A researcher controls for a mediator variable between treatment and outcome in a regression. What bias does this introduce?