Mediators and Colliders
What you'll learn: How to distinguish between mediators and colliders in causal graphs, two fundamentally different variable types with opposite roles in causal inference.
Understanding Mediators
A mediator sits on the causal path between treatment and outcome. It's a mechanism through which the treatment exerts its effect.
Think of it like this: Exercise → Weight Loss → Lower Blood Pressure. Weight loss is a mediator—exercise causes weight loss, which in turn causes lower blood pressure. The effect flows through the mediator.
In a DAG:
X → M → Y
Where X is treatment, M is the mediator, and Y is the outcome. Controlling for a mediator blocks part of the causal effect you're trying to measure—usually something you want to avoid if you're estimating total treatment effects.
Understanding Colliders
A collider is caused by two or more variables pointing into it. It's where causal arrows collide.
Analogy: Talent → NBA Selection ← Height. Both talent and height cause selection into the NBA. Among NBA players only, you might observe a negative correlation between talent and height (shorter players need exceptional talent to compensate)—even though talent and height aren't causally related in the general population!
In a DAG:
X → C ← Y
Here C is a collider. Critical insight: Controlling for a collider opens a spurious path between X and Y, creating bias where none existed. This is the opposite of what you do with confounders.
The Key Difference
- Mediators: On the path (X → M → Y). Controlling blocks causal flow.
- Colliders: Where paths meet (X → C ← Y). Controlling creates false associations.
Key Takeaway: Mediators transmit causal effects and should often be left uncontrolled; colliders block spurious paths naturally but create bias when controlled—recognize them in your DAG to avoid conditioning mistakes.