How Does Bayesian Causal Discovery Fail? Characterising Structural Consequences in Linear Gaussian Networks under Latent Confounding
The paper investigates how Bayesian causal discovery behaves when latent confounding is present in linear Gaussian networks. It identifies a correlation threshold that lowers with larger sample sizes, causing the posterior to favor spurious edges between confounded variables. The authors also describe two distinct failure regimes of the posterior and validate their analysis with exact computations on various graph structures.
- ▪Bayesian causal discovery quantifies epistemic uncertainty over directed acyclic graphs through posterior inference.
- ▪The study focuses on linear Gaussian causal models with additive latent confounding between exactly two observed variables.
- ▪A critical correlation threshold is derived, which decreases as sample size increases, leading to a preference for spurious edges beyond the threshold.
- ▪Two distinct posterior failure regimes are characterized based on the local structure around the confounded variables.
- ▪The theoretical findings are supported by exact posterior calculations on multiple graph structures.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2607.09449 (cs) [Submitted on 10 Jul 2026] Title:How Does Bayesian Causal Discovery Fail? Characterising Structural Consequences in Linear Gaussian Networks under Latent Confounding Authors:Debargha Ghosh, Silja Renooij, Anna Kononova View a PDF of the paper titled How Does Bayesian Causal Discovery Fail? Characterising Structural Consequences in Linear Gaussian Networks under Latent Confounding, by Debargha Ghosh and 2 other authors View PDF HTML (experimental) Abstract:Bayesian causal discovery is widely used for its ability to quantify epistemic uncertainty over directed acyclic graphs (DAGs) through posterior inference.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
