Federated Causal Inference from Observational Data

Abstract

Decentralized data sources are prevalent in real-world applications, posing a formidable challenge for causal inference. These sources cannot be consolidated due to privacy constraints, and differences in data distributions and missing values across them can introduce bias into causal estimands. In this article, we propose a unified framework for estimating causal effects from decentralized observational data without sharing raw data. This contributes to privacy-preserving causal learning across heterogeneous and incomplete environments. The framework includes three novel instances, each tailored to address a distinct challenge in federated causal inference. First, we introduce a Bayesian framework based on Gaussian processes that estimates posterior distributions of causal effects and computes higher-order statistics to quantify uncertainty. Second, we develop an adaptive transfer algorithm that uses Random Fourier Features to learn similarities among data sources and disentangle the loss function into source-specific components–without requiring prior knowledge of similarity metrics. Third, we present a method for federated causal inference from incomplete data, enabling the estimation of causal effects under the missing-at-random assumption while also capturing higher-order uncertainty in the estimands. Together, these components address core limitations of existing approaches, which often handle privacy, heterogeneity, and missingness in isolation. Our framework offers a principled and scalable solution for robust, privacy-aware causal inference across decentralized data landscapes.