466 Chapter 10:Analysis of Variance
Suppose now that we want to test the hypothesis that there are no row and column
interactions — that is, we want to test
H 0 int:γij=0, i=1,...,m, j=1,...,n
Now, ifH 0 intis true, then the random variablesXij. will be normal with mean
E[Xij.]=μ+αi+βj
Also, since each of these terms is the average oflnormal random variables having variance
σ^2 , it follows that
Var(Xij.)=σ^2 /l
Hence, under the assumption of no interactions,
∑n
j= 1
∑m
i= 1
l(Xij.−μ−αi−βj)^2
σ^2
is a chi-square random variable withnmdegrees of freedom. Since a total of 1+m−
1 +n− 1 =n+m−1 of the parametersμ,αi,i=1,...,m,βj,j=1,...,n, must be
estimated, it follows that if we let
SSint=
∑n
j= 1
∑m
i= 1
l(Xij.−ˆμ−ˆαi−βˆj)^2 =
∑n
j= 1
∑m
i= 1
l(Xij.−Xi..−X.j.+X...)^2
then, underH 0 int,
SSint
σ^2
is chi-square with (n−1)(m−1) degrees of freedom
Therefore, under the assumption of no interactions,
SSint
(n−1)(m−1)
is an unbiased estimator ofσ^2
Because it can be shown that, under the assumption of no interactions,SSeandSSintare
independent, it follows that whenH 0 intis true
Fint=
SSint/(n−1)(m−1)
SSe/nm(l−1)
is anF-random variable with (n−1)(m−1) numerator andnm(l−1) denominator
degrees of freedom. This gives rise to the following significance levelαtest of
H 0 int:allγij= 0