Research map / Probabilistic and scientific modeling / Causal discovery

Identifying Structural Biases from Causal Mechanism Shifts

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Identifiability with hidden confounders · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes StruBI, which uses dependent mechanism shifts across multi-context data (tested with KCI and adjusted mutual information over partitions) plus ancestral coverage to classify variables as unbiased, hidden-confounded, or selection-biased.

Problem

Causal discovery usually assumes i.i.d. data and causal sufficiency, so hidden confounding and selection bias produce spurious dependencies; the paper asks whether these biases can be detected and told apart from multi-context data.

Approach

Intrinsic mechanism shifts of a latent confounder appear as observed shifts in its children, whereas a latent selection collider's shift propagates to all its observed ancestors. Non-parametric KCI tests per mechanism and context pair give partitions; AMI between partitions finds jointly shifting sets, and ancestral coverage separates selection (coverage near 1) from confounding (coverage below 1)

Key results

• StruBI given true partitions classifies structural bias perfectly on synthetic data, and StruBI with inferred graph performs essentially as well as with the true graph.
• StruBI outperforms all five baselines (FCI-P, JCI-FCI, CoCo, LS-P, LS-C) by a wide margin on latent discovery, subset recovery, and bias classification on synthetic data.
• On the Sachs case study, StruBI correctly classifies confounding and selection bias for PKC, retrieving three confounded edges and six of eight ancestors under selection.

Surprising findings

• CoCo was the only baseline that improved in subset recovery as DAG size increased.
• The paper reports that with higher observed-node shift ratio r, confounding can be misclassified as selection bias due to spurious backdoor effects.

Limitations

• Requires Markov and faithfulness assumptions, so cyclic systems are not covered.
• Relies on sparse, independent mechanism shifts and on a sufficient number of samples per context and number of contexts.
• Depends on a correct observed causal graph; learned graphs may introduce errors.

Experimental setup

Datasets: Sachs flow cytometry dataset synthetic Erdos-Renyi SEM data

Benchmarks: synthetic Erdos-Renyi DAGs (3-13 observed variables) Sachs flow cytometry

Metrics: macro-F1 accuracy precision recall

Baselines: FCI-P JCI-FCI CoCo LS-P LS-C

Builds on: JCI CoCo FCI Latent-Selection KCI test

Closest papers by meaning

• Latent Variable Causal Discovery under Selection Bias (ICML 2025 · similarity 0.90)
• Detecting and Measuring Confounding Using Causal Mechanism Shifts (NeurIPS 2024 · similarity 0.89)
• On Causal Discovery in the Presence of Deterministic Relations (NeurIPS 2024 · similarity 0.89)
• CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.88)
• Causal learning with the invariance principle (NeurIPS 2026 · similarity 0.88)
• Bivariate Causal Discovery using Bayesian Model Selection (ICML 2024 · similarity 0.88)