Federated Adaptive Causal Estimation (FACE) of Target Treatment Effects.

Area of Science:

• Biostatistics
• Epidemiology
• Machine Learning

Background:

• Federated learning offers potential for efficient causal inference by pooling data from multiple sites.
• Ensuring robustness to data heterogeneity and model misspecification is crucial for valid causal estimands.

Purpose of the Study:

• To develop a robust and efficient federated learning framework for causal effect estimation.
• To address heterogeneity across study sites and ensure privacy-preserving data sharing.

Main Methods:

• Developed the Federated Adaptive Causal Estimation (FACE) framework.
• Utilized density ratio weighting to account for covariate distribution heterogeneity.
• Implemented an adaptive weighting procedure via penalized regression for consistency and efficiency.
• Employed a communication-efficient and privacy-preserving strategy sharing only summary statistics.

Main Results:

• FACE demonstrated increased precision in treatment effect estimates compared to traditional methods.
• Reductions in standard errors ranged from 26% to 67% in the vaccine effectiveness study.
• The framework proved robust to site-level heterogeneity and model misspecifications.

Conclusions:

• FACE provides a valid, efficient, and robust approach for federated causal inference.
• The framework enhances precision and allows for flexible target population specification.
• FACE is applicable to real-world comparative effectiveness research using electronic health records.