Content uploaded by Emil O. W. Kirkegaard
Author content
All content in this area was uploaded by Emil O. W. Kirkegaard on Sep 24, 2024
Content may be subject to copyright.
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
ISSN: 0025-2344
Does Cognitive Ability Mediate Black-White Income
Disparities in the USA?
Simon Wright∗Bryan J. Pesta†Emil O. W. Kirkegaard‡
Abstract
Much research has focused on the putative causes of income differentials between American Blacks and Whites. Few
studies, however, have explored whether average cognitive ability differences between these groups can explain the gaps.
This is despite IQ (intelligence quotient) scores being a potent predictor of various measures of well-being, including
both personal and household income. Thus, we investigated whether average intelligence differences between American
Blacks and Whites can mediate income differentials, over and above the test-takers’ self- and/or other-perceived race
(i.e., a control for “Colorism” as an explanation for disparate treatment). We did so by examining data from five large,
nationally representative US datasets (e.g., the General Social Survey), and by using structural equation modeling
(SEM) to first construct a general factor of intelligence (
g
) corrected for measurement error. Except with the NLSY97
dataset, we found that
g
fully mediated the personal income difference between Blacks and Whites, but that race
gaps in household income still persisted. Our results suggest that race may also influence spousal income and marital
choice, and our data underscore the need to include group intelligence differences when attempting to understand
race-based income disparities in the USA.
Keywords: Black-White differences, Income, Intelligence, Discrimination, IQ, USA, Structural equation
modeling
1 Introduction
Despite the enactment of civil rights legislation in 1964, income disparities between Black and White Ameri-
cans remain large and ubiquitous. In 2022, for example, the US Census reported household median incomes of
$81,060 for non-Hispanic Whites but only $52,860 for Blacks (
https://www.census.gov/data/tables/
2023/demo/income-poverty/p60-279.html
). Personal incomes show a similar pattern, with median earn-
ings gaps at $22,600 and $8,000 for men and women, respectively (
https://www.ssa.gov/policy/docs/
factsheets/at-a-glance/earnings-women-race-ethnicity.html
,
https://www.ssa.gov/policy/
docs/factsheets/at-a-glance/earnings-men-race-ethnicity.html
). Likewise, 8.6 percent of Whites
had income levels below the poverty line in 2022 versus 17.1% of Blacks (
https://www.census.gov/data/
tables/2023/demo/income-poverty/p60-280.html
). These differences have remained stable since at
least 1968 (Manduca 2018, Figure 2).
At least two potential factors may contribute to these income differentials. The first hypothesis
postulates differences in average human capital among Blacks and Whites, while the second is based on the
presumption of widespread systemic bias and disparate treatment against Blacks in employment settings
(Gaddis 2013). Our aim here is to test whether human capital differences can explain race/income gaps
over and above measures of both self- and other-perceived “race,” including measures of skin color in some
of our datasets. Controlling for skin color allows us to test “Colorism” as a potential explanation for income
gaps (Hunter 2007). Colorism predicts that darker-skinned Americans will realize lower average incomes
relative to lighter-skinned Americans of the same race due to pervasive racism and disparate treatment
∗Independent researcher, corresponding author; Email: simonwright50392@protonmail.com
†Independent researcher, North Olmsted, Ohio, USA
‡Arthur Jensen Institute, Denmark; Email: emilowk@proton.me
111
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
(Keith & Herring 1991). In other words, a darker-skinned “Black” individual should be expected to earn
less than a lighter-skinned “Black” individual, all other things being equal.
Conversely, one line of evidence supporting the discrimination hypothesis features survey-based data.
The approach involves sending fictitious job applications to businesses at large, varying the applications by
nothing except the race of the applicant. Such studies typically find that White applicants are more likely
to receive callbacks than are Black or Latino applicants (Quillian et al. 2017). Importantly, however, the
literature demonstrates substantial evidence of publication bias in favor of finding such an effect (Zigerell
2017). Additionally, even if a true difference in callback rates existed, it would not necessarily be caused by
irrational preference-based discrimination (Kirkegaard 2015).
