A trend that exists in every single subgroup you examine can completely vanish—or flip into the opposite direction—the moment you combine them all together. This is not a computational error. The math works. This is Simpson's Paradox, and it lives in the weird space between intuition and reality where our brains stop making sense.
Most people assume that if something is true for group A and group B separately, it must be true for A plus B combined. It feels obvious. It feels like arithmetic. But statistics doesn't care about how things feel. A 2023 analysis in Scientific American highlights how this paradox has derailed drug trials, skewed hiring investigations, and made researchers draw conclusions that completely evaporate when someone bothers to check the underlying data structure. The paradox isn't rare—it's lurking in any dataset where subgroup sizes or distributions differ enough to create what statisticians call "confounding variables."
The classic example, documented on Wikipedia's exhaustive treatment of the topic, comes from a 1973 UC Berkeley gender discrimination study. When admissions were broken down by department, women's acceptance rates were either equal to or slightly higher than men's in almost every single program. Yet when all departments were combined, the university appeared to be discriminating against women. How? The departments with lower overall acceptance rates happened to have proportionally more female applicants. The aggregate trend reversed entirely because of the compositional difference between groups, not because of bias within them.
Here's what makes this particularly insidious: the paradox doesn't require anyone to be lying or the data to be wrong. Both the subgroup trends and the aggregate trend are mathematically correct. According to research highlighted in The Conversation, the paradox emerges when three conditions align: the subgroups have different sizes, the distributions of outcomes differ between groups, and—crucially—you're comparing ratios or percentages rather than absolute numbers. A treatment can improve outcomes for men and for women separately, but when you weight the improvement by how many men versus women were in each study, the overall trend can reverse.
Medical researchers face this constantly. Imagine a drug trial where men show improvement at a 70% success rate in one dosage group and 80% in another. Women show 60% and 90% in the same groups respectively. Both sexes improve with the higher dose. But if there were far more women in the low-dose group, the aggregate data could show the higher dose actually performing worse overall. The subgroup trends are real. The aggregate reversal is also real. They're just telling you different stories about the same numbers.
The paradox persists because we're bad at holding multiple distribution patterns in our heads simultaneously. When you see a headline about a trend, you rarely see the granular breakdown of how many people were in each category or how those categories were composed. Simpson's Paradox doesn't prove that statistics is useless—it proves that you can't understand data without understanding its structure. The numbers aren't lying. But without knowing how those numbers are organized, you might be.