Research map / Probabilistic and scientific modeling / Causal discovery
Causal Discovery via Conditional Independence Testing with Proxy Variables
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 a discretization-based, parametric-free proxy hypothesis test for causal relations between continuous variables under latent confounding/mediation, with asymptotic level and power guarantees.
Problem
Conditional independence testing for causal discovery is biased when unobserved confounders or mediators exist; prior proxy-based tests were limited to discrete variables or needed strong parametric assumptions.
Approach
Discretize X, W (and implicitly U) under completeness conditions so that the probability matrices satisfy a linear equation with identifiable coefficients under the causal null; test via squared least-squares residual, with bin count growing with sample size.
Key results
- Type I error of the proposed test approximates the nominal 0.05 level, while KCI, ICM, and Miao's procedure show considerably higher type I error.
- Type II error of the proposed method approaches 0 as sample size grows (n up to 1200).
- On the sepsis MIMIC-III data, the proposed method matches both RCT findings (Vancomycin->WBC yes, Morphine->WBC no), whereas KCI and ICM falsely detect morphine effect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed proxy-based test instead of KCI | Causal discovery with blood pressure as proxy · Sepsis: Morphine→WBC | agreement with RCT causal relation | Helped |
| Use the proposed proxy-based test instead of ICM | Causal discovery with blood pressure as proxy · Sepsis: Morphine→WBC | agreement with RCT causal relation | Helped |
| Use the proposed test with careful discretization instead of Miao | Proxy-based causal discovery · Sepsis: Vancomycin→WBC | agreement with RCT causal relation | Helped |
| Increase bin number and sample size (ablation) | Proposed test · Synthetic data, confounding and mediation settings | type I error rate | Helped |
| Increase bin number and sample size (ablation) | Proposed test · Synthetic data, confounding and mediation settings | type II error rate | Helped |
| Increase bin number at small sample sizes (ablation) | Proposed test · Synthetic data (e.g., n=200) | type II error rate | Hurt |
| Use a smooth model satisfying the TV smoothness condition rather than a nonsmooth model (ablation) | Discretization procedure · Synthetic data, confounding and mediation settings | discretization error | Helped |
Surprising findings
- Miao's discrete procedure fails to detect the vancomycin effect on WBC, attributed to coarse discretization.
- KCI conditioning on the proxy does not remove confounding bias, leading to false positives.
Limitations
- Discretization may lose power when the conditioning dimension is high (curse of dimensionality).
- Power result requires a local alternative restriction (integral equation has no solution), not guaranteed in general.
- Theory is bivariate and requires completeness and smoothness assumptions.
Experimental setup
Datasets: MIMIC-III (sepsis ICU cohort, 3,251 patients) synthetic ANM data
Benchmarks: synthetic confounding and mediation graphs sepsis causal discovery vs RCT results
Metrics: type I error rate type II error rate
Baselines: KCI ICM Miao et al. procedure
Builds on: proximal causal learning Miao et al. discrete proxy test KCI integrated conditional moment test
Closest papers by meaning
- On Causal Discovery in the Presence of Deterministic Relations (NeurIPS 2024 · similarity 0.90)
- Latent Variable Causal Discovery under Selection Bias (ICML 2025 · similarity 0.90)
- Bivariate Causal Discovery with Proxy Variables: Integral Solving and Beyond (ICML 2025 · similarity 0.89)
- Causal Discovery with Fewer Conditional Independence Tests (ICML 2024 · similarity 0.89)
- Testing Causal Models with Hidden Variables in Polynomial Delay via Conditional Independencies (AAAI 2025 · similarity 0.89)
- PAIR-CI: Calibrated Conditional Independence Testing for Causal Discovery with Incomplete Data (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.