Previous regression models have attempted to control for human capital and test for remaining
racial differences, but have come to mixed results (Herrnstein & Murray 1994, p. 323; Kanazawa 2005;
Tomaskovic-Devey et al. 2005). However, these papers suffer from potential weaknesses. Kanazawa (2005)
relies on measures of human capital which have the potential to be tainted by reverse causation, e.g., Black
people being less able to join unions because they are discriminated against in unionized occupations. The
conclusions of Tomaskovic-Devey et al. (2005) apply only if African Americans are not paid above what their
human capital would justify when they first enter the labor market. This may not be true due to Affirmative
Action (Arcidiacono et al. 2015). Finally, all three of these articles fail to correct for underestimation of
intelligence gaps resulting from attenuation bias (Kirkegaard 2022).
Further, we present novel data in that we use cognitive ability as one aspect of human capital. We
utilize this measure of human capital for three reasons. Firstly, the Black-White gap in measured intelligence
is substantial, approximately a one standard deviation gap in the general factor of intelligence, providing a
plausible candidate to explain income gaps (Frisby & Beaujean 2015; Fuerst et al. 2021; Hu et al. 2019;
Lasker et al. 2019; Roth et al. 2001). Secondly, intelligence appears to be mostly unchanging through
adulthood, correlating at 0.945 between the ages of (approximately) 20 and 37 (Lasker & Kirkegaard 2022)
and 0.78 between 17 and 70 (Deary 2014), with the one standard deviation Black-White IQ gap emerging
by the age of 3 years (Malloy 2013; Rushton & Jensen 2005). A recent meta-analysis has found that after
adjusting for test reliability, the rank order correlation between the general factor of intelligence at age
20 and any much older age is 0.9, and that most components of gdo not significantly differ from that
correlation (Breit et al. 2024). Further, one’s employment does not appear to affect intelligence measured
in later life (Feinkohl et al. 2021). This leaves limited scope for reverse causation: discrimination in the
labor market causing group differences in intelligence.
Thirdly, intelligence appears to be an important component of human capital. A correlation between
intelligence and income is well attested to in the literature, with a meta-analysis finding a correlation of
0.21 (Strenze 2007). When long-run average income (permanent income) is used as a measure, removing
year-to-year fluctuations, this correlation is even higher at
r
= 0
.
36 (Dalliard 2016). Intelligence also beats
education and parental education combined as a correlate of income (r= 0.41 vs r= 0.24) (Marks 2022).
For the aforementioned reasons, reverse causation (occupation causing intelligence) is not a major
concern. Additionally, Jensen (1998) provides an overview of evidence concerning IQ and job performance,
revealing that IQ is positively associated with learning speed, the rate at which employees acquire new skills,
and supervisor-rated job performance. The general factor of intelligence (g) is also positively associated with
math, reading, and writing abilities (Caemmerer et al. 2018; Rohde & Thompson 2007). These findings
strongly imply that intelligence is likely to be positively associated with productivity and consequently higher
wages.
Lastly, Danish cohort data show that IQ differences have the same effect on income differences
between siblings as between two random individuals from the general population, accounting for other
factors such as region of birth (Hegelund et al. 2019). Because siblings generally share the same common
environment in childhood, this reduces the chances of omitted variable bias. IQ differences between siblings
predicting income differences have also been found in the United States (Murray 2002).
Hence, by leveraging cognitive ability and income data from five large datasets, we aim to test four
hypotheses stemming from the discrimination model and the colorism model:
112
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
H1: Interviewer perception of race (IP) should predict income controlling for self-perception (SP). For
instance, an individual who is perceived by an interviewer as White but self-identifies as Black should have
higher earnings than those who are both perceived and identify as Black.
H2: Self-perception of race should not predict income controlling for interviewer perception. This is a
strong-form test of a discrimination model, as it assumes that self-perception of race is not associated with
any human capital differences that are not absorbed by interviewer perception. However, it serves as a test
for the statistical precision of H1. The high correlation between SP and IP is likely to cause multicollinearity
(across all datasets, there are 306 cases where SP and IP differ), leading to large standard errors in both
H1 and H2 tests. A statistically significant result in H2 would demonstrate that our sample size is large
enough to overcome multicollinearity.
H3: Those who are perceived by an interviewer (IP) to be Black should have lower earnings than those
who are IP White after controlling for cognitive ability. Cognitive ability represents only one aspect of human
capital. Therefore, results supporting H3 would not fully rule out the human capital model. However, if
controlling for a single aspect of human capital fully mediates racial earnings differences, then this would
provide strong evidence against the discrimination model.
