Research map / Learning theory and algorithms / Sample complexity
Active Treatment Effect Estimation via Limited Samples
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 RWAS, an iterative re-weighting active sampling estimator for ATE with finite-sample guarantees (sample complexity O(d/ε + m')), extended to network interference, and compares against HT, Hajek, CRA, GSW, RAHT, SC, and 4-Vectors on synthetic and real datasets.
Problem
Estimating average treatment effects accurately with a limited sample budget and non-asymptotic guarantees, including under network interference where SUTVA is violated.
Approach
Adapts the iterative re-weighting active regression procedure of Chen & Price (2019) to a design-based finite population setting (IRD), uses its weights to regression-adjust a GSW-based Horvitz-Thompson estimator with sample splitting, and claims lower sample complexity than leverage-score sampling.
Key results
- On synthetic ATE data with n=1000 and 80% proportion, RWAS reports error 0.40 (sd 0.26), versus 0.40 for GSW and 0.86 for HT.
- On IHDP at 20% sampling, RWAS reports 0.07 (sd 0.06) versus 0.06 for CRA, which is comparable.
- Paper notes GSW and CRA sometimes match or beat RWAS on Lalonde and Twins, so RWAS is not uniformly best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RWAS instead of the strongest listed baseline | ATE estimation · Synthetic (n=1000, prop. 0.2) | ATE estimation error: 1.81 -> 1.84 | Mixed |
| Use RWAS instead of the strongest listed baseline | ATE estimation · Synthetic (n=1000, prop. 0.5) | ATE estimation error: 0.80 -> 0.71 | Helped |
| Use RWAS instead of the strongest listed baseline | ATE estimation · Synthetic (n=2000, prop. 0.2) | ATE estimation error: 1.43 -> 1.29 | Helped |
| Use RWAS instead of the strongest listed baseline | ATE estimation · Synthetic (n=2000, prop. 0.5) | ATE estimation error: 0.58 -> 0.59 | Mixed |
| Use RWAS instead of CRA | ATE estimation · Boston (prop. 0.2) | ATE estimation error: 0.80 -> 0.99 | Hurt |
| Use RWAS instead of CRA | ATE estimation · Boston (prop. 0.5) | ATE estimation error: 0.68 -> 0.65 | Helped |
| Use RWAS instead of CRA | ATE estimation · IHDP (prop. 0.8) | ATE estimation error: 0.02 -> 0.05 | Hurt |
| Use RWAS instead of CRA | ATE estimation · Lalonde (prop. 0.5) | ATE estimation error: 1.15 -> 1.13 | Helped |
| Use RWAS instead of CRA | ATE estimation · Twins (prop. 0.5) | ATE estimation error: 1.35 -> 1.21 | Helped |
| Use IRD reweighting active sampling rather than leverage-score sampling | ATE estimation sample complexity · Theoretical sample-complexity comparison | sample complexity: O(dlog(d)/ε2 + m′) -> O(d/ε + m′) | Helped |
Surprising findings
- CRA performs comparably to or better than RWAS on several real datasets, including IHDP and Boston.
- GSW outperforms RWAS on some datasets such as Lalonde and Twins.
Limitations
- Under network interference the estimator is not guaranteed unbiased, leading to a bias-variance trade-off.
- Lower bound is shown only up to constant factors and relies on a construction; optimality is nearly rather than exactly tight.
- CRA and GSW match or outperform RWAS in several real-world settings, so the advantage is not uniform.
Experimental setup
Datasets: Boston Housing IHDP Twins LaLonde
Benchmarks: synthetic ATE data Boston IHDP Lalonde Twins
Metrics: relative absolute error standard deviation of estimation error
Baselines: Horvitz-Thompson estimator Hajek estimator Classic Regression Adjustment (CRA) GSW design RAHT SC estimator 4-Vectors
Builds on: Chen & Price (2019) Harshaw et al. (2024) GSW design Ghadiri et al. (2024) RAHT Addanki et al. (2022) Kandiros et al. (2024)
Closest papers by meaning
- Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect (ICML 2025 · similarity 0.88)
- Off-policy estimation with adaptively collected data: the power of online learning (NeurIPS 2024 · similarity 0.87)
- Higher-Order Causal Message Passing for Experimentation with Complex Interference (NeurIPS 2024 · similarity 0.86)
- Online Experimental Design With Estimation-Regret Trade-off Under Network Interference (NeurIPS 2025 · similarity 0.86)
- Optimal Experiments for Partial Causal Effect Identification (NeurIPS 2026 · similarity 0.86)
- It’s Hard to Be Normal: The Impact of Noise on Structure-agnostic Estimation (NeurIPS 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.