Conditional diffusion for causal inference with state space representation

doi:10.1016/j.knosys.2026.116800

Conditional diffusion for causal inference with state space representation

2026-08-01

0

PRE

AI

DOI：10.1016/j.knosys.2026.116800

Original

Original request for help

Save

Abstract

En 中文

Causal inference from observational data, particularly for estimating individual-level effects, is vital in high-stakes domains such as healthcare and economics. Many established approaches to potential outcome prediction provide a single estimate under each treatment, which does not describe outcome variability or distributional heterogeneity across individuals. Recent generative approaches have moved beyond such point estimates by estimating conditional potential outcome distributions, but effectively incorporating structural and temporal dependencies among observational covariates remains a key challenge. To address these gaps, we propose CoDiS, a conditional diffusion framework for estimating potential outcome distributions through a structure-aware state space denoising architecture. Specifically, CoDiS jointly processes the covariate-treatment condition, the noisy outcome state, and diffusion timestep with an SSM-based backbone during reverse denoising. This design allows the reverse denoising process to exploit latent dependencies among observational variables, thereby improving the potential outcome distribution generation. Extensive evaluations on a COVID-19 dataset and standard causal benchmarks demonstrate the effectiveness of CoDiS in both distributional and point estimation. The code is available at https://github.com/JustinLiam/CoDiS.

Keywords:

Causal inference

Potential outcome prediction

State space model

Conditional diffusion

Distributional estimation

Conditional diffusion for causal inference with state space representation

Abstract

Journal

Organization

Cited Papers