496 Chapter 11:Goodness of Fit Tests and Categorical Data Analysis
The different members of the population will be assumed to be independent. Also, let
pi=P{X=i}=
∑s
j= 1
Pij, i=1,...,r
and
qj=P{Y=j}=
∑r
i= 1
Pij, j=1,...,s
That is, pi is the probability that an arbitrary member of the population will have
X-characteristici, andqjis the probability it will haveY-characteristicj.
We are interested in testing the hypothesis that a population member’s X- and
Y-characteristics are independent. That is, we are interested in testing
H 0 :Pij=piqj, for all i=1,...,r
j=1,...,s
against the alternative
H 1 :Pij=piqj, for some i,ji=1,...,r
j=1,...,s
To test this hypothesis, suppose thatnmembers of the population have been sampled, with
the result thatNijof them have simultaneously hadX-characteristiciandY-characteristic
j,i=1,...,r,j=1,...,s.
Since the quantitiespi,i=1,...,r, andqj,j=1,...,sare not specified by the null
hypothesis, they must first be estimated. Now since
Ni=
∑s
j= 1
Nij, i=1,...,r
represents the number of the sampled population members that haveX-characteristici,
a natural (in fact, the maximum likelihood) estimator ofpiis
pˆi=
Ni
n
, i=1,...,r
Similarly, letting
Mj=
∑r
i= 1
Nij, j=1,...,s