Berkson's Paradox: Why Hospital Data Makes Healthy Smokers Look Fine

Imagine you're a researcher in the 1940s, studying whether smoking causes lung disease. You pull patient records from the hospital where you work. You compare smokers and non-smokers among the admitted patients. And you find... almost no relationship. Maybe a slight negative one.

Photo by Negative Space on Pexels.

You didn't find nothing because smoking is safe. You found nothing because you were looking in exactly the wrong place.

This is Berkson's paradox, named after Joseph Berkson, a biostatistician who published a quietly devastating paper in 1946. The core problem: when you select your study population from a place that requires some kind of health problem to get in, you've already filtered out a huge chunk of the healthy population. Everyone who shows up at a hospital has something wrong. Smokers with lung disease are there. But so are non-smokers with broken legs, appendicitis, and everything else. Suddenly your sample looks like smokers have a better respiratory track record than they do, because the non-smokers in your dataset are there for other reasons.

Berkson called this "bias of selection." Statisticians today call it collider bias when they're being precise about causal graphs. The hospital admission is a collider: a variable caused by both smoking status and health problems. Conditioning on a collider (by studying only hospital patients) opens a spurious statistical pathway between variables that might be independent or even positively correlated in the general population.

Here's a cleaner version of the same trap. Suppose attractive people and talented people are roughly independent in the general population. No real correlation. Now imagine you're casting for a television show, and you only see candidates who passed an audition that selects for either attractiveness or talent. In your casting pool, if someone is not particularly talented, they probably got in because they're attractive. And vice versa. Suddenly attractiveness and talent look negatively correlated in your data. You'd walk away thinking beautiful people tend to be less skilled. That conclusion is real within your sample and completely wrong about the world.

This is why the paradox stings. You didn't make an error in your math. Your calculations are correct. The dataset is the problem.

graph TD
    A[Smoking] --> C{Hospital Admission}
    B[Other Illness] --> C
    C --> D[/Study Sample/]
    A --> E((Lung Disease))
    E --> C

Conditioning on hospital admission (the collider) creates a false link between smoking and other illnesses in your observed data, even if none exists in reality.

Berkson's paradox shows up far outside medicine. Online dating platforms see it constantly: users rate profiles after filtering for people they'd swipe right on, which means they're evaluating a pre-selected pool where looks and personality are already correlated in a distorted way. Hiring managers who only interview candidates from elite universities see a compressed range of credentials, which can make GPA look less predictive than it actually is. Any time your data comes from a filtered funnel, you risk the same problem.

The diagnosis is straightforward to describe and genuinely hard to execute: ask where your data comes from before you trust what it says. If the population you're observing had to satisfy some condition to appear in your dataset, ask whether that condition is related to the variables you're studying. If it is, you might be conditioning on a collider.

The fix requires either broadening your sample to include people who didn't make it through the filter (harder than it sounds), or using causal modeling to explicitly account for the selection process. Berkson himself suggested that hospital-based studies needed careful comparison against population-level data. Seventy-plus years later, researchers still don't always do that.

What makes this particularly sneaky is that researchers who understand selection bias still fall into it. Knowing that biased samples are bad doesn't automatically tell you your sample is biased in this particular way. The hospital feels like a natural place to study disease. Of course you'd go there. That intuition is exactly the trap.

Your data was collected somewhere, by someone, using some process. That process left fingerprints all over what you're looking at. The numbers themselves won't tell you. You have to know the story of how those numbers were born.