H4: After controlling for race and cognitive ability, those with darker skin tones should earn less than
those with lighter skin tones.
More specifically, we first employed structural equation modeling (SEM) on five nationally representa-
tive US datasets (e.g., The General Social Survey, GSS) to estimate the influence of racial differences in
intelligence on both personal and household income. By using SEM, we were able to leverage the statistical
advantage of correcting our g estimates for attenuation bias, which can bias
g
estimates towards finding
smaller group differences (see, e.g., Kirkegaard, 2022).
2 Method
2.1 Estimating IQ
We coded cognitive ability data from five large, nationally representative samples within the USA:
1. National Longitudinal Survey of Youth (NLSY) 1979 (BLS, 2016)
2. NLSY 1997 (BLS, 2017)
3. General Social Survey (GSS: 2019) (Smith et al., 2019)
4. Add Health (2015-2018) (Harris & Udry, 2015a,b,c, 2018, 2020)
5. American National Election Studies (ANES, 2014)
The sections below detail how we derived IQ scores from each of these five datasets. For brevity
here, tables containing the original names of variables used and how we coded them, along with summary
statistics, are available in the supplementary materials (SM) file.
One alternative aspect of human capital not used in this paper is years of education. We elected
to use cognitive ability measures over this for two reasons. Firstly, because cognitive ability is a stronger
predictor of income than years of education (Marks, 2022). And secondly, because of the presence of
affirmative action in universities (Arcidiacono et al., 2015). Because Black candidates are more likely to
be accepted to universities than White candidates at the same level of measured ability, including years of
education as an aspect of human capital may create bias.
113
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
2.1.1 National Longitudinal Study of Youth (1979)
NLSY79 (National Longitudinal Study of Youth, 1979) comprises a large dataset of approximately 12,000
individuals whose survey responses were collected in 1979. We estimated IQs via reported scores on the
subtests of the Armed Services Vocational Aptitude Battery (ASVAB; BLS, 2015). The ASVAB is a
well-researched and highly valid measure of general cognitive ability (Koenig et al., 2008). NLSY79 also
reports race from self-report and coded by the NLSY79 survey administrator.
2.1.2 National Longitudinal Study of Youth (1997)
NLSY97 (National Longitudinal Study of Youth, 1997) is a relatively recent replication of NLSY79. The
1997 data feature a substantially younger cohort, as well as additional cognitive ability scores from the
Peabody Individual Achievement Test (
https://www.bls.gov/nls/nlsy79-children/topical-guide
-to-data/home.htm
). The NLSY97 also reports self- and other-perceived race. A skin color card used for
color coding is available at:
https://www.nlsinfo.org/sites/default/files/attachments/140114/
skin%20color%20card%20for%20NLSY97.pdf.
2.1.3 General Social Survey (Smith et al., 2019)
Our GSS intelligence estimates were derived from scores on the 10-item Wordsum vocabulary test. This is a
multiple-choice test where individuals are given a word and then asked to choose which other word is most
similar in meaning. Because the GSS intelligence metric was binary (individuals answered the vocabulary
items either correctly or incorrectly), we used a DWLS estimator, as opposed to the ML estimator we
employed for our other gestimates. The reliance on Wordsum is a limitation of the GSS dataset. Although
it has been estimated that Wordsum is highly g-loaded (0.93 when corrected for reliability; Woodley of
Menie et al., 2015), this result originates from a sample dating to the 1940s (Kirkegaard, 2019). It is also
solely a measure of verbal intelligence which, whilst highly correlated with general intelligence (Woodley of
Menie et al., 2015), is conceptually only one aspect of g(Kan et al., 2013). Further research is needed to
establish the validity of the Wordsum test as a proxy for intelligence in a modern context.
2.1.4 Add Health (2015 - 2018)
Add Health data are collected under the umbrella of the NLSY, although income data for this survey exist
only for its last two data-collection waves. Regarding gestimates, Add Health reports scores from the PIAT
(taken in Waves 1 and 3), as well as a knowledge quiz (Waves 1 and 2), and digit span/word recall tests
(Waves 4 and 5). Add Health also codes participants by self- and other-perceived race.
