Researchers have worked out how to run causal inference across multiple data silos without any site surrendering its raw records.
A new paper proposes estimating the Average Treatment Effect - a standard measure of whether an intervention actually works - from observational data distributed across multiple sites. Instead of pooling individual records, participating sites compute local propensity scores (the likelihood that a given observation received a treatment), then share only those aggregates. A weighting scheme called Membership Weights adjusts for the fact that each site's population differs from the whole, yielding two new estimators - Federated Inverse Propensity Weighting and an augmented variant - that the authors test on simulated and real-world data, outperforming meta-analysis approaches in both settings.
Most data that could answer important questions in medicine, policy, and social science sits locked in silos: hospital networks barred from sharing patient records, agencies that will not export individual data across jurisdictions, or organizations that cannot justify the legal exposure. Standard meta-analysis collapses when any single site has a coverage gap, a condition researchers call positivity violation; this method turns cross-site heterogeneity into an asset rather than a problem by exploiting the fact that different sites may assign treatments differently.
This is academic research, not a production tool, and federated methods have a long history of promising more than they deliver outside the lab. Still, the specific gap being addressed here - extracting causal claims, not just correlations, from siloed observational data - is one with genuine demand in healthcare and policy research that existing approaches handle badly.