Using IQ as a predictor

Back in October, CIP posted some stuff about IQ as a predictor, and in particular, this quote:

The upshot of this research is that general mental ability (IQ and related tests) was the best predictor and work sample tests (e.g., seeing if people can actually do key elements of a job -- if a secretary can type or a programmer can write code ) were the second best of the 19 examined.

Well, if you examine the comments thread, you will find if you chase down the original research results, IQ is second best, after work sample tests. But there was something else, that did not strike me with full force back then.

Look up the Handbook of Industrial, Work and Organizational Psychology: Personnel psychology,by Neil Anderson on books.google.com (g is the supposed real life "thing" that IQ purports to measure.)

Schmidt, Gooding, Noe and Kirsch (1984) performed a 'bare bones' meta-analysis (McDaniel, Hirsh, Schmidt, Raju & Hunter, 1986) of the predictive efficiency of g for job performance. Schmidt et. al. reported an average validity of .248. Ree and Carretta (1998) corrected this value for range restriction and predictor and criterion unreliability using the meta-analytically derived default values in Raju, Burke, Normand and Langlois (1991). After correction, the estimated true correlation between g and job performance was .512.

What is the "corrected for range restriction"? A comment on this blogpost explains (context: a study of eight graders at a school showed that school performance correlated much better with measures of self-discipline than with IQ).

this study covered only eighth graders in a relatively privileged school.
I was thinking this during the whole post, and then here's confirmation. In statistics, this is called "restriction of range" for some variable, here IQ (because the school is privileged, indicating that students of low or average IQ are unlikely to be found in the expected proportions). The less a variable varies, the less power it has to account for some outcome.

The point is that if you include the full range of human IQs, from the severely retarded to the highest intelligence, there is a stronger correlation between IQ and performance (at work, or at school or whatever) than in a limited population.

But think about it - when using IQ as a practical predictor, e.g., for college admissions or for selection for a job, you always, in practice have "restriction of range" - the candidates you are considering all are pretty much very similar. So, unless you are dealing with conscripts - involuntary applicants - or are a social engineer on a grand scale, IQ is, precisely because of the range restriction, a pretty useless predictor. This sentence is thus total B.S.

After correction, the estimated true correlation between g and job performance was .512.