Beruflich Dokumente
Kultur Dokumente
It may be noted that it has already been described how to develop the likelihood ratio tests for the
testing the hypothesis of equality of more than two means from normal distributions and now we
will concentrate more on deriving the same tests through the least squares principle under the setup
of linear regression model.The design matrix is assumed to be not necessarily of full rank and
consists of 0’s and 1’s only.
where
is the general mean effect.
- is fixed.
- gives an idea about the general conditions of the experimental units and treatments.
i is the effect of ith level of the factor.
- can be fixed or random.
1
Example: Consider a medicine experiment in which there are three different dosages of medicines
- 2 mg., 5 mg., 10 mg. which are given to patients for controlling the fever. These are the 3 levels of
medicines, and so denote 1 2 mg., 2 5 mg., 3 10 mg. Let Y denotes the time taken by the
medicine to reduce the body temperature from high to normal. Suppose two patients have been given
2 mg. of dosage, so Y11 and Y12 will denote their responses. So we can write that when 1 2mg is
given to the two patients, then
E (Y1 j ) 1 ; j 1, 2.
Similarly, if 2 5 mg. and 3 10 mg. of dosages are given to 4 and 7 patients respectively then
the responses follow the model
E (Y2 j ) 2 ; j 1, 2,3, 4
E (Y3 j ) 3 ; j 1, 2,3, 4,5,6, 7.
Here denotes the general mean effect which may be thought as follows: The human body has
tendency to fight against the fever, so the time taken by the medicine to bring down the temperature
depends on many factors like body weight, height, general health condition etc. of the patient. So
denotes the general effect of all these factors which is present in all the observations.
In the terminology of linear regression model, denotes the intercept term which is the value of the
response variable when all the independent variables are set to take value zero. In experimental
designs, the models with intercept term are more commonly used and so generally we consider these
types of models.
indicates the variations due to uncontrolled causes which can influence the observations. We assume
that ij ’s are identically and independently distributed as N (0, 2 ) with E ( ij ) 0, Var ( ij ) 2 .
2
Note that the general linear model considered is
E (Y ) X
E (Yij ) i .
When all the entries in X are 0’s or 1’s, then this model can also be re-expressed in the form of
E (Yij ) i .
E (Y ) X * *
Cov (Y ) 2 I
1
1 X
X*
1
is a n ( p 1) matrix, and X denotes the earlier defined design matrix in which
- first n1 rows as (1,0,0,…,0),
- second n2 rows as (0,1,0,…,0)
- …, and
- last np rows as (0,0,0,…,1).
Analysis of Variance | Chapter 3 | Experimental Design Models | Shalabh, IIT Kanpur
3
3
We earlier assumed that rank X p but can we also say that rank X * is also p in the present
case?
Since the first column of X* is the vector sum of all its remaining p columns, so
rank X * p .
It is thus apparent that all the linear parametric functions of 1 , 2 ,..., p are not estimable. The
question now arises that what kind of linear parametric functions are estimable?
with
ni
Ci aij
j 1
Now
p ni
E ( L) aij E (Yij )
i 1 j 1
p ni
aij ( i )
i 1 j 1
p ni p ni
aij a ij i
i 1 j 1 i 1 j 1
p p
( Ci ) Ci i .
i 1 i 1
p
Thus C
i 1
i i is estimable if and only if
C
i 1
i 0,
p
i.e., C
i 1
i i
is a contrast.
estimable.
4
This effect and outcome can also be seen from the following explanation based on the estimation of
parameters , 1 , 2 ,..., p .
S p ni
(a) 0 ( yij i ) 0
i 1 j 1
S ni
(b) 0 ( yij i ) 0, i 1, 2,..., p.
i j 1
Note that (a) can be obtained from (b) or vice versa. So (a) and (b) are linearly dependent in the
sense that there are (p + 1) unknowns and p linearly independent equations. Consequently
ˆ , ˆ1 ,..., ˆ p do not have a unique solution. Same applies to the maximum likelihood estimation of
, 1 ,... p . .
niˆi 0 or
i 1
n i 1
i i 0
yio yoo
p
where n ni .
i 1
p p
In case, all the sample sizes are same, then the condition niˆi 0
i 1
or n
i 1
i i 0 reduces to
p p
ˆi 0 or
i 1
i 1
i 0.
