What is Berkson's paradox?

Berkson's paradox is a spurious negative association created between two unrelated traits when the sample is selected on something that both traits influence. Conditioning on that shared outcome — a collider — makes the traits look inversely related even though they are independent in the wider population. It is why, for instance, talent and looks can seem negatively correlated among celebrities when they aren't in general.

Berkson's Paradox — unspurious: a compendium of statistical illusions

01 · What just happened

Inside the selected group, one virtue explains away the other

Think about who gets through the casting door. Someone short on screen presence can only make the cut by being unusually talented; someone short on talent can only make it by being magnetic on camera. The doorway forbids being mediocre at both — and so, among those inside, low scores on one trait guarantee high scores on the other.

That manufactured trade-off is what the downward-sloping line is reporting. It's a fact about the doorway, not about actors. Yet to anyone who only ever observes the selected group — a casting director, a film critic, a fan — it looks exactly like a law of nature: the beautiful ones can't act.

Simpson's paradox punished you for failing to split the data. Berkson's punishes you for the opposite sin: your data arrived pre-split, by a filter you forgot existed.

02 · The original case

Berkson's hospital, 1946

Joseph Berkson, head of biometry at the Mayo Clinic, noticed the trap in the standard study design of his day: comparing diseases among hospital patients. Suppose two conditions are completely unrelated in the general population. Each, on its own, can put you in hospital. Then among hospital patients the two become statistically linked — because a patient without one disease must, more often than not, be there for the other.

The arithmetic is unglamorous and devastating. In the population below, disease B is found in 10% of people whether or not they have disease A: textbook independence. Walk the same people through a hospital's front door and the association snaps into existence.

The same 10,000 people, before and after the front door Column width = group size · Shaded share = how many have disease B

Fig. 4 — Admission manufactures association. In the population, B's share is 10% in both columns — flat, independent, boring. Inside the hospital, patients without A carry B at nearly seven times the rate of patients with A. Having one disease "explains" your admission; lacking it demands another explanation. (Admission rule: anyone with A or B, plus 5% of everyone else.)

03 · The triangle returns

Simpson's diagram, with the arrows reversed

Readers of № 1 in this series will recognise the shape. In Simpson's paradox, the third variable pointed at both of yours — a confounder — and the cure was to split the data by it. Here the arrows run the other way: talent and presence both point into selection. Causal-inference people call such a variable a collider.

And colliders invert the rule. Splitting — or sampling, or filtering — on a collider is precisely what creates the phantom correlation. The cure is to leave it alone: study the whole population, or at least model how the doorway works.

This is why "control for everything" is bad advice. Adjusting for a confounder removes bias; adjusting for a collider injects it. The triangle looks identical in a spreadsheet. Only the direction of the arrows — a claim about how the world works, not a property of the data — tells you which move is safe.

The collider

Fig. 3 — Where the arrows collide. Disease A and disease B each push people into hospital. Look only inside the hospital — condition on the collider — and a path opens between A and B that does not exist outside.

04 · Field notes

Colliders you have personally met

Dating. Your shortlist is a casting door: people make it by being some mix of kind, attractive, funny and available. Inside the shortlist, those traits trade off — hence the eternal complaint that the charming ones are unreliable. The population at large has no such rule; your filter does.

Hiring. Elite firms select on a blend of brilliance and polish, then their managers "discover" that the most brilliant hires are the least polished. Restaurants that survive in terrible locations turn out to have wonderful food — because surviving there required it.

Pandemic data. Several early-2020 studies of hospitalised COVID patients found smokers strangely underrepresented, briefly fuelling a "protective nicotine" theory. Hospitalisation was a collider: there were many routes into the sample, and conditioning on it could distort the smoking–COVID relationship without any biology at all.

The catch-question is the mirror image of the survivorship test: not "who left the sample?" but "how did anyone get in?" If entry depended on the very variables you're correlating, the correlation is the doorway's autograph — interesting evidence about the filter, and none at all about the world.

Filter on two virtues, and they become rivals.