The late Daniel Seligman, who more or less invented the kind of heavily quantitative journalism I practice, loved gambling. (Here's his 1997 City Journal story about the regular poker game he'd been in for 43 years.)
In this 1989 column, he makes an important point about hiring discrimination by invoking his favorite hobby:
April 10, 1989
BIAS IN THE CASINO
Although generally loath to talk up the competition, we find ourselves again recommending instant acquisition of the Journal of Vocational Behavior, last plugged here on August 3, 1987. This scholarly bimonthly has brought forth another blockbuster special issue on employment testing. For those predisposed to think of this subject as not too sexy, we should add that certain themes elaborated in the Journal are rousing animal passions along the Potomac.
A thought you could take away from the special issue is that policymakers will soon be driven to choose between two competing values in the workplace. Value No. 1 is economic efficiency. Value No. 2 is increased opportunity for minority workers. Under the gun nowadays is the U.S. Department of Labor, which seems reluctant to admit there is any tension between those values.
The Labor Department's problem begins with the fact that its U.S. Employment Service (USES) is the country's leading impresario of job testing. The service has developed the General Aptitude Test Battery (GATB), a quasi-IQ test used to predict workers' performance. The subtests that make up the battery are widely used by state employment services in deciding which workers to refer to employers with job openings. But there is a large embarrassment about the results: the black-white gap. On one major subtest, for example, the average white is a bit over the 50th percentile, the average black around the 35th.
For the past 20 years or so, the public-policy response to such data has been to assume, or possibly the word is "pretend," that the tests are flawed or biased. The assumption is built into guidelines adopted by the Equal Employment Opportunity Commission: These state firmly that group differences on employment tests constitute prima facie evidence of bias. The same assumption is also discernible in the adoption by USES of "race norming," which means putting white and minority testees into separate pools and ranking each individual only in relation to other members of the pool. Race norming is intended to raise minority scores, and does.
The Labor Department is in a somewhat anomalous position here. In supporting race norming, it seems to be implicitly admitting that the GATB is indeed biased. In fact, however, USES strongly endorses the tests and denies they are biased. Obvious underlying reality: USES wants to do as much employment testing as possible but thus far, at least, senses a political need for its tests to generate more minority hires.
The logical case for race norming is meanwhile getting weaker all the time, as the case for the tests' validity gets stronger. Researchers at USES and academic scholars have lately poured forth an avalanche of data supporting "validity generalization." That clunky phrase refers to several interrelated propositions: that tests of cognitive ability like the GATB are superior predictors of job performance, that they work equally well for whites and blacks, and that they predict better than job-specific aptitude tests. The new support for validity generalization is not just academic. Employers charged with adverse impact have recently been winning cases in which they rebutted the presumption of discrimination by pointing to their use of cognitive tests. The Journal includes an article by James C. Sharf of the U.S. Office of Personnel Management pointing up a string of wins for validity generalization in federal courts.
Now about the animal passions. These too are in evidence in the special issue, which features a presentation by Richard T. Seymour of the Lawyers' Committee for Civil Rights Under Law. Seymour derides and excoriates the experts supporting validity generalization and puts forward data to support his claim of boundless test bias. His claim is shot down (on our scoring, anyway) elsewhere in the issue, in a statistical argument unfortunately impossible to render
unless the management around here gives us another 100 lines. [Forget it -- The Management.]
Perhaps Mickey Kaus grew up on Keeping Up too?
However, we do insist on noting one small corner of the argument.
Seymour makes much of the fact that observed correlations between test scores and worker performance are not terribly high. In general, the correlations run around 0.3 (on a scale where 0 means no relationship at all between the two variables and 1.0 means a perfect relationship). With a correlation of 0.3, the test scores "explain" only 9% (the square of that number) of the variability in performance. Seymour argues that this percentage is far too small to be useful and says nobody in his right mind would invest in the stock market with so small an edge.
The answer (delivered by Robert A. Gordon of Johns Hopkins, Mary A. Lewis of PPG Industries, and Ann M. Quigley of the City of Tulsa Personnel Department) directs your attention to the economics of the casino industry. In roulette, the house edge at Monte Carlo is 2.7%. That is equivalent to a correlation of only 0.027 when a player bets red or black. And nobody in his right mind would expect the house to lose.