New Test Statistics for MANOVA/Descriptive Discriminant Analysis

Abstract

The univariate Brown-Forsythe test of 1974 and the multivariate two-sample Nel-van der Merwe test of 1986 are generalized to form five new multivariate alternatives to one-way multivariate analysis of variance (MANOVA) for use when dispersion matrices are heteroscedastic. Type I error rates for the Johansen test of 1980 and four of the five new test statistics were estimated using simulated data. Only conditions in which dispersion matrices were heterogeneous and sample sizes were unequal were included.In controlling Type I error rates, the four new multivariate generalizations gready outperformed the Johansen test.