In the fall of 1973, the University of California, Berkeley admitted 44% of the men who applied and only 35% of the women, and the numbers looked like an open-and-shut case of bias.
Then the same data was examined department by department. In the six largest departments, the gap did not just shrink. It flipped, with women holding a small but statistically significant edge.
A gap that vanished on closer inspection
How can women do better inside the departments and worse overall? The answer is not hidden in the math. It is hidden in where people applied.
Women had applied in greater numbers to the most competitive, most selective departments, the ones that turned down most candidates whoever they were. Send more women through the hardest doors, and their overall rate drops, even if each door treats them fairly or slightly better.
The combined figure was real. It simply answered a different question from the one everyone was asking.
The same flip shows up in a hospital
A well-known medical study compared two treatments for kidney stones. Treatment A had a better success rate than treatment B for small stones. It also did better for large stones.
Yet once every patient was pooled together, treatment B looked like the more effective option. The catch: doctors were reserving treatment A for the most serious cases. Hard cases drag down any average, so A was punished in the totals for taking on the toughest jobs.
A patient reading only the combined figure would pick the worse treatment. That is why this quirk matters far beyond a statistics classroom.
Two baseball players, two opposite answers
In 1995, David Justice had a better batting average than Derek Jeter. In 1996, Justice was ahead again. Season by season, Justice wins every time.
Combine the two seasons, though, and Jeter comes out on top. The combined average quietly gives more weight to whichever season a player had more chances in, and that weighting is enough to tip the result.
It sounds like a freak coincidence. Mathematician Ken Ross estimated that this kind of reversal would show up about once a year among all the possible pairs of players.
Not really a paradox at all
The effect has a name: Simpson’s paradox. Statistician Edward H. Simpson described it formally in a 1951 paper in the Journal of the Royal Statistical Society, and Colin R. Blyth attached Simpson’s name to it in 1972. The idea is older still. Karl Pearson noticed similar effects in 1899 while studying heredity, and Udny Yule described analogous cases in 1903.
Despite the label, nothing here contradicts itself. The flip comes from a confounding variable, a hidden factor that shapes both who ends up in each group and how things turn out. Department choice at Berkeley. Severity of the case at the hospital.
Statisticians call the deeper issue non-collapsibility: some relationships cannot be squashed into one combined number without changing their meaning. One economic dataset even showed prices and demand rising together, against basic economic theory. Time was the hidden variable. Split the data by period, and the expected pattern, higher prices with lower demand, came straight back.
So which number should you trust?
The instinct is to trust the big total, since it uses all the data. Causal inference researcher Judea Pearl argues the opposite: the data split into the right subgroups is what reveals the real cause and effect.
The key word is “right”. Pearl’s so-called back-door criterion is a method for checking whether a way of splitting the data actually blocks the hidden influences, rather than just slicing it at random.
And this is not a once-in-a-century curiosity. In a random table crossing two options, two outcomes and two groups, the chance of the reversal appearing purely by accident is exactly 1 in 60. In certain models with two predictors and a single outcome, it climbs to about 12.8%.
The story depends on how you slice it
Every headline statistic is an average of something. Most of the time, the average tells the truth. Sometimes, as Berkeley found out, it tells the exact opposite of what is happening underneath.
So the next time a number announces that one group beats another, the most useful question is not whether the figure is correct. It is what that figure has been averaged over.