2.1.5 American National Election Studies (2014)
Data from the American National Election Studies (ANES, 2014) are somewhat limited in that they report
household but not personal income data. For intelligence estimates, we relied on a Wordsum-type test
administered to survey participants. This survey has a measure for skin color (Figure 1).
Figure 1: Skin color coding in the ANES dataset. Interviewers were asked to give a number based on this key
(ANES, 2012).
114
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
2.2 Statistical Approach
In datasets where it is recorded (NLSY 1979, 1997, and Add Health), we measured race by the interviewer’s
perception (IP), as opposed to the respondents’ self-perception (SP). We did so because only IP, not SP,
can logically cause labor market discrimination: an employer can only discriminate based on their own
beliefs about an individual’s race. Also, some datasets reported the respondents’ skin color, which we coded
here to control for predictions based on Colorism Theory (H4). In all of our analyses, we treated “White”
as the omitted variable, and therefore the regression coefficients reported below for Blacks estimate the
income gap between this group relative to non-Hispanic Whites. Similarly, darker skin tones are always
coded as higher numbers. To concentrate our analysis on the Black-White income gap, members of other
minorities were removed from our datasets. For all datasets, the logarithm of the income measure is taken
and then normalized to have a mean of 0 and standard deviation of 1.
Additionally, in three datasets (NLSY79, NLSY97, Add Health) income measures were available
for the same individuals across multiple years. We thus used SEM to construct a “permanent income”
latent variable, with income for each of the individual years serving as indicator variables. We then used
“permanent income” as a dependent variable in the models reported below.
In order to test H1, we aim to test for an effect of IP race on income when controlling for SP. Data
for both SP and IP are available in three datasets (NLSY79, NLSY97, Add Health). The discrimination
model would predict that an individual perceived by others as white (black), but self-identifying as black
(white), should have an income closer to the average for whites (blacks). However, introducing both of
these variables into a model threatens to introduce multicollinearity, which will increase the standard error
of the estimated effect of IP race. This is the motivation for conducting a meta-analysis of our results,
giving more statistical precision than any individual dataset can. Effectively, we combined the statistical
precision from the 165 total cases where IP and SP differ.
Figure 2 summarizes the base model used to test H3, together with the indicator variables we used
to estimate latent g. Figures for all variants of this model are available in the SM file. Recall that our
latent variable approach to measuring gallowed us to correct for attenuation bias resulting from random
measurement error. Failure to correct for this bias leads to underestimating standardized differences in
mean intelligence across Black and White Americans (see, e.g., Kirkegaard, 2022; Kline, 2016, p. 127).
Finally, our analyses involved regressing income on both racial identity and gwhile also controlling
for each participant’s sex and age (or birth year for datasets with income for one individual across multiple
years). Thereafter, we conducted a meta-analysis based on our overall findings to statistically estimate a
meta-analytic value summarizing all the effects. The primary focus of this study takes personal income as
our outcome measure. Although we also report household income, we leave a more thorough investigation
of those results to future research.
3 Results
3.1 Main Analysis
The grand mean for our Black gestimate was
Z
=
−
1
.
09, which corresponds to an IQ estimate of 83.7
(relative to the White mean, set to equal 100). The IQ difference here is remarkably consistent with Roth et
al.’s (2001) large-scale meta-analysis of these effects (with particular focus on high-stakes testing scenarios).
To wit, these authors reported a meta-analytic effect size of -1.10 (IQ = 83.5) for Black Americans, relative
to White Americans, with analyses featuring well over one million participants.
We next turn to reporting the specifics of our regression analyses within each of our five datasets.
For each analysis, we attempted to predict personal and/or household income based on our IQ estimates,
SP and IP (when available), age, and sex.
3.1.1 National Longitudinal Study of Youth (1979)
Table 1 and Table 2 display regression results featuring race and gas predictors of income in this dataset,
controlling for age and sex. For brevity, the full models for all our results are available in the supplementary
materials. As expected, prior to controlling for g, being perceived as Black is associated with lower personal
115
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Figure 2: Base model for our analysis. Rectangles represent manifest variables. Rounded rectangles represent
(manifest) indicator variables. Circles represent latent variables. Solid lines represent regression paths. Dashed lines
represent factor loadings.
income. However, this association becomes insignificant when self-perceived race is introduced as a control,
providing evidence against H1. Interestingly, despite concerns of multicollinearity, self-perceived race is
statistically significant and negatively associated with income in this model, providing evidence against H2.
