Professional Documents
Culture Documents
THE
STATISTICS LIBRARY
UNIVERSITY OF GLASGOW
'
. . / \
r
\
J
Design and Analysis
of Experiments
M. N. DAS
Central Water Commission, New Delhi
N. C. GIRI
University of Montreal, Quebec, Canada
10 98765432
F
Published by Mohinder Singh Sejwal for W/iley Eastern Limited
4835/24 Ansari Road, Daryaganj, New Delhi 110002, and
printed by Mohan Makhijani at ReJchaPrinters Private Limited
A 102/1 Okhla Industrial Estate, Phase A New Delhi 110020.
Printed in India. fh .
Preface
The material presented in the book is the product of the experience gained
by the authors while offering courses on design and analysis of experiments
to graduate and post-graduate students and applied workers in the
Institute of Agricultural ResearchStatistics,New Delhi and Department of
Mathematics, University of Montreal, Canada. The book has been written to
suit the academic persons including teachers and students and the applied
workers who have to apply statistical principles for the collection and
interpretation of experimental data. Almost all the commonly adopted
designs have been included in the book. Most of the advanced techniques
and methodology available for designing and analysis of experiments are
included here together with discussion of elementary concepts and pre¬
liminary treatments of different topics. A number of new concepts and
alternative methods of treatment of several topics have been presented.
Simple andj convenient methods of construction of confounded symmetri¬
cal and asymmetrical factorial designs, alternative methods of analysis of
missing observations, orthogonal latin squares, designs for bio-assays and
weighing designs are some of the examples. The book contains in all nine
chapters including a chapter on basic statistical methods and concepts. A
chapter on designs for bio-assays required for pharmaceutical investigations
and response surface designs has also been added.
Complicated mathematical treatment has been avoided while presenting
the results in the different chapters. No emphasis has been given on
combinatorics while presenting the methods of construction of various
designs. More emphasis has been laid on common sense, experience and
intuition while introducing the topics, providing proofs of main results
and discussing extension and application of the results.
The book can serve as a textbook for both the graduate and post¬
graduate students in addition to being a reference book for applied
workers and research workers and students in statistics.
The authors acknowledge gratefully the facilities provided by the
University of Montreal for writing the book.
M. N. Das
N. C. Giri
May 1979
'
, , /
. \
■ '
'
V
■
Contents
CHAPTER 1
Concepts of Experiments: Design and Analysis .
1.1 Design of Experiments and Collection of Data 1
1.2 Experiments and their Designs 2
1.3 Methodology for Making Inferences 3
1.4 Three Principles of Designs of Experiments 7
1.5 Experimental Error and Interpretation of Data 8
1.6 Contrasts and Analysis of Variance 10
1.7 Models and Analysis of Variance 16
1.8 Two-Way Classified Data 26
1.9 Assumptions of Analysis of Variance 43
CHAPTER 2
Complete Block Designs
2.1 Completely Randomized Designs 48
2.2 Randomized Block Designs 52
2.3 Latin Square Designs 56
2.4 Missing Observations in Randomized Block Designs 64
2.5 An Illustration 74
CHAPTER 3
Factorial Experiments
3.1 Characterization of Experiments 78
3.2 Factorial Experiments 79
3.3 Factorial Experiments with Factors at Two Levels 80
3.4 Finite Fields and Designs of Experiments 83
3.5 Grouping for Interaction Contracts 85
3.5 Confounding 86
3.7 Confounding in more than Two Blocks 89
3.8 Experiments with Factors at Three Levels Each 91
3.9 A General Method of Construction of Confounded
Factorials 97
3.10 Maximum Number of Factors to Save Interactions upto a
Given Order for a Given Block Size 103
3.11 Analysis of Factorial Experiments 105
3.12 Fractional Factorials 107
VI
CHAPTER 4
Asymmetrical Factorial and Split-Plot Designs
4.1 Asymmetrical Factorial Designs 120
4.2 Confounded Asymmetrical Factorials 120
4.3 Construction of Balanced Confounded Asymmetrical
Factorials 122
4.4 Construction of Confounded Asymmetrical Factorial
vx22 in 2v plot Blocks 131
4.5 Analysis of Balanced Confounded Asymmetrical Factorials 133
4.6 Split-Plot Designs 143
4.7 Analysis 144
CHAPTER 5
Incomplete Block Designs
5.1 Varietal Trials 152
5.2 Incomplete Block Designs 153
5.3 Balanced Incomplete Block Designs 156
5.4 Construction of B.I.B. Designs 158
5.5 Analysis 166
5.6 Analysis with Recovery of Inter-Block Information 168
5.7 Youden Squares 172
5.8 Lattice Designs 172
5.9 Partially Balanced Incomplete Block Designs 176
5.10 Analysis of P.B.I.B. Designs 183
5.11 Analysis with Recovery of Inter-Block Information 186
5.12 Optimality of Designs 198
CHAPTER 6
Orthogonal Latin Squares
6.1 Orthogonal Latin Squares 205
6.2 Maximum Number of Orthogonal Latin Squares 205
6.3 Construction of Orthogonal Latin Squares 206
6.4 Construction of Orthogonal Latin Squares by Using
Pairwise Balanced Designs 209
CHAPTER 7
Designs for Bio-assays and Response Surfaces
7.1 Bio-assays 277
7.2 Direct Assays 218
7.3 Indirect Bio-assays 220
7.4 Parallel Line Assays 224
CONTENTS vii
CHAPTER 8
Analysis of Covariance and Transformation
8.1 Analysis af Covariance 263
8.2 Analysis of Covariance for Randomized Block Designs 263
8.3 Analysis of Covariance of Completely Randomized and
Latin Square Designs 267
8.4 Analysis of Covariance of Non-Orthogonal Data and Designs
in Two-Way Classification 267
8.5 Analysis of Covariance with Two Ancillary Variates 272
8.6 Covariance and Analysis of Experiments with Missing
Observations 274
8.7 Transformations 275
CHAPTER 9
Weighing Designs
9.1 Introduction 280
9.2 Definition 280
9.3 Method of Estimation 281
9.4 Incomplete Block Designs as Weighing Designs 282
9.5 Two Pan Weighing Designs from B.I.B. Designs 296
9.6 Two Associate P.B.I.B. Designs as One Pan Weighing
Designs 289
9.7 Weighing Designs from Truncated Incomplete B.I.B.
Designs 290
9.8 Efficiency 290
v; 1 , -&C '■c-5
■
• ' ■ LI l&i ,
I
'
r ;i:
Concepts of Experiments:
Design and Analysis
and number of the experimental units, the manner in which the treatments
are allotted to the units and also the appropriate type and grouping of the
experimental units. These requirements of a design ensure validity,
interpretability and accuracy of the results obtainable from an analysis of
the observations. These purposes are served by the principles of
(i) randomization which defines the manner of allocation of the treatments
to the experimental units, (ii) replication which specifies the number of
units to be provided for each of the treatments and (iii) error control
which increases the precision by choosing appropriate type of experimental
units and also their grouping.
at \/2tt
where y denotes the variable of yield of the crop, pi is the mean yield and
ffi2 is the variance of the ith variable.
4 DESIGN AND ANALYSIS OF EXPERIMENTS
the hypothesis is rejected at 5 per cent level of significance. This shows that
we are likely to reject a correct hypothesis in 5 per cent of the samples.
If, again, the calculated value of the test statistic is greater than its tabulated
value at 1 per cent level of significance, the probability of getting such
samples as has given the value is less than 1 per cent if the hypothesis is
correct. In this case the hypothesis is rejected at 1 per cent level of
significance. Evidently, rejection at 1 per cent level of significance carries
more definite information than rejection at 5 per cent level of significance,
because when we reject an hypothesis at 1 per cent level we are likely
to reject a correct hypothesis for 1 per cent samples while in the case of
5 per cent level rejection we are likely to reject a correct hypothesis in
5 per cent cases. Thus, in the 1 per cent case we are likely to be in less
error by rejecting an hypothesis.
Though we have discussed the test procedure with reference to 5 per
cent and 1 per cent levels of significance, it is not necessary that testing
should be restricted to only these two levels. Tables are available for
testing at other levels of significance if so required. For example, in a
situation where more definite information is required, a test at 0.1 per cent
level of significance can be applied.
In relation to design and analysis of experiments, two test statistics,
viz. (i) /-test for testing the significance of the difference between two linear
estimates, and (ii) F-test or variance ratio test for testing equality of two
variances, are usually applied. Some of the hypotheses tested by, these
tests are discussed below:
Case I: One sample from a normal population with mean g.. When the
problem is to know if p. differs from a given or conjectured value the
hypothesis (x=^0 is made. The test statistic to be applied in this case is
called one sample /-test and is taken as
y-Fo
VsZln
where y is the sample mean based on n observations from a normal
population and s2 is the error mean square given by (l(y—y)2/(n—l) and
has («—1) degrees of freedom (d.f.). The concept of degrees of freedom
has been discussed in a later section.
Here, / is said to have the degrees of freedom of s2, that is, n— 1. The
5 per cent or 1 per cent values of / depend on the d.f. of / and are tabulated
for different values of d.f. Usually, /-table is prepared by showing the
5 per cent level of significance as /.975. This is done in consideration of
the fact that / has a symmetrical distribution and can be both positive and
negative as the mean of the /-distribution is zero. Actually, /.025 is negative
and the probability of / being less than /.025 is .025. Thus, the 5 per cent
significance value of / is a value, say, /„' such that the probability of having
samples to give values of / exceeding /„' is 2.5 per cent and that of having
DESIGN AND ANALYSIS OF EXPERIMENTS
6
samples to give values of t less than —/„' is 2.5 percent so that the
probability of t being greater than t0 or less than —10 is 5 per cent. This
type of test which takes into account the probability of occurrence of
extreme values of a test statistic on either direction of its mean value is
known as two-tailed test. In such cases the alternatives to the hypothesis,
viz. do not specify the nature of difference, that is, whether the
difference should be only positive or only negative. The hypothesis
which specify the direction like the alternative being \x > {*„, leads
to one-tailed test. In such tests the 5 per cent level of significance corres¬
ponds to t.05. Though we have discussed the test with reference to 5 per
cent level of significance, the same considerations and procedures apply to
tests at other levels of significance.
Case 2: Two samples from two normal populations having means Ui and
{a2 and a common variance. Setting the problem to test if the estimates
of mean from the two samples are homogeneous, the hypothesis
is made. The test statistic / is defined as
y*—y
where yx and y2 are the means of the two samples of sizes ftj and n2
respectively,
v2-.^ Q’lf "ki)2+
«l + «2 —2
Case 3: Variance ratio or F-test. Let mean squares, sx2 and s22 be two
estimates of variance a2 given by
We have seen that randomization, replication and error control are the
three main principles of designs of experiments. The roles they play in
data collection and interpretation are discussed below.
Randomization
After the treatments and the experimental units are decided the treatments
are allotted to the experimental units at random to avoid any type of
personal or subjective bias whiqh may be conscious or unconscious. This
ensures validity of the results. It helps to have an objective comparison
among the treatments. It also ensures independence of the observations
which is necessary for drawing valid inference from the observations by
applying appropriate statistical techniques.
We shall see subsequently that depending on the nature of the experi¬
ment and the experimental units, there are various experimental designs.
Each design has its own way of randomization. We shall, therefore,
discuss the procedure of random allocation separately while describing
each specific design. However, for a thorough discussion on the subject
the reader may see Fisher (1942), Kempthorne (1952) and Ogawa (1974).
Replication
error variance the more is the precision. Once a measure of error variance
is available for a set of experimental units, the number of replications
needed for a desired level of sensitivity can be obtained as below.
Given a set of treatments an experimenter may not be interested to
know if two treatments differ in their effects by less than a certain
quantity, say, d. In other words, he wants an experiment which should be
able to differentiate two treatments when they differ by d or more.
As discussed in the previous section the significance of the difference
between two treatments is tested by /-test where
t _ yt—yj
V 2 s2/r
Here, yt, and yj are the arithmatic means of two treatment effects each
based on r replications, s2 is a measure of the error variation.
Given a difference d, between two treatment effects such that any
difference greater than d should be brought out as significant by using a
design with r replications, the following equation provides a solution of r:
r=-L£L
0 V2s2/r ’
X
where /0 is the critical value of the /-distribution at the desired level of
significance, that is, the value of / at 5 or 1 per cent level of significance
read from the table. If s2 is known or is based on a very large number
of observations, made available from some pilot pre-experiment investi¬
gation, then / is taken as the normal variate. If .s2 is estimated with
n degrees of freedom (d.f.) then /„ corresponds to n d.f.
When the number of replications is r or more as obtained above, then
all differences greater than d are expected to be brought out as significant
by an experiment when it is conducted on a set of experimental units
which has variability of the order of s2.
For example, in an experiment on wheat crop conducted in a seed
farm in Bhopal, India to study the effect of application of nitrogen and
phosphorus on yield, a randomized block design with three replications
was adopted. There were 11 treatments two of which were (i) 10 lb/acre
of nitrogen (ii) 20 lb per acre of nitrogen. The average yield figures for
these two applications of the fertiliser were 1438 and 1592 Ibs/acre
respectively and it is required that differences of the order of 150 lb/acre
should be brought out significant. The error mean squares 02) was
12134.88. Assuming that the experimental error will be of the same
order in future experiments and t0 is of the order of 2.00, which is likely
as the error d.f. is likely to be more than 30 as there are 11 treatments,
we have all the information to obtain the numbers of replications r from
the relation:
/ - |J|-
#“V2Wr
10 DESIGN AND ANALYSIS OF EXPERIMENTS
Error Control
The main technique adopted for the analysis and interpretation of the data
collected from an experiment is the analysis of variance technique which
essentially consists of partitioning the total variation in an experiment into
components ascribable to different sources of variation due to the controlled
factors and error.
The following discussion attempts to relate the technique of
analysis of variance to comparisons among treatment effects which in terms
of symbols can be called contrasts.
CONCEPTS OF experiments: design and analysis 11
Contrast
Let ylt y2,... ,y„ denote n observations or any other quantities. The linear
n »
function C=X 1 tfi where l/’s are given numbers such that Y 1<=0, is
i-i i
n
When there are more than two contrasts, they are said to be mutually
orthogonal, if they are orthogonal pairwise.
For example, when there are four observations yx, y2, y2 and yi} we can
write the following three mutually orthogonal contrasts:
(t) Jl+^2-.V*
(u) yi-y2-ya+y*
(iii) yi—yi+yz—y*
Ci=Z lj iyi
1
n
c2=Z iztyi
i
Z 1 mi l/=0
i
(i) yi-yz
(“) Ti+Tg—2ys
(iii) Ti+T2+T3-3v4
and so on.
CONCEPTS OF experiments: design and analysis 13
Measures of Variation
TABLE 1.1
Yield Data from the Irrigation Experiment (lb/acre)
Irrigation levels
Treatments Marginal total
h h h
h 3*u 3*12 3*13 * +3*12+3*13
3 11
(837) (804) (843) (2484)
Here, yu (/= 1,2, y=l, 2, 3) denotes the observation from the ith treatment
and yth level of irrigation.
It will be seen that the following two orthogonal contrasts are free
from the effects of the two controlled factors viz. treatments and irrigation:
(i) 0'u->,2i)-0'i2-J’22)
00 O'ls-Jaaj
Each of these contrasts is actually a contrast of contrasts. In a contrast
of two observations in the same column, the effect of irrigation is removed.
Again, by taking contrast of such contrasts the effect of treatment is
eliminated. Evidently, it is not possible to obtain any more error contrast
which is orthogonal to each of the above two error contrasts. Hence, the
s.s. due to error variation, in short, error s.s. can be obtained by adding
the squares of these two contrasts. This s.s. has two d.f.
Again, let us take the contrast
O'l+J'a+J's) ~ O^+Ts+ya)
It is easily seen that this contrast gives a comparison between the effects
of the two treatments. This contrast is not affected by the effects of the
irrigation levels as they are evenly distributed in the positive and negative
parts of the contrast. There is no other contrast orthogonal to the above
which also represents purely treatment comparison. This contrast is also
orthogonal to each of the two error contrasts presented earlier.. We get the
treatment s.s. by obtaining the square of the above contrast and dividing
it by the appropriate factor as indicated earlier. It has evidently 1 d.f.
On the hypothesis that there is no variation among the treatment effects,
the treatment s.s. is distributed as c2/2 withl d.f. Thus, on the above
hypothesis both the error m.s. and treatment m.s. have the same expected
value, a2. They are independent as they are obtained from orthogonal
contrasts. Their ratio, F, can, therefore, be used to provide a test of the
hypothesis of equality of the treatment effects.
To complete the analysis of variance, we have yet to account for two
more orthogonal contrasts each of which is orthogonal to the three
contrasts already discussed. We can write these two contrasts as below
from the irrigation marginal totals:
0) (>;U+^2l)-(>’x* + ^22)
(ii) (>;iX+3;2x) + (>,ia+>’22)-2 Oia+J^)
They represent comparison of irrigation effects and their s.s. gives a
measure of variation due to irrigation effects.
Table 1.2 shows in a compact form the details of the analysis of
variance obtainable from the contrast approach.
As the figures in bracket in Table 1.1 are averages based on 4 observations
each, the actual divisors for getting the s.s. is 1/16 times each of the
divisors shown in col (3) of table 1.2 for the different s.s. The s.s. obtained
CONCEPTS OF experiments: DESIGN AND ANALYSIS 15
X yi?~
Analysis of Variance Table
Total
16 DESIGN AND ANALYSIS OF EXPERIMENTS
are the normal equations by solving which estimates of the effects can be
obtained. It can be shown that all these normal equations are not
independent. As such unique solutions for the effects are not available.
Solutions are, however, available only when the sum of the effects of each
factor is assumed to be zero. The implication of such assumptions has
been discussed subsequently.
Once the estimates are available as above, we can get the value of E
when the estimates of the effects are substituted in (1.1). The value of E
thus obtained is called the error s.s. It is said to have degrees of freedom
equal to N—k where N is the total number of observations and k is the
total number of independent normal equations. Clarification of the
concept of degrees of freedom has been given at the end of the section
1.7.2.
The next problem is to obtain the variability due to the effects of a
factor and compare it with the error variability. An hypothesis about the
effects of this factor is then made. The hypothesis is usually that the
effects of the different levels of the factor are equal, that is, each effect is
zero by virtue of the fact that their sum is zero. On this hypothesis we
get another model in which the effects of the factor under test is absent.
On this model also we can get another error s. s., say, Ex. Actually, Ex
contains the variability due to the effects of the factors under test and also
the error variability. Thus, Ex—E gives a measure of the variability due
to effects of the factor under test and is called the s. s. due to the factor.
Its d.f. is one less than the number of levels of the factor. Next, the mean
squares (m. s.) for each of error s. s. and the s. s. due to the factor are
obtained by dividing the s. s. by their respective d. f. These two m. s’s are
independent and hence their ratio is tested by the variance ratio test. The
hypothesis here is that both these mean squares have the same expected
values which is possible only when all the effects under test are equal.
After this general discussion of analysis of variance we have discussed
below specific types of data and the details regarding their analysis.
yu |’ 2’ ' ,Tjfj against the zth level. Then the observations yu classified
in k groups according to the k levels of the factor are said to form one¬
way classified data.
Two-way classified data: Similarly, if we take two factors simultane¬
ously, say, A and B at numbers of levels k and r respectively, then there are
kr cells each of which is defined by one level of A and one level of B. For
example, the data presented in Section 1.6 are arranged according to the
levels of two factors, treatments and irrigation at levels 2 and 3 respectively.
There are 6 cells each having one observation.
Let there be njj observations in the (/, y')th cell defined by the zth level
of A and jth level of B. Let yuk denote the ftth observation in the (z, y)th
cell. Then the data
(*=1,2, .. .,nu \
j 1, 2,... ,r I
i— 1, 2,... ,k J
arranged in the rk groups are called two-way classified data.
Similarly, in-general, m-way classified data are defined using levels of
m factors simultaneously.
Let yu denote the jth observation in the zth level of a factor, say, A and
Y{j denote the corresponding random variate. We make the assumption
that Yij=y.+tl+elj where [x represents the general mean which is fixed and
stands for general condition of the experimental units and treatments, U
denotes the effect of the zth level of the factor (z=l, 2,... ,k). It can be
both random and fixed depending on how it has been chosen. For example,
if U denotes the effect of a crop variety which has been chosen at random
at the zth draw from a large number of varieties, its effect is a random
variable. If, again, U denotes the effect of one treatment in a given
complete set of varieties, it is a fixed effect. When ti s are fixed the model
is called fixed eflect model and the variation due to the treatments,
tt (/= 1, 2,.. .,k) is said to be controlled. We shall discuss here the fixed
effect model.
The last component in the model, eu is a random variable and is called
the error component. It makes Yi} a random variable. The variation in
it is due to all the uncontrolled causes which can influence the observation.
It has a hypothetical distribution, because there are a large number of
uncontrolled factors which contribute to the realization of a particular yu.
Actually, an indefinite number of possibilities of any particular ei} arises
from an infinite number of possible combinations of the many levels of an
unspecified number of uncontrolled factors which influence y{J. The one
that is realized is the outcome of the particular combination of these
20 DESIGN AND ANALYSIS OF EXPERIMENTS
£=z (yu-it-uy
U
(1.4)
»•/ V , m n )
G2
The first part of (1.4), viz. -is called corrected total sum of
u n'
squares.
For obtaining a measure of variation due to the treatments, we make
the hypothesis that all the treatments have equal effects. The model on
this hypothesis becomes
Yij — {a -f etj
Proceeding as before we get the error s.s., Ex on the hypothesis of equal
G2
treatment effects as EX = E. yu2-• It actually contains the variation
a n-
due to both treatments and errors. Hence, we get a measure of treatment
variation from
Ti2 G2
Ei — E= Y — - (1.5)
, ^ n-
(that is with the same d.f. as the two m.s.) table value of F. If the
calculated value of F is larger than the tabulated value, the hypothesis is
not tenable and is rejected. The table values of F are available at 1, 5 and
10 per cent levels of significance.
If the F-test rejects the hypothesis, it is necessary to find an index for
comparing the effects two by two to have a better picture about their
distribution. That is, given two effects, the problem is to know if they
differ significantly. It is also necessary to have a measure of the precision
of their estimates. For a measure of precision first the variance of the
estimates of individual effects is obtained from
where a2 is the error variance and is estimated by s2, the error m.s., that is,
V(u)=s2lni
and can apply the /-test to test the significance of the difference. The
d.f. of / is the same as that of s2. If the table value of / at a given level
of significance is /„, then CD (ij)—t0\/s2 {l/nt + l/rij) gives an index such
that if a difference between two effects // and tj estimated with m and nj
observations, using the above method, is greater than CD (ij) then the
treatments differ significantly, otherwise not. If all the effects are estimated
with the same number of observations, say, n, then CD{ij)=y/2s2jn and is
independent of the pair of effects chosen. In this case the index is called
critical difference and is denoted by CD. If the difference between any
two effects is greater then CD then the effects differ significantly. Using
CD a bar diagram can be drawn which shows all effects which do not
differ among them significantly under one bar.
CONCEPTS OF experiments: design and analysis 23
When F-test reveals that the hypothesis can not be rejected, no bar
diagram is necessary. Other types of procedures for such testing the signifi¬
cance of the difference between pairs of effects are, among others, Student
Newman-Keuls range test, multiple range test by Duncan(1955), and S' and
T method by Scheffe (1953). We shall discuss here only the first two of the
above procedures. For a detailed discussion on these procedures and
their relative merits and demerits the reader is referred to Federer (1955).
Though we shall discuss these procedures for comparing means in two-
way classified data, they can be applied in other similar situations without
any complexity.
Let the observed estimates yx, y2,..., yk which are the means in the
present case be arranged from the highest to the smallest as Xk, Xk-i.. .xx,
where xk denotes the highest among yx, y2,..., yk and xt is the ith largest
among yx, y2.. .yk, i=( 1, 2.. .k). The quantity Xk—xx is called the range
of the set of means and
__(xk—Xi)\/n
4k p ~
where nx—n2.. ,np=n (by assumption) and p is the d.f. of error m.s. (s2)
is called the Studentized range of the observed means yx, y2.. .yk- Let
qa,k,p denote the upper a per cent point of the Studentized range. The
values of q<x,k,P for different values of a, k and p are available in statistical
table in Federer (1955).
Student-Newman-Keuls test: We compute the upper a per cent point
of Studentized Critical Range Wk given by
Next, we compare the observed range Xk—xx with Wk- If it is less than
Wk, the process of comparison stops and we conclude that the means are
all of the same order. If Xk—xx is greater than Wk, we devide Xk,. .., xx
into two subgroups, one containing Xk,..., x2 and the other from Xk-i to
xx. Next, each of the ranges in the two subgroups viz Xk—X2 and
Xk-1—JCj, is compared with Wk-j. If either range does not exceed, Wk-i,
then the means in each of the groups are equal. If either or both ranges
exceeds Wk-u then the k^-\ means in the group concerned are divided
into two groups of k—2 means each and the ranges for these subgroups
are compared with Wk-2- This procedure is continued until a group of
i means is found whose range does not exceed W/. In other words, by
this method the difference between any two means is significant when the
range of the observed means of each and every sub-group containing the
two means under test is significant according to Studentized Critical
Range.
Duncan’s multiple range test: This test procedure is essentially the same
as the previous procedure except the observed ranges are compared here
24 DESIGN AND ANALYSIS OF EXPERIMENTS
range table.
The results obtained from the analysis of one-way classified data are
presented in form of an analysis of variance table as shown below. This
method has also been illustrated by analysis of a set of actual data. The
estimates, SE's and CD’s and other calculations have also been presented
there.
TABLE 1.3
Sources of s.s.
d.f. s.s. F
variation m-s-=dT
TABLE 1.4
Data on Increased Body Weight of each Set (oz) Classified According to Feeds
Feeds
(iv) sawan; and (v) meat meal. The cockerels under study were allotted
to the feeds at random and the increase in their body weight over a period
of 10 weeks was recorded. The data of increase in body weight in oz of
individual birds are shown in Table 1.4 separately for the different
treatments.
Denoting the totals of the observations of the different feed treatments
by T0, Tj, T2, T3 and T4 in the same order as above, we have
7’0=199.7(7)
7\=249.9 (9)
J3=287.8(10)
r8=327.3(12)
7>=284.9(11)
The figures in bracket indicate the numbers of observations («,) in the
different totals so that
n=X 49
Uncorrected total s.s.=]T j</=38056.8800
TABLE 1.5
Analysis of Variance Table of One-Way Classified Data
Total 48 885.0400
26 DESIGN AND ANALYSIS OF EXPERIMENTS
As the variance ratio is less than unity, the variability among the feed
effects is not even as much as can be expected due to chance causes,
that is, the error variance. This fact is expressed by saying that the feed
effects are not significantly different.
The mean effects of the feeds along with their standard errors
are presented below:
Here, G, Ai, Bj, Ttj are as defined earlier. These equations are not all
independent. The p equations at (1.7) when added give equation (1.6).
Similarly, the q equations at (1.8) when added also give (1.6). The
equations at 1.9 give the equation at 1.7 when added over j and those
at (1.8) when added over i. In order to get unique solutions of the effects,
we impose the restrictions that
2 fl/=0, 2bj=0,2 cij=2cij=0
i i i I
The implications of the first two of these restrictions are that effects are
estimated as deviates from their means; ct/s for a given i are estimated
as deviates from the effect of the ith level of A and so on. Under these
restrictions we get
[i,—G/npq
at=(Ai/nq)—G/npq
f>J—{B)lnp)—Gjnpq
Tij * £ y\
CONCEPTS OF experiments: design and analysis 29
=Z yuk2-—- £ Si^,- Z M;
i J
i--2 A. A \
-(z^ Z Z bjBj — ^G ) (1.10)
vfr » i J '
npq
with npq—pq d.f.
It is seen that the error s.s. £ has come out as within cell s.s, so that
variations due to the other causes are contained in the between-cell sum
of squares. In order to estimate these variations, we shall treat the cell
totals Ti/s as the observations and the model becomes
Tn=nii-\-nai-\-nbi-\-nci)-}-Z euk
k
T• 1
That is, —=v.+at+bj+Cij+Eij where Eij=~Y eW
n n kj
In the model ctj and E,j get mixed up and the error s.s. in this model
includes the interaction s.s. In this model the error s.s., say, E' comes
out as
Tn2 G2
=z Y a‘Ai — Z bjBj
n npq
This s.s. is the interaction s.s. with (p—1) (q— 1) d.f. Making the
hypothesis that Cj=0, (i=l,2,...,/>) we get
B,'=Z
Tu2
=V [iG— Y bjBj where p. is
n npq
nq °a
npq
^pn npq
TABLE 1.6
2 A(* G2
Due to A P-1 V Sa2ISab2
nq npq
2 Bf2 G2
Due to B q-1 So8 sb2isab2
pq npq
/S7V,2 E AS S G2 ^
Int. AB (P—1) (?-l) \ n nq pq Vnpqj
Sab2 sabw
Error pq (n—1) V
For such data first the interaction m. s. is tested against the error m. s.
If the test shows that interaction variation is not significant, the error s. s.
and the interaction s. s. are added and then divided by their total d. f. to
get the pooled m, s. which is then used to test the significance of the
variation due to A and B. If interaction s.s. comes out significant, then
the interaction m. s. alone is used to test the significance of the variations
of A and B.
The rest of the details regarding presentation of the results is the same
as discussed in the case of one-way classified data. When there is only
one observation per cell, the within cell s. s. is not available and in such
cases the interaction s. s. is used as error. If it is known in advance that
there is no interaction, the results can be derived on similar lines. Actually,
in such cases the error s. s. becomes the sum of the interaction s. s. and
the error s. s. as obtained from the model with interaction. In other
words, the error s. s. under this model is given by
Error s. s. = Total s. s. — s. s. due to A — s. s. due to B
where the total s. s. and the s. s. due to A and B are the same as obtained
earlier.
The method of analysis has been illustrated in the next section by
analyzing an actual set of data.
CONCEPTS OF experiments: design and analysis 31
1.8.2. Example
The data presented below are the wool yield in (oz) of individual clips of
24 ewes belonging to three crops (generation) of progenies of farm bred
stock maintained in a sheep breeding research station in Rajasthan,
India. For each ewe clip yield was recorded for 6 clips at intervals
of 6 months. There were 8 animals in each crop.
TABLE 1.7
Data of Clip Yield (in oz per Clip)
Clip Nos. 1 2 3 4 5 6
16 12 12 38 24 18
First crop 15 18 32 24 23 18
progeny 16 12 23 30 24 19
19 12 20 32 30 15
16 26 16 22 32 20
17 18 22 25 28 18
18 24 20 30 20 24
14 16 22 26 20 17
22 28 26 38 30 22
Second crop 20 26 24 30 32 20
progeny 16 20 24 36 29 24
15 24 20 26 40 22
19 30 28 38 22 20
17 20 24 36 22 26
16 23 30 30 20 18
19 20 28 30 24 22
15 16 22 25 24 30
14 16 27 27 19 24
Third crop
16 26 25 35 24 26
progeny
17 30 19 35 21 16
18 20 20 39 18 18
20 18 21 40 26 20
15 15 45 39 27 16
27 38 28 24 28
16
TABLE 1.8
Table of Cell Totals (Each Total Based on 8 Observations)
Crops 1 2 3 4 5 6
TABLE 1.9
Analysis of Variance Table
Sources d. f. s. s. m. s. F
Pooled=
Total 143 6661.75 24.43
Clips
t—A
©t—A
II
Generation
Means 21.10 24.91 23.85
S. E. == 0.07
k q<xkP Wk=qakp
2 3.15 3.26
3 3.88 4.02
4 4.33 4.48
5 4.65 4.82
6 4.91 5.09
As the range Xk—Xj for k=6 is 14.71 and the value of Wk for k=6 is
5.09, the 6 averages are different. Accordingly, they are broken into the
following two groups:
16.91, 20.70, 20.81, 24.50, 25.12
20.70, 20.81, 24.50, 25.12, 31.61
As W5=4.S2 is less than the ranges in the above two groups viz.
