Analysis of Variance and Analysis of Variance and Analysis of Variance and Analysis of Variance and

Design Design of Experiments of Experiments--II

MODULE MODULE III III
gg pp
EXPERIMENTAL DESIGN EXPERIMENTAL DESIGN
LECTURE LECTURE - - 16 16
EXPERIMENTAL DESIGN EXPERIMENTAL DESIGN
MODELS MODELS
Dr. Shalabh
Department of Mathematics and Statistics
Indian Institute of Technology Kanpur
Two-way classification with interactions
2
Consider the two-way classification with an equal number, say K observations per cell. Let : k
th
observation in
(i, j)
th
cell , i.e., receiving the treatments i
th
level of factor A and j
th
level of factor B,
and are independentlydrawn from so that the linear model under consideration is
ijk
y
1,2,..., ; 1,2,..., ; 1,2,..., i I j J k K = = =
ijk
y
2
( , )
ij
N and are independently drawn from so that the linear model under consideration is
ijk
y ( , )
ij
N
ijk ij ijk
y = +
where are identically and independently distributed following Thus
ijk

2
(0, ). N
( )
( ) ( ) ( )
where
ijk ij
oo io oo oj oo ij io oj oo
i j ij
E y


=
= + + + +
= + + +
where
oo
i io oo
j oj oo



=
=
=
1 1 1 1
0, 0, 0, 0.
with
ij ij io oj oo
I J I J
i j ij ij
i j i j


= = = =
= +
= = = =

1 1 1 1 i j i j
Assume that the design matrix X is of full rank so that all the parametric functions of are estimable.
ij

: 0
The null hypothesis are
= H = = =
3
0 1 2
0 1 2
0
: ... 0
: ... 0
: 0 , . all All for
=
=
I
J
ij
H
H
H i j

= = =
= = =
=
1
1
: ,
: ,
The corresponding alternative hypothesis is
At least one for
At least one for


i j
i j
H i j
H i j

1
1
: , . At least one for
i j
ij ik
j
H j k

Minimizing the error sum of squares
I J K
2
1 1 1
( ) ,
I J K
ijk i j ij
i j k
E y
= = =
=

0
the normal equations are obtained as
E
= 0,
0 ,
0
for all
for all and


i
E
i
E
j

= 0
0 .
for all and
for all and


j
ij
j
E
i j

1
The least squares estimates are obtained as
I J K
4
1 1 1
1
1

I J K
ooo ijk
i j k
I
i ioo ooo ijk ooo
i
y y
IJK
y y y y
JK

= = =
=
= =
= =

1
1

J
j ojo ooo ijk ooo
j
ij ijo ioo ojo ooo
y y y y
IK
y y y y

=
= =
= +

1 1
1

I J
ijk ioo ojo ooo
i j
y y y y
K
= =
= +

The error sumof square is

2

, , ,
1 1 1
2
( )

( )

The error sum of square is
i j
I J K
ijk i j ij
ij i j k
I J K
SSE Min y

= = =
=

2
1 1 1
2
1 1 1
( )
( )
ijk i j ij
i j k
I J K
ijk ijo
i j k
y
y y

= = =
= = =
=
=

2
2
~ ( ( 1)). with
SSE
IJ K

Now minimizing the error sum of squares under , i.e., minimizing
0 1 2
... 0

= = = = =
I
H
5
2
1
1 1 1
( )
= = =
=

I J K
ijk j ij
i j k
E y
with respect to and and solving the normal equations ,
j

ij
1 1 1
0, 0 0 for all and for all and
j ij
E E E
j i j


= = =

gives the least squares estimates as gives the least squares estimates as

ooo
j ojo ooo
y
y y

=
=
.
ij ijo ioo ojo ooo
y y y y = +
The sum of squares due to is
0
H

2
, ,
1 1 1
( )
j ij
I J K
ijk j ij
i j k
Min y


= = =

2
1 1 1
2 2

( )
( ) ( )
I J K
ijk j ij
i j k
I J K I
ijk ijo ioo ooo
y
y y JK y y

= = =
=
= +

1 1 1 1
2
1
( ) .
ijk ijo ioo ooo
i j k i
I
ioo ooo
i
SSE JK y y
= = = =
=
= +

Thus the sum of squares due to deviation from or the sum of squares due to effect A is
0
H

6
2
0
1
2
2
( )
~ ( 1).
Sum of squares due to
with


I
ioo ooo
i
SSA H SSE JK y y
SSA
I

=
= =

Minimizing the error sum of squares under i.e., minimizing

0 1 2
: ... 0
J
H

= = = =
2
2
1 1 1
( )
I J K
ijk i ij
i j k
E y
= = =
=

and solving the normal equations
2 2 2
0, 0 0 for all and for all and
i ij
E E E
j i j


