1.49 1.49 1.48 1.48 .75
2.24 2.20 2.18 2.16 2.13 2.12 2.11 2.09 2.08 2.07 2.06 2.06 .90 10
2.85 2.77 2.74 2.70 2.66 2.64 2.62 2.59 2.58 2.56 2.55 2.54 .95
4.56 4.41 4.33 4.25 4.17 4.12 4.08 4.01 4.00 3.96 3.93 3.91 .99

1.50 1.49 1.49 1.48 1.47 1.47 1.47 1.46 1.46 1.46 1.45 1.45 .75
2.17 2.12 2.10 2.08 2.05 2.04 2.03 2.00 2.00 1.99 1.98 1.97 .90 11
2.72 2.65 2.61 2.57 2.53 2.51 2.49 2.46 2.45 2.43 2.42 2.40 .95
4.25 4.10 4.02 3.94 3.86 3.81 3.78 3.71 3.69 3.66 3.62 3.60 .99

1.48 1.47 1.46 1.45 1.45 1.44 1.44 1.43 1.43 1.43 1.42 1.42 .75
2.10 2.06 2.04 2.01 1.99 1.97 1.96 1.94 1.93 1.92 1.91 1.90 .90 12
2.62 2.54 2.51 2.47 2.43 2.40 2.38 2.35 2.34 2.32 2.31 2.30 .95
4.01 3.86 3.78 3.70 3.62 3.57 3.54 3.47 3.45 3.41 3.38 3.36 .99

385
 df for numerator

1 2 3 4 5 6 7 8 9 10 11 12
1.45 1.54 1.54 1.53 1.52 1.51 1.50 1.49 1.49 1.48 1.47 1.47
13 3.14 2.76 2.56 2.43 2.35 2.28 2.23 2.20 2.16 2.14 2.12 2.10
4.67 3.81 3.41 3.18 3.03 2.92 2.83 2.77 2.71 2.67 2.63 2.60
9.07 6.70 5.74 5.21 4.86 4.62 4.44 4.30 4.19 4.10 4.02 3.96

1.44 1.53 1.53 1.52 1.51 1.50 1.48 1.48 1.47 1.46 1.46 1.45
14 3.10 2.73 2.52 2.39 2.31 2.24 2.19 2.15 2.12 2.10 2.08 2.05
4.60 3.74 3.34 3.11 2.96 2.85 2.76 2.70 2.65 2.60 2.57 2.53
8.86 6.51 5.56 5.04 4.69 4.46 4.28 4.14 4.03 3.94 3.86 3.80

1.43 1.52 1.52 1.51 1.49 1.48 1.47 1.46 1.46 1.45 1.44 1.44
15 3.07 2.70 2.49 2.36 2.27 2.21 2.16 2.12 2.09 2.06 2.04 2.02
4.54 3.68 3.29 3.06 2.90 2.79 2.71 2.64 2.59 2.54 2.51 2.48
8.68 6.36 5.42 4.89 4.56 4.32 4.14 4.00 3.89 3.80 3.73 3.67

1.42 1.51 1.51 1.50 1.48 1.48 1.47 1.46 1.45 1.45 1.44 1.44
16 3.05 2.67 2.46 2.33 2.24 2.18 2.13 2.09 2.06 2.03 2.01 1.99
4.49 3.63 3.24 3.01 2.85 2.74 2.66 2.59 2.54 2.49 2.46 2.42
8.53 6.23 5.29 4.77 4.44 4.20 4.03 3.89 3.78 3.69 3.62 3.55

1.42 1.51 1.50 1.49 1.47 1.46 1.45 1.44 1.43 1.43 1.42 1.41
17 3.03 2.64 2.44 2.31 2.22 2.15 2.10 2.06 2.03 2.00 1.98 1.96
4.45 3.59 3.20 2.96 2.81 2.70 2.61 2.55 2.49 2.45 2.41 2.38
8.40 6.11 5.18 4.67 4.34 4.10 3.93 3.79 3.68 3.59 3.52 3.46

1.41 1.50 1.49 1.48 1.46 1.45 1.44 1.43 1.42 1.42 1.41 1.40
18 3.01 2.62 2.42 2.29 2.20 2.13 2.08 2.04 2.00 1.98 1.96 1.93
4.41 3.55 3.16 2.93 2.77 2.66 2.58 2.51 2.46 2.41 2.37 2.34
8.29 6.01 5.09 4.58 4.25 4.01 3.84 3.71 3.60 3.51 3.43 3.37

1.41 1.49 1.49 1.47 1.46 1.44 1.43 1.42 1.41 1.41 1.40 1.40
19 2.99 2.61 2.40 2.27 2.18 2.11 2.06 2.02 1.98 1.96 1.94 1.91
4.38 3.52 3.13 2.90 2.74 2.63 2.54 2.48 2.42 2.38 2.34 2.31
8.18 5.93 5.01 4.50 4.17 3.94 3.77 3.63 3.52 3.43 3.36 3.30

1.40 1.49 1.48 1.46 1.45 1.44 1.42 1.42 1.41 1.40 1.39 1.39
20 2.97 2.59 2.38 2.25 2.16 2.09 2.04 2.00 1.96 1.94 1.92 1.89
4.35 3.49 3.10 2.87 2.71 2.60 2.51 2.45 2.39 2.35 2.31 2.28
8.10 5.85 4.94 4.43 4.10 3.87 3.70 3.56 3.46 3.37 3.29 3.23

1.40 1.48 1.47 1.45 1.44 1.42 1.41 1.40 1.39 1.39 1.38 1.37
22 2.95 2.56 2.35 2.22 2.13 2.06 2.01 1.97 1.93 1.90 1.88 1.86
4.30 3.44 3.05 2.82 2.66 2.55 2.46 2.40 2.34 2.30 2.26 2.23
7.95 5.72 4.82 4.31 3.99 3.76 3.59 3.45 3.35 3.26 3.18 3.12
1.39 1.47 1.46 1.44 1.43 1.41 1.40 1.39 1.38 1.38 1.37 1.36
24 2.93 2.54 2.33 2.19 2.10 2.04 1.98 1.94 1.91 1.88 1.85 1.83
4.26 3.40 3.01 2.78 2.62 2.51 2.42 2.36 2.30 2.25 2.21 2.18
7.82 5.61 4.72 4.22 3.90 3.67 3.50 3.36 3.26 3.17 3.09 3.03
1.38 1.46 1.45 1.44 1.42 1.41 1.40 1.39 1.37 1.37 1.36 1.35
26 2.91 2.52 2.31 2.17 2.08 2.01 1.96 1.92 1.88 1.86 1.84 1.81
4.23 3.37 2.98 2.74 2.59 2.47 2.39 2.32 2.27 2.22 2.18 2.15
7.72 5.53 4.64 4.14 3.82 3.59 3.42 3.29 3.18 3.09 3.02 2.96
1.38 1.46 1.45 1.43 1.41 1.40 1.39 1.38 1.37 1.36 1.35 1.34
28 2.89 2.50 2.29 2.16 2.06 2.00 1.94 1.90 1.87 1.84 1.81 1.79
4.20 3.34 2.95 2.71 2.56 2.45 2.36 2.29 2.24 2.19 2.15 2.12
7.64 5.45 4.57 4.07 3.75 3.53 3.36 3.23 3.12 3.03 2.96 2.90
 Statistical tables 387

Table D {Continued)

df for numerator
1 —a df for
15 20 24 30 40 50 60 100 120 200 500 00 denom.

1.46 1.45 1.44 1.43 1.42 1.42 1.42 1.41 1.41 1.40 1.40 1.40 .75
2.05 2.01 1.98 1.96 1.93 1.92 1.90 1.88 1.88 1.86 1.85 1.85 .90 13
2.53 2.46 2.42 2.38 2.34 2.31 2.30 2.26 2.25 2.23 2.22 2.21 .95
3.82 3.66 3.59 3.51 3.43 3.38 3.34 3.27 3.25 3.22 3.19 3.17 .99

1.44 1.43 1.42 1.41 1.41 1.40 1.40 1.39 1.39 1.39 1.38 1.38 .75
2.01 1.96 1.94 1.91 1.89 1.87 1.86 1.83 1.83 1.82 1.80 1.80 .90
2.46 2.39 2.35 2.31 2.27 2.24 2.22 2.19 2.18 2.16 2.14 2.13 .95 14
3.66 3.51 3.43 3.35 3.27 3.22 3.18 3.11 3.09 3.06 3.03 3.00 .99

1.43 !.41 1.41 1.40 1.39 1.39 1.38 1.38 1.37 1.37 1.36 1.36 .75
1.97 1.92 1.90 1.87 1.85 1.83 1.82 1.79 1.79 1.77 1.76 1.76 .90 15
2.40 2.33 2.29 2.25 2.20 2.18 2.16 2.12 2.11 2.10 2.08 2.07 .95
3.52 3.37 3.29 3.21 3.13 3.08 3.05 2.98 2.96 2.92 2.89 2.87 .99

1.41 1.40 1.39 1.38 1.37 1.37 1.36 1.36 1.35 1.35 1.34 1.34 .75
1.94 1.89 1.87 1.84 1.81 1.79 1.78 1.76 1.75 1.74 1.73 1.72 .90 16
2.35 2.28 2.24 2.19 2.15 2.12 2.11 2.07 2.06 2.04 2.02 2.01 .95
3.41 3.26 3.18 3.10 3.02 2.97 2.93 2.86 2.84 2.81 2.78 2.75 .99

1.40 1.39 1.38 1.37 1.36 1.35 1.35 1.34 1.34 1.34 1.33 1.33 .75
1.91 1.86 1.84 1.81 1.78 1.76 1.75 1.73 1.72 1.71 1.69 1.69 .90 17
2.31 2.23 2.19 2.15 2.10 2.08 2.06 2.02 2.01 1.99 1.97 1.96 .95
3.31 3.16 3.08 3.00 2.92 2.87 2.83 2.76 2.75 2.71 2.68 2.65 .99

1.39 1.38 1.37 1.36 1.35 1.34 1.34 1.33 1.33 1.32 1.32 1.32 .75
1.89 1.84 1.81 1.78 1.75 1.74 1.72 1.70 1.69 1.68 1.67 1.66 .90 18
2.27 2.19 2.15 2.11 2.06 2.04 2.02 1.98 1.97 1.95 1.93 1.92 .95
3.23 3.08 3.00 2.92 2.84 2.78 2.75 2.68 2.66 2.62 2.59 2.57 .99

