I’ve discussed this before but given that I now have better data on the old SAT, I thought I’d revisit it.

Because the SAT has generally only been given to the college bound elite, nobody really knows how well it correlates with IQ in the general U.S. population.

But because it’s generally assumed that roughly 100% of Americans capable of scoring very high on the SAT actually take the SAT, I decided that the best way to estimate the SAT’s correlation with IQ among all American young adults is to find a subset of Americans selected for extremely high SAT scores and then see how they score on official IQ tests.

Because people selected for high SAT scores will regress to the U.S. mean on IQ in direct proportion with the correlation between the SAT and IQ in the full U.S. population (if all Americans young adults took the test), I use the degree to which they regress as an estimate for said correlation (assuming a bivariate normal distribution).

Perhaps the single best study was referred to me by a commenter named Andrew.  In this study, data was taken from the older more difficult SAT, and participants took the full-original WAIS.  In this study, six samples of  seniors from  the extremely prestigious Dartmouth (the 12th most selective university in America) averaged 1357 on the SAT just before 1974. Based on my latest research, an SAT score of 1357 if all American 17-year-olds circa 1974  had taken the SAT, the mean score would have been 770 (SD = 223) which means 1357 would have equated to an IQ of 139 (U.S. norms); 138 (U.S. white norms).

Assuming these students are typical of high SAT Americans, it is interesting to ask how much they regress to the mean on various subtests of the WAIS.

Averaging all six samples together, and then adjusting for the yearly Flynn effect from the 1950s through the 1970s (see page 240 of Are We Getting Smarter?) since the WAIS was normed circa 1953.5 but the students were tested circa 1971.5, then converting subtest scaled scores to IQ equivalents, in both U.S. norms and U.S. white norms (the 1953.5 norming of the WAIS included only whites), we get the following:

 iq equivalent (u.s. norms) iq equivalent (u.s. white norms) estimated correlation with sat in the general u.s. population inferred from regression to the mean from SAT IQ 44 points above U.S. mean. sat score 139 138 39/39 = 1.0 wais information 128.29 127.2 28.29/39 = 0.73 wais comprehension 122.22 120.9 22.22/39 = 0.57 wais arithmetic 120.37 119 20.37/39 = 0.52 wais similarities 119.16 117.75 19.16/39 = 0.49 wais digit span 117.37 115.9 17.37/39 = 0.45 wais vocabulary 125.93 124.75 25.93/39 = 0.66 wais picture completion 105.87 104 5.87/39 = 0.15 wais block design 121.82 120.5 21.82/39 = 0.56 wais picture arrangement 108.33 106.55 8.33/39 = 0.21 wais object assembly 113.65 112.05 13.65/39 = 0.35 wais verbal scale 126 125 26/39 = 0.67 wais performance scale 116 114 16/39 = 0.41 wais full-scale 123 122 23/39 = 0.59

Conclusion

The degree of regression from the SAT to the WAIS in an extreme sample suggests a 0.59 correlation between the two tests in the general U.S. population.

The SAT seems to be a reasonable proxy for verbal IQ (an implied correlation of 0.67) but only a moderate proxy for performance IQ (an implied correlation of 0.41). There are generations of visually brilliant people who did not get into a good college thanks to the SAT, but nonetheless may have earned high incomes as artists, tradesmen, fashion designers and film makers.

Not surprisingly, the SAT correlated best with the Information and Vocabulary subtests (implied correlations of 0.73 and 0.66 respectively) since these are direct measures of facts taught in school.

The real question is how well does the SAT correlate with g (the general factor of IQ tests)

Arthur Jensen argued that the correlation between two tests is a product of their factor loadings, so if the WAIS correlates 0.9 with g, and correlates 0.59 with the SAT, then the SAT should correlate 0.59/0.9 = 0.66 with g.

However since the SAT and WAIS share other factors besides just g (i.e. verbal), then 0.66 might overestimate the g loading.

On the other hand some would argue that deducing the regression slope from only high ability people might underestimate the correlation between the SAT and WAIS (ceiling bumping, Spearman’s Law of Diminishing Returns), and correcting WAIS scores for the Flynn effect might be overcorrections particularly at the high end.

Perhaps both factors cancel out, leaving 0.66 as a good estimate of the SAT’s g loading in the general U.S. population.

Is the SAT an IQ test?

If one defines an IQ test as any psychometric test with a high g loading in a general national or global population, and if one defines a high g loading as 0.7+, I’d say the SAT is a borderline IQ test.