Discrimination Relative to Measures of Non-Normality

The robustness of discriminant functions to nonnormality is investigated. The performance of procedures relative to measures of the difference between the actual distribution of the observations and the usual assumption of normal densities is assessed. For example, the two population, mixed distributions problem with equal costs of misclassification are considered. The parameters are estimated by maximum likelihood and recently proposed robust methods.