8.21 and 10.91, the averages in each of the groups are different. These
are subdivided again into the following three distinct groups each of
size 4:
16.91 20.70 20.81 24.50
20.70 20.81 24.50 25.12
20.81 24.50 25.12 31.61
As Wt is 4.48 while the ranges in the above groups are 7.59, 4.42 and
10.80 respectively, the range for the second group only is less than Wt.
The averages in this group, thus, do not differ among themselves.
The averages in the other two groups are therefore divided into the
following two distinct groups each of three:
16.31 20.70 20.81
24.50 25.12 31.61
34 DESIGN AND ANALYSIS OF EXPERIMENTS
As Wa=4.02 and the ranges for the above two groups are 3 90 and
7.11 the range for the first group is less than 4.02 and so the average in
this group are of the same order while those in the other are different.
Proceeding similarly it is seen that the last two averages viz. 25.12 and
31.62 are significantly different. The final position is summed up below
by putting the averages which do not differ under the same bar
16.91, 20.70, 20.81 24.50, 25.12, 31.62
The average 16.91 is significantly lower than each of 24.50, 25.12 and 31.63,
is significantly higher than each of all the other averages.
TABLE 1.10
Form of Three-way Classified Data (The figures in the cells indicate the number
of observations in the cells.)
Levels of A
<*i a2
Levels of C
Cl c2 Cl c2 ci c2
(1) (2) (3) (4) (5) (6)
bi n n n n n n
Levels of B b* n n n n n n
ba n n n n n it
From the three-way ABC table, we can form 3 two-way tables, that is,
AB table, AC table and BC table. The AB table is formed by putting
together all the observations under c1 and c2 occurring with each combi¬
nation (aftj). The BC table is obtained similarly by putting together all
the observations under the three levels of A occurring with each combi¬
nation (bicj), that is, by putting together the observations in columns 1, 3
and 5 and writing them in the same row on which they occur, we get the
observations in the cells bxcv bic1 and b3cx. Similarly, by putting together
the observations in the columns 2, 4 and 6, we get the observations in the
remaining cells of the BC table, that is, bxc2, b2c2 and bzc2. Similarly, by
CONCEPTS OF experiments: design and analysis 35
TABLE 1.11
Two-Way AC Table Formed from the ABC Table Showing the Cell Frequencies
a2 a3
Cl c2 Cl c2 Cl c2
3n 3n 3n 3n 3n 3n
To get the usual form of the table, it can be written after rearranging the
cells as shown in Table 1.12.
TABLE 1.12
- Levels of A
«i o2 «s
Cl 3n 3n in
Levels of C
c2 ‘in in in
Next, each of these tables can be analyzed as discussed for two-way data.
From these analyses we get the s.s. due to A, B, AB, C, AC, and BC. For
three-way data there is one more component called three-factor interaction
denoted by ABC. To get the s.s. due to ABC, first the s.s. due to the
cell totals of the three-way table is obtained from
Z Ti jk2/n — G2I npqr
where Ttjk is the total of the cell defined by the ith level of A, jib. level of
B and fcth level of C and p, q and r are the number of levels of A, B and
C respectively. Next, by subtracting from the between cell s.s. all the s.s.
due to A, B etc., we get the s.s. due to ABC with
(p-l) (q-1) (r-1) d.f.
This technique can be extended in a straight forward manner to m-way
classified data.
(a,- (1.15)
j n’J J V j n-j J mj*i \j n-J /
Substituting Qi for ( At
v r n'j J
C„ forn,.-Z—!
j nJ
ijk
ijk i j
ijk I n'J I
0-1&)
uk i n‘- j
R,=v,-Y.n~
i Hi-
,*
As 1.17 and 1.18 are both the same error s.s., we have
,
E—’-Z—*jrj
j nJ r
Next, for getting the s.s. of A we make the hypothesis ax=a2=... =a*=0
0-19)
-L-.
ijk J
TABLE 1.13
Analysis of Variance of Non-Orthogonal Data
B (unadj) r—1 2 QI
J n.j
Error n. ,—k—r+1 by subtraction
k J nJ i
Hence, the interaction s.s.
=e-e2=zn ~—~y.
ntJ j n.j
T• 2
———x: diQ,.
B -2
,
Its d.f. is (r— \)(k— 1), if there is at least one observation in each cell;
otherwise, it is reduced by the number of cells having no observation.
— C/m cr2
=ct2 (X qi X qmCim)
t m
=a2 X
/ « '
TABLE 1.14
3 7 1
Periods Sahiwal Total (jB,-) Average Bj/n.
4 8 2
/ 42 , 72 262
Cu =53- + —+ -^=26.404
V 8 +23 40 46 )
35 , 286 256 )
— Ci2
=G+ 23 40 46 j
— C13=5.802; —Cl4=5.867; -C23=4.237
|=14.737
TABLE 1.15
It has been seen that the main difficulty in the analysis of non-orthogonal
data is the solution of the reduced normal equations. There are, however,
certain types of data as discussed below for which the solution problem
does not exist.
Proportionate Cell FrequenciesOne such type of data is obtained when
the cell frequencies are the same in a column (row), though it may vary
from column to column. Suppose the frequencies in each row of the ith
column be n/. Then
rnp
Cti=rnr
'~N
where JV=2 m and r is the number of rows.,
rnm„
Qj
N
The reduced normal equations become
r <5<~~N ^
That is,
. «,• „ Qi
—2 nmam=-j
y il_G?
mi rN
TABLE 1.16
Frequency Distribution of Observations in a Two-Way Classified Balanced Data
Treatments
Groups l 2 3 4 5 6 7
1 l 1 0 1 0 0 0
2 0 1 1 0 1 0 0
3 0 0 1 1 0 1 0
4 0 0 0 1 1 0 1
5 1 0 0 0 1 1 0
6 0 1 0 0 0 1 1
7 1 0 1 0 0 0 1
r-4fiO
u
Til
E
2oi~— 51 am=Qi
mj&i
z' 3 Qt
Solutions are: ai=-y-
We have seen that the data should satisfy certain assumptions before these
methods can be applied. One of the assumptions is that the variance of
44 DESIGN AND ANALYSIS OF EXPERIMENTS
the observations should be constant. There are certain types of data for
which this assumption does not hold. For example, if the observations
are the number of plants (r) affected by a disease among a given number
of plants 1(aj) in a plot,then the fraction (r/n) which is taken for analysis is
binomially distributed and therefore does not have a constant variance.
In such a case the observations are suitably transformed such that the
variance of the transformed observation is constant. The methodology
for obtaining appropriate transformation indifferent situations is discussed
in a subsequent chapter.
We have seen that classification of data is a way of reducing the error
variance because the variation due to a factor according to which the data
have been classified can be taken out of the error variance by the technique
of analysis of variance. There are other methods of reducing the error
variance; one such method is called analysis of covariance for which
observations on another variate called concomitant variate from each of
the experimental units are used. It is, however, necessary that there is a
linear relation connecting the variate under study (>>) and the concomitant
variate (x). By this method the portion of variability in y which is due to
the variation in x, is eliminated from the error variance. The method will
also be discussed in another chapter together with the technique of
transformation of data.
EXERCISES
(0 7i+72-2y3
(ii) 71+4^2-2^3-3^4
2. Write down the complete set of mutually orthogonal contrasts among
7i> y2,- ■ ,78 using only +1 and —1 as coefficients.
3. Show that the sum total of the s.s due to all the (n—1) mutually orthogonal
contrasts among n observations is equal to the s.s obtained by summing the
squares of their deviations from the mean.
Solution: Write the contrasts as
C1=2Lk7<
C2='E,Ltiyi
C„-i — 2Z.(„-i)<7<
where 2Lmi =0
CONCEPTS OF experiments: design and analysis 45
m , r L- ^ X1+J2+. • •yn
Take one more function C0=-7=-
Vn
Then the following matrix is an orthogonal matrix:
L 21 ■^'22 • • •
j_ jl _i
— y/n \/n Vn
m***± m*=*L
71-1 1
71-1 1
771-1
4. Show that two-way classified data where the cell frequencies are proportional,
that is, mj: nij'^Cj.C., (/,/=!, 2,...r) are orthogonal.
5. In a set of two-way data classified according to k levels of factor A and r
levels of B, there is one observation in each cell. Show that the total
number of error contrasts is (r-1) (A:—1). Describe a convenient method of
writing them.
6. If one observation is absent in one cell of the data in problem 5, obtain the
reduced normal equations for the effects of A and get their algebraic
solutions. Find the variance of the contrast estimate t-tm where the cell
with no observation belongs to /,.
7 Show that the /-value for testing the significance of any contrast estimated
from non-orthogonal data is unaffected when the contrast is multiplied by a
constant.
8 in a two-way classified data, the cell frequencies are shown below. Analyze
the data completely taking the general observations as y,j. Show that the
variance of the estimates of all treatment contrasts of the type t(—tm is
constant.
46 DESIGN AND ANALYSIS OF EXPERIMENTS
Treatments
1 2 3 4 5 6 7
Blocks 1 1 1 0 1 0 0 0
2 0 1 1 0 1 0 0
3 0 0 1 1 0 1 0
4 0 0 0 1 1 0 1
5 1 0 0 0 1 1 0
6 0 1 0 0 0 1 1
7 1 0 1 0 0 0 1
REFERENCES
2.1.1 Randomization
k
The 2 ri=R (say) experimental units are also numbered suitably.
i-i
One of the methods uses the random number table as discussed below.
Any page of a random number table is taken. If k is a one-digit number,
then the table is consulted digit by digit. If k is a two-digit number then
two-digited random numbers are consulted. All numbers which are greater
than k along with zero are ignored.
Let the first number chosen be nlt then the treatment numbered is
allotted to the first unit. If the second number is «2, which may or may
not be equal to«j, then the treatment numbered n2 is allotted to the second
unit. This procedure is continued. When the ith treatment number has
occurred rt times, (/= 1,2,..., k) this treatment is ignored subsequently.
This process terminates when all the units are exhausted.
One drawback of the above procedure is that sometimes a very large
number of random numbers may have to be ignored because they are
greater than k. It may even happen that the random number table is
exhausted before the allocation is complete. To avoid this difficulty the
following procedure is adopted. We have described the procedure by
taking A: to be a two-digit number.
Let P be the highest two-digit number divisible by k. Then all numbers
greater than P and zero are ignored.,If a selected random number is less than
Jfc,then it is used as such. If it is greater than or equal to k, then it is divided
by k and the remainder is taken to be the random number. When a
number is completely divisible by k, then the random number is k. If k is
a n-digit number, then P is taken to be the highest /i-digit number divisible
by k. The rest of the procedure is the same as above.
tail TT. Similarly, when the coin is thrown three times, there are the
following eight possible events:
HHH, HHT, HTH, HTT, THH, THT, TTH, TTT.
Similar events can be written easily for four or more numbers of throws
of the coin.
The five treatments are now labelled not by serial numbers as earlier
but by any five of the above eight events obtainable by tossing three coins.
Let us use the first five events and omit THT, TTH and TTT.
A coin is now thrown three times and the event noted. If the event is
any of the first five, the treatment labelled by it is allotted to the first plot.
If the event is any of the last three, it is ignored. The coin is tossed again
three times and this event is used to select a treatment for the second plot.
If the same event occurs more than once, we are not to reject it till the
number of times it has occurred equals the number of replications of the
treatment it represents. This process is continued till all the plots are
exhausted.
2.1.4 Analysis
., k
• ,rt )
where Yi} is the random variable corresponding to the observation yi}
obtained from theyth replicate of the ith treatment, ^ is the general mean,
U is the fixed effect of the z'th treatment and etJ is the error compo¬
nent as described in Chapter 1. The error components are assumed
to be independently and normally distributed with 0 mean and constant
variance a2.
Let j y,j=Ti be the observation total of the ith treatment. Following
the method of analysis suggested for the one-way classified data in Chapter 1
the following sums of squares are computed first.
COMPLETE BLOCK DESIGNS 51
K T,2 G*
1. Treatment sum of squares =2]-—
<=1 ri R
k k
where G=X Tt, R=Y. ri-
1 (-1
G2
2. Total sum of squares=]T ytJ2——.
u R
The analysis of variance is given in Table 2.1.
TABLE 2.1
y Ti 2 G2
Treatments k-1 St2 St2/S2
,-1 "
Error
(within treatments) R-k By subtraction s2
V 0 G2 * '
Total R-1 2- y*>2—jr
u R
The hypothesis that treatments have equal effects is tested by the F-test.
If F is not significant at the desired level of significance, the treatments
can be considered to have equal effects. If F is significant, the treatment
effects are not equal. In such cases it becomes necessary to estimate and
test individual treatment contrasts in which the experimenter may be
interested.
k
Estimate of any treatment contrast ]T hti where tt denotes the effect
i= 1
v Cl Uyi)=°2 X~
, f n
where a2 is the error variance which is estimated by the error mean squares,
s2 in Table 2.1.
52 DESIGN AND ANALYSIS OF EXPERIMENTS
I £ lm I
other /’s.
Presentation of Results
general, such groups are formed with units having common characteristics
which are known to have influence on the variate under study.
The number of blocks in the design is the same as the number of
replications. The k treatments are allotted at random to the k plots in
each block.
This type of homogeneous grouping of the experimental units and the
random allocation of the treatments separately in each block are the two
main characteristic features of randomized block designs.
Actual number of replications in the design is determined by the
availability of resources and considerations of cost and precision. A
method of finding the number of replications from considerations of
sensitivity of comparison of the treatments has been given in Chapter 1.
The treatments are first numbered from 1 to k in any order. The units in
each block are also numbered, conveniently from 1 to k. The A: treatments
are then allotted at random to the k units in each block. Random allocation
can be made either by consulting a random number table or by drawing of
lots or by coin tossing as described for completely randomized designs. If
k is not very large, not more than two digited numbers need be consulted.
A column is first chosen. If the first number in the column is k or less,
the treatment corresponding to this number is allotted to the first plot in
the first block. If it is zero or greater than ky it is omitted. Similarly, if
the next number is not greater than k and has not been chosen earlier,
the treatment corresponding to it is allotted to the second plot. All
numbers greater than k are ignored. When all the units in the first block
are allotted treatments, by following the above procedure, units in the second
and subsequent blocks are allotted treatments exactly similarly.
When all numbers greater than k are rejected, there may be too many
rejections. This difficulty can be avoided by consulting all two-digited
numbers except 00 and numbers greater than P where P is the highest
two-digit number exactly divisible by k. When a number is greater than
k but less than P, it is divided by k and the remainder gives the number
to be chosen. If the number is non-zero and exactly divisible by k, the
number chosen is k.
2.2.3 Analysis
The data collected from experiments with randomized block designs form
a two-way classification, that is, classified according to the levels of two
factors, viz., blocks and treatments. There are kr cells in the two-way
table with one observation in each cell. The data are orthogonal and
therefore the design is called an orthogonal design.
We take the model
Let further
Z Ti=E B)—G.
i s
We shall call G2/rk as correction factor denoted by C.F.
T,2
Sum of squares due to treatment = Z-C.F.
i r
The hypothesis that the treatments have equal effects is tested by F-test
where Fis the ratio st2/sz with (fc—T) and (r — 1) (&—1) d.f. If F is not
significant the data do not suggest that the treatment effects are different.
When F is significant, we conclude that the treatment effects are different.
We may then be interested to either compare the treatments in pairs or
evaluate special contrasts depending on the objectives of the experiments.
This can be done as follows.
Let 2 Utt denote a general contrast among the treatment effects.
i
Estimate of 2 hti is given by
I
Z liu— Z hyt
i t
where yt=Ti/r,
nz hu)=7 z 42
i I
where a2 is estimated by the error mean squares s2 (in Table 2.2).
The critical difference for testing the significance of estimate of the
differences like — tm is obtained from
C.D. — fl—a/2, (r—lHk-1) J i2s2
56 DESIGN AND ANALYSIS OF EXPERIMENTS
It was pointed out earlier that the randomized block designs are
improvements over completely randomized designs in the sense that they
provide error control measures for the elimination of block variation.
This principle can be extended further to improve randomized block
designs by eliminating more sources of variation. Latin square design is
one such improved design with provision for the elimination of two
sources of variation. This design is discussed below.
Let there be k treatments each replicated k times so that the total
number of experimental units required is k2.
Let P and Q denote two factors whose variabilities are to be eliminated
from the experimental error by having a suitable design. Evidently, both
these factors should be related to the variate under study so that their
variability may influence the variability of the variate under study. These
two are actually the controlled factors. Each of the factors P and Q is
COMPLETE BLOCK DESIGNS 57
Levels of Q Levels of P
Pi P2 Pz Pi P5
<h A B C D E
?2 D E A B C
#3 B C D E A
E A B C D
<li
C D E A B
(7 5
58 DESIGN AND ANALYSIS OF EXPERIMENTS
It has been pointed out earlier that this design is effective if the two-
factors P and Q can cause variability in the variate under study. If one
of the factors does not have substantial influence on the variate under
study, elimination of its variance may not reduce the error variance. In
such a situation a latin square design is no improvement over the rando¬
mized block designs. So, unless it is known that both the factors cause
sufficient variation in the variate under study, it is better to adopt a
randomized block design. In agricultural experiments if there is soil
fertility in two mutually perpendicular directions, then the adoption of a
latin square design with rows and columns along the directions of fertility
gradients, proves useful.
2.3.1 Randomization
TABLE 2.3
A 5 X 5 Latin Square Written Systematically
A B C D E
B C D E A
C D E A B
D E A B C
E A B C D
For the purpose of randomization rows and columns of the square are
rearranged at random. Both the rows and columns are, therefore,
numbered separately. Keeping the first row as such the remaining
rows are rearranged and numbered. If the latin square is of order r, then
random numbers less than or equal to r—1 are selected by consulting-a
random number table or otherwise as indicated earlier. If the first number
chosen is nlr then the «xth row of the initial square is written as the second
row. If n2 is the second number and not equal to nv then the n2th row is
written as the third row. This procedure is continued till all the rows of
the initial square are placed at random to form another square. After
row randomization is over, the columns of the row randomized squares
COMPLETE BLOCK DESIGNS 59
Next, the columns of the above row randomized square have been
rearranged randomly to give the following random square:
E B C A D
A C D B E
D A B E C
C E A D B
B D E C A
In latin square designs there are three factors. These are the factors P, Q
and treatments. The data collected from this design are, therefore,
analyzed as a three-way classified data. Actually, there should have been
k3 observations as there are three factors each at k levels. But because of
the particular allocation of treatments to the cells, there is only one
observation per cell instead of k in the usual three-way classified orthogonal
data. As a result we can obtain only the sums of squares due to each of
the three factors and error sum of squares. None of the interaction sums
of squares of the factors can be obtained. Accordingly, we take the
model
Let the data be arranged first in a row X column table such that
denotes the observation of (z, y)th cell of the table.
G2
Correction factor, C.F.=-^.
r2
Treatment sum of squares= £ ~j^—C.F.
Rt2 /
Row sum of squares= ]T —C.F.
C/2
Column sum of squares = X —C.F.
TABLE 2.4
Ri2
Rows k—1
Cj 2
Columns k—\ jT-C-F.
T. 2
Treatment k-\ St2
52
r—1
Error By subtraction
1
A latin square design is not normally used when the number of treatments
is eight or more, because a design of higher order requires too many
replications and it may not also be possible to get the required type of
experimental units.
Again, if the number of treatments is very small, then also there are
difficulties for adopting a latin square design. When there are two treat¬
ments a latin square design cannot be adopted, because from a 2x2 latin
square design error variance is not estimable. For a 3x3 latin square
design the degrees of freedom for error sum of squares is 2 and for a 4 x 4
latin square it is 6. In both these cases the degrees of freedom for error
sum of squares are two small. To make the design more effective in such
cases, the latin square design may be repeated, that is, instead of taking
one latin square, a number of, say, P latin squares each of the same order,
is taken for the experiment. The treatments are the same in each square
but each has a separate set of units and a separate randomization. The
data obtained from such repeated latin squares are analyzed as below.
Each of the P latin squares is first analyzed separately by following the
method described above. The corresponding sums of squares obtained
from the different squares are then added. This gives the pooled row,
column, treatment and error sums of squares. The pooled row sum of
squares is also called ‘between rows within square sum of squares’ and
similarly for the other pooled sums of squares.
From each of the p squares, the k treatment totals are obtained. With
these totals apxk table of squares X treatment is prepared.
and Tis denote the total of all observations of 5th treatment in z'th square.
The square X treatment table is formed with these Tia totals. The totals
Pi and Ts are the marginal totals of the Square X Treatment Table.
Next, the following sums of squares are obtained:
where yijst denotes the observation from tth square in its ith row, jth
column and under 5th treatment. The following partitioning of the
degrees of freedom of analysis of variance then follows:
TABLE 2.5
Partitioning of Degrees of Freedom in the Analysis of Variance of
Repeated Latin Squares
Squares P—1
Treatments k—1
Total pk*-l
The testing and other procedures are the same as in ordinary latin square
designs.
We have seen that by using latin square designs treatment effects can be
estimated by eliminating two sources of variation. This technique can
be extended further to eliminate more sources of variation. A Graeco
latin square is one such design where three sources of variation can be
eliminated. This design can be defined formally as below.
An orthogonal design by which three sources of variation can be
eliminated by using only k1 units where k is the number of treatments is
called a k X k Graeco latin square design.
A design is said to be orthogonal when the data obtained from it are
orthogonal. When all the levels of a controlled factor occur with each level
of any other controlled factor in a design, the data obtained from such a
design are always orthogonal.
In order to make the data from aGraeco latin square orthogonal, we
have to take another controlled factor R at k levels in addition to the two
controlled factors P and Q introduced for the latin square design, such
that each level of R occurs with each level of P and also of Q only once.
If treatments in an ordinary latin square design are taken to be levels of
the factor R, then such an arrangement is provided by the latin square
design. In that case we have to allocate treatments to the units such that
each treatment occurs once with each level of the controlled factors P, Q
COMPLETE BLOCK DESIGNS 63
and R in the k2 units. For example, let us replace the five treatments
symbols A, B, C, D and E in the example of the latin square design written
systematically by the five Greek letters a, (3, y, § and 0 respectively. These
letters represent the five levels of the factor R. We then get the following
arrangement of five levels of each of three factors in 52 units:
Levels of P
a (3 y S 0
(3 y 8 0 a
Levels of Q Y 8 9 a (3
8 0 a (3 y
0 a ,3 y 8
Now, five treatments denoted by A, B, C, D and E are to be allotted to
the above 25 units such that each treatment occurs once in each row, each
column and with each of the Greek letters.
It is not easy to allot the treatments to the units so as to satisfy the
above requirements of a Graeco latin square design. The problem can,
however, be solved easily by exploiting the properties of orthogonal latin
squares described below.
Two latin squares each of the same order, say, r are said to be orthogonal
if when one is superimposed on the other each symbol of one falls on each
symbol of the other once and only once.
For example, the following two latin squares are orthogonal.
Square I Square II
A B C a P Y
B C A Y a (3
C A B P Y «
We shall discuss the methods of construction of orthogonal latin squares
in Chapter 6. Orthogonal latin squares are also available in the Tables
for Statisticians and Biometricians by Fisher and Yates (1948).
For construction of a Graeco latin square design we have to take two
orthogonal latin squares each of order k. One of these squares is written
by using Latin letters and the other by using greek letters. By super¬
imposing one square over another and by treating the Greek letters as the
levels of the third controlled factor R, we get the Graeco latin square design.
The rows and columns of the square denote levels of the other two
controlled factors.
DESIGN AND ANALYSIS OF EXPERIMENTS
64
For example, the following is a graeco latin square with five treatments.
aA YD 8C 6B
YA 8E 6Z) aC
ye 8B 6A aE
9E aD pc YB SA
It has been pointed out earlier that sometimes observations from experiments
get lost due to unavoidable circumstances. Consequently, the data become
non-orthogonal and they cannot be analyzed according to the method of
analysis suggested for the design. Starting with Yates (1937) several
attempts have been made to simplify the method of analysis of experiments
with missing observations avoiding the complication of analysis of non-
orthogonal data. Some of these methods are discussed below.
Let
where the totals Tu Tm, Bj, By and G have the same definitions as in
Section 2.4.1 Tp is the observation total of pth treatment. By following
the method of analysis of non-orthogonal data as discussed in Chapter 1,
the reduced normal equations for estimating treatment effects are as follows
(^denotes the estimate of 1/ for all i):
r—2 2
(r k k-l
COMPLETE BLOCK DESIGNS 67
- r-2- ^
~jr~ tm—Qi
H-rraM- k ' k ^r)?r'-r-
k ~^k \)^tp
rJr+li±^=Qr
(r-l)(k—l)ti+tm=(k—l) Qt
(r— 1){k— \)Tm+7t=(k—1) Qm.
The solutions of these equations are
7>=T-r>-0^-"l)+l} (P=2.k’P*U
k—\ J Ql~hQm Qi Qm
2 1Xr-l)(fc-l)+l ! (r—\)(k— 1)—
k-1 1f Qi^rQm Qi-Qm
' 2 |[(r-l)(/c-l)+l (r-1)^-1)-
The adjusted treatment sum of squares is obtained from
2 t pQp-\-imQm-\-tiQi
pj*i, m
The unadjusted block sum of squares is obtained from
y B? B?+B,,2
TABLE 2.6
Source of
d.f. s.s. m.s. F
variation
Similarly,
T„- -x ,fl , r(fc-l)+2 Ar—1 1
6)-ci [_7+2r{(r-l)(A:-l)+l}+2{(r-l)(Ar-l)-l>J
There are two other possible ways of two missing observations in a design,
that is, (i) when they are both in the same block and (ii) when they are
the two replicates of the same treatment. The analyses of all these
cases together with the variance of estimates of treatment contrasts
following this method have been given by Das (1954).
treatment and column are non-orthogonal while the row X treatment and
row X column classifications are orthogonal. The treatment sum o
squares has, therefore, to be adjusted for the columns by following the
method of analysis of non-orthogonal data. The reduced normal equations
for this purpose are as follows. Let n denote the effect of ith treatment
(i = l, 2,..., r) and Qi denote the adjusted total of ith treatment and is
given by Qi = Ti-Si where S( is the sum of the averages of those columns
in which ith treatment is not missing and Tt is the total of ith treatment.
r (r-l)Qi
U~r(r-2) ‘
The analysis of variance results are then obtained as below. Let i?„ Cj,
etc. denote respectively the row and column totals and C.F. = G2/r(r 1)
where G is the grand total.
TABLE 2.7
*<2
Rows r—2 S —-C.F.
i r
Cj2
Columns r-1 S ——C.F.
J r
(r-D ^ ^ „ st2ls2
Treatments r-1 r (r—2) * Qs **2
The variance of the estimate of the difference between any two treatments is
obtained as below:
R==i- 1r 1+
—
k
L
rk
and
M TiS- Bt G
M,-7+k'7k
The missing values are estimated by minimizing this error s.s. with respect
to the xi’s. The equations, obtained by partial differentiation of the error
sum of squares with respect to xt, are
rkRxt+ 2 Xj—rkMi (i=l, 2(2.2)
R2x?+2 2
i i>j rlc
say, is the error sum of squares obtained by putting the existing
observations as zero and the missing observations as xt (i= 1, 2
Eliminating Mi from (2.1) with the help of (2.2), we get
=C.-CX.
Though the result (2.3) is obtained for the particular case described in
72 DESIGN AND ANALYSIS OF EXPERIMENTS
the beginning of this section, it is true in general for all designs including
the incomplete block and latin square designs even when the observations
are missing in a general fashion.
Adding the n equations of (2.2) we get
(rkR—1)2 X/+/J 2 Xi—rk'Z Mi
/ i i
that is,
rk'ZMi
^]Xl~{rkR—\+n)‘
From (2.2) we have
{rkR-\)xi+'Zxi=rkM, (2.4)
where i?!=(A:—l)/k and Xj' (*=1, 2,...,«) denote the unknowns substi¬
tuted for the missing values. Obviously, these are different from the xt s
(because of ignoring the treatment classification).
C3 is free from xx (i=l, 2,..., ri) and is actually (Treatment s.s.+
Error s.s.) obtained by putting zero for the missing values. We put
C3=(T.s.s.)0+(E.s.s.)0 where (T.s.s.)0 denotes the treatment s.s. obtained
by putting zero for the missing values. Similarly, (E. s. s.)0 is defined.
The equations for obtaining xt (i=l, 2,... ,n) are
RlXl -k
R y'—Bi
or,
2»-- •’ n)-
From (2.3)
E~C%—CX
=(E.s.s.)0— Cx, since C,=(E.s.s.)0.
The correct treatment s. s.=EX—E
X W
=(T.s.s.)0- *k (A:_1)+C*
Though the same symbols R and Mi have been used for both randomized
block designs and latin square designs, there is no chance of confusion.
The equations for estimating Xi (z= 1, 2,. .., n) are as below:
f 22Mi |
(2.9)
X,=F^3{ Mr~r3-3r + 2/z J (,==1, 2>'' *’ ")
£=C2'—2 XtMi
( 2 (2 M,Y \
= Ca'-^3UW'a-7C3^iJ
By following a procedure similar to that for the randomized block designs
the correct treatment s.s. can be obtained as below:
/ (2 Nif \
Correct treatment s.s
2 a Mi? \
—3{^^-3r+2n
' )
where
N, =r — (* = 1, 2,.. .,ri).
The formulae for the variances of estimates of different types of treatment
contrasts have been given in Section 2.4.4 for this general case.
2.5 An Illustration
TABLE 2.6
1 27 12 X(0) —11 30 16 17
2 21 6 20 9 14 0 15
4 13 -2 8 —3 7 —7 —12
5 22 7 5 —6 6 —8 —7
6 23 8 18 7 20 6 21
7 11 —4 23 12 18 4 12
8 3 —12 4 —7 5 —9 —28
9 11 —4 20 9 18 4 9
10 19 4 11 0 19 5 9
Mi M2
rkR—\ +m
=3.47
30/50 59 \
JCa—17 V 3 3x19/
=27.59
2x3.47x27.59
=0.6 (3.472+27.592M 30
=470.16
B *4-B *
Corrected T,.s.s. =(Tr.s.$.\-1 * +CX
= 1178-384.44+470.16
= 1263.72
TABLE 2.7
- - 2a2
v (t2—tu) = — = 16.67 (non affected)
=24.99
\3^6x 19^2x17/
=21.30 (one affected)
EXERCISES
1. When there are n missing observations in a randomized block design such that
each observation is missing in a different block affecting each time a different
treatment, find the variances of the estimates of the missing observations obtained
by minimizing the error variance. Find also the covariance between any two
estimates.
COMPLETE BLOCK DESIGNS 77
2. An experiment was conducted in two different places adopting the same latin
square design in each place. Give a method of combined analysis of the data
collected from the two places, assuming that the error variances are the same in
both the places.