1.38 1.37 1.36 1.35 1.34 1.33 1.33 1.32 1.32 1.31 1.31 1.30 .75
1.86 1.81 1.79 1.76 1.73 1.71 1.70 1.67 1.67 1.65 1.64 1.63 .90 19
2.23 2.16 2.11 2.07 2.03 2.00 1.98 1.94 1.93 1.91 1.89 1.88 .95
3.15 3.00 2.92 2.84 2.76 2.71 2.67 2.60 2.58 2.55 2.51 2.49 .99

1.37 1.36 1.35 1.34 1.33 1.33 1.32 1.31 1.31 1.30 1.30 1.29 .75
1.84 1.79 1.77 1.74 1.71 1.69 1.68 1.65 1.64 1.63 1.62 1.61 .90 20
2.20 2.12 2.08 2.04 1.99 1.97 1.95 1.91 1.90 1.88 1.86 1.84 .95
3.09 2.94 2.86 2.78 2.69 2.64 2.61 2.54 2.52 2.48 2.44 2.42 .99
/
1.36 1.34 1.33 1.32 1.31 1.31 1.30 1.30 1.30 1.29 1.29 1.28 .75
1.81 1.76 1.73 1.70 1.67 1.65 1.64 1.61 1.60 1.59 1.58 1.57 .90 22
2.15 2.07 2.03 1.98 1.94 1.91 1.89 1.85 1.84 1.82 1.80 1.78 .95
2.98 2.83 175 2.67 2.58 2.53 2.50 2.42 2.40 2.36 2.33 2.31 .99

1.35 1.33 1.32 1.31 1.30 1.29 1.29 1.28 1.28 1.27 1.27 1.26 .75
1.78 1.73 t70 1.67 1.64 1.62 1.61 1.58 1.57 1.56 1.54 1.53 .90 24
2.11 2.03 1.98 1.94 1.89 1.86 1.84 1.80 1.79 1.77 1.75 1.73 .95
2.89 2.74 2.66 2.58 2.49 2.44 2.40 2.33 2.31 2.27 2.24 2.21 .99

1.34 1.32 1.31 1.30 1.29 1.28 1.28 1.26 1.26 1.26 1.25 1.25 .75
1.76 1.71 1.68 1.65 1.61 1.59 1.58 1.55 1.54 1.53 1.51 1.50 .90 26
2.07 1.99 l95 1.90 1.85 1.82 1.80 1.76 1.75 1.73 1.71 1.69 .95
2.81 2.66 2.58 2.50 2.42 2.36 2.33 2.25 2.23 2.19 2.16 2.13 .99
1
1.33 1.31 1.30 1.29 1.28 1.27 1.27 1.26 1.25 1.25 1.24 1.24 .75
1.74 1.69 1.66 1.63 1.59 1.57 1.56 1.53 1.52 1.50 1.49 1.48 .90 28
2.04 1.96 1.91 1.87 1.82 1.79 1.77 1.73 1.71 1.69 1.67 1.65 .95
2.75 2.60 2.52 2.44 2.35 2.30 2.26 2.19 2.17 2.13 2.09 2.06 .99
 388 Statistical tables

Table D {Continued)

df for numerator
df for 1 —a
denom. 1 2 3 4 5 6 7 8 9 10 11 12

.75 1.38 1.45 1.44 1.42 1.41 1.39 1.38 1.37 1.36 1.35 1.35 1.34
30 .90 2.88 2.49 2.28 2.14 2.05 1.98 1.93 1.88 1.85 1.82 1.79 1.77
.95 4.17 3.32 2.92 2.69 2.53 2.42 2.33 2.27 2.21 2.16 2.13 2.09
.99 7.56 5.39 4.51 4.02 3.70 3.47 3.30 3.17 3.07 2.98 2.91 2.84

.75 1.36 1.44 1.42 1.40 1.39 1.37 1.36 1.35 1.34 1.33 1.32 1.31
40 .90 2.84 2.44 2.23 2.09 2.00 1.93 1.87 1.83 1.79 1.76 1.73 1.71
.95 4.08 3.23 2.84 2.61 2.45 2.34 2.25 2.18 2.12 2.08 2.04 2.00
.99 7.31 5.18 4.31 3.83 3.51 3.29 3.12 2.99 2.89 2.80 2.73 2.66

.75 1.35 1.42 1.41 1.38 1.37 1.35 1.33 1.32 1.31 1.30 1.29 1.29
60 .90 2.79 2.39 2.18 2.04 1.95 1.87 1.82 1.77 1.74 1.71 1.68 1.66
.95 4.00 3.15 2.76 2.53 2.37 2.25 2.17 2.10 2.04 1.99 1.95 1.92
.99 7.08 4.98 4.13 3.65 3.34 3.12 2.95 2.82 2.72 2.63 2.56 2.50

.75 1.34 1.40 1.39 1.37 1.35 1.33 1.31 1.30 1.29 1.28 1.27 1.26
120 .90 2.75 2.35 2.13 1.99 1.90 1.82 1.77 1.72 1.68 1.65 1.62 1.60
.95 3.92 3.07 2.68 2.45 2.29 2.17 2.09 2.02 1.96 1.91 1.87 1.83
.99 6.85 4.79 3.95 3.48 3.17 2.96 2.79 2.66 2.56 2.47 2.40 2.34

.75 1.33 1.39 1.38 1.36 1.34 1.32 1.31 1.29 1.28 1.27 1.26 1.25
200 .90 2.73 2.33 2.11 1.97 1.88 1.80 1.75 1.70 1.66 1.63 1.60 1.57
.95 3.89 3.04 2.65 2.42 2.26 2.14 2.06 1.98 1.93 1.88 1.84 1.80
.99 6.76 4.71 3.88 3.41 3.11 2.89 2.73 2.60 2.50 2.41 2.34 2.27

.75 1.32 1.39 1.37 1.35 1.33 1.31 1.29 1.28 1.27 1.25 1.24 1.24
00 .90 2.71 2.30 2.08 1.94 1.85 1.77 1.72 1.67 1.63 1.60 1.57 1.55
.95 3.84 3.00 2.60 2.37 2.21 2.10 2.01 1.94 1.88 1.83 1.79 1.75
.99 6.63 4.61 3.78 3.32 3.02 2.80 2.64 2.51 2.41 2.32 2.25 2.18
 Statistical tables 389

Table D* {Continued)

df for numerator
df for
15 20 24 30 40 50 60 100 120 200 500 00 1 — a denom.

1.32 1.30 1.29 1.28 1.27 1.26 1.26 1.25 1.24 1.24 1.23 1.23 .75
1.72 1.67 1.64 1.61 1.57 1.55 1.54 1.51 1.50 1.48 1.47 1.46 .90 30
2.01 1.93 1.89 1.84 1.79 1.76 1.74 1.70 1.68 1.66 1.64 1.62 .95
2.70 2.55 2.47 2.39 2.30 2.25 2.21 2.13 2.11 2.07 2.03 2.01 .99

1.30 1.28 1.26 1.25 1.24 1.23 1.22 1.21 1.21 1.20 1.19 1.19 .75
1.66 1.61 1.57 1.54 1.51 1.48 1.47 1.43 1.42 1.41 1.39 1.38 .90 40
1.92 1.84 1.79 1.74 1.69 1.66 1.64 1.59 1.58 1.55 1.53 1.51 .95
2.52 2.37 2.29 2.20 2.11 2.06 2.02 1.94 1.92 1.87 1.83 1.80 .99

1.27 1.25 1.24 1.22 1.21 1.20 1.19 1.17 1.17 1.16 1.15 1.15 .75
1.60 1.54 1.51 1.48 1.44 1.41 1.40 1.36 1.35 1.33 1.31 1.29 .90 60
1.84 1.75 1.70 1.65 1.59 1.56 1.53 1.48 1.47 1.44 1.41 1.39 .95
2.35 2.20 2.12 2.03 1.94 1.88 1.84 1.75 1.73 1.68 1.63 1.60 .99

1.24 1.22 1.21 1.19 1.18 1.17 1.16 1.14 1.13 1.12 1.11 1.10 .75
1.55 1.48 1.45 1.41 1.37 1.34 1.32 1.27 1.26 1.24 1.21 1.19 .90 120
1.75 1.66 1.61 1.55 1.50 1.46 1.43 1.37 1.35 1.32 1.28 1.25 .95
2.19 2.03 1.95 1.86 1.76 1.70 1.66 1.56 1.53 1.48 1.42 1.38 .99

1.23 1.21 1.20 1.18 1.16 1.14 1.12 1.11 1.10 1.09 1.08 1.06 .75
1.52 1.46 1.42 1.38 1.34 1.31 1.28 1.24 1.22 1.20 1.17 1.14 .90 200
1.72 1.62 1.57 1.52 1.46 1.41 1.39 1.32 1.29 1.26 1.22 1.19 .95
2.13 1.97 1.89 1.79 1.69 1.63 1.58 1.48 1.44 1.39 1.33 1.28 .99

1.22 1.19 1.18 1.16 1.14 1.13 1.12 1.09 1.08 1.07 1.04 1.00 .75
1.49 1.42 1.38 1.34 1.30 1.26 1.24 1.18 1.17 1.13 1.08 1.00 .90 00
1.67 1.57 1.52 1.46 1.39 1.35 1.32 1.24 1.22 1.17 1.11 1.00 .95
2.04 1.88 1.79 1.70 1.59 1.52 1.47 1.36 1.32 1.25 1.15 1.00 .99

* Abridged from E. S. Pearson and H. Hartley (Eds.), Biometrika Tables for Statisticians,
vol. 1 (2nd Ed.). New York; Cambridge University Press (1958), Table 18. Reproduced by
permission of E. S. Pearson and the trustees of Biometrika.
 r- Tj- m ON NO r~- to to ,—1 Y—1 to NO O ON lo T—1 ON OO NO ■tt to
o NO r- q (N to y—i q NO ■tj- to fN t-H O ON ON OO q r- q to q q q
NO q td id id
(N ON od K K NO NO N6 NO NO NO NO td td td td td id td id id to
ON
T™^ ON
fj
.j r- (N ■ri- OO O OO l-H to NO 00 ON O NO to to <N o ON r- ON
ON
00 m l-H q to q OO to t;}- (N i-H O ON OO r' q
q q q Tj- to (N q ON to
6
od NO' q ON od t--' r-' NO NO’ NO' NO NO’ NO td td td td id id id td td td td id q .
IT)
"o ci
V) >
NO OO m to m ro .-Htt o ON o fN to ON Tj- ON lO O 00 to q to
OC q ON O q q q to to q o o ON OO q q q NO q q to q q q q _E
od NO' d ON K r-' NO NO NO NO NO NO id id td id id id td td id td td id Tj- 4—»
r;. •r^ TH c/3
<U

