Strenze_Intelligence-and-socioeconomic-success_Intelligence-2007

Préparation du document



Intelligence 35 (2007) 401–426
Intelligence and socioeconomic success: A meta-analytic
review of longitudinal research ☆
Tarmo Strenze ⁎
Department of Sociology and Social Policy, University of Tartu, Tiigi 78-227, 50410 Tartu, Estonia
Received 5 April 2006; received in revised form 19 September 2006; accepted 20 September 2006
Available online 3 November 2006
Abstract
The relationship between intelligence and socioeconomic success has been the source of numerous controversies. The present
paper conducted a meta-analysis of the longitudinal studies that have investigated intelligence as a predictor of success (as measured
by education, occupation, and income). In order to better evaluate the predictive power of intelligence, the paper also includes metaanalyses
of parental socioeconomic status (SES) and academic performance (school grades) as predictors of success. The results
demonstrate that intelligence is a powerful predictor of success but, on the whole, not an overwhelmingly better predictor than parental
SES or grades. Moderator analyses showed that the relationship between intelligence and success is dependent on the age of the
sample but there is little evidence of any historical trend in the relationship.
© 2006 Elsevier Inc. All rights reserved.
Keywords: Intelligence; Socioeconomic success; Career success; Status attainment; Meta-analysis
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
2. A brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
3. Previous reviews . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
4. Topics addressed in the present paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
4.1. The size of the correlation between intelligence and success . . . . . . . . . . . . . . . . . . . . . . . . . . 404
4.2. Intelligence and other predictors of success. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
4.2.1. Intelligence versus parental SES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
4.2.2. Intelligence versus academic performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
4.3. Moderators of the correlation between intelligence and success . . . . . . . . . . . . . . . . . . . . . . . . 405
4.3.1. Age at the time of testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
4.3.2. Age at the measurement of success . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
4.3.3. Year of the measurement of success . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
☆ A detailed Appendix to this meta-analysis is available at www.zone.ee/tstrenze/meta.xls. The author wishes to express his gratitude to the
reviewers for their helpful comments.
* Tel.: +372 55 57 16 40; fax: +372 737 5900.
E-mail address: tstrenze@ut.ee.
0160-2896/$ - see front matter © 2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.intell.2006.09.004
402 T. Strenze / Intelligence 35 (2007) 401–426
5. Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
5.1. Definition of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
5.1.1. Socioeconomic success . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
5.1.2. Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
5.1.3. Parental SES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
5.1.4. Academic performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
5.2. Collection of data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
5.3. Correcting for unreliability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
5.3.1. Socioeconomic success . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
5.3.2. Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
5.3.3. Parental SES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
5.3.4. Academic performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
5.4. Correcting for range restriction and dichotomization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
5.5. Moderator variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
5.6. Meta-analytic calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
6. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
6.1. The meta-analytic database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
6.2. Predictors of socioeconomic success. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
6.3. Moderator analysis using subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
6.4. Moderator analysis using multiple regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
7. Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
7.1. Intelligence as a predictor of socioeconomic success . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
7.2. Possible limitations and implications for future research . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
1. Introduction
Decades of research on human mental abilities have
demonstrated that the scores of intelligence tests are
positively correlated with several desirable outcomes
and negatively correlated with several undesirable
outcomes. One of the central and personally most relevant
desirable outcomes is socioeconomic success (or
career success), which is usually measured by the educational
level, occupational prestige, and income of an
individual in adulthood. Although it is sometimes claimed
in popular press and textbooks that intelligence has
no relationship to important real-life outcomes (see
Barrett & Depinet, 1991, for a review of such claims), the
scientific research on the topic leaves little doubt that
people with higher scores on IQ tests are better educated,
hold more prestigious occupations, and earn higher
incomes than people with lower scores (Gottfredson,
1997, 2003; Jensen, 1980, 1998; Schmidt & Hunter,
2004).
