More Data, More Confusion: New Study Shows Bayesian Causal Discovery Can Break Down as Sample Size Grows

A team behind a new arXiv paper has pinpointed exactly when and why Bayesian causal discovery algorithms get fooled by hidden variables, and the answer runs counter to a common assumption in machine learning: that more data reliably means better inference.

The paper, "How Does Bayesian Causal Discovery Fail? Characterising Structural Consequences in Linear Gaussian Networks under Latent Confounding," was submitted to arXiv on July 10, 2026. It tackles a persistent weak spot in causal inference: what happens when the algorithm doesn't know that two observed variables are both being influenced by a third, unmeasured factor. This scenario, known as latent confounding, is common in real-world data where researchers can't measure every relevant variable, from genomics to economics to observational studies of any kind.

A Threshold That Moves in the Wrong Direction

Working within linear Gaussian causal models, a standard and mathematically tractable setting for this kind of analysis, the authors derive a critical correlation threshold. Once the correlation between two confounded variables crosses this line, the scoring function used in Bayesian causal discovery starts favoring graphs that include a spurious, non-causal edge directly connecting them. In other words, the model mistakes a shared hidden cause for a direct causal link.

What makes the finding notable is the direction the threshold moves as sample size increases. Rather than becoming more forgiving of noisy correlations, the threshold actually drops. That means less correlation is needed to trigger the spurious edge as more observations are added, so scaling up a dataset doesn't wash out the confounding artifact. It can make the model more confident in the wrong structure.

The authors back this up with exact posterior computations run across multiple graph configurations, rather than relying purely on asymptotic approximations or simulation. This lets them characterize the failure mode with mathematical precision instead of just demonstrating it empirically.

Two Regimes, Not One

The paper also draws a distinction between two separate posterior failure regimes, each shaped by how the confounded variables sit within the broader graph structure. This local-structure dependence suggests there isn't a single, uniform way Bayesian causal discovery breaks down under confounding: the failure pattern depends on the topology surrounding the affected variables, not just the strength of their correlation.

For practitioners using causal discovery tools in applied settings, the implications are direct. Latent confounding is rarely something you can rule out with certainty, and this work suggests that simply gathering more data isn't a safe hedge against it. Understanding the specific graph structures around suspected confounders may matter as much as sample size when deciding how much to trust an inferred causal edge.