Discrete discriminant models: a performance simulation with reference to expert systems' applications

Abstract

Monte Carlo simulation experiments are used to assess performance of seven models for discrete discriminant analysis. Discriminant rules are obtained for binary observation vectors from training samples drawn from two-group populations. Performance is evaluated primarily in terms of the error rate observed in the training and test samples for populations characterized by small and unequal training sample sizes, log-likelihood reversals, various correlation structures, and missing values. Some models are superior for certain population structures, and an attempt is made to identify data set characteristics that favor a particular model. Applications of the models are discussed, including their use as components in expert systems, and an expert system for consulting on discriminant analysis problems is reviewed.