Research map / Probabilistic and scientific modeling / Treatment effects
Proximal Causal Inference With Text Data
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 proximal causal inference with two zero-shot LLM-inferred text proxies from distinct pre-treatment notes, plus an odds-ratio falsification heuristic, evaluated on synthetic and MIMIC-III semi-synthetic data.
Problem
Causal effect estimation from observational data when an important confounder is completely unobserved (no labels), so supervised text proxies are unavailable; proximal causal inference needs two valid proxies that are hard to find among structured variables.
Approach
Infer two proxies for the unmeasured binary confounder by applying two different zero-shot LLM classifiers to two distinct pre-treatment text instances, so proxies satisfy proximal conditions (P1-P4) by design; add an odds-ratio heuristic on observed data to flag violations before estimation.
Key results
- In fully synthetic data with true ACE 1.3, P1M passing the heuristic gives bias 0.004 with CI coverage, while using the same instance of text for both proxies gives bias 0.130 and fails coverage.
- In semi-synthetic MIMIC-III experiments, P2M (Flan and OLMo) passed the odds ratio heuristic for two settings (A-Sis with Echo and Radiology; Heart with Echo and Nursing) and yielded low bias, while failing settings were often biased.
- Using a single inferred LLM proxy directly in backdoor adjustment produced large bias across all semi-synthetic settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the same text instance to infer both proxies (ablation) | P1M · Fully synthetic | bias: 0.004 -> 0.130 | Hurt |
| Use the same text instance to infer both proxies (ablation) | P2M · Fully synthetic | bias: 0.043 -> 0.107 | Hurt |
| Use two distinct text instances rather than the same instance for both proxies (ablation) | P1M · Fully synthetic | confidence interval coverage: No -> Yes | Helped |
| Use two distinct text instances rather than the same instance for both proxies (ablation) | P2M · Fully synthetic | confidence interval coverage: No -> Yes | Helped |
| Use a second zero-shot model instead of using FLAN-T5 XXL for both proxies (ablation) | Proximal inference with text proxies · MIMIC-III semi-synthetic; three settings | bias | Helped |
| Use an inferred proxy directly in backdoor adjustment instead of the proximal g-formula | Text-based causal effect estimation · MIMIC-III semi-synthetic; all settings of U | bias | Hurt |
| Use two distinct pre-treatment text instances and two different zero-shot models | P2M · A-Sis; Echo and Radiology notes | odds ratio CI heuristic pass: (2.381, 2.962) -> (1.372, 1.995) | Helped |
| Use two distinct pre-treatment text instances and two different zero-shot models | P2M · Heart; Echo and Nursing notes | odds ratio CI heuristic pass: (2.298, 2.676) -> (1.152, 1.376) | Helped |
| Use two distinct pre-treatment text instances and two different zero-shot models | P2M · A-Sis; Radiology and Nursing notes | odds ratio CI heuristic pass: (4.337, 5.266) -> (2.050, 2.663) | No clear effect |
Surprising findings
- With real clinical notes, the single-model P1M had higher bias than the two-model P2M across all three settings, which the authors attribute to Flan over-relying on pretraining.
- A setting failed the heuristic even without text dependence because proxy accuracies (0.49 and 0.52) were below majority-class accuracy.
Limitations
- Odds ratio heuristic requires domain-specified gamma_high threshold; misspecification can yield false positives.
- Requires pre-treatment text and metadata to split into conditionally independent instances, which may not be available.
- Linear two-stage estimator and binary U restrict generality; experiments use linear proximal estimator.
Experimental setup
Datasets: MIMIC-III fully synthetic DGP
Benchmarks: fully synthetic DGP MIMIC-III semi-synthetic DGP
Metrics: bias confidence interval coverage odds ratio gamma_WZ.C accuracy of proxies
Baselines: P1M (single zero-shot model) P1M same instance P2M same instance backdoor adjustment with inferred proxy W (Gotcha 1)
Builds on: proximal causal inference proximal g-formula two-stage regression estimator for proximal inference odds ratio sensitivity analysis
Closest papers by meaning
- A Proxy Variable View of Shared Confounding (ICML 2021 · similarity 0.91)
- End-To-End Causal Effect Estimation from Unstructured Natural Language Data (NeurIPS 2024 · similarity 0.90)
- Automating the Selection of Proxy Variables of Unmeasured Confounders (ICML 2024 · similarity 0.89)
- Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · similarity 0.88)
- Detecting and Measuring Confounding Using Causal Mechanism Shifts (NeurIPS 2024 · similarity 0.87)
- Doubly robust identification of treatment effects from multiple environments (ICLR 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.