Examining the outcome measure of personal income, controlling for g, our findings reveal no statistically
significant impact of being perceived as Black (therefore, not supporting H3). This lack of association holds
both before and after controlling for an individual’s self-perception. There does appear to be a negative
association between interviewer perception as Black and household income; this effect disappears when
self-perception is controlled for.
3.1.2 The National Longitudinal Study of Youth (1997)
Table 3 and Table 4 display regression results featuring race and gas predictors of income in this dataset,
controlling for sex and birth year (full results available in supplementary materials). Here, without controlling
for g, neither being perceived as Black nor self-identifying as Black is statistically significant in model 2,
although both coefficients are negative.
Unlike the NLSY79 dataset, the estimated effect of being perceived as Black on personal income,
whilst controlling for g, is statistically significant and negative when self-perception is not controlled for,
providing some evidence against H3. However, being perceived as Black explains little of the overall
variation in income between people (only 0.4 % of the variance not explained by other variables). This
effect disappears once an individual’s self-perception is controlled for. Similar results are obtained when skin
color is used as a proxy for others’ perception of race. The results for household income are more closely
aligned with the NLSY79.
116
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 1: The impact of race on personal and household income in the NLSY79 dataset, no controls for intelligence.
*
p < .
05; **
p < .
01; ***
p < .
001.
β
= standardized beta (path coefficient). All latent variables have a mean of 0
and variance of 1.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.092
IP_Black -0.177*** -0.406 0.026 0.033
Personal income 2 0.123
IP_Black -0.039 -0.087 0.165 0.000
SP_Black -0.210** -0.463 0.166 0.002
Household income 1 0.115
IP_Black -0.327*** -0.763 0.028 0.108
Household income 2 0.173
IP_Black -0.197* -0.447 0.201 0.002
SP_Black -0.217* -0.494 0.204 0.002
Table 2: The impact of race on personal and household income in the NLSY79 dataset, controlling for g. *
p<.
05;
**
p<.
01; ***
p<.
001.
β
= standardized beta (path coefficient). All latent variables have a mean of 0 and variance
of 1.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.187
g 0.377*** 0.344 0.014 0.106
IP_Black 0.017 0.041 0.032 0.000
Personal income 2 0.234
g 0.422*** 0.380 0.020 0.126
IP_Black 0.070 0.164 0.150 0.000
sex_Male 0.208*** 0.475 0.029 0.053
SP_Black -0.069 -0.164 0.152 0.000
Household income 1 0.329
g 0.562*** 0.565 0.016 0.242
IP_Black -0.040** -0.108 0.033 0.002
Household income 2 0.379
g 0.575*** 0.576 0.019 0.249
IP_Black -0.044 -0.116 0.182 0.000
SP_Black -0.032 -0.083 0.187 0.000
3.1.3 General Social Survey (2019)
Table 5 displays regression results featuring race and gas predictors of income in this dataset, controlling
for age and sex (full results available in supplementary materials). For personal income, perceiving oneself
as Black is not a significant predictor of income, and neither is skin color. For household income, however,
there is a negative effect of either when used in the model individually, although neither is statistically
significant when both are used.
117
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 3: The impact of race on personal and household income in the NLSY97 dataset, no controls for intelligence.
*
p < .
05; **
p < .
01; ***
p < .
001.
β
= standardized beta (path coefficient). All latent variables have a mean of 0
and variance of 1.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.136
IP_Black -0.265*** -0.598 0.038 0.075
Personal income 2 0.137
IP_Black -0.104 -0.236 0.200 0.000
SP_Black -0.163 -0.368 0.200 0.001
Personal income 3 0.137
Color -0.087** -0.085 0.030 0.003
SP_Black -0.191*** -0.427 0.068 0.012
Personal income 4 0.126
Color -0.248*** -0.241 0.018 0.066
Household income 1 0.172
IP_Black -0.407*** -0.941 0.037 0.167
Household income 2 0.175
IP_Black -0.087 -0.200 0.200 0.000
SP_Black -0.326*** -0.752 0.222 0.004
Household income 3 0.184
Color -0.049 -0.050 0.032 0.001
SP_Black -0.382*** -0.877 0.070 0.049
Household income 4 0.142
Color -0.371*** -0.364 0.017 0.138
3.1.4 Add Health (2015 - 2018)
Table 6 and Table 7 display regression results featuring race and gas predictors of income in this dataset,
controlling for birth year and sex (full results in supplementary materials). The results from Add Health are
similar to the results of the previous surveys. Without controls for g, IP Black is not statistically significantly
associated with personal income, whilst SP has a significant negative association, providing evidence against
both H1 and H2.
