Beyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data
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.
Related
- Seq2Synth: Benchmarking Temporal Fidelity in Synthetic Sequential Tabular Data
- Reconciling Interpretability with Covariate-Dependent Shape Flexibility in Penalized Transformation Models for Distributional Regression
- Challenges of analyzing synthetic tabular data generated from 115 phase 3 oncology trials
- Classical methods for modern applications: synthetic tabular data generation using lightweight discrete response continuous latent variable model
- SynthGuard-ReleaseBench: Locked-Audit Evidence for Synthetic Tabular Data Releases
BibTeX
@misc{tao2026beyond,
title = {{Beyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data}},
author = {Lan Tao and Yongxian He and Shirong Xu and Yidong Ouyang and Guang Cheng},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.11407},
eprint = {2610.11407},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.11407},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Beyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data
AU - Lan Tao
AU - Yongxian He
AU - Shirong Xu
AU - Yidong Ouyang
AU - Guang Cheng
PY - 2026
DA - 2026/10/08
PB - arXiv
DO - 10.48550/arxiv.2610.11407
UR - https://arxiv.org/abs/2610.11407
AB - 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.
ER -