= = =

yields the least squares estimators

ooo
i iooo ooo
y
y y

=
=
.
ij ijo ioo ojo ooo
y y y y = +
The minimum error sum of squares is
I J K J
2 2
1 1 1 1
( ) ( )
ijk i ij ojo ooo
i j k j
y SSE IK y y
= = = =
= +

and the sum of squares due to deviation from or the sum of squares due to effect B is
0
H

7
0
2
1
2
( )
( 1)
Sum of squares due to
ith


J
ojo ooo
j
SSB H SSE
IK y y
SSB
J

=
=
=

Next, minimizing the error sum of squares under for all i, j, i.e., minimizing
2
2
~ ( 1). with J

0
: 0
ij
H all

=
2
3
1 1 1
( )
I J K
ijk i j
i j k
E y
= = =
=

with respect to and and solving the normal equations ,
j

yields the least squares estimators as

3 3 3
0, 0 0 for all and for all
i j
E E E
i j


= = =

.
ooo
i ioo ooo
j ojo ooo
y
y y
y y

=
=
=
j ojo ooo
y y
The sum of squares due to is
0
H

8
2
, ,
1 1 1
2
( )

( )
i j
I J K
ijk i j
i j k
I J K
ijk i j
Min y
y



= = =

=

1 1 1
2

1 1
( ) .
i j k
I J
ijo ioo ojo ooo
i j
SSE K y y y y
= = =
= =
= + +

Thus the sum of squares due to deviation from or the sum of squares due to interaction effect AB is
0
H

0
2
( )
Sum of squares due to
I J
SSAB H SSE
K

=

2
1 1
2
2
( )
~ (( 1) 1)). with

ijo ioo ojo ooo
i j
K y y y y
SSAB
I J

= =
= +


The total sum of squares can be partitioned as
TSS = SSA + SSB + SSAB + SSE
where SSA, SSB, SSAB and SSE are mutually orthogonal. So either using the independence of SSA, SSB, SSAB and
2
SSE as well as their respective distributions or using the likelihood ratio test approach, the decision rules for the
null hypothesis at level of significance are based on F - statistics as follows:
2

( 1) IJ K SSA
9
[ ]
[ ]
1 0
2 0
( 1)
. ~ ( 1, ( 1) ,
1
( 1)
. ~ ( 1, ( 1) ,
under
under

IJ K SSA
F F I IJ K H
I SSE
IJ K SSB
F F J IJ K H

=
[ ]
[ ]
2 0
3 0
( , ( ) ,
1
( 1)
. ~ ( 1)( 1), ( 1) .
( 1)( 1)
and
under
J SSE
IJ K SSAB
F F I J IJ K H
I J SSE

=

[ ]
0 1 1
0 2 1
( 1), ( 1)
So
Reject
Reject
if
if
H F F I IJ K
H F F

>
>
[ ]
( 1), ( 1) J IJ K

If or is rejected, one can use t -test or multiple comparisontest to find which pairs of are

[ ]
0 3 1
( 1)( 1), ( 1) . Reject if H F F I J IJ K

>
0
H
0
H

' '
i j
s s or
If or is rejected, one can use t test or multiple comparison test to find which pairs of are
significantly different.
If is rejected, one would not usually explore it further but theoretically t- test or multiple comparison tests can be
0
H
0
H

i j
s s o
0
H

used.
It can also be shown that
2 2
1
( )
1
I
i
i
J
JK
E SSA
I

=
= +

10
2 2
1
2 2
1 1
( )
1
( )
( 1)( 1)
J
j
j
I J
ij
i j
IK
E SSB
J
K
E SSAB
I J


=
= =
= +

= +

2
( ) . E SSE =
The analysis of variance table is as follows:
Source of Degrees of Sum of squares Mean squares F - value Source of
variation
Degrees of
freedom
Sum of squares Mean squares F value
Factor A ( I - 1) SSA
1
MSA
F
MSE
=
1
MSA
MSA
I
=

1
( , ) C F p n p

=
Factor B (J - 1) SSB
2
MSB
F
MSE
=
1
MSB
MSB
J
=

0
0
: H =
Interaction AB (I - 1) (J - 1) SSAB
3
MSAB
F
MSE
=
( )( 1)
SSAB
MSAB
I I J
=

SSE
Error I J (K - 1) SSE
Total (I J K 1) TSS
( 1)
SSE
MSE
IJ K
=