When controls for gare introduced, being perceived as Black surprisingly is associated with higher
personal income, absent controls for SP. Skin color is never negatively associated with personal income.
3.1.5 American National Election Studies (2014)
Lastly, the ANES results (Table 8) corroborate our findings that being perceived as Black is negatively
correlated with household income (col. 1). Skin color presents a negative relationship with income, both
before and after controlling for self-perceived race (cols. 2 & 3).
3.2 Meta-analysis
We conduct meta-analyses of our measures of race on personal income. The effect sizes shown are the
standardized beta effect estimates estimated in the above sections. These analyses were conducted using
the Meta and Metafor packages in R (Schwarzer, 2023; Viechtbauer, 2023). We further conducted
meta-analyses for the impact of skin color on personal income, with and without controls for self-perception.
118
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 4: The impact of race on personal and household income in the NLSY97 dataset. *
p < .
05; **
p < .
01; ***
p<.001.β= standardized beta (path coefficient). All latent variables have a mean of 0 and variance of 1.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.241
g0.396*** 0.373 0.021 0.122
IP_Black -0.068*** -0.163 0.044 0.004
Personal income 2 0.241
g0.395*** 0.371 0.021 0.121
IP_Black -0.007 -0.017 0.200 0.000
SP_Black -0.063 -0.150 0.200 0.000
Personal income 3 0.256
Color -0.010 -0.010 0.031 0.000
g0.422*** 0.398 0.023 0.137
SP_Black -0.046 -0.112 0.070 0.001
Personal income 4 0.255
Color -0.046* -0.049 0.020 0.002
g0.428*** 0.414 0.023 0.146
Household income 1 0.294
g0.425*** 0.417 0.020 0.148
IP_Black -0.193*** -0.483 0.041 0.037
Household income 2 0.296
g0.422*** 0.414 0.020 0.147
sex_Male -0.072*** -0.172 0.031 0.007
Household income 3 0.311
Color 0.028 0.031 0.032 0.000
g0.437*** 0.429 0.022 0.156
SP_Black -0.226*** -0.566 0.071 0.020
Household income 4 0.299
Color -0.145*** -0.157 0.019 0.022
g0.472*** 0.472 0.022 0.182
Our estimates test H1 and H2 in Figure 3 and Figure 4, respectively, showing the impact of interviewer-
and self-perceived race on income. The results fail to support either H1 or H2. Interviewer-perceived race
does not significantly affect income when it differs from self-perception, whereas self-perception is strongly
negative and statistically significant. The finding of a strong and significant negative effect, combined with
relatively narrow confidence intervals, reduces the concern that multicollinearity makes results uninterpretable.
Across all our datasets where this test was available, there are 54 individuals where SP is Black but IP
is White and 252 where the opposite is true. Figure 5 further reinforces the lack of support for H1 as
controlling for both gand self-perception fails to find any significant results.
Next, H3 is tested in Figure 6. In this model, there is no control for self-perception. It is purely the
estimated impact of race (interviewer perception in datasets where available) on personal income, controlling
for g. The results show no estimated effect of being Black on personal income, failing to support H3.
We also fail to find any evidence to support the colorism hypothesis (H4): With and without controls
for SP, there is no significant impact of having a darker skin color, net of intelligence (Figure 7 and Figure 8).
119
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 5: The impact of race on personal and household income in the GSS dataset. *
p<.
05; **
p<.
01; ***
p < .
001.
