Thursday, January 02, 2003

Measuring within and between-group variances well 

No. 2 Pencil has disected perfectly the argument for the efficacy of affirmative action admissions in the Univ. of California system. (Click here and scroll to the entry if the previous link doesn't work -- her archives weren't working last I checked.) Discussing the claim that achievement test scores don't measure academic success, Kimberly Swygert, a psychometrician, demonstrates the difference between "within-group and between-group variances" that Dave discussed earlier today.
Ms. Guerrero [who was arguing for affirmative action in law school admissions] claims, as do many affirmative action proponents, that SAT scores should be weighted less because they are not valid predictors of performance in higher education. But ethnicity is? If the purpose is to better predict college performance, then the variable of ethnicity fails miserably, which reveals that affirmative action proponents aren't really interested in improving prediction. They aren't really interested in developing a set of standards that will truly determine who will do best in college.
She also gets third-factor effects down cold; discussing how minority students who went to the same colleges and had the same high grades (perhaps inflated?) Swygert demonstrates the care one needs in interpreting the data.
possibly it demonstrates that grade inflation that more commonly affects minorities. Or it demonstrates that stereotype threat is a reality, and that minority test takers are so convinced that they will do poorly on the exam that they do poorly on the exam. Or it demonstrates the effectiveness of affirmative action in convincing minorities that they do not need to study hard for the LSAT, because they will be admitted with lower scores than will whites. Or it demonstrates that minority students don't prepare correctly for the LSAT, and other students do. Yes, test bias is a possibility, but it's tiring to see it constantly presented as the only possible explanation for the data.