Causal Inference Across Private Data Silos

33 concepts30 papers

Standard machine learning excels at identifying correlations, mapping inputs to outputs by modeling the conditional probability $P(Y|X)$. Causal inference, however, demands more: it seeks to uncover the underlying structural mechanisms—often represented as a Directed Acyclic Graph (DAG)—or to predict the outcomes of active interventions, denoted by the do-calculus expression $P(Y|\text{do}(X))$.

To reliably map these cause-and-effect relationships, algorithms traditionally require massive, centralized datasets. Yet in high-stakes domains like healthcare, finance, and genomics, data is rarely centralized. Instead, it is locked within isolated organizational silos due to strict privacy regulations (such as HIPAA and GDPR) and competitive boundaries, a challenge highlighted early on by Shichao Zhang et al..

To bridge this gap, a major research direction has emerged at the intersection of distributed computing, cryptography, and statistics. Initiated by foundational work from Matt J. Kusner et al., who first formalized the integration of causal discovery with differential privacy, this field answers a critical question: How can we discover global causal structures and estimate treatment effects across decentralized networks without ever exposing raw, individual-level data?

Decomposing the Graph: Federated Structure Learning

Causal structure learning aims to discover the DAG that represents the data-generating process. There are two primary ways to do this without centralizing data: constraint-based methods and score-based continuous optimization.

Constraint-Based Aggregation

Constraint-based algorithms, like the classic Peter-Clark (PC) algorithm, rely on a sequence of conditional independence (CI) tests to prune edges from a fully connected graph. The key intuition behind decentralizing this process is that a CI test—asking whether variable $X$ is independent of $Y$ given a set of variables $Z$—does not require looking at individual records across the entire network simultaneously.

Instead, as shown by Jianli Huang et al., clients can perform local CI tests or compute local statistics independently. A central coordinator can then aggregate these local insights using layer-wise strategies to systematically reconstruct the global causal skeleton. Because only test statistics or graph updates are shared, the underlying patient or customer records remain completely localized.

Score-Based Continuous Optimization

Score-based methods treat graph discovery as an optimization problem, searching for a DAG that maximizes a score function (like likelihood or score-based loss). Modern breakthroughs have transformed this combinatorial search into a continuous optimization problem by utilizing differentiable characterizations of acyclicity.

This continuous formulation fits naturally into the Federated Learning (FL) paradigm. Global loss functions can often be mathematically decomposed into a sum of local losses across different data sites. As explored by Ignavier Ng et al. and Erdun Gao et al., clients can compute the gradients of these causal score functions on their local data and transmit only the gradients to a central server. The server aggregates these gradients, updates the global adjacency matrix, and sends it back to the clients for the next iteration.

[Client 1: Local Data] ──> Compute Local Gradient ──┐
                                                    ▼
[Client 2: Local Data] ──> Compute Local Gradient ──┼─> [Central Server] ──> Update Global DAG
                                                    ▲   (Aggregates Gradients)
[Client 3: Local Data] ──> Compute Local Gradient ──┘

To prevent adversarial observers from reconstructing private data from these shared gradients, Aurélien Bellet et al. demonstrated that greedy coordinate-wise updates can be combined with Differential Privacy (DP). By injecting precisely calibrated noise into the aggregated statistics, the algorithm guarantees that the presence or absence of any single individual's record will not significantly alter the final output graph, providing mathematical privacy without sacrificing structural accuracy.

Overcoming Data Heterogeneity

In real-world federated settings, data distributions vary drastically between sites (non-IID data). For instance, a hospital in one city might have an older patient demographic than a hospital in another, introducing severe selection bias and unmeasured environmental confounding.

To prevent these distribution shifts from distorting the global model, Loka Li et al. introduced frameworks that leverage non-parametric tests and surrogate variables to absorb site-specific, non-linear variations. This ensures that the global model captures true invariant causal mechanisms rather than local statistical anomalies.

Alternatively, David Kaltenpoth et al. bypassed gradient sharing entirely by introducing a "regret"-based scoring mechanism. In this setup, clients evaluate how much a proposed global causal model underperforms compared to their own locally optimized model. By sharing only these localized "regret" scores, the network can collaboratively vote on the validity of causal edges with high communication efficiency and robust privacy.

