Research map / Probabilistic and scientific modeling / Treatment effects
Inverse-Variance Weighting for Estimation of Heterogeneous Treatment Effects
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.
Argues that the choice of weights, not the pseudo-outcome transformation, drives performance in weighted pseudo-outcome regression for CATE estimation, showing R-Learning uses inverse-variance weights and deriving oracle-efficient convergence rates for IVW-based DR-learners.
Problem
CATE estimation via pseudo-outcome regression (POR) is unstable when propensity scores are extreme, and prior comparisons focused on pseudo-outcome transformations rather than the regression weights.
Approach
Shows that R-Learning is an implicit weighted POR with inverse-variance weights (approximately (A-pi)^2 or pi(1-pi)), which stabilize the pseudo-outcome regression without reintroducing confounding since the CATE is conditional on X; proposes applying such weights to DR and covariance-based pseudo-outcomes and derives convergence rates for the resulting local polynomial estimator.
Key results
- Weighted POR matched or outperformed unweighted POR in every simulation setting across 600 iterations, with sample sizes 250 to 1000.
- The T-Learner performed comparably to weighted POR in settings D, E and F but dramatically underperformed in settings A, B and C.
- Derives that the IVW local polynomial DR estimator attains oracle-efficiency when sπ + sµ > c_oracle, and is the first shown oracle-efficient without knowing the covariate distribution for sπ + sµ < c_oracle.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use R-Learner weights (A − π̂(X))² instead of uniform weights (ablation) | Spline-based pseudo-outcome regression predicting fU,θ · Simulation F | rMSE: 1.38 -> 0.288 | Helped |
| Use inverse-variance weights instead of uniform weights (ablation) | Pseudo-outcome regression across simulation settings A–F · Simulations A–F | mean squared error | Helped |
| Use a T-Learner instead of weighted pseudo-outcome regression | CATE estimation · Simulations A–C | mean squared error | Hurt |
| Use a T-Learner instead of weighted pseudo-outcome regression | CATE estimation · Simulations D–F | mean squared error | No clear effect |
Surprising findings
- The T-Learner performed comparably to weighted POR in several settings, despite the weighting argument.
Limitations
- Most powerful results require higher-order (nested) cross-fitting, which is impractical at finite samples.
- Weighting changes the estimand when the target is an aggregated effect over a subset of covariates V that is a strict subset of X.
Experimental setup
Datasets: simulated data (settings A-F from Nie & Wager 2020 and Kennedy 2023)
Benchmarks: simulation settings A-F
Metrics: mean squared error (log scale) rMSE convergence rate
Baselines: U-Learner T-Learner R-Learner unweighted DR-Learner HORL
Builds on: R-Learner DR-Learner U-Learner T-Learner higher-order R-Learner
Closest papers by meaning
- Treatment Effects Estimation By Uniform Transformer (ICLR 2024 · similarity 0.88)
- Collaborative Heterogeneous Causal Inference Beyond Meta-analysis (ICML 2024 · similarity 0.87)
- Estimating Heterogeneous Treatment Effects by Combining Weak Instruments and Observational Data (NeurIPS 2024 · similarity 0.87)
- A Non-parametric Direct Learning Approach to Heterogeneous Treatment Effect Estimation under Unmeasured Confounding (NeurIPS 2024 · similarity 0.86)
- A Meta-learner for Heterogeneous Effects in Difference-in-Differences (ICML 2025 · similarity 0.86)
- Model-agnostic meta-learners for estimating heterogeneous treatment effects over time (ICLR 2025 · similarity 0.86)
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.