Primary Submission Category: Generalizability/Transportability
Multiply robust federated estimation of targeted average treatment effects
Authors: Zhu Shen, Larry Han, Jose Zubizarreta,
Presenting Author: Zhu Shen*
Multi-site studies have advantages including increased generalizability, the ability to study underrepresented populations, and the opportunity to study rare exposures and outcomes. However, these studies are challenging due to the need to preserve the privacy of each individual’s data and the heterogeneity in their covariate distributions, treatment guidelines, and conditional outcome models. We propose a multiply robust federated estimator to derive valid causal inferences for a target population using multi-site observational data. Our proposed estimator is more flexible than current methods as it relaxes the requirement of homogeneous model specifications across sites; investigators from different sites can leverage site-specific knowledge such as patient preferences and treatment guidelines to propose multiple different models for the outcome and treatment. Our proposed estimator adopts a model mixing approach and is consistent for the target average treatment effect if either one of the outcome models is correctly specified or one of the propensity score models and the density ratio model are correctly specified. The estimator data-adaptively down-weights source sites that are sufficiently different from the target site to avoid the negative transfer. Numerical evaluations show that our estimator can produce comparable results to doubly robust federated estimators when models are correctly specified and offer more robustness when models are misspecified.