Comparative performance of four discriminant analysis methods for mixtures of continuous and discrete variables

The present study investigates the performance of four classification rules with respect to discriminatory ability for data consisting of a mixture of continuous and discrete variables. The four discriminant analysis methods are Fisher's linear discrimination, logistic discrimination, quadratic discrimination and a kernel model. Four measures of performance for evaluation of the classification rules are used: the error rate, the quadratic scoring rule, the modified logarithmic scoring rule and a doubt-based scoring rule. The mixed data are obtained by generating from the fourdimensional normal distribution. Three of these variables were discretized. The results show that Fisher's linear discrimination and logistic discrimination have an alomost similar performance. In most of the situations model seems to be appropriate as far as discriminatory ability is concerned.