(N NO Tj- to to <N to NO ON to OO to ON NO to <N T-H O ON s o

n
r- (N OO q q m q q q q O q OO q q q q q lo q q q q OO u
r-’ NO d od r^' NO NO NO NO NO NO td id td td td id td id td id td id td q o
VN
-C
ON o to Tj- r-- NO o NO to NO ON T—1
(N NO to O OO NO to q o
Sj \o q NO q q q q q CO q q q OO q q q q to lo q to q q q q q GX) c
>/d td id td id id td q q’
c *^
Jr NO
»r)
d od NO NO NO NO NO td td id id td id td id 03
Di: B
o
. •TN (N to OO OO (N OO OO ON (N to ON to o NO to (N T—lO o o T3 TO
-rf q to q q q q <N q q OO q q q to q lO q q to q q q q q <U o
V >o to d od r^' NO NO NO NO td td •o’ id td td id td td id td td td td q’ •
N <u
,1-'
>0 '—1 '-H c
IHH

!U C*H
OO to (N (N ■ttto ON ON to O o (N to ON NO to to to Tj- q T3 o
q
i
Tj- m ro ro to to q NO <o q q q q q q q to ■q ro cq <N q q oq q t/5
(U
VN d od fd NO NO NO NO td td td id id id td td id td id td td Tj- Tj-’ q’ (U
\
a> GO
'CS x: (U
] OO NO r-- ON (N ON ON to to NO ON ■tj- ON to tN 00 00 OO OO 00 OO OO TD
m (N q q m ON to q q q OO q q q ■q to to to q q q q oq q NO
V o o
rd td o od K NO NO NO NO td td td id id id td id td id id id id q on
X
+-*
>o
_C _C/5

? (N 0\
to to to ON (N OO
q ON q q q tt q
OO
q
to
OO q q
to
to
NO o to
q
to O O O
to CO <N q <N q
o
q oq q q
(N 'o
s;
<N
, PH
■'^" ON od r^' NO' NO' NO to td td td td td td id td td id id td td •rf •'tf q <u ,

* lo qj
* _>
r- t-- q ■rt lO 3
ON (N <o o to ON to tN to NO NO (N (N co g o
NO m q q q (N q OO q q q q to to q q q q q q OO q q to 0) >
d ON od NO' NO NO to’ td td td id id id td td id id lo id q ■^’ q'
C_)
_c
m (U
OH (/5
Vo
NO m m ON
ON to (N o ■tt ON O (N to O to TO <N to Tj- to NO i-,
tu
q 00 ON q q q q to q q q q q q q q to q 0)
10

q OO GO
"H*
ON rd ON r-‘ NO NO
a
NO to to' td td td td td td id td id id td TJ-' q q’ a
c
*
VH
D
0)
d) OO o <N ON o NO to ON to OO CO ON NO <N O <N to to ON <u .O </3
bfi q to q q m q q q q to q q q q q q q oq q q q q to x; • •»«» o
■'tf
ON »
C r-‘ rd ON r-‘ NO NO' to’ to’ to’ td td td id td td td ''ti- q q q' o:) X
03
o CHH
o
c/5
cn to to <N rs| O o to O O CN to ON ■tt" O ND to ON OO o <N NO ON (U
c/5 c/5

<u OO
q q m to OO q q to q q q q q q 00 oq q q q q q q cq q 3
<U
(U
c
o3
_N iri rd od r~' NO NO to’ to id td td to td ■t;}- -td T}-' q’ 03 <u
-4->
c
H
C/5
P
VH
s
T3
d) m t^- NO OO ON ON o ■ttt (N to to OO to OO o (N NO ON q r- 0) (U
(U
T3
r- ’'t 'tt o (N q to q q q q OO OO q q q q q q lO q to cq q HH
ax:
O
3 fd <N’ od r-’ NO to' to' to td id id ■tj-’ q ■tj- ■tt ■rj- "tt ''tf ti- q’ o p
GO c*-
u. o c3
O X
o m ro ON m to (N ^-H (N to ON ■'tj" ON NO tN ON to r-. o to o c (L>
>o q o q q q CO q q q OO q q q to to to q q q to q q
NO to
q U _o u.
O) NO to’ td ■';}■ ■rj- ■rt- TO
d »-H od NO to to’ to’ q q "tt’ q' c c/5 3
o3 'H
c/5
ON .-H
■td- NO ON NO NO OO NO _
ooNO ■ttj- ^H ttl- OO (N NO T3 c 1.-
OH
(N 00 to CO q ro q OO q q to' to q to to to q q q q ON q OO
<U (U +->
TD a
-(->

C
r-' d r-‘ NO to to to ■'ti- ■tj- 'tt ■'tj- ■tti-' rj- rd rd rd c c
d)
m (U 03
d) X
X) 3
o ro NO OO Q OO to tN to NO o to OO to tN O OO NO o Tf ON ON CO UJ 'T3 cr
d)
OO OO OO q q q to q to q q q q q q q q q q q q q q (U CHH
OH
to ■rj- Tt- q •r}- O O
(N
fO
ON NO to' "tt rd rd rd rd rd td rd td 3
»o >T TO VH
o3 o (U
X
d) OO OO o NO to OO (N r-- to o r-- to tN ON OO to OO o NO tN S SH a c
cx q q q q q q q q q q q q q q q q c
26.7

OO to ro OO OO to to to
a r<^ od to* to’ rd rd rd rd rd rd rd rd rd td rd rd rd rd rd rd rd rd
(U 3
0^ C
D
O
rd
ON o <o NO NO O to OO NO to O OO NO to (N ON NO to o ON X
p q to q q q ro q q q q q p q q. ON q ON ON q OO oq OO r- £ t/5
(N 7
U OO NO •rt- rd ro CO rd rd rd rd to rd rd rd rd tN (N IN tN tN (N tN tN (N tN o CN
o ON
mm iJs
*
<N
1—i (N m to NO r- OO ON O <N to q NO OO ON O O O O o o * *
a tN (N to •t;!- NO tN o a.
' '
H \ CL.

390
c
 CO VO CO CM VO CO VO Ov VO VO VO VO oo CM T-H T-H T—H CM CO VO
Ov oo q VO VO O q q Ov c- VO CO CM T-H o GO oo OO VO CM q 00 q

20
00 r~' Ov T-H d Ov Ov od od r^’ cd cd cd cd cd cd vd vd vd vd vd vd vd vd vd
o^ m T-H T-H
(N

T-H CO VO ov VO oo VO oo CO o OV O VO VO VO C' oo ov
CM OO VO q T-H oo VO CO CM o O oo oo c- q CO T—H q q VO
K Ov T-H d Ov od od od u cd cd cd cd cd cd vd vd vd vd vd vd vd vd vd
Os T-H T-H

p is the number of quantities (for example, means) whose range is involved, tiz is the degrees of freedom in the error estimate.
(N

oo CM VO VO r-- T-H ov CM c- CO VO oo T-H CM CO VO C-

O m q CO q o oo q CM O OO oo c- c^ VO CO T-H q q VO
00
O r'* Ov T—T d ov od od od u C^’ cd cd C^’ cd vd vd vd vd vd vd vd vd vd vd
o^ T-H
(N

VO VO VO CM ov CO CM o C- c- c- oo CM VO VO VO Ov T-H
IT) ov VO CM CO q CO ov VO CO CM o q oo C' C^ VO CM q OO q q
'O vd Ov cd T—i d Ov od Ov U c-’ cd cd cd cd vd vd vd vd vd vd vd vd vd vd vd
oo m T-H T-H T-H T-H
(N

o oo Tt VO CO VO r-- CM o o o CM VO oo oo o CM Tj- VO Ov
q q q o CM q q Ov VO CM T-< o Ov oo C' VO VO CO CM O q q
vd od cd T-H d ov od od U K cd cd cd cd vd vd vd vd vd vd vd vd vd vd vd
00 T-H T-H
fN

VO CM VO CO VO VO Ov VO CO CM CO VO oo CM CO VO ov VO
m VO CM ov T—< q q oo VO CO T-H o Ov OO c- VO VO VO CO oo q
r-- lO oo' cd T-H Ov Ov od od c-’ u cd cd cd vd vd vd vd vd vd vd vd vd vd vd vd
(N m T-H
rsl

oo T-H o CO VO VO o VO VO oo T-H VO VO 00 o CO VO O
q <N CO q q o q; O c~ CM T—1 Ov oo VO VO VO TC CM o q q
oo’ cd HH Ov Ov od od u C-’ cd cd .vd vd vd vd vd vd vd vd vd vd vd vd vd
r-- ro T-H
(N

ov VO VO T-H o VO c- c- VO VO c- o CO C' Ov T-H VO

ov oo q q q ov VO CO T-H o oo c- VO VO VO CO o oo q VO CO
ro
MO cd d Ov od od C"’ C-’ c-’ cd cd vd vd vd vd vd vd vd vd vd vd vd vd vd
MO T-H
(N

o Ov T-H oo oo oo VO VO o c- VO VO 00 oo CO c- O Ov
lO oo q c- CM O Ov C' VO VO TC CO CM 1-H q q q Tj- q
O fd rd CM’ d ov od od K C-’ K cd vd vd vd vd vd vd vo’ vd vd vd vd vd vd vd
MD m T-H T-H T-H
CM
oo o VO CO VO VO CO ov VO V) VO oo VO oo CM VO oo CO oo CO
VO q CO VO O VO CO T—1 q c- VO q CO CO CM O q q VO CO CM
m r4 r--’ CM’ d ov od od c-’ u cd vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd
IT) T-H
CM

O c- c- ov T-H ov T-H n- VO c- o oo CM VO o VO O VO
r-; r- CO CM CO 00 CM Ov 00 VO VO CO CM CM o OO c- VO CO q
o
VO 1-H ■sd CM’ d Ov od C-’ u r-‘ vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd
T-H
Table E.2 Upper 1 Percent Points of Studentized Range q

(N

r- 00 CM VO c- CO T-H CM VO oo CM C" VO o VO 00
CM q Ov oo T-H VO CO O oo q VO CO CM T-H o o oo oo q q CO q o
ON
r- o vd HH Ov od od C-’ vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd
m cn r—( T-H
(N

c- CO c- c- VO VO oo Tl- oo TC oo Ov VO CM Ov
m VO q VO q Ov Tj- q 00 MD VO CO CM T-H o o q q oo q VO CO CM T-H q
00
Ov vd T-H Ov od r--’ U vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd
CM <N T-H
CM

