Categorical variable selection based on entropy reduction

This paper presents the derivation of a categorical variable selection technique which utilizes the entropy function as a measure of variability for nominally scaled variables. The selection criterion uses likelihood ratio statistics which, for the hypotheses under consideration, are identical to minimum discrimination information statistics. Thus, the paper provides an alternative motivation for a selection technique based on discriminatory power, and it provides an extension of that technique to the multipopulation discrimination problem. The selection technique is illustrated for a study in which we discriminate among three populations: cervical cancer patients, population-based controls, and hospital-based controls.