Causal inference—the science of determining not just that two things happen together, but that one causes the other—traditionally requires massive, centralized datasets. To identify a treatment effect or map a causal graph, researchers usually need to pool individual-level records into a single database. However, in sensitive domains like healthcare and finance, privacy regulations such as GDPR and HIPAA, along with institutional data silos, make this centralization impossible.
Federated Causal Learning resolves this tension. It allows multiple parties to collaboratively discover causal structures (Directed Acyclic Graphs or DAGs) and estimate treatment effects without ever sharing raw, private data. Instead of moving the data to the model, the model's "understanding" of causal relationships is aggregated across decentralized sites.
Learning the Causal Map without Raw Data
The first challenge is Causal Structure Learning: identifying the "who-causes-what" in a system of variables. Traditionally, algorithms like the Peter-Clark (PC) algorithm or score-based methods require direct access to all samples to perform conditional independence tests.
In a federated setting, researchers like Jianli Huang and Kui Yu have adapted these methods into frameworks like FedPC Towards Privacy-Aware Causal Structure Learning in Federated Setting. Rather than sharing data, clients perform local independence tests and share only the resulting "skeletons" or statistical summaries. The server then aggregates these to form a global DAG.
A more modern approach treats structure learning as a continuous optimization problem. Ignavier Ng and Kun Zhang leverage the "NOTEARS" formulation, which turns the search for a graph into a calculus problem that can be solved via the Alternating Direction Method of Multipliers (ADMM) Towards Federated Bayesian Network Structure Learning with Continuous Optimization. Similarly, Erdun Gao and colleagues developed FedDAG, which uses a functional characterization of DAGs to ensure the global model remains acyclic while learning from distributed gradients FedDAG: Federated DAG Structure Learning.
Estimating Effects Across Silos
Once the causal structure is known, the next goal is Causal Inference: measuring the strength of a specific cause (e.g., a drug) on an effect (e.g., patient recovery). This is often expressed as the Average Treatment Effect (ATE).
Aurélien Bellet, Julie Josse, and Rémi Khellaf have shown that federated methods can achieve the same statistical precision as centralized pooling Federated Causal Inference: Multi-Centric ATE Estimation beyond Meta-Analysis. Their approach involves Propensity Score Aggregation, where each site calculates how likely a patient is to receive a treatment based on local covariates. These scores are then combined globally to adjust for confounding variables that might otherwise bias the results.
For more complex scenarios where data distributions vary wildly between sites (non-IID data), Thanh Vinh Vo and Young Lee introduced FedCI Bayesian federated estimation of causal effects from observational data. By using Gaussian Processes and variational inference, FedCI provides not just a point estimate of the treatment effect, but also a measure of uncertainty, which is critical for clinical decision-making.
Overcoming Unobserved Confounders
A major hurdle in causal discovery is "endogeneity"—when hidden factors influence both the cause and the effect. In centralized science, researchers use Instrumental Variables (IV) to bypass these hidden factors.
Translating IV analysis to a federated environment is difficult because it requires satisfying complex "moment conditions" across different datasets. Geetika and Bapi Chatterjee addressed this with FEDIV, which formulates the problem as a federated zero-sum game Federated Instrumental Variable Analysis via Federated Generalized Method of Moments. This allows the system to identify parameters that satisfy causal conditions across all clients without any raw data exchange. To ensure these gradients don't leak information, Haodong Liang and colleagues developed DP-2S-GD, a differentially private version of two-stage gradient descent for IV regression Differentially Private Two-Stage Gradient Descent for Instrumental Variable Regression.
Handling Heterogeneity and Quality
In the real world, decentralized data is rarely uniform. One hospital might have more elderly patients, while another has different sampling rates.
- Data Heterogeneity: Loka Li and Biwei Huang developed FedCDH, which uses surrogate variables to handle non-linear causal models across heterogeneous sites Federated Causal Discovery from Heterogeneous Data.
- Sample Quality: Not all clients provide equally reliable data. Xianjie Guo and Hao Wang proposed FedACD, which adaptively masks unreliable local results to prevent low-quality data from poisoning the global causal graph Discrete Budget Aggregation: Truthfulness and Proportionality.
- Temporal Drift: For time-series data, Jiegang Xu introduced Fed-CAW, which uses "credibility scores" to align mismatched sampling rates across different institutions Credibility-Aware Weighting Federated Causal Discovery for Time Series.
By combining these decentralized optimization techniques with rigorous privacy guarantees like Differential Privacy, Federated Causal Learning enables large-scale scientific discovery that respects individual confidentiality.