Research map / Probabilistic and scientific modeling / Treatment effects
Allocation Requires Prediction Only if Inequality Is Low
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.
Proves, in a stylized model of units (hospitals, schools) with individual welfare, that unit-level allocation dominates individual prediction-based allocation when between-unit inequality is high or the budget is low, and that prediction is only worthwhile under low inequality.
Problem
Whether individual-level welfare prediction is necessary for efficient allocation of interventions, compared with allocating by aggregate unit-level statistics, given the cost of prediction and between-unit inequality.
Approach
Introduces a dominance notion over unit profiles and a stylized model comparing individual-level allocation (ILA, with costly prediction) against unit-level allocation (ULA, with within-unit delegation), giving sufficient and necessary conditions in terms of the Gini index of unit-level statistics, budget, and price of prediction.
Key results
- ULA dominates ILA under high between-unit inequality (normalized Gini above a threshold) as long as the budget is not excessively large, for sufficiently costly prediction.
- Necessary conditions for ILA to dominate ULA are significantly stronger; in the Gini ~1/2 regime ULA remains nondominated even with q=0 (no within-unit screening).
- In a real-world LA school-district example with high inequality, ULA outperforms ILA; in Utah (low inequality) ILA comes out ahead.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use individual-level prediction-based allocation instead of unit-level allocation when between-unit inequality is low and the intervention budget is high | ILA vs. ULA · Theoretical allocation model | social welfare | Helped |
| Use unit-level allocation rather than individual-level prediction-based allocation in high-inequality settings when the budget is not excessively large | ULA vs. ILA · Theoretical allocation model | social welfare | Helped |
| Use unit-level allocation rather than individual-level prediction-based allocation when prediction has a non-negligible cost | ULA vs. ILA · Theoretical allocation model | social welfare | Helped |
| Use unit-level allocation rather than individual-level prediction-based allocation in the Utah school-district example | ULA vs. ILA · Utah school districts | allocation value | Hurt |
| Use unit-level allocation rather than individual-level prediction-based allocation in the Los Angeles school-district example | ULA vs. ILA · Greater Los Angeles school districts | allocation value | Helped |
| Use unit-level allocation rather than individual-level prediction-based allocation in the New York school-district example | ULA vs. ILA · New York State school districts | allocation value | Helped |
| Use a within-unit allocation that effectively excludes high-welfare individuals | ULA · Theoretical allocation model | social welfare | Helped |
| Increase between-unit inequality | ULA · Theoretical allocation model | social welfare | Helped |
| Increase the intervention budget | ULA vs. ILA · Theoretical allocation model | relative allocation value | Hurt |
| Allow heterogeneous treatment effects while assuming a Lipschitz effect curve and limited welfare concentration | ULA vs. ILA · Theoretical allocation model with heterogeneous treatment effects | social welfare | No clear effect |
| Learn unit-level statistics instead of assuming them freely available | ULA · Theoretical allocation model | allocation value | No clear effect |
Surprising findings
- Unit-level allocation can beat individual prediction even when within-unit allocation treats everyone, under high inequality.
- Costly prediction can make individual-level allocation worse than simple allocation even with perfect predictions after payment.
Limitations
- Stylized model with fixed or Lipschitz treatment effects and uniform error bounds on predictions.
- Assumes unit-level statistics (rho_k) are known or cheaply learnable; learning costs analyzed only partially.
- Real-world illustration uses income as a proxy for welfare and uses a limited set of districts.
Experimental setup
Datasets: American Community Survey Children's Education Tabulation (2017-2021)
Benchmarks: simulated beta-distributed units ACS school district household income data (Los Angeles, New York, Utah)
Metrics: allocation value (welfare gain) Gini index of unit-level statistics
Baselines: individual-level allocation (ILA)
Closest papers by meaning
- The Hidden Cost of Waiting for Accurate Predictions (ICLR 2025 · similarity 0.89)
- The Relative Value of Prediction in Algorithmic Decision Making (ICML 2024 · similarity 0.88)
- Evaluating Index-based Treatment Allocation in Underresourced Communities (AAAI 2025 · similarity 0.88)
- Position: Scarce Resource Allocations That Rely On Machine Learning Should Be Randomized (ICML 2024 · similarity 0.85)
- The Value of Prediction in Identifying the Worst-Off (ICML 2025 · similarity 0.85)
- Comparing Targeting Strategies for Maximizing Social Welfare with Limited Resources (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.