Semiparametric principal stratification analysis beyond monotonicity

Abstract: Intercurrent events, common in clinical trials and observational studies, affect the existence or interpretation of final outcomes. Principal stratification addresses these challenges by defining local average treatment effects within latent subpopulations, but often relies on restrictive assumptions such as monotonicity and counterfactual intermediate independence. To address these limitations, we propose a unified semiparametric framework for principal stratification analysis leveraging a margin-free, conditional odds ratio sensitivity parameter. Under principal ignorability, we derive nonparametric identification formulas and develop efficient estimation methods, including a conditionally doubly robust parametric estimator and a de-biased machine learning estimator with data-adaptive nuisance estimators. Simulations show that incorrectly assuming monotonicity can often lead to suboptimal inference, while specifying non-trivial odds ratio sensitivity parameter can enable approximately valid inference under monotonicity. We apply our methods to a critical care trial and further suggest a semiparametric sensitivity analysis approach under violation of principal ignorability.