Measuring the Impacts of Teachers I: Evaluating Bias in Teacher Value-Added Estimates
Chetty R, Friedman JN, Rockoff JE · 2014
grade Bnatural-experimentindependentmixed
Sample
~1,000,000 students, grades 3-8, large urban school district
Population
US grades 3-8 in one large urban district, linked to administrative records.
Design
Validates prior-score-controlled value-added with a teacher-switching quasi-experiment: when a high-VA teacher moves into a school, the cohort's scores move by the predicted amount. The forecast coefficient is ~1, i.e. VA estimates are close to unbiased for the score outcome they are built on.
Key findings
Teacher value-added, estimated with prior-score controls, forecasts actual score changes roughly one-for-one. This is the paper that made VA defensible as a measurement rather than a ranking.
Genetic confound
Dynamic sorting of students on partly heritable ability is exactly the threat this design is built to bound; the switching variation is at the school-cohort level.
Replication notes
Agrees with Kane & Staiger's random-assignment validation. Contested by Rothstein 2017 on placebo-test grounds; see that record and the authors' reply.
Effects
| Outcome | Metric | Value | Measure | Timing | Vs | Horizon | Class |
|---|---|---|---|---|---|---|---|
| Forecast bias in prior-score-controlled value-added | forecast coefficient | ~1:1 — teacher-switching changes cohort scores by approximately the amount VA predicts | standardized | same year | none | end-of-treatment | domain-skill |
Cited by
- Teacher quality — selection over credentials and workshopsstrong supportconf: highgc: low