CM CM 00 00 CM ov oo Ov CM VO Oo CO oo o CO oo
CM q T—^ CO CO VO q ov q q o q Ov q q q q VO ■q q T-H q q
VO oo’ vd T—T ov od U U vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd
<N r-H T-H
CM

T-H c- VO VO CO VO o oo 00 O CM VO o VO C' Tj- Ov r-- VO

VO CM VO q ov CO OV VO CM Ov q oo q q q VO V) CO CM T^ q q q
\o vd vd vd vd vd vd vd vd vd vd vd V) vd vd vd vd Tj-’
CM vd d od r-’ u
O CM T-H
(N

VO CM VO CO VO c- CO CO VO Ov CO oo CO oo C- VO CO CM T-H o
t-- CO ov q O VO CO q q q q VO CO q q q q q q q
VO cd OV od U U vd vd vd vd vd vd vd vd vd vd vd vd vd vd vd
oo CM

C' o CO o VO c- CM O o CM VO Ov TC oo VO CM O o O o o
m CM T-H q O VO CM q q VO VO q CO q T-H T-H q o O q q q q q •q
CM CM ov u u vd vd vd vd vd vd vd vd vd vd vd vd vd vd ■rr
VO CM
T—(

CM c- CO CM CO CO c- 'Tj- VO oo CO oo TC O C- TC VO oo O CM
O VO T-H ov CO Ov q q T-H q q q q q q q q q VO q q q T-H
rn
V-) OV d oo’ vd vd vd vd vd vd vd vd Tf
tn

VO o VO O oo Ov CM VO T-H c- CO o V) CM VO Ov CM VO o
o q CM VO r- CM q q q ■q q CO q CM q q q q q q q q q q
(N
o 00 vd vd vd Tj- Tf Tj-’ cd cd cd cd cd cd
OV

T-H rN CO VO VO 00 ov o CM CO VO VO c~ oo oo o Tt- O O O O Q
N T-H T-H T-H T-H T-H T-H T-H TT—1 —H T-H CM CM CO VO CM o

391
392 Statistical tables

Table F Coefficients of Orthogonal Polynomials

k Polynomial 1 2 3 4 5 6 1 8 9 10 Sc,? 2

3 Linear -1 0 1 2 1
Quadratic 1 -2 1 6 3
Linear -3 -1 1 3 20 2
4 Quadratic 1 -1 -1 1 4 1
1 0
Cubic -1 3 -3 1 20 3

Linear -2 -1 0 1 2 10 1
5 Quadratic 2 -1 -2 -1 2 14 1
5
Cubic -1 2 0 -2 1 10 6
3 5
Quartic 1 -4 6 -4 1 70 12

Linear -5 -3 -1 1 3 5 70 2
3
6 Quadratic 5 -1 -4 -4 -1 5 84 2
5
Cubic -5 7 4 -4 -7 5 180 3
7
Quartic 1 -3 2 2 -3 1 28 12

Linear -3 -2 -1 0 1 2 3 28 1
7 Quadratic 5 0 -3 -4 -3 0 5 84 1
Cubic -1 1 1 0 -1 -1 1 6 1
6
Quartic 3 -7 1 6 1 -7 3 154 12

Linear -7 -5 -3 -1 1 3 5 7 168 2
Quadratic 7 1 -3 -5 -5 -3 1 7 168 1
8 Cubic -7 5 7 3 -3 -7 -5 7 264 2
3
Quartic 7 -13 -3 9 9 -3 -13 7 616 7
12
Quintic -7 23 -17 -15 15 17 -23 7 2184 7
1 0

Linear -4 -3 -2 -1 0 1 2 3 4 60 1
Quadratic 28 7 -8 -17 -20 -17 -8 7 28 2772 3
9 Cubic -14 7 13 9 0 -9 -13 -7 14 990 5
6
Quartic 14 -21 -11 9 18 9 -11 -21 14 2002 7
12
Quintic -4 11 -4 -9 0 9 4 -11 4 468 3
20

Linear -9 -7 -5 -3 -1 1 3 5 7 9 330 2
Quadratic 6 2 -1 -3 -4 -4 -3 -1 2 6 132 1
2
10 Cubic -42 14 35 31 12 -12 -31 -35 -14 42 8580 5
3
Quartic 18 -22 -17 3 18 18 3 -17 -22 18 2860 5
12
Quintic -6 14 -1 -11 -6 6 11 1 -14 6 780 1
1 0
 Answers to Odd-Numbered Problems

CHAPTER 2
2.1 Y = 18,472.9, = 41.7.
2.3 For a two-sided alternative, some points are
fi = 18,473; 18,472; 18,471; 18,470; 18,469; 18,468; 18,467
= = 0.08, 0.39, 0.80, 0.95, 0.80, 0.39, 0.08.
2.5 n = 1.
2.7 Do not reject hypothesis, since = 17.05 with 11 df.
2.9 Do not reject hypothesis, since F = 1.8.
2.11 Reject hypothesis at 1 percent level, since \t\ = 13.4.
2.13 1. Reject hypothesis, since z = 4.0.
2. Do not reject hypothesis, since z = 1.71.
3. Do not reject hypothesis z, since z < 1.
2.15 For = 13 15 17 19
Power (1 - 13) = 0.04 0.27 0.70 0.95.
2.17 Reject hypothesis at a = 0.05, since t = 8.32.
2.19 Reject hypothesis of equal variances at a = 0.05, since F = 17.4. Do not reject
hypothesis of equal means at a = 0.05, since t' is <1.
2.21 Reject hypothesis at a = 0.05, since t = 2.68.
2.23 Do not reject hypothesis at a = 0.05, since t = 1.62.
2.25 Reject hypothesis at a = 0.05, since z = — 1.96.

CHAPTER 3
3.1

Source df SS MS

Between A levels 4 253.04 63.26

Error 20 76.80 3.84

Totals 24 329.84

Significant at the 1 percent level.-

393
394 Answers to odd-numbered problems

3.3 Two such contrasts might be

SS
= TA - Tc = 31 60.06
C2 = T^- 2T, + TC^ -79 130.02
190.08

Neither is significant at the 5 percent level.

3.5
B A C
Yj = 25.4 22.4 18.5

None significantly different.

3.7 Two sets and their sums of squares might be

SS
Set 1 : Cl = 2Ti — 275 =
-28 49.0
C2 = 4T2 — 674 =
-108 48.6
C3 = llTi + IIT2 - 1473 + 1174 + 1175 =
71 1.3
C4 = lOTi 4T2
— —
4T4 + 1075 8 0.1
99.0
SS
Set 2 : Cl = 6T1 2T2
— =
-18 3.4
C2 = llTi + 1172 - 873 =
-237 33.6
C3 = 4Ti + 472 + 473 — 1974 = -252 36.3
C4 - 2Ti + 272 + 273 + 2T4 — 2375 = -172 25.7
99.0

3.9 75.6 percent idue tc1 levels of A, 24.4 ]percent due to error.
3.11 Two might be C 1 —^ C.i -- T 2 and C2 = 27.1 - T.2 - C3 and C
which is less than = 60.65, so is not significant at 5 percent, and C2 = 7,
which is less than = 105.07, so is not significant at 5 percent.

3.13

Source df SS MS 7

Between temperatures 4 1268.54 317.14 70.3***

Error 25 112.83 4.51

*** Significant at the .1 percent level.

3.15

Source df SS MS F

Between coatings 3 1135.0 378.3 29 g***

Error 16 203.2 12.7
 Answers to odd-numbered problems 395

3.17 Newman-Keuls tests show:

Coating IV III II I
Means 42.0 43.6 57.2 58.4

3.19 One case based on n — 1 df and the other on Newman-Keuls. The latter is
better if variances are homogeneous.
3.21 One set of orthogonal contrasts might be
Cl = 4T, - T, - 7,3 - - T,
C2 = Ti - T3

C3 = T4 - T,
C4 = Ti + T3 - - Ts
3.23 For in Problem 3.21, = 24.5.
3.25 Litter to litter differences account for 28.6 percent of the total variance.

CHAPTER 4

4.1

Source df SS MS

Between coater types 3 1.53 0.51*

Between days 2 0.21 0.10

Error 6 0.54 0.09

Totals 11 2.28

* Significant at the 5 percent level.

4.3
K A L M
Y_j\ 5.27 4.87 4.73 4.27

K and M are significantly different at the 5 percent level.

4.5 Four contrasts might be

= TM — TA Nonsignificant
C2 = TM — Significant at 5 percent by Schelfe
C3 = TM ~ Nonsignificant
C4 = — Tf^ Nonsignificant
4.7

Source df SS MS

Electrodes 4 4.2966 1.0742**

Positions 4 0.5696 0.1424
Strips 4 0.3136 0.0784
Error 12 0.8058 0.0672

** Significant at the 1 percent level.

4.9 Proof
396 Answers to odd-numbered problems

4.11 Error df = 0.
4.13 Three groups (treatments) and six lessons (blocks) ANOVA

Source df

Groups 2
Lessons 5
Error 10

Newman -Keuls on means:

Group C B A
Means 12.33 19.17 22.17

a. Y,j, = /r T iSj T fj T Df^ + ^ijk 3^i^d ANOVA

Source df

Scales 4
Times 4
Days 4
Error 12

b. ANOVA;

Source df

Scales 4
Days 4
Error 16

4.19 HQ. Fj ^ 0 for all j. Model: Yfj = Fj + + 8^. F = 8.75, which is signifi-
cant at the 1 percent level.
4.21 From —0.38 to 0.88.
4.23 For example,

Cl = T.i - T,
C2 = T.i + 7.2 - 27.3
C3 = F3 - 7.4

4.25

Source df SS MS F

Vendors 2 10.89 16.44 3.02

Scales 2 32.89 0.11 <1
Inspectors 2 0.22 0.11 <1
Error 2 0.22 5.44

Nothing significant at the 5 percent level.

 Answers to odd-numbered problems 397

CHAPTER 5

5.1

Source df SS MS F

Phosphors 2 1,244.44 622.22 8.96**

Glass 1 13,338.89 13,338.89 192.09***
P X G interaction 2 44.45 22.22 <1
Error 12 833.33 69.44

5.3 G better than

2 —lower current. For phosphors, Newman-Keuls:

Phosphor: CAB
Means: 253.33 260.00 273.33

so A or C is better than B.

