A comparison of models for detecting discrimination: An example from medical school admissions

Abstract

Detecting bias in admissions to graduate and professional schools presents important problems to the data analyst. In this paper some traditionally used methods, such as multiple regression analysis, are compared with the newer methods of logistic regression and structural equations models. The problems faced in modeling decision rules in this situation are (1) a dichotomous dependent variable, (2) nonlinear relationships between independent variables and the probability of being admitted, (3) omitted variables, and (4) errors in variables. Each method used involves an attempt to solve one or more of these problems, but each has its own drawbacks. Using multiple methods, and finding several areas of agreement in the results among the methods, makes the conclusions stronger than had only one method been used.