Primary Submission Category: Mediation Analysis, Mechanisms
Semiparametric Inference for Causal Path-Specific Effects in Longitudinal Studies
Authors: Xiaxian Ou, Razieh Nabi, Xinwei He,
Presenting Author: Xiaxian Ou*
We develop a semiparametric framework for estimating and conducting inference on causal path-specific effects in longitudinal studies involving multiple or repeatedly measured treatments, mediators, and confounders. The framework accommodates general longitudinal structures, including differing measurement intervals and varying numbers of mediators across time. Our framework focuses on estimation and inference for identifiable effects under the edge g-formula, allowing for binary, continuous, or multivariate mediators and certain patterns of unmeasured confounding among treatments, mediators, and the outcome. We propose multiply robust estimators derived from influence function theory that integrate data-adaptive machine learning techniques, and we establish rate conditions for their asymptotic linearity, efficiency, and robustness to nuisance misspecification. We further extend the framework to include a sensitivity analysis procedure that evaluates the impact of violations of cross-world ignorability assumptions. Simulation studies and an empirical application using the Framingham Heart Study demonstrate the method’s performance. An accompanying R package, flexPaths, was developed to facilitate implementation of the proposed methods.