3. Derive the method of analysis of a latin square design in which a row and a
column are missing.
4. In a randomized block design with k treatments and r replications r>k, a
replication of ith treatment is missing in ith block (i=l, 2, ..., k). Derive the
expressions for estimates of the missing observations and hence give the
computational procedure for the analysis of the data.
5. An experiment was conduted at Indian Agricultural Research Institute, New Delhi
during rabi 1970-71 under irrigated condition to determine the optimum dose of
ammonium sulphate for Kalyan Sona variety of wheat. Details of the experiment
are given below:
(i) Design—Latin square design
(ii) Treatment—A— 0 kg N per hectare
B— 40 99
C— 80 99
D—120 99
E—160 99
F—200 99
REFERENCES
1. Das, M.N. (1954). “Missing plots and randomized block designs with balanced
incompleteness”. Jour. Ind. Soc. Agri. Stat. 6, 58-76.
2. Das, M.N. (1955). “Latin squares with several missing observations”. Jour. Ind.
Soc. Agri. Stat. 7, 46-56.
3. Duncan, D.B. (1953). “Multiple range and multiple F-tests”. Mimeo. Tech,
Report no. 6, Va. Polytechnic Inst.
4. Fisher, R.A. and Yates, F. (1948). Statistical Tables for Statisticians and
Biometricians. Hafner Publishing Co. New York.
5. Giri, N.C. (1963). “A note on the combined analysis of Youden squares and
latin square designs with some common treatments”. Biometrics 19.
6. Yates, F. (1933). “Analysis of replicated experiments when the field results are
incomplete”. Emp. J. Exp. Agri 1, 235-44.
7. Yates, F. (1936). “Incomplete latin squares”. Jour. Agri. Sci. 26,301-15.
CHAPTER 3
Factorial Experiments
TABLE 3.1
Replications r—1
Treatments 3
AB 1 (Tn0-T01-Tl0+Tn)ZAr
Error 3 (r—1)
Total 4/-1
Combinations Contrasts
A B AB
a0b0 - - +
a0b i - + -
a,b0 + - -
aA + + +
The estimate of the main effect of A is the contrast among the treatment
totals: -T00-2’01+ri0+T11 where To0, etc. denote the observation totals
of the treatments. This contrast is evidently a comparison between two
group totals. In one of these groups, the factor A occurs at level a0 and
in the other at level ax. The two levels of B are evenly distributed in each
of these groups. The main effect of B is also a similar contrast. The AB
interaction is the difference between two contrasts. The first of these
contrasts is the difference between the first two combinations in the above
table and gives the effect of B at a0 level. The second contrast between
the last two combinations, gives the effect of B at ax level. So, the interac¬
tion component indicates if the effect of B is the same or different at the
two levels of A. If the effects of B are different at the two levels of ,4
the two factors are said to have interaction.
The s.s. due to each of A, B and AB is obtained by first obtaining the
contrast from the treatment totals and then dividing its square by the
total number of observations in the experiment. As the three contrasts
are mutually orthogonal, we can get the treatment s.s. from the total of
their sums of squares.
82 DESIGN AND ANALYSIS OF EXPERIMENTS
We can get the seven mutually orthogonal contrasts representing the main
effects and interactions in 23 factorial from Table 3.2.
TABLE 3.2
Main Effect and Interaction Contrasts in 2s Factorials
OoboCo — — — + + + —
aab0cx — — + + — — +
<*oVo — + — — + — -f
&obxcx — 4* + — — + —
axbQc0 + — — — — + +
axb0cx + — + — + — —
axbxc0 + + — + — — —
axbxcx + + + + + + +
The elements of mod. 2, that is, 0 and 1 are taken as the level codes of
each factor at two levels. The different treatment combinations are then
formed by taking one level from each factor. When there are n factors,
the total number of treatment combinations is 2". We shall introduce the
various procedures for obtaining the interaction groups and their contrasts
with reference to the 23 factorial.
The treatment combinations of the 23 factorial are written as 000, 001,
010, 011, 100, 101, 110, 111. Let us now define a variate *i(i=1, 2, 3)
such that it takes as its values the level codes of the z'th factor in the
different factorial combinations. Though this variate is defined with
reference to 28 factorial, it gets generalized immediately.
We write the following pairs of linear equations using the elements,
0 and 1 as coefficients:
1 *x,+0-x2+0-a:8=0|
O-^+l'Xa+O-Xa^Oj
O'Xa+Oxj+l-Xa^Oj
l-xx + l'x2+0-x3=0|
1 -Xj+O-^-fl 'x3=0|
86 DESIGN AND ANALYSIS OF EXPERIMENTS
O-Xi+l-Xj + l-Xg^Oj
3.6 Confounding
Example: We can take for 24 factorial two blocks each of eight units
and divide the 16 treatment combinations into two groups of eight each
for allotment to the two blocks.
In general, the block size for 2n factorials is of the form 2r. There are
many ways of grouping the treatments into as many groups as the number
of blocks per replication. We have already seen that for obtaining an
interaction contrast the treatment combinations are divided into two
groups. Two such groups representing a suitable interaction, say, P, can
be taken to form the contents of two blocks, each containing half the total
number of treatments.
In such cases the contrast of the interaction Pand the contrast between
the two block totals are given by the same function. They are, therefore,
mixed up with block effects and cannot be separated. In other words, the
interaction P, has been confounded with the blocks. Evidently, P has
been lost but the other interactions and main effects can now be estimated
with better precision because of reduced block size. This device of
reducing the block size by making one or more interaction contrasts
identical with block contrasts, is known as confounding. Though we have
introduced the concept of confounding by having only two blocks, the
number of blocks can be any power of 2 as we shall see subsequently.
Preferably, only higher order interactions, that is, interactions with
three or more factors are confounded, because their loss is immaterial.
As an experimenter is more interested in main effects and two factor
interactions, these should not be confounded as far as possible.
The designs for such confounded factorials can be called incomplete
randomized block designs. The treatment groups are first allotted at random
to the different blocks. The treatments allotted to a block are then
distributed at random to its different units.
Example: We can get the two groups of eight treatments each for 24
factorial from the solutions of the equations:
* +* +x3+xi=0"]
1 2
= 1J
Block 1 Block 2
0 0 0 0 0 0 0 1
0 0 1 1 0 0 1 0
0 1 0 1 0 1 0 0
0 1 1 0 0 1 1 1
1 0 0 1 1 0 0 0
1 0 1 0 1 0 1 1
1 1 0 0 1 1 0 1
1 1 1 1 1 1 1 0
Block 1 Block 2
0 0 0 0 0 0 1 0
0 1 1 1 0 1 0 1
1 1 0 0 1 1 1 0
1 0 1 0 1 0 0 0
1 6 1 1 1 0 0 1
1 1 0 1 1 1 1 1
0 1 1 0 0 1 0 0
0 0 0 1 0 0 1 1
We notice that one of the two equations for getting the two block
contents is homogeneous. So, all linear conbinations of the solution vectors
of the homogeneous equation are also its solutions. By obtaining three
independent solutions, all the other solutions can be obtained from their
FACTORIAL EXPERIMENTS 89
Example: We get the key block of size 23in 25 factorial from the solutions
of the equations:
Xj+Xj+Xg^O
*l+*4+*5 = 0
These equations indicate that ABC and ADE interactions are confounded
simultaneously in the same replication.
Now, any solution of the above two homogeneous equations is also a
solution of the equation which is obtained from a linear combination of
the equations. This shows that some more interaction is also confounded.
Adding the above equations we get
*2 + *3 + *4+*5 = 0
No other equation is possible from their linear combination. Hence, in
addition to interactions ABC and ADE, the interaction BCDE is also
confounded. This last interaction is called the generalized interaction of
the previous two independent interactions. The generalized interaction
BCDE can also be obtained from the product of the independent interactions
90 DESIGN AND ANALYSIS OF EXPERIMENTS
Key block
0 0 0 0 0
0 1111
10 10 1
10 110
110 10
110 0 1
0 0 0 1 1
0 110 0
There are three more blocks. These are obtained from the solutions of
the following three sets of equations. These equations have the same left
hand side, but the right hand side changes from equation to equation.
This ensures that the solutions of the three pairs of equations are non¬
overlapping. '
Sets of Equations
1 2 3
-^l+^2 + X3 — 1 0 1
Xl + X4+X5= 0 1 1
*J+*2+*3=l
*l + *4 + *5 = 0
Then, by adding (su s2,..., j5) in turn to all the treatments in the key
block, we get the block corresponding to the above equations.
Similarly, the two blocks for the other two equations are obtained.
All the four blocks forming the full replication have been obtained as
above and are presented in Table 3.3.
FACTORIAL EXPERIMENTS 91
TABLE 3.3
0 0 0 0 0 0 0 1 0 0 0 0 0 0 J 1 0 0 0 0
0 1 I 1 1 0 1 0 1 1 0 1 1 1 0 1 1 1 1 1
1 0 1 0 I 1 0 0 0 1 1 0 1 0 0 0 0 1 0 1
1 0 1 1 0 1 0 0 1 0 J 0 1 1 1 0 0 1 1 0
1 1 0 1 0 1 1 1 1 0 1 1 0 1 1 0 1 0 1 0
1 1 0 0 1 1 1 1 0 1 1 1 0 0 0 0 1 0 0 1
0 0 0 1 1 0 0 1 1 1 0 0 0 1 0 1 0 0 1 1
0 1 1 0 0 0 1 0 0 0 0 1 1 0 1 1 1 1 0 0
We have seen that when there are four blocks in a replication, three
interactions are confounded, two of the which are independent and one
generalized. When there are 2k blocks in a replication, 2*-l interactions
are confounded of which k are independent and the remaining 2k—k— 1
are generalized. We can have enough control over the choice of indepen¬
dent interactions in so far as their order is concerned, but we can have very
little control on the generalized interactions which are consequential. This
fact creates difficulty in the construction of confounded factorials so as to
save all lower order interactions including main effects. Lots of trial and
error in the choice of independent interactions become necessary to get the
desired type of confounding.
For example, in 27 confounded factorial using blocks of size 23, there
are 2* blocks. So, 15 interactions are confounded of which four are
independent and 11 generalized. If we are to confound only three factor
and higher order interactions, it is very difficult to get a set of four
independent interactions each of two or more order such that all their
generalized interactions'are of the desired order. Das (1964) gave a method
of construction of confounded factorials. This method mostly eliminates
the trial and error procedure. We have presented this method in a general
way covering factors with different numbers of levels after discussing
experiments with factors at three levels each in Section 3.8.
When factors are taken at three levels instead of two, the scope of the
experiment increases. It becomes more informative. Further, when the
levels of a factor are quantitative, the pattern of change of response with
92 DESIGN AND ANALYSIS OF EXPERIMENTS
the increase of the level values of the factor can be studied in a better way.
A study to investigate if the change is linear or quadratic is possible when
the factors are at three levels, though from this point of view the more
the number of levels the better. However, the number of levels of the
factors cannot be increased too much as the size of the experiment increases
too rapidly with them.
Let us begin with two factors A and B each at three levels. We shall
use 0, 1 and 2 as the level codes and treat them as the three elements of
mod. 3. In future, all mathematical treatments on these elements will be
in the finite field.
Taking one level from each factor we get the following treatment
combinations or simply treatments:
00, 01, 02, 10, 11, 12, 20, 21, 2±
A randomized block design can be adopted for the experiment. If there
are r replications, the analysis of variance partitioning for the design is as
given in Table 3.4.
TABLE 3.4
Analysis of Variance for 32 Factorial
Sources d.f.
Replication r—1
Treatments 8
A 2
B 2
AB 4
Error 8 (r—1)
Total 9 r—1
The letters a and b have been written with level codes to denote the
contrasts more clearly. The AlBl contrast is actually the contrast obtained
by multiplying algebraically (a2—a0) by (b2—b0), that is, a2b2—a2b0—a0b2+
a0b0. Similarly, the other contrasts are obtained. The estimate of the
contrast is obtained from the corresponding treatment totals.
Though this way of subdivision of the interaction effect is more
informative, such components cannot be used for the construction of 3"
factorials. For this purpose alternative types of components having some
sort of geometrical meaning, have been used. This procedure is described
below. Its development is due to Fisher, Yates and Bose.
Let us define, as earlier, one variate for each factor in a factorial such
that the ith variate xi denotes the levels of the z'th factor in the different
combinations of the factors. In the 3a factorial we have two variates xx
and x2 corresponding to A and B respectively.
Using the solutions of the equations:
DESIGN AND ANALYSIS OF EXPERIMENTS
94
=2 J
we get the main effect of B.
We can write two more such sets of equations as below:
*1 + *2==0
=1 ■
=2 .
*2+2*2=0 '
= 1 •
=2 .
Each of these sets gives three groups of solutions—one coming from each
equation in a set. Comparison among these three groups represents a two
factor interaction component. Thus, there are two 2-factor interaction
components each having 2 d.f. These two interaction contrasts are
orthogonal between them and also to those of the two main effects.
The interaction component obtained from the equations
*2 + *2 I (/=0, 1, 2)
is denoted by AB.
The component corresponding to the equations
*2+2*2=/ (i—0, 1, 2)
A 2
B 2
AB 2
AB2 2
Treatments 8
Components like AB, AB2 have great advantages for construction and
FACTORIAL EXPERIMENTS 95
analysis of these and other factorials, but they do not have any physical
meaning.
\
When there are three factors we get on the same analogy the following
main effects and interaction components each having 2 d. f. The equations
from which they are obtainable are also shown beside them.
TABLE 3.5
Main Effect and Interaction Components of 33 Factorial
A *1=*' 0=0, 1, 2)
B *2=»
C *3=*'
AB *l+*2 = *
AB2 •*1 + 2*2=/
AC *l + *3 = *
AC2 *1 + 2*3=/
BC *2+*3=*
BC2 *2 + 2*3=/
ABC *l + *2 + -*3 = I
ABC2 *1+ *2 + 2*3 = /
AB2C *l + 2*2+ *3 = /
The 26 d.f. due to 27 treatments are divided into the above 13 components
each having 2 d.f. A group contrast for any of the components is ortho¬
gonal to the group contrasts of any other component. The equations
having t non-zero coefficients of the variates give the groups of an inter¬
action component of the corresponding t factors in the general case.
Evidently, a randomized block design is not suitable for a 33 factorial
We have, therefore, to resort to confounding. The number of blocks per
replication in 3" factorials is a power of 3. In the present case we have three
blocks. The contents of these blocks can be obtained from the three treat¬
ment groups obtained from any of the last four equations which represent
three factor interactions. Here also, confounding can be complete or partial.
If we take four replications, each time confounding a different three factor
95 DESIGN AND ANALYSIS OF EXPERIMENTS
TABLE 3.6
Blocks of the 33 Factorial
0 0 0 0 0 1 0 0 2
0 1 2 0 1 0 0 1 1
1 0 2 1 0 0 1 0 1
1 1 1 1 1 2 1 1 0
1 2 0 1 2 1 1 2 2
0 2 1 0 2 2 0 2 0
2 0 1 2 0 2 2 0 0
2 2 2 2 2 0 2 2 1
2 1 0 2 1 1 2 1 2
When there are fpur factors we have to get nine blocks of nine plots each.
The 34 combinations have to be made into 32 groups. We have to use two
homogeneous equations simultaneously to get contents of the key block.
For example, we take the following two equations:
x1+x2-]-x8=0
+x4—0
This amounts to confounding ABC and AB2D. The solutions of these
equations give the key block contents. As these are homogeneous equa¬
tions any solution is also a solution of an equation obtained from their
linear combinations. Hence, two more interactions corresponding to the
following two resulting equations are also confounded:
FACTORIAL EXPERIMENTS
97
*1+2*3+2*4=0
*2 + 2*3+*4 = 0
Therefore, the interactions AC2D2 and BC2D are also confounded. These
two interactions are the generalized interactions. They can also be
obtained—one by multiplying ABC with ABZD and taking out the cubed
letters and another by multiplying ABC with (AB2D)a.
The other blocks are obtained from the solutions of the following eight
equations:
2 3 4 5 6 7 8 9
*l+*2 + *3=: 0 0 1 l 1 2 2 2
*1 + 2*j + *4 = 1 2 0 1 2 0 1 2
They can also be obtained from the key block by an exactly similar
procedure as described in 2" factorials. This method has, however, been
discussed in detail subsequently.
Here, again, by obtaining only two independent solutions of the
homogeneous equations, we can get the others from their linear
combinations.
In general, if in 3" factorial block size is 3r, the key block can be
obtained from the linear combinations of r independent solutions of n—r
homogeneous equations corresponding to confounded interactions. Again,
out of the (3n-r—1)/2 confounded interactions, (n—r) are independent
and $(3"-r—1)—(«—r) are generalized.
Following the above procedures the nine blocks obtained by confounding
the independent interactions, ABC and AB2D are shown in Table 3.7.
Here, also, a considerable amount of trial and error is required to obtain
a desired type of confounding. We have presented below, as stated earlier
also, a general technique of construction of confounded factorials which
eliminates considerably trial and error in the search of a desired type of
confounding.
Let there be n factors each at s levels where s=pm and p is a prime. The
total number of treatments in the factorial is sn. There are (i*— 1) d.f.
due to the treatments. These are divided into (sn—l)/(s— 1) main effect
and interaction components each of (s— 1) d.f. There are (s — l)*-1
components belonging to interaction among a given set of k factors.
The components can be represented by equations of the following
form:
98 DESIGN AND ANALYSIS OF EXPERIMENTS
cn fN o o T-H CN o
o\
u CN O CN o o pH cn T-H
o o CN
o
o T-, CN CN tH © T-H
o —H T-H CN o CN cn t-H o
«
00 pH _ cn cn © -1 o o CN
cn o CN o o cn tH *“
o
o O pH CN o fS cn pH © pH
s o pH pH CN ©
CN CnJ pH O
cn CN o o pH CN T-H T-H o
h- o
cn tH CN cn tH o o
u © ph fS O CN cn t-H o T-H
o O cn CN T-H
5
o pH CN ©
PH pH CN CN o T“H O o
VO
T-H CN pH cn cn o T-H o ©
o
o o P*H CN O cn CN T-H o
3 o pH T-H CN o CN cn ©
CN o CN O O t-H cn T-H tH
o
O o H CN o cn CN t-H o T-H
3 o CN O CN CN T-H ©
.
Tf o o CN o CN CN
© o PH PH cn o CN CN T-H O
cn CN O o T-H CN T-H o
mQ o pH © T-H T-H CN O CN CN
JD o pH CN o CN CN T-H o tH
3 o pH CN O <N fS T-H O
CN pH CN cn O T-H O o CN
o o pH O T-H T-H CN O CN CN
w
_o © — CN o cn CN T-H o pH
© o pH CN o CN CN TH
o
o T-H o P* ^H CN O CN CN
5 © tH CN O
0)>>
CN CN T-H O T-H
© T-H tH CN -© CN CN pH O
FACTORIAL EXPERIMENTS 99
0 0 0...1
All the treatment combinations of the r factors can be generated from
the Unear combinations of these r treatments. Next, to accommodate
more factors we go on annexing further columns, one by one, to the nghfi
100 DESIGN AND ANALYSIS OF EXPERIMENTS
Example: As the factors are each at four levels, we use as level codes the
four elements of the field 22:
0, 1, a, a8.
As the block size is 42 we start with the 2 x2 unit matrix shown in the
first two columns in Table 3.8. The last three columns are taken sub¬
sequently.
TABLE 3.8
Key Block of the 46 Factorial in 42 Plot Blocks
0 1 1 1 0 a2 a
1
<£ a2 1 1 a2 0
0 1
The last two rows of Table 3.8 give the two independent treatments
of the five factors in the key block. The interaction confounded due to
the column below A3 is
A1A2aA3
Similarly, the interactions confounded for the other two columns under
At and AB are . . .
A1AiaAl and A1A2A5.
The generalized interactions for the three derived columns shown in the
table are
A2A3A4, A*A*A3Ap
and Af-A3Af respectively.
The next 15 generalized interactions can be obtained from the other linear
combinations of the last three columns.
102 DESIGN AND ANALYSIS OF EXPERIMENTS
A\ A2 A9 A, ab
0 0 0 0 0
1 0 1 1 1
0 1 a a2 1
1 1 a2 a 0
1 a a 0 a2
1 a2 0 a2 a
The last three combinations are obtained from the first two, say, ax and a2
from ax -f pa2 by taking for p the values 1, a and a2 in turn. Another set of
five combinations is obtained by multiplying each of the last five treatments
by a and a third set of five, by a2. These give all the 16 treatments in the
key block.
The other blocks can be obtained by taking a treatment which has not
occurred in the key block and adding this treatment vector to each of the
treatments in the key block. To get the other blocks a similar operation
is done, each time taking a new treatment which has not occurred earlier
in any block. Sometimes it happens that a treatment which has already
occurred in some blocks is taken to get a block. This results in the
repetition of a block which is not always apparent as the treatments are
arranged in different order in the two blocks. This risk can be avoided
by following the method described below:
TABLE 3.9
Independent
confounded Confounding vectors Independent increment vectors
interactions
0 0 00 a a 1 000 0 0 10 0 a 1 0 0»
h h h h
AtA2*A3 1 a 1 0 0 (Ci) 0 0 0 0
Ax A2*' Ai 1 a2 0 1 0 (C2) 0 0 1 1
Ax A2 As 110 0 1 (C3) a a2 0 1
Example: When the block size is 23, we can accommodate seven factors
to save all main effects and two factor interactions. The independent
treatments in the key block are shown below:
A B C D E F G
10 0 110 1
0 10 10 11
0 0 10 111
This has been obtained by annexing four possible non-null columns to the
3x3 unit matrix under ABC.
In the example of 45 factorial in 42 plot blocks, discussed earlier, five is
the maximum number of factors for 42 plot blocks to save all main effects
and two factor interactions. Therefore, it is not possible to annex any
further column to the key block presented in Table 3.8.
When the factors are at two levels each and the block size is 2r, let us
choose columns containing only odd numbers of unity in all possible ways.
All these columns are distinct and the sum of any two cannot be another
of these columns, because the sum of two columns each containing an odd
number of unity, necessarily contains an even number of unity. So this
type of choice of columns saves three factor interactions also. The
maximum number of such columns is
rc1+rca+,'c#+ • • •
That is, the maximum number of factors each at two levels that can be
accommodated in a block of size 2r saving up to three factor interactions
is 2'-1.
Recently, Biyani (1973) in his unpublished thesis has shown that as
many as 5 • 2r-3 factors each at three levels can be accommodated in a block
FACTORIAL EXPERIMENTS
105
The procedure of getting the sum of squares due to the main effects and
interaction components is straight-forward though laborious. For each
interaction the treatment groups are obtained from the solutions of its
equations. Then, from the treatment totals of the observations, these group
totals are obtained. The s.s. due to these totals is then obtained by the
deviation square method.
In case of complete confounding the s.s. of the completely confounded
interactions are not obtained. In case of partial confounding, the s.s. of
a partially confounded interaction is obtained by using the observation
totals from replications in which it is not confounded.
For 2" factorial there is a very convenient and compact method of
analysis due to Yates (1937). It consists of first writing the treatment
combinations in a column beginning with the control treatment (0 0.. .0)
and then introducing the factors one by one as shown in Table 3.10
in the case of 24 factorial. A factor is introduced by a row which
contains the element 1 in one position only and 0 in all other positions.
This row is then added to the previous rows to get new treatments. When
all the previous rows have been added another factor is introduced. The
observation totals of the treatments are entered in the second column.
These totals are then added two by two and entered in the third column.
This way half the third column is filled up. The difference between the
same two totals which have been added (second minus the first) is now
obtained and entered in order to fill the rest of the third column.
Thereafter, the fourth column is obtained from the third column in exactly
the same way as the third column is obtained from the second. As many
columns are obtained in this way as the number of factors. In the present
case four such columns are obtained. The entries in the last column give
the main effect and interaction contrasts. The first entry in this column is
106 DESIGN AND ANALYSIS OF EXPERIMENTS
the grand total of the observations. If the treatment in z'th position of the
first column is (a/i, a/2... a/4), then zth contrast belongs to the interaction,
Jan ga-ii C<XI» D*lt.
The s.s. due to an individual contrast is obtained by squaring a contrast
and dividing it by the total number of observations in the experiment. If
an interaction is confounded, its s.s. is ignored. If an interaction is
partially confounded, its contrast is also obtained from those replications
in which it is confounded and then this contrast is subtracted from its
corresponding contrast in the table. This adjusted contrast is then squared
and divided by the number of observations in the replications in which it
is not confounded.
TABLE 3.10
Analysis of Variance of 24 Factorial by Yates’ Algorithm
Treatments Totals
Col. 1 Col. 2 Col. 3 Col. 4 Col. 5 Col. 6
0 0 0 0
10 0 0
0 10 0
110 0
0 0 10
10 10
0 110
1110
0 0 0 1
10 0 1
0 10 1
110 1
0 0 1 J
10 11
0 111
1111
This technique can be applied in 3" series as well. Here also, the
treatments are written by introducing the factors one by one as shown for
32 factorial:
FACTORIAL EXPERIMENTS 107
A B
0 0
1 0
2 0
0 1
1 1
2 1
0 2
1 2
2 2
In the next column the treatment totals are written. The operations on
the total column are as follows:
(i) The totals are added three by three and one third of the next
column is filled up with these sums.
(ii) The first total in each of the above triplets is subtracted from the
third and another one third of the next column is filled up with
such differences.
(iii) In each of the above triplets twice the second total is subtracted
from the sum of the first and third totals. These differences fill up
the remaining one third of the next column.
This procedure is continued as many times as the number of
factors. The last column gives the contrasts of their interactions of
the linear and quadratic components of the factors. If the treatment
in /th position is, say, (an a/2 a/3) the contrast in zth position gives
the interaction component, AXil BxlJ Caf8. When any at is zero, the
corresponding factor goes out of the interaction. The value 1 or 2
of a/ indicates respectively the linear or quadratic component of A.
The same rule holds for B and C as well. The s.s. due to a
contrast is obtained by dividing the square of the contrast by a
suitable divisor. If ,in the treatment in column 1 against a contrast
there are p unity and a two, the division for the contrast is
(2p+6a) r where r is the number of observations in each total
entered in column 2.
When there is a large number of factors even at two levels each, the total
number of treatment combinations becomes so large that it is very difficult
to organize an experiment involving these treatments. The demand on the
resources is so great that it may not, always, be possible for the experi-
108 DESIGN AND ANALYSIS OF EXPERIMENTS
Pnxi+Pizx2+ ■. • +Pi„*n=0
P2lXl'^~Pi2X2~^~ • • • fiPlnXn= 0
PkiX1-\~Pk2X2Jr . ■ . -rPknXn—0
P\\XlJrP\2X2JC • • • +7’lnXn = 0
Pkix1+pk2X2+... +pknX„=0
Pm\X^ -\-pm2X2~\- • • .-f•PmnXn — i (/— 0, 1, a2, a3, . .
Any contrast between these groups is not only a contrast for the inter-
action/>mlx1-Fpff,2^2+... •3rpmnXn—i but also for all interactions corres¬
ponding to equations which are obtained from its linear combinations
with the rest of the equations in the above set and their combinations. All
these interactions form the aliases of 2 pmjXj=i. This shows that all the
FACTORIAL EXPERIMENTS 109
The main effect of A is obtained from the difference between two group
totals, the groups being obtained from the solutions of the following two
sets of simultaneous equations:
*1 + + X» + + X + "I" =
*2 *4 5 *8 *7 01
*1 = 0 J
*i+Xt4-.x3+.x4+.X 5+.x8+x7=01
*i=l)
It is seen that any solution of the first set of equations is also a solution
of their linear combination which gives
x% "P *3+*4+x6 -f-x6+x7=0
As the block size is 23, we begin with the 3x3 unit matrix.
A B C D E F G
10 0 1 1 0 1
0 10 10 11
0 0 10 1 11
TABLE 3.11
Choice of Increment Vectors in £ (27) with 23 Units per Block
Confounded
interactions
with the Confounding
defining contrast vectors Increment vectors
0000011 0000101 0 0 0 10 0 1
ABD 1101000 0 0 1
ACE 1010100 0 1 0
BCF 0110010 1 0 0
ABCDEFG 1111111 0 0
0
FACTORIAL EXPERIMENTS
111
The rest of the increment vectors are obtained from the linear combinations
of the above vectors.
The full partitioning of the degrees of freedom for the analysis of the
design is shown below.
TABLE 3.12
Partitioning of d.f. for \ (2?) Factorial
d.f.
Sources
7
Main effects
(Aliases: 6factor interactions)
21
Two factor interactions
(Aliases: 5 factor interactions)
35
Three factor interactions
(Aliases: 4 factor interactions)
63
^rthree factor interactions, there are two types: ABC and A B C*.
112 DESIGN AND ANALYSIS OF EXPERIMENTS
The type with two squared letters does not arise because it is the square
of the second type above.
Similarly, for four and five letter interactions there are three types
each. For six factors there are four types. It is seen that the aliases of
four factor interactions with two squared letters are interactions of the
same type. Hence, confounding should be limited to them as far as possible-
The key block is obtained below by beginning with a 3 x3 unit matrix:
ABC D EF
1 0 0 1 1 0
0 1 0 1 2 2
0 0 1 2 1 2
The columns under D and E have been used for confounding and that
under F for keeping the block within the fraction. The element 2 has to
be taken once in each of the columns under D and E to get four factor
interactions with two squared letters.
The interactions confounded are shown below:
ABC2 D2—A2B2 EF—C2D2EF
A B2 C E2=A2 C2 D F=B2 D E2 F
B2 C D2 E=A C2 E2 F—A B2 D2 F
A D E=A2 BCD2 E2F=B CF
The other blocks can be obtained from the key block by increasing the
key block treatments with the following increment vectors.
Confounded
interactions and Confounding Independent
defining contrasts vectors increment vectors
000012 000102
ABC2 D2 1 1 2 2 0 0 0 2
AB2C E2 121020 2 0
ABCDEF 111111 0 0
• 6i means one letter in a six factor interaction is squared and similarly for others.
FACTORIAL EXPERIMENTS 113
(Aliases: 5X and 5J
Three factor, ABC type 20
(Aliases: 6, and 30)
Three factor, ABC2 type 120
(Aliases: 52 and 4j)
Four factor, A B C2 D2 type 30
(Aliases: 42 and 4S)
242
There are 40 d.f. due to all the interactions of type ABC, as there are
20 such combinations of three letters out of six. But since in the same alias
group there are two interactions of the same type, we get only ten such
alias groups. So, the total d.f. of the type is 20. Two of these d.f. are
confounded.
Given a set of four letters, say, ABCD, there are three components
each having two square letters. These are:
ABC2 D2
A B2 CD2
ABAC2 D
Hence, for q. given set of letters, there are six d.f. due to the interactions
of 42 type, that is, the type of four factor interactions with two squared
letters. As the total number of such combinations is 15, the total d.f. of
this type of interactions is 90. But since each of their aliases contains
interactions of the same typet the total d.f. of the type becomes 30. Six
of them are confounded.
The rest of the interaction which will contain the ignored factors also will
form aliases of the above interactions.
The fractions of 3" factorial can also be analyzed on the same analogy.
Here, also, while writing the treatments, factors are suppressed first and
then they are written by introducing the factors one by one as described
in full factorial. The operations and the correspondence of contrasts and
interactions are also similar when the non-suppressed factors alone are
considered. It is, however, not possible to write the aliases of such
interaction components, But this does not create any serious problem.