5.5

Source df SS MS

Humidity 2 9.07 4.53

Temperature 2 8.66 4.33
T X i/interaction 4 6.07 1.52
Error 27 28.50 1.06

Totals 35 52.30

5.7 At the 5 percent significance level reject H and 2 but not

5.9 Plot of cell totals versus feed for the two material lines are not parallel (interaction),
and both material and feed effect are obvious.

5.11

Source df SS MS

■^/Temperature 2 600.09 300.04**

\ Mix 2 1.59 0.80*
,T X M 4 0.98. 0.25*
v^'Laboratory 3 3.85 1.28**
X L 6 1.54 0.26*
\/M X L 6 0.71 0.13*
T X M X L 12 0.95 0.07*
Error 36 0.67 0.02

Totals 71 610;38
398 Answers to odd-numbered problems

5.13

Source df SS MS

Soil 4 41.47 10.37

Fertilizer 2 68.87 34.44
Interaction 8 16.13 2.02
Error 15 52.50 3.50

5.15 Newman-Keuls on fertilizers:

Types: 3 2 1
Means: 3.6 5.2 7.3

Hence type 1 gives best yield with any of the five soil types.

5.17 a. No. b. Yes. c. Yes.

5.19 Two-dimensional stimulus more variable than three dimensional.

J by A means:

1.04 1.14 1.17 1.20 1.24 1.42 1.58 1.59 1.70 1.70

Source df SS MS

Thickness A 1 846.81 846.81**

Temperature B 1 5041.00 5041.00**
AB 1 509.63 509.63**
Drying condition C 1 5.88 5.88
AC 1 1.44 1.44
BC 1 15.21 15.21*
ABC 1 0.14 0.14
Length of wash L 3 69.76 23.25*
AL 3 15.79 5.26
BL 3 3.04 1.01
ABL 3 11.45 3.82
CL 3 9.65 3.22
ACL 3 7.57 2.69
BCL 3 6.43 2.14
ABCL 3 5.66 1.89
Error 32 87.21 2.73

Totals 63 6636.67
 Answers to odd-numbered problems 399

5.23 a.

Source df

Waxes 3
Times 2
W X T interaction 6
Error 12

b. Changes in gloss index due to the four waxes are different for the three polishing
times.

5.25 a. Significants are A, B, C, and the ABC interaction.

b. Run Newman-Keuls on 24 cell means
c. 4 X 3 X 2 layout with five observations per treatment.
d. 7.0 to 12.2.

CHAPTER 6

Source df SS MS
Factor A 1 0.08 0.08
Factor B 1 70.08 70.08**
A X B interaction 1 24.08 24.08
Error 8 36.68 4.58

Totals 11 130.92

Source df SS MS

A 1 2704.00 2704.00
B 1 26,732.25 26,732.25
AB 1 7744.00 7744.00
C 1 24,025.00 24,025.00
AC 1 16,256.25 16,256.25
BC 1 64,516.00 64,516.00
ABC 1 420.25 420.25
Error 8 246,284.00 30,785.50

Totals 15 388,681.75

Nothing significant at the 5-percent level.

6.5 No plots, since there are no significant effects at the 5 percent level.
400 Answers to odd-numbered problems

6.7

Source df SS MS

A 1 13,736.5
B 1 6188.3
AB 1 22,102.5 Same
C 1 22.7 as
AC 1 22,525.0 SS
BC 1 12,051.3
ABC 1 20,757.0
D 1 81,103.8
AD 1 145,665.0*
BD 1 9214.3
ABD 1 126,630.3*
CD 1 148.8
ACD 1 6757.0
BCD 1 294.0
ABCD 1 19,453.8
Error 16 431,599.4 26,975.0

Totals 31 918,249.7

6.9 a. (1), a, b, ab, c, ac, be, abc, d, ad, bd, abd, cd, acd, bed, abed.

b. Source df

Main effects 1 each for 4

Two-way interaction 1 each for 6
Three-way interaction 1 each for 4
Use as error with 5 df
Four-way interaction 1 each for 1

Total 15

c. Find A effect by plus signs where a is present in treatment combination and

minus signs where absent. Same for D and then multiply signs to get AD signs.
6.11 a. 32.

Source df

Main effects 1 each for 5

Two-way interaction 1 each for 10
Three-way interaction 1 each for 10
Four-way interaction 1 each for 5
Use as error.
Five-way interaction 1 each for 1

Total 31
 Answers to odd-numbered problems 401

c. Get signs for A, C, and E and multiply.

d. True if no interactions are present. Otherwise use Newman-Keuls on treatment
means.
6.13 Feed rate, 0.015; tool condition, new; tool type, precision.
6.15 2^ factorial. Probably use four-, five-, and six-way interactions as error term with
22 df.

6.17

Source df SS and MS

Drop-out temperature [A) 1 3.00

Back-zone temperature (B) 1 4.06*
A X B interaction 1 1.20
Atmosphere (C) 1 3.00
A X C interaction 1 1.71
B X C interaction 1 4.65*
A X B X C interaction 1 15.96**

Significant based on MS for error of 0.37 with 3 df.

6.19

Source df SS and MS

A 1 45.60*
B 1 1.36
AB 1 53.56*
C 1 0.01
AC 1 0.00
BC 1 0.01
ABC 1 0.04

A and AB significant based on MS for error of 4.41 with 3 df.

Newman-Keuls on/4, 5 means: 0.00 4.20 4.35 9.95.

No significant gaps as error is too large.

6.21 Plot the following: (1) = 0, a = —3,h— —2,ab = 5,c— —l,ac = 0,bc= 1,
and abc = — 2.

CHAPTER 7

7.1 Y'^ = 6.57 + 0.50A^.

7.3 ri^ = 0.5981, r" = 0.5023, Sy.;, = 0.756.
7.5 Scattergram indicates quadratic.
7.7 y; = 144.65 -H 3.22 -f 1.23^^ - 0.46^^ - 1.05w^.
402 Answers to odd-numbered problems

7.9 a. Second degree.

b. y;, = 96.19 + 2.69u + 1.13M^

7.11 = 0.22, rj^ 0.27.

7.13

Source df SS

Between ages 4 82.28

Linear 1 72.25**
Quadratic 1 9.78*
MS.rror = 0.39
Cubic 1 0.25
Quartic 1 0.00
Error 45 17.50

Curve is Y'^ = 3.91 + 0.85M — 0.264M^ and R^ ■ = 0.82.

7.15

Source df SS MS

Gate 2 10.58 5.29**

Resin 1 0.01 0.01
G X R interaction 2 0.00 0.00
Error 6 0.07 0.01

Linear only on gate setting giving: y[, = 2.517 + 1.15M Oi X = 0.217 + 0.575A,.

7.19

Source df Source df

R 1 G X W 4
G 2 Gz. X 1
GL 1 GL X WQ 1
GQ 1 GQ X W^ 1
R X G 2 GQ X WQ 1
R X Gi^ 1 R X G X W 4
R X GQ 1 R X X ILL 1
W 2 R X GL X WQ 1
1 R X GQ X Wj^ 1
% 1 R X GQ X WQ 1
R X W 2 Error 18
R X 1
R X WQ 1 Total 35

7.21 Graph.
 Answers to odd-numbered problems 403

12?>

Source df Source df

Radius (R) 3 RL X FL 1
RL 1 RL X FQ 1
RQ 1 RL X Fc 1
Rc 1 RQ X FL 1
Force (F) 3 RQ X Fa 1
FL 1 RQ X Fc 1
FQ 1 Rc X FL 1
Fc 1 Rc X FQ 1
R X F interaction 9 Rc X Fc 1
Error 32

Total 47

7.25 Plot.

CHAPTER 8

8.1 a. 7 = 9.2875, Sy = 3.5292; = 1.5250, Si = 0.2252; X2 = 10.0125, 53 = 4.3066;

= -0.7940, ry2 = 0.5012, ri2 = -0.1904.
b. r = 28.2641 - I2.443IA1.
8.3 Y' = 68.78 + Y92Xi adequate, ryi = 0.34.
For both x’s: Y' = 64.69 + \.92Xi + O.OI27A2.
Note: ri2 = 0 as and X2 are taken orthogonal.
8.5 a. One with largest correlation with Y.
b. One with largest partial correlation with Y after first X is removed.
c. When no more partials are significant.
d. ^2,20 = 3.33 < 3.49 at 5 percent, so stop here.
8.7 a. Ry,i234 = 0.71.
b. X2 and X3 only.
c. ± 12.89 (two standard errors).
8.9 Adequate equation: Y' = —0.66 + 0.29Xi + 0.45X3, where = ethnic group
and X3 = interview score. Ry.13 = 0.65.
8.11 a. X14.
b. X\ = 156.18 + 0.89X4
c. rf 13 = 0.89.
d. S5 14 = 443.02.
8.13 Stop now and use equation

X5 = 260.26 + 0.71X14 + 81.74X5 — 27.22Xg + 0.43X402

- O.8OX9 - 79.01X15.
= 0.93.
404 Answers to odd-numbered problems

8.15 a. Scattergram looks like quadratic.

b. r = 52.40 + 9.03X - 0.22^^ and = 0.91.
c. Linear: Y' = 132.77 + 0.41A and = 0.12.

CHAPTER 9

Source df SS MS

Temperature 2 348.44 174.22

Humidity 2 333.77 166.88
T X i/interaction 4 358.23 89.56
Error 9 680.00 75.66

Totals 17 1720.44

None significant at the 5 percent level.

9.3 AB = 320.111
358.22
AB^ = 38.11J

9.5

Source df SS MS

V 2 2.79 1.40*
T 2 9.01 4.50**
V X T 4 1.99 0.50
Error 9 2.15

9.7 VT = 0.94 and VT^ = 1.04, each with 2 df.

9.9 Any combination except TfjVfj.

9.11 After coding data by subtracting 2.6 and multiplying by 10, we have

Source df SS MS

Surface thickness A 2 2544.71 1272.35

Base thickness B 2 4787.37 2393.68
A X B interaction 4 185.18 46.29
Subbase thickness C 2 4165.15 2082.57
A X C interaction 4 189.72 47.48
B X C interaction 4 27.74 6.93
A X B X C interaction 8 100.06 12.51
Error 27 71.50 2.65
 Answers to odd-numbered problems 405

All except B x C are significant at the 5 percent level, but the main effects pre-
dominate.

9.13

Source df SS

AB 2 66.931
185.19
AB^ 2 118.26)
AC 2 58.931
189.75
AC^ 2 130.82)
BC 2 5.821
27.75
BC^ 2 21.93)
ABC 2 17.603
ABC^ 2 8.26 [
100.05
AB^C 2 53.48
AB^C^ 2 20.71 /

9.15 Plots show strong linear effects of A, B, and C. They show AB and AC inter-
action—but it is slight compared with the main effects. A slight quadratic trend
in A is also noted.

