Causal Robustness from Synthetic Environments
Abstract
We study robust prediction when the training data come from a single observational distribution. Causality-oriented methods for robust prediction usually use heterogeneity across observed environments both to learn an invariant prediction function and to determine the perturbations against which it is robust. We show that both can be recovered from one distribution under a structural additive-noise mechanism for the response. Under the structural model considered here, the causal prediction function has constant conditional mean squared error given its prediction. We generate synthetic environments by reweighting the covariate distribution along the prediction and use a Cramér--von Mises criterion to measure departures from this invariance. The resulting penalised risk admits a regularised worst-case representation, linking the criterion to robustness against covariate shifts induced by reweighting observations according to their predicted values. Under an additional orthogonality condition, we identify the causal prediction function as the unique risk minimiser subject to this invariance restriction. We establish consistency and an explicit rate, derive a goodness-of-fit test for the invariance restriction, and characterise the population target under misspecification. With multiple observed environments, the construction applies environment by environment and can be combined with Distributional Robustness via Invariant Gradients (DRIG) to control two complementary classes of shifts.
Related
- A Multiverse of Good and Bad Controls: Candidate Causal Graphs for Interpreting Model Robustness Analysis
- Minimum Specification Perturbation: Robustness as Distance-to-Falsification in Causal Inference
- Debiasing and t -Tests for Synthetic Control Inference on Average Causal Effects
- Beyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data
- Robustness intervals for competing risks analysis with causes of failure missing not at random
BibTeX
@misc{yang2026causal,
title = {{Causal Robustness from Synthetic Environments}},
author = {Linying Yang and Cyrill Scheidegger and Xinwei Shen and Peter Bühlmann},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.26250},
eprint = {2610.26250},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26250},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Causal Robustness from Synthetic Environments
AU - Linying Yang
AU - Cyrill Scheidegger
AU - Xinwei Shen
AU - Peter Bühlmann
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.26250
UR - https://arxiv.org/abs/2610.26250
AB - We study robust prediction when the training data come from a single observational distribution. Causality-oriented methods for robust prediction usually use heterogeneity across observed environments both to learn an invariant prediction function and to determine the perturbations against which it is robust. We show that both can be recovered from one distribution under a structural additive-noise mechanism for the response. Under the structural model considered here, the causal prediction function has constant conditional mean squared error given its prediction. We generate synthetic environments by reweighting the covariate distribution along the prediction and use a Cramér--von Mises criterion to measure departures from this invariance. The resulting penalised risk admits a regularised worst-case representation, linking the criterion to robustness against covariate shifts induced by reweighting observations according to their predicted values. Under an additional orthogonality condition, we identify the causal prediction function as the unique risk minimiser subject to this invariance restriction. We establish consistency and an explicit rate, derive a goodness-of-fit test for the invariance restriction, and characterise the population target under misspecification. With multiple observed environments, the construction applies environment by environment and can be combined with Distributional Robustness via Invariant Gradients (DRIG) to control two complementary classes of shifts.
ER -