β
= standardized beta (path coefficient). All latent variables have a mean of 0 and variance of 1. Year
fixed effects. HAC (heteroskedasticity and auto-correlation-robust) standard errors.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.154
g0.209*** 0.217 0.014 0.049
SP_Black -0.004 -0.013 0.031 0.000
Personal income 2 0.116
Color -0.022 -0.013 0.020 0.000
g0.225*** 0.249 0.041 0.059
SP_Black 0.008 0.025 0.103 0.001
Personal income 3 0.124
Color 0.004 0.003 0.017 0.000
g0.248*** 0.273 0.042 0.074
Household income 1 0.161
g0.353*** 0.322 0.011 0.126
SP_Black -0.080*** -0.222 0.024 0.003
Household income 2 0.160
Color -0.034 -0.019 0.014 0.000
g0.364*** 0.367 0.031 0.147
SP_Black -0.046 -0.128 0.072 0.000
Household income 3 0.169
Color -0.055* -0.030 0.012 0.000
g0.385*** 0.389 0.030 0.161
Figure 3: Meta-analysis of the impact of interviewer-perceived race on personal income, controlling for self-perception
of race.
120
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 6: The impact of race on personal and household income in the Add Health dataset, no controls for intelligence.
*
p < .
05; **
p < .
01; ***
p < .
001.
β
= standardized beta (path coefficient). All latent variables have a mean of 0
and variance of 1.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.119
IP_Black -0.112*** -0.263 0.037 0.014
Personal income 2 0.120
IP_Black 0.057 0.135 0.127 0.000
SP_Black -0.174** -0.413 0.128 0.002
Personal income 3 0.122
Color -0.115*** -0.092 0.027 0.003
SP_Black -0.015 -0.036 0.078 0.000
Personal income 4 0.122
Color -0.128*** -0.103 0.014 0.018
Household income 1 0.138
IP_Black -0.366*** -0.870 0.040 0.134
Household income 2 0.141
IP_Black -0.138* -0.328 0.141 0.001
SP_Black -0.234*** -0.564 0.143 0.003
Household income 3 0.148
Color -0.193*** -0.157 0.028 0.009
SP_Black -0.196*** -0.478 0.083 0.009
Household Income 4 0.140
Color -0.368*** -0.299 0.015 0.136
Personal income 1 0.119
IP_Black -0.112*** -0.263 0.037 0.014
Personal income 2 0.120
IP_Black 0.057 0.135 0.127 0.000
SP_Black -0.174** -0.413 0.128 0.002
Personal income 3 0.122
Color -0.115*** -0.092 0.027 0.003
SP_Black -0.015 -0.036 0.078 0.000
121
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 7: The impact of race on personal and household income in the Add Health dataset. *
p < .
05; **
p < .
01;
*** p<.001.β= standardized beta (path coefficient). All latent variables have a mean of 0 and variance of 1.
Independent variable βbeta SE beta R2/partial r2
Personal income 1 0.266
g0.441*** 0.459 0.034 0.174
IP_Black 0.077*** 0.199 0.059 0.006
Personal income 2 0.266
g0.441*** 0.457 0.034 0.173
IP_Black 0.090 0.231 0.151 0.001
SP_Black -0.012 -0.032 0.155 0.000
Personal income 3 0.271
Color -0.040 -0.035 0.029 0.000
g0.451*** 0.462 0.038 0.176
SP_Black 0.125*** 0.328 0.091 0.004
Personal income 4 0.276
Color 0.077** 0.068 0.023 0.006
g0.455*** 0.472 0.038 0.182
Household income 1 0.312
g0.475*** 0.513 0.035 0.208
IP_Black -0.145*** -0.386 0.056 0.024
Household income 2 0.312
g0.474*** 0.509 0.035 0.206
IP_Black -0.074 -0.195 0.148 0.000
SP_Black -0.074 -0.200 0.200 0.000
Household income 3 0.320
Color -0.107** -0.098 0.030 0.003
g0.478*** 0.512 0.038 0.208
SP_Black -0.043 -0.117 0.094 0.001
Household income 4 0.328
Color -0.134*** -0.124 0.022 0.021
g0.495*** 0.538 0.039 0.224
122
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Table 8: The impact of race on personal and household income in the ANES dataset. *
p < .
05; **
p < .
01; ***
p<.001.β= standardized beta (path coefficient). All latent variables have a mean of 0 and variance of 1.
Independent variable βbeta SE beta R2/partial r2
Household income 1 0.159
g0.376*** 0