Thus, the existence of an overall positive correlation
between intelligence and socioeconomic success
is beyond doubt. But quite surprisingly, the mere
existence of this correlation seems to be the only fact
that is established beyond doubt after many decades of
research. Several major questions are still without
definite answers and continue to arouse heated debates
(the debate about The Bell Curve being a prominent
example in recent decades; see Herrnstein & Murray,
1994; Fischer et al., 1996). First, what is the
approximate size of the correlation between intelligence
and success? Is it large enough to be of any
practical importance? While some researchers have
said that this correlation is “larger than most found in
psychological research” (Schmidt & Hunter, 2004:
162), others are convinced that “IQ is just not an
important enough determinant of economic success”
(Bowles & Gintis, 2002: 12). Second, how does the
predictive power of intelligence compare to the
predictive power of other variables, such as parental
socioeconomic status (SES) or school grades? On the
one hand there are studies showing that “individual
ability is by far the strongest influence on occupational
achievement” (Bond & Saunders, 1999: 217).
And yet other studies conclude that “the effect of
socioeconomic background on each of the three adult
status variables – schooling, income, and occupational
status – is greater than the effect of childhood
IQ” (Bowles & Nelson, 1974: 44). Third, are there
any age-related or historical changes in the relationship
between intelligence and success? The question
of historical changes in the importance of IQ has been
particularly controversial with some authors warning
against increasing cognitive stratification (Herrnstein
T. Strenze / Intelligence 35 (2007) 401–426
403
& Murray, 1994) and others trying to disprove these
claims (Hauser & Huang, 1997).
The present paper will address these questions by
conducting a meta-analysis of the longitudinal research
on the relationship between intelligence and socioeconomic
success. I will concentrate on longitudinal studies
(where intelligence is measured before the actual success)
because only longitudinal research design allows
us to make conclusions about the possible causal impact
of intelligence on success.
2. A brief history
Longitudinal studies on the relationship between intelligence
and career success have been conducted since
the first decades of the 20th century (Ball, 1938; Thorndike
et al., 1934). And these studies have invariably
uncovered a positive relationship. The early studies,
however, did not consider other possible determinants of
success, most importantly parental SES. Therefore, they
were open to the criticism that the positive relationship
between intelligence and success might actually be the
result of parental SES influencing them both (Bowles &
Gintis, 1976; McClelland, 1973). At the end of 1960s,
with the inception of the status attainment research
paradigm, investigators started to construct more sophisticated
models of career advancement that considered
several determinants of success at the same time
(Duncan, 1968; Jencks et al., 1972; Sewell, Haller, &
Ohlendorf, 1970).
But it was with the publication of The Bell Curve in
1994 (Herrnstein & Murray, 1994) that the question of
intelligence and socioeconomic success really came to
public attention. Analyzing a representative longitudinal
data set from the United States, Herrnstein and Murray
found that intelligence is a better predictor of several
desirable outcomes (e.g., not living in poverty, not being
arrested) than is parental SES. They also found evidence
that the role of intelligence in status attainment has been
growing throughout the 20th century and concluded that
the social structure of American society is increasingly
based on mental ability. The ideas of The Bell Curve
have been severely criticized for a number of reasons.
Fischer et al. (1996) argued that Herrnstein and Murray
used an inappropriate measure of parental SES and,
therefore, underestimated its importance. Hauser and
Huang (1997) argued that the claim about the growing
importance of intelligence is simply a misinterpretation
of previous research. Other researchers have, however,
supported the ideas of The Bell Curve (Gottfredson,
2003; Jensen, 1998) saying that its central claims have
been convincingly confirmed (Nyborg, 2003: 459).
At the same time in Great Britain, a similar discussion
was inspired by the work of Saunders who,
analyzing a representative longitudinal data set from
Great Britain, found that intelligence is a better
predictor of occupational attainment than is parental
SES and concluded that England is, to a large extent,
a meritocratic society (Bond & Saunders, 1999;
Saunders, 1997, 2002). These conclusions were
challenged by Breen and Goldthorpe (1999, 2001)
who argued that Saunders greatly overestimated the
importance of intelligence by using inappropriate
analytic techniques.
3. Previous reviews
There have been surprisingly few attempts to systematically
review the literature on intelligence and
socioeconomic success. Reviewers typically cite only a
couple of studies (see e.g., Brody, 1997; Farkas, 2003;
Schmidt & Hunter, 2004). Some of the most comprehensive
reviews have been conducted by Jencks (see
Jencks et al., 1972, 1979). Two meta-analyses have so
far addressed the relationship between intelligence and
socioeconomic success. Both of them used income as a
measure of success. The more comprehensive one of
the two was conducted by Bowles, Gintis, and Osborne
(2001). They assembled 65 estimates from 24 studies
to estimate the relationship between intelligence and
income. The mean standardized regression coefficient
of intelligence on income is .15 according to their study
(p. 1154). In addition to that, Bowles et al. (2001)
reported that there is no time trend in the size of the
coefficients between the years 1960 and 1995 and that
the age of the sample at the time of ability testing has
no effect on the results.
