Causal Inference

Mediator

A mediator sits on the causal pathway between treatment and outcome: the treatment changes the mediator, and the mediator in turn changes the outcome. It is part of how the effect happens, not a nuisance to remove.

The key practical rule is the opposite of confounders: you should usually not control for a mediator if you want the total effect, because doing so blocks the very pathway you are trying to measure and shrinks the estimated effect. Deliberately splitting effects into direct and indirect (through the mediator) parts is the separate field of mediation analysis.

Exercise lowers heart-disease risk partly through blood pressure: blood pressure is a mediator, so controlling for it hides part of exercise's benefit.