Efficient collaborative learning of the average treatment effect.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Generating real-world evidence from multisite studies is crucial but challenging due to data-sharing constraints.
• Existing methods for integrating data across sites in causal inference often require iterative communication or struggle with distributional shifts.

Purpose of the Study:

• To introduce ECO-ATE (Efficient Collaborative learning to Evaluate Average Treatment Effect), a federated learning approach for multisite causal inference.
• To develop an efficient estimator for average treatment effect on a target population using individual-level data and summary statistics from other populations.
• To enable robust causal inference in multisite settings without iterative data exchange.

Main Methods:

• ECO-ATE employs a federated learning strategy, utilizing target population's individual data and source populations' summary statistics.
• The approach does not require iterative communication between research sites, facilitating resource-limited consortia.
• It is designed to accommodate distributional shifts in outcomes, treatments, and baseline covariates.

Main Results:

• Simulation studies demonstrated significant efficiency gains by incorporating additional data sources with ECO-ATE.
• The method showed robustness against varying distributional shifts and overparameterization compared to existing benchmarks.
• ECO-ATE achieved semiparametric efficiency bounds under appropriate conditions.

Conclusions:

• ECO-ATE provides an efficient and robust solution for estimating average treatment effects in multisite studies under data-sharing constraints.
• The federated approach is suitable for research consortia, enabling causal inference without complex data-sharing infrastructure.
• The method's ability to handle distributional shifts enhances its applicability to diverse real-world electronic health record data.