Small sample longitudinal clinical trials with missing data: A comparison of analytic methods

The statistical power of 6 analytic methods was compared: t test and Mann-Whitney-Wilcoxon test on mean least squares regression slopes estimated for each participant, t test and Mann-Whitney-Wilcoxon test on pre-post differences, least squares multivariate repeated measures analysis of variance (MANOVA), and unbalanced repeated measures (URM) model with a 1st-order autoregressive covariance structure using restricted maximum-likelihood estimation. Factors included effect size, attrition pattern, sample size, and visit-to-visit correlation levels under sample size and dropout rates common to drug abuse treatment trials. In addition to its known effects of increasing N and effect size, the classic MANOVA approach was not as powerful as the other methods. Both t tests on pre-post change and the URM model performed well, and the results of using individual slope estimates as a summary statistic under an intent-to-treat model were surprisingly poor.