Handling baselines in repeated measures analyses with missing data at random

In longitudinal clinical studies, after randomization at baseline, subjects are followed for a period of time for development of symptoms. A mixed model for repeated measures (MMRM) can be used to analyze data from such studies. Fitzmaurice et al. (2004) outlined five approaches for handling baseline responses in an MMRM analysis. They are: (1) Retain the baselines as part of the outcome vector and make no assumptions about group differences in the mean response at baseline. (2) Retain the baselines as part of the outcome vector and assume the group means are equal at baseline. (3) Subtract the baselines from all of the remaining post-baseline responses, and analyze the differences from baseline. (4) Use the baselines as a covariate in the analysis of the post-baseline responses, assuming homogeneous regression slopes. (5) Use the baselines as a covariate in the analysis of the post-baseline responses, allowing different regression slopes. In this paper, we evaluate these five approaches in the presence of data missing at random. We evaluate the approaches based on the bias of the estimate and the coverage accuracy of the confidence interval. The results suggest that strategies 2 and 5 are recommended.