9.17

Source df SS MS

Resin type R 1 0.050 0.050*

Weight fraction S 2 0.540
SL 1 0.540 0.540**
SQ 1 0.000 0.000
Gate setting G 2 30.300
GL 1 30.150 30.150**
G<2 1 0.150 0.150*
RS 2 0.040 0.020
RG 2 0.060 0.030*
SG 4 0.343
SLGL 1 0.280 0.280**
SLGQ 1 0.002 0.002
SQGL 1 0.010 0.010
SQGQ 1 0.051 0.051*
RSG 4 0.050 0.013
Error 18 0.150 0.0083

Totals 35 31.530

9.19 y; = 3.20, y; = 3.02.

406 Answers to odd-numbered problems

CHAPTER 10

10.1

Source EMS

(Tg + lOo-J
+ 2a o A + 6^^-

OA,j (T^ +
Wg <— J ^
%ij)

Tests are indicated by arrows. None significant at the 5 percent level.

10.3

Source EMS

A + UCGIB + nbcG^
Bj + nca^B + nacai
AB,j + UCGIB
Q + ^^ABC + naaic + nba^c + nab(l)c
AC,, + ^^ABC + nbalc
BCjk + ^^ABC + naalc
+ ^^ABC

Tests are obvious.

No direct test on C.
10.5

Source EMS

A, (^E + nbcG^j) + nbda^c + nbcd(l>^

Bj + naalcD + nacGBD + nadaic + nacd(l)B
AB,j + f^^ABCD + ncalBD + nda^Bc + ncd(l>AB
2
Q 0-£ + nabGQB + nabdGc
2
AC„ CTE + nbalcD + nbda^c
2
BCjk (TE + naGBCD + nadGBc
2
ABCijk (7E + f1(^ABCD + nda^BC
2
D„ 0-£ + nabacD + nabcGB
2
AD,^ CTe + + nbcGAD
BDj^ 2
C^E + fJacricD + nacGBD
ABD,j„ 2
^E + nc^ABCD +
2
CDkm <^E + nabacD
ACD,km + nbalcD
BCDjkm + naGBCD
2
ABCDijkm ^E + ^^ABCD
^q(ijkm)

No direct test on A, B, or AB.

 Answers to odd-numbered problems

10.7 a. Yij,. _ /i + T,- + Sj + TSij +

b.

Source df EMS

Ti 3 + 2GJB T 10(^E\
Sj 4 (^8 + 8cr| )
TSu 12 <y^
20

c. See arrows ;above.

Source EMS

0. + +
2OQI 8(TQ

ij (^8 + +
2GQI 8(TJ

0/,, + 2^01

Instruments; 40 percent
Operators: 32 percent
Operator/
instrument interaction; 15 percent
Error; 13 percent

10.11 UQ = 4.44, Sg = 1.16, = 5.32.

10.15 To test A, MS = + MS^c - ^^ABC

B, MS = MS^B + MSBC - MS^BC
C, MS = + MS^c - MS^BC

CHAPTER 11

11.1

Source df SS MS EMS

Li 2 27.42 13.71 + 3crJ + 12^2.

^/(i) 9 36.38 4.04 erf + 3af
24 21.80 0.91

Totals 35 85.60

Difference between rolls within lots is significant at the 1 percent level.

408 Answers to odd-numbered problems

Source EMS

Ai (T^ + 2(TC + 6cr| + 24^^

^Ki) (7^ + 2GC + 6o-|
+ 2GC

(^8

Tests are obvious.

11.5

Source df SS MS EMS

Ti 1 5,489,354 5,489,354 G^ + 6(7^ + 12^4j-

^Ki) 2 408,250 204,125 G^ + 6GM

1 8971 8971 (^8 + 3aj^s + 12^5

TS,, 1 37,445 37,445 G^ + 3(7MS + ^</>TS

^^kKi) 2 31,520 15,760 G^ + 3(T^S

^m(ijk) 16 316,930 19,808

11.7

Source EMS

Ti + 18cr^ + 36
^Ki) + 18a^
s. + + 36
TSi, + +
^^kj(i) + 9cr^5
Pm + + 24 (/>p
TPim <y^ + ^^mp + l2(/)fp
+ ^^mp
SP,m (^8 +I 3a
^^nxsp + 12<^sp
'3^2
TSPij^^ + ^^msp "b ^^tsp
'3—2
+ -^^msp
^q(ijkm) (^8

11.9 a, b.

Source df EMS

c, 1 + 2ol + 8^2

Tj(i)
6 + 2aj
Vk 1 + ajv + 4ocv
CV,k 1 + Gpy + 4OQV
FKm 6 (^1 + olv

c. 60 percent.
 Answers to odd-numbered problems 409

11.11 Newman-Keuls shows experimental posttest group means better than control
and one experimental group better than the other.

11.13 Newman-Keuls gives; 54.66 54.96 57.00 65.66

So posttest experimental is best.

11.15 Discuss based on EMS and results.

11.17 Gi ^ Cj + GCij + + 7"^ + GTi^ + CTp^

+ GCTij^ +

11.19

Source df EMS

Di 2 2cr| + llaj)
Gj 1 ^SG + llo-oG + 33^;^G
DG,j 2 ^SG + 11^DG
^k(i) 30 2GI

30 ^SG

11.21

Source df EMS

Ri 1 + 6cri + 36^1
Hjw 10 + 6aH
Pk 2 + 2(JpH + \2opp + 24</>p
RPiK 2 + 2(7pH + I2app
P^kKi) 20 + 2GPH
^m(ijk) 36

11.23

Source df EMS

F, 4 (^e + 9(T|+ 54-(pp

Cj 1 + 9ol + 135(/)c
FC, 4 + 9ol + 21
^k(ij) 20 (^s + 9GI
P
^ m 2 + 3GSP + 90(/)p
FP 8 3GSP + lS(j)pp
CP
^ ^ jm 2 + 3GSP + 450CF
FCP
^ ^ 1 jm 8 + 3GSP + ^4*FCP
^Pmk(ij) 40 + 3GSP -

^q(ijkm) 180
410 Answers to odd-numbered problems

11.25 Recommend 1:1 concentration and any of the fillers—iron filings, iron oxide, or
copper.

CHAPTER 12

12.1

Source df SS MS EMS

2 337.15 168.58 + 18o-|

Sj 1 16.66 16.66 + ll(t>s

RS,j 2 126.78 63.39 CTg + 9<7^5

2 15.59 7.80 + 6cr|jj + 18^jj
HSj, 2 80.12 40.06 + 3GSRH +
RH,, 4 62.85 15.71 + 6cr|^
SRH,j, 4 229.44 57.36 + ^G^RH
36 200.00 5.56

Totals 53 1068.59

Replications (blocks) and replications by all other factors are significant

at the 5 percent level; however, these are not the factors of chief interest.

12.3

Source df SS EMS

Ri 1 As before G^ + ^(^RM + 12GR

Tj 1 with R in G^ + ^(^RM + 6G^ + 6GRJ^ + 12</>R

RTij 1 place of S (^E + “f RT

^kU) 2 GE + ^GRf^ + 6G^

R^ik(J) 2 G^ + ^(^RM

^miijk) 16

Total 23

No direct test on types, but results look highly significant. No other signifi-
cant eflfects.

12.5

Source df EMS

Ri 2 lOai No test
4
RM,j 8 ^^RM No test
S, 1 5GIS + 150s\ Poor test
RStk 2 ^^RS No test
MS,, 4 '^RMS + 30MS)
2
8 ^RMS No test
 Answers to odd-numbered problems 411

Source df EMS

R, 2 30(7^
yj 4 6(TRy + 18(/)^\
R|/, 8 6GIV ^
5 Sals +
RSik 10 54s ^
VSj, 20 ^RVS + 3</)J/5\
RVS,j, 40 ^RVS

Source df EMS

^1 1 4 + 364
Pi 2 GI + 12(T4 + 24(/)p^ Poor test
RPij 2 + 124P
H, 5 G} + ^4//

RHi, 5 + ^^RH
PHj, 10 + ^^RPH + 4(/)p^'\
RPH,j, 10 + ^^RPH 2F
c 36
^m(ijk) <^c

12.11 Compare models.

12.13 Complete: Yijt„ = + R, + M, + RM,,, +
Reduced: = fi + R,, + M, +
Take fewer readings per head.
12.15 Complete: + M, + RMi^ + Gj + RGj^ + MG^j + RMG^^

+ ^kU) + ^T^qkU) + + ^m(ijkq)

Reduced: Omit all interactions with R.

CHAPTER 13

Source df EMS

Whole Ri 1 4 + 64s + 184

plot Hj 2 + 24HS + 44s + 64H + 124
RHij 2 + 24HS + 64ir

Split- 2 4 -f- 64s + 12cr|

plot RSik 2 + 64 s
ffSjk 4 _j_
1 24HS + 44s
RHSij, 4 4 + ^^RHS
*^m(ijk) 18

Total 35

13.3 a. Fill nine spaces in random order.

 Answers to odd-numbered problems

b. Choose one of the orifice sizes at random, randomize flow rate through this
orifice, then move to a second orifice size, and so on.

Source df

2
Oi 2
RO,j 4

Fk 2
RF,, 4
OFj, 4
ROF iji^ 8

d. Take more replications.

Source df EMS

Whole Ri 3 -f AOal
plot Sj 1 ol + 4cr|5^ + 16cr|^ + 20cr|s + 80 ^5
RS,j 3 + ^^IsA + 20cris
Jk 1 0^ + + 16(TJ^ + IOORJ + ^0(J)j

RJik 3 + ^^RJA + 20a|j

SJjk 1 (^1 + '^^RSJA + lOo-isj + + A0(t>sj
RSJij, 3 + ^^RSJA + lOcrisj

Split- 4 GI + 8cr|^ + 32crJ

plot RAm 12
4 GI + 4G^SA + 16crs^
RSAij,„ 12 + 4^7|S^
4 G^ + ^(^RjA +
RJAii^fjj 12 G^ + ^(^RJA
SJAji^ffj 4 GI + 2(7|SJ^ + SGSJA

RSJAiji^^ 12 GE + ^(^RSJA
2
^qiijkm) 80 f^E

Source df MS EMS

Whole Ri 2 60.46 (^E + 5(7^0

4- 20G^
plot Dj 4 276.19 f^E + CFIDO + 3(7^0 + 4cr^£) -f 12^^

RDij 8 40.63 (^E + CT^DO + ^^RD

Split- 0, 3 16.73 (^E + + 1 5(TQ

plot ROi, 6 57.44 f^E +
2 1 2
DOji^ 12 12.78 ^E + (^RDO + 3(TOO
RDO,j, 24 46.83 (^E + (^RDO

Total 59

Days are significant by an F' test.