The linear and quadratic contrasts for a suppressed factor, L, come from
contrasts involving those factors with which L is in alias.
EXERCISES
1. Show that the total number of solutions of the following k equations is s”-fc
for every choice of (ax, a2- • •**):
TllXl +Pl$-2T ■ • ■ +/7l»Xn = al
/72lXl+/?22Xg+ • • • +/>2nXn = «2
110 1
10 11
0 111
The choices being Cx and C2 for the first replicate, C2 and C3 for the second,
C3 and C4 for the third and C4 and Cx for the fourth, the fifth replicate is
obtained from the three remaining interactions.
4. Given that the following three independent interactions ABCD, BEFK and
ABGK are confounded in 2® factorial having blocks of 26 plots, find the
independent treatments in the key block by applying the general method
discussed in the text.
Hints: Find three letters each occurring in a different confounded interaction
but no two or more of which occur together in any interaction. These can be
FACTORIAL EXPERIMENTS 115
C, F and G. Write all the letters in a row and then write five columns of the
5x5 unit matrix below the five letters other than C, Fand G. Put below C a
column which confounds A BCD. Similarly, put below F a column which
confounds BEFK and so on.
5. Given the following key block of 28 plot factorial in 28 blocks find the inter¬
actions confounded:
A B C D E F G K
0 110 110 1
1 1 0 0 0 0 1 1
1 11110 0 0
1 0 0 10 10 1
0 10 10 11 0
1 0 10 111 0
0 0 1110 1
0 0 0 0 0 0 0
Hint: Choose three rows from the above eight rows such that three columns
of these rows give three different columns of the 3x3 unit matrix. Then, with
the help of the remaining columns find the interactions confounded.
6. Find the confounded interactions separately for four replications of 34 factorial
in 32 plot blocks so that confounding is balanced for three factor interactions.
Hints: Take a 3x3 unit matrix. Annex two columns taken consecutively from
the following four columns:
112 2
12 2 1
7Write the independent treatments in the key block of size 33 of 313 factorial
such that no main effect and two factor interactions are confounded. Obtain
the independent interactions confounded.
8. Describe the method of analysis of 4» factorials by applying Yates’ algorithm.
9 . If the fraction 1/23 (2») is obtained from the confounding as shown in problem 4
and if the key block in problem 5 is the key block of the fraction, obtain the
three alias groups of interactions for the three confounded interactions.
10 Show that it is not possible to construct the half fraction of 2® factorial using
blocks of 23 plots and saving all main effects and two factor interactions.
11 The following table gives data relating to yields of dry herbage (cwt/acre) from
an experiment on grazed grass. Three fertilizers nitrogen at three levels,
phosphorus at two levels and potassium at two levels were used. Yields are
given for three years 1958-1960.
116 DESIGN AND ANALYSIS OF EXPERIMENTS
Years
1958 1959 1960
N P K
Analyze the data, assuming all factors to be fixed and give your comments.
12 The table below gives the yields and the layout of a 25 factorial experiment on
beans conducted in Rothamsted in 1935. A single replication in four blocks of
eight plots each was used. The five factors and their levels were as follows:
(s) spacing of rows at 18 inch or 24 inch
(d) dung at none or 10 tons per acre
(n) nitrochalk at none or 0.4 cwt N per acre
(p) super phosphate none or 0.6 cwt P205 per acre
(k) muriate of potash at none or 1.0 cwt KgO per acre
As usual, absence of a symbol denotes the lower level of the corresponding,
factor and (1) denotes the treatment with all the factors at the lower level.
Analyze the data, using higher-order interactions as error.
Block I Block II
13. Th© following table gives the plan and the yields of a 23 experiment involving
three factors N, P, K, in block size of four plots each.
Analyze the data and give your comments.
1 I 54 1 2 k 98
pk 90 P 182
pk 188 n 65
np 136 npk 201
2 1 122 3 4 P 156
k 65 nk 210
npk 145 pk 143
np 88 n 187
3 P 68 5 6 n 109
npk 73 k 93
I 70 pk 142
nk 103 np 154
14. A 33 experiment was conducted to study the effects of the three factors (each at
three levels 0,1, 2): sowing date (d), spacing of rows (s) and sulphate of ammonia (n)
on sugar beets. Two replications each consisting of three blocks of nine plots each
were used. The table below shows the plan and the percentage of sugar. Analyze
the data and write a report!
0 1 2 14.07 1 2 1 19.67
2 0 2 15.06 1 0 0 18.68
0 0 1 16.00 1 1 2 16.93
2 i 0 15.07 0 0 2 18.16
1 2 2 17.96 0 1 1 15.63
1 1 1 17.50 0 2 0 16.12
2 1 2 19.36
1 0 0 18.65
2 2 1 16.63 2 0 1 16.39
118 DESIGN AND ANALYSIS OF EXPERIMENTS
2 0 1 15.63 2 1 1 14 63
0 0 0 17.69 2 0 2 16.24
0 1 1 17. 50 1 2 2 16.39
0 2 2 19.58 0 1 2 16 54
2 2 2 16.59 1 0 2 14.66
2 0 0 15.96 2 1 2 16 24
1 0 1 16.74 2 0 0 16.72
2 1 1 19.66 1 1 1 16.40
1 1 2 16.38 0 1 0 16.63
1 2 0 15.66 1 2 0 17.30
0 0 2 18.56 2 2 1 16.86
0 2 ,1 17.06 0 0 1 16.72
15. Obtain the layout of a design for a 25 factorial in a 4x8 rectangle, confounding
ABD, ACE, BCDE with rows and AD, CE, ABD with columns.
16. Obtain the partially .confounded design for a 33 factorial with four replicates, each
in three blocks of nine plots each. Confound ABC with blocks in the first repli¬
cate, ABC2 in the second, AB2C in the third, and AB?C2 in the fourth.
REFERENCES
1. Biyani, S.H. (1973). “Maximum number of factors saving up three factor inter¬
action”. Unpublished thesis submitted to I.A.R.S., New Delhi.
2. Bose, R.C. (1947). “Mathematical theory of symmetrical factorial design.”
Sankhya, 8.
3. Carmichael, R.D. (1937). Introduction to Theory of Groups of Finite Order.
Ginn, Boston.
4. Cochran, W.G. and Cox, G.M. (1957). Experimental Designs, Wiley, New York.
5. Das, M.N. (1964). “A somewhat alternative approach for construction of
symmetrical designs and obtaining maximum number of factors.” Cal. Stat. Assoc.
Bull. 13.
6. Finney D.J. (1945). “The fractional replication of factorial arrangements.”
Ann. Eugen. 12.
7. Fisher, R.A. (1947). The Design of Experiments. 4th Ed., Oliver and Boyd.
Edinburgh.
8. Kshirsagar, A.M. (1966). “Balanced factorial designs”. J.R.S.S. (B), 28,
ppi 559-67.
FACTORIAL EXPERIMENTS 119
It has been mentioned earlier that factorial experiments are of two types,
symmetrical and asymmetrical. In symmetrical factorial experiments the
factors occur at the same number of levels. The designs for these
experiments have been discussed in Chapter 3. In asymmetrical factorial
experiments all the factors are not at the same number of levels. The designs
for such experiments are discussed in this chapter. The designs for asym¬
metrical factorial experiments will also be called asymmetrical factorials.
Symmetrical factorial experiments are somewhat inflexible because all
the factors have to be at the same number of levels. This may be in
conflict with the requirements of the experimenter. It may even be
unrealistic in some situations to take all the factors under investigation at
the same number of levels.
The above drawbacks can be removed by adopting asymmetrical factorials
as they accommodate factors with different numbers of levels. These designs
are, therefore, more flexible to meet the requirements of the experimenters.
Construction of confounded asymmetrical factorials and their analysis
are, however, somewhat complicated. As a matter of fact very little was
known about their construction through any systematic method till Kishen
and Srivastava (1959) and Das (1960) gave two convenient methods of their
construction. The previous works on asymmetrical factorials were due to
Yates (1937) and Li (1944) who gave several confounded designs which
were obtained individually rather than devising general methods for such
constructions.
Of the two methods given by Kishen and Srivastava (1959) and Das
(1960) it appears that the method of Das is more general, in the sense that
by using this method many more designs are obtainable than by the method
of Kishen and Srivastava. This is so because for the construction of designs
of the form in siXsk plot blocks, the method of Kishen and
Srivastava requires that ^ < 5, while the method of Das is not subjected
to any such restriction.
The simplest of the asymmetrical factorial experiments is the one with two
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 121
factors, say, A and B where A is at two Tevels and B is at three levels. This
is denoted by 2 x 3 factorial. Denoting the levels of A by a0 and ax and
those of B by b0, b1 and b2, the six treatments combinations of the factorial
are
^0^2’ ^1^1* ^1^2”
TABLE 4.1
Partitioning of the Total d.f. in the Analysis of Variance of 2x3 Factorial
Source d.f.
Block r— 1
Treatments 5
A 1
B 2
AB 2
Error 5 (r—1)
The sums of squares due to blocks, treatments and error can be obtained
exactly in the same way as in the analysis of randomized block designs.
The sums of squares due to the main effects A and B and their interaction
AB can be obtained by forming the following table with six treatment
totals, Tij, i=0, 1,7=0, 1, 2.
TABLE 4.2
Table of Treatment Totals of the 2 x 3 Factorial
Levels of A
oo *i Total
bi Tor Tii B1
b2 T02 T12 b2
Total Ag A\ G
122 DESIGN AND ANALYSIS OF EXPERIMENTS
A*'+A*_G*
Sum of squares due to A =
3r 6r
When there are three factors at the lowest possible numbers of levels,
viz., 3, 2 and 2, they generate 12 treatment combinations. In such
situations a randomized block design or a latin square design may not
always be suitable. It is, therefore, necessary to adopt confounding for
reducing the block size.
Normally, the block size in confounded asymmetrical factorials is a
factor of the total number of treatment combinations in the experiment.
This makes the design resolvable. A design is said to be resolvable when
it is possible to divide all its blocks into several groups such that each
group represents a full replication of all the treatments. Resolvability,
however, does not seem to serve any useful purpose in our case.
Confounding of effects in asymmetrical factorials is somewhat different
from what we have seen in symmetrical factorials. When an interaction is
confounded in a replication in these designs, it is not necessary that it is
completely confounded with the blocks in the sense that the block contrasts
and the interaction contrasts become identical. Though these two sets of
contrasts are not identical, they are dependent so that the usual contrasts
for obtaining a confounded interaction from the treatment totals is not
free from block effects. It is, therefore, necessary to take more than one
replication in confounded asymmetrical factorials so as to make the design
balanced, in the sense that, (i) any contrast of a confounded interaction is
estimable independently of any other contrast belonging to any other
confounded interaction and (ii) the loss of information of each degree of
freedom of any confounded interaction is the same. The difficulty of
construction of confounded asymmetrical factorials is mainly due to
achieving such a balance. We have discussed below a method of cons¬
truction of balanced confounded designs by linking them with the fractions
of suitable symmetrical factorials. This method is due to Das (1960).
TABLE 4.3
Block Contents of the Confounded Asymmetrical Factorial 3x2x2 in Six Plot Block.
(The Factors in the Treatment Combinations Appear in the Order Xly X2, A and B)
0 1 1 0 1 0 0 1 1 0 1 0 0 1 1 0 1 0
1 00 1 0 1 1 0 1 1 00 1 0 1 1 00
1 1 1 1 1 0 1 1 0 1 1 1 1 1 0 1 1 1
20 1 200 2 1 1 2 10 2 1 0 2 1 1
Factors
X A B
Levels 0 0 0 0 0
0 0 1 1 1
0 1 0
0 1 1
1 0 0
1 0 1
separately XxX%AB, X2AB and XxAB. The interactions which are con¬
founded here are XAB and. AB. We, thus, see that the two different choices
of interactions for confounding to obtain the initial replication have given
two different designs. One design requires four replications confounding
only XAB while the other requires three replications but confounding
XAB and AB. The block contents for the three replications corresponding
to the second choice are shown in Table 4.4.
TABLE 4.4
Block Contents of the Design 6x22 in 12 Plot Blocks in Three Replications. The
Factors in the Different Combinations Appear in the Order Xx, X2, A and B.
give the balanced design. The interactions confounded are XAB and AB
each with two degrees of freedom.
One more design is possible by choosing XxAB2 for confounding in the
first replication. This design is also balanced in two replications obtained
by confounding XxAB2 and Xx2 AB2. The interactions confounded are
XAB2 and AB2.
Combining these two designs a third design can be obtained in four
replications. In this design the interactions XAB, XAB2, AB and AB2 are
confounded. This design is balanced for the interaction of the factors
A and B.
(i) XxAB=AB=XxX2AB=X2AB
XAB AB XAB XAB
(ii) X2AC=XxX2AC=AC=X1AC
XAC XAC AC XAC
(iii) X1X2BC=X2BC=X1BC=>BC
XBC XBC XBC BC.
So we have to take other replications such that the interactions
XxX2AB, x2ab
XxX2AC, XxAC
XxBC, x2bc
occur equally frequently in the confounded interactions. Such a set of
confounded interactions is shown below:
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 131
Replacing the association la by 100 and 111 and so on for the other
associations and writing the transpose of the resultant matrix we get the
following seven blocks of the design:
Block
1 2 3 4 5 6 7
1 0 0 1 0 1 1 0 1 1 0 1 t 0 0 1 0 1 1 0 0
1 1 1 1 1 0 1 1 0 1 1 0 1 1 1 1 1 0 1 1 1
2 0 0 2 0 0 2 0 1 2 0 1 2 0 1 2 0 0 2 0 1
2 1 1 2 1 1 2 1 0 2 1 0 2 1 0 2 1 1 2 1 0
3 0 1 3 0 0 3 0 0 3 0 1 3 0 1 3 0 1 3 0 0
3 1 0 3 1 1 3 1 1 3 1 0 3 1 0 3 1 0 3 1 1
4 0 0 4 0 1 4 0 0 4 0 0 4 0 1 4 0 1 4 0 1
4 1 1 4 1 0 4 1 1 4 1 1 4 1 0 4 1 0 4 1 0
5 0 1 5 0 0 5 0 1 5 0 0 5 0 0 5 0 1 5 0 1
5 1 0 5 1 1 5 1 0 5 1 1 5 1 1 5 1 0 5 1 0
6 0 1 6 0 1 6 0 0 6 0 1 6 0 0 6 0 0 6 0 1
6 1 0 6 1 0 6 1 1 6 1 0 6 1 1 6 1 1 6 1 0
7 0 1 7 0 1 7 0 1 7 0 0 7 0 1 7 0 0 7 0 0
7 1 0 7 1 0 7 1 0 7 1 1 7 1 0 7 1 1 7 1 1
The other seveni blocks can be obtained from the above seven blocks by
replacing 00 combination of AB by 01; 11 by 10; 01 by 00 and 10 by 11.
Other incomplete block designs can also be utilized likewise to get
partially balanced asymmetrical factorials in the sense that though the
interactions AB and XAB can be estimated independently, the loss of
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 133
Example 10: The six blocks of the balanced 3x22 design in six plot
blocks are as shown:
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 135
The frequencies of p0, px and p2 defined above for the XAB interaction
in the different blocks of the above design are obtained below.
The R-treatments of the six observations variables in the first block
are in the order of the treatment combinations in that block p0, pQ, px, pu
—p2, —p2. Hence the frequencies of the R-treatments in this block are
as shown in Table 4.5. The frequencies of these treatments in the other
blocks have also been shown in Table 4.5.
TABLE 4.5
Frequencies of /^-Treatments for XAB in the 3 x 22 Design
/{-treatments
Blocks Po Pi P2
/
1 2 2 —2
2 -2 -2 2
. 3 2 -2 2
4 -2 2 -2
5 2 -2 -2
6 -2 2 2
which has the value p0, that is, the sum of the first two observations in the
block. Similarly, the total in the cell defined by block 1 and p2 is the sum
of the last two observations in the block multiplied by 1, as the co-efficient
of p2 value of each of these two observations is —1.
Both the cell frequencies and cell totals of the block X R-treatment
table are now obtained. These data in the table are now analyzed just
like the analysis of non-orthogonal data discussed in Chapter 2. The total
frequency of zth i?-treatment is taken as w/=2 | atjk |. This frequency is
also equal to P/k where P is the total number of observations in the design
being analyzed and k is the number of i?-treatments in the interaction
under consideration. The block size, n.j is the same as in the design. It
is not to be obtained by adding the frequencies in the block.
For finding the sums of squares of balanced interactions the problem
of solution of the reduced normal equations of the i?-treatments is very
much simplified as we shall see in the following examples.
Example 11: We shall analyze the 3 X 22 design in six plot blocks presented
above by the method described here. In this design the interactions XAB
and AB are confounded. The ^-treatments for XAB have been defined as
pt^iOQ+i Il-z01-il0(*=0, 1, 2)
The ^-treatments for AB are
m0=000+011 + 100+111 +200+211
and ,^=001 + 010+101 + 110+201+210.
The frequency tables for these two interactions are shown in Table 4.6.
TABLE 4.6
Frequency Tables of the jR-treatments of XAB and AB Interactions
Blocks 1 2 2 -2 4 2 6
2 -2 -2 2 2 4 6
3 2 -2 2 4 2 6
4 -2 2 -2 2 4 6
5 2 -2 -2 2 4 6
6 -2 2 2 4 2 6
Total n<. 12 12 12 18 18 36
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 137
(l2~) Pi+!(Po+/>t)-e*
The adjusted totals QPoetc. are found exactly as described in the analysis of
non-orthogonal data excepting that the marginal totals of the reconstrained
table are not taken as the block totals. The totals of the observations in
the blocks are taken as the block totals. The marginal jR-treatment totals
are, however, used as such for obtaining their adjusted total. These totals
are actually the algebraic sum of the requisite cell totals.
Here, QP(i=TP*-2f6 (B1-B.i-\-B3-BA+B5-B6) where Tp,is the margi¬
nal total of the p0 treatment obtained from the table and B1 etc. are the
block totals and not the marginal totals of the table. The other adjusted
totals for pr and p2 are ob+ained similarly.
The adjusted total Qm0—Tma —1/6 (4B1A-2B%+^B3-\-2Bi+2B6-\-4B6)
and 0Ml=Tmi-l/6(251+4J524-258+454+456+256) where Tm and Tm
are the marginal 7?-treatment totals.
The reduced normal equations for XAB can be written as
20 ,2y_
X Po+t
20 ,8v
y Pi "h g y Pi ~ Qpi
y 7^2+^ J Pi—Qpt
3 n
P-2 9Q Up*
xs /v. 2a2
Variance of (pn—pi)=20"*
T
138 DESIGN AND ANALYSIS OF EXPERIMENTS
20 5
Hence, the fraction of information recovered is
3xl2~9’
The solution of the normal equations for AB gives us
Qmo
°~ 16
m =Q”X
1 16
2a2
var (w0—wx)=
18
,16 8
Hence, information recovered=jg=^-
,s
,f VQpiY
Qpi2~±~ir~
s.s. (AB)=±2Qm,\
Since 2 Qmi=0, the correction factor has not been subtracted while
I
4 5.2 An Illustration
> o •n r- CN •n
CN
CN CN
& 1 i
1
.2 OO 00
*>> 00
cn
cn
cn
in
r-
VO
in
cn
vo
cn
oc rf
cn
cn
© © o
c 8 o o «—i
o « 8 o
V—
CN cn
O CN cn VO CN
> CN CN CN
a •8 1 1
pd 2
O 00 cn VC CN O
©O **">
VO vr> *n
cn CN r-
CN
The Treatment Combinations and Plot Yields of the 3x2* Experiment
8
cq
i—i
o 8 O
CN
©
i-H
CN
5
VO »T) ©V ON cn
> O
'O 7 7
.2 o in . %o VO
E VO
CN CN
VO
Tt Tt
©V
cn cn
cn
cn
,
o
o
JO O
o
>»—«
o 8 T_M
—
o
cn
O
i—«
cn
c P2
o
iT) cn CO cn cn VOi
ccJ >
O
T3 7 7
E
T3
tn
.2 cn
cn CN
VO
in
O
TT Tt
cn
TT <n|<
<N
, _
o o o
o *—« o O —■1
2 o O *—< CN CN
pq
OO in cn os o r-<
>
<t>
T3
1
7 •
2 o 00
2 OS CN vO t"-
Cn
cn r-
E CN CN Tt Vn CO cn cn
cn
CN
o o i— o L t
O © o o *—<
C2 2 o o ,“l
CN <N
.2 PP
aj VO On
o>>
CN O in OO
.2 i—<
1
E 1
pd I
.2 rt CN o *n 00 OO
T^- Tt
in Tt Tt
CN CO Tt
<N 0
bO
r— <3
(Tt
140 DESIGN AND ANALYSIS OF EXPERIMENTS
factors by 0 and 1. These codes have also been indicated above while
describing the levels of the factors. The design together with the yield of
hay in kg (1000 gram) per plot are presented in Table 4.7. The deviation
of each observation from its block average has also been presented in
this table.
These data are analyzed first by the method described in 4.5.1. The first
step in this method is to get the deviation of each observation from its
block average. These deviations have been presented along with the
observations. Next, a three-way table corresponding to the three factors
has been prepared. In each cell the totals of the deviations corresponding
to the different treatment combinations have been entered. There are
three observations in each cell. Hence each cell total is based on three
deviate values. The table is presented below:
TABLE 4.8
Three Way Table of the Experiment with the Deviate Totals in the Cells
0 1 2
0
-51 -56 19 0 8 13
(3) (3) (3) (3) (3) (3)
Levels of fall (F)
, -38 -44 47 5, 20 31
1 (3) (3) (3) (3) (3) (3)
s.s. (iVSF)=25.38-^=45.70
s.s. (SF)=21.78^|=24.50
TABLE 4.9
Analysis of Variance of the Experiment Obtained through Observation Deviate
Source of
variations d.f. s.s. m.s.
Blocks 5 2357.00
N 2 4549.50 2274.75**
S 1 2.78 2.78
F 1 498.78 498.78**
NS 2 47.39 23.69
NF 2 148.39 74.20
SF 1 24.50 24.50
The variations due to nitrogen and fall clipping effects are significant at
1 per cent level. The average values of the effects of the three levels of
nitrogen and two levels of fall clipping are given below.
N0 Nt Nt S.E. F0 S.E.
27.4 52.9 49.2 1.8 39.4 46.9 1.5
These averages show that application of nitrogen enhanced the yield very
much. The increase in yield due to the increase of dose of nitrogen
from 30 to 60 is not significant. Similarly, fall clipping has better effect
on the yield.
These data have now been analyzed to get the sums of squares due to
NSF and SF which are confounded by the alternative method discussed
in 4.5.2.
The i?-treatments for SNF are shown below
DESIGN AND ANALYSIS OF EXPIRIMENTS
142
^=000+011-010-001
^=100+111-110-101
^=200+211-210-201
130(4) 216 36 6
>
4
1
1
nt 12 12 12 18 18
Q«,=782-^ (2x240+4x222+4x228+2x216+4x300+2x348)
= 14
(/«1=772—^(4x240+2x222+2x228+4x216+2 x 300+4x348)
= -14
s.s. (tfSF)=^l2+23*+62-^=45.70
These two sums of squares are exactly the same as obtained earlier. The
sums of squares of other interactions which are not confounded are
obtained by the usual method. The unadjusted block sum of squares and
total sum of squares are obtained as earlier. The error sum of squares is
obtained by subtraction.
2 fjkfj'k
k
is constant for different choices of ^-treatments j and j' of the same zth
interaction, then the design is balanced in the sense that the loss of inform¬
ation of each degree of freedom of the interaction is the same. This
criterion, however, does not ensure that different confounded interactions
are estimable independently. For this purpose the following criterion
should be satisfied. Let
Uiip —'ZfijkZ'j'k •
k
If Uu, is constant for different choices of y'th and /th /^-treatments in the
zth and z'th confounded interactions respectively, then the zth and z'th
interactions are estimable independently.
For this purpose first a design with bigger plots is taken to accom¬
modate the factors which require bigger plots. Next, each of the bigger
plots is split into as many parts as the number of treatments coming
from the other factor. The bigger plots are called main plots and the
treatments allotted to them are called main plot treatments or simply main
treatments. The constituent parts of the main plots are called subplots
and the treatments allotted to them are called subplot treatments. The
different types of treatments are allotted at random to their respective
plots. Such designs are called split-plot designs.
For example, let there be three levels of irrigation prescribing three
different amounts of water per plot and four doses of nitrogen fertilizer.
First a randomized block design with a suitable plot size is taken with the
three levels of irrigation as treatments. Let there be, say, five replications
in this design. The irrigation treatments are then allotted at random to
each of the five blocks of three bigger plots each. Next, each of these
main plots is split into four parts or subplots to accommodate the four
levels of nitrogen. The 15 main plots actually serve as 15 replications
of the subplot treatments. The subplot treatments are allotted at random
to the subplots in each of the main plots.
The split-plot designs are thus a combination of two or more
randomized block designs (depending on the number of factors) such that
the plots of one design form the block of another design. It is not
necessary that a split-plot design has always two sizes of plots. The
above analogy can be extended to have three or more plot sizes to
accommodate different types of treatments. Though in the above
example, the main plot treatments are the levels of only one factor, such
treatments can also be the combinations of the levels of different factors
each of which require similar plot size. Similarly, the subplot treatments
can also be the combinations of the levels of different factors. Accordingly,
the device of confounding can be applied to reduce the block sizes of
each of the constituent randomized block designs.
4.7 Analysis
We shall discuss the analysis of simple split-plot designs with two sizes
of plots. It is further assumed that both the main plot and subplot
treatments are the levels of single factors. Further the design adopted
for the main plots is a randomized block design.
Let the number of main plot treatments be a and that of the subplot
be (3. Let further the number replications in the randomized block design
for the main treatments be r.
First the observations in the p subplots in each main plot are added.
This will give ar totals distributed in r blocks of a units each. These
totals are then analyzed by the method of analysis of variance appropriate
for the randomized block design excepting that each sum of squares has
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 145
Then the different sums of squares for the main plot analysis are shown
in Table 4.11.
TABLE 4.11
Anaiyis of Variance of Main Plot Totals
Rl_&
Replications r—1 7 ra£ . .
A,*__ G*
Main plot treatments a— 1
J Pr rap
-TV* R* yAl+&
Error (r_l)(a_l)
f, P t 3r rap
rw« C2
Total ra —1
5 e
For the analysis of subplot observations we treat the r« main plots as the
replications of the subplot treatments.
A two-way main plot treatment X subplot treatment is first made
entering the appropriate observation totals in each cell. There are a,3 cells
in this table and each cell contains the total of r observations coming
from the r replications of each combination of the two factors. An
analysis of variance of these totals gives the sum of squares due to subplot
treatments and interaction between subplot and main plot treatments in
addition to the sum of squares due to main plot treatments which has been
obtained in the previous analysis as well. The error sum of squares in the
subplot analysis is obtained by subtraction as all the sums of squares
146 design and analysis of experiments
including those in the main plot analysis add to the total sum of squares.
All these sums of squares are shown in Table 4.12.
TABLE 4.12
d.f. s.s.
Sources of variation
Tu 2 G2
Replication (main plot totals) ra—1
5 P
5fc2_ G2
Subplot treatments P-1 \ ret rap
(r—1)(P—D By subtraction
Error (subplot)
G2
rap—1
Total
It is seen that in the analysis of split-plot designs there are as many error
variances as the number of plot sizes—one coming from each plot size and
its associated randomization. In the present case the error for testing the
interaction sum of squares is the subplot error because the interaction
contrasts are functions of subplot contrasts.
The above method of analysis has been derived from the following
model;
Yijk—y.-\-ri-\-a.jJr$k+IjkJr etjk
where Yi]k is the observation variable corresponding to the observation
yiJk coming from the fcth subplot treatment of theyth main plot treatment
in the ith replication; p. is the general mean, r/, a/ and [3& are the fixed
effects of the ith replication, yth main treatment and kth subplot treatment
respectively: Ijk is interaction effect and etJk are the error components
which are assumed to be independently and normally distributed with zero
mean and constant variance a2. As the subplots in a main plot are close
together, being located in the same main plot, two error components
corresponding to two subplots in the same plot are taken to be correlated
with pas the correlation coefficient. This type of association of observations
in the same main plots do not affect the usual analysis presented above.
Assuming that E(eJkjz)=c" and E{eijk eijk.)^9o* the expected values of
the two error sum of squares can be obtained as below:
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 147
Let Ea denote the main plot error mean squares and Eb, the subplot
error mean squares.
Then the expected value of isa=<T2{l-f ([}—l)p}and expected value of
Eb=o2 (1 - p).
From these the estimates of p and a2 come out as
)Eb
P
—_ Ea—Eb
9~Ea+$-l)£b
Expectedly p is positive. Hence Ea is expected to be greater than Eb.
Case 1 The estimate of the effect of any main plot treatment is the
average of all the observations coming from the treatment. Hence, the
variance of the estimator of the difference between two main plot treatments
say a0—at is given by
. 2 Ea
V(a0-0!)=—
Case 2 Similarly, the variance of the estimator of the difference between two
O 17
subplot treatments, say, bl}—b1 is given by V (b0—b})=—-.
rJ.Y
Case 3 If a0, av etc., denote the main plot treatments and b0, bu etc.,
denote the subplot treatments, then the variance of the estimator of the
difference between two combinations like aibj and ajhji is given by
V&bj-^bj,)-2-- •
V(aibj—att bji)=~- •
This is because the two averages which estimate the above two combination
effects are independent being based on observations coming from different
main plots. Substituting the estimate of c2 we get
2{Ea+{$-\)Eb)
V (aibj—aj> bj>)-
When there are more than two sizes of the plots the same technique of
148 DESIGN AND ANALYSIS OF EXPERIMENTS
analysis can be extended in a straight forward way. We may assume that only
observations coming from the smallest plots in the next bigger plots are
correlated. This assumption seems to be more reasonable than assuming
an association between observations coming from two different plot sizes.
If the subplot or main plot treatments are combinations of the levels of
several factors, then the above technique of analysis can be extended in the
usual manner, excepting that the ((3— 1) degrees of freedom for the
subplot treatments have to be partitioned according to the factors whose level
combinations build up the p combinations. The same considerations hold
for the main treatments as well.
EXERCISES
1. Construct the confounded design 6x23 in 12 plot blocks and indicate the
confounded interactions. Outline the method of analysis of the design.
2. Construct the 3x22 design in three plot blocks. Indicate the interaction that are
confounded. Find the loss of information on the different confounded
interactions.
3. Examine if the following 5 x 22 design in 10 plot blocks and five replications fa
balanced.
la 2a 3P 43 5p
Ip 2a 3a 4P 53
IP 2P 3a 4a 53
IP 2P 3P 4a 4a
la 2p 3p 4P 5a
IP 2a 3P 4a 5a
la 2p 3a 43 5a
la 2a 3p 4a 53
IP 2a 3a 4P 5a
la 2p 3a 4a 53
Here, the combination /a means iOO and ill and ip means id and *10 (*=1, 2y
3, 4, 5).
4. Discuss why asymmetrical factorials in one replication are not considered suitable
designs:
5. Obtain first a balanced confounded factorial 5 x 3 in three plot blocks. Next,
associate each combination of the above design with the two levels 0 and 1 of a
third factor so that out of one combination of 5 x 3 factorial, we get two combi'
nations of the 5x3x2 factorial and the block size becomes six. Show that the
resultant 5x3x2 design in blocks of six plot and four replications, is balanced.
