be a unitary trait that is highly heritable and normally distributed in the population, just like
height and weight.
Herrnstein and Murray (1994) extract this measure of general intelligence from the
ASVAB Test in the NLSY79. However, instead of using g directly in their regressions,
Herrnstein and Murray use the Armed Services Qualification Test (AFQT), which is composed
of four subtests of the ASVAB Test,7 as the measure of cognitive ability. They justify this
approximation of g by arguing that the AFQT tests are highly g-loaded, i.e. tests that have high
correlations with the extracted g.8 Thus, the AFQT scores behave sufficiently similar to g.
However, the assumption that human intelligence is unidimensional has been questioned
by other psychologists;9 economists have also cautioned that IQ test scores are at best an
imperfect proxy for human intelligence. Neal and Johnson (1996) demonstrate that, contrary to
the assumption that intelligence is exogenous and immutable, the AFQT test scores can be
affected by additional schooling. Heckman (1995) asserts that “AFQT is an achievement test that
can be manipulated by educational interventions” (Heckman [1995], p.1103). Fisher et al. (1996)
take a further step and assert that the AFQT is really a test that “largely tapped school and
school-like learning” (Fisher et al. [1994], p.62).
While the debate over g and its legitimacy in approximating innate intelligence is still
active today, most economists agree that IQ tests reflect some information about individual
abilities and are the best available measures of intelligence.10 Fisher et al. (1996) and Cawley et
7 The AFQT score is the sum of four ASVAB subtests, i.e., Word Knowledge, Paragraph Comprehension, Arithmetic Reasoning,
and Mathematics Knowledge.
8 Herrnstein and Murray (1994): 583. Correlations of AFQT subtests with g: Word Knowledge (.87), Paragraph Comprehension
(.81), Arithmetic Reasoning (.87), and Mathematics Knowledge (.82).
9 While recognizing the significance of the general intelligence g, psychologists also discovered the group factors of intelligence
that some, but not all, of the tests shared in common (see Thurstone [1947], Carrol [1993]). These group factors have less
explanatory power than g but nonetheless are both statistically and numerically significant in predicting test scores. Even
Spearman himself reluctantly admitted the existence of group factors : “We have now arrived at the ‘ group factors ’ which... are
of immense importance, not only theoretically, but also practically” (Spearman [1927], p.222-223).
10 See the discussions in Cawley et al. (1996), Ashenfelter and Rouse (1999), and Heckman (1995).