Evaluation of overall treatment effect in MMRM

Abstract

In longitudinal clinical trials for drug development, the study objective is often to evaluate overall treatment effect across all visits. Despite careful planning and study conduct, the occurrence of incomplete data cannot be completely eliminated. As a direct likelihood method, the mixed-effects model for repeated measures (MMRM) has become one of the preferred approaches for handling missing data in such designs. MMRM is a full multivariate model in nature, which avoids potential bias as a predetermined model, and operates in a more general missing-at-random (MAR) framework. However, if treatment effect is constant over time, overparameterization of treatment by time interaction in MMRM could result in loss of power. In this article, we utilize MMRM estimates and propose an optimal weighting method for combining visit-specific estimates to maximize the power under MAR mechanism. For a special case where the underlying covariance is compound symmetry, we show that the optimal weighting method is asymptotically equal to MMRM. In other words, MMRM has optimal power under this special case. When the underlying covariance is of an unstructured pattern, the optimal weighting method has increased power under MAR and missing-not-at-random (MNAR) mechanisms, and can lead to bias reduction under MNAR. This is especially true when the variance is greater at later time point, which could lead to a smaller weight. We present practical examples using the optimal weighting method to analyze two cystic fibrosis clinical trial data sets.