Keywords: Estimand, causal inference, Bayesian analysis, clinical trial, identification, subgroup
In a randomized controlled trial, occurrence of post-randomization events associated with treatment and the primary endpoint may complicate the interpretation of the overall treatment effect. In this presentation, we discuss how these events may be accounted for at the estimand and the estimator level in the context of a recent case study. We define a principal stratification estimand derived from the scientific question of interest – specifically, we consider the treatment effect in the principal stratum of patients where the intercurrent event would not occur regardless of treatment assignment. We discuss the extent to which this estimand can be identified from the data and the potential identifying assumptions that may be considered. A Bayesian model is proposed in which we estimate the estimand of interest by jointly modeling the principal strata proportions and the primary endpoint with mixture distributions. In order to account for variable follow-up time, estimates are standardized over the empirical covariate distribution. Finally, we discuss the use of sensitivity analyses to assess robustness to departures from the key identifying assumptions.