5
So the model yij i ij needs to be rewritten so that all the parameters can be uniquely
estimated. Thus
Yij i ij
( ) ( i ) ij
* i* ij
where
*
i* i
1 p
i
p i 1
and
p
i 1
*
i 0
Thus in a linear model when X is not of full rank, then the parameters do not have unique estimates.
p p
In such conditions, a restriction
i 1
i 0 (or equivalently n
i 1
i i 0 in case all ni’s are not
same) can be added and then the least squares (or maximum likelihood) estimators obtained are
unique.
The model
p
E (Yij ) * i* ;
i 1
*
1 0
Let us now consider the analysis of variance with additional constraint. Let
Yij i ij , i 1, 2,..., p; j 1, 2,..., ni
( i ) ij
i ij
with
1 p
p
,
i 1
i i i ,
p p
ni i 0, n ni .
i 1 i 1
and ij ’s are identically and independently distributed with mean 0 and variance 2 .
Analysis of Variance | Chapter 3 | Experimental Design Models | Shalabh, IIT Kanpur
6
6
The null hypothesis is
H 0 : 1 2 ... p 0
This model is a one-way layout in the sense that the observations yij ' s are assumed to be affected
by only one treatment effect i . So the null hypothesis is equivalent to testing the equality of p
population means or equivalently the equality of p treatment effects.
E p ni
0 2 (y
i 1 j 1
ij i ) 0
or
p p ni
n ni i yij (1)
i 1 i 1 j 1
E ni
0 2 ( yij i ) 0
i j 1
or
ni
ni ni i yij (i 1, 2,..., p). (2)
j 1
p
Using n
i 1
i i 0 in (1) gives
1 p ni G
ˆ
n i 1 j 1
yij yoo
n
p ni
where G yij is the grand total of all the observations.
i 1 j 1
7
Substituting ̂ in (2) gives
1 ni
ˆ i
ni
y
j 1
ij ˆ
Ti
ˆ
ni
yio yoo
ni
where Ti yij is the treatment total due to ith effect i , i.e., total of all the observations receiving
j 1
1 ni
the ith treatment and yio yij .
ni j 1
Now the fitted model is yij ˆ ˆ i and the error sum of squares after substituting ̂ and ˆ i in E
becomes
p ni
E ( yij ˆ ˆi ) 2
i 1 j 1
p ni
( yij yoo ) ( yio yoo )
2
i 1 j 1
p ni p ni
( yij yoo ) ( yio yoo ) 2 2
i 1 j 1 i 1 j 1
p ni G 2 p Ti 2 G 2
yij2
i 1 j 1 n i 1 ni n
G2
and is called as correction factor CF .
n
8
Minimizing the error sum of squares
p ni
E1 ( yij )2
i 1 j 1
E1 p ni
0 2 (y ) 0
ij
i 1 j 1
or
G
ˆ y oo .
n
Note that
E1: Contains variation due to treatment and error both
E: Contains variation due to error only
So E1 E : contain variation due to treatment only.
9
These sum of squares forms the basis for the development of tools in the analysis of variance. We
can write
TSS SSTr SSE.
( y
i 1 j 1
ij yoo ) 0 so TSS carries (n 1) degrees of freedom.
The sum of squares due to the treatments is based on p quantities subject to the constraint
p
n (y
i 1
i io yoo ) 0 so SSTr has ( p 1) degrees of freedom.
( y
j 1
ij yio ) 0, i 1, 2,..., p
the TSS has been divided into two orthogonal components - SSTr and SSE. Moreover, all TSS, SSTr
and SSE can be expressed in a quadratic form. Since ij are assumed to be identically and
N ( i , 2 ).
Now using the theorems 7 and 8 with q1 SSTr, q2 SSE , we have under H 0 ,
SSTr
~ 2 ( p 1)
2
and
SSE
~ 2 ( n p ).
2
10
The mean squares is defined as the sum of squares divided by the degrees of freedom. So the mean
square due to treatment is
SSTr
MSTr
p 1
and the mean square due to error is
SSE
MSE .
n p
Thus, under H 0 ,
MSTr
2
F ~ F ( p 1, n p).
MSE
2
The decision rule is that reject H 0 if
F F1 , p 1, n p
at % level of significance.