 Answers to odd-numbered problems 413

13.9 In a nested experiment the levels of B, say, are different within each level of A.
In a split-plot experiment the same levels of B (in the split) are used at each level
of A.

13.11 Laboratory means: 1 3 2

11.22 11.25 12.53

Mix means: B A C
10.29 11.58 13.13

Laboratory x mix means all different in decreasing order: 165, 155, 145 with
B, A, and C.

13.13

Source EMS

Ri + 9GRL + 27cr|
j-j -f 9(7iL + 36O-^
RLij + 9^RL

Tk -f ^^RLT -h 12(7^J + 9G^Y + 36

RT,, + ^^RLT + 9G^T
TTjk + ^^RLT -f \2GIT
RTTijk + ^^RLT

Mm + ^^RLM + 9(7RM +
RMim + ^^RLM + 9ORM
TMjm + ^^RLM +
RLMijm + ^^RLM
TM,m + ^RLTM + ^^LTM + ^^RTM +
RT Mi,m + ^RLTM + ^^RTM
^2
LT Mj,m ol ^RLTM + ^^LTM
2
RTLMijkm CTE

13.15 a. Split-plot.

Source df EMS

R. r — 1 -|- 15<T^
Mj 2 (^e + ‘^^RM + IQ^A^
RMij 2(r - 1) + ‘^^RM

Tk 4 Gt + 3(TR7’
RT -I- 6^^
2
RT,, 4(r - 1) a; -I- 3(T RT
c • r > 4.
MTj, 8 G; + (j RMT + 2(/) MT]
RMT,j, 8(r - 1) G; G RMT
414 Answers to odd-numbered problems

13.17

Source df EMS

Ri 1 + 18(7^
Mj 2 +
RM,j 2

T, 5 + T
RTik 5
MTj, 10 + ^RMT +
RMT,j, 10 ^6 + ^RMT ^

13.19 Take another replication.

CHAPTER 14
14.1 Show. SSbiocks ~ 10.125.
14.3 replication I replication II replication III
confound AB confound AC confound BC

(1) a a (1) (1) b

ab b c ac be c
c ac ab b a ab
abc be be abc abc ac

14.5 BC.
14.7 SSbiocks ~ 29,153.5 — 884^ + + ^^BCD-
14.9 The principal block contains: (1), ab, ad, bd, c, abc, acd, bed.
14.11 One scheme is to confound ABC, CDE, and ABDE. The principal block contains
(1), ab, de, abde, ace, bee, acd, bed.

Source df

Main effects 5
Two-way interaction 10
Three-way interaction 8
Four-way interaction 4
Five-way interaction 1
Blocks (or ABC, CDE, ABDE) 3

Total 31

14.13 One scheme is to confound ABCD^. The principal block contains 0000, 1110,
1102, 1012, 2111, 2122, 0221, 0210, 0120, 0101, 1211, 1200, 2220, 0022, 0011,
0202, 0112, 1121, 1020, 1001, 1222, 2010, 2201, 2100, 2002, 2021, 2212.
14.15 Verify.
 Answers to odd-numbered problems 415

14.17 Confounded ABC, CDE, ADF, BEF, ABDE, BCDF, ACEF gives a principal
block of (1), acd, hce, abf, abde, bcdf, acef, def.

Source df

Main effects 1 each for 6

Two-way interaction 1 each for 15
Three-way interaction 1 each for 16
(except confounds)
Four-way interaction 1 each for 12
(except confounds)
Five-way interaction 1 each for 6
Six-way interaction 1 each for 1
Blocks (or confounds) 7

Total 63

14.19 Yates with seven columns—pool to get temperature and its interactions.
14.21 Confound TABC with blocks (mod 2), L = + Y2 + Y3 + Y4.
14.23 Principal blocks are

Replication I (1), ab, de, abde, bed, acd, bee, ace

Replication II (1), be, cd, bade, ace, abc, ade, abd
Replication III (1), be, abe, ace, ede, bde, abed, ad

Source df

Replication 2
Blocks in replication 9 11

Main effects 1 each for 5

Two-way interaction 1 each for 10
Three-way interaction 1 each for 10
Four-way interaction 1 each for 5
Five-way interaction 1 each for 1
Replications x all 53 84

Total 95

CHAPTER 15

15.1 Aliases:

Source df SS

A = C 1 30.25
B = ABC 1 2.25 No tests.
AB = BC 1 0.25
416 Answers to odd-numbered problems

15.3 Aliases are

A ^ BC = AB^C^
B = AB^C - AC
C = ABC^ = AB
AB^ = AC^ = BC^

15.5 A = BD,B = AD, C = ABCD, D = AB, AC = BCD, BC = ACD, CD = ABC.

15.7 A = AB^C^D = BCD^

B = AB^CD^ = ACD^
C = ABC^D^ = ABD^
D = ABC = ABCD
AB = ABC^D = CD^
AB^ = AC^D = BC^D
AC = AB^CD = BD^
AC^ = AB^D = BC^D^
AD = AB^C^ = BCD
BC = AB^C^D^ = AD^
BC^ = AB^D^ = AC^D^
BD = AB^C = ACD
CD = ABC^ = ABD

15.9

Source df

Main effects with 5- or 6-way 7

2- way interaction with 4- or 6-way 21
3- way interaction, and so on as error 35

Total 63

15.11

Source df SS MS

Heats (or 2-way) 6 51.92 8.65*

Treatments (or part of H) 1 12.04 12.04*
Positions (or 3-way) 2 0.58 0.29
PH (or 3-way) 121 ^^•^^|33 42 2 39
PT (or 2-way) A 3.59) ■

Totals 23 97.96
 Answers to odd-numbered problems 417

15.13 Combining three- and four-way interactions as error with 5 df gives

Source MS

A 12.44
B 53.40*
AB 0.00
C 33.44
AC 7.06
BC 0.00
D 31.22
AD 42.48*
BD 1.62
CD 0.07
Error 5.73

B and AD significant at 5 percent level.

15.15 a. Replication I: (1), ab, de, abde, bed, acd, bee, ace.
b. BC = DE = ABD = ACE.
c. Replications x three-way = 14 df.

15.17 a. ACD.

Source df MS

A (or CD) 1 5729.85*

B (or ABCD) 1 91.80
C (or AD) 1 4199.86*
D (or AC) 1 1217.71
AB, BC, BD, and
their aliases 3 281.88

Some bad aliases, better to confound ABCD.

15.19 BCDEF confounded.

Source df

Main (or four-way) 1 each for 5

Two-way or three-way 1 each for 10

No good error term unless one assumes no interactions.

CHAPTER 16

16.1 Before covariance: Eg ^QO ^ 5-0 and significant.

After covariance: Eg 99 = 16.5 and still significant.
418 Answers to odd-numbered problems

Adjusted
Source df Z xy Z^" Z>'^ df MS

Between lots 3 500 -620 790

Error 16 1182 428 406 251.02 15 16.73

Totals 19 1682 -192 1196 1174.08 18

923.06 3 307.69*

End of cycle Estimate of 5 Significance

2, block 1 0.60 None

2, block 2 0.64 All except AB
3, block 1 0.60
3, block 2 0.61 All

Source df MS

Machines M 1 0.0136
Lumber grade L 1 0.0003
ML 1 0.0066
Replications R 1 0.0001
MR 1 0.0028
LR 1 0.0021
MLR 1 0.0015

Source df SS MS

Days 3 34,292 11,431

Treatments (adjusted) 3 508 169
Error 5 5025 1005

Treatments not significant at 5 percent.

16.11 If seven were used,

Source df

Treatments 1
Blocks (men) 6
Error 6

Total 13
 }

* • '
I ■ . Index
M , I . . • •

\ , /

- I
I

•.- ' . '. ,1

«
< . •

f S

•4 ‘I

^ •

• *
> \

/ -^1

*♦ ,

r I-,. . .-if*

•> '.^W^va
 Index

a, probability of type I error, 20 Change in mean effect (CIM), 341, 345

AB^ and AB components of A x B in- Chi square
teraction, 195 factorial, 351-354
Aliases, 304—306 table, 382
Analysis, 2, 6 Cochran’s C test, 60
Analysis of variance (ANOVA), 37, 351 Coefficient of determination r^, 135
covariance analysis and, 319-321 Completely randomized design, 11, 67-
factorial experiments rationale, 98-103 76, 266, 319
fundamental equation, 40, 73 distinguished from split-plot design,
nested experiments rationale, 232-233 270
one-way, 37-^6, 50, 53, 54, 56, 57, Components of variance, 55-59
59, 61 Computer programs
randomized block design rationale, 72- factorial experiments, 96-97
73 homogeneity of variance, 60-61
rationale, 38-46 multiple regression, 167, 175-176
two-way, 68, 73 for nested or nested-factorial experi-
Anderson, 270 ments, 238-239
Attribute data, 318, 351-360 Newman-Keuls test (SNK), 53-54
philosophy, 351 for one-way ANOVA, 43-44, 50
problems with quantitative levels of a
P, probability of type II error, 20-23 factor, 144, 145
Barnett, E. H., 341 Concomitant variable X, 318
Bartlett-Box F test, 60 Confidence
Batson, H. C., 351, 355 interval, 19
Behrens-Fisher problem, 28 limits, 19, 25, 26, 54
Bennett, C. A., 210, 222 Confounding in blocks, 281-299
Binomial variable, 7 complete, 290-291
Blocks, 68 confounding systems, 282-285
See also Randomized block design partial, 291-294
Box, G.E.P., 339, 341 in 2^ factorials, 285-290
in 3^ factorials, 294—299
Canonical form, 340 Consistent statistic, 18
Central composite design, 340 Contrasts, 46-53, 114

