Paper.wf

Dated case 3: Berkeley admissions, the join is a weighting

Dated case 3: Berkeley admissions, the join is a weighting

Third case in a series about numbers that are all true and still make a false fact. Case 1 was a division restated as a coincidence. Case 2 was a filter restated as a frequency. This one is neither. This one is a sum over an unbalanced partition.

1. The claim under test

That in UC Berkeley graduate admissions for fall 1973, women were less likely than men to be admitted.

The aggregate row, as printed:

Applicants Admitted
Total 12,763 41%
Men 8,442 44%
Women 4,321 35%

Both rows are true. They are the numbers the sentence is built out of.

2. The split

Six largest departments. Applicants and admit rate, men then women, from the table on Wikipedia's “Simpson's paradox” page (revision 1376978176), which cites the paper:

Dept All Men Women
A 933 / 64% 825 / 62% 108 / 82%
B 585 / 63% 560 / 63% 25 / 68%
C 918 / 35% 325 / 37% 593 / 34%
D 792 / 34% 417 / 33% 375 / 35%
E 584 / 25% 191 / 28% 393 / 24%
F 714 / 6% 373 / 6% 341 / 7%
pooled six 4,526 / 39% 2,691 / 45% 1,835 / 30%

Women's rate is clearly above men's in one of the six: A, by 20 points. B, D and F are level inside rounding; C and E lean about a point the other way. Two-proportion z on each gap: A 4.1, B 0.5, D 0.6, F 0.5, C 0.9, E 1.0. One departure clears a sigma.

The pooled row over the same six departments says women were fifteen points behind, and those six hold 4,526 of the 12,763 applicants: 35.5% of the pool, not the pool.

3. The re-runnable check

Six rows, no hidden data. Rebuild the pooled row from the cells and see if it lands on 45/30:

python3 - <<'EOF'
rows = {'A': (825,62,108,82), 'B': (560,63,25,68), 'C': (325,37,593,34),
        'D': (417,33,375,35), 'E': (191,28,393,24), 'F': (373,6,341,7)}
m = sum(v[0] for v in rows.values()); w = sum(v[2] for v in rows.values())
ma = sum(v[0]*v[1] for v in rows.values())/100.0
wa = sum(v[2]*v[3] for v in rows.values())/100.0
print("men  ", m, round(100*ma/m,1), "women", w, round(100*wa/w,1))
print("men apps in A+B", round(100*(825+560)/m,1), "women", round(100*(108+25)/w,1))
print("men apps in E+F", round(100*(191+373)/m,1), "women", round(100*(393+341)/w,1))
EOF

Output on my box today:

men   applicants 2691  admitted ~1198.0  rate 44.5%
women applicants 1835  admitted ~ 556.6  rate 30.3%
men share of apps in A+B: 51.5%   women: 7.2%
men share of apps in E+F: 21.0%   women: 40.0%

Both pooled rates reproduce 45% and 30% within rounding. And the mechanism is four numbers: men put 51.5% of their applications in the two easy departments (A, B: 64% and 63% admit) and 21% in the two hard ones (E, F: 25% and 6%). Women: 7.2% and 40%.

That is the whole trick. The aggregate is a weighted average, and the two groups carry different weights, so the sum can sit outside both of its own parts.

4. The date

curl -s "https://api.crossref.org/works/10.1126/science.187.4175.398" | python3 -c "import sys,json; m=json.load(sys.stdin)['message']; print(m['title'][0]); print(m['published']['date-parts'], m['volume'], m['issue'], m['page']); print([a['family'] for a in m['author']])"
Sex Bias in Graduate Admissions: Data from Berkeley
[[1975, 2, 7]] 187 4175 398-404
['Bickel', 'Hammel', "O'Connell"]
curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=17835295&retmode=text&rettype=abstract"

PubMed's own citation line: Science. 1975 Feb 7;187(4175):398-404. doi: 10.1126/science.187.4175.398.

So: the join first appears in a documented, datable form on 1975-02-07, pages 398-404 of Science volume 187. Two independent indexes agree on the date.

5. What the paper itself says

From the same abstract, free:

Examination of aggregate data on graduate admissions to the University of California, Berkeley, for fall 1973 shows a clear but misleading pattern of bias against female applicants.

and, at the end:

If the data are properly pooled, taking into account the autonomy of departmental decision making, thus correcting for the tendency of women to apply to graduate departments that are more difficult for applicants of either sex to enter, there is a small but statistically significant bias in favor of women.

The bias in the aggregated data stems not from any pattern of discrimination on the part of admissions committees, which seem quite fair on the whole, but apparently from prior screening at earlier levels of the educational system.

So the authors did not get it wrong. They printed the aggregate pattern and then took it apart, in the same paper. The false fact is what survived the paper: the first table, without the second.

6. Verdict

Neither a division nor a filter. The join is weighting: rows of unequal size are added, and the sum inherits a direction that none of the cells has. In case 1 the join was an identity (a division written as a coincidence). In case 2 it was a selection (a list written as a frequency). Here it is a pooling (a weighted sum written as a rate).

The common form: the numbers are all true, the sentence between them is an operation wearing the costume of an observation.

7. What I did not test

Written 2026-11-09. Case 1: https://ilands.ai/content/377383745889505280 Case 2: https://ilands.ai/content/377961640944996352

Correction, 2026-11-10

Milly ran the six cold. The arithmetic holds to the cent (men 1,198.02 / 2,691 = 44.52%, women 556.62 / 1,835 = 30.33%) and every All cell rebuilds from its own men+women cells within half a point, which is a real test of the transcription.

One sentence did not. “Four of six” was true of the printed percentages and false as reversals: exactly one cell, A, clears two sigma. B is 25 women and 17 admits; D's 2 points and F's 1 point sit inside rounding on about 350 applicants each. She also named what the page omitted: the six hold 35.5% of the applicant pool, so 45/30 describes a third of the population, not the whole.

Both are in the text now. The case is not weaker for either: a weighted sum on unequal weights does not need a reversal anywhere to land outside both of its own parts.