Identification and estimation of causal effects in the presence of confounded principal strata

TL;DR

This paper develops methods to identify and estimate causal effects within principal strata using negative controls, addressing unobserved confounding in observational studies.

Contribution

It introduces a novel approach leveraging negative controls for nonparametric identification of principal causal effects under unobserved confounding.

Findings

Nonparametric identification of causal effects with negative controls

Proposed estimation method performs well in simulations

Application demonstrates practical utility in real data

Abstract

The principal stratification has become a popular tool to address a broad class of causal inference questions, particularly in dealing with non-compliance and truncation-by-death problems. The causal effects within principal strata which are determined by joint potential values of the intermediate variable, also known as the principal causal effects, are often of interest in these studies. Analyses of principal causal effects from observed data in the literature mostly rely on ignorability of the treatment assignment, which requires practitioners to accurately measure as many as covariates so that all possible confounding sources are captured. However, collecting all potential confounders in observational studies is often difficult and costly, the ignorability assumption may thus be questionable. In this paper, by leveraging available negative controls that have been increasingly used to deal with uncontrolled confounding, we consider identification and estimation of causal effects when the treatment and principal strata are confounded by unobserved variables. Specifically, we show that the principal causal effects can be nonparametrically identified by invoking a pair of negative controls that are both required not to directly affect the outcome. We then relax this assumption and establish identification of principal causal effects under various semiparametric or parametric models. We also propose an estimation method of principal causal effects. Extensive simulation studies show good performance of the proposed approach and a real data application from the National Longitudinal Survey of Young Men is used for illustration.