Maimonides' Rule and its Redux: the class-size RD replication failure
Angrist, J., & Lavy, V. (1999); Angrist, J., Lavy, V., Leder-Luis, J., & Shany, A. (2019) · 2019
grade Breplicationindependentfailednumbers spot-checked
Sample
1991 data vs 2002-11: 243,213 students / 8,944 classes
Population
Israeli 4th-5th graders; 40-pupil cap RD.
Design
Same rule, same country, same estimator, two decades apart.
Key findings
Class-size effects are context-dependent, not a constant of nature: the founding RD's large effect became a precise zero in the same country two decades later.
Genetic confound
Enrollment-cutoff variation orthogonal to family background; sorting detected but shown innocuous.
Replication notes
One of the most consequential replication failures in education economics. (Note: units differ across eras — class-level vs student-level SD — so the figures aren't directly comparable, but the zero is unambiguous.)
DOI / URL
10.1257/aeri.20180120
Effects
| Outcome | Metric | Value | Measure | Timing | Vs | Horizon | Class |
|---|---|---|---|---|---|---|---|
| 1991: 5th-grade reading | ES | ~0.25 SD (class-level) per 8-student reduction | standardized | end of year | business-as-usual | end-of-treatment | domain-skill |
| 2002-11: same design | ES | PRECISELY ESTIMATED ZEROS (SE ~0.002 SD/student; excludes effects 1/10th the 1991 size) | standardized | end of year | business-as-usual | end-of-treatment | domain-skill |
Cited by
- Class-size reductionmixedconf: highgc: low