Imagine two hospitals. Hospital A has a higher survival rate for its patients than Hospital B. The choice seems obvious—if you need surgery, go to Hospital A. But what if I told you that Hospital B is actually better for both healthy patients and critically ill patients, considered separately? Both statements can be true at the same time.

This isn't a trick or a misprint. It's called Simpson's Paradox, and it's one of the most unsettling phenomena in statistics. It reveals how the way we group data can flip our conclusions upside down. Understanding this paradox isn't just an academic exercise—it's a defense against being misled by numbers that seem, on their face, to tell a clear story.

The strangest feature of Simpson's Paradox is that a trend observed in every subgroup can completely reverse when those subgroups are combined. In 1973, the University of California, Berkeley was accused of gender bias in graduate admissions. The aggregate numbers looked damning: men were admitted at significantly higher rates than women.

But when researchers examined the data department by department, a surprising picture emerged. In most individual departments, women were actually admitted at slightly higher rates than men. How could both things be true? The answer was that women tended to apply to more competitive departments with lower overall acceptance rates, while men applied more often to less competitive ones.

The aggregate number wasn't lying, exactly—it was measuring something different from what people assumed. This is the heart of the paradox. Combined data and segmented data can each be mathematically correct while pointing in opposite directions. Which one reflects reality depends entirely on what question you're actually trying to answer.

Takeaway

A statistic is not a fact about the world—it's a fact about a particular slice of the world. Always ask which slice you're looking at, and whether it's the right one for your question.

Simpson's Paradox almost always has a culprit lurking behind it: a hidden variable, sometimes called a confounder. This is a factor that influences both the groups being compared and the outcome being measured, quietly distorting the relationship between them.

In the Berkeley case, the hidden variable was department choice. In the hospital example, it's often patient severity—a hospital that treats sicker patients will naturally have worse aggregate outcomes, even if it provides superior care. In both examples, ignoring the hidden variable produces a misleading picture. Accounting for it reveals the truth.

This is why good reasoning about data requires more than mathematical skill. It requires curiosity about causes. You have to ask: what else might be driving this pattern? What information am I not seeing? A statistician can calculate averages, but only someone who understands the underlying system can identify which variables actually matter. Numbers alone cannot tell you what they mean.

Takeaway

Correlation without context is not evidence—it's a clue. The real work of understanding begins when you ask what invisible factors might be shaping the pattern you see.

If aggregation can mislead, why not always break data into its finest components? Because segmentation has its own dangers. Slice data too finely and patterns dissolve into noise. Slice it along the wrong dimensions and you create new distortions. The question isn't whether to segment—it's how to segment meaningfully.

The guiding principle is causal reasoning. Ask yourself: is the variable I'm segmenting by a genuine cause of the outcome, or just a coincidental grouping? In the Berkeley case, department selection was causally relevant because different departments genuinely had different admission standards. Segmenting by, say, the applicant's favorite color would add nothing.

This is where science and philosophy meet. Statistics can describe relationships, but deciding which relationships matter requires judgment about how the world actually works. Karl Popper argued that good theories make risky predictions—they say something specific enough to be wrong. Good data analysis works the same way. It commits to a model of causation, then tests whether that model holds up.

Takeaway

Data doesn't interpret itself. Every analysis rests on hidden assumptions about causation—and the honest analyst makes those assumptions visible, then puts them at risk.

Simpson's Paradox is a humbling reminder that numbers can be simultaneously accurate and misleading. The same dataset, examined differently, can tell opposite stories—and both stories can be technically correct.

The defense isn't to distrust statistics but to interrogate them. Ask what's being combined, what's being separated, and what hidden variables might be pulling the strings. Good thinking isn't about finding the right answer immediately—it's about learning to ask better questions of the evidence in front of you.