Power Differences Between the Modified Brown-Forsythe and Mixed-Model Approaches in Repeated Measures Designs

Abstract. This article compares the sensibility of the modified Brown-Forsythe (MBF) approach developed by Vallejo and Ato (2006) and a modified empirical generalized least squares (EGLS) method adjusted by the Kenward-Roger solution available in the SAS Institute's (2002) Proc Mixed program to detect the presence of an interaction effect under departures from covariance homogeneity and multivariate normality. Although none of the approaches demonstrated superior performance in all situations, our results indicate that the so-called EGLS method, based on the Akaike's Information Criterion or based on always assuming a unstructured between-subjects heterogeneous covariance pattern, was the most powerful alternative. Results also indicate that little power can be achieved with the EGLS method if the covariance matrix is specified correctly.