Describe some of the relationships and trade-offs between validity and adverse impact.

Question

What will be an ideal response?

Answer 1

So the presence of adverse impact is not a sufficient basis for a claim of unfair discrimination (Drasgow, 1987). However, apparent, but false nondiscrimination may occur when the measure of job success is itself biased in the same direction as the effects of minority group status on predictor performance (Green, 1975). Consequently, a selection measure is unfairly discriminatory when some specified group performs less well than a comparison group on the measure, but performs just as well as the comparison group on the job for which the selection measure is a predictor. This is precisely what is meant by differential prediction or predictive bias (i.e., different regression lines across groups based on the intercepts, the slopes, or both). We hasten to emphasize, however, that the very same factors that depress predictor performance (e.g., verbal ability, spatial relations ability) also may depress job performance. In this case, slopes may be identical across groups, and only intercepts will differ (i.e., there are differences in the mean test scores across groups). Gottfredson (1988) summarized the following problem based on the finding that the mean score in cognitive ability tests is typically lower for African Americans and Hispanics as compared to Whites: “The vulnerability of tests is due less to their limitations for measuring important differences than it is to their very success in doing so . . . The more valid the tests are as measures of general cognitive ability, the larger the average group differences in test scores they produce” (p. 294). Given differences in mean scores for cognitive ability tests across subgroups, and the consequent adverse impact, does this statement mean that there is an inescapable trade-off between validity and adverse impact?
Fortunately, the belief that there is a negative relationship between validity and adverse impact is incorrect in many situations. Specifically, Maxwell and Arvey (1993) demonstrated mathematically that, as long as a test does not demonstrate differential prediction, the most valid selection method will necessarily produce the least adverse impact. Hence to minimize adverse impact, HR researchers and practitioners should strive to produce unbiased, valid tests. However, a problem occurs when, unbeknownst to test users and developers, a test is biased and it is nevertheless put to use.