The meta-analysis of Bowles et al. is a valuable
contribution but it suffers from several shortcomings.
First, it considered only one measure of success, income,
thereby ignoring education and occupation. Second,
the meta-analytic estimate of .15 was not derived
from zero-order correlations as is usually required by
the textbooks of meta-analysis (see Hunter & Schmidt,
2004: 475) but from regression equations that included
several predictors in addition to intelligence. Peterson
and Brown (2005) have recently suggested that the use
of partial effect sizes, instead of zero-order ones, does
not affect the meta-analytic results very much but it is
nevertheless obvious that the use of disparate studies
makes the results difficult to interpret. Third, the metaanalysis
of Bowles et al. was not based on independent
samples. The authors stated that they used 65 estimates
from 24 studies (p. 1154) but neither of these figures
404 T. Strenze / Intelligence 35 (2007) 401–426
represents the number of independent samples. Inspection
of the appendix (not published but available from
the authors) leaves no doubt that some samples
contributed more than one coefficient to the final
meta-analytic estimate thereby ignoring the requirement
of independent data (see Hunter & Schmidt,
2004, chapter 10). Fourth, their meta-analysis mixed
cross-sectional and longitudinal studies. The distinction
between cross-sectional and longitudinal study design
is vital in the present context because only the latter can
answer questions about the causal impact of intelligence
on career success.
Another, more recent, meta-analysis was conducted
by Ng, Eby, Sorensen, and Feldman (2005) who
collected 8 studies and found an average correlation of
.27 between intelligence and salary. The meta-analysis
of Ng et al. (2005) was, unlike the one by Bowles et al.,
based on zero-order correlations and avoided the use of
non-independent samples but it failed to separate crosssectional
and longitudinal studies.
4. Topics addressed in the present paper
4.1. The size of the correlation between intelligence and
success
Meta-analyses are often conducted with the aim to
determine if a statistical relationship between two variables
is significantly different from zero. This cannot be
the only aim of the present meta-analysis because very
few social scientists would doubt that there is a positive
correlation between intelligence and socioeconomic
success. Having acknowledged that, the next logical
question is: what is the approximate size of the correlation?
Answers to this question are far from uniform.
Take the correlation between intelligence and income:
Jensen has suggested that it is somewhere around .40
(Jensen, 1998: 568) while Bowles et al. (2001) have
found that it is only about .15. That is why the first aim
of the present meta-analysis is to estimate the
approximate sizes of the correlations between intelligence
and measures of success. The importance of the
correlations can be evaluated using Cohen's classification
scheme which classifies correlations as small if
they are below .30, medium-sized if they are between
.30 and .50, and large if they are over .50 (Cohen,
1988). Knowing the size of the correlation between
intelligence and career success would allow us to
compare it to other, well-established, correlations in the
social scientific literature; e.g., the correlation of .51
between intelligence and job performance (Schmidt &
Hunter, 1998).
4.2. Intelligence and other predictors of success
It is difficult to evaluate the importance of a predictor
in isolation; it would be informative to compare the
predictive power of intelligence to the predictive power
of other relevant predictors of socioeconomic success.
This paper will, therefore, analyze two additional
predictors – parental SES (e.g., father's occupation)
and academic performance (e.g., school grades) – with
the aim to determine if intelligence is a better predictor
of success than the other two variables. Parental SES
and academic performance have often been treated as
the main “competitors” of intelligence in predicting
career success because, as explained shortly, they
represent different views about a typical path to success.
Including them in this paper will, consequently, allow us
to better evaluate the role of intelligence in people's
career.
4.2.1. Intelligence versus parental SES
The question about the relative importance of intelligence
and parental SES in predicting success is one of
the central questions of status attainment research. This
is a question about the nature of the society we live in:
whether a typical western society rewards people for
their own abilities or their social background (Saunders,
1997; Turner, 1960)? But we are far from having a
definite answer to this question. Some authors have
found that intelligence outcompetes parental SES as a
predictor (Herrnstein & Murray, 1994; Murray, 1998;
Saunders, 1997). Others have replied that parental SES,
if properly measured, is actually a better predictor
(Bowles & Nelson, 1974; Fischer et al., 1996). The
seemingly greater predictive power of intelligence in
some studies results from the failure to correct for measurement
error in the measures of parental SES (Bowles
& Nelson, 1974) and the failure to include important
aspects of parental status (most importantly, parental
income) among the predictors (Bowles & Nelson, 1974;
Fischer et al., 1996).
