Efficiency and adverse impact of general classification decisions

Classification decisions relate to situations in which a battery of predictors is used to assign individuals to a number of different trajectories. De Corte (2000) proposed a method to estimate the classification efficiency in case the assignment of individuals to trajectories is based on least square criterion estimates. The current paper extends this method to the case where the applicants come from several subpopulations and estimates are no longer only regression weighted. The extension is motivated by the fact that using other than regression based criterion estimates for assigning applicants to the different trajectories may result in classification decisions that show substantially less adverse impact as compared to classifications in which regression based criterion estimates govern the allocation process (De Corte, Lievens & Sackett, 2007). An application of the new analytic method indicates that while classifications based on regression weighted criterion estimates lead to optimal classification efficiency, they also yield substantial adverse impact because many of the most valid predictors, and cognitive ability predictors in particular, show large effect sizes in favor of the so-called majority applicants. Alternatively, general (non regression based) classification decisions lead to a wide range of possible trade-offs between efficiency and diversity where concessions in terms of classification efficiency are compensated by more advantageous levels of adverse impact. The proposed method may be used by practitioners to alleviate the quandary between efficiency and adverse impact in a classification context.

Publication year:2010

Accessibility:Open