A simulative comparison of linear, quadratic and kernel discrimination

Abstract

Three discriminant analysis models: linear discriminant analysis, quadratic discriminant analysis and a model using kernel estimation, are compared in a simulation study. Data are generated from multinormal, lognormal and mixtures of distributions. Some measures of performance that are based upon posterior probabilities are used for the comparison. The influence on performance of sample size, dimensionality and distance between populations is investigated. An algorithm to estimate the smoothness parameters for the kernel model is evaluated. Finally some remarks are made with respect to the so-called variable kernel model.