Derive the method of its analysis.
6. In an asymmetrical factorial with the factors X, A and B at levels 3, 2 and 2
respectively, the three levels of X are coded by the four combinations of two
factors Xi and X% each at two levels as below. The combination 00 is used to
code the first level of X, the combination 01, the second level and each of 10 and
11 is used to code the third level. This way the asymmetrical factorial is.
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 149
x
converted to a symmetrical factorial. Obtain the confounded factorial 3x2x2
in four plot blocks in one replication and outline the method of analysis of the
design. Indicate what interactions are confounded.
7. In a factorial with two factors each factor requires bigger plots. It is required
to study their interaction more accurately. Suggest a suitable design for the
experiment and indicate its method of analysis, (strip- plot designs)
8. To study the effect of application of different doses of nitrogen, phosphorus and
potash with different doses of farm-yard manure on cabbage an experiment was
conducted at the Indian Agricultural Research Institute, New Delhi during
1961-62. The design adopted was split-plot-cum-confounding. The details
of the experiment are given below:
I. Treatment
Main plot: All combinations of
Nitrogen Phosphorus Potash
$
o
00
O
^5
Na—150 P2-80
II
99 99
Subplot treatments:
Fi=2.5 tons/acre
F2—5 tons/acre
2. Numbers of replications—Two
3. Plot size: Gross: 17'xl5'
Net: 17'x 10'
The layout plan and yield of cabbage in (Oz) per plot are given below:
The lower figures indicate the yield and the upper ones the treatment combinations.
Analyze the data appropriately. The yields are below the treatments.
REPLICATION I
Block I Block II Block III
1001 1002 3121 3122 2002 2001
378 535 4428 803 855 315
2102 2101 2111 2112 1202 1201
1024 1122 622 744 626 368
1212 1211 1221 1222 2212 2211
867 510 418 599 658 354
1121 1122 3211 3212 1111 1112
REPLICATION II
Block I Block II Block III
2022 2021 1011 1012 3222 3221
337 506 338 195 346 293
3201 3202 2221 2222 3112 3111
591 598 867 396 235 294
2101 2102 2001 2002 2012 2011
951 425 442 426 171 97
3121 3122 2111 2112 3001 3002
894 393 368 638 192 258
mi 1112 3102 3101 1212 1211
470 516 300 127 263 343
2211 2212 1122 1121 1102 1101
572 482 379 68 479 521
3012 3011 1201 1202 1021 1022
464 253 381 306 369 453
1001 1002 3021 3022 2202 2201
386 160 58 191 850 707
1221 1222' 3212 3211 2121 2122
509 365 375 389 276 372
«o «1 «0 «i H «i no ni
104 105 117 129 12 3 123 105 137
vi *1 Vl n
n2 «3 «2 «3 «2 «3 «2 nz
112 146 153 139 151 164 129 143
«o «i «o «i n0 «i no «i
112 109 111 123 117 109 124 129
v2 v2 v3 V2
«2 «3 «2 «3 «2 n3 «2 «3
125 161 134 141 159 157 133 139
«o «i «o «i no «i no
116 119 119 132 102 116 135 143
v3 V3 V3 v3
'*2 «3 *2 n3 "2 n3 «i nz
121 159 143 149 167 161 142 158
ASYMMETRICAL FACTORIAL AND SPLIT-PLOT DESIGNS 151
REFERENCES
1. Banerjee, A.K. and Das, M.N. (1969). “On. a method of construction of con¬
founded asymmetrical factorial designs.” Cal. Stat. Asso. Bull. 18, pp. 163-77.
2. Chakrabarti, M.C. (1962). Mathematics of Design and Analysis of Experiments,
Asia Publishing House.
3. DAS, M.N. (1960). Fractional replicates as asymmetrical factorial designs. J.Ind.
Soc. Agri. Stat. 12.
4. Kishen, K. and Srivastava, J.N. (1959). “Mathematical theory of confounding
in asymmetrical and symmetrical factorial designs.” J. Ind. Soc. Agri. Stat. 11,
pp. 73-110.
5. Kishen, K. and Tyagi, B.N. (1964). “On the construction and analysis of some
balanced asymmetrical factorial designs.” Calcutta Statistical Association Bull.,
13, pp. 123-49.
6. Kshirsagar, A.M. (1966). “Balanced factorial designs.” Jour. Royal Stat .
Soc. B, 28, pp. 559-67.
7. Li, J.C.R. (1944). “Design and analysis of some confounded factorial experi¬
ments,” Res. Bull. 333, Iowa State College of Agriculture.
8. Muller, E.R. (1966). “Balanced confounding of factorial experiments.” Bio-
metrika, 53, pp. 507-24.
9. Nair, K.R. and Rao, C.R. (1948). “Confounding in asymmetrical factorial
experiments.” J. Roy. Stat. Soc. B. 10,109-31.
10. Raghavarao, D. (1971). Construction and Combinatorial Problems in Design of
Experiments, Wiley, N.Y.
11. Sardana, M.G. and Das, M.N. (1965). “On the construction and analysis of
some confounded asymmetrical factorial designs.” Biometrics 21, 948-56.
12. Shah, B.V. (1960). “On a 5x2^ factorial design.” Biometrics 16, 115-18.
13. Yates, F. (1935). “Complex experiments.” Sup pi. J. Roy. Stat. Soc. 2,
181-247.
14. Yates, F. (1937). “The design and analysis of factorial experiments.” Imperial
Bureau of Soil Science Tech. Comm. 35.
CHAPTER 5
the same. We shall, therefore, discuss the designs for all such experiments
in a general way. The term “treatment” will accordingly be used in
place of “variety”.
of the balanced and other incomplete block designs and was not necessarily
constrained by the considerations of practical utility.
The various constructional and analytical methods of the incomp e e-
block designs have been discussed separately in the subsequent sections.
It was found that the balanced incomplete block designs were not
always suitable for varietal trials because these designs require large
numbers of replications and further, suitable designs are not available tor
all numbers of treatments. To overcome such difficulties Yates (193 )
evolved a series of incomplete block designs which he called lattice designs.
Subsequently, Bose and Nair (1939) evolved another type of incomplete
block designs which they called partially balanced incomplete block
(P.B.I.B.) designs. . . ,
One type of lattice designs is available for numbers of varieties which
are perfect squares. These are, therefore, called square lattice designs.
These can accommodate a flexible number of replications beginning with
two. The number of replications of pair of treatments is one for certain
pairs and zero for others. These designs are actually a type of confounded
symmetrical factorial designs involving two factors as we shall see after¬
wards. Likewise, there are lattice designs corresponding to confounded
symmetrical factorial designs involving three factors. These are called
cubic lattice. As these designs fellow from factorial designs these are
also called quasi factorial designs. .
The partially balanced incomplete block designs are available with
smaller numbers of replications for many more numbers of treatments. The
number of replications of pairs of treatments in these designs is not
constant. These were defined in general for m types of replications of the
different pairs of treatments where m is any integer. That is, in these
designs some pairs of treatments can be replicated, say, \ times, some
others A2 times and so on up to m such type s of pairs. From considera¬
tions of analysis of these designs, they have to possess some more
properties as well. These are taken care of by defining some more
parameters which we shall discuss subsequently.
In order to make incomplete block designs available for all number of
treatments with smaller numbers of replications two more series of
incomplete block designs were defined. These are re-inforced incomplete
block designs (Das, 1958 and Giri, 1958) and circular designs (Das, 1958).
The re-inforced incomplete block designs are actually augmented in¬
complete block designs and are obtained by including in each block of
any usual incomplete block design a certain number of additional treat¬
ments, say, a so that the block size becomes k+a and the number of
treatments v+a. By suitably choosing v and a any numbers of treatments
can be accommodated in such designs.
Circular designs are a particular type of P.B.I.B. designs which are avail¬
able for any number of treatments. They have no constructional problems.
When the block size is 2 or 3, their analysis is also not complicated. For v
INCOMPLETE BLOCK DESIGNS
155
treatments and block size ti, these are obtained by developing the initial
block 1,2,3, n mod v.
Even before Yates used the B.I.B designs for agricultural experiments,
some of them were known in combinatorial mathematics in the form of
groups possessing special properties. Bose exploited the group properties
for the construction of these designs. Since then considerable amount
of work has been done in the name of incomplete block designs.
But, it seems, much of these activities has very little to do with designs
required for actual experiments. They are more curiosity guided than
problem oriented. The present activities have penetrated much into
combinatorial mathematics and are, therefore, more concerned with group
properties rather than problems of precision or availability of designs
to suit particular experiments.
We have so far tried to view the incomplete block designs from the
angle of precision as related to the replication of pairs of treatments. But
actually, analytical considerations have played a greater role in the
development of various incomplete block designs. From incomplete
block designs we get two-way data classified according to blocks and
treatments. If there are b blocks in a design with v treatments and k as
block size, then there are b xv cells in the two-way table with frequencies
0 or 1. Since k of the v treatments occur in a block, the frequencies in
the k corresponding cells in the row for the block, are unity and those in
the remaining cells of the row are zero. The data obtained from such
designs are, therefore, non-orthogonal. The b X v cell frequency table is
called the incidence matrix of the corresponding design, As the cells can
take two values 0 or 1 these designs are called binary designs.
While obtaining the method of analysis of non-orthogonal data in
Chapter 1, we have seen that the analysis involves the solution of as many
reduced normal equations as the number of treatments. In balanced
incomplete block designs, these equations take the form
V
,
So, taking the restriction 2 tm=0 the solution of these equations becomes
m
very easy.
While evolving new designs similar simplicity of the solutions of the
normal equations has also been kept in view along with the other
considerations discussed previously. A further consideration that is kept
in view is that all the treatment effects should be estimable as deviates
from their means. This in turn implies that the reduced normal equations
should be solvable taking one additional restriction among the treatment
effects. Usually, this criterion is called “connectedness” of the design.
If a design is not connected, all the (v— 1) mutually orthogonal contrasts
of the v treatments effects cannot be estimated. Hence, it is necessary to
know if a design is connected before it is recommended. Efforts have,.
156 DESIGN AND ANALYSIS OF EXPERIMENTS
=&(r—1)+&(k—1)(A—1)—&2(r—1)2/(6—1)
(5.6)
~(r—k) (r-A)(v-k)/(6—1)
after simplification subject to (5.1) and (5.2).
Being sum of squares of deviates, (5.6) is evidently zero or positive.
But since (r—A) and (v—k) are necessarily positive, (r—k) must be zero
or positive.
Since 6/v=r/A:, we get 6 > v.
When (5.6) is zero, all the x/s are equal. This also makes r=k or 6 = v.
A much simpler and more elegant proof of the result 6 ^ v which makes
use of the incidence matrix N of the design is as follows. It is easy to
observe that for a B.I.B. design N'N=(r—A)/,+A£W so that |N'N\—
(r+A(v—1)} (r—A)?-1=rAr(r—A)*-1 # 0. This means that N’N is non-
design and analysis of experiments
158
follows from the fact that (£_l) (r-*)>0 which implies ~-v > r-k
i q, b ^ v-\~k k.
The class of designs in which b = v is called symmetrical B.I.B. designs.
In these designs, therefore, the number of treatments common between
any two blocks is constant.
Since, 2xi=k(r— 1) the value of this constant is
7 '
*(r-l) r{k-1)_,
b~\- b-1 “ b-1
because in these designs r=k and b=v. It is found that many designs
where b is greater than v are impossible. For example, the design with
the parameters v =36, b = 42. r=l, k = 6 and A=1 is impossible. Many
more designs where b ^ v have since been proved impossible. A sufficient
condition for the availability of B.I.B. designs is yet to be found. The
designs which are impossible or unsolved have been pointed out in the
tables for such designs as indicated in the next paragraph.
The block size in this series of designs is 2. All non-repeated designs with
Jc=2, belong to this series. They are constructed easily when all possible
combinations of v treatments taken two at a time are used as blocks.
By repeating p times the b blocks of a B.I.B. design with parameters
v, b, r, k and A, another B.I.B. design with parameters v, pb, k, pr, p\ is
obtained. Such designs are called repeated designs.
The above series of designs with k=2, is suitable when vis not very
large because it requires v—1 replications which for large v may be
prohibitive. When twin calves or litter mates are used as blocks for
testing a smaller number of treatments, these designs are more suitable.
This method of construction can be extended to construct designs
with parameters
INCOMPLETE BLOCK DESIGNS 159
Point no. 1 2 3 4 5 6 7 8
Points 00 01 0a Oa2 10 11 la la2
Point no. 9 10 11 12 13 14 15 16
Points a0 al aa aa2 a20 a2l a2a a2 a2
x=cc2
(2) y=i
(3) x+y~i
(4) x + aty=i
(5) x+«*y=i
160 DESIGN AND ANALYSIS OF EXPERIMENTS
Thus there are five sets of lines with four paraUel lines in each set.
There is no common point in any two lines belonging to a set.
In general, there are (5+1) such sets of lines each set containing
s parallel lines. So the total number of lines is 5(5+1).
The number of points in a line, say,
x+py=i (5.8)
is 5, where p and / are two elements in the Galois field. That is because
for each value of x in the field, one value of y is obtainable from (5.8+
These two values of x and y give a point on the line (5.8). As x can
be any of the 5 elements in the field, there are in all 5 such points on
the line.
Given a point there is one line in each set which passes through the
point. As there are (5+1) sets, (5+1) lines passes through a point. Hence
the replication of each variety is 5 + 1. As only one line can pass through
two points, a pair of treatments occurs in one block only. That is, A=l.
The points on the 20 lines of Example 2 obtained as indicated above
are shown below. Only the point numbers have been shown (not the
coordinates of the points).
When 5=6, no Galois field exists. Hence, the design for 5 =6, v=36,
b=42, r=7, k—6, A=1 is not obtainable by this method. Actually, it
has been shown that this design is impossible.
By taking 5+1 more treatments and by putting the ith of these
treatments (when arranged in some order) in each of the 5 blocks in the
rth set of blocks, (1= 1,2,. ., 5+1) we get 52+5+1 treatments distributed
over 52+5 blocks each of size 5+1. These blocks together with another
block formed of the new (5 1) treatments give a series of symmetrical
B.I.B. designs with parameters
v=Z>=52+5+1, r=fc=5+l, A = 1 (5.9)
Evidently, each of the new treatments occurs in(5+l) blocks and each
of them occurs with each of the previous treatments in a block only once.
INCOMPLETE BLOCK DESIGNS 161
Resolvable Designs
It has been seen that the blocks of the designs
v=s2, b—s2-\-s, r—s-\-1, k—s, A—1
can be divided into (j+ 1) groups of s blocks each, such that in each group
each of the treatments is replicated once. Designs for which this type of
grouping of the blocks is possible are called resolvable designs. If, again,
the number of treatments common between any two blocks belonging to
two different groups of a resolvable design is constant, then such designs
are called affine resolvable designs.
It can be shown that for resolvable designs no solution exists if
b < v—r+1.
The proof follows exactly the same line as adopted for showing b > v.
B.I.B. Designs
=(j'2+j-f-i)
ls2—l\ /s* — 1
possible ways of choosing two interactions such that each choice generates
a separate replication. Hence, the number of blocks is The
block size is evidently (a2-1)/(j-1)=(a+1). The number of replications
of the factorial in which a particular interaction is always confounded is
(s2— 1)A>— 1)=CH-1), because given one interaction, a second interaction
can be chosen in (s2- 1)/(a— 1) ways. Similarly, A==(a — 1)/(a—1)=1. This
method was given by Kempthorne (1952).
0 1 3 9
1 3 9
12 10 4
2 8
11 5
6
7
Blocks Contents
1 0 1 3 9
2 1 2 4 10
3 2 3 5 11
4 3 4 6 12
5 4 5 7 0
6 5 6 8 1
7 6 7 9 2
8 7 8 10 3
9 8 9 11 4
10 9 10 12 5
11 10 11 0 6
12 . 11 12 1 7
13 12 0 2 8
1,4, 5, 9, 3
The actual design is shown below.
Blocks Contents
1 4 5 9
2 5 6 10 4
3 6 7 0 5
4 7 8 1 6
5 8 9 2 7
6 9 10 3 8
7 10 0 4 9
8 0 1 5 10
9 1 2 6 0
10 2 3 7 1
0 3 4 8 2
Dual Designs
Let 0l5 02, .. .,0„ denote the treatments in a B.I.B. design with b blocks.
Let the blocks be numbered by 1,2, ..., b. Any treatment 0/ (i= 1, 2,..., v)
occurs in r blocks. Let a block (3/ be formed with these r block numbers.
Then the design with the blocks (3j, (i2, ..., (3„ is called the dual design of the
original B.I.B. design. The parameters of the design are v'= b, b'—v, r'=k,
k'—r. The resultant design is not, however, a B.I.B. design always. If
the original design is a symmetrical B.I.B. design, then the dual design is
also a B.I.B. design with the same parameters.
5.5 Analysis
It has been pointed out earlier that the data obtainable from incomplete
block designs are non-orthogonal. They are, therefore, analyzed by
following the method of analysis of non-orthogonal data described in
Chapter 1. The method as appropriate for B.I.B. designs is discussed
below briefly.
The incidence matrix of a B.I.B. design is of the following type.
TABLE 5.1
Incidence Matrix of B.I.B. Designs
Block
1 1 0 1 . .. 0 k Bx
2 0 0 1 . .. l k b2
3 1 1 1 . .. 0 k b3
b 0 1 0 . .. I k Bb
Total
no. of rep. (n4.) r r r ... r
Observation total Tx Tt T3 ... Tv
Adjusted total 01 0S 03 •• • Qv
INCOMPLETE BLOCK DESIGNS
167
where nu denotes the entry in the above table in the cell defined by the ith
treatment and j'th block.
The reduced normal equations obtainable from the model
are as follows
2 (5.11)
\ k) k m7it
Taking the restriction 2 tm—0, the solutions of the treatment effects are as
m
follows:
f,=gwherc£=^1(i=1.2,....v).
s uq,=~e 2 e<*.
Next, the following analysis of variance table is prepared.
TABLE 5.2
Analysis of Variance of B.I.B. Designs
Sources of
s.s. m.s. t
variation d.f.
V
Block
b-1
(unadjusted) f k
Treatments St2
v—1 4rS Qi*
rE "
(adjusted)
vr—b—v+1 By subtraction s2
Error
where c2 is estimated by the error mean square, j*. This design is called
balanced because the variances of the estimates of all elementary treatment
contrasts like (ti—tm) ure all equal.
168 DESIGN AND ANALYSIS OF EXPERIMENTS
rE
where cr„2 is the error variance from a randomized block design with v as
block size and a- the error variance from B.I.B. designs with k as block
size.
The ratio 0„2)/a2 is expected to be greater than unity while E is less
than unity as shown below. If the product Ex(<jf)/a2 is greater than
unity, then the B.I.B. design is more efficient than the corresponding
randomized block design. E does not indicate the real efficiency of a B.I.B
design but is only a factor in the efficiency. It is, therefore, called the
efficiency factor of the design.
We have
v’A v(k—1) k— 1
because-=
rk k (v—fj r "v-1
k— 1
k
< 1 since k < v.
'El
V
In all incomplete block designs there are two types of blocks. One type
of blocks consists of sets of k treatment numbers. Another type of
blocks are the actual blocks, each consisting of a set of k experimental
units. We shall call the two types of blocks as treatment blocks and unit
TABLE 5.3
Layout and Yield from the B.I.B. Design v=b~13, r=k=4, X=1
Block numbers
1 2 3 4 5
(1) (2) (3) (1) (2) (3) (1) (2) (3) (1) (2) (3) (1) (2) (3)
3 50 1 3 45 —3 10 31 --15 2 54 1 7 47 --11
6 39 --10 4 38 —10 11 37 --9 5 54 1 8 60 2
9 58 9 8 65 17 12 63 17 8 70 17 9 61 3
11 49 0 12 44 —4 13 53 7 11 34 —19 10 64 6
Block numbers
8 9 10
6 7
(2) (3) (1) (2) (3) 0) (2) (3)
(1) (2) (3) CD (2) (3; (1)
3 68 0 1 77 7 2 57 —3
4 60 —1 1 69 8
5 65 —3 2 65 —5 4 60 0
5 65 4 5 '62 1
7 66 —2 3 75 5 9 53 —7
6 54 —7 9 52 —9
0 13 73 5 10 63 —7 13 70 10
10 65 4 12 61
Block numbers
li 12 13
(3) (1) (2) (3) (1) (2) (3)
a) (2)
72 9 1 63 —4 2 61 —3
i
62 —1 6 67 0 6 63 —1
4
62 —1 8 54 —13 7 79 15
7
56 —7 13 84 17 12 53 —11
11
The treatment totals as obtained from the deviates are shown below.
Treatments 1 2 3 4 5 6 7 8 9 10 11 12 13
Incidentally, the above totals are also the adjusted totals of the treatments.
Taking each total to be based on four observations the treatment sum
of squares=1106.5
The efficiency factor =11
16
1106.5x16
Hence, the adj. treat. s.s. =- = 1361.85
13
INCOMPLETE BLOCK DESIGNS 171
Table 5.4
Analysis of Variance of the B.I.B. design v=6=13, r=&=4, A=1
Total . 53 188052.00
*not significant
The total sum of squares has been obtained by addition. The variance of
the estimate of the difference between two treatment effects is
2x77.04 4x2x77.04
=47.41
rE 13
3072
The grand mean of all observations—^ -59.1
Treatments 1 2 3 4 5 6 7
Treatments 8 9 10 11 12 13
The relative performances of the varieties can be assessed better from the
■estimates increased by the grand mean as presented in the last row.
5 8.1 Analysis
TABLE 5.5
Incidence Matrix of m-ple Lattice Designs
Block 1 1 1 0 . 1 k
2 0 0 1 • . 1 k
3 1 1 0 . . 0 k b3
mk 0 1 0 . . 1 k Bmjc
Total (Rep. «<.) m m m m
Treatment totals n T2 Tz Tk*
Adjusted totals Q\ Q, Qz Qk*
ro, if the ith treatment does not occur in the /'th block,
where
(.1, if the /th treatment occurs in the 7th block.
Further,
2 nijnmJ=\1’ if the Ith and wth treatments occur together in a block,
j 11 mJ [0, if these two treatments do not occur together in a block.
The reduced normal equations come out as below.
where SR(Q,) has similar meaning as SR(ii). The other terms vanish due
to 2,ti—0.
Similarly, Sc (ti)=S£^Ql)
m— 1
7:=gf+^(^)+‘s'c(g/)+^, (Qi)+...+Sm^(o,)
m mk(m-1) - (5-13)
The adjusted sum of squares due to the treatments can now be obtained
from 2 ti Qi.
i
TABLE 5.6
Analysis of Variance of m-pie Lattice Design
b,‘ «
Blocks (unadjusted) mk—1
7
s,
by subtraction
1
5^
i
1
The estimate of the difference between the effects of the /th and the i'th
treatments is given by
(5.14)
DESIGN AND ANALYSIS OF EXPERIMENTS
176
If the ith and the i'th treatments do not occur together in any block, then
(5.15)
Lattice designs have the drawback that they are available only for numbers
of treatments which are perfect squares or cubes. This limitation was
‘removed considerably by Bose and Nair (1939) by evolving partially
balanced incomplete block (P.B.I.B.) designs. They defined an m-associate
design as follows.
A set of v treatments distributed in b blocks each containing k distinct
treatments {k < v), is said to form a P.B.I.B. design with m-associate
classes when
(1) every treatment occurs in r blocks
(2) two treatments occur together in Aa, X8, ...» or Am blocks
(3) given a treatment 0 each of nt treatments occur with 0 in A,
blocks (i=l,2,..., w) so that 2 n,=v-1 and 2 n,A,=r (k-1).
1 i
Relative to 0 the remaining treatments are, thus, grouped into m groups
such that the ith group contains n/ treatments each of which occurs with 0
in A, blocks. These nt treatments are said to be the ith associate treatments
or simply ith associates of 0.
Given two treatments 0 and <f> which are ith associates, let the number
of treatments common between the jth associates of 0 and /cth associates
of<t> be denoted by p‘k•
INCOMPLETE BLOCK DESIGNS 177
Then, the P.B.I.B. designs require further that p\k is constant for any
choice of 0 and 6 so long as they are z'th associates.
In the above definition the association between two treatments has been
defined in relation to the number of replications of these two treatments
together. Later on, Bose and Shimamoto (1952) defined association
between two treatments purely from group considerations without relating
it to the number of replications of the pair of treatments. According to
this definition, two treatments or symbols are either first associate or second
associate, etc. or mth associate, and any treatment has m z'th associates
(z'= 1,2,..., m). Given any two treatments 0 and <f» which are z'th associates,
the number of treatments which are y'th associate to 0 and kth associate to
<f> is pjk where p)k is independent of the choice of 0 and <f> so long as they
are z'th associates.
A display of the associates of each of v treatments satisfying the above
requirements forms an association scheme. The parameters of the scheme
are nlt «2.nm, Pjk(i,j, k—1,2,,. .,m). If with v treatments such a
scheme is available then the P.B.I.B. designs are so defined that the require¬
ments in (1) and (2) of the previous definition are satisfied together with
the fact that two treatments which are z'th associates should occur in A<
blocks (z'= 1,2, ..., m).
Ai=2 4, 7, 6, 9
1, 3, 5, 8
A, =2 1,3, 7,9
Aa=»l 2, 4, 6, 8
Taking the treatments 1 and 2 which are second associates we find p*x=2,
because 6 and 9 are common between the first associates of 1 and 2.
Similarly,
*-(! a
5.9.1 Some Facts About P.B.I.B. Designs
Many incomplete block designs including the B.I.B. and lattice designs are
particular cases of P.B.I.B. designs. When all the Vs are equal, it
becomes a B.I.B. design. The square lattice designs are two associates
P.B.I.B. designs. If there are m replications in a square lattice design with
k2 treatments, then the parameters of the corresponding P.B.I.B. design
are as given below:
v—A:2, b=mk, r—m, k—k, nx=m (k— l),n2=(fc—l)(fc+l— m), A1=l,Aa=0
(1) The relation jPjk—i^k) always holds because of symmetry, that is, from
the fact that if 0 is an /th associate <f>, then<£ is an fth associate of 0.
=nj—1 if i—j.
(3) n,pjk =njPjk (5.17)
Taking 0 and ^ as two /th associate treatments, the /^treatments which
are yth associates of 0 are distributed in the different associate classes of
^ excepting that when i=j, <f> is one of the /th associate treatments of 0
and is not present in any of the associate classes of <j>. The relation at
(5.16) follows immediately from this consideration.
INCOMPLETE BLOCK DESIGNS 179
(1) 2 Ai=EVfV=the vxv matrix all of whose elements are unity. From
z-o
the same consideration, it follows that
Iff A
(2) 2 aiAi=0,,y implies and is implied by a0=a1«=... =am=0.
' Z-0
(2) and (3) mean that the set of linear functions of A0, Alf..., Am
forms a vector-space of dimension (m+1) with basis A0, Alt..Am
and is closed under multiplication. Multiplication of the A-matrices
is commutative also. This is because
Pm, P/rtf mi
' ' • Pmi
(6) is equivalent to
The P.B.I.B. designs with two associate classes are the simplest for analysis.
INCOMPLETE BLOCK DESIGNS 181
0 n— 1
n— 1 n(m—2)
(i) the positions in the principal diagonal of the scheme (upper left
to lower right) are left blank,
(ii) the n(n—l)/2 positions above the principal diagonal are filled by
the treatment numbers 1, 2.n(n—1)/2,
(iii) the n (n—1)/2 positions below the diagonal are so filled that the
array is symmetrical about the principal diagonal and
(iv) for any treatment i the first associates are exactly those treatments
which lie in the same row as i.
Here, nx=2n—4, n2=(n—2)(n—3)/2
1 _ 1f n—2 n—3
Pi i |\ n—3 {n—3) (n—-4)/2
2/1—8
Pi]— I((2n-8
4
(71— 4) (77 —-5)/2
L(L-l) L(n—L)
L(n—L) {n-L)2+(L-2)
TABLE 5.7
Incidence Matrix of P.B.I.B. Designs
Blocks 1 1 0 o . . . 1 k Bx
2 0 1 l . . . 0 k b2
3 1 1 o . . . 1 k Bz
b 1 1 l . . . 0 k Bb
Treatment totals Tx t2 t3 . . . T,
Adjusted totals Qx Q2 Qz Qv
where Qt—Tt—^nuBj
In this design «,.=r, n.j=k and 2 mj ni>j—'kp when the ith and the i'th
j
treatments are pth associates (p — \, 2,..., m).
Let ti (i= 1, 2,..., v) denote the effect of the ith treatment. Let further
Sp (ti) denote the sum of the np treatment effects which are pth associates
to U(p—1,2,..m).
Similarly, Sp(Qi) denotes the sum of the adjusted totals of these np
treatments. Let further 0', ©/,..., 0j,p denote the effects of the np
treatments which are pth. associates to the ith treatment.
The reduced normal equation for ti is
A j=A2 Aj
’^l) P] 2
B2=r (*-!)+A2+(A2-Aj) (Pl-Pl)
186 DESIGN AND ANALYSIS OF EXPERIMENTS
In Section 5.6 it was pointed out that block effects can be considered to
be random variables in incomplete block designs. Accordingly, using
block effects as another source of error, there can be one more estimate of
the treatment effects (based on block totals). Such estimates are called
inter-block estimates./ Yates (1940) obtained inter-block estimates of
treatment effects for balanced incomplete block designs as functions of
block totals. Then by suitably combining the inter-block estimates and the
usual intra-block estimates the combined estimates of the treatment effects
were obtained. In P.B.I.B. designs it is not always possible to get
separate inter-block estimates as functions of the block totals because in
these designs the number of blocks can be less than the number of treat¬
ments. However, Rao (1947) gave a method of obtaining combined
estimates of treatment effects for incomplete block designs in general. This
method has been described below.
In the model
Yu—(x -j- ti+bj+Cij
INCOMPLETE BLOCK DESIGNS 187
The reduced normal equations for intra-block estimates which are obtained
by minimizing
2 (y>j—[>■—U—bj)2 (5.27)
u
are r(k — 1) ti— 2 —kQi(i= 1,2, ..., v) (5.28)
iV(
1 W' 1
Let W=~ and -=*-(kab2+o2)-\
Then linearly combining the two equations (5.26) and (5.28) with W'Jk
and W respectively as weights we get the following equation from which
the combined estimates can be obtained:
The equations (5.29) can also be obtained by first combining the two
least-squares set-ups (5.23) and (5.27) with W'\k and W respectively as
weights and then minimizing with respect to the parameters.
The equations (5.29) are the general reduced normal equations for all
incomplete block designs. For P.B.I.B. designs with m-associate classes,
these equations take the same form as in (5.18) obtained for intra-block
analysis of P.B.I.B. designs. Subsequent procedures for obtaining the
combined estimates are, therefore, the same as in the case of intra-block
analysis.
188 DESIGN AND ANALYSIS OF EXPERIMENTS
lr{W(k-l)+W'}+(W-W')K] t,]
Adding such equations for each of the treatments in S1 (u) and eliminating
S2 (U) we get
• 2 kB2 , , . .
