Current omnibus procedures for the analysis of interaction effects in repeated measures designs which contain a grouping variable are known to be nonrobust to violations of multisample sphericity, particularly when group sizes are unequal. An alternative approach is to formulate a comprehensive set of contrasts on the data which probe the specific nature of the interaction. Six interaction contrast procedures are compared via Monte Carlo methods. Two test statistics are considered, one relying on an estimate of the standard error of the contrast formed by pooling across levels of the groups and trials factors, and the other employing a nonpooled estimate based on only that data used in defining a contrast. A Scheffe, Studentized maximum modulus, and Hochberg step-up Bonferroni critical value are paired with each statistic. Only the Studentized maximum modulus and Hochberg nonpooled procedures provided acceptable rates of familywise Type I error control under departures from multisample sphericity when the data was nonnormal. Six tables present data from the analyses. (Contains 23 references.) (Author)
Authors
- Peer Reviewed
- F
- Publication Type
- ['Reports - Evaluative', 'Speeches/Meeting Papers']
- Published in
- United States of America
