Effects of kurtosis for the error rate estimators using resampling methods in two class discrimination

Abstract

In preceding studies, error rate estimators have been compared under various conditions and in most cases the population distribution was assumed to be normal. Effects of non-normality of the population have therefore not been studied sufficiently. In this study, we focused on kurtosis as a measure of non-normality and examined the effects of kurtosis for error rate estimators, especially resampling-based estimators. Our simulation results in two-class discrimination using a linear discriminant function suggest that it is necessary to consider non-normality of the population in comparison of estimators. © 2009 Springer Berlin Heidelberg.