—-mu—m*> when the treatments are second associates
AxB2—A.ifBi
It is seen that by putting Ai=A2 in equation (5.30) we get the combined
-estimate of the treatment effect for B.I.B. designs as
Q,+ P, VW
WkQi+W'Pi
where p= (5.34)
r{W(k-\) + W } + (W-W')X W
rE+ p
(^)
2 Wk
var{ti—tm)-
r{W(k-\)+W') + (W-W')A~r£[ p(«-A)
INCOMPLETE BLOCK DESIGNS 189
The variance and the efficiency formula remain of the same form as in
the corresponding incomplete block designs. No exact F-test is valid for
testing the equality of the treatment effects based on the combined estimates
particularly because W and W' are unknown. Approximate F and f-test
are, however, possible when the treatment effects are estimated from (5.34),
after substituting for estimates of W and W'.
W' is estimated by using the adjusted block mean squares. The adjusted
block sum of squares can be obtained from the relation (1.19) which was
derived in Section 1.8.4 of analysis of variance of non-orthogonal data in
Chapter 1.
Adjusted block r.r^fadj. Tr. j’.5,.)-|-(unadj. block s.s.)—(unadj. Tr. s.s.)
(5.35)
Let S&* denote the adjusted block mean squares for any incomplete block
design. Then
k(b-\)sb2-(?-k) s2
(5.36)
v(r—1)
These estimates remain the same for both B.I.B. and Youden Square
designs.
For resolvable incomplete block designs in which the blocks can be
made into r groups such that each group contains a complete replication
of the treatments, the within group adjusted block mean squares can be
used to estimate W more precisely. The within group adjusted
block sum of squares can be obtained by subtracting the between group
190 DESIGN AND ANALYSIS OF EXPERIMENTS
sum of squares from the adjusted block sum of squares. It has (b—r)
degrees of freedom. Let the within group adjusted block mean squares
be denoted by Sb'2.
(r- -l)(v—Ac) 2
Then E(sb’*>-oa- i---- Go
E{Yiu)=\i+ 2 tj rrijiu
j=i
if i—i’, u=u
n * - g* i
6(fc-l)
(5.39)
(6-1) 2 ylu*-2 2 tjO Tj+r2(tjW)*-l 2 B,a(f<»))l
L i,u j J *'/=■! J
where B/(if)=i?,—2 mjiutj and 7<") is the «th approximation for t, obtained
J,u _ _
by solving the equations (5.38) taking r=f<"-1). As a first approximation
for T, we may take its intra-block estimate.
Individual maximum likelihood estimates of a2 and o*2 have also been
provided by Roy and Shah under the normality assumption. But since
these are somewhat difficult to compute, they also derived estimates of a2
and (a2+kcb2) by restricting to the class of quadratic estimators. In case
when the ratio p=WfW' is large, the optimum unbiased estimates of o2
and (kab2+o2) obtained by minimizing the term involving p2 in the
expression for the respective variances of the estimates have been provided
by them and are given by
± S **(»)-£] (5.40)
where Bi(i)=Bi— 2 mjlu ~tj and tj is the intra-block estimate ofT/ obtained
from (5.28) and ci=(rk—<f>i)/<f>i with <£,’s as the q positive latent roots of
NN\ which are smaller than rk. This estimate (5.40) was also suggested
by Shah (1962) from intuitive considerations. Starting with these
estimators, we get as an unbiased estimator of p the following:
2 \(a2+W) 2(y-l) (\
bk-b-v+l) (bk-b-v+l)(b-l)\E )
(5.41)
For details of the derivations of the above results, the reader is referred
to the relevant paper by Roy and Shah (1962) cited at the end of this
chapter.
Another kind of results in this direction has been obtained by
Zyskind and Martin (1966) while developing conditions for combinations
of information from uncorrelated linear models by simple weighting. It
is clear from the work of these authors and also from the work of Roy
and Shah (1962) that the reduced intra-block and inter-block normal
equations (5.28) and (5.26) respectively for estimating the same set of
treatment effects can be regarded as originating from two uncorrelated
sources of data. This fact often makes it possible, under certain conditions,
to obtain the best linear unbiased estimate of an estimable linear function
of treatment effects (e.g., estimable treatment contrasts) just from one or the
other set of normal equations alone or by simple weighting of the solutions
from both the equations. Zyskind and Martin (1966) have derived
various such conditions in terms of the form of the matrix NN' where N
is the incidence matrix of the incomplete block design in question. For
example, for incomplete block designs in general, if we are interested in
estimating a treatment contrast V t, then it is best estimated from intra-
block information alone, if, and only if, NN'I=0 and from inter-block
information alone, if NN'l=rkl so that in either case combination is no
longer necessary. Moreover, in general, a treatment contrast I't, estimable
from both the intra-block and inter-block information, is best combined
by simple weighting (b.c.s.w.) in the sense that its b.l.u.e. is obtained by
simple weighting of its intra-block and inter-block estimates if, and only
if, / is an eigenvector of NN’. In particular, in an incomplete block
design all treatment contrasts are b.c.s.w. if, and only if, the design has
the structure of a balanced incomplete block design. For the derivation
of these and other very similar interesting results the reader is referred
to the relevant paper of Zyskind and Martin (1966) cited at the end of
this chapter.
The data along with the block contents and the deviate of each observation
from its block average are shown in Table 5.8.
INCOMPLETE BLOCK DESIGNS 193
TABLE 5.8
Layout Plan and the Observation (K.G. per plot) from the above P.B.I.B. Design
Blocks
1 2 3
3 59 3 2 35 -1 1 48 4
8 56 0 7 33 -3 7 42 -2
4 53 -3 4 40 4 5 42 -2
Blocks
4 5 6
Blocks
7 8 9
1 54 -4 • 2 45 —1 1 31 0
8 58 0 9 46 0 2 27 -4
6 62 4 6 47 1 3 35 4
The sum of the deviates in the above table against the treatment
number / gives Qt(i—1, 2,..9).
Thus gi=4—4+0=0.
Similarly, 02=—6
03=8
04=3
05=-3
06=1
07= 10
08 = 5
09=2.
The association scheme of the design is as shown below.
V=1
< -- 2, 3, 5, 6, 7, 8 (first associates)
- 4, 9 (second associates)
At=0
1. 3, 4, 6, 7, 9
5, 8
1,2,4, 5, 8,9
6,7
2, 3, 5, 6, 7, 8
1,9
1, 3, 4, 6, 7, 9
2, 8
1, 2, 4, 5, 8, 9
3,7
1, 2, 4, 5, 8, 9
3, 6
1, 3, 4, 6, 7, 9
2,5
2, 3, 5, 6, 7, 8
1, 9
INCOMPLETE BLOCK DESIGNS 195
TABLE 5.9
Analysis of Variance of the Data Presented in Table 5.8
For combined
Source
d.f. s.s. m.s. F analysis
of
s.s.
variation
Blocks
Blocks
64476.00 1295.00 (adj.)
(un. cor) 8
(unadj.)
121.67 15.21 1.5 n.s. 63302.67 Treatments
Treatments 8
(unadj.)
(adjusted)
„i=!
- 54 9
=c
(|x«42)
Av. var.
11 .
=12 °
Combined Estimates
The weight, W is estimated by 1 Is2 where 52 is the intra-block error
m.s. in Table 5.9.
That is, W=0.1.
To estimate W’, we require within group adj. block s.s.
This s.s.=(Adj. block 5.5.)—(Bet. group s.s.)
= 1495—14=1481
Hence W'=■■-■■& 1 -2
rst,2—52
2
"3x185.125-10.033
=0.004
To obtain the combined estimates we require further
A^riWik-V+W^+iJV-W) Aa=0.612
A2'=(\-\){W-W')
= -0.096
■®i =(^a ^i) Pu*
=0
INCOMPLETE BLOCK DESIGNS 197
Bt’=r{W{k-\)+W')+(W-W) (p\-*np)}
=0.900
Again, Ri=Wk Qt + W'Pi (/= 1,2.9)
and P/=F,~G.
b
In the subsequent calculation Pi has been taken to be equal to
Pi(*=l» 2,.. .,9) and the term rjb G has been ignored. This will result
in a shift of the origin of the estimate tt of the treatments. Since 2 f/=0
i
without the above manipulation, we can easily get the actual solution
from the following:
2//
T *' l
where U' is the estimate of tt obtained after ignoring rjb G while calculating
Pi(i— 1, 2,..., v).
Accordingly,
PX=VX= 132+174+93 =399=Sum of totals of blocks with tx.
Pa=339
Ps=414
P4=453
P6=462
Pe=4 89
P7=393
P8=495
P„=444
Hence P,= WkQt+ W'Pi (/= 1, 2,....9)
= .3<2,+ .004P,
is said to be Z)-optimum in E if det (cov ^ det (cov («)<*) for any other
d^D.
the C-matrix of the design plays a very important role. In fact, all the
optimum properties are derived from some special forms of the C-matrix
as defined in Chapter 1. For further study and proofs of the results stated
here the reader is referred to the references given in this chapter.
EXERCISES
Prove that for a symmetrical B.I.B.D. with v even, (r—A) must be a perfect
square. Also, show that for a symmetrical B.I.B. design,
and hence that NN'^N'N. Interpret the last result in terms of the number of
treatments common between any two blocks.
2. Prove that for a resolvable B.I.B.D. rank N < b—(r—1).
3. Prove that the C-matrix of a B.I.B.D. is given by
vk A
C=~j~ Iv Evv
and hence show that using a B.I.B.D. all elementary treatment contrasts are
estimated with the same precision.
4. Show that for an affine resolvable B.I.B.D. the parameters can be expressed in
terms of two integers n and t (n > 2, t > 0) as
v=nk=ri2‘[(n—\) f+1], 6 = «/-<=« (rfi + nt+l), A=/tf+l.
where Qi' is the adjusted total of the z'th treatment obtained by taking zero for
the missing values and 2 Qi is the sum of adjusted total of those treatments
an
which are present in the block in which the z'th treatment is missing. Show
further that the true error sum of squares is equal to
where (E.S.S.)o is the error sum of squares obtained by taking zero for the
missing observations.
8 rfi numbers from I to are arranged in the form of an rtxn square. The n
elements occurring in the left to right diagonal of the square are. omitted. Next,
INCOMPLETE BLOCK DESIGNS i201
from each row of the square a block of size («—1) is formed. Similarly, from
each column of it a block is formed. The design in these 2n blocks each of (n-/l)
units is called a simple rectangular lattice design. Show that this design is a four
associate P.B.I.B design.
9. If in each block of a P.B.I.B. design with two associate classes a new treatment
is included, derive a method of analysis of the resultant design.
10. Let one of the b blocks of k units each of a B.I.B. design with v treatments be
drawn at random with equal probability. If these k units are taken to provide
a sample of k units from a population of v units, show that it provides the same
unbiased estimate of the population mean and its variance, as is provided by the
technique of simple random sampling without replacement. Show further that
the usual simple random sampling without replacement corresponds to drawing of
one of the blocks, of the following B. I. B. design at random.
Vl« 36 V7 24
32 Vio 22
v6
Block 3 Block 4
v2 36 v# 41
13 V8 29
v16
31 vu 3&
^12
15 Vs 25
V6
Block S Block 6
Replication II
18 VlO 18
v8
41 v. 21
v13
23 Vl 1+
V8
39 Vl2 44
Vl2
Block 7 Block 8
13 v# 27
v6
40 v7 21.
H
25 V2 30
vn
26 Vl« 31
Vll
202 DESIGN AND ANALYSIS OF EXPERIMENTS
REFERENCES
;
INCOMPLETE BLOCK DESIGNS 203
5. Bose, R.C., Clatworthy, W.H. and Shrikhande, S.S. (1954), “Tables of parti¬
ally balanced designs with two associate classes.” North Carolina Agricultural
Experiment Station. Tech. Bull, no 107.
Bose, R.C. and Mesner, D.M. (1959). “On linear associative algebras corres¬
ponding to association schemes of partially balanced designs.” Ann. Math. Stat.
30, 21-38.
7. Chakrabarti, M C. (1963). “On the C-matrix in design of experiments.” Jour.
Indian Stat. Assoc. 1, 8-23.
8. Connor, W.S., Jr. (1952). “On the structure of balanced incomplete block
designs.” Am. Math. Stat. 23, 57-71.
9. Connor, W.S. and Clatworthy, W.H. (1954). “Some theorems for partially
balanced designs.” Am. Math. Stat. 25, 100-112.
10. Das, M.N. (1958). “Reinforced incomplete block designs.” Journ. Ind. Soc. Agri.
Stat. 10, 73-77.
11. Das, M.N. (1958). “Circular designs.” Journ. Ind. Soc. Agri. Stat. 12, 45-56.
12. Ehrenfeld, S. (1953). “On the efficiency of experimental designs.” Ann. Math,
Stat. 26, 247-55.
13. Fisher, R.A. and Yates, F. (1948). Statistical Tables. Hafner Publishing Co.
New York. _
14. Fisher, R.A. (1940). “An examination of possible different solutions of a
problem in incomplete blocks. Ann. Eugen.” 9,353-400.
Giri, N.C. (1958). “On reinforced P.B.I.B. designs.” Journ. Ind. Soc. Agri.
15.
Stat. 12, 41-51.
16. Giri, N.C. (1957). “On row balance in P.B.I.B. designs.” Journ. Ind. Soc.
Agri. Stat. 11, 168-78.
17. Giri, N.C. (1965). “On the F-test in the intra-block analysis of a class ot two
associate P.B.I.B. designs.” Journ. Amer. Stat. Assn. 60, 309.
Harshbarger, B.(1947). “Rectangular lattices.” Virginia Agricultural Experi¬
18.
mental Station, Memoire, 1. „
19. John, P.W.M. (1964). “Balanced designs with unequal number of replications.
Ann. Math. Stat. 35, 897-99. v ,
Kempthorne, O. (1952). The Design and Analysis of Experiments. Wiley, New York.
20.
Kempthorne, O. (1956). “The efficiency factor of an incomplete block design.
21.
Ann. Math. Stat. 27, 846-49. ... , , . .
22. Kiefer, J. (1958). “On the nonrandomized optimality and randomized no
optimality of symmetrical designs.” Ann. Math. Stat. 29, 675-99.
Kiefer, J. (1959). “Optimum experimental designs. ’ Jour. Royal Stat. Soc. (»;,
23.
21 272-319.
Kshirsagar, A.M. (1958). “A note on incomplete block designs.” Ann. Math.
24.
KulshLhthI0 A.c. and Das M.N. (1968). “On derivation of initial Nocks of
25.
BIB designs with more than one initial block.” Aust. J. Stat. 10, 75-8Z.
Martin, F.B. and Zyskind, G. (1966). “On combinability of information from
26.
uncorrelated linear models by simple weighting. Ann. Math. Stat. 37, 1321 ■■ •
Mote, V.L. (1958). “On a minimax property of a balanced incomplete b
27.
design. “Ann. Math. Stat. 29, 910-14. / ,i i
Nair, K.R. and Rao, C.R. (1942). “A note on partially balanced incomplete block
28.
designs.” Science and Culture, 7, 568-69. . „ . «
Nandi, H.K. (1950). “On the efficiency of experimental designs. Cal. Stat.
29.
32. Rao, M.B. (1966). “A note on equireplicate balanced designs with 6=v.”
Calcutta Stat. Assoc. Bull. 15, 43-44.
33. Rao, V.R. (1958). “A note on balanced designs.” Ann. Math. Stat. 29,290-
294.
34. Raghavarao, D. (1971). Constructions and Combinatorial Problems in Design of
Experiments. Wiley, New York.
35. Roy, J. (1958). ‘On the efficiency factor of block designs.” Sankhyd. 19,181-
188.
36. Roy, J. and Shah K.R. (1962). “Recovery of interblock information.” Sankhya
(A), 24, 269-280.
37. Saha, G.M. and Das, M.N. (1971). “Construction of P.B.I.B. designs through
2" factorials and some new designs.” Journ. Comb. Theory, 11, 282-95.
38. Shah, K.R. (1962). “An estimate of inter-group variance in one and two-way
designs,” Sankhyd (A), 24, 281-86.
39. Sinha, B.K. and Sinha, B.K. (1970). “Comparison of relative efficiency of some
classes of designs.” Cal. Stat. Assoc. Bull. 19, 97-122.
40. Sinha, B.K. (1971). “Contribution to comparison of experiments: Optimum
experiments for linear inference.” Unpublished Ph. D. thesis, Calcutta
University.
41. Smith, C.A.B. and Hartley H.O. (1948). “The construction of Youden squares.”
Journ. Roy. Stat. Soc. B, 10, 262-63.
42. Sprott, D.A. (1962). “Listing of B.I.B. designs from r=16 to 20.” Sankhyd 24,
203-04.
43. Vartak, M.N. (1963). “Disconnected balanced designs.” Jour. Indian Stat.
Assoc., 1, 104-07.
44. Wald, A. (1943). “On the efficient design of statistical investigations.” Ann
Math. Stat. 14, 134-40.
45. Yates, F. (1936). “Incomplete randomised blocks.” Ann. Eugen. 7,121-40.
46. Yates, F. (1936). “A new method of arranging variety trials involving a large
number of varieties.” Ann. Eugen. 7. 319-22.
47. Yates, F. (1940). “The recovering of inter-block information in balanced incom¬
plete block designs. Ann. Eugen. 10, 317-25.
48. Youden, W.J. (1940). “Experimental designs to increase accuracy of greenhouse
studies.” Cont. Boyce Thompson Inst., 11, 219-28.
CHAPTER 6
We assume that all the orthogonal latin squares are in the substandard
form.
In an orthogonal latin square of order r any of r— 1 symbols can be
written in the second position of the first column, the remaining symbol
occurs in the first position of this column. As in each square a different
symbol is to occur in this position, the number of such squares can be at
the most r— 1.
When r is either a prime or a prime power all the (r— 1) orthogonal
latin squares can be constructed. When r is the product of different
primes or prime powers, MacNaish (1922) and Mann (1942) gave methods
of construction of (5— 1) orthogonal squares of order r where sis the
minimum prime or prime power factor of r.
We have given below a unified method of construction of orthogonal
latin squares of order r, by which (t— 1) orthogonal squares can be obtained
where, (i) t=s if r is the product of different primes and 5 is the minimum
of the prime or prime power factors of r and (ii) t=r if r is a prime or a
prime power.
The elements in Galois field G.F. (2*) are 0, 1, a, a2 with a2-f a-4-1 as the
minimum function.
1 1 0 7.2 a
a a .2
0 1
a2 1 0
Second principal Third principal
column obtained column obtained
by multiplying by a by multiplying by a2
0 0 1 a 0 0 1 a CL*
a (X2 0 1 a2 a2 x. 1 0
a1 a 1 0 1 1 0 a2 a
0 a a a a2 0 1
The three summation tables give three orthogonal latin squares.
208 DESIGN AND ANALYSIS OF EXPERIMENTS
The two factors of 12 are 3 and 4. The elements in the two corresponding
fields are 0, 1, 2 and 0, 1, a, a2.
The 12 combinations which will be used as symbols for constructing
the latin squares are
00, 01, 0a, Oa2, 10, 11, la, la2, 20, 21, 2a, 2a2.
Let us now obtain the summation table.
First
principal Principal row
column 00 01 0a Oa2 10 11 la la2 20 21 2a 2a2
We can take that the multiplier in the first square is 11 and that in the
second square is 2a. Hence, no further multiplier is possible as in the
first position 1 and 2 have appeared in the two multipliers and there is no
other multiplier left in the first field of 3.
It is easily seen that such summation tables always give latin squares.
As each row or column is obtained by adding a particular combination to
all the combinations, so no combination can be repeated in a row or
column.
Let a/ be the combination in the z'th position of the principal row and
be the combination in the yth position of the first principal column.
Let further c, and c2 denote two multiplier combinations for obtaining
two latin squares denoted by Lx and L2. The combinations in the (/, y)th
cell of the two latin squares are
These two symbol combinations can be equal only when (3, is the null
combination. In all other cases the symbols in the (/, y)th cell in the two
latin squares are different. But if cx and c2 have a common element in a
field with elements 0,1, U, t2, etc., then the combinations like (00- 010- -00),
(00- -0^00- -00), (00- -OLjOO- -00) etc. will occupy identical positions in the
two squares. Hence, the multipliers should contain each time a different
element in each field.
then <*s +
But this is not possible. This shows that any two cells having identical
symbols in one square, cannot have identical symbols in another square.
Let the two symbols in the different columns of the first two rows be used
to denote the cell positions in a v x v square. Then by putting the symbol
which is present in a column in the third row, in the cell defined by the
preceding two symbols in the same column and repeating this process for
all the columns we get a latin square. Similarly, one latin square can be
obtained for each row of the array beginning with the third. These latin
squares are orthogonal.
If all the blocks of a particular size are distinct so that no two blocks of
this size have any common treatment then such a set of blocks in a pairwise
balanced design of index unity is called a clear set.
When in a pairwise balanced design there is a clear set, the previous
method can be modified as below. For each block of a clear set, the Q
array may be taken instead of the P array. For all the other blocks P-type
arrays are used. The rest of the procedure is the same excepting that Tvqm
is replaced by T(v-s) qm where s is the number of treatments in the clear set
of blocks.
Similarly, if there be more than one clear set, 0-type of array is taken
for each block in the different clear sets. For the other blocks P-type
arrays are taken. The rest of the procedure is the same excepting that
Tyqm is to be replaced by -^(v—2 Si) qm where 2 si is the number of treatments
in all the clear sets.
This method is due to Bose, Shrikhande and Parker (1960). They showed
by these methods that it is possible to get at least two orthogonal latin
squares of any order other than 6. They used several other types of arrays
also to get the desired orthogonal latin squares. One of these methods is
through symmetrical differences and is described below.
The (2m+ 1) residue classes mod(2m+l) along with xlt x2, ..., xm form
the 3m+l symbols of the. latin squares.
First, the following 4 vectors are taken:
=(0,0,0,,.. .0)
P2=(l,2,3,.. .m)
^4==(^'l, * • • j Xm).
Next, the following latin square is written with the above vectors:
212 DESIGN AND ANALYSIS OF EXPERIMENTS
7?2
J?3 i?4 /Jj iv2
_/?4 /?3 .Z?2 ^1
0 1 2 ... 2m
2m 0 1 ... 2m— 1
1 2 3 ... 0
and by B the matrix of order (2m +1) whose first (m+1) diagonals (counted
from the principal diagonal onwards from left to right) are respectively the
first (m+1) rows of + and the remaining m diagonals have the constant
elements 2m+l, 2m+ 2, ...,3m respectively, i.e..
ORTHOGONAL LATIN SQUARES 213
Be
2m—5 2m—6 2m—7 ... 2m+2 2m+3 2m +4 2m+5 ... 2m—3 2m—4
2m—3 2m—4 2m—5 ... 2m+1 2m+2 2m+3 2m+4 ... 2m—1 2m—2
2m—1 2m—2 2m—3 ... m 2m+1 2m+2 2m+3 ... 3m 2m
Let Cbe the (2m+l)xm-matrix whose columns are the last m rows of
.4 reckoned from the bottom upwards:
1 2 ... m ;
2 3 ... m+1
3 4 ... m+2
2m 0 ... m—2
0 1 ... m—1
2 3 4 ... 2m 0 1
4 5 6 ... 1 2 3
• •••••• • *
Finally, let Li, Z,a be two mutually orthogonal latin squares of order m
formed by 2m+1, 2m+2, ...,3m. This is always possible as mis odd.
Then it can be verified that the matrices
Eleven symbols are obtained from the residue classes mod 11 and xu x2
and x3 are three more symbols. The two 14 x 14 orthogonal latin squares
are obtained by using these 14 symbols.
214 DESIGN AND ANALYSIS OF EXPERIMENTS
S2 S3
$2 S3 S*
S3 iS4 5, S2
S1 S2 sa
By adding ith residue class mod 11 to each of the residue classes in 0 and
keeping xltx2 and x3 the same always, we get the matrices
0i(/=0, 1, 2,.10).
The symbols for these latin squares are the 23 residue classes mod 23
along with xlt x2 and xa.
Let us define four vectors as below:
5'1=0, 0, 0, 0, jcj, x2, x3
S2—3, 6, 2, 1, 0, 0, 0
S3=8, 20, 12,16, 20, 17, 8
S4=12, 16, 7, 2, 19, 6, 21
The matrix 0 is then formed with these vectors as in the case of 14x 14
latin squares as below.
Si S2 Sa S4
St S3 Si
s3 Si Si s2
Si Si S2 S3
■ORTHOGONAL LATIN SQUARES 215
By adding *th residue class mod 23 to each residue class in 0 and keeping
^i, *2, x8 the same, another matrix is obtained. Let it be denoted by
Of (1=0, 1,2,..., 22). Then, the matrix
(0o»0j,0S, ...,022, QT(2i)4)
is an orthogonal array of strength two with four rows and 262 columns.
Here Q is the same as used while obtaining the 14x14 latin squares.
From this array two orthogonal latin squares of order 26 can be
obtained as described in the case of 14x 14 latin squares.
EXERCISES
1. If r—2 mutually orthogonal latin squares of order r are available, show that the
complete set of (r— 1) orthogonal latin squares can be obtained.
2. It is known that in a sm factorial where s is a prime or a prime power, if the
block size is s2, a maximum of (s2—l)/(s—1) factors can be accommodated in the
blocks so as to save all main effects and two factors interaction. Treating such a
principal block as an orthogonal array of s+1 constraint and size s2, show that
(5—1) orthogonal latin squares can be obtained from it.
3. Show that using any three orthogonal latin squares of order 5 as three
symbols for the 3x3 latin squares, we can get two orthogonal latin squares of
order 15 from the two orthogonal latin squares of order 3 when each symbol of
5x5 latin squares is prefixed by the symbol of the 3x3 latin squares 0, 1, 2 to
which the 5x5 latin square is made to correspond.
4. Show that the procedure of obtaining orthogonal latin squares as described in
problem 2 can be extended to get all the (sm—1) orthogonal latin squares of
order r where r=slt s2 ..sp and sm is the minimum of the prime or prime
power factors of r.
5. Four persons bought 16 cows at the rate of four per person from four markets
such that each person bought one cow from each market. The cows belonged
to 4 breeds equally, such that each person bought a cow of each breed and the
four cows sold in each market belonged to different breeds. The cows belonged
to four different age groups equally such that each person purchased cows of
different ages, the cows sold in a market belonged to different age groups and
the four cows of every breed were of different age groups. Lastly, the cows
were marked by using four different colours such that four cows were marked
by each of the four colours, each cow belonging to a person was marked by
different colours, the cows purchased from each market were marked by different
colours and the cows belonging to each breed and age group were also marked
differently.
Describe the 16 cows in respect of the owner, market of purchase,breed,
age group and colour marking of each individual cow.
6. Show that it is impossible to get a set of nine or 36 cows of similar groupings,
that is, changing the group size to three or six but keeping the owner, market,
etc. classification the same.
REFERENCES
1. Bose, R.C. and Shrikhande, S.S. (1959). “On the falsity of Euler’s conjecture
on the non-existence of two orthogonal latin squares of order 4«+2.” Proc.
,
Natl. Acad. Sci. U.S., 45 734.
216 DESIGN AND ANALYSIS OF EXPERIMENTS
7.1 Bio-assays
z.
An assay with two preparations which have a common effect but do not
contain the same effective ingredient, is called a comparative dillusion
assay. The results obtained from analytical dillusion assays hold in general
and are not necessarily limited only to experimental conditions, while the
results obtained from comparative dillusion assays are limited by the
experimental conditions. For example, two preparations which do not
contain the same effective ingredient may produce a common effect on a
certain organism, say A, while they may not produce the same effect on a
certain other organism say, B. Then the potency estimated from such an
assay holds only for A organism and not for B. This is, however, not the
case with analytical dillusion assays. We shall discuss here only analytical
dillusion assays.
The purpose of bio-assays is ultimately to conduct an experiment for
estimating the doses of the standard and test preparations, such that each
of them produces the same response. Depending on the nature of the
preparations, the experimental subjects, that is, the type of living being or
organism being used as experimental units and the type of response, such
two doses can be estimated by various methods. One of these methods
attempts to estimate such doses directly and is called direct bio-assay. In
many situations this technique is not applicable and indirect methods are
adopted for the estimation of such doses. A detailed discussion of the
application of statistical methods to bio-assays is available in the works of
Finney (1964). We have discussed here only some of the topics which
are necessary for statistical designs for bio-assays.
Fieller's Theorem
Let a and b be two random variables distributed normally such that
E(a)=a
E(b)=£
a
v(a) = vus2
Hb)=v22s2
cov (a, b) = v12s2
The fiducial limits of [x are obtained from the following
response under consideration. These animals are then divided into two
random groups for allocation to the two preparations. It is, however,
risky to choose two homogeneous groups of subjects without regards to
the between group variation and then allot them to the two preparations.
This procedure may increase the precision of the estimates and }’t but
it introduces bias in the estimation of relative potency when the animals
in the two groups differ materially.
Fieller’s theorem gives a measure of precision of R by obtaining
fiducial limits of R. This method makes the assumption that tolerance is
normally distributed. The expression for the fiducial limits of R are some¬
what complicated. If, however, the logarithm of tolerance is normally
distributed, estimation of precision of R gets simplified as shown below.
When the logarithm of tolerance is normally distributed, the individual
tolerance values collected from the assay are transformed to the logarithmic
scale. If xs and Xt denote the averages of the logarithm of the tolerance
values for the standard and test preparations respectively, then
log R=Xs—xt.
TABLE 7.1
Responses from k Doses of the Standard Preparation
Doses
Response dx d<L d$ . .. dk
Let 2 Si = G.
TABLE 7.2
Analysis of Variance for Testing the Linearity of Dose Response Relation
s.s.
Sources of variation df. s.s. F
m-S
5,2 <72
Between doses k-1
Tn=D
C.
(Ydi)G
Regression of response on dose 1 if
T A.
UjUt =>/v
/
k
G2
Total kn—1 s yrf--kn=T
u
Hence, the relationship for the test preparation is also linear like that of
the standard preparation for the same transformation.
An examination of the two equations for the two preparations shows
that the lines have the same slope and are, therefore, parallel.
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 223
As at=as+b log p
Of Qs
we gei log p
~Y~
where a, and at are the intercepts of the lines on the y axis and b is the
common slope.
If in an assay k doses are taken for each of the two preparations and
xs and xt denote the averages of the dose metameters and ys and yt are the
average responses for the preparations, then it is known that
as=y,—bxs
and at—yt—bxt.
Substituting these values in
From equations (7.2) and (7.3) it is seen that the two lines for the two
preparations should be parallel when the dose metameter is log (dose).
The assays corresponding to this transformation are, therefore, called
parallel line assays. Before going to estimate the relative potency, it is
desirable to test with the help of the data collected from the assay if the
two lines are parallel. Such a test is called validity test. It is also desirable
to test if the relation between response and dose metameter is linear. This
gives rise to another validity test. While planning the assay it is necessary
that suitable data are available for conducting these validity tests before
relative potency is estimated.
(1) y—a-\-btXs
y=a+b,xt
(2)
where bt=b,px.
(7.5)
Hence,
Since the relative potency is estimated from the ratio of the slopes of the
two preparations, the assays corresponding to the transformation zx is
called slope ratio assays.
An examination of the equations of the two lines shows that they
224 DESIGN AND ANALYSIS OF EXPERIMENTS
intersect on the response axis. A test of this fact gives rise to one validity
test. The other validity test is to see if the relation between the response
and dose metameter is linear. It is seen that the first validity test for slope
ratio assays is different from the first validity test for the parallel line assays,
though the second test is the same for both the types. So the plans for
the two types of assays have to be different. The plans for parallel line
assays have been discussed in Section 7.4.
