Implications of chance baseline differences in repeated measurement designs

Datasets representing randomized, parallel-groups designs were analyzed by repeated measurements ANOVA and linear trend analysis with and without baseline values being covaried. ANOVA tests for the between-groups main effect, groups X times interaction, and differences in linear trends across time periods are shown to be seriously conservative or seriously nonconservative, depending on the direction and significance of chance baseline mean difference. Inclusion of baseline scores as a covariate in the repeated measurements ANOVA provides appropriate correction for the between-groups (average) effect across time, but the covariate provides no correction for the within-subject effects that are concerned with differences in the rates or patterns of change across time. If one desires to evaluate differences between patterns of treatment-induced change, tests of significance for differences in group means on composite trend scores with covariance correction for baseline are recommended. If covariance correction is not or cannot be employed, the potentially "favorable" or "unfavorable" influence of chance baseline differences on tests of significance needs to be explicitly recognized.