Heterogeneity of variance in clinical data.

No one robust method is ideal for all situations, but such methods are superior to the traditional tests and should be widely used to control rate of Type I error and maintain power.

Abstract

Traditional parametric (t, F) and nonparametric (Mann-Whitney-Wilcoxon U, Kruskal-Wallis H) statistics are sensitive to heterogeneity of variance (heteroscedasticity). Moreover, there are theoretical reasons to expect, and empirical results to document, the existence of heteroscedasticity in clinical data. Transformations to reduce heteroscedasticity are problematic. This article reviews the literature on robust methods that are available and that should be widely used to control rate of Type I error and maintain power. No one robust method is ideal for all situations, but such methods are superior to the traditional tests. Specific recommendations are made for application under various conditions of heteroscedasticity.