Therefore, it is necessary to compare the correlation
between intelligence and success with the correlation
between parental SES and success. To accomplish that,
the present paper will include a meta-analysis of the
relationship between the different aspects of parental
SES (parental education, occupation, and income) and
socioeconomic success. Research on this relationship is,
of course, voluminous and several narrative and quantitative
reviews of it are available (see Ganzeboom,
Luijkx, & Treiman, 1989; Haveman & Wolf, 1995;
Mulligan, 1999). Ganzeboom et al. (1989), for instance,
gathered 149 studies from 35 countries to analyze the
T. Strenze / Intelligence 35 (2007) 401–426
405
association between father's occupation and son's occupation,
and concluded that the association is stronger in
non-industrialized societies and has been weakening
during the 20th century. But none of these reviews has
presented the results in a manner that would make them
directly usable in this paper, hence the need for a
separate meta-analysis. 1
4.2.2. Intelligence versus academic performance
The question about the relative importance of mental
ability and academic performance in predicting success
has also been recognized as important (see Jencks &
Phillips, 1999). It is a question of what really matters for
career success: is it one's general ability (as measured
by IQ tests) or the things one has learned at school and
motivation to learn (as measured by school grades)? Not
many studies have explicitly compared the predictive
power of IQ scores and school grades (e.g., Taubman &
Wales, 1974, chapter 3). But the more general question
about the usefulness of grades as predictors of success
has been the object of considerable debate (see Roth,
BeVier, Switzer, & Schippmann, 1996; Roth & Clark,
1998). The meta-analysis by Roth and Clark (1998), for
instance, found an average correlation of .28 between
grades and salary. Thus, contrary to some earlier claims
(e.g., McClelland, 1973), grades have turned out to be
good predictors of success. This literature is somewhat
limited by being almost exclusively restricted to college
grades. If the purpose is to compare grades and IQ test
scores as predictors of career success, then high school
grades would be a better choice because college
students constitute a rather selected group that does
not represent the full range of career attainments in
society. High school grades have been meta-analytically
related to job performance (Dye & Reck, 1988) and
college grades (Robbins et al., 2004) but there is
currently no meta-analysis about the relationship
between high school grades and general socioeconomic
success (as measured by education, occupation, and
income). The present paper will, thus, conduct such a
meta-analysis.
4.3. Moderators of the correlation between intelligence
and success
In order to further clarify the role of intelligence in
people's career, the effects of three moderator variables
1 Ganzeboom et al. (1989) analyzed social mobility tables, Mulligan
(1999) analyzed bivariate unstandardized regression coefficients; for
this paper, however, bivariate standardized regression coefficients
(i.e., correlations) are needed.
(age at testing, age at success, and year of success) on
intelligence–success correlation will be studied. These
moderator variables have been analyzed in several
studies but with rather conflicting results.
4.3.1. Age at the time of testing
The first moderator analysis concerns age at testing
(age of individuals at the time the IQ test was taken)
and how it affects the correlation between intelligence
and success. Analysis of the effect of age at testing
reveals something about the mechanism behind the
intelligence–success correlation. If intelligence predicts
success irrespective of the age at which it is measured,
then there is reason to believe that the differences in
people's career success are the result of the stable
individual differences measured by IQ tests (Jencks &
Phillips, 1999). If however, the predictive power of IQ
tests changes with age, then different interpretations
are possible depending on how we believe the test
score to be affected by genes and environment.
According to the standard sociological interpretation,
the test scores of older individuals should be more
affected by life experiences than the scores of children
(because older individuals simply have had more
experiences) and consequently, if intelligence tested
at an older age should turn out to be a better predictor
career success, then it would mean that the test scores
probably reflect some career-relevant experiences
which the older individuals have had more time to
accumulate (Jencks & Phillips, 1999). The other
interpretation is based on behavior genetic research
which has found that genetic influences on IQ scores
increase with age and environmental influences
decrease (McCartney, Harris, & Bernieri, 1990); from
these results one can conclude that if the test scores of
older individuals are better predictors of success, then
it can be attributed to the growing effect of some
career-relevant genes. 2
Empirical evidence concerning age at testing is rather
contradictory. A study by McCall (1977) found a clear
upward trend in the correlations between intelligence
and success; that is, correlations grew stronger as age at
testing increased. Some of the studies reviewed by
Jencks and Phillips (1999) have found a similar trend.
