New Method Creates Private Synthetic Data for Causal Inference

A new research paper proposes a method for generating synthetic data that is both differentially private and optimized for causal inference. Traditional differentially private synthetic data methods focus on general data distributions, but this new approach specifically targets the preservation of causal relationships, such as the average treatment effect (ATE), which are crucial for understanding cause-and-effect. The core of the method involves defining "causal workloads" – sets of private queries designed around the specific statistical moments used in robust causal estimators. These workloads can either be directly used by estimators or transformed into reusable synthetic data through maximum-entropy calibration. The framework also includes Causal-AIM for adaptive workload selection and a noise-aware multiple-imputation procedure for accurate confidence intervals. The study highlights a trade-off: while generic privacy methods might offer better point accuracy for overall data distribution, this causal workload approach excels at strict privacy budgets and provides more reliable uncertainty quantification for causal analyses. This allows the same private synthetic dataset to support various causal analyses without additional privacy costs.