421
422 Index

Correlation, 2 Expected value, 17-19, 58, 59, 218-220

coefficient (r), 135, 168, 177 Experiment
ratio R and R^, 139 defined, 1
Covariance analysis, 318-328 factorial, 3, 86-104, 112-124, 187-
philosophy, 318-319 206, 252-259, 265-276, 281-299
Critical region (of test statistic), 20, 21 general, 2-A
Curvilinear regression, 137-139 nested, 227-233, 238-244, 259
nested-factorial, 233-244, 259
Davies, O. L,, 25, 297 single-factor, 36-61, 67-80, 103-104,
Decision rule, 20, 21, 25 211-212, 224, 312-313, 319
Defining contrast, 283 Experimental research, 1, 164
Degrees of freedom, 73, 75, 77, 78 Ex-post-facto research, 2, 164
Dependent variable, 3
Design F distribution, 26-27, 38
completely randomized, 11, 67-76, table of, 383-388
266, 270, 319 F” test, 222-223
general, 4—6 pseudo-F, 221-223
Graeco-Latin square, 79-80 Factorial chi square, 351-354
incomplete block, 80, 355-360 table, 382
randomized block, 67-76 Factorial experiments, 3, 86-104, 112-
split-plot, 265-276 124
split-split plot, 273-276 advantages, 88
Youden square, 80, 360-362 analysis, 91-94
Developmental research, 2 ANOVA rationale, 98-103
computer program, 96-97
Effect of a factor, defined, 114 confounding in blocks, 281-299
Error defined, 88-89
restriction, 270 interpretations, 94—96
split-plot, 268 in Latin square design, 258-259
type I, 20 in randomized block design, 252-258
type II, 20-23 in split-plot design, 265-276
whole-plot, 268 2^, 112-119
Estimate 2^ 10, 120-123
interval, 17, 19 3^ 187-198
point, 17 3^ 198-206
Estimation, 17-19 Factors. See Levels of factors
Eta squared, 135 Fisher, R. A., 77, 140, 355
Evolutionary operation (EVOP), 318, Fixed model, 210-214, 220
341-351 Fractional factorial, 304—314
form rationale, 346-351 Fractional replication, 303-314
philosophy, 341 aliases, 304—306
Expected mean square (EMS), 56-59, 2\ 306-310
212-223 3^ 310-314
derivations, 218-220 Franklin, N. L., 210, 222
pseudo-F test, 221-223 Fundamental equation of analysis of
rules, 214—218 variance, 40, 73
 Index 423

Graeco-Latin square design, 79-80, 259 qualitative and quantitative, 3, 5, 6,

14^158
Hierarchical experiments, 3, 227 random and fixed, 36-37, 210-224
Historical research, 2 Linear regression, 129, 130-137
Homogeneity of variances tests, 60-61
Hypotheses McLean, 270
statistical, 19 Manipulation, 1, 2
tests of, 19-21, 25-27, 206 Mathematical model, 5, 6
Max-min-con principle, 6
I and J components of interaction, 194 Mean of population, 11, 16
Incomplete block design, 80, 355-360 of sample, 17
Independent variables, ways of handling, Mean square (MS), 18, 41
5-6 Means
Inference. See Statistical inference confidence limits on, 54
Interaction tests on, 27-30, 47, 51
A X B, AB components of, 195 Minimum variance, 18
factorial experiments, 88-90, 93-95, Minimum-variance estimate, 18
97, 99, 100, 102-104 Missing values, 73-76
factorial experiments in randomized Mixed model, 210, 211, 213, 214, 220,
block design, 253, 256, 258 223
high-order, increase in, 303 Model
I and J components of, 194 fixed, 210-214, 220
nested-factorial experiments, 234 mathematical, 5, 6
with quantitative and qualitative fac- mixed, 210, 211, 213, 214, 220, 223
tors, 149-151, 153, 154, 156 random, 55, 210-214, 220, 223
2^ factorial experiments, 113-116 Modulus of 3, 195
2^ factorial experiments, 120, 121 Multiple regression, 129, 164—177 , 331
3^ factorial experiments, 189-198 computer programs, 167, 175-176
3^ factorial experiments, 198-206
Intercorrelation matrix, 168 Nested experiments, 3, 227-233
Interval estimate, 17, 19 computer programs, 238-239
in randomized blocks or Latin square
J and I interaction, 194 design, 259
repeated-measures designs as, 239-244
Kempthome, O., 283, 294 Nested-factorial experiments, 233-238
Keuls. See Newman-Keuls computer programs, 238-239
in randomized blocks or Latin square
Latin square design, 76-80, 328, 360 designs, 259
factorial experiments in, 253, 258-259 repeated-measures designs as, 239-244
fractional replication and, 312, 313 Newman-Keuls range test, 51-54, 78,
Graeco-Latin squares, 79-80 94-95, 198, 231, 238
nested or nested-factorial experiments NID, 36, 37
in, 259 Normal curve, table of, 379-380
Youden square, 80, 360-362 Null hypothesis form, 5
Least squares method, 132, 164
Least squares normal equations, 132 Operating characteristic (OC) curve, 22-
Levels of factors 23
424 Index

Orthogonal contrasts, 47-51, 359 Regression, 2, 319, 321

Orthogonal polynomials, 129, 139-144, covariance analysis and, 319, 321,
159 324, 326, 328
coefficients of, table, 391 curvilinear, 137-139
Ostle, B., 319, 328 linear, 129-137
Owen, D. B., 25 multiple, 129, 164-177 , 331
orthogonal polynomials, 139-144
Repeated-measures designs, 71
Parameters, 16
as nested and nested-factorial experi-
Path of steepest ascent method, 330
ments, 239-244
Pearson product-moment correlation
Research
coefficient (r’s), 135, 168, 177
defined, 1
Plots, 266
types, 2
Point estimates, 17
Research hypothesis, 5
Population standard deviation, 16
Response-surface experimentation, 318,
Population variance, 16
328-340, 341
Power curve, 23
more complex surfaces, 338-340
Power of a test, 23
philosophy, 328-329
Principal block, 284
two-fold problem, 329-338
Probability, 15
Response variable Y, 3
Proportions, tests on, 30-32
Restriction error, 270
Pseudo-F test, 221-223
Risks of making incorrect decisions, 20
Rotatable design, 339
Quadratic regression, 137-139
Qualitative/Quantitative factors, 129,
144-159 Sample, random, 16
two factors, both quantitative, 152-158 Sample size, determining, 4, 23-25, 224
two factors, one qualitative, one quan- Sample standard deviation, 17
titative, 145-152 Sample statistics, 16, 19
Satterthwaite, 222
Random model, 55, 210-214, 220, 223 Scheffe’s test, 51, 52-54
Random order, 4—5 Single-factor experiments, 36-61, 312-
Random samples, 16 313
Random variable, average or expected compared with factorial experiments
value of, 16 with one observation per cell,
Randomization, 1, 2 103-104
of order of experimentation, 4—5 completely randomized, 36-59
Randomized block design covariance analysis and, 319
complete, 67-76 Latin square designs, 76-80
factorial experiments in, 252-258 levels of treatment, 211-212
incomplete, 80, 355-360 randomized block design, 67-76
nested or nested-factorial design in, sample size, 224
259 Snedecor, G. W., 319
Randomized blocks, 328 SNK, 53
Split-plot design, 265-276
 Index 425

Split-plot error, 268 Test statistics, 19

Split-split-plot design, 273-276 Treatment effect, 36, 211
SPSS (Statistical Package for the Social Type I error, 20
Sciences), 43, 44, 53 Type II error, 20-23
Standard deviation
of population, 16
Unbiased statistic, 17
of sample, 17
Uncertainty, 15
Standard error of estimate, 135
Statistical hypothesis, 19
Statistical inference, 16-32 Variance
application to tests on means, 27-30 components of, 55-59
application to tests on proportions, 30- of population, 16
32 of sample, 16, 17
application to tests on variances, 25-27
Variance ratio, 26
estimation, 16-19 Variance tests, 25-27
operating characteristic curve, 21-23 homogeneity, 60-61
sample size, determining, 23-25 maximum/minimum, 60
tests of hypotheses, 19-21 Variances, homogeneous, tests for, 60-
Statistics 61
defined, 15
sample, 16
Stepwise Regression program, 168, 173, Whole-plot error, 268
176 Winer, 243
Studentized ranges, table of, 389-390
Student’s t distribution, 54, 381 Yates, F., 77, 118, 140, 355
Sum of squares (SS), 17, 18, 41-43, 56 Yates method, 118, 119, 121-123, 288,
degrees of freedom, 73 308, 309, 352, 353
Survey research, 2 Youden square, 80, 360-362
 ■n ,-
' .■
1

I S) , ■^-,1
. •,» \ •.'r ’ .•
s\

A i'#. * -' •
.’.'r

•*, - - ♦
♦ » #
. * *t\’,/'■ V; , 1 1 f'w

r*t \ ■
.■.'Si

'-,*i- i 'itc ’ ■ •_ •'•’^ V’J

< ‘-•'i
.V ^

J
< ■■■ V

= ^ '.? li* ■'

.♦ .. . >•'''?■
■” ''\[li .). 'j*. •'■-'i- r ' --•■t.* - ‘ * •^',.-1- V--

:■^'• W'. 'V,.-■

' r '

.. '.A' . '■>>;■■ ■: v»*T-jr jri«i<*’W'‘.’. • . ' '.i ^ ^ .'

■ Tv ‘ .. ’'' ■ -

W!>O'L».' ; T'fflHP -...v 'V • ' •.■■ •• ■ n* ->■»*

r,...: ' V .' ^ : ,,, ,''

^ . I

,: A*'- ' .' «r^

’ '^2^
' : .:^v• - V i,r^‘-
•^-. a ■ \ '4m
■ ^i»r.T

iCSkMp , -■.<(» * f V, ' "

iV^;--;, .,^.1 .*'^

'3T ' k'-r" ’• j, ^-n-iT : .''V.v . H.' ..

r _J»'W • .*“. • t • • A
.:5.' /' s . ..-
'y'l '•*
■'ii ;■' • . * .'T«f -'^ '-o
-;,'T
■ , = 3»- .' -'*• V.,,
■'’, * ,'•' ‘ *.- ■ •’ Jt-'> ' V*I "i.! » t
•’•: ’ ■•,■» iifj.,1 .;- sJ-f.-'-'r-

•r K'' ■
iT' f ■ . . A
• ■ . .• * '■ ' \ ^ L ’
' V '■^'' t V i- . .* ^*.'V' *2^1

•-■ ' -■• ■* '-'... ’ >•' ^

'■'' *-’•'■♦ , ’ ■'. ' m ‘'i
^ *•* > ‘jj’. ' .*-■: ■*>!':
<*'» .. .V,>y. ^

'■■ ■■ ;, " *c< '■

mmi0t
 V
■«<
%
I I A

)
I'
IT
i
9 4

ft.

1 ,

■ r“
■'i

.9
4^:
iP-'

«r^ -

V -*«

S>

l‘

' f

<s I