Quantifying the Effect: Private Treatment Estimation

Discovering the causal graph is often just the first step; scientists frequently need to measure the exact strength of a causal link, such as the Average Treatment Effect (ATE) of a clinical intervention.

When data is fragmented across silos, estimating these effects is complicated by "covariate mismatch" (where different sites record different sets of patient characteristics) and varying sample sizes. Larry Han et al. solved this by developing a multiply robust federated estimation framework. By combining local nuisance models (which estimate propensity scores and outcome regressions locally) with adaptive global weighting, they proved that accurate treatment effects can be synthesized across sites without centralizing patient-level data.

Similarly, Yuxuan Liu et al. proposed a disentanglement strategy. Their method separates data features into two distinct spaces: a shared causal space that remains invariant across all silos, and a private space that captures site-specific nuances. This allows cross-silo models to generalize effectively across heterogeneous populations.

Private Instrumental Variable Regression

A persistent challenge in causal inference is endogeneity, which occurs when unmeasured confounders influence both the cause and the effect. To obtain unbiased estimates, researchers use an Instrumental Variable (IV)—a variable $Z$ that influences the treatment $X$ but has no direct causal link to the outcome $Y$ except through $X$.

  [Instrument: Z] ──> [Treatment: X] ──> [Outcome: Y]
                             ▲                ▲
                             └──── [Hidden] ──┘
                                 Confounder

In centralized settings, this is typically solved via Two-Stage Least Squares (2SLS) regression. However, performing 2SLS across private data silos introduces severe privacy risks, as intermediate regression steps can leak individual identities.

To address this, Haodong Liang et al. reformulated the classical two-stage approach into an iterative, two-stage gradient descent algorithm (DP-2S-GD). By injecting noise directly into the gradient updates of both regression stages, the authors established the first differentially private IV regression algorithm that maintains provable finite-sample convergence rates. This allows researchers to correct for hidden confounding while maintaining strict mathematical privacy boundaries.

Bypassing Raw Data: Summary Statistics and Cryptography

In some scenarios, establishing an active federated learning infrastructure is impossible due to organizational friction or extreme data residency laws. Researchers have therefore developed methods that bypass the need for raw data entirely, relying instead on public summary statistics or secure cryptographic handshakes.

Causal Inference from Summary Statistics

In genetic epidemiology, individual-level genomic data is highly protected. However, large-scale studies frequently publish summary statistics—such as the marginal correlations between specific genetic variants and traits.

Work by Ioan Gabriel Bucur et al. and Shimeng Huang et al. demonstrated that these public summaries contain sufficient information to reconstruct complex causal pathways. By assuming sparsity—the principle that out of thousands of potential genetic markers, only a small subset truly drive a specific biological outcome—their methods can estimate causal effects and account for unmeasured confounding using only public, aggregated summary data.

Cryptographic Valuation of Data Merges

Finally, organizations often face a catch-22: how do you determine if another institution's siloed data is valuable enough to justify the administrative and legal hurdles of a collaborative analysis, without seeing their data first?

Jake Fawkes et al. resolved this dilemma by introducing a privacy-preserving framework based on Secure Multi-Party Computation (MPC). Their protocol allows two organizations to jointly compute the Expected Information Gain (EIG) for specific causal parameters before any data is shared or merged. By executing the calculation inside a cryptographic wrapper, both parties can verify whether a data partnership will significantly reduce the uncertainty of their causal estimates, ensuring complete privacy up until the moment collaboration begins.

Key Foundational Work

Multi-Database Mining (2003) — contributed 1 concept elements to this direction.
Private Causal Inference (2016) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_privacy_preserving_causal_discovery_b12198a6,
  title = {privacy-preserving causal discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/privacy-preserving-causal-discovery/rdr_b12198a6dcdf475cb923667af720255a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_privacy_preserving_causal_discovery_b12198a6_v1,
  title = {privacy-preserving causal discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/privacy-preserving-causal-discovery/rdr_b12198a6dcdf475cb923667af720255a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake