Unrelated in the crowd. A trade-off inside the gate.

Nine hundred applicants. A test score and a sports rating, dealt out completely independently — nothing whatever links them. Admit on the sum of the two, then look only at who got in.

What you're seeing

Every dot is an applicant. Across the bottom is how they did on a test; up the side is how good they are at sport. The two were dealt out with a coin that has no memory: knowing one tells you exactly nothing about the other. That is why the cloud is round and has no tilt. The gauge on the right says so — the correlation across the whole pool is zero.

The college admits on the sum. Be brilliant at the test, or brilliant at sport, or decent at both, and you are in; the diagonal line is the bar. Now look at the amber dots on their own. They lean. They lean down: among the students who got in, the better ones at the test are the worse ones at sport. Push the bar up and the lean gets steeper.

Nobody's qualities changed. No applicant is any different than they were a moment ago. The trade-off is a property of the gate, and it is easy to see why once you look at the corner. To be admitted with a poor test score you must be exceptional at sport — that is the only way to clear the bar. The top-left and bottom-right corners of the admitted region are full; the bottom-left corner has been sliced off entirely, because being mediocre at both is the one combination the gate refuses. Remove a corner from a round cloud and what is left is an ellipse lying on its side.

Now press show only who got in. The rejected applicants disappear, and with them every trace of the evidence. What is left looks exactly like a discovery: a clean negative relationship in a sample of real people. This is the position you are actually in whenever you study a group that was selected — hospital patients, customers who complained, published papers, funded startups, anyone who got in anywhere. You never see the rejected.

Then press admit on test score alone and watch the effect vanish. This is the part worth keeping. Selection is not the problem. Selecting on a consequence of both things you are studying is the problem — statisticians call such a variable a collider, and the college's admission decision is one. Cut on one quality and the other is untouched. Cut on the sum and you have manufactured a relationship out of nothing.

Finally, drag the second slider up. Give the two qualities a genuine positive link — make them properly twinned — and the gate still drags the picture downward. It can turn a real positive relationship into a flat one, and a flat one into an apparent trade-off. Berkson noticed this in 1946 in hospital records, where two unrelated diseases appeared to protect against one another, for no better reason than that having either one was a way to end up in the building.

The rule, exactly. Each applicant gets a test score x and a sports rating y, both standard normal, correlated by however much you ask for — zero, by default. Admission is x + y > c. The trick to solving it is to stop looking at x and y and look at their sum and difference instead. For a round Gaussian cloud those two are independent, so cutting on the sum truncates the sum's spread and leaves the difference completely alone: S = x+y, D = x−y ⇒ corr(x,y | admitted) = (VarS − VarD) / (VarS + VarD) Var D stays at 2 no matter where the bar goes, while Var S is squeezed below 2 the moment you cut — so the correlation among the admitted is negative for every bar, and heads for −1 as the bar rises. At a bar of exactly zero the whole expression collapses to a constant: corr(x,y | x+y > 0) = −1 / (π − 1) = −0.4669… Verified in node (improve/verify/99-berkson.js). The estimators are proved first on data whose answers are known — a perfect line must return ±1, a symmetric V must return exactly 0, and the truncated-variance formula is checked against 400,000 brute-force draws. Then the closed form is matched against 300,000 simulated applicants at seven different bars, agreeing everywhere to within 0.012, including the exact −1/(π−1). That constant was wrong in the first draft of the check, typed in as a hand-rolled decimal; the gate caught it, which is what the gate is for. Four negative controls. Admitting on one quality alone leaves the correlation at 0, and so does admitting the same fraction at random — so it is the collider and not the selection. The rejected applicants show the identical illusion, mirrored, so both halves of a split can lean while the whole leans not at all. The estimator is shown not to be wired to say “negative”: handed a pool built at ±0.6 and −0.3 it reports those numbers back. And two plausible-but-wrong closed forms — both of which predict no effect at all — are refuted by the measurement.