The Use of a Probit Model for the Validation of Selection Procedures

Abstract

After a review of the disadvantages of linear models for estimating the probability of academic success from previous school records and admission test results, we propose the use of a probit model. In contrast to a number of linear regression models which use academic results as a dependent variable, the probit model allows the retention of drop-out cases within the sample; in addition, the coefficients as well as the probability estimates obtained from the probit model by the maximum likelihood method have satisfactory asymptotic properties. The model is calibrated on admissions data for the Ecole des Hautes Etudes Commerciales in Montreal for the years 1979 and 1980. It is noted that both of the variables used in selection have a significant impact on the probability of success, and each variable seems to be associated with a specific aspect of the phenomenon. Furthermore, when the candidates are grouped into quintiles, the deviations of the estimated probabilities from actual success rates are extremely small.