Research map / Probabilistic and scientific modeling / Treatment effects
Comparing Targeting Strategies for Maximizing Social Welfare with Limited Resources
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.
Empirically compares risk-based targeting (prioritizing high baseline risk) with treatment-effect-based targeting across five real RCTs, under simulated confounding and several social welfare functions.
Problem
Policymakers allocating limited interventions often target by predicted baseline risk rather than heterogeneous treatment effects, yet there is little empirical evidence on which targeting objective maximizes social welfare, especially under confounding and inequality-averse preferences.
Approach
Uses five real RCT datasets to compare risk-based and treatment-effect-based targeting, simulating accurate CATE estimates via partially observed pseudo-outcomes and simulating confounding by systematically removing data, evaluated under utilitarian, weighted utilitarian, and Nash social welfare functions.
Key results
- On STAR, treatment-effect targeting significantly outperforms risk-based targeting with no confounding, but the advantage shrinks toward parity as confounding increases.
- Across most datasets, targeting on biased treatment effects remains at least as good as risk-based targeting under utilitarian welfare, even at relatively severe confounding.
- When risk-based targeting wins under weighted utilitarian welfare, the required alpha is almost always at least 3 (about 4.5x weight ratio), i.e., extreme welfare weights.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use treatment-effect targeting instead of risk-based targeting when treatment effects are predicted from features | Targeting policies in the five RCT datasets · STAR | utilitarian welfare | Helped |
| Use treatment-effect targeting instead of risk-based targeting when treatment effects are predicted from features | Targeting policies in the five RCT datasets · NSW | utilitarian welfare | No clear effect |
| Use treatment-effect targeting instead of risk-based targeting when treatment effects are predicted from features | Targeting policies in the five RCT datasets · Ultra Poor, Postoperative Pain, and Acupuncture | utilitarian welfare | No clear effect |
| Use risk-based targeting instead of random allocation | Targeting policies in the five RCT datasets · Five RCT datasets | welfare | No clear effect |
| Rank individuals by observed pseudo-outcome treatment effects rather than predicted effects from features | Treatment effect targeting · Five datasets, no confounding | utilitarian welfare | Helped |
| Introduce confounding by systematically removing data while estimating treatment effects | Treatment effect targeting using observed pseudo-outcomes · Five datasets | utilitarian welfare | Hurt |
| Use biased treatment-effect estimates instead of risk scores | Treatment effect targeting using observed pseudo-outcomes · Five datasets | utilitarian welfare | Helped |
| Give higher baseline-risk individuals greater welfare weight while comparing risk-based and treatment-effect targeting | Targeting policies using observed pseudo-outcomes · Five datasets | weighted utilitarian welfare | Helped |
| Use treatment-effect targeting rather than risk-based targeting under Nash welfare | Targeting policies using observed pseudo-outcomes · Ultra Poor at high confounding | Nash social welfare | Hurt |
| Use treatment-effect targeting rather than risk-based targeting under Nash welfare | Targeting policies using observed pseudo-outcomes · NSW, Postoperative Pain, Acupuncture, and STAR | Nash social welfare | Helped |
| Use a feature-predicted treatment-effect score rather than observed pseudo-outcome values to rank individuals | Treatment effect targeting · STAR, no confounding | utilitarian welfare | Hurt |
| Increase the proportion of systematically removed data (confounding) | Treatment effect targeting compared with risk-based targeting · Five datasets | relative welfare benefit | Mixed |
Surprising findings
- Risk-based targeting often fails to outperform random allocation strongly; baseline risk is only weakly and noisily related to treatment effects.
- Learned treatment effect models on real RCT data often do not beat risk-based targeting, so the bottleneck is data/feature reliability rather than confounding.
Limitations
- Assumes a consequentialist policymaker; non-consequentialist preferences for vulnerability are not modeled.
- Treatment-effect targeting with observed pseudo-outcomes is not attainable in practice; the high-accuracy setting is semi-synthetic.
- Several datasets are small (236 to 796 examples), leading to wide confidence intervals.
Experimental setup
Datasets: Targeting the Ultra Poor (TUP) India RCT NSW job training RCT Postoperative pain licorice RCT Acupuncture chronic headache RCT Tennessee STAR class size RCT
Benchmarks: TUP NSW STAR Postoperative Pain Acupuncture
Metrics: treatment-on-the-treated welfare estimate utilitarian welfare weighted utilitarian welfare Nash social welfare minimum alpha at which risk-based targeting outperforms
Baselines: risk-based targeting random allocation
Builds on: doubly-robust estimator random forest rank-weighted average treatment effect
Closest papers by meaning
- Evaluating Index-based Treatment Allocation in Underresourced Communities (AAAI 2025 · similarity 0.86)
- On the Actionability of Outcome Prediction (AAAI 2024 · similarity 0.86)
- PairNet: Training with Observed Pairs to Estimate Individual Treatment Effect (ICML 2024 · similarity 0.86)
- Benchmarking Estimators for Natural Experiments: A Novel Dataset and a Doubly Robust Algorithm (NeurIPS 2024 · similarity 0.85)
- On Efficient Estimation of Distributional Treatment Effects under Covariate-Adaptive Randomization (ICML 2025 · similarity 0.85)
- AutoCATE: End-to-End, Automated Treatment Effect Estimation (ICML 2025 · similarity 0.85)
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.