Discriminant Analysis with Discrete and Continuous Variables

Abstract

A paradigmatic methodologic approach does not exist for the problem of discriminant analysis with both discrete and continuous explanatory variables. Procedures that have been employed in this situation include Fisher's linear discriminant function, the quadratic discriminant function, the linear discriminant furlction with higher-order terms, discriminant functions with logistic regression estimates of the coefficients, and the location model. Average optimal error rates for these procedures are reported for the case of one normal variable and two binary variables. These approaches are also compared for the prediction of two-year survival following recovery from myocardial infarction, employing resubstitution, jackknife and independent-sample estimates of the error rates. Recommendations for applications are presented.