Predictive validity under partial observability

Abstract

To assess the predictive capacity of selection tests is a challenge because the response variable is observed only in selected individuals. In this paper we propose to evaluate the predictive capacity of selection tests through marginal effects under a partial identification approach. Identification bounds are defined for the marginal effects under monotonicity assumptions of the response variable. The performance of our method is assessed using a real data set from the university selection test applied in Chile and compared with the marginal effect of the traditional model used in Chile to evaluate the predictive capacity of the selection test.