A parallel line assay in which each of the preparations has an equal number
of doses and an equal number of subjects is allotted to each of the doses,
is called a symmetrical parallel line assay. We shall discuss here only
symmetrical parallel line assays.
Let the number of doses of each of the preparations be k. As there
are in all 2k doses in this assay, it is called a 2&-point symmetrical parallel
line assay or simply 2A>point assay.
Let n subjects be allotted to each of the doses and a suitable response
be measured from each subject. Suppose further that jj, s2,..., s* denote
the doses of the standard preparation and tx, t2,..., tk the same for the
test preparation. Denoting the response of the rth subject allotted to the
pth dose of the standard preparation by ytpT and the rth response from the
<?th subject of the test preparation by jtqr, the response data are first
arranged as in Table 7.3.
TABLE 7.3
Response Data from 2A>Point Assay
Total Si sa .. . s* Tv t2 ,. ■ Tk
>
The analysis of the assay for conducting validity tests and for estimating
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 225
relative potency becomes very much simplified when the doses of each of
the preparations are taken in geometric progression as shown below:
s, cs, c2s,..c*-1 s and t, ct, c*t,...,c*-11
where s and t are suitable starting doses of the standard and test preparation
respectively and c is a constant which is the same for both the preparations.
A further precaution necessary while choosing the doses is that the doses
should be evenly distributed in the range of response in which the dose
response relationship was investigated for obtaining the linearing trans¬
formation.
Let xs, — log c'.y=log s-f-i log c (/=0, 1, 2,..k— 1)
and xtt = log c‘t=\og t+i log c (i=0, 1, 2,..k— 1).
Denoting by xs and x, the averages of the doses of the two preparations,
we get
i . 1 i
x,=logs-l-Y log c (7.6)
- . . .k—\.
and X(= log t-)—— log c. (7.7)
Test preparation
k-1 k—3 k-1
2 ’ 2 *
, , ,.
10 1
’ 2 '
By choosing the base of the logarithm as above, these deviate values could
be made integers.
Analysis
As stated earlier the purpose of analysis of indirect bio-assays is two fold.
First, it is tested through the analysis of variance technique if, (i) the dose
metameter and response relationship is linear and (ii) the two lines for the
two preparations are parallel. If the tests reveal that the relationship is
linear and the lines are parallel, then the relative potency of the test
preparation is estimated from the relation
log R=xs—xt—^---Jt.
k-1
(LJ- ~ to+ro--y-3 (S2+T2)~, (Sk+Tk)
when k is odd,
Combined regression contrast
-'LY (Sk-T£i
when k is odd,
= -(fc-l) (S1-T1)-(k-3) (Sa-T,)-...-^-!) (,sk-Tk)
when k is even.
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES
227
Again, when k is odd
4i(V)+(V)+---+i!}
6L,
kn(k2—])
when k is even
b=_hi_
4{(fc—l)2+(fc—3)*+... + p}
3L,
~2kn (k2—1)'
The following analysis of variance table is then written for the validity
tests and estimation of error variance.
TABLE 7.4
Analysis of Variance in 2&-Point Assays for Validity Tests
L
i
S<2+s
i
Ti2 ,
{sos«+r,)>>
Doses 2k-1
n 2k n
Within doses (error) 2k («—1) By subtraction j2
„ {S(5f+r<)}2
Total 2kn— 1 S J^spr + 2 y2tqr ?»
pr qr n
The value of D, the divisor for the regression and the parallelism sums;of
squares in the above table is (J<n (k2—1))/6 when k is odd and (2fcn(A:8—1))/3
when k is even.
For testing the linearity of regression, the mean squares for the deviation
from regression is tested by the F-test using the within mean squares as
error. For testing parallelism, the “parallelism” component is tested.
If both these tests are not significant, then the relative potency can be
estimated as below.
log R=xs—xt—
Lp kn(k2—l)
— log y—log t+
kn 6 L,
DESIGN AND ANALYSIS OF EXPERIMENTS
228
when k is odd.
=log:
v
(£2-i)
6 ’ L,
y
or R=S- antilog x'}
t
where d=logi0 c.
When k is even
s Lp 2kn(k2—\)
log R—log —
kn 3L1
That is.
„ * , •, \d(k 2_ 1) Lp)
R-- antilog <-j
lS'
Precision of R can be estimated through Fieller’s theorem.
TABLE 7.5
Contrasts of Validity Tests, Regression and Preparation in 6-Point Assay
X.i-Xs
—1 0 1 -1 0 1
TABLE 7.6
Analysis of Variance
Difference between
quadratics 1 (S1-2S2+S3-7’i+272-r3)2/12/i SDQ sys*
Within dose (error) 6(n—1) By subtraction S2
. ayf,)
Total
• 6" 1 f/h 6n
When the number of doses is large it may not always be possible to get
230 DESIGN AND ANALYSIS OF EXPERIMENTS
Blocks
1 ^1^2
2 $1^3
3 *^1*^4
4 s.zs3
5 s3sA
6 s3st
The final design in four plot blocks along with the doses of the test
preparation is then the following design.
Blocks
1
2
3
4
5
6
232 DESIGN AND ANALYSIS OF EXPERIMENTS
The coefficients of the dose total in the different contrasts of the 8-point
assay are shown in Table 7.7. These, excepting Lp are taken from the
orthogonal polynomial tables of Fisher and Yates (1963).
TABLE 7.7
Coefficients of the Dose Total in the Different Contrasts of an 8-Point Assay
Contrasts L„ -1 —1 —1 —1 1 1 1 1
Lx -3 —1 1 3 3 1 —1 -3
L\ —3 —1 1 3 —3 —1 1 3
L* 1 —1 —1 1 1 —1 —1 1
l'2 1 —1 —1 1 —1 1 1 -1
l3 1 —3 3 —1 —r 3 —3 1
K 1 -3 3 —1 1 —3 3 -1
Blocks
1
2 JjJtMs
3
TABLE 7.8
The Design and the Data of a 6-Point Parallel Line Assay (Doses in \ig)
Blocks
1 2 8 — — 9 7 26
2 6 — 9 3 — 8 26
3 — 6 12 4 6 — 28
4 9 11 — — 14 13 47
5 10 — 17 8 — 10 45
6 — 7 5 6 9 — 27
7 4 10 — — 11 13 38
8 11 — 9 3 — 15 38
9 — 9 14 5 8 — 36
10 4 7 — — 10 10 31
11 12 — 9 15 — 15 51
12 — 8 IT 7 8 — 34
13 4 4 — — 5 9 22
14 7 — 8 3 — 9 27
15 — 15 10 6 8 —■ 39
16 2 4 — — 6 6 18
17 4 — 13 5 — 12 34
18 — 10 13 4 18 — 45
TABLE 7.9
Contrasts, Divisors and Sum of Squares for Data of Table 7.8
4 6x12 0.22
Lp
Affected contrast
-3.00 4X9 0.25
w
TABLE 7.10
Analysis of Variance of Data in Table 7.8
Total 71 996.00
236 DESIGN AND ANALYSIS OF EXPERIMENTS
All the validity tests are satisfied as in the original data. It is also
seen that the error variance has not increased due to regrouping of the
observations.
The estimate of relative potence is
=anti!og {3~|.
the test preparation by 7i(i=l, 2,..k). Let C denote the total of the
observations from the blank dose. Now, the data, as shown in Table
7.11 can be used to fit the above relation by the usual method of partial
regression fitting.
TABLE 7.11
Doses and Data of a 2A:-|-1 Point Slope Ratio Assay for Fitting the
Partial Regression Equation
xt y
Si
l- »
r o *5*2
• • • • ■
l 0 •S*
1 V .. Hi:
0 Ti
k
2
0
k
• • * , .
0 1 Tk
0 0 Ci
Validity Tests
While investigating the dose response relation the blank dose is not usually
considered. When the blank dose is taken in the assay a question arises,
viz., does the linearity of relation hold up to the zero dose? It is, therefore,
necessary to test if the relation y=a+bsxs+btxt holds up to the zero dose.
This provides one validity test. For this test the difference between d and
o' is tested where d is the estimate of a, the constant term in (7.9) when the
relation is fitted by including the blank dose and a' is the estimate of a
when the relation is fitted by omitting the observation from the blank dose.
This contrast is called the blank contrast. A general expression for this
contrast in terms of the dose total has been given afterwards.
Another fact to be verified is that whether the two lines intersect on the
response axis. For this the two lines are fitted individually ignoring the
blank dose and then their intercepts on the response axis are compared.
This provides another validity test. The corresponding contrast is called
intersection. The expression for this contrast in terms of the dose totals
has also been given below.
238 DESIGN AND ANALYSIS OF EXPERIMENTS
Intersection contrast=L/=(2A:—2)(S1—7\)+(2fc—5)(S2—T2)
+(2fc—8)(S3—r3)+ ...-(£-1) (Sk-Tk).
Its divisor is nk(k— 1).
Designs
We shall consider here the following three designs for symmetrical slope
ratio assays with 2k-f-l doses.
The first two designs are the usual ones and can be obtained without
difficulty. The third design is obtained by first taking a balanced incomplete
block design taking the A: doses of the standard preparation as the treatments.
Let k' denote the block size of this initial design. The final assay design
is obtained by including in each block of the above design k' doses of the
test preparation. The rule for choosing these test preparation doses is that
if the dose st of the standard preparation is included in a block then the
dose ti is also included in the same block (i=l, 2,..., k). The blank dose
is included in each block. Here the doses of both the standard and test
preparations are arranged in the same order of magnitudes but not in
opposite orders as in parallel line assays. We shall now discuss the analysis
of these three designs in a general way.
Analysis
Let yui where one of i or j is zero, denote the observation in the /th block
coming from /th dose of the standard preparation and y'th dose of the test
preparation. y0o/ denotes the observation from the blank dose in the lth
block. Let further xsi and xtJ denote in general the variables representing
the doses of the standard and test preparations respectively.
We take the following model for the analysis:
Yiji—a-ybsXsiA-btXtiA-ri-j-eui (7.10)
where h denotes the effect of the /th block and em is the error component
assumed to be normally and independently distributed with a common
variance, as.
The details of estimation of these contrasts have been given by Das and
Kulkarni (1966) in a general way. We have given the final results for each
of the above three designs.
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 239
Cib'+Cfr^l 2 iQis
K i=i
C26,+CA=4 2 iQiT.
K i=l
The values of Cx, Ct, vx and v2 for the different designs are shown below.
where M=3k/r(k+l)(2k+l).
Here CXR, C,,R, \\R and v.2R denote the values of Cx, C2, vx and va respectively
for the randomized block design.
V\=ViM=M (5kk'+2k'+3?c+2)/(kk'+k+1)
v2*±v2m=M (3kk'+k+2k’)/(kk'+k+l).
Since bs and b, cannot be estimated independently, we cannot get their sum
of squares by adding their individual sum of squares.
The combined sum of squares of bs and b, is as below
Example 4: Bliss (1952, p. 568) has reported the data from a 9-point
slope ratio symmetrical assay on riboflavin content of yeast. The data
were collected by adopting a randomized block design with two replications.
For the present illustration we have omitted the observations from the
highest dose of each preparation. The remaining data have been taken as
if they resulted from a modified balanced* ncomplete block design from a
7-point assay in blocks of five units with two replications of each non¬
blank dose. Taking as the initial design the following balanced incomplete
block design
v=b=3, r—k—2, X=1
the final assay design was obtained as below.
Blocks
1 ij <&2 12 C
2 sx s3 t3 c
3 s2 s3 t2 t3 c
The observations from the original assay were fitted into the above design
maintaining the dose block relation. The observations of titer per tube
arranged according to this design are shown in the following table.
Originally, there were only two observations from the blank dose. But as
the present design requires three observations from the blank dose, the
third observation from this dose has been taken arbitrarily as 0.76.
The coded doses for each preparation have been taken as 1/3, 2/3
and 1.
TABLE 7.12
Data and Assumed Design as Example of a Modified B.I.B. Design
Doses
Blank Standard Test
Blocks c *2 **3 h h h Total
l(Qls+2Qls+3Qas)=4A23
1
(0ir+222r+30sr)=4.5O7
TABLE 7.13
Contrasts and Divisors for Validity Tests and their Sum of Squares
Blank ( Lb ) 6 -4 —1 2 75
—4 —1 2 —0.53 0.008
2
Quadratic (£t) 0 1 —2 1 1 15
—2 1 —0.96 0.123:
2
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 243
The divisors for the affected contrasts have been obtained as below.
Lb — 6C—4(^4-tj — (s2-\- f2)+2 Cs3-W3)
e»-f Qi-Iq,+\q*
Hence the divisor
75
12
15
= 2 •
TABLE 7.14
Analysis of Variance for the Modified B.I.B. Design
Blocks 2 6.380
Total 14 52.064
All the validity tests are satisfied as in the original data. The ratio of
the regression coefficients is
5.037
= 1.005
5.010
244 DESIGN AND ANALYSIS OF EXPERIMENTS
To obtain the relative potency this ratio has to be multiplied by the ratio
of the two highest doses of the two preparations included in the assay.
Q1
v(5,)=v(Zu)=-jj2xO.08=0.058
The fiducial limits of the estimated relalive potency can now be obtained
through Fieller’s theorem.
7.7.1 Introduction
The objectives with which the factorial experiments are interpreted are
two fold. One objective consists of studying the main effects and interac¬
tions of the factors. In this approach it is postulated that the response
from a treatment combination is a function of the main effects and interac¬
tions of different factors. However, these assumptions are not adequate
when the factors in the experiment are quantitative in nature. The second
approach of interpretation of data from factorial experiments consists of
studying the relationship between the response as a dependent variate on
the factor levels. This approach helps to understand better how with the
change in the levels of application of a group of factors, the response
changes. A combination of the levels of the factors which leads to
certain optimum response can also be located through this approach.
In order to get the relation between the response and the factor levels
the treatment combinations have to be suitably chosen. The treatment
combinations of the ordinary full factorial need not be the best for fitting
the relation. It is, therefore, necessary to search for a suitable set of
treatment combinations by using which a stipulated relation can be fitted.
Such a set of treatment combinations is called a response surface design.
Actually the factor-response relationship is called response surface. These
surfaces can be linear, second degree and third degree polynomials, and
so on. We shall discuss here only the first and second degree surfaces,
construction of designs suitable for fitting them and methods of
interpretation of the data.
In all the above equations the summation is over the design points. One
of the purposes of choosing a suitable design is to simplify the solution of
the above normal equations. The other purpose is to increase the precision
of the estimates of the parameters b0, blt... ,bv.
As we have to fit a linear relation the minimum number of levels of
each of the factors is 2. There is a series of linear response surface
designs with each of the factors at two levels. For each factor first a
range of the levels is fixed. For a factor like nitrogen fertilizer a range
like 0 to 80 lbs. of nitrogen per acre may be fixed. In specific cases the
range has, however, to be fixed according to the prevailing conditions,
taking into account the experimenter’s demand and available resources.
Preferably the two extreme doses defining the range are taken for the
experiment and denoted by the codes -1 and 1 for each of the factors.
Each of the treatment combinations will, therefore, be of the iorm
/ , 1 _L.\ that is, consisting of only 4-1 and — 1 as elements.
To simplify the solution of the normal equations the design has to be so
obtained that all the following relations hold:
(7.14)
2 Xiu xju— 0 (i^ji Uj 2,... v)
u-l
Under these conditions the normal equations reduce to
Nb0='Zyu
bi ^Xiif—^Xtu yu (,=2,.. .,v) (7.15)
246 DESIGN AND ANALYSIS OF EXPERIMENTS
XN2 ■ • X^y
(7.19)
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 247
Let ylf y2,...yN be the N observations obtained from the above N points.
To obtain a fit of the surface through the method of least squares, we
get the following normal equations:
27=iY7>0+612x1+^22x2+ ... +6112x12+fc222x22-f... +7>122x1;t2+...
2x1_F=&02Xi4-7?12.x'i‘'-f b^2x^x2-\-... —|-~L • • • T-
Z>122x12x2-|-...
• •• ••• ••• •••
2x1zy=b02x]J+b12x13+b22x12x2+ ... +bn2x1i+b^x12x22->r ... +
'Lx1x2y=b02x1x2+b12x12x2+... +Z>n2x18x2+... +
22xx2x22+32x12^2^3 + •••• (7.20)
All the summations in the above expressions are over the design points.
In order to simplify the solution of the above normal equations, the
design points have to be so chosen that all sums like 2xif 2XiXj, Ixf-Xj,
2x,3, IXiXjXk, 'Zx? ■ xj ■ Xk, 2xt3Xj, 2xiXjXkX/ are equal to zero. That is, all
sums of products in which at least one of the x's is with an odd power
are zero.
The other simplifying conditions are given below:
(i) 2xj2=constant - 7VA2, say, that is same for each factor
(ii) 2xj4=constant=cAA4
(iii) 2xi2Xj2=constant=iVA4
where c, A2 and A4 are constants.
For a design satisfying the above conditions, the normal equations reduce
to the following:
2y=Nb0-\-N'\i (bu-\-ba2~\~ •.. +^v») (1)
2Xiy=N?i2bi (z —1,2,..., v) (2)
2x{Xjy=N\bij (zV7=l, 2,..., v) (3)
2x12y=N’A2b0JrcNXib11JrN'Xi (b22-\-b23-\- ... +7>yy)
2x22y--NX2b0-\-cN\b22+N‘\t • •. +6VV)
* i • ••• •••
(4)
2xy2y= N\2b0+cAA47jv»+iVA4 {bllJrb22+... +&y-i,v-i)
The equations at (2) and (3) provide the solution of bi and btj immediately.
To obtain the solutions of the other equations let us put
t>ll +^22"t" • • •
Now equation (1) becomes 2y=Nb0-\-N'\2B0 (5)
Again, summing all the equations at (4) we get
2 (2x/2j) = vNX2b0+cAA4 I?0+TV A4 (v— 1) B0
i
=vN\b0+N\ (c+ v— 1) B0 (6)
248 DESIGN AND ANALYSIS OF EXPERIMENTS
For finding .the variance of y0, we have to obtain the variance of and
covariances between the estimated surface parameters. Now from the
theory of least squares it follows that the variance of any parameter
estimate, say, bu, is given by a2 times the coefficient of Zxfy (i.e. the left
hand side expression in the normal equations obtained by differentiating
with respect to bu) in the solution of bu. Similarly, the covariance between
any two estimates, say, cov (bu, bjj) is a2 times the coefficient of lx/y in
the solution of bu or o2 times the coefficients of Ixfy in the solution of
bjj. Here c2 is the error variance.
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 249
Applying this principle we get the expressions for variance and covariances
as follows:
r//2\ A4(c+v—l)<r2 T,/£x a2
(7.22)
F(*»)=s—m-’V{bl)=Ia2
K(ft;)=<r*m4
_z!_xr^ (c+v—2)-(v—1)A22
A
(c -1)7VA4X
A
L ]
Cov (b0, bn)=—(\<j*/Nk),
2
*">-(£-1)^A
Other covariances are zero.
An inspection of the variance of b0 shows that a necessary condition for
the existence of a second order design is A > 0. This actually leads to
the further condition that
W>v/(c+v-1). (7.23)
The variance of ,v0 is then given by
A second degree response surface design will be calleda second order rotatable
design if in this design c=3 and all the other conditions enumerated earlier
hold. Formally a second order rotatable design can be defined as below.
A design is said to be a second order rotatable design if a second degree
surface of the form
y=b0+b1x1+b2x2+... +iHx<a+... +bijXiXj+...
of the response as obtained from the design points on the variates
*;(/=l, 2, ..., v) with some suitable origin and scale can be so fitted that
the variance of the estimate of the response from any treatment combina¬
tion is a function of the sums of squares of the levels of the factors in that
treatment combination (see Box and Hunter, 1957).
After the surface has been fitted, it is necessary to obtain the analysis of
variance so as to test how good is the fit as also to obtain an estimate of
the error variance. The appropriate analysis of variance partitioning is
shown in Table 7.15.
TABLE 7.15
Analysis of Variance Table
Sources d.f. s.s.
Due to regression
2,+^ i| b02y+2bi(lx{y)+2bii
coefficients
J
(Z*<2jO+ 2 %ti($xixty)
11 —C.F. ^
Deviation from regression V'—2v—v. —1 ('Obtained by subtraction normally,
c2
or lack of fit j TV'=Number of district treatment
[combinations.
Error N—N' Obtained independently adopting
methods according to the actual
design.
-1-
Total N— 1 C.F.
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 251
7.7.7 Uses
One of the uses of the surface is to obtain the optimum dose combination
at which the response is either maximum or economically optimum. Spch
combinations are obtained by partially differentiating the surface relation
with respect to the x,'s and then equating such differentials to zero. There
will be v such equations, solving which the combination is obtained. For
obtaining the economical optimum dose combination the partial differentials
are equated to the price ratios px,/py where px, is the price per unit of the
jth input variate. In addition there are some other uses of the surface.
For example, with the help of such relations it is possible to investigate the
nature of the surface in specific ranges of the input variates x/s.
7.7.8 Construction
A full factorial of the series 3" or a suitable fraction of the same, in which
no interaction with less than 5 factors is confounded, will always provide a
second order response surface design when the origin is shifted to the
middle of the range of each factor and then a suitable scale change is
made. These designs are not rotatable but will have a value of 1.5 for c.
Method I
Let there be v factors. Each of these factors is taken at a certain number
of levels. Some of these levels are denoted by one or more unknowns like
a, b, etc. Thus, we shall have treatment combinations like aao.. .b. We
shall not take all such combinations but several of them as the situation
demands. Next, another design 2V is taken. In this design each of the
factors is taken at levels +1 and —1.
Each of combinations of the unknown levels, a, b, o, etc. which have
been taken for constructing the design is then associated with each of the
combinations of the 2V design. The rule of association is that the corres¬
ponding entries in the two combinations, viz., one involving the unknowns
and the other coming from the 2V factorial are multiplied and the products
are written in the same order. This last combination obtained from the
products is a design point. Thus out of each combination involving the
unknowns without sign we shall get design points after association. For
example, if we have the combination aab involving the unknowns a and b,
we have first to take a 2s factorial written by the levels +1 and — 1 for
each factor. This factorial is
252 DESIGN AND ANALYSIS OF EXPERIMENTS
-1 -1 -1
-1 -1 1
-1 1 -1
-1 1 1
1 —1 -1
1 -1 1
1 1 -1
1 1 1
Now by associating the combination aab with each of the above combina¬
tions of the 23 factorial we get
—a —a -b
—a —a b
—a a -b
—a a b
a —a -b
a —a b
a a -b
a a b
0 —a -b
0 —a b
0 a -b
0 a b
0 —a -b
0 —a b
o a -b
0 a b
It is seen that these eight points are not distinct but the last four are
repetitions of the first four. This has happened because by associating
+ 1 or —1 to zero we do not get different levels. Thus to get distinct
design points which is normally required, we have to associate a
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 253
o a b
a b o
b 0 a
0 a b
0 a -b
o —a b
0 —a -b
a b o
—a -b 0
—a b 0
a -b 0
b 0 a
b 0 —a
-b 0 a
-b o —a
254 DESIGN AND ANALYSIS OF EXPERIMENT
It is easily seen that the above points satisfy each of the conditions
2Xf*=»constant, 2;q4=constant and 2x/2x/*=constant. For the above¬
combinations we find
2x,2=4(a2+b2), 2x,4=4(a4+Z>4) and 2x,2*;2=4a2 b\
In order to satisfy the condition that 2 */4=32 x? xj2 we require that
4 (u4-fZ>4)=3x4a262
b2
or x2—3x+l=0 where x=-9.
a2
This condition thus gives an equation involving two unknowns and one
of the unknowns can therefore be chosen so as to satisfy the above relation.
The other unknown can be fixed arbitrarily.
It will be seen that the condition Xj\2 > v/(v+2) which follows from
VV > v/(v+c—1) for the general response surface designs, is not
satisfied in the above design. As a matter of fact if all the points in a
design are equidistant from the origin as in this design, then the above
condition is not satisfied. As a remedy, a central point of the form (0, 0, 0)
is taken and the above condition then holds. Addition of such central
points does not disturb any other requirement.
Thus, by taking say “a” as 1 arbitrarily, then obtaining b from the
above equation and finally taking a central point we get a second order
rotatable design in 13 points with each factor occurring at the five levels
-1.62, -1, 0, 1, 1.62
In order to transform the above levels to actual levels first the range
of the levels of each factor is fixed. Let such ranges for a factor be R„
and Rm. Thus —b corresponds to the minimum Z?0and b corresponds to
Rm. As the coded doses are linear transforms of the real doses, we
assume that the coded dose, say, y is connected with the real dose, say x,
through the relation y=a+(3x. Now two points on the line, viz. (R0, —b)
and (Rm, b) are known. Hence substituting these two points in the
equation a and p can be obtained. Once the relation j=a+px is thus
established the other real doses corresponding to —a,o and a can easily
be obtained from it.
a, a, a,..., a
b, o, o,..., o
o, b, o,..., o
o, o, b,..., o
o, o, o,..., b
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 255
(a a a a a)
(b o o o o)
o b o o o
o o b o o
o o o b o
o o o o b
We shall associate a (1/2) (2s) fraction with the first combination. The
\ fraction is obtained by confounding the 5-factor interaction in 2s. This
design will have thus 16+10=26 design points.
Then 2x,4=16a4+2 b*
and 2x;txi2=16a4
Hence we have the equation
16o4+2Z>4=3 x 16a4
giving bl—l6a*
i.e. b=2a
Taking, a=l, we get 6=2 and hence the five levels are —2, —1, 0, 1 and 2.
Method III
Use of balanced incomplete block designs for construction of second order
rotatable designs (Das and Narasimhan, 1962).
First the incidence matrix of a balanced incomplete block design is
taken. The element 1 in it is replaced by an unknown a. From the rows
of this matrix involving 0 and the unknown a, we shall get b combinations.
Each of these combinations is then associated with 2* factorial or a suitable
fraction, say 2*, of it. We shall thus get 2 x,4=r 2* a4 (taking k=p when
a fractional factorial is used) and 2x,2x/=A 2* a4. It is now required
that
2 X/4=32 x,2 x;2. That is, r 2r a*=3A 2* a4 or r=3A.
It may happen that such a BIB design actuaUy exists. In that case
no further combinations need be taken excepting a central point, which is
necessary as all such points are at a distance equal to ka2 from the origin.
256 DESIGN AND ANALYSIS OF EXPERIMENTS
For example in the BIB design v=7, b—7, r=k=3, A=l, r=»3A.
Hence, from the incidence matrix of this design we can get the
following design. The incidence matrix is given by
1 1 0 1 0 <f 0
0 1 1 0 1 0 0
0 0 1 1 0 1 0
0 0 0 1 1 0 1
1 0 0 0 1 1 0
0 1 0 0 0 1 1
1 0 1 0 0 0 1
a 0 a 0 0 0 a
lxi2xj2=8a*
Hence the condition 2 xf=yixt2 x? is automatically satisfied. The
unknown “a” can now be fixed arbitrarily.
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 257
If however, the relation r=3A does not hold in a BIB design, we have
to take further combinations of another unknown 6 as below.
If r < 3X, we have to take the combinations
b 0 0 ... 0
0 b 0 ... 0
0 0 b ... 0
0 0 0 ... b
For example in the design v=6=3, r—k=2, A=l, we have r < 3A.
The incidence matrix of this design is given by
1 1 0
0 1 1
1 0 1
a a 0
0 a a
a 0 a
6 0 0
0 6 0
0 0 6
EXERCISES
Standard Test
preparation preparation
(yg)
1. 840 580
2. 350 500
3. 420 570
4. 400 700
5. 540 800
6. 600 600
7. 560 550
8. 550 470
Estimate the potency of the test preparation of flaxedil relative to the standard
one together with its fiducial limits assuming that (1) tolerance is normally
distributed and (2) log-tolerance is normally distributed.
2. For the investigation of linearizing transformation of dose and response of the
stimulus “Gonadothorpic hormone” an experiment was conducted at the Central
Drug Laboratory, Calcutta during the year 1956-57. Immature rats of 30-40
grams weight were taken and injected with hormone doses of some specified
dilution for five consecutive days. The rats were killed on the sixth day and the
organs (uterus) were weighed. Four doses of a standard hormone preparation
were each tried on a batch of six rats, and the weights of uterus, so observed,
are given below:
Doses (iv) Weight of uterus (gm.)
2.5 18, 12, 13, 15, 14, 20
5.0 34, 40, 42, 50, 38, 28
7.5 35, 40, 39, 42, 45, 47
12.5 56, 50, 44, 40, 41, 50
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 259
Find metametric transformations for dose and response so that the relations
between the metameters is linear.
3. A 4-point parallel line assay on pituitary extracts in uterus of rats was conducted
at the Central Drug Laboratory, Calcutta in the year 1972. The effect studied
was the contraction of a horn of rats uterus. A horn of uterus of a rat was
suspended in a bath containing a solution of the following composition: (Nad,
9.0 gm/L, KC1, 0.42 gm/L, CaCl2, 0.12 gm/L, NaHC03, 0.5 gm/L, Dextrose,
0.25 gm/L, MgCl2, 0.0025 gm/L and water 1000 ml.). The bath was maintained
at a temperature of 32°C. A dose was then added to the bath and the maximum
contraction of the uterus was measured on a kymograph. For the other doses
also, the same horn was used after washing and giving rest. Each dose was
tested four times. The effect of each dose expressed as the height of contraction
of uterus in m.m. are given below:
Responses in m.m.
(0.01 of standard) 33.0 35.0 35.0 34.0
S2 (0.02 of standard) 46.0 43.0 44.0 44.0
Ti (0.01 of test) 36.0 37.0 36.0 35.0
Tg (0.02 of test) 46.0 44.0 45.0 46.0
Assuming that the response is linearly related to log dose, estimate the potency
of the test preparation with its fiducial limits after applying necessary validity
tests.
An experiment on rats was conducted at the Central Drugs Laboratory, Calcutta
during 1963 in assaying the potency of corticotrophi (ACTH). Fortyeight rats
nearly of the same age and weight were taken and made into six equal groups
at random. Morphine sulphate or hydrocortisone was first injected to the rats
and after an interval of two to three hours the Requisite dose of the drug was
injected. After a certain specified interval the adrenal glands of the rats were
taken out and ascorbic acid contents was estimated. The observations in regard
to ascorbic acid were expressedas mg/100 gram of the adrenal gland and are as
given below:
Si s2 s3 Ti T% t3
. 25 mu. .50 mu. 1.0 mu. .25 mu. .50 mu. 1 .Omu.
298.7 259.7 230.4 274.5 176.7 212.6
280.5 250.1 228.1 252.5 231.2 209.7
279.2 193.5 199.1 245.5 252.5 223.6
240.5 259.2 216.9 206.1 219.2 171.4
362.1 256.7 201.8 325.7 218.2 189 3
304.0 228.1 238.8 339.4 222.1 172.9
233.8 272.2 225.1 210.4 189.6 168.0
304.6 284.1 189.8 247.0 210.0 190.7
Estimate the potency of the test preparation with its fiducial limits after applying
the necessary validity test.
260 DESIGN AND ANALYSIS OF EXPERIMENTS
5. For estimating the riboflavin content of malt the following common zero 5-points
slope-ratio assay was designed. At each dose four tubes for the riboflavin content
of malt were tested. The dosages tried were:
(i) Control or blanks
(ii) 0.10 gg and 0.20 [xg of riboflavin as two doses of standard preparations
(denoted by Xs) and
(iii) 0.025 fig and 0.050 [xg of malt as two doses of test preparation (denoted
by XT).
