| Abstract: | Presents a perspective on test and item bias in predicting the performance of individuals in majority and minority groups. It is suggested that a difference between groups in the slope of the regression of the criterion on a test used for prediction represents bias that can frequently be corrected. Given a criterion with adequate measurement properties, the selection of items with essentially equivalent discrimination indices is expected to produce approximately parallel slopes. Accordingly, items having different discrimination indices are themselves biased. Intercept bias, on the other hand, is essentially not under the control of the test constructor. It is argued that the goal of producing a test without intercept bias is an unreasonable one. These problems are illustrated by factor models of intercept bias. It is concluded that relative differences in item-difficulty levels from group to group, given items with adequate discrimination indices, should not be labeled as bias. Such items are not automatic candidates for discard. (13 ref) (PsycINFO Database Record (c) 2010 APA, all rights reserved) |
