Multiply Robust Federated Estimation of Targeted Average Treatment Effects

Larry Han, Zhu Shen, José R. Zubizarreta

Introduction

In modern medical and social science research, the ability to draw causal conclusions from diverse datasets is paramount. However, data is often siloed across different institutions—such as hospitals or government agencies—due to privacy regulations and logistical barriers. Federated learning offers a solution by allowing researchers to learn from decentralized data without direct access to raw records. This paper introduces a novel approach to estimating the Targeted Average Treatment Effect (TATE) in a federated setting, ensuring robust and accurate causal inference even when data is heterogeneous or models are imperfectly specified.

The Challenges of Federated Causal Inference

Estimating causal effects—determining if a specific treatment directly caused an outcome—requires accounting for various biases. In a federated environment, three primary challenges arise:

  1. Covariate Shift: The population at a source site (e.g., a specialized clinic) may differ significantly from the target population (e.g., a general hospital) in terms of age, health status, or demographics.
  2. Covariate Mismatch: Different sites may record different variables. One site might track BMI and smoking status, while another only records age and weight, complicating the alignment of datasets.
  3. Model Misspecification: Causal models rely on mathematical assumptions about treatment assignment and outcomes. If a model at one site is incorrect, it can "poison" the global estimate.

The Multiply Robust Framework

The authors propose a "multiply robust" federated estimator designed to act as a statistical safety net. The framework integrates several advanced techniques to address the aforementioned challenges.

Multiply Robust Nuisance Estimation

Traditional estimators often rely on a single model for treatment assignment (propensity score) or outcome regression. If that model is wrong, the estimate is biased. The proposed method allows each site to propose multiple candidate models. As long as at least one model in the ensemble is correctly specified, the final TATE estimate remains consistent. This "any-is-enough" property is crucial in federated settings where researchers cannot manually inspect data at remote sites.

Adaptive Ensembling (MR-L1)

To handle covariate shift and prevent "negative transfer"—where irrelevant data from a source site degrades the target estimate—the researchers employ an L1-penalized regression (Lasso-style) on influence functions. This adaptive ensembling technique automatically weights the contributions of different sites. If a source site’s data is too dissimilar or unhelpful for the target population, the algorithm shrinks its influence to zero, effectively filtering out noise.

Privacy-Preserving Alignment

To align populations across sites without sharing raw patient data, the method utilizes an "exponential tilt" model for density ratios. Sites only need to share summary statistics, such as covariate means and variances. Furthermore, to address covariate mismatch, the authors use hyperplane projection to map site-specific outcome regressions onto a common set of variables shared with the target site.

Experimental Evaluation

The methodology was tested using complex simulated datasets designed to mimic real-world irregularities, including non-linear transformations and skewed distributions. The setup involved five sites with varying sample sizes (300 for the target site and up to 1000 for source sites).

The results demonstrated significant advantages over standard methods:

  • Robustness to Error: While standard Augmented Inverse Probability Weighting (AIPW) failed when models were misspecified, the MR-L1 estimator maintained low bias and high accuracy.
  • Efficiency Gains: By successfully leveraging relevant source sites, the federated estimator produced much tighter confidence intervals than those achieved using the target site's data alone.
  • Handling Mismatch: In covariate mismatch tests, the MR-L1 estimator significantly outperformed simpler weighting schemes like Inverse Variance Weighting (IVW), which produced highly biased results when variables did not match perfectly.
  • Statistical Reliability: The method consistently achieved approximately 95% coverage for confidence intervals, a key metric for scientific reliability.

Conclusion and Future Directions

This research provides a robust path forward for privacy-preserving causal inference. By combining multiple robustness with adaptive ensembling, the proposed estimator allows researchers to harness the power of multi-site data while guarding against the common pitfalls of data heterogeneity and model error.

Future work may focus on extending this framework to high-dimensional data settings where the number of covariates is very large. Additionally, applying this method to real-world federated clinical networks, such as PCORnet or OHDSI, will be essential to validate its performance against the inherent noise of electronic health records.

Summary updated

BibTeX
@inproceedings{han2023multiply,
  author = {Larry Han and Zhu Shen and José R. Zubizarreta},
  title = {{Multiply Robust Federated Estimation of Targeted Average Treatment Effects}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=M6UccKMFGl},
}
Figure