MSTr
~ noncentral F ( p 1, n p, )
MSE
p
ni i2
where is the noncentrality parameter.
i 1 2
MSTr
Note that the test statistic can also be obtained from the likelihood ratio test.
MSE
If H 0 is rejected, then we go for multiple comparison tests and try to divide the population into
several groups having the same effects.
11
The analysis of variance table is as follows:
MSTr
Treatment p 1 SSTr MSTr
MSE
Total n 1 TSS
p
E ( SSTr ) E ni ( yio yoo ) 2
i 1
p
2
E ni ( i io ) ( oo )
i 1
where
1 ni
1 p ni p
ni i
io
ni
j 1
ij , oo ij and
n i 1 j 1
i 1 n
0.
p 2
E ( SSTr ) E ni i ( io oo )
i 1
p p
ni E (i2 )
i 1
n E (
i 1
i io oo ) 2 0.
Since
p p
1 1
E ( SSTr ) ni i2 2 ni
i 1 i 1 ni n
p
ni i2 ( p 1) 2
i 1
p
SSTr ni i2
2 i 1
or E
p 1 p 1
p
n i i
2
or E ( MSTr ) 2 i 1
.
p 1
12
1 ni
1 2
E ( io2 ) Var ( io ) Var ij ni2
ni 2
ni j 1 ni
1 p ni
1 2
E ( ) Var ( oo ) Var ij 2 n
2
oo
2
n i 1 j 1 n n
E ( io oo ) Cov( io , oo )
1 ni p ni
Cov ij ij
ni n j 1 i 1 j 1
ni 2 2
.
ni n n
Next
p ni
E ( SSE ) E ( yij yio ) 2
i 1 j 1
p ni 2
E ( i ij ) ( i io )
i 1 j 1
p ni
E ( ij io ) 2
i 1 j 1
p ni
E ( ij2 io2 2 ij io )
i 1 j 1
p ni
2 2 2 2
=
i 1 j 1 ni ni
p
ni 1
ni
2
i 1 j 1 ni
p
n (n 1)
2 i i
i 1 ni
p
2 (ni 1)
i 1
(n p) 2
SSE
2
or E
n p
or E ( MSE ) 2 .
13
Two way classification under fixed effects model
Suppose the response of an outcome is affected by the two factors – A and B. For example, suppose I
varities of mangoes are grown on I different plots of same size in each of the J different locations.
All plots are given same treatment like equal amont of water, equal amount of fertilizer etc. So there
are two factors in the experiment which affect the yield of mangoes.
- Location (A)
- Variety of mangoes (B)
Such an experiment is called two – factor experiment. The different locations correspond to
the different levels of A and the different varities correspond to the different levels of factor B. The
observations are collected on the basis of per plot.
The combined effect of the two factors (A and B in our case) is called the interaction effect (of A
and B).
Mathematically, let a and b be the levels of factors A and B respectively then a function f ( a , b ) is
called a function of no interaction if and only if there exists functions g ( a ) and h (b ) such that
f ( a , b ) g ( a ) h (b ) .
14
We consider here two cases
1. One observation per plot in which the interaction effect is zero.
2. More than one observations per plot in which the interaction effect is present.
factor, say B. So assume Yij are independently distributed as N (ij , 2 ) i 1, 2,..., I , j 1, 2,..., J .
where
oo
i io oo
i oj oo
ij ij io oj oo
with
I I
i (io oo ) 0
i 1 i 1
J J
(
j 1
j
j 1
oj oo ) 0
Here
i : effect of ith level of factor A
or excess of mea nof ith level of A over the general mean.
j : effect of jth level of B
We also assume that the model E (Yij ) ij is a full rank model so that ij and all linear parametric
15
The total number of observations are I J which can be arranged in a two way calssified I J table
where the rows corresponds to the different levels of A and the column corresponds to the different
levels of B.
