Research map / Probabilistic and scientific modeling / Treatment effects

Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation

NeurIPS 2026 · Probabilistic and scientific modeling › Treatment effects · approach: Latent-variable causal inference · 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.

Shows that the local covariate-selection rules R1/R2 (with Markov-blanket search) of an acyclic method remain sound and complete for simple cyclic SCMs via sigma-acyclification, and validates them on synthetic linear and nonlinear data.

Problem

Covariate (adjustment set) selection and causal effect estimation from observational data when latent confounding and feedback cycles are present; existing local methods assume acyclicity.

Approach

Extends the acyclic local adjustment rules (R1/R2 over Markov blankets, from LSAS/EHS) to simple cyclic SCMs by showing that sigma-separation and Markov blankets are invariant under sigma-acyclification, so the DAG guarantees transfer without modification.

Key results

• Relative error of the local adjustment method is lower in cyclic than acyclic linear SCMs, with the same qualitative trend as sample size increases.
• Cyclic graphs achieve higher edge fraction accuracy than acyclic graphs, often near-perfect at larger sample sizes.
• Acyclic graphs obtain higher precision of returned adjustment sets than cyclic graphs, while the sample size has little effect on precision.

Surprising findings

• Cyclic settings show lower relative error and higher edge recovery than acyclic settings despite the method being designed for DAGs; the authors leave the mechanism unexplained.
• Cyclic graphs have lower precision of adjustment sets than acyclic graphs.

Limitations

• Requires the pre-treatment assumption (Y has no descendants, X has no children other than Y).
• Precision of returned adjustment sets is lower in cyclic graphs and the mechanism of the trade-off is unexplained.
• No direct comparison baselines exist for cyclic settings; evaluation is purely synthetic.

Experimental setup

Datasets: synthetic linear and nonlinear SCMs with latent confounding and cycles

Benchmarks: synthetic SCMs (8 to 250 nodes)

Metrics: relative error precision edge fraction accuracy empty fraction

Baselines: acyclic case of the same method (DAG setting)

Builds on: LSAS EHS sigma-separation sigma-acyclification Fast-IAMB

Closest papers by meaning

• Local Learning for Covariate Selection in Nonparametric Causal Effect Estimation with Latent Variables (NeurIPS 2025 · similarity 0.91)
• Bivariate Causal Discovery using Bayesian Model Selection (ICML 2024 · similarity 0.89)
• Data-Driven Selection of Instrumental Variables for Additive Nonlinear, Constant Effects Models (ICML 2025 · similarity 0.88)
• Latent Variable Causal Discovery under Selection Bias (ICML 2025 · similarity 0.88)
• Continuous Bayesian Model Selection for Multivariate Causal Discovery (ICML 2025 · similarity 0.88)
• Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · similarity 0.88)