Differentiable Causal Discovery of Linear Non-Gaussian Acyclic Models Under Unmeasured Confounding

TL;DR

This paper introduces ABIC LiNGAM, a score-based causal discovery method that uses continuous optimization to identify causal structures in linear non-Gaussian models with unmeasured confounders, providing theoretical guarantees and practical effectiveness.

Contribution

It extends LiNGAM to handle unmeasured confounding by assuming multivariate generalized normal errors and enables direct recovery of causal graphs with theoretical identifiability.

Findings

Successfully recovers causal structures in simulations

Demonstrates effectiveness on real-world datasets

Provides theoretical guarantees on identifiability

Abstract

We propose a novel score-based causal discovery method, named ABIC LiNGAM, which extends the linear non-Gaussian acyclic model (LiNGAM) framework to address the challenges of causal structure estimation in scenarios involving unmeasured confounders. By introducing the assumption that error terms follow a multivariate generalized normal distribution, our method leverages continuous optimization techniques to recover acyclic directed mixed graphs (ADMGs), including causal directions rather than just equivalence classes. We provide theoretical guarantees on the identifiability of causal parameters and demonstrate the effectiveness of our approach through extensive simulations and applications to real-world datasets.