A new computational framework enables causal inference across distributed clinical databases by combining federated aggregation with differential privacy, allowing propensity score subclassification and matching without exposing individual patient records. Applied to Long COVID and very preterm birth multicenter cohorts, the method produced treatment effect estimates closely matching centralized analyses, with quantified bias-variance trade-offs introduced by privacy noise. Two open-source R packages implement the approach within the established DataSHIELD platform.
Multicenter observational studies are epidemiology's workhorse for rare conditions and heterogeneous populations, yet covariate imbalance between treated and control groups routinely threatens causal validity — and privacy regulations like GDPR increasingly prohibit the data pooling that traditional balancing methods require. This framework directly attacks that bottleneck. If validated, it could unlock genuinely large, diverse patient populations for causal analysis in areas like post-infectious syndromes, neonatal outcomes, and oncology where no single center has sufficient sample size.
That said, this is a methods paper, not a clinical discovery. Its health implications are indirect: better analytical infrastructure may eventually yield more reliable treatment effect estimates from real-world data. Key limitations include reliance on differential privacy parameters that researchers must tune carefully — too much noise degrades estimates — and the simulation scenarios may not capture the full complexity of real institutional heterogeneity. As a preprint posted on medRxiv and not yet peer-reviewed, the statistical guarantees and implementation correctness require independent scrutiny. The contribution is genuinely useful infrastructure, but incremental within the growing federated learning literature.