La. Goodman, ON QUASI-INDEPENDENCE AND QUASI-DEPENDENCE IN CONTINGENCY-TABLES, WITH SPECIAL REFERENCE TO ORDINAL TRIANGULAR CONTINGENCY-TABLES, Journal of the American Statistical Association, 89(427), 1994, pp. 1059-1063
In the usual statistical analysis of the entries (frequencies) in a co
ntingency table having I rows and J columns (I greater than or equal t
o 2, J greater than or equal to 2), an important role is often played
by the usual null hypothesis of independence (i.e., the hypothesis tha
t the row variable and the column variable are statistically independe
nt of each other). In the situation considered in this article, the en
tries in some of the I X J cells of the contingency table are omitted
from the analysis (because these entries are either missing, unreliabl
e, void, or restricted in certain ways); and in this situation, the us
ual null hypothesis of independence cannot be applied in a meaningful
way. The concept of ''quasi-independence'' was introduced for this sit
uation; and an important role can often be played in this situation by
the null hypothesis of quasi-independence (a generalization of the us
ual null hypothesis of independence). For this situation, we introduce
also the concept of likelihood-ratio ''quasi-dependence'' (a generali
zation of the usual concept of likelihood-ratio dependence in the I X
J table), and we consider the null hypothesis of null quasi-dependence
and the alternative hypotheses of positive quasi-dependence and negat
ive quasi-dependence. The relationship between quasi-independence and
null quasi-dependence is considered, and methods are introduced for te
sting these null hypotheses against the aforementioned alternative hyp
otheses. Special attention is given to the triangular contingency tabl
e; and some of the results presented for this special case can also be
applied more generally to the I x J contingency table. With the appro
ach presented in this article, we are able to simplify and gain furthe
r insight into some formulas appearing in the earlier literature on th
e quasi-independence model applied to contingency tables. This approac
h has certain other advantages as well. For example, with this approac
h we are able to introduce various tests of the quasi-independence mod
el against alternative hypotheses of positive or negative quasi-depend
ence, and these tests will in some cases be more powerful than other t
ests proposed in the earlier literature.