Primary Submission Category: Generalizability/Transportability
Population-Level Causal Effect Estimation in the Presence of Noncompliance: A Bayesian Approach Integrating RCT with Observational Studies
Authors: Yueying Hu, Yajuan Si, Michael Elliott,
Presenting Author: Yueying Hu*
RCTs are the gold standard for causal inference but are often challenged by treatment noncompliance and limited generalizability when trial participants do not represent the target population. This work is motivated by the REFLUX trial, which compared laparoscopic fundoplication with medical management for gastroesophageal reflux disease (GERD). Of 810 enrolled patients, 453 expressed strong treatment preferences and were excluded from randomization but followed in a parallel preference arm. Both arms collected identical covariates and outcomes, while treatment uptake followed different mechanisms.
To jointly address noncompliance and generalizability, we combine the RCT and preference arm as a proxy for the target population and develop an integrative Bayesian framework for population-level complier average causal effect (CACE) estimation. Using the RCT data, we estimate a covariate-dependent latent compliance class model via principal stratification. This model is then used to infer compliance class membership for preference-arm participants from baseline covariates. Conditional on inferred classes and observed treatment uptake, outcomes from both arms are used to estimate population-level CACE by averaging class-specific treatment effects over the target covariate distribution. We evaluate the proposed methods through simulations and an application to REFLUX, and compare them with weighting-based alternatives in settings where outcomes may not be available for benchmark.