Controlled multiple imputation methods for sensitivity analyses in longitudinal clinical trials with dropout and protocol deviation

Sensitivity analyses are commonly requested as part of the analysis of longitudinal clinical trials when data are missing. There are many ways in which such sensitivity analyses can be constructed. This article focuses on one particular approach, so-called controlled imputation. This combines two statistical ingredients, pattern-mixture models and multiple imputation. The aim is to assess sensitivity of the original conclusions to alternative assumptions about the statistical behavior of the patients’ outcomes following dropout and withdrawal. Such assumptions must reflect postulated treatment compliance when intention-to-treat-like inferences are required. Many such scenarios could be considered, depending on the clinical setting. The advantage of this approach is that it makes such assumptions explicit in the sensitivity analysis and hence readily accessible to the user.