The observations on Y and the design matrix X in this case are
Y 1 2 I 1 2 j
y11 1 1 0 0 1 0 0
y12 1 1 0 0 0 1 0
y1J 1 1 0 0 0 0 1
yI 1 1 0 0 1 1 0 0
yI 2 1 0 0 1 0 1 0
yIJ 1 0 0 1 0 0 1
If the design matrix is not of full rank, then the model can be reparameterized. In such a case, we can
start the analysis by assuming that the model E (Yij ) i j is obtained after
reparameterization.
against
H1 : at least one i (i 1, 2,..., I ) is different from others
Now we derive the least squares estimators (or equivalently the maximum likelihood estimator) of
, i and j , i 1, 2,..., I , j 1, 2,..., J by minimizing the error sum of squares
I J
E (y ij i j )2 .
i 1 j 1
16
The normal equations are obtained as
E I J
0 2 (y
i 1 j 1
ij i j ) 0
E J
i
0 2 (y
j 1
ij i j ) 0 , i 1, 2,..., I ,
E I
j
0 2 (y
i 1
ij i j ) 0 , j 1, 2,..., J .
I J
Solving the normal equations and using
i 1
i 0 and
j 1
j 0 , the least squares estimator are
obtaianed as
1 I J
G
ˆ
IJ
yi 1 j 1
ij
IJ
yoo
1 J
T
ˆi
J j 1
yij yoo i yoo = yio yoo i 1, 2,..., I
J
1 I Bj
ˆ j
I i 1
yij yoo
I
yoo yoj yoo , j 1, 2,..., J
where
Ti : treatment totals due to ith effect, i.e., sum of all the observations receiving the ith treatment
effect.
B j : block totals due to jth effect, i.e., sum of all the observations in the jth block.
( y
2
= ij yoo ) ( yio yoo ) ( yoj yoo )
i 1 j 1
I J
( yij yio yoj yoo ) 2
i 1 j 1
I J I J
= ( yij yoo ) 2 J ( yio yoo ) 2 I ( yoj yoo ) 2
i 1 j 1 i 1 j 1
which carries
IJ ( I 1) ( J 1) 1 ( I 1)( J 1)
degrees of freedom.
Analysis of Variance | Chapter 3 | Experimental Design Models | Shalabh, IIT Kanpur
17
17
Next we consider the estimation of and j under the null hypothesis H 0 : 1 2 ... I 0
Sum of squares due to factor A Error sum of squares
Thus the sum of squares due to deviation from H 0 (or sum of squares due to rows or sum of squares
are to factor A)
I I
SSA J ( yio yoo ) 2 J yio2 IJyoo2
i 1 i 1
and carries
18
The normal equations are
E2 E2
0 and 0, i 1, 2,..., I
i
which on solving give the estimators as
ˆ yoo
ˆi yio yoo .
The minimum value of the error sum of squares is obtained by
I J
Min E2 ( yij ˆ ˆ i ) 2
, j
i 1 j 1
I J
( yij yio ) 2
i 1 j 1
J I J
I ( yoj yoo ) 2 ( yij yio yoj yoo ) 2
j 1 i 1 j 1
Sum of squares due to factor B Error sum of squares
The sum of squares due to deviation from H 0 (or the sum of squares due to columns or sum of
i 1 j 1
I J I J
J ( yio yoo ) 2 I ( yoj yoo ) 2 ( yij yio yoj yoo ) 2
i 1 j 1 i 1 j 1
19
Note that SSA, SSB and SSE are mutually orthogonal and that is why the degrees of freedom can be
divided like this.
Now using the theory explained while discussing the likelihood ratio test or assuming yij ' s to be
SSE
~ 2 (( I 1)( J 1)).
2
SSA / 2
F1 I 1
SSE / 2
( I 1)( J 1)
( I 1)( J 1) SSA
.
( I 1) SSE
MSA
~ F (( I 1), ( I 1) ( J 1)) under H 0
MSE
where
SSA
MSA
I 1
SSE
MSE .
( I 1)( J 1)
Same statistic is also obtained using the likelihood ratio test for H 0 .
20
Similarly, the test statistic for H 0 is obtained as
SSB / 2
J 1
F2
SSE / 2
( I 1)( J 1)
( I 1)( J 1) SSB
( J 1) SSE
MSB
~ F (( J 1), ( I 1)( J 1)) under H 0
MSE
SSB
where MSB .
J 1
The decision rule is
Reject H 0 if F2 F1 (( J 1), ( I 1)( J 1)) .
The same test statistic can also be obtained from the likelihood ratio test.
