Categories → Variables, Causal Inference, Causal Wizard Concept
A collider is a variable caused by both the treatment and the outcome. Controlling for a collider can create a spurious association between treatment and outcome that didn't exist before - the opposite of what controlling for a confounder does.
A collider is a variable that is directly caused by both the Treatment and the Outcome (or by variables that themselves cause Treatment and Outcome). In a Causal Diagram, this looks like two arrows colliding head-to-head at the same node:
Treatment → Collider ← Outcome
This is the opposite arrow pattern to a confounder (where the arrows point out of the confounder, toward both Treatment and Outcome).
Confounders need to be controlled for. Colliders are the opposite: do not control for a collider. Conditioning on (controlling for, or restricting your sample by) a collider can create a spurious statistical association between Treatment and Outcome, even when no causal relationship between them exists at all.
This is often called collider bias or selection bias, because it very commonly shows up as a side-effect of how a sample was selected, not just an explicit modelling choice.
Suppose Talent and Looks are completely independent in the general population - knowing someone is talented tells you nothing about how they look. Now suppose that both Talent and Looks independently increase a person's chance of being cast in a movie.
Talent → Cast in a movie ← Looks
Cast in a movie is a collider of Talent and Looks. If you now only look at actors (i.e. you've implicitly conditioned on "cast in a movie" = yes, by restricting your sample to actors), you'll typically find that Talent and Looks are negatively correlated - the untalented actors who made it tended to be very good-looking, and the less conventionally attractive ones who made it tended to be very talented. This apparent relationship is completely spurious: it's an artefact of the selection process, not a real effect of one on the other.
The same mechanism explains other well-known examples, like the observation that among hospitalized patients, having one disease can appear negatively associated with having another (Berkson's paradox) - hospitalization is a collider of the two diseases (either one, or both, can land you in hospital), and conditioning on it (by only studying hospitalized patients) distorts the relationship between them.
Two common ways a collider sneaks into an analysis:
Causal Wizard identifies collider variables automatically from your Causal Diagram during the Check process, and excludes them from the set of variables it controls for. You should still think carefully about whether your sample's selection process (how rows ended up in your dataset in the first place) might itself be conditioning on a collider - that's something no software can detect for you, since it depends on how the data was collected, not just what's in your Causal Diagram.