Optimizing Efficiency and Convergence in MMRM: Practical Considerations for Longitudinal Data Analysis
arXiv:2607.26362
The paper studies how to improve convergence and efficiency of mixed models for repeated measures (MMRM) in longitudinal clinical trial data, recommending specific covariance structures and bias‑reduced adjustments for different sample sizes.
Abstract
Mixed models for repeated measures (MMRM) are a popular method for analyzing longitudinal data in clinical trials. However, practical challenges, such as small sample sizes, large numbers of time points, and selection of variance-covariance structure for within-subject errors, often present barriers to model convergence and valid inference. This article evaluates different options for using MMRM in different scenarios through extensive simulation studies and an application on diabetes trial data. We demonstrate that the empirical bias-reduced coefficient covariance adjustment with the heterogeneous autoregressive covariance structure yields near-nominal coverage with high convergence rates for moderate and large sample designs. For small sample sizes, the simple model with baseline covariates and treatment by time point interaction achieves good efficiency and high probability of convergence. Based on these results, we provide practitioners with actionable guidance for applying MMRM to clinical trial data.
10 pages