The Bell Curve: Intelligence and Class Structure in American Life (15 page)

We begin with a review of the evidence that an important statistical link between IQ and job performance does in fact exist. In reading the discussion that follows, remember that job performance does vary in the real world, and the variations are not small. Think of your own workplace and of the people who hold similar jobs. How large is the difference between the best manager and the worst? The best and worst secretary? If your workplace is anything like ours have been, the answer is that the differences are large indeed. Outside the workplace, what is it worth to you to have the name of a first-rate plumber instead of a poor one? A first-rate auto mechanic instead of a poor one? Once again, the
common experience is that job performance varies widely, with important, tangible consequences for our everyday lives.

Nor is variation in job performance limited to skilled jobs. Readers who have ever held menial jobs know this firsthand. In restaurants, there are better and worse dishwashers, better and worse busboys. There are better and worse ditch diggers and garbage collectors. People who work in industry know that no matter how apparently mindless a job is, the job can still be done better or worse, with significant economic consequences. If the consequences are significant, it is worth knowing what accounts for the difference.

Job performance may be measured in many different ways.
¹⁵
Sometimes it is expressed as a natural quantitative measure (how many units a person produces per hour, for example), sometimes as structured ratings by supervisors or peers, sometimes as analyses of a work sample. When these measures of job productivity are correlated with measures of intelligence, the overall correlation, averaged over many tests and many jobs, is about .4. In the study of job performance and tests, the correlation between a test and job performance is usually referred to as the
validity
of the test, and we shall so refer to it for the rest of the discussion.
¹⁶
Mathematically, validity and the correlation coefficient are identical. Later in the chapter we will show that a validity of .4 has large economic implications, and even validities half as large may warrant worrying about.

This figure of .4 is no more than a point of reference. As one might expect, the validities are higher for complex jobs than for simple ones. In Edwin Ghiselli’s mammoth compilation of job performance studies, mostly from the first half of the century, a reanalysis by John Hunter found a mean validity of .53 for the job family labeled “manager” and .46 for a “trades and crafts worker.” Even an “elementary industrial worker” had a mean validity of 37.
¹⁷

The Ghiselli data were extremely heterogeneous, with different studies using many different measures of cognitive ability, and include data that are decades old. A more recent set of data is available from a meta-analysis of 425 studies of job proficiency as predicted by the General Aptitude Test Battery (GATB), the U.S. Labor Department’s cognitive ability test for the screening of workers. The table below summarizes the results of John and Ronda Hunter’s reanalysis of these databases.
¹⁸

The average validity in the meta-analysis of the GATB studies was 45
¹⁹
The only job category with validity lower than .40 was the industrial
category of “feeding/offbearing”—putting something into a machine or taking it out—which occupies fewer than 3 percent of U.S. workers in any case. Even at that bottom-most level of unskilled labor, measured intelligence did not entirely lose its predictiveness, with a mean validity of .23.

The third major database bearing on this issue comes from the military, and it is in many ways the most satisfactory. The AFQT (Armed Forces Qualification Test) is extracted from the scores on several tests that everyone in the armed forces takes. It is an intelligence test, highly loaded on g. Everyone in the military goes to training schools, and everyone is measured for training success at the end of their schooling, with “training success” based on measures that directly assess job performance skills and knowledge. The job specialties in the armed forces include most of those found in the civilian world, as well a number that are not (e.g., combat). The military keeps all of these scores in personnel files and puts them on computers. The resulting database has no equal in the study of job productivity.

We will be returning to the military data for a closer look when we turn to subjects for which they are uniquely suited. For now, we will simply point out that the results from the military conform to the results in the civilian job market. The results for training success in the four major
job families are shown in the table above. These results are based on results from 828 military schools and 472,539 military personnel. The average validity was .62. They hold true for individual schools as well. Even the lowest-validity school, combat, in which training success is heavily dependent on physical skills, the validity was still a substantial .45.
²⁰

The lowest modern estimate of validity for cognitive ability is the one contained in the report by a panel convened by the National Academy of Sciences,
Fairness in Employment Testing.
²¹
That report concluded that the mean validity is only about .25 for the GATB, in contrast to the Hunter estimate of .45 (which we cited earlier). Part of the reason was that the Hartigan committee (we name it for its chairman, Yale statistician John Hartigan), analyzing 264 studies after 1972, concluded that validities had generally dropped in the more recent studies. But the main source of the difference in validities is that the committee declined to make any correction whatsoever for restriction of range (see above and note 6). It was, in effect, looking at just the tackles already in the NFL; Hunter was considering the population at large. The Hartigan committee’s overriding concern, as the title of their report
(Fairness in Employment Testing)
indicates, was that tests not be used to exclude people, especially blacks, who might turn out to be satisfactory workers. Given that priority, the committee’s decision not to correct for restriction of range makes sense. But failing to correct for restriction of range produces a misleadingly low estimate of the overall relationship of IQ to job performance and its economic consequences.
²²
Had the Hartigan committee corrected for restriction of range; the estimates of the relationship would have been .35 to .40, not much less than Hunter’s.