MSA
Factor A (or rows) ( I 1) SSA MSA F1
MSE
MSB
Factor B (or column) ( J 1) SSB MSB F2
MSE
Error ( I 1)( J 1) SSE MSE
(by subtraction)
Total IJ 1 TSS
It can be found on similar lines as in the case of one way classification that
J I 2
E ( MSA) 2 i
I 1 i 1
I J 2
E ( MSB) 2 j
J 1 j 1
E ( MSE ) 2 .
21
If the null hypothesis is rejected, then we use the multiple comparison tests to divide
the i ' s (or j ' s) into groups such that i ' s (or j ' s) belonging to the same group are equal and
those belonging to different groups are different. Generally, in practice, the interest of experimenter
is more in using the multiple comparison test for treatment effects rather on the block effects. So the
multiple comparison test are used generally for the treatment effects only.
yijk are independently drawn from N (ij , ) so that the linear model under consideration is
2
yijk ij ijk
where ijk are identically and independently distributed following N (0, 2 ). Thus
E ( yijk ) ij
oo ( io oo ) ( oj oo ) ( ij io oj oo )
i j ij
where
oo
i io oo
j oj oo
ij ij io oj oo
with
I J I J
i 1
i 0, j 0, ij 0, ij 0.
j 1 i 1 j 1
Assume that the design matrix X is of full rank so that all the parametric functions of ij are
estimable.
22
The corresponding alternative hypothesis is
H1 : At least one i j , for i j
H1 : At least one i j , for i j
H1 : At least one ij ik , for j k .
1 I
ˆi yioo yooo
JK
y
i 1
ijk yooo
1 J
ˆ j yojo yooo
IK
y
j 1
ijk yooo
I J K
( yijk ˆ ˆ i ˆ j ˆij ) 2
i 1 j 1 k 1
I J K
( yijk yijo ) 2
i 1 j 1 k 1
SSE
with ~ 2 ( IJ ( K 1)).
2
23
Now minimizing the error sum of squares under H 0 1 2 ... I 0 , i.e., minimizing
I J K
E1 (y ijk j ij ) 2
i 1 j 1 k 1
I J K
( yijk ˆ ˆ j ˆij ) 2
i 1 j 1 k 1
I J K I
( yijk yijo ) 2 JK ( yioo yooo ) 2
i 1 j 1 k 1 i 1
I
SSE JK (y
i 1
ioo yooo ) 2 .
Thus the sum of squares due to deviation from H 0 or the sum of squares due to effect A is
I
SSA Sum of squares due to H 0 SSE JK ( yioo yooo ) 2
i 1
SSA
with ~ 2 ( I 1).
2
24
ˆ yooo
ˆ i yiooo yooo
ˆij yijo yioo yojo yooo .
The minimum error sum of squares is
I J K
( y
i 1 j 1 k 1
ijk ˆ ˆ i ˆij ) 2
J
SSE IK ( yojo yooo ) 2
j 1
and the sum of squares due to deviation from H o or the sum of squares due to effect B is
J
SSB Sum of squares due to H 0 SSE IK ( yojo yooo ) 2
j 1
SSB
with ~ 2 ( J 1).
2
Next, minimizing the error sum of squares under H 0 : all ij 0 for all i, j, i.e., minimizing
I J K
E3 (y ijk i j )2
i 1 j 1 k 1
I J K
( yijk ˆ ˆ i ˆ j ) 2
i 1 j 1 k 1
I J
SSE K ( yijo yioo yojo yooo ) 2 .
i 1 j 1
25
Thus the sum of squares due to deviation from H 0 or the sum of squares due to interaction effect
AB is
I J
SSAB Sum of squares due to H 0 SSE K (y
i 1 j 1
ijo yioo yojo yooo ) 2
SSAB
with ~ 2 (( I 1) J 1)).
2
SSA, SSB , SSAB and SSE as well as their respective distributions or using the likelihood ratio
2
test approach , the decision rules for the null hypothesis at level of significance are based on F-
statistic as follows
IJ ( K 1) SSA
F1 . ~ F ( I 1, IJ ( K 1) under H 0 ,
I 1 SSE
IJ ( K 1) SSB
F2 . ~ F ( J 1, IJ ( K 1) under H 0 ,
J 1 SSE
and
IJ ( K 1) SSAB
F3 . ~ F ( I 1)( J 1), IJ ( K 1) under H 0 .
