A Likelihood-Based Approach for the Analysis of Longitudinal Clinical Trials with Return-to-Baseline Imputation

Discontinuation of assigned therapy in longitudinal clinical trials is often inevitable due to various reasons such as intolerability or lack of efficacy. When the primary outcome of interest is the mean difference between treatment groups at the end of the trial, how to deal with the missing data due to discontinuation of assigned therapy is critical. The draft ICH E9 (R1) addendum proposes several strategies for handling intercurrent events, such as discontinuation of assigned therapy, under the estimand framework. The “hypothetical strategy”, in which the outcomes after discontinuation are envisioned under the hypothetical condition that patients who discontinued assigned therapy had actually stayed on assigned therapy, is commonly employed but requires untestable assumptions about the distribution of the post-discontinuation data. Return-to-baseline (RTB) is an assumption recently suggested by at least one regulatory agency. RTB assumes that any treatment effects observed prior to discontinuation are washed out, such that the mean effect at the end of the study among discontinued patients is the same as that at baseline. Multiple imputation (MI) may be used to implement this method but may overestimate the variance. In this paper, we propose a likelihood-based method to get the point estimate and variance for the treatment difference directly from a mixed-model for repeated measures (MMRM) analysis. Simulations are conducted to evaluate its performance as compared to other approaches including MI and MI with bootstrap. Two clinical trials are used to demonstrate the application.