Federated causal inference in heterogeneous observational data.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Estimating treatment effects is crucial in healthcare and research.
• Data privacy constraints often limit centralized data analysis.
• Heterogeneity across sites (populations, treatment assignment) poses analytical challenges.

Purpose of the Study:

• To develop federated methods for estimating average treatment effects (ATE) across multiple sites.
• To address privacy concerns by analyzing local data without direct sharing.
• To account for population and treatment assignment heterogeneity in federated analysis.

Main Methods:

• Utilized propensity score methods for local summary statistic computation.
• Developed aggregation schemes to combine site-specific statistics.
• Ensured aggregation accounts for heterogeneity in treatment assignment and outcomes.

Main Results:

• The proposed federated estimators are consistent and asymptotically normal.
• Aggregation schemes successfully addressed site-specific heterogeneity.
• Demonstrated validity using a comparative study on two large medical claims databases.

Conclusions:

• Federated methods provide a viable solution for multi-site treatment effect estimation under privacy constraints.
• Accounting for heterogeneity is essential for robust asymptotic properties in federated learning.
• The developed methods are effective and validated on real-world healthcare data.