The response measured as titers in milliliters0. IN (NaOH) are given below:
*,=0
II
I
Assuming that the relationship between dosage and response is linear, estimate
the riboflavin content of malt through the estimation of its relative pontency
after applying the necessary validity tests. Find the fiducial limits of the estimate.
6. Consider the problem of estimating the parameters in the equation
*i *2 *3 y
—1 —1 —1 50
1 —1 —1 46
—1 1 —1 29
1 1 —1 47
—1 —1 1 49
1 —1 1 52
—1 1 1 51
1 1 1 36
•—2 0 0 45
2 0 0 52
0 —2 0 49
0 2 0 55
0 0 —2 37
0 0 2 42
0 0 0 44
DESIGNS FOR BIO-ASSAYS AND RESPONSE SURFACES 261
8. The following are the design matrix and vector of responses of a 22 factorial
design used to fit a planar response function.
(a) Estimate the first order response function and plot lines of constant response
in the (xj, x2) plane.
(b) Provide an analysis of variance table showing significance tests on each
coefficient:
*1 A'2
-1 -1 y= r 54.7
—1 -1 50.3
1 —1 59.7
1 —1 60.3
—1 1 63.7
—1 1 64.0
1 1 67.5
1 1 69.2
0 0 61.5
0 0 62.3
0 0 62.0
0 0 _ 61.9 _
9. A simplex design in two variables is used to measure the effect of the variables
on a single response. The following represent the design matrix and the vector
of observations.
“ x2 y
1.0 0 96.5
—1.0 0 55.4
0 0 98.6
0 0 94-6
0 0 94.4 _
(a) Estimate the response function and determine the stationary point.
(b) Plot contour of constant response in two dimensions.
(c) Perform an analysis of variance including the sources of variation due to:
linear terms, quadratic terms and lack-of-fit.
REFERENCES
2. Buss, C.I. (1952). The Statistics of Bio-Assay with Special Reference to Vitamins
(reprinted with addition from vitamin methods) Vol. II, Academic Press,
New York.
3. Box, G.F. and Hunter, J.S. (1957). “Multifactor experimental design for
exploring response surfaces”, Ann. Math. Stat. 28, 195-241.
4. Das, M.N. (1957). “Bio-assays with non-orthogonal data”. Jour. Inc. Soc.
Agri. Stat., 9, 67-81.
5. Das, M.N. (1958). “Reinforced incomplete block designs”. Jour. Ind. Soc.
Agri. Stat., 10, 73-77.
6. Das M.N. (1958). “Circular designs”. Jour. Ind. Soc. Agri. Stat., 12, 45-56.
7. Das, M.N. (1964). “A somewhat alternative approach for construction of sym¬
metrical designs and obtaining maximum number of factors”. Cal. Stat.
Assoc. Bull. 13.
8. Das, M.N. and Kulkarni, G.A. (1966). “Incomplete block designs for bio-
assays”. Biometrics, 22, 706-29.
9. Das, M.N. and Mehta, J.S. (1968). “Asymmetric rotatable designs and ortho¬
gonal transformations”. Technometrics, 10, 313-22.
10. Das, M.N. and Narasimhan, V.L. (1962). “Construction of rotatable designs
through B.I.B. designs”. Ann. Math. Stat., 33, 1421-39.
11. Das, M.N. and Nigam, A.K. (1966). “On a method of construction of rotatable
designs with smaller number of points controlling the number of levels”. Cal.
Stat. Ass. Bull., 15, 147-57.
12. Draper, N.R. (1960a). “Second order rotatable designs in four or more
dimensions”. Ann. Math. Stat., 31, 23-33.
13. Draper, N.R. (1960b). “Third order rotatable designs in three dimensions”.
Ann. Math. Stat., 31, 865-74.
14. Draper, N.R. (1960c). “A third order rotatable design in four dimensions”.
Ann. Math. Stat. 31, 875-77.
15. Draper, N.R. (1961). “Missing values in response surface designs”. Techno¬
metrics, 3, 389-98.
16. Finney, D.J. (1964). Statistical Methods in Biological Assays, 2nd ed., Griffin,
London.
17. Fisher, R.A. and Yates, F. (1963). Statistical Tables for Biological, Agricultural
and Medical Research, 6th ed., Oliver and Boyd, Edinburgh.
18. Gardiner, D.A., Grandage, A.H.E. and Hader, RJ. (1959). “Third order
rotatable designs for exploring response surface”. Ann. Math, Stat., 30,
1082-96.
19. Herzberg, A.M. (1966). “Cylindrically rotatable designs”. Ann. Math. Stat.,
37, 242-74.
20. Herzberg, A.M. (1967). “Cylindrically rotatable designs of types 1, 2 and 3”.
Ann. Math. Stat., 38, 167-76.
21. John, P. W.M. (1971). Statistical Design and Analysis of Experiments, Macmillan,
New York.
22. Myers, R.H. (1971). Response Surface Methodology, Allyn and Bacon, Inc.,
Boston.
23. Plackett, R.L. and Burman, J.P. (1946). “The designs of optimum multi-
factorial experiments”. Biometrika, 33, 305-25.
24. Raghavarao, D. (1963). “Construction of second order rotatable designs
through incomplete block designs”. J. Ind. Stat. Ns\s., 1, 221-25.
25. Tocher, K.D. (1952). “The design and analysis of block experiments”. Jour.
Roy. Stat. Soc., 13, 14, 45-100.
26. Tyagi, B.N. (1964). “On construction of second and third order rotatable
designs through pair-wise balanced and doubly balanced designs”. Cal. Stat.
zDj. Bull., 13, 150-62.
CHAPTER 8
We have seen that the devices of local control, confounding and increase
in the number of replications are adopted to increase the precision of the
estimates of treatment contrasts. Analysis of covariance is another device
with the same objective. This technique can be applied when observations
on one or more correlated variables are available from each of the experi¬
mental units in a design, along with the observations on the variate under
study. These additional variates are called ancillary or concomitant
variates. It is necessary that these variates are associated with the variate
under study.
The rationale behind the technique is as follows. Let x denote an
ancillary variate and y, the variate under study. When y and x are
associated, a part of the variance of y is due to the variation in x oyer the
experimental units. So, had the variate x been constant over the units,
there would have been a corresponding reduction in the variance of y.
The purpose of analysis of covariance is to effect such reduction in the
estimate of the variance of y by analytical means when the observations
on the x-variate are available along with the observations on the y-variate.
Basically, what is done is, to estimate the relationship between x and y,
taking x as the independent variate, after eliminating the effects of blocks
and treatments and then obtain the values of y as expected from this
relation corresponding to each of the values of x. These expected values
of y are then analyzed to get the error variance of y.
For the purpose of the present chapter we shall assume that the relation
between x and y is linear and the values of x are not subject to error.
For the application of this technique it is necessary that the x-variate
is not affected by the experimental treatments, otherwise a part of the
treatment effect gets eliminated along with the elimination of the effect of
the x-variate.
This technique will be discussed in Section 8.2 with the help of an
appropriate model.
T _Bj-bBj^x) ~
b,=—^
v
f/—2 \Ti(X)
—-2-p--v BjB](X) GGx
S xuyij
ij rk Exy
T2 say.
Hx) ~B*n
m ,Gi eTx
2 xJj — 2
u •' i r rk
The error sum of squares adjusted for the variation of x is obtained by
substituting the above estimates in
2 (yu —bj - bxu)2
ij
and is given by
E=2(yu-y.- -ti—bj—bxij)2 (8.2)
u
rr2 n2 /^»2
i Y-H+Tk-~bE- 'xy-
and 4=2V‘2%.‘
y A:
The adjusted treatment sum of squares is now Ex—E with k—l degrees of
freedom.
The various calculations needed for the covariance analysis are shown in
the following tables.
In Table 8.2 Txx, Txy and Tyy stand for the sum of squares for x, sum of
products of x and y and sum of squares for y, respectively, in the “treatment
line” of Table 8.1.
From Table 8.2 the mean square for the adjusted treatment and error
sums of squares are obtained. The hypothesis of equal treatment effects
is tested by F-test where F is the ratio of these two mean squares.
The estimates and the variance of different treatment contrasts are
obtained as below.
=yt—y—b(xi—x), (say)
for i=l,..., k.
Estimate of any contrast 2 Ut{ is then given by
i
r -2 (2/,*,)>
_ (j*
F(2/,r,)=o* 2
i L/ r
—
]
because V(b)=p-.
In particular,
(8.5)
and <j2 is estimated by the error mean square obtainable from Table 8.2 as
E
Em= (8.6)
(r—1)(*— i)-r
_ It is seen that the variance of the estimate of the difference between any
pair of treatments is not constant, as we get in analysis of variance of this
design. Sometimes to make the critical difference (C.D.) a constant for
each pair of treatments, (xt—xm)2 in the above variance of if
replaced by T^/fJc— 1), the treatment mean square for the x-variate.
DESIGN AND ANALYSIS OF EXPERIMENTS
266
ANALYSIS OF COVARIANCE AND TRANSFORMATION 267
For both these designs the technique is exactly the same as the one
developed for the randomized block designs. The procedure indicated
for randomized block design holds, as a matter of fact, for any design
once the treatment and the error sum of squares for the x-variate, the
sums of products for the two variates and the sum of squares of the
j-variate are obtained according to the appropriate analysis of variance
of the design. For these two designs or rather for any orthogonal design
the following table (8.3) is first prepared. Only the treatment and error
sums of squares have been shown in the table, as these alone are needed.
Estimate of treatment effects
U=yt—y -b (xt—x)
— f /i2 ^
(8.7)
C/5 ■*
4 s
>1 »3
E
(N
l 1 -*
-d ts f
“a
>»
X
5S
,bq ii»
ha
>-Cj 1
|
1 »s ss
hT h!T ha
II
II —4
ha ha
eae
x
u hI *
X X
O halha
II 1
.0 I-CJ
w z
« 2
>>
Eyy
£ U4
»> X
X
X
ha
4 U
8 X X
s? *3
+
/—S
*—«•
1 ^3 t
'w' +
CO
■*-*
c
u |
O 1
to
1
1
E ^ +
3
tS o S
O u
©
uU.
E tH
1/3 H W W
ANALYSIS OF COVARIANCE AND TRANSFORMATION 269
TABLE 8.4
270 DESIGN AND ANALYSIS OF EXPERIMENTS
TABLE 8.5
Adjusted Sums of Squares
- Exy
Error Exx EXy Eyy Os=s'Er~ E=Eyy—~bExy n..'-m.—n.j
JZxx
r, E'*> E1=E'yy-b'E'xy n.
Treatment+error E’xx E’xy E’yy
b ~ E’xx
2lii,='Zl,(iKy)-btt{x)) (8.8)
i i
where t{(,) and *«*) denote the estimates of the ith treatment effect
obtained by using respectively the y and x variates. The variance of
2 U U can be obtained by the method indicated earlier in connection with
the analysis of variance of non-orthogonal data in Chapter 1.
Actually
f (2 htj(x))2)
hh)=°21,q,+ 1 Ex—\ (8.9)
«s
'w'
1
■*-*
£
II §
c? 2
i C* %
W* w- m w=s
S CD el
J >— * v-/
Q |
« w- +-*
B §
15. eg 3
S 3 -O
oo <5 dC/5
J9 eg I'i* >» X
w- W' « W=>
TABLE 8.6
J
«! d
H I! .2
CO H
V—'
o
<5 a
2
eg v—' 3
ic: Cfi
X
w-> W M M
+
I I I
-o o
I
a
.o
!
o d
4>
e
§ o c3 o
s o <u fi O
o
CO 5 w H
272 DESIGN AND ANALYSIS OF EXPERIMENTS
If ancillary data are available from each experimental plot for more than
one ancillary variate and each of them is linearly related to the variate y
under study, then the covariance technique can be extended to eliminate
the portion of variation in y which is due to the ancillary variables. We
shall discuss here the case of two ancillary variables denoted by x and z.
The data from a randomized block design with k treatments and r repli¬
cations have been analyzed. This technique can be modified to suit other
designs exactly as in one ancillary variate.
We take the following model:
Ytj={*+tt-f-b)+PiX/;++eu
where (jl, tt, bj and etj have the same meaning as in the model for one
ancillary variate, xtj, ztj are the observations corresponding to ytj from the
x and z ancillary variates respectively; and p2 are the partial regression
coefficients of y on x and z respectively.
Let
2 yu—G, 2 Xij=Gx, 2 zi)=Gz
U U U
2 yu—Bj, 2 xij=Bj(X), 2 zi)=Bj(z)
1 I i
Let further the corrected sums of squares and products for x, z and y
variates be denoted as shown
Block s.s. (x) =BXX Treatment s.s. (x) =TXX Error s.s. (x) =EXX
Block s.s. (z) =Bzz Treatment s.s. (z) —Tzz Error s.s. (z) =Etz
Block s.s. (y) =Byy Treatment s.s. (y) —Tyy Error s.s. (y) —Eyy
Block s.p. (xz)=Bxz Treatment s.p. (xz)—Txz Error s.p. (xz)=Exz
Block s.p. (xy)—Bxy Treatment s.p. (xy)=Txy Error s.p. (xy)=Exy
Block s.p. (zy)=Bzy Treatment s.p. (zy)=Txy Error s.p. (zy)=Ezy
By the method of least squares the estimates of the fixed constants and
the regression coefficients are given by
{*—y -PiX—(J*z
, — G — Gx , — Gg
where , =^. * and
ANALYSIS OF COVARIANCE AND TRANSFORMATION 273
where & and jf8 are obtained by solving the following two equations:
with r(k— 1)—2 degrees of freedom where p'x and p'2 are obtained from
the following equations
$\E'xx + $\E,xz=E'xy (8-11)
V^E’xz+ViEzz^E'xz.
The adjusted treatment sum of squares is obtained from Ex—E with k—1
degrees of freedom.
The hypothesis of equal treatment effects can be tested by the F-test
where F is based on the mean squares corresponding to the error and
treatment adjusted sums of squares.
Estimate of U—tm is given by
ti-tm=yx-ym—Pi(x, - xm) - p2(z,—? m)
-ExzO2
(8.12)
V^)=sExxEzz-E^’ C0V (Pl’Pa) EzxEzz-Exz2 •
Treatment Error
Error+Treatment
r 1 I
J *X = -r p 1 1 1 , 1
r rk E’xx= I-p
r k v rk k
T _J 1 E _i 1 1.1
** r rk
T 1
X2 rk •£»*=0
rp Tm\ G
Ixy— -r p _ (Tm i Bmi
p, BmL
r rk
Ex'>—y—+—- rk) E
rp 7m, G
p _ (Tm% Bmt
F, _Bmt
zv“ r ~7k
rkJ ** T
Tyy, Eyy and E'yy are obtained as usual for the design. The equation foi
the partial regression coefficients can now be written and solved as indicated
earlier. The rest of the analysis can also be completed likewise.
. Estimate of the difference between two treatments effects, one of which
is affected, is shown below.
=yi-ym+f-1.
v (Ji~Tm)~a2 +_ El
r%(ExxEzz:~EX/)}
=-124- — (r—0(^—1) )
r 1 ^ (r—1)2 (/t—1)2_ i J*
These are obtained after substituting the values of Exx, etc. as given above.
Pi-P
—Tmi Jm,'f" (8.13)
r y^rk-r-k)'
8.7 Transformations
We have seen that the main technique for the analysis of the data collected
from various designs is the method of analysis of variance. This method
is applicable to data which satisfy the following assumptions:
DESIGN AND ANALYSIS OF EXPERIMENTS
276
=sin“1 f*
v n
when c is chosen suitably. The asymptotic variance of the transformed
variable is —. Anscombe (1948) proved that a slightly better transforma-
4n
tion is
.v-1-3/8
F(x)=sm~1
5T+3/4
1
which has asymptotic variance .
A2
1. A2m2 (A constant) log* (X), log* (JV+1)
logjo (X), logio (■X'+l)
A2
2. A2m2 (1 —m)2 log e (x)/(l —X)
1 In— 3
3. (1 — m2)2/n—1 £loge(l+ *)/(!-*)
1/4
4. m+A2m2 A-1 sin /r1 (AVX)
F(r)= [ YjlSjL-tanh-» r
' J (1-/*)
choosing c=\/y/n it has a constant asymptotic variance which is equal to
2
c2= -1
n'
For the purpose of analysis the data are first transformed appropriately
and then the transformed data are analyzed according to the design. For
example, if the data correspond to binomial proportion then these are
transformed by sin-1 \/p. The values of sin-1 y/p for different p have
been tabulated by Fisher and Yates (1948). For this transformation it is
necessary that if />=-, then n should be the same for all the observations,
otherwise the variance does not remain constant.
The inferences in regard to different hypotheses can be drawn from
the analysis of the transformed data. The different averages of the
original data which are needed for a meaningful interpretation are
obtained by the inverse transformation of the averages obtained from the
transformed data.
EXERCISES
E( Yij)=+ 0xt- j
V(Yii) = <5 2
where x,/s are observed fixed quantities, n is a constant, «,• is the ith treatment
etiect, 0 is the regression coefficient.
(u) Find the distribution of 0, a2 and show that they are statistically independent
(m) Find the confidence interval of a2 w|th confidence coefficient 1-a.
(iv) Test the hypothesis /70; 0—0.
(v) Test the hypothesis H0 : a1==... =«(.
2. If the relation between an
ancillary variate x and the study variate y is quad¬
ratic, deduce the computational procedure of the analysis of covariance for
adjusting y for x.
3. Obtain the computational procedure for the analysis of covariance of a B.I B
design with one missing observation.
4. With a view to making a selection for high yielding hybrid maize, a varietal trial
was conducted in 1966 at the Vivekananda Laboratory in Almora, Uttar Pradesh
The trial was laid out in randomized block design with four replications At the
time of harvest the number of plants harvested for recording the yield were also
noted. The data obtained are given below;
Grain yield (in lb.) and the number of plants (given in parentheses) harvested ner
plot (48'x 12') are given below. y
ANALYSIS OF COVARIANCE AND TRANSFORMATION
279
Analyze the yield data by using information about the number of plants harvested
and draw conclusions. Calculate the gain in efficiency as a result of using addi¬
tional information of the number of plants harvested. Give the standard error
of the difference of the means of varieties N.C. 27 and V.L. 23 after adjusting the
means for the unequal number of plants harvested.
5. Let X be a Poisson variate with mean m. Expanding y/7+b in powers of
(*—I*) — t obtain first a few terms of the mean and variance of y/x+b. Hence
show that the choice of 6=3/8^ makes the asymptotic (for large (jl) variance of
y/x+b smaller than that of ^x.
REFERENCES
♦ _
1. Anscombe, F.J. (1948). “The transformation of Poisson, binomial and negative-
binomial data.” Biometrika, 35, 246-54.
2. Bartlett, M.S. (1937). “Some examples of statistical methods of research in
agriculture and applied biology.” J. Roy. Stat. Soc. Suppl., 4, 137-83.
3. Bartlett, M.S. (1936). “The square root transformation in analysis of variance.”
J. Roy. Stat. Soc. Suppl., 3, 68-78.
Weighing Designs
9.1 Introduction
Yates (1935) showed that if several light objects are weighed in groups
rather than individually as customary and next the weights of the indiyidua
objects are estimated by the method of least squares, then the precision ol
the estimates increases quite considerably. In the scheme suggested y
Yates the objects are always placed on the same pan in a chemica
balance or on a single pan in a spring balance. Hotelling (1944) showe
that Yates’ scheme can be improved by distributing the objects on both
the pans in a chemical balance. In this case the precision of the
estimates of the weight of the objects increases further. Yates illustrated
the technique by providing a scheme for weighing seven objects in eight
weighings and showed that the variance is £ times of that obtainable from
the customary method when correction for the needle deviating from zero
position when the pans are empty, is needed. By modifying the scheme
of Xates s0 as to place each individual weighing the group of objects
as in Yates scheme on one pan and the remaining objects on the other pan
and then balancing by placing suitable weight on one of the pans,
Hotelling showed that the variance of the estimate of the weights of
individual objects becomes half of that obtainable from the Yates scheme
for seven objects. In the scheme of Hotelling eight objects were taken
instead of seven, the extra object being placed always on the same pan
together with the others as indicated earlier.
Following Yates and Hotelling several authors provided methods of
construction and analysis of such designs, together with investigation of
the precision of such designs. Prominent among them are the works of
Banerjee (1948, 1949, 1950), Kempthome (1949), Kishen (1945), Mood
(1946), Raghavarao( 1959, 1960), Dey(1969, 1971) and others.
9.2 Definition
pan and the third of size nn is omitted from the weighing. There will thus
be one weighing for each of the N groupings.
As indicated earlier there are two types of balances, the chemical
balance and the spring balance. In chemical balance the objects can be
placed either on one pan or on both for each weighing. In spring balance
there is only one pan for placing the objects. If the objects are placed on
•one pan only, say^ the right one in a chemical balance, then n2~0 and
n0=p—n1. If the objects are distributed on the two pans, then n0=0 and
n1=p—n2. In Spring balance n2 is always zero and n0=p—n1. When the
objects are placed on one pan only in a chemical balance, we shall call
the weighings one pan weighing and the design one pan design. If,
again, the objects are distributed over both the pans, we shall call the
corresponding designs two pan designs. Evidently, the spring balance
designs are one pan designs just like the chemical balance design when
the objects are placed on one pan only. Instead of classifying the
weighing design as chemical balance and spring balance designs we shall
classify them as one pan and two pan designs.
As the weights of individual objects have to be estimated by the
method of least squares, the groupings involved in a weighing design are
to satisfy certain conditions to make the weights estimable. Thus,
construction problems arise for obtaining the groups, both for ensuring
estimability and increase of precision.
where aj, takes the values +1, -1, or 0 according as the ith object is on
the left pan, right pan or does not occur in the 7th weighing and e is the
error component having a constant variance, a2.
In a two pan weighing if the balancing weight is placed on the right
pan then the observation consists of this weight as such. If the
balancing weight occurs on the left pan, the observation then consists of
minus one time the balancing weight. Thus, Yt denotes an observation
an dp 1 '
\
an #22 ••• #p2 f
•
alN #2 N • • • apN /
/
W is the column vector of the weight of the objects and E is the vector
of the error component in the different observations.
The least squares estimates of W are given by
fV^CA'A)-1 (A'Y)
The normal equations from the least squares theory are as below:
(vA+r—A) 2 w/=2 Tj
«r^(r'-T) <9-4>
of wt, because this coefficient occupies the diagonal position in the ith row
of (A'A)-1
In some balances the pointer needle does not settle on the zero position
when both the pans are empty. Such balances are said to need zero
correction and an observation from a weighing on it is said to be biased.
If for a one pan design the estimate of weight comes out as a < contrast
among the observations, this bias which is constant in all the weighings
in a balance, gets automatically eliminated and hence no separate
correction for zero bias is needed.
As the estimate at (9.4) is not a contrast it is not free from any
possible bias.
where Tt is the sum of the observations from those weighings which involve
the zth object including R.
DESIGN AND ANALYSIS OF EXPERIMENTS
284
* ro2 J. k r+k 1
v<w')“(7--W
The estimate at (9.8) though a contrast, is not the least square estimate.
Its variance is greater than the variance of the estimate at (9.7) which is
the least squares estimate.
By taking the B.I.B. design with
v=b=7, r=k=3, A=l,
WEIGHING DESIGNS 285
along with an extra weighing involving all the objects, we get the design
used by Yates (1935) as illustration. The estimate provided by Yates
follows from that at (9.9). Thus, the estimate suggested by Yates is not
the least squares estimate, but it has the advantage that it is free from
zero bias.
It would be noticed that for a B.I.B. design without the extra weighing
involving all the objects the expression for the estimate of weight and that
for the inter-block estimate of B.I.B designs are the same. In both these
situations the same model is used.
where s denotes the weight of the extra object and 5 in the sum of all the
(6 + 1) observations.
Equation (9.11) can be written as
(r+1)s+(r—A) W/+(A+1) 2 = (9.12)
Adding these equations over i
v(r+l) 5+{r(X+l)+r—A} 2 w,=^T, (9.13)
r+k-\-\—v)
Var.
f r(v-Ar)2 f
iy+rk)
Var. (j)=o2
(b-r) (v-k)
It is seen that for the series of original B.I.B. designs having parameters
v=4A-|-3, r=k=2h-\-\, A, Var. (s) is greater than Var. (w,) even though
the extra object is involved in b+1 weighings which is larger than the
number of weighings in which other objects are involved, viz. (r+1).
This seems to be contrary to expectation.
where T\ is the sum of the observations from those weighings in which the
ith object is involved, minus the sum of the remaining observations.
These equations can also be written as
that is v * 2T,
bv—4 (r—A)(v—1)
G
“(2 r-by
since 2 T,=(2k-v) G
and 6v—4(r—A) (v-l)=(2A:_v) (2r-b)
6—4r+4A 4A 1
A)=4<^d l-
(b-2r) (v--2k)\
It will be seen that no estimation of the weight is possible for such
B.I.B. designs with v=2k, as estimate of 2 w} is not possible. If, however,
an extra weight involving all the objects be taken, the estimates are available
for all B.I.B. designs. The normal equations in this situation are
We have already discussed one pan reinforced designs using B.I.B. designs.
These designs can also be used for two pan weighing designs. One
extra object is taken in each weighing and is placed on the same pan every
time. Further, one more weighing involving all the (v+1) objects is taken.
The normal equations for these designs are
v=b=4X+3, r=k=2X+1, X
Substituting the parameters of the B.I.B. desings of the series v=6=4X+3,
r_£_-2X+l, X in the results obtained in Section 9.5.1 we get the estimate
of vvj as
/v T, __ Tt
w,~4 (r-X) 4(X+1)
since for this series 2r-b+l=0 and &-4r+4X + l=0,
- S
4 (X+l)
2
v(»,)='’W=4^pjy-
A
It will be seen that for this series of designs the estimates, wt are con¬
trasts, but not s. Even though vv/’s are contrasts, zero bias, if any, is not
eliminated automatically from the estimate because the bias does not occur
with the same sign in all the observations. As in a two pan design the
balancing weight is put on any of the pans depending on the weights of
the individual objects, the bias does not remain constant in the different
weighings but changes sign. If in a biased balance the weight of an object
is observed as w+B when placed on the left pan, it will be observed as
w—B when placed on the right pan. Here w is the true weight of the
object and B stands for the bias. One way out of this difficulty is to take
a known weight which is big enough so that when it is put always on the
same pan, say, the left pan, the balancing weight always comes on the
other pan. In this situation the estimates w.-’s are unbiased but not s
which has to be adjusted for the extra known weight and possible bias in
the balance.
For these series of designs the variance of any of the weights is g2/N.
As for this series of designs the number of objects is equal to the number
WEIGHING DESIGNS 289
of weighings, it is not possible to get the estimate of error from it. To get
the estimate of error variance, the designs may be repeated.
When 4A+3 is a prime or a prime power, a method of construction of
such designs has been given by Bose (1939). For other cases these designs
can be obtained from Plackett and Burman (1946). These designs are also
provided by the Hadamard matrices by taking the objects corresponding to
+1 on one pan and those with — 1 on the other.
Two associate P.B.I.B. designs with b ^ v can be used as one pan weighing
design just as in the B.I.B. designs.
Denoting the parameters of the P.B.I.B. designs as usual, the normal
equations for such designs are
2 m-w+S,
where G is the total of all the observations.
Similar modifications are needed for the other series of designs also. It
has to be kept in mind that while estimating 2 wj the summation of the
normal equations is to be taken over the existing v—q treatments
(objects) As a result the forms of the estimates will change.
Through this device the number of weighings can be increased relative
to the number of objects. As such when symmetrical B.I.B. designs are
used, error d.f. can be obtained through such truncation.
9.8 Efficiency
REFERENCES
2. Banerjee, K.S. (1949). “On certain aspects of spring balance designs.” Sankhya,
9, 367-76.
3. Banerjee,K.S. (1950). “How balanced incomplete block designs may be made
to furnish orthogonal estimates in weighing designs.” Biometrika, 37, 50-58.
4. Banerjee, K.S. (1975). Weighing Designs. Marcel Dekker, New York.
5. Bose, R.C. (1939). “On the construction of balanced incomplete block designs.”
Ann. Eugenics, 9, 353-99.
6. Das, M.N. (1958). “Reinforced incomplete block design.” Jour. Ind. Soc.
Agri. Stat., 10, 73.
7. Dey, A. (1969). “A note on weighing designs. Ann. Inst. Stat, Math., 21,
343-46.
8. Dey, A, (1971). “On some chemical balance weighing designs.” Aust. J. Stat.,
13, No. 3, 137-41.
9. Hotelling, H. (1944). “Some improvements in weighing and other experimental
techniques.” Ann. Math. Stat., 15, 297-306.
10. Kempthorne, O. (1949). “The factorial approach to the weighing problem.”
Ann. Math. Stat., 19, 238-48.
11. Kishen, K. (1945). “On the design of experiments for weighing and making
other types of measurements.” Ann. Math. Stat., 16, 294-300.
12. Mood, A.M. (1946). “On Hotelling’s weighing problem.” Ann. Math. Stat.,
17, 432-46.
13. Plackett, R.L. and Burman, J.P. (1946). “The design of optimum multifactorial
experiments.” Biometrika, 33, 305-25.
14. Raghavarao, D. (1959). “Some optimum weighing designs.” Am. Math. Stat.,
30, 295-303.
15. Raghavarao. D. (1960). “Some aspects of weighing designs.” Ann. Math.
Stat., 31, 878—84.
16. Yates, F. (1935). “Complex experiments.” J. Roy. Stat. Soc. Suppl., 2,
181-247.
Index
accuracy, 3 one-way, 19
additivity, 17 two-way, 19
adjusted sum of squares, 37 comparative dillusion assay, 218
totals, 36 complementary design, 165
affine resolvable designs, 161 completely rendomized design, 48
alias, 108, 109 analysis of, 50
analytical dillusion assay, 217 design, 48
analysis of covariance, 44, 108, 109, 263 concomitant variate, 44
analysis of variance, 8, 10, 50, 54, 59, 82, confounding, 86, 95
105, 166, 173, 183, 185 partial, 87, 95
ancillary variate, 263 total, 87, 95
A-optimum, 198 connected design, 155
association, matrix, 179 contrast, 10, 11
scheme, 177 analysis of variance, 10
assumptions, analysis of variance, 43,276 maximum number, 12
asymmetrical factorials, 120 method of writing, 12
analysis of, 133 orthogonal, 11
balance in, 122 sum of squares, 11
confounding, 120 control treatment, 80
correction term, 25
balanced data, 42 critical difference, 22, 56
B.I.B. designs, 153, 156 cyclic design, 182
analysis of, 166
construction of, 158 defining contrast, 108
efficiency, 167, 168 degrees of freedom, 5, 18, 21
tinter-clock analysis, 168 direct assay, 218
parameters, 156 dose response relation, 221
parametric relations, 157 dual design, 166
randomization, 154 Duncan’s test, 23, 56
repeated designs, 158
symmetrical design, 158 efficiency factor, 167, 168, 176
bar diagram, 22, 23 E-optimum, 199
bias, 7 error control, 37, 10
binary designs, 155 estimation,
bio-assays, 78 maximum likelihood method, 3
designs, 217, 219 theory of, 3
direct assays, 218 experiment, principal, 3, 7
indirect assays, 220 experimental error, 8
standard preparation, 220 experimental units, 2
test preparation, 227 groupings of, 3,10