Comparaison empirique des estimateurs des taux d'erreur en analyse discriminante

Monte Carlo experiments are performed to compare twenty estimators or error rates in discriminant analysis in the case of two populations and linear classification rule. Several distributions (normal, chi-square and beta) have been considered. Estimators eOS (for actual and expected error rate) and eB (for optimal error rate) are the best, except for distributions very different from normal distributions. The e632 estimator is the best non parametric estimator. This estimator is better than parmetric estimators for distributions very different from normal distributions.