Missing data is a common problem in longitudinal clinical trials, potentially causing the loss of statistical efficiency and biased inferences. It complicates the statistical analysis of clinical trials further as the missingness mechanism is unknown and unverifiable in most situations. Usually, a primary analysis is based on a missing at random (MAR) assumption, which may not be conservative when informative censoring occurs. To quantify the robustness of inferences on the primary study objective to departure from the MAR assumption, selected missing not at random (MNAR) models will be used to stress test the primary result.
In this case study, we propose a sensitivity analysis using a placebo-based multiple imputation (MI) framework with Bayesian Piecewise Exponential Model of K intervals (PEM[K]) as the imputation model for recurrent events and other similar outcomes. The methods are illustrated by intensive simulation studies.
|