Comparison of Tests for Heteroscedasticity in Between-Subjects ANOVA Models

Abstract

Several tests for heteroscedasticity in a two-group between-subject variances were compared with a simulation study. Two common rank-based procedures inflated test size with skewed error distributions. Nonparametric Levene test performed well but has notable limitations. Tests based on the absolute value of OLS residuals also inflated test size with skewed error distributions. Procedures based on squared OLS residuals performed better; however, the original Breusch-Pagan and Variance Function Regression are sensitive to even slight departures from the normality assumption. The Brown-Forsythe test based on taking the absolute value of median centered data performed the best; however, generalization to more complex analyses would not be straightforward.