Primary Submission Category: Principal stratification
Semiparametric principal stratification analysis without monotonicity
Authors: Jiaqi Tong,
Presenting Author: Jiaqi Tong*
Intercurrent events, common in clinical trials and observational studies, affect the existence or interpretation of final outcomes. Principal stratification addresses these challenges by defining average treatment effects within subpopulations using counterfactual intermediate outcomes as pre-treatment covariates. However, most methods rely on strong assumptions, such as monotonicity and counterfactual intermediate independence. To relax these assumptions, we consider a margin-free, and variation-independent framework for principal stratification analysis based on a conditional odds ratio sensitivity parameter. Under principal ignorability, we derive non-parametric identification formulas for principal causal effects and propose weighting and regression approaches for estimation. We further derive the efficient influence function to construct a conditionally doubly robust estimator, as well as a de-biased machine learning estimator. We use extensive simulations to illustrate the consequence of incorrectly assuming monotonicity, and the implications of misspecifying the sensitivity parameter under non-monotonicity. We apply our methods to two critical care clinical trials comparing active treatments where monotonicity is deemed questionable.