The meta-analysis of Bowles et al. (2001), however,
found that age at testing has no effect on the association
between intelligence and income. Jencks et al. (1979)
reached a similar conclusion in their review.
2 I am grateful to a reviewer for pointing this interpretation out to
me.
406 T. Strenze / Intelligence 35 (2007) 401–426
4.3.2. Age at the measurement of success
A related issue concerns the age of the individuals at
the time their career success is measured. According to
the so-called gravitational hypothesis, the correlation
between intelligence and success should grow stronger
as individuals grow older because (as a result of selfselection
and competition) individuals “gravitate”
towards the positions that correspond to their ability
levels as they progress in their careers. This reasoning
has been used to support the idea that intellectual
differences cumulate over life course and become
progressively more important (see Gottfredson, 2003;
Wilk & Sackett, 1996). Other researchers have
suggested that exactly the opposite is true: the
predictive validity of IQ scores should decline as time
goes by because less able people have time to
accumulate skills to compensate for their initial lack
of ability (Ackerman, 1987; Keil & Cortina, 2001).
These opposing views can be reconciled by saying that
the idea of declining importance of IQ applies to the
performance of specific tasks that become automatic
after some practice and the idea of growing importance
applies complex long term activities, such as attaining
and maintaining social status, that never cease to be
cognitively demanding. But so far as socioeconomic
success can depend on the performance of specific work
tasks, the possibility of declining validity of IQ is not
completely ruled out.
Several studies have correlated intelligence with
success at different points in people's life course. Some
of them have found that the correlations indeed increase
with age as predicted by the gravitational hypothesis
(Brown & Reynolds, 1975; Deary et al., 2005; Wilk &
Sackett, 1996), others have found no clear trend (Hauser,
Warren, Huang, & Carter, 1996; Warren, Sheridan,
& Hauser, 2002). The reviews by Hulin, Henry, and
Noon (1990) and Keil and Cortina (2001) found support
for the declining validity thesis but it should be noted
that many of the studies reviewed in these papers used
specific laboratory tasks as dependent variables and are,
therefore, not directly comparable to the studies reviewed
in the present paper.
4.3.3. Year of the measurement of success
A particularly controversial issue concerns the historical
changes in the relationship between intelligence and
success. It was one of the central claims of The Bell
Curve that the association between mental ability and
career success in western societies has been growing
throughout the 20th century (Herrnstein & Murray,
1994). The logic behind this idea is similar to the gravitational
hypothesis, discussed in the previous section –
in both cases individuals are increasingly drawn towards
the positions that correspond to their ability as time goes
by – but in this case the gravitation does not take place
during a life course of a single individual but over several
generations.
Several studies have investigated changes in the
association between intelligence and success during past
decades. Although Herrnstein and Murray concluded
that “the main point seems beyond dispute” (1994: 52)
and some studies have found support for this point
(Murnane, Willett, & Levy, 1995), there are still serious
reasons to doubt that the importance of intelligence is or
has been growing. Neither the meta-analysis by Bowles
et al. (2001) nor the review by Jencks et al. (1979) found
any clear trend in the correlations between intelligence
and success. The same conclusion was reached by Flynn
(2004) and Hauser and Huang (1997). Breen and
Goldthorpe (2001) found that the association between
intelligence and occupational status in England is, if
anything, declining.
5. Method
5.1. Definition of variables
The present meta-analysis investigated the relationship
between three measures of socioeconomic success
(educational level, occupational level, and income) and
three predictors (intelligence, parental SES, and academic
performance). The operationalization of these variables
is described next.
5.1.1. Socioeconomic success
Educational level was measured by the number of years
spent in full time education or the highest level of education
completed. Occupational level was typically measured by
such occupational scales as Duncan Socioeconomic Index,
International Socioeconomic Index of Occupational Status,
NORC prestige scale, etc. These scales provide detailed
numerical measures of occupational status (see Ganzeboom
& Treiman, 1996a, for a general discussion). In some
studies, less detailed occupational classifications were
used. Irrespective of the level of detail, all the occupational
variables in this paper had a common property of ordering
occupations on a single hierarchical dimension with higher
values designating more desirable and prestigious occupations.
