Error Rates Stability of The Homoscedastic Discriminant Function

Abstract

In this study the stability of the observed error rates of the homoscedastic discriminant function relative to the number of parameters in the model using simulated data from multivariate normal populations was investigated.   Three models were considered, the four, six and eight variables models, each having four values of the separator function (). Equal and unequal prior probabilities were considered for the different number of parameter and separator function configurations. The asymptotic performance of the models was considered using the cross validation error rate estimation procedure. Results indicate the six variable models as being more stable (displaying less variability in the estimated error rates) than the other models under consideration. Less deterioration was observed for the six-variable model specification as was evident in the other models and this was more pronounced for smaller values of.