Synthetic Data Risks for Causal Inference

A recent study published on arXiv (2604.23904) indicates that generative synthetic data, such as those derived from GAN and LLM models, may skew causal estimands like the average treatment effect (ATE), despite achieving high predictive accuracy. The researchers articulate this issue through sensitivity analyses and tradeoff findings, demonstrating that maintaining ATE necessitates managing both the generated covariate distribution and the treatment-effect difference. To address this, they introduce a hybrid framework that generates covariates independently from treatment and outcome processes, employing distance-to-closest-record diagnostics alongside distinct nuisance models.