Income was measured by salary or total monetary
income, which had to refer to the personal income of an
individual, not to family or household income. If possible, I
preferred income measured on a logarithmic scale because
logarithmic transformation removes the skew typically
found in income distribution.
T. Strenze / Intelligence 35 (2007) 401–426
407
5.1.2. Intelligence
Intelligence, or general mental ability, of an individual
was measured by a score on a test of intelligence. It
is not always easy, however, to decide if a given test is a
test of intelligence. The definitions of intelligence state
that it is an abstract ability that is not tied to any specific
domain of knowledge. Therefore, only the tests that are
designed to measure such ability should be used in the
meta-analysis. If we take the traditional threefold
distinction between ability, aptitude, and achievement
tests (Jensen, 1981), then the present study should use
only ability and aptitude test scores. Although some
researchers have contended that achievement tests can
also be treated as measures of general ability (Boudreau,
Boswell, Judge, & Bretz, 2001), and even everyday life
can be interpreted as an IQ test (Gordon, 1997), the
present study took a more conservative approach and
included only those tests that are generally regarded as
tests of intelligence (see e.g., Anastasi & Urbina, 1997;
Jensen, 1980, chapter 7, for a discussion and classification
of different tests).
There are numerous “classical” tests (e.g., Henmon–
Nelson, Lorge–Thorndike, Otis–Lennon, Raven Progressive
Matrices, Stanford–Binet, Wechsler tests) for
which there seems to be a general consensus that
these are indeed tests of general mental ability. Such
multiple aptitude test batteries as Armed Services
Vocational Aptitude Battery or General Aptitude Test
Battery are also often treated as measures of general
ability. The most problematic ones are the tests that
are specifically constructed for use in a single data
set. Such unique tests have been used in several large
and influential data sets (e.g., National Child
Development Study, National Longitudinal Survey of
High School Class 1972, Panel Study of Income
Dynamics, Project Talent). In these cases I consulted
the manuals of the data sets and studies that are based on
the data. If the test was derived from other IQ tests or if it
was described as a test of intelligence, then I included it
in my study. Studies using well-known achievement
tests, such as Iowa Test of Basic Skills (Smokowski,
Mann, Reynolds, & Fraser, 2004), were excluded. The
names of all the IQ tests used in this paper are listed in the
Appendix.
5.1.3. Parental SES
Five measures of parental socioeconomic status
(SES) were used in this paper; the first four were
father's education, mother's education, father's occupation,
and parental income. The measurement of these
variables was similar to the measurement of respondent's
own education, occupation, and income (see
above). Parental income refers to father's income or
total income of parents. Because too few studies
reported data on mother's occupation, this variable
was not included. In addition to these four, I also used
a general index of SES, which combines several
parental characteristics into one variable. A number of
studies have used a composite index on the assumption
that it is a better indicator of social advantages
than any of the single variables that make up the index
and, therefore, also a better predictor of success (see
White, 1982, for supporting evidence). A correlation
with SES index was included in the present metaanalysis
if the index was composed of the following
components — parental education (education of one
or both parents), parental occupation (occupation of
one or both parents), and material well-being of the
parental home. The latter was measured by parental
income or by a “possession index” which indicates
how many of the valued items (e.g., a car, TV set,
computer) were present at home. If a study did not use
an index of SES but presented intercorrelations among
the necessary variables, then I used the formulas
reported by Hunter and Schmidt (2004: 433) to
calculate a composite score correlation between SES
index and success.
5.1.4. Academic performance
Academic performance was in most studies measured
by a grade point average (GPA) obtained in high
school or the years preceding high school. In some
studies, rank in class (i.e., how well the student
performed in comparison with other students in the
class) was used instead of GPA. Rank is generally used
interchangeably with GPA (see Kuncel, Crede, &
Thomas, 2005), therefore, these studies were also
included.
5.2. Collection of data
Studies were identified for inclusion in the metaanalysis
by searching computerized databases (such as
JSTOR, PsycINFO) using terms like “status attainment”,
“educational attainment” “occupational attainment”,
“socioeconomic achievement” as keywords.
Reference sections of review papers were also searched.
