Collider Bias

A collider is a variable that is caused by two (or more) other variables — in a causal diagram, two arrows "collide" into it. Conditioning on a collider, whether by adjusting for it in a model or by selecting study subjects based on it, opens a non-causal path between its causes and creates an association where none existed (or distorts a real one). This is the formal mechanism behind many puzzling findings.

Understanding colliders matters because the instinct to "control for more variables" can actively introduce bias. Unlike confounders, which should be adjusted for, colliders should generally be left alone. Selection bias is often collider bias in disguise: selecting a sample on a common effect of exposure and outcome conditions on a collider.

The classic pitfall is the "obesity paradox" pattern: among hospitalized patients, a risk factor can appear protective because hospitalization is a collider influenced by both the risk factor and other causes of admission. The fix is to draw a DAG and identify which variables are colliders before modeling.

For example, suppose both genetic risk and a lifestyle factor independently increase the chance of being enrolled in a cancer registry. Studying only enrolled patients can create a false inverse association between genetics and lifestyle, even if they are unrelated in the general population.