We introduce a novel estimator for predicting outcomes in the presence of hidden confounding across different distributional settings without relying on regularization or a known causal structure. Our approach is based on parametrizing the dependence of the covariates with response noise, ensuring optimal prediction and favorable asymptotic properties. We achieve identifiability under lean assumptions that have direct empirical translation, enabling the incorporation of causal parameters into a generative model that replicates the true conditional distribution of a test environment. This method achieves probabilistic alignment with test distributions uniformly across interventions, offering robust predictions without the need for worst-case optimization or specific assumptions about the strength of perturbations at test. Our findings represent a significant advancement in the statistical understanding of causality, providing a robust and flexible framework for predictive modeling in varied domains.
翻译:暂无翻译