A metric measuring how far a subgroup's actual outcome deviates from what an additive model of independent attributes would predict.