- Researchers have released a federated approach for causal inference that works on decentralized observational data.
The paper proposes estimating propensity scores by sharing only aggregate statistics. Each site trains a local model; the results are combined using Membership Weights that reflect the probability a record belongs to a given site given its covariates. These federated scores feed into two estimators—Federated Inverse Propensity Weighting (Fed‑IPW) and its augmented version (Fed‑AIPW). Experiments on simulated and real‑world datasets show the method handles differing sample sizes, treatment mechanisms, and covariate distributions better than standard meta‑analysis, which often breaks down when any site lacks positivity.
The advance matters because many health, finance, and social science studies are siloed by regulation or logistics. By avoiding raw data exchange, the technique respects privacy constraints while still exploiting heterogeneity that traditionally harms pooled analyses. It also offers a concrete alternative to the ad‑hoc workarounds researchers have used for years.
In short, federated propensity‑score aggregation provides a privacy‑preserving path to reliable average treatment effect estimates, but questions remain about scalability to many sites, robustness to model misspecification, and real‑world deployment hurdles.