Model averaging using likelihoods that reflect poor outcomes for clinical trial dropouts

Despite efforts undertaken to prevent missing data in clinical trials, it is still inevitable to have some missing data due to various reasons. For example, some patients drop out due to lack of efficacy or tolerability issues. Their ‘true’ but unobserved end-of-study outcomes are likely worse than the observed outcomes for completers because dropouts stop taking their assigned therapy. For a binary endpoint, a ‘dropout equals failure’ approach has been widely applied. However, there is no similar approach for a continuous endpoint. Commonly used mixed model repeated measures (MMRM) analyses or multiple imputation methods require a missing at random assumption which may not realistically reflect the poor response for dropouts. We propose a model averaging approach using likelihoods that assume all missing outcomes are worse than the observed outcomes for each treatment group. The estimated treatment difference is obtained as a weighted average of the estimates derived from likelihood functions with a single normal distribution and a mixture of two normal distributions. Simulations are used to compare our proposed method with a quantile regression approach using trimmed means and medians. Applications to clinical trial examples are presented for illustration.