Project STAR and Krueger's reanalysis (Experimental Estimates of Education Production Functions)
Krueger, A. B. (Tennessee STAR RCT, 1985-89) · 1999
grade Arctindependentreplicatednumbers spot-checked
Sample
11,571 students, 79 Tennessee schools, randomized K-3 to ~15 vs ~22
Population
Tennessee K-3 students.
Design
The only large US class-size RCT. Hanushek's audit (attrition correlated with treatment, ~10% crossover, no baseline test, unblinded teachers; first-year-only pattern suggesting a level effect) is contested: Krueger's IV on initial random assignment addresses crossovers, his attrition reanalysis leaves effects intact, and Chetty re-verified randomization balance.
Key findings
The anchor: small K-3 classes buy ~0.2 SD in the first exposure year (larger for disadvantaged students), then little more. Real, front-loaded, and expensive.
Genetic confound
Within-school randomization removes ability selection; residual threats are implementation, not genetics.
Replication notes
Numbers verified. The Hanushek-Krueger dispute also covers the 277-estimate vote count: counting ESTIMATES gives 'no effect'; reweighting by STUDY flips it positive — the old consensus was a weighting choice.
DOI / URL
Effects
| Outcome | Metric | Value | Measure | Timing | Vs | Horizon | Class |
|---|---|---|---|---|---|---|---|
| Achievement, small vs regular class (first exposure year) | percentiles | +4 (~0.19-0.20 SD); ~+1 percentile/yr thereafter | standardized | K-3 | business-as-usual | end-of-treatment | domain-skill |
| Subgroups | pattern | roughly 2x larger for Black and free-lunch students | standardized | K-3 | business-as-usual | end-of-treatment | domain-skill |
| Teacher-aide arm | ES | ~0 | standardized | K-3 | business-as-usual | end-of-treatment | domain-skill |
| Cost-benefit | IRR | ~6% (Krueger 2003) for 22->15 reduction (~47% more teacher cost) | mixed | lifetime projection | business-as-usual | adulthood | attainment |
Cited by
- Class-size reductionmixedconf: highgc: low