To be included in the meta-analysis, the following
general criteria had to be met. First, the measurement of
the variables had to correspond to the descriptions
presented in Section 5.1. Second, the data had to be
longitudinal; that is, the predictors (intelligence, parental
SES, and academic performance) had to be measured at
an earlier time and career success (education, occupation,
408 T. Strenze / Intelligence 35 (2007) 401–426
and income) at a later time. 3 Third, the interval between
the measurement of predictors and dependent variables
had to be at least 3 years because studies with shorter
intervals would have very little advantage over crosssectional
studies. Fourth, the study had to report a zeroorder
correlation between the variables and another
measure of association transformable into a zero-order
correlation.
Fifth, majority of individuals in the sample had to be at
least 20 years old at the time the career success was
measured because it makes little sense to talk about the
career success of individuals younger than 20. Sixth,
majority of individuals had to be less than 25 years old at
the time the IQ test was taken because, to properly
investigate the effect of intelligence on career, intelligence
should be tested before the individuals start a career.
Obviously, even individuals tested in their early twenties
might already have started a career, but since these
individuals can be used for comparison with younger
individuals, they were included. Information on parental
SES and academic performance had to refer to the time the
respondents were approximately 12–18 years old (the
time these variables presumably have their greatest impact
on subsequent career). Seventh, the study had to be
conducted in a “western” society; that is, in the United
States, Canada, Europe, Australia, or New Zealand. Additional
criteria are described in Section 5.4.
It is rather common for published studies not to
report the information necessary for meta-analysis (the
lack of zero-order correlations is a typical problem). But
fortunately, the raw data of several well-known data sets
(e.g., General Social Survey, National Longitudinal
Survey of Youth) are available for public use. Because it
would be a serious waste of information to leave these
sources unused, I decided to use the available raw data
to calculate the correlations if none of the published
sources reported the necessary information or if the
information in the published source was deficient in
some way (e.g., if the correlations were reported separately
for men and women but not for the complete
3 That does not mean that all the studies had to actually include at
least two waves of measurement because one of the predictors,
parental SES, can be measured retrospectively (by asking adult
respondents questions like “what was your father's occupation at the
time you were 16?”). It is important, however, that the information
about father's occupation or parental income obtained from adult
respondents refers to parents' past (not current) occupation or
income. The latter requirement was not applied to parental education
because parents' education is unlikely to change while children grow
up. In some studies (e.g. Duncan & Hodge, 1963), father's occupation
was rather vaguely referred to as father's “usual occupation” or
“longest occupation”. These studies were also included.
sample). Most of the raw data sets had been prepared for
public use and contained all the necessary variables in a
ready-to-use form. In some cases, minor statistical procedures
were implemented before calculating the correlations
(e.g., summing the standardized scores of subtests
to obtain the score of general intelligence; transforming
the original occupational variable into a more appropriate
prestige scale using the methodological tools provided by
Ganzeboom and Treiman (1996b). The raw data sets used
in this paper are listed in the Appendix.
Several longitudinal data sets contain data from more
than one follow up. Career success has been measured
repeatedly for the same individuals in these data sets
(up to 20 times in some cases). In some data sets,
the predictors (intelligence, parental SES, or academic
performance) have also been measured repeatedly. In
order to ascertain that every sample contributed only
one correlation to one analysis, I averaged all the correlations
that were derived from the same sample. If
the sample sizes of the averaged correlations were
different, mean sample size was used. The procedures
for moderator analyses are described in Sections 6.3
and 6.4.
5.3. Correcting for unreliability
Ideally, every correlation should be corrected with the
reliability coefficients obtained from the same sample as
the correlation that needs to be corrected (Hunter &
Schmidt, 2004, chapter 3). However, such reliability
coefficients were available for only a small minority of
studies included in the present meta-analysis. The correlations
from these studies were corrected with these
reliability coefficients. But for the majority of studies,
mean reliabilities (estimated from various sources, as
described below) were used. Each study was then corrected
individually with the appropriate reliability coefficient.
The nature and sources of reliability information
are described next.
5.3.1. Socioeconomic success
Information on education, occupation, and income
can be obtained from three sources. The first source is
institutional record (e.g., tax records of income). Following
the common practice, data from such objective
sources were assumed to have a reliability of 1. The
second and by far the most common source is self-report.