( I 1)( J 1) SSE
So
Reject H 0 if F1 F1 ( I 1), IJ ( K 1)
Reject H 0 if F2 F1 ( J 1), IJ ( K 1)
Reject H 0 if F3 F1 ( I 1)( J 1), IJ ( K 1) .
If H 0 or H 0 is rejected, one can use t -test or multiple comparison test to find which pairs of
26
It can also be shown that
JK I 2
E ( SSA) 2 i
I 1 i 1
IK J 2
E ( SSB ) 2 j
J 1 j 1
I J
K
E ( SSAB ) 2
( I 1)( J 1) i 1 j 1
ij2
E ( SSE ) 2 .
SSA MSA
Factor A ( I 1) SSA MSA F1
I 1 MSE
SSB MSB
Factor B ( J 1) SSB MSB F2
J 1 MSE
SSAB MSAB
Interaction AB ( I 1)( J 1) SSAB MSAB F3
( I 1)( J 1) MSE
SSE
Error IJ ( K 1) SSE MSE
IJ ( K 1)
27
Tukey’s test for nonadditivity:
Consider the set up of two way classification with one observation per cell and interaction as
I J
yij i j ij ij , i 1, 2..., I , j 1, 2,..., J with i 0,
i 1
j 1
j 0.
Error 0
_________________________________
Total IJ 1
____________________________________
There is no degree of freedom for error. The problem is that the two factor interaction effect and
random error component are subsumed together and cannot be separated out. There is no estimate
for 2 .
If no interaction exists, then H 0 : ij 0 for all i ,j is accepted and the additive model
yij i j ij
is well enough to test the hypothesis H 0 : i 0 and H 0 : j 0 with error having ( I 1)( J 1)
degrees of freedom.
If interaction exists, then H 0 : ij 0 is rejected. In such a case, if we assume that the structure of
interaction effect is such that it is proportional to the product of individual effects, i.e.,
ij i j
then a test for testing H 0 : 0 can be constructed. Such a test will serve as a test for nonadditivity.
It will help in knowing the effect of presence of interact effect and whether the interaction enters
into the model additively. Such a test is given by Tukey’s test for nonadditivity which requires one
degree of freedom leaving I -1 J -1 -1 degrees of freedom for error.
28
Let us assume that departure from additivity can be specified by introducing a product term and
writing the model as
I J
E ( yij ) i j i j ; i 1, 2,..., I , j 1, 2,..., J with
i 1
i 0,
j 1
j 0.
When 0, the model becomes nonlinear model and the least squares theory for linear models is
not applicable.
I J
Note that using i 0,
i 1
j 1
j 0 , we have
1 I J
1 I J
yoo
IJ
yij
i 1 j 1 IJ
i 1 j 1
i j i j ij
1 I 1 J I J
i J
I i 1
j 1
j ( j ) oo
)(
IJ i 1 j j 1
oo
E ( yoo )
ˆ yoo .
Next
1 J 1 J
yio
J j 1
yij i j i j ij
J j 1
1 J 1 J
i
J j 1
j i j io
J j 1
i io
E ( yio ) i
ˆi yio ˆ yio yoo .
Similarly
yoj j
ˆ j yoj ˆ yoj yoo
Thus ˆ , ˆi and ̂ j remain the unbiased estimators of , i and j , respectively irrespective of
whether 0 or not.
29
Also
E yij yio yoj yoo i j
or
E ( yij yoo ) ( yio yoo ) ( yoj yoo ) i j .
S ( yij i j i j ) 2
i j
Sij2 .
i j
ˆ yoo
S J
0 (1 j ) Sij 0
i j 1
S I
0 (1 i ) Sij 0
j i 1
S I J
0 i j Sij 0
i 1 j 1
I J
or i j ( yij i j i j ) 0
i 1 j 1
I J
i j yij
(say)
i 1 j 1
or
2 2
I J
i j
i 1 j 1
which can be estimated provided i and j are assumed to be known.
