Mitigating Unobserved Confounding via Diffusion Probabilistic Models

ICLR 2026 Conference Submission22532 Authors

Keywords: causal inference, Diffusion Probabilistic Models, treatment effect estimation

Abstract:

Learning Conditional average treatment effect estimation from observational data is a challenging task due to the existence of latent covariates. Previous
methods mostly focus on assuming the ignorability assumption ignoring the latent covariates or overlooking the impact of an apriori knowledge on the generation process of the latent variable, which can be quite impractical in real-world scenarios. We introduce a novel framework that mitigates unobserved confounding by generating the latent covariates using a conditional diffusion probabilistic model. This model first infers a causal anchor variable from the observed data, and then uses this variable to guide a reverse diffusion process that synthesizes the unobserved covariate. We render this architecture tractable by deriving a closed-form variational lower bound for its optimization. To ensure causal validity, we theoretically analyze that the latent variable $z$ learned by our model is orthogonal-identifiable. In the experiments, we compare our model with the state-of-the-art methods based on two standard benchmarks, demonstrating consistent improvements of our model.

Primary Area: causal reasoning

Submission Number: 22532