This study compared three approaches to the two-factor experiment with repeated measures on one factor: (1) the conventional mixed model analysis of variance, (2) the Greenhouse-Geisser conservative analysis of variance, and (3) multivariate extensions of analysis of variance. Computer simulated data were used in a total of 96 sets of covariance structures and sample size ratio to determine type I error rates when the usual covariance matrix restrictions of compound symmetry and equality across levels of the nonrepeated factor are violated jointly and separately. Practical implications of the results are discussed. (Author)