Generative Synthetic Data for Causal Inference: Pitfalls, Remedies, and Opportunities

arXiv stat.ML / 4/28/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper shows that fully generative tabular synthetic data methods (including GAN- and LLM-based synthesizers) can look strong on predictive performance while still significantly distorting causal estimands like the average treatment effect (ATE).
It formalizes why this happens: preserving ATE requires controlling not only predictive fidelity, but also the generated covariate distribution and the treatment-effect contrast in the outcome regression.
The authors propose a hybrid synthetic-data framework that generates covariates separately from the treatment and outcome mechanisms, then uses diagnostics (distance-to-closest-record) plus separately learned nuisance models to build (W, A, Y) triplets.
They also study targeted synthetic augmentation for positivity/overlap problems and introduce a synthetic simulation engine to evaluate causal estimators (OR, IPW, AIPW, TMLE) in finite samples.
Experiments indicate that the hybrid approach improves ATE preservation compared with fully generative baselines and provides practical tools for more robust causal analysis under synthetic data settings.

Abstract

Synthetic data offers a promising tool for privacy-preserving data release, augmentation, and simulation, but its use in causal inference requires preserving more than predictive fidelity. We show that fully generative tabular synthesizers, including GAN- and LLM-based models, can achieve strong train-on-synthetic-test-on-real performance while substantially distorting causal estimands such as the average treatment effect (ATE). We formalize this failure through sensitivity and tradeoff results showing that ATE preservation requires control of both the generated covariate law and the treatment-effect contrast in the outcome regression. Motivated by this observation, we propose a hybrid synthetic-data framework that generates covariates separately from the treatment and outcome mechanisms, using distance-to-closest-record diagnostics to monitor covariate synthesis and separately learned nuisance models to construct (W, A, Y) triplets. We further study targeted synthetic augmentation for practical positivity problems and characterize when added overlap support helps by improving conditional-effect estimation more than it shifts the covariate distribution. Finally, we develop a synthetic simulation engine for pre-analysis estimator evaluation, enabling finite-sample comparison of OR, IPW, AIPW, and TMLE under realistic covariate structure. Across experiments, hybrid synthetic data substantially improve ATE preservation relative to fully generative baselines and provide a practical diagnostic tool for robust causal analysis.