In the early 1970s Berkeley was sued for gender discrimination in graduate admissions. Of 8,442 male applicants for the fall of 1973, 44 percent were admitted, against 35 percent of 4,351 women. Assuming similar qualifications, the aggregate numbers look like exactly what the suit alleged.
Yes, and it happened. Looking within specific departments, the bias against women went away and even reversed in several cases. Women disproportionately applied to departments with low while men disproportionately applied to departments with high ones, so the totals said the opposite of the parts.
In the early 1970s the University of California, Berkeley, was sued for gender discrimination over graduate admissions. Of the 8,442 male applicants for the fall of 1973, 44 percent were admitted, but only 35 percent of the 4,351 female applicants.
The departments told a different story. Looking more closely within them, the bias against women went away and even reversed in several cases, because women disproportionately applied to departments with low acceptance rates while men applied to departments with high ones.
The same shows up in education data. National math scores for 17-year-olds barely moved between 1992 and 2012, dipping by a point. Average scores rose slightly for white, black and Hispanic students separately, because the of the student body changed.
Earnings repeat the pattern. Inflation-adjusted for men aged 25 to 44 fell by about $1,000 between 1982 and 2013. The same measure rose by more than $3,000 for white men, just under $1,000 for black men, and about $10,000 for other men, with Hispanic men flat.
wat. Every part can improve while the whole gets worse.
wat wat
The aggregate number is the one that gets quoted, and it can point the other way from every group inside it.
How do we know?
Brad Hershbein sets the Berkeley case out for Brookings alongside two modern examples, national math scores and median earnings, where the same reversal appears.
What is still unknown
A commentary using worked examples, which shows the pattern rather than measuring how often it occurs.
Learn something?
Where to next
Can the first digit of a number give away a fraud?statistics
Can one ball be cut up and reassembled into two of the same size?paradox
After the host opens a door, does switching help?paradox