Why are job performance and cognitive ability correlated? Surgeons, for example, will be drawn from the upper regions of the IQ distribution. But isn’t it possible that all one needs is “enough” intelligence to be a surgeon, after which “more” intelligence doesn’t make much difference? Maybe small motor skills are more important. And yet “more” intelligence always seems to be “better,” for large groups of surgeons and every other profession. What is going on that produces such a result?

As we begin to explore this issue, the story departs more drastically from the received wisdom. One obvious, commonsense explanation is that an IQ test indirectly measures how much somebody knows about the specifics of a job and that that specific knowledge is the relevant thing to measure. According to this logic, more general intellectual capacities are beside the point. But the logic, however commonsensical, is wrong. Surprising as it may seem, the predictive power of tests for job performance lies almost completely in their ability to measure the most general form of cognitive ability,
g,
and has little to do with their ability to measure aptitude or knowledge for a particular job.

S
PECIFIC
S
KILLS
V
ERSUS G IN THE
M
ILITARY.
The most complete data on this issue come from the armed services, with their unique advantages as an employer that trains hundreds of thousands of people for hundreds of job specialties. We begin with them and then turn to the corresponding data from the civilian sector.

In assigning recruits to training schools, the services use particular combinations of subtests from a test battery that all recruits take, the Armed Services Vocational Aptitude Battery (ASVAB).
²³
The Pentagon’s psychometricians have tried to determine whether there is any practical benefit of using different weightings of the subtests for different jobs rather than, say, just using the overall score for all jobs. The overall score is itself tantamount to an intelligence test. One of the most comprehensive studies of the predictive power of intelligence tests was by Malcolm Ree and James Earles, who had both the intelligence test scores and the final grades from military school for over 78,000 air force
enlisted personnel spread over eighty-nine military specialties. The personnel were educationally homogeneous (overwhelmingly high school graduates without college degrees), conveniently “controlling” for educational background.
²⁴

What explains how well they performed? For every one of the eightynine military schools, the answer was
g—
Charles Spearman’s general intelligence. The correlations between
g
alone and military school grade ranged from an almost unbelievably high .90 for the course for a technical job in avionics repair down to .41 for that for a low-skill job associated with jet engine maintenance.
²⁵
Most of the correlations were above .7. Overall,
g
accounted for almost 60 percent of the observed variation in school grades in the average military course, once the results were corrected for range restriction (the accompanying note spells out what it means to “account for 60 percent of the observed variation”).
²⁶

Did cognitive factors other than
g
matter at all? The answer is that the explanatory power of
g
was almost thirty times greater than of all other cognitive factors in ASVAB combined. The table below gives a sampling of the results from the eighty-nine specialties, to illustrate the two commanding findings:
g
alone explains an extraordinary proportion of training success; “everything else” in the test battery explained very little.

An even larger study, not quite as detailed, involving almost 350,000 men and women in 125 military specialties in all four armed services, confirmed the predominant influence of
g
and the relatively minor further predictive power of all the other factors extracted from ASVAB scores.
²⁷
Still another study, of almost 25,000 air force personnel in thirty-seven different military courses, similarly found that the validity of individual ASVAB subtests in predicting the final technical school grades was highly correlated with the
g
loading of the subtest.
²⁸

E
VIDENCE FROM
C
IVILIAN
J
OBS.
There is no evidence to suggest that military jobs are unique in their dependence on
g.
However, scholars in the civilian sector are at a disadvantage to their military colleagues; nothing approaches the military’s database on this topic. In one of the few major studies involving civilian jobs, performance in twenty-eight occupations correlated virtually as well with an estimate of
g
from GATB scores as it did with the most predictively weighted individual sub test scores in the battery.
²⁹
The author concluded that, for samples in the range of 100 to 200, a single factor,
g,
predicts job performance as well as, or better than, batteries of weighted subtest scores. With larger samples, for which it is possible to pick up the effect of less potent influences, there may be some modest extra benefit of specialized weighted scores. At no level of sampling, however, does
g
become anything less than the best single predictor known, across the occupational spectrum. Perhaps the most surprising finding has been that tests of general intelligence often do better in predicting future job performance than do contrived tests of job performance itself. Attempts to devise measures that are specifically keyed to a job’s tasks—for example, tests of filing, typing, answering the telephone, searching in records, and the like for an office worker—often yield low-validity tests, unless they happen to measure
g,
such as a vocabulary test. Given how pervasive
g
is, it is almost impossible to miss it entirely with any test, but some tests are far more efficient measures of it than others.
³⁰