A simulation study of a nested sequence of binomial regression models

Abstract

The inference problem we consider is that of model choice from a nested sequence of binomial regression models. The approach we take is to test successively, from most general to most specific, the corresponding sequence of composite hypotheses. This approach is based on the very general class of divergence measures, the phi-divergence. An approximation to the power function of the new family of test statistics proposed is obtained. An extensive simulation study is carried out by obtaining new test statistics that are a good alternative to the traditional loglikelihood test statistic.