Research map / Probabilistic and scientific modeling / Treatment effects
Quantifying Treatment Effects: Estimating Risk Ratios via Observational Studies
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 and analyzes estimators of the risk ratio (RR) in observational studies (IPW, G-formula, one-step, and AIPW-based RR-AIPW), proving asymptotic normality and CIs, and recommends the doubly-robust RR-AIPW.
Problem
Estimating the causal risk ratio (relative treatment effect) from observational data, where no non-parametric or doubly-robust estimators with theoretical guarantees existed, and confounding must be handled.
Approach
Adapts IPW, G-formula, and AIPW-style efficient influence-function estimators to the nonlinear risk ratio: the RR-AIPW is a ratio of two AIPW estimates of E[Y(1)] and E[Y(0)], with asymptotic normality and CIs derived via M-estimation and influence functions, and shown to be efficient under weaker conditions than the one-step RR-OS.
Key results
- RR-AIPW and RR-OS (linear nuisance) are unbiased even at small sample sizes and converge quickly to the true RR in the linear/logistic DGP.
- In the linear/logistic DGP with n=1000, only the Linear RR-AIPW achieves coverage above 95% among the compared CIs, while RR-N has nearly zero coverage in observational settings.
- On the Traumabase real-world data, most estimators give RR above one (suggesting harmful effect of TXA on mortality), with Forest RR-IPW the only one indicating benefit.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use linear nuisance models for RR-AIPW instead of forest-based nuisance models | RR-AIPW · Linear/Logistic DGP | bias | Helped |
| Use linear nuisance models for RR-OS instead of forest-based nuisance models | RR-OS · Linear/Logistic DGP | bias | Helped |
| Use linear nuisance models for RR-G instead of forest-based nuisance models | RR-G · Linear/Logistic DGP | bias | Helped |
| Use random forests to estimate nuisance functions rather than parametric models | RR estimators · Linear/Logistic DGP | convergence rate | Hurt |
| Use linear outcome models for RR-G when the baseline response is nonlinear | RR-G · Non-Linear/Logistic DGP | bias | Hurt |
| Use forest-based nuisance models for RR-AIPW rather than linear models | RR-AIPW · Non-Linear/Logistic DGP | bias | Hurt |
| Use linear nuisance models for RR-AIPW when at least one model is well specified | RR-AIPW · Non-Linear/Logistic DGP | bias | Helped |
| Use linear nuisance models for RR-OS when at least one model is well specified | RR-OS · Non-Linear/Logistic DGP | bias | Helped |
| Use RR-N confidence intervals in observational data | RR-N · Linear/Logistic DGP | coverage | Hurt |
| Use linear rather than forest-based nuisance models for RR-AIPW confidence intervals | RR-AIPW · Linear/Logistic DGP | coverage | Helped |
| Use forest-based nuisance models for RR-AIPW confidence intervals | RR-AIPW · Linear/Logistic DGP | coverage | Hurt |
| Use linear nuisance models for RR-AIPW confidence intervals | RR-AIPW · Non-Linear/Logistic DGP | coverage | Helped |
Surprising findings
- RR-AIPW, a simple ratio of two AIPW estimators, is efficient and requires weaker assumptions than the one-step RR-OS, challenging expectations about the ratio of two one-step estimators.
- Estimating the propensity score reduces the asymptotic variance of RR-IPW compared with the oracle propensity score.
Limitations
- Relies on standard causal assumptions (unconfoundedness, overlap) that are hard to verify in practice.
- Non-parametric nuisance estimation (random forests) converges slowly and yields poor coverage for n=1000.
- Estimators for the non-RCT and generalization settings remain open; guidelines for choosing linear vs non-parametric nuisance models are lacking.
Experimental setup
Datasets: Traumabase (8,270 TBI patients, Mayer et al. 2020) semi-synthetic and real subsamples Simulated linear/logistic and non-linear DGPs
Benchmarks: Traumabase TXA 28-day mortality (real-world subsample)
Metrics: bias variance/efficiency coverage of 95% confidence intervals CI length convergence to true risk ratio
Baselines: RR-N (Neyman risk ratio) RR-IPW RR-G RR-OS Forest RR-G Forest RR-AIPW Linear RR-AIPW
Builds on: AIPW Inverse Propensity Weighting Katz et al. 1978 risk ratio CI Kennedy 2022 influence function estimators cross-fitting (double machine learning)
Closest papers by meaning
- A Unified Framework for the Transportability of Population-Level Causal Measures (NeurIPS 2025 · similarity 0.87)
- Benchmarking Estimators for Natural Experiments: A Novel Dataset and a Doubly Robust Algorithm (NeurIPS 2024 · similarity 0.86)
- Unveiling the Potential of Robustness in Selecting Conditional Average Treatment Effect Estimators (NeurIPS 2024 · similarity 0.85)
- B-CALM: Bias-Limited Bayesian Borrowing for RCT-Anchored Treatment Effects under Covariate Mismatch (NeurIPS 2026 · similarity 0.85)
- Generalizing Causal Effects from Randomized Controlled Trials to Target Populations across Diverse Environments (ICML 2025 · similarity 0.85)
- Constructing Confidence Intervals for Average Treatment Effects from Multiple Datasets (ICLR 2025 · similarity 0.84)
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.