Self-reports are not perfectly reliable, however. The
amount of error is usually measured by asking the
same individuals to report their socioeconomic characteristics
again after a few months and then correlating the
first and second reports (producing a test–retest
T. Strenze / Intelligence 35 (2007) 401–426
409
correlation). Several estimates of these correlations,
derived from nationally representative samples, are
available. Using the data presented by Bowles (1972)
and Jencks et al. (1979), I calculated the average test–
retest correlations for educational level (.89), occupational
status (.88), and income (.83). These values were used to
correct the correlations with self-reported socioeconomic
success. The third possibility is to obtain the information
from the spouse, parent, sibling, or child of the focal
individual. Because these sources would introduce
unnecessary complications and an unknown degree of
error, the studies using these sources were excluded
unless they contained correlations between intelligence
and success (these correlations are too valuable to be
discarded). 4
5.3.2. Intelligence
When correcting for unreliability in the test scores,
test–retest alternate-form reliability (the correlation
between parallel forms of the same test administered
on two separate occasions) is generally considered to be
the most appropriate form of reliability (Schmidt &
Hunter, 1999). But since these coefficients are rarely
available, simple test–retest reliability is often taken as
the second best option in meta-analytic studies of the
predictive power of IQ scores (Judge, Colbert, & Ilies,
2004; Salgado, Anderson, Moscoso, Bertua, & Fruyt,
2003). Because test–retest reliability coefficients were
reported in only a few studies, an average test–retest
coefficient, obtained from the meta-analysis of Salgado
et al. (2003), was used for most of the studies. Salgado
et al. averaged 31 test–retest correlations of different
general mental ability tests (the mean interval between
test and retest being 6 months) and obtained an average
coefficient of .83. This value is similar to average test–
retest correlations obtained in other reviews: e.g., .82 in
Parker, Hanson, and Hunsley (1988) or .85 in Kuncel,
Hezlett, and Ones (2004). Thus, the reliability of .83
seems to be a representative estimate and was used in the
present study.
4 In the study by Vroon, Leeuw, and Meester (1986) the dependent
variable (occupation) was reported by the child of the focal
respondent. For this study the reliability of children's report on
father's occupation was used.
5.3.3. Parental SES
The information on parental education, occupation,
and income can come from three sources. First, it can
be reported by the parents themselves. If this was the
case, then the correlations were corrected with the
same reliability coefficients that were used for selfreported
education, occupation, and income. Second, it
can be reported by the children. Children's reports on
parental characteristics are known to suffer from
considerable error. Probably the best estimate of this
error is the correlation between child's report and
parent's own report on a given characteristic. Looker
(1989) has presented a comprehensive review of these
correlations for father's education, mother's education,
and father's occupation. Using the information in Table
3 in Looker's paper, I calculated the average correlations
between child's report and parent's report. The
average correlations are .80 for father's education, .79
for mother's education, and .78 for father's occupation.
These values were used to correct the correlations that
involved children's reports on parental SES. 5 Information
on the reliability of children's reports on parental
income is harder to find. I could locate two studies
(Bell, Senese, & Elliott, 1984; Massagli & Hauser,
1983) that provided reliability estimates from three
samples. The estimates ranged from .45 to .59. with an
average of .51 that was used in the present paper. The
third source of information on parental SES is
objective data (e.g., tax records of income) that was
assumed to have a reliability of 1. Internal consistency
method (Cronbach alpha) was used to correct for
unreliability in the SES index. This method was
recommended by Hunter and Schmidt (2004: 438)
for composite variables. For all but two studies, the
alpha value of the SES index was obtained from the
same sample as the correlation itself. For the remaining
two, the average alpha of all the other studies (.71) was
used.
5.3.4. Academic performance
If the information on academic performance (GPA
or class rank) was obtained from school records, it was
assumed to have a reliability of 1. Students' selfreports
on their GPA or rank are, of course, not
perfectly reliable. The reliability of self-reports is
assessed by the correlation between self-reported GPA
(or rank) and GPA (or rank) obtained from school
records. A recent meta-analysis by Kuncel et al. (2005)
found that this correlation is .82 for high school GPA
and .77 for high school rank. These values were used
to correct the correlations that involved self-reported
GPA or rank.
5 Children's reports on parental SES can further be divided
according to the age of the child at the time of reporting. When
calculating the average reliabilities from Looker's (1989) data, I
excluded the samples of children younger than 9th grade because the
reports of such children were not used in the present meta-analysis.