Berkson's Paradox

In plain terms

Berkson's paradox is a trade-off that isn't real. Two qualities have nothing to do with each other out in the world. Then you look at a group that was filtered in a way that either quality could get you into, and inside that group the two look like opposites. Having more of one seems to mean having less of the other.

Nothing caused that. The filter did it. If either trait alone is enough to clear the bar, then anyone who's weak on one must be strong on the other, otherwise they wouldn't be in the group you're looking at. So the group fills up with lopsided cases, and the balanced-but-mediocre ones never appear.

The statistician Joseph Berkson described it in 1946 using hospital patients. Two unrelated conditions each make admission more likely. Among admitted patients, they show up as negatively correlated, because having one already explains why the patient is in a hospital bed, so they need the other one less. Study patients instead of people and you'll find relationships between diseases that don't exist anywhere outside the building.

Why it matters

This is a selection effect, and it's the sneakiest kind, because the sample looks fine. Nobody cherry-picked. The data is accurate. The correlation is genuinely there in the numbers you have. It simply doesn't survive contact with the population you actually care about.

It shows up wherever a gate exists. Hires, admissions, funded startups, playoff teams, published papers, people you've dated. Any pool assembled by a rule that rewards two things at once will show those two things trading off inside the pool. The conclusion people draw from it, that one quality comes at the cost of the other, is a fact about the gate, not about human beings.

Canonical example

"Every genuinely talented person we hire coasts. The hard workers here are the ones with less raw ability. Talent makes people lazy."

Run the numbers. Say 400 people apply. Talent and work ethic are unrelated in that pool, and each is high in half the applicants, so the four combinations split evenly at 100 each. The firm hires anyone strong in at least one, which is 300 people, and rejects the 100 who are low in both.

Now look inside the firm. Of the 200 hires with high talent, 100 also have a high work ethic, so half. Of the 100 hires with low talent, every single one has a high work ethic, because that's the only way they got through the door. From the inside, talent looks like it predicts laziness. In the applicant pool it predicted nothing at all. The hiring rule manufactured the entire pattern.

The dating version is the same arithmetic. If you'll date someone who is either very kind or very attractive, the very attractive people in your history didn't need to be as kind to make the cut, and the kindest ones didn't need to be as attractive. Hence the familiar conclusion that attractive people are colder, drawn from a sample of one person's dating history.

Counter-example (not a fallacy)

"Across every car model sold last year, not just the ones we stock, engine power and fuel economy trade off against each other. That trade-off is physical, and it holds in the full catalog."

This isn't Berkson's paradox, because the measurement wasn't taken inside a filtered pool. The full population of models shows the same negative relationship, and there's a mechanism, burning more fuel per second to make more power. A real inverse relationship that persists in the unselected population is just an inverse relationship.

The line: does the negative relationship survive outside the gate, or does it only exist among the ones that got through?

How to fix it

If you've been linked here, ask what filter produced the group you're describing. Hires, matches, admits and finalists all passed a test, and if two qualities each helped pass it, you'll see them trading off among the survivors no matter what's true of everyone else. The fix is to look at the applicant pool rather than the hired pool, or at least to say the claim out loud with its boundary attached: "among the people we hired" rather than "in general". If the wider data isn't available, treat the trade-off as a property of your selection process until something else confirms it.

If you're on the receiving end, ask about the people who aren't there. "Who got filtered out before this group existed, and what would they have looked like?" The pattern usually dissolves the moment the rejected cases are put back in.