Covariance Matrix Selection for Mixed Models Repeated Measurements in Clinical Trials

Abstract

Inference for Mixed Models Repeated Measurements (MMRM) is strongly dependent on the assumed within-subject covariance, particularly when data are missing. Data collected in clinical trials usually exhibit missing data (dropout), and guidelines for covariance matrix selection are required. Under specific conditions, the mechanism generating the missing data can be ignored without affecting the inferences of the end-of-trial treatment effect. These conditions can be difficult to test and sometimes misinterpreted in practice. We conducted a simulation study fitting different models covering multiple sample sizes, dropout rates, and underlying covariances and quantified how model covariance misspecification affects estimation of the end-of-trial treatment effect. Bias, Type I error, and power are evaluated for the MMRM fitted models assuming the missing data mechanism as ignorable as well as for pattern-mixture models (PMM), bypassing the assumptions of MAR and ignorability. Simulations show that when treating the mechanism as ignorable, covariance misspecification can bias the estimated end-of-trial treatment effect. When the generating covariance is unstructured and the sample size is large, structured models tend to inflate the Type I error while maintaining similar or higher power relative to correctly specified unstructured models. Conversely, structured covariances (e.g., Toeplitz) can be preferred for small samples and high dropout, offering better Type I error control and lower bias variance, with small impact on mean bias or power. Modeling dropout via PMM showed inconsistent behavior: bias depended on the generating covariance, and, even when correctly specified, structured covariances often yielded inflated Type I error compared to when ignoring the mechanism. These inconsistencies might be due to not having selected an optimal PMM. Until further investigation, we recommend using an unstructured covariance and treating the missingness mechanism as ignorable for large samples, and to consider the Toeplitz for small samples with high dropout. These choices are safer and simpler than adopting a PMM, which offers little added benefit in this setting.