Multiple- versus Single-Index Predictors of Dangerousness, Suicide, and other Rare Behaviors

Abstract

The principle that prediction of low base-rate events based on test signs may be grossly in error in the direction of overprediction is argued to be limited to situations where single-index predictors are used. An example is presented of a three-variable Boolean function which accurately distinguishes 70% of true positive instances and 100% of true negative instances in a situation where the base rate of the predicted event is 100 per 100,000.