Do small classes help young children learn?
Independent rebuild
The study said first graders put in small classes scored about 8.6 percentile points higher. The Lab retraced it: ✓ same result.
Whose code is this?
Our own small program, run on a public copy of the data, which has only reading and maths scores. It isn't the author's code, and it can't match exactly. It got 8.52 against 8.57 for first grade. The other grades are 4.93 against 4.82 in kindergarten, 5.95 against 5.93 in second grade, and 5.68 against 5.32 in third, which is outside the margin we used. We chose first grade because it's the largest gain. It's a settled showcase, so there was no forecasting.
See the codeTennessee's experiment is the study people reach for whenever schools argue about class sizes.
Cited by 1,116 studies
We test one thing: does this code and data give this number? That's not the same as the science being right. The forecast price never changes the verdict.
See the proof
Counts as holding up
Where the number comes fromSmall Class 8.57 (1.97)Table 6, panel B, column 1, of the working paper (NBER 6051), read from the page imageThe exact claim as posted
Do small classes help young children learn? Tennessee's experiment found 8.57 percentile points in first grade
Krueger (Quarterly Journal of Economics, 1999) studied Tennessee's STAR experiment, which put thousands of young children into small or normal-size classes at random. First graders in small classes scored 8.57 percentile points higher, before other factors are held equal. This runs our own small program on a public copy of the data, which has reading and maths scores only, so it can't match exactly. The rebuild isn't the author's code, so a mismatch ends as inconclusive.
Do small classes help young children learn?
- Measured in
- percentile points
- Paper says
- 8.57
- Counts if
- 8.32 to 8.82
- Lab got
- 8.5152
This result is final
Paper said 8.57. We got 8.5152.
The code and data produced the same result. That doesn't prove the study's method is right.
This slip is the Lab's .
✓ Verified on NEAR. See the contract's history
- People forecast
- Forecasting ends
- The Lab retraces it
- Verdict posted, open to challenge
- Winners collect
What we didn't check
- Whether the sample was fair or large enough.
- Whether test data leaked into training or tuning.
- Whether the result holds on new data.
- Whether the study's method was the right one for its question.
- Whether the gain lasts, or what happens in other places.
- The published journal table, which isn't free to read. We used the working paper's.
No forecasting on this claim. It was posted with the result already known.
The evidence
See the full receiptHow much evidence is behind this?
Each step is a tougher test. Only the first one decides who wins the market.
- 1. ✓
The Lab's retrace of the code and data named in the claim
The Lab retraced the code and data and got a number inside the claim's range. YES wins.
- 2. –
Stress test: does the number depend on luck?
Fetching the extra runs…
- 3. –
Rebuilt from the description (experimental)
Fetching the AI's rebuild…
Coming next: a data-leakage check, for the same rows in training and test data, and a maths check on the statistics a paper reports. Neither has run on this claim.
Even all three steps can't prove the science is true. They show how well this result stands up to checking.
Verified
Your browser just re-checked these itself, so you don't have to trust this site.
Checking…
What the Lab retraces
- Source
- https://doi.org/10.1162/003355399556052
A. B. Krueger; rebuild by the RERUN project, Rebuild: MIT. Data: GPL (R package AER), fetched from Rdatasets, not redistributed. - Code, at this exact version
- https://github.com/Maveri-Q/rerun-rebuilds
commit 175eddea88cc8adab8dd20f2100b5b453d359244 - Commands
- python3.11
$ python class-size-1999/rebuild.py STAR.csv > rerun_output.txt && cat rerun_output.txt - Where the result is read
- rerun_output.txt using regex first_grade_gap (-?[0-9.]+)
- Data files
- STAR.csv
fingerprint 0e8b179ea3d883730b25008ca1293c4c8f886aa8d5d299c03ab6ff05d0123ae6 - Limits
- 2 CPU, 4 GB, 15 minutes
- 9dec5a93332271dd854ae721e75ba32855be56a09e8781adaa1407fec5f90597
Show the claim exactly as stored on the blockchain
{"artifact":{"commit":"175eddea88cc8adab8dd20f2100b5b453d359244","data":[{"path":"STAR.csv","sha256":"0e8b179ea3d883730b25008ca1293c4c8f886aa8d5d299c03ab6ff05d0123ae6","url":"https://vincentarelbundock.github.io/Rdatasets/csv/AER/STAR.csv"}],"entrypoint":"python class-size-1999/rebuild.py STAR.csv > rerun_output.txt && cat rerun_output.txt","output":{"extractor":{"expr":"first_grade_gap (-?[0-9.]+)","type":"regex"},"path":"rerun_output.txt"},"repo_url":"https://github.com/Maveri-Q/rerun-rebuilds","runtime":"python3.11","setup":[]},"metric":{"comparison":"within_tolerance","name":"percentile points gained in a small class, first grade","reported_value":"8.57","tolerance":{"type":"abs","value":"0.25"}},"private_data":null,"resources":{"cpu":2,"max_minutes":15,"mem_gb":4},"source":{"authors":"A. B. Krueger; rebuild by the RERUN project","doi":"10.1162/003355399556052","license":"Rebuild: MIT. Data: GPL (R package AER), fetched from Rdatasets, not redistributed.","type":"paper","url":"https://doi.org/10.1162/003355399556052"},"statement":"Krueger (Quarterly Journal of Economics, 1999) studied Tennessee's STAR experiment, which put thousands of young children into small or normal-size classes at random. First graders in small classes scored 8.57 percentile points higher, before other factors are held equal. This runs our own small program on a public copy of the data, which has reading and maths scores only, so it can't match exactly. The rebuild isn't the author's code, so a mismatch ends as inconclusive.","title":"Do small classes help young children learn? Tennessee's experiment found 8.57 percentile points in first grade","version":1}Community notes
Notes are people's own words, kept on the blockchain. They never change a verdict, and the Lab's AI never reads them. Posting costs a little NEAR for storage, and what isn't used comes back.