Causal penalized training improves synthetic data for causal analysis
Beyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data
Machine Learning
Summary
Synthetic data generators usually focus on making fake data look like real data by matching overall statistics. However, this does not guarantee that cause-and-effect relationships stay the same. The authors define a way to measure how well synthetic data preserves these causal effects and show that making data statistically similar alone is not enough. They design a new training method that adds a penalty for differences in causal effects to improve causal accuracy in synthetic data generation. Their tests show this approach can better preserve cause-and-effect information without losing overall data quality.
What this means in practice
- •For healthcare data teams: Create synthetic patient records that better preserve causal relationships for safer and more accurate medical analysis.
- •For financial data engineers: Generate synthetic financial datasets that maintain causality for improved testing of risk models without exposing sensitive data.
Authors
Lan Tao, Yongxian He, Shirong Xu, Yidong Ouyang, Guang Cheng
Abstract
Synthetic tabular generators are commonly optimized for distributional fidelity, but statistical similarity alone does not guarantee preservation of causal effects. In this paper, we study whether causal fidelity can be improved directly within a fully generative tabular model. Causal Fidelity is defined with respect to a target estimand as the discrepancy between inferential distributions obtained from real and synthetic data, and theoretical results show that high statistical fidelity does not generally imply high causal fidelity. We then propose a causal-fidelity-aware training framework which adds a causal discrepancy penalty to the generative objective. The framework is instantiated with a causal-penalized TabDDPM and optimized using an on-policy score-function estimator. We further establish conditions under which causal regularization improves expected causal fidelity. Experiments across diverse treatment-effect simulations and two benchmark datasets evaluate the ability of our method to improve causal fidelity while preserving competitive statistical fidelity.