Collider Bias

mechanism

Collider bias appears when you study only cases admitted through a shared gate. Inside that selected group, independent causes can look related—or a real relationship can appear weaker or reversed—because either cause helped the case become visible.

Look only at hospital patients and smoking can appear less associated with heart attacks. The cigarette has not become protective; the hospital door has distorted the comparison.

The gate manufactures the pattern

Suppose both smoking and heart attacks can help put someone in hospital, while smoking can also lead to admission through other conditions. Once you restrict the sample to people already hospitalized, you condition on their shared result—the collider. A patient who entered through one route has less need to have entered through the other, so the causes become statistically entangled inside the ward even if they were not related that way outside it.

This is a particularly treacherous form of selection bias: the filter does more than make the sample unrepresentative. It can create a correlation between variables upstream of selection, conceal a genuine association, or point it backwards.

Where it shows up

Inside the hospital ward

Among admitted patients, smokers include people hospitalized through many smoking-related routes besides heart attack. Comparing only this selected population can therefore make smoking and heart attack look less connected than they are in the wider population.

Not every filtered sample contains a collider

Selection alone does not guarantee this paradox. The distortion requires the gate to be influenced by the variables being compared, or by their consequences. And the hospital pattern does not establish that smoking protects against heart attacks; it shows why a relationship observed after admission cannot automatically be exported beyond admitted patients.

Draw the door before reading the room

Before trusting a correlation in any selected group, write down the rule that made its members observable. Then ask whether either variable you are comparing could have helped a case cross that gate. If both arrows point into selection, seek data from before the gate—or compare with people who never passed through it.

Episodes that teach this

Why Hot Guys are Jerks and Cute Girls Are Dumb? Berkson's Paradox 9,292 views
The hospital example says that among only hospital patients, smokers may appear less associated with heart attacks because smoking also selects people into hospital through many other conditions.