A drug treatment fails for men. The same drug fails for women. Yet when you combine the data, it works. This isn't a paradox in the sense of being impossible—it's a paradox in the sense of being deeply, unsettlingly real.
We tend to assume that if something is good for Group A and good for Group B, it must be good overall. That's intuitive. That's also wrong. Most people, including many statisticians and medical researchers, believe that aggregated data simply represents subgroups added together. Zoom out, the logic goes, and you see the full picture. But Simpson's Paradox—a phenomenon first formally described in 1951 but dating back further—reveals that zooming out can flip the story entirely. A trend that reverses when you break data into categories suggests hidden structures in how those categories distribute. The subgroups aren't separate views of the same truth. They're evidence that something else is happening underneath.
The most famous real-world example comes from UC Berkeley admissions in 1973. The university's graduate programs admitted 44% of male applicants but only 35% of female applicants—a stark gap suggesting discrimination. When the university's statisticians looked at individual departments, though, nearly every one admitted women at a higher rate than men. Some departments favored women by as much as 10 percentage points. How could this be? Women disproportionately applied to competitive departments with lower acceptance rates across the board. Men clustered in less competitive fields. The overall trend was real, but its apparent cause was an illusion created by different group sizes and different base rates.
Medical trials have fallen into this trap repeatedly. According to a 2019 analysis in Scientific American, a treatment can appear beneficial when data from men and women are combined, yet show harm when separated by gender. The reversal occurs when one group has higher underlying risk or a larger sample size, and the treatment effect interacts with that risk differently. A 2004 analysis published in the Stanford Encyclopedia of Philosophy notes that Simpson's Paradox isn't a flaw in arithmetic—the aggregate numbers are correct—but rather a failure of causal reasoning. The aggregate tells you what happened. The breakdown tells you why.
COVID-19 case fatality rates between Italy and China illustrate this vividly. When comparing raw death counts, Italy's rate appeared higher. But when age-stratified data emerged, China showed higher mortality in nearly every age group. Italy's population simply skewed older, masking the true relative severity within comparable cohorts. The paradox doesn't mean the aggregate data was wrong. It means the aggregate data was measuring something different than what intuition suggests it should measure.
Why does this happen? Subgroup analysis requires careful attention to what statisticians call "confounding by group size" and "effect modification." When the sizes of subgroups differ substantially, or when a variable affects subgroups differently, aggregate trends can reverse. It's not a quirk—it's a fundamental feature of how proportions work at different scales. A treatment might have a small positive effect in a large, low-risk group and a large negative effect in a small, high-risk group. The aggregate could swing either way depending on which dominates.
The implication is humbling: published aggregate results, especially in medicine and policy, deserve skepticism until you see the breakdown. Not because researchers are dishonest, but because our brains aren't wired to expect reversals at different zoom levels. The UC Berkeley case wasn't hidden discrimination—it was hidden structure in how different groups applied. That's often more interesting than the surface story. Asking "why did this paradox occur?" frequently reveals the real mechanism behind the data, the one that actually matters for understanding what's happening and what to do about it.