Mixed models for repeated measures should include time-by-covariate\n interactions to assure power gains and robustness against dropout bias\n relative to complete-case ANCOVA

In randomized trials with continuous-valued outcomes the goal is often to\nestimate the difference in average outcomes between two treatment groups.\nHowever, the outcome in some trials is longitudinal, meaning that multiple\nmeasurements of the same outcome are taken over time for each subject. The\ntarget of inference in this case is often still the difference in averages at a\ngiven timepoint. One way to analyze these data is to ignore the measurements at\nintermediate timepoints and proceed with a standard covariate-adjusted analysis\n(e.g. ANCOVA) with the complete cases. However, it is generally thought that\nexploiting information from intermediate timepoints using mixed models for\nrepeated measures (MMRM) a) increases power and b) more naturally "handles"\nmissing data. Here we prove that neither of these conclusions is entirely\ncorrect when baseline covariates are adjusted for without including\ntime-by-covariate interactions. We back these claims up with simulations. MMRM\nprovides benefits over complete-cases ANCOVA in many cases, but covariate-time\ninteraction terms should always be included to guarantee the best results.\n