Covariance Matrix Selection for Mixed Models Repeated Measurements in Clinical Trials

Sammanfattning

Inference for Mixed Models Repeated Measurements (MMRM) is strongly dependenton the assumed within-subject covariance, particularly when data are missing.Data collected in clinical trials usually exhibit missing data (dropout), and guidelinesfor covariance matrix selection are required. Under specific conditions, themechanism generating the missing data can be ignored without affecting the inferencesof the end-of-trial treatment effect. These conditions can be difficult to testand sometimes misinterpreted in practice. We conducted a simulation study fittingdifferent models covering multiple sample sizes, dropout rates, and underlying covariancesand quantified how model covariance misspecification affects estimationof the end-of-trial treatment effect. Bias, Type I error, and power are evaluatedfor the MMRM fitted models assuming the missing data mechanism as ignorable aswell as for pattern-mixture models (PMM), bypassing the assumptions of MAR andignorability. Simulations show that when treating the mechanism as ignorable, covariancemisspecification can bias the estimated end-of-trial treatment effect. Whenthe generating covariance is unstructured and the sample size is large, structuredmodels tend to inflate the Type I error while maintaining similar or higher powerrelative to correctly specified unstructured models. Conversely, structured covariances(e.g., Toeplitz) can be preferred for small samples and high dropout, offeringbetter Type I error control and lower bias variance, with small impact on mean biasor power. Modeling dropout via PMM showed inconsistent behavior: bias dependedon the generating covariance, and, even when correctly specified, structured covariancesoften yielded inflated Type I error compared to when ignoring the mechanism.These inconsistencies might be due to not having selected an optimal PMM. Untilfurther investigation, we recommend using an unstructured covariance and treatingthe missingness mechanism as ignorable for large samples, and to consider theToeplitz for small samples with high dropout. These choices are safer and simplerthan adopting a PMM, which offers little added benefit in this setting.