Research map / Probabilistic and scientific modeling / Treatment effects
Local Learning for Covariate Selection in Nonparametric Causal Effect Estimation with Latent 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 LSAS, a fully local, sound and complete algorithm that finds valid covariate adjustment sets for a treatment-outcome pair using Markov blanket discovery and conditional independence rules, allowing latent variables.
Problem
Selecting a valid covariate adjustment set for estimating the causal effect of a treatment on an outcome from observational data with latent variables, without learning the full global causal graph.
Approach
Locally searches adjustment sets inside the Markov blanket of the outcome (in a MAG) using two testable independence rules (R1 finds a valid adjustment set and effect; R2 detects zero effect), avoiding global structure learning and exhaustive search over all covariates.
Key results
- On the Cattaneo2 real-world dataset, LSAS's estimated effect falls within the 200-250 g baseline interval using 158 conditional independence tests, versus 1284 for CEELS and 266 for LDP.
- LSAS outperforms compared methods on relative error and nTest across specific structures and benchmark networks at all sample sizes (1K-15K), per the paper's claims.
- LSAS needs far fewer independence tests than LV-IDA and EHS, which require global structure learning or exhaustive search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LSAS instead of LDP | Cattaneo2 · Cattaneo2 | nTest: 266 -> 158 | Helped |
| Use LSAS instead of CEELS | Cattaneo2 · Cattaneo2 | nTest: 1284 -> 158 | Helped |
| Use LSAS for estimating the effect of maternal smoking on birth weight | Cattaneo2 · Cattaneo2 | causal effect agreement with baseline interval | Helped |
Surprising findings
- The paper argues that if no subset of MB(Y)\{X} is an adjustment set, then no subset of the full observed covariates is either.
Limitations
- Assumes the pretreatment condition (no descendants of treatment among covariates); the method may select invalid sets when violated.
- Cannot determine causal effect when neither rule R1 nor R2 applies (scenario S3).
- Worst-case exponential complexity in the size of MB(Y).
Experimental setup
Datasets: Cattaneo2 INSURANCE MILDEW WIN95PTS ANDES synthetic linear Gaussian data from specific structures and random graphs
Benchmarks: Cattaneo2 INSURANCE MILDEW WIN95PTS ANDES specific M-structure graphs
Metrics: relative error number of conditional independence tests
Baselines: LV-IDA with RFCI EHS CEELS LDP
Builds on: generalized adjustment criterion EHS CEELS LV-IDA TC (total conditioning) Markov blanket discovery
Closest papers by meaning
- Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation (NeurIPS 2026 · similarity 0.91)
- Latent Variable Causal Discovery under Selection Bias (ICML 2025 · similarity 0.89)
- Score-Based Causal Discovery of Latent Variable Causal Models (ICML 2024 · similarity 0.88)
- Local Identifying Causal Relations in the Presence of Latent Variables (ICML 2025 · similarity 0.88)
- Local Causal Structure Learning in the Presence of Latent Variables (ICML 2024 · similarity 0.88)
- PCM Selector: Penalized Covariate-Mediator Selection Operator for Evaluating Linear Causal Effects (AAAI 2025 · similarity 0.87)
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.