Causal Canonical Modeling for Confounding Robust Treatment Evaluation

ICLR 2026 Conference Submission19638 Authors

Keywords: Causal Inference, Graphical Models

TL;DR: This paper develops a novel family of causal generative models that allow one to derive robust evaluation of unknown causal effects provided with confounded observations in complex domains.

Abstract: The chance of confounding bias presents one of the central challenges for policy evaluation, since target causal effects of actions are not computable (i.e., under-determined) from the observational data. This paper investigates treatment/causal effect evaluation over the continuous action-reward domain from confounded observations, while requiring only basic temporal ordering between the treatment and the outcome, and Lebesgue integrability over the target treatment effect. We introduce a novel family of causal canonical models that can effectively approximate the observational and interventional distributions of any causal model consisting of continuous action and reward variables. Building on this newfound universal approximation property, we develop a novel family of generative models via a mixture of Gaussian processes that allow one to derive posterior distributions over unknown causal effects provided with confounded observations.

Primary Area: causal reasoning

Submission Number: 19638