Collapsibility of likelihood ratio tests in multidimensional contingency tables
Linda June Davis · Communication in Statistics- Theory and Methods · 1990
A pair of nested hierarchical log linear models is defined to be collapsible onto a subset of factors if the likelihood ratio statistic computed using the original table of observed counts is identical to the likelihood ratio statistic computed on the corresponding marginal table. Conditions implying this notion of collapsibility are derived. Also, this notion of collapsibility is related to collapsibility as defined in Whittemore (1978) and Asmussen and Edwards (1983).