A distance based approach to discriminant analysis and its properties

Abstract

A general distance based method for allocating an observation to one of several known populations, on the ha.sis of both continuous and discrete explanatory variables, is proposed and studied. This method depends on a given statistical distance between observations, and leads to sorne classic discriminant rules by taking suitable distances. The error rates can be easily computed and, unlike other distance classification rules, the prior probabilities can be taken into account.