Since i and j can be estimated by ̂ i yio yoo and ˆ j yoj yoo irrespective of whether 0
30
I J I J
2 ˆ2 2 ˆ2
I J I J
ˆ i j i j
J ˆ I
i 1 j 1 i 1 j 1
I J
IJ ( yio yoo )( yoj yoo ) yij
i 1 j 1
S ASB
I I
where S A J ˆ i2 J ( yio yoo ) 2
i 1 i 1
J J
S B I ˆ j2 I ( yoj yoo ) 2 .
j 1 j 1
When i and j are estimated by ̂ i and ˆ j , then substitute them back in the expression of
2
Var (ˆ )
I 2 J ˆ2
ˆ i j
i 1 j 1
IJ 2
S ASB
31
Note that if 0, then
I J
i j yij
E ˆ / ˆ i , ˆ j for all i, j E i 1I j 1 J
2
i j
2
i 1 j 1
I J
i j ( i j 0 ij )
E 1 1
i j
I J
( i )( j )
2 2
i 1 j 1
0
I J
0.
( i2 ) ( j2 )
i 1 j 1
Var ( )ˆ S ASB
2
I J
IJ ( yio yoo )( yoj yoo ) yij
i 1 j 1
2 S ASB
2
I J
IJ ( yio yoo )( yoj yoo )( yij yio yoj yoo )
i 1 j 1
2 S ASB
SN
2
follows a 2 - distribution with one degree of freedom where
2
I J
IJ ( yio yoo )( yoj yoo )( yij yio yoj yoo )
SN
i 1 j 1
S ASB
is the sum of squares due to nonadditivity .
32
Note that
I J
S AB
(y
i 1 j 1
ij yio yoj yoo ) 2
2
2
follows 2 (( I 1)( J 1)).
S SAB
so N2 2 is nonnegative and follows 2 ( I 1)( J 1) 1 .
Moreover SN (SS due to nonadditivity) and SSE are orthogonal. Thus the F-test for nonadditivity is
SN / 2
1
F
SSE / 2
( I 1)( J 1) 1
SSN
( I 1)( J 1) 1
SSE
~ F 1, ( I 1)( J 1) 1 under H 0 .
33
So the decision rule is
Reject H 0 : 0 whenever
The analysis of variance table for the model including a term for nonadditivity is as follows:
SA
A I 1 SA MS A
I 1
S
B J 1 SB MS B B
J 1
MS N
Nonadditivity 1 SN MS N S N
MSE
SSE
Error ( I 1)( J 1) 1 SSE MSE
( I 1)( J 1) 1
(By substraction)
Total IJ 1 TSS
Comparison of Variances
One of the basic assumptions in the analysis of variance is that the samples are drawn from different
normal populations with different means but same variances. So before going for analysis of
variance, the test of hypothesis about the equality of variance is needed to be done.
We discuss the test of equality of two variances and more than two variances.
34
Case 1: Equality of two variances
H 0 : 12 22 2 .
Suppose there are two independent random samples
A : x1 , x2 ,..., xn1 ; xi ~ N ( A , A2 )
B : y1 , y2 ,..., yn2 ; yi ~ N ( B , B2 )
(n2 1) s y2
~ 2 (n2 1).
2
Moreover, the sample variances sx2 and sy2 and are independent. So
(n 1) s 2
1 2 x
n1 1 s2
x2 ~ Fn1 1, n2 1.
(n2 1) s y s y
2
2
1
n 2
So for testing H 0 : 12 22 versus H1 : 12 22 , the null hypothesis H 0 is rejected if
F F
1 ; n1 1, n2 1
2
or
F F
1 ; n1 1, n2 1
2
where
1
F .
; n1 1, n2 1; F
2
1 ;n2 1, n1 1
2
If the null hypothesis H 0 : 12 22 is rejected, then the problem is termed as the Fister-Behren’s
problem. The solutions are available for this problem.
Analysis of Variance | Chapter 3 | Experimental Design Models | Shalabh, IIT Kanpur
35
35
Case 2: Equality of more than two variances: Bartlett’s test
H 0 : 12 22 ... k2 and H1 : i2 2j for atleast one i j 1, 2,..., k.
Let there be k independent normal population N (i , i2 ) each of size ni , i 1, 2,..., k . Let
s12 , s22 ,..., sk2 be k independent unbiased estimators of population variances 12 , 22 ,..., k2 respectively
with 1 , 2 ,..., k degrees of freedom. Under H 0 , all the variances are same as 2 , say and an
unbiased estimate of 2 is
k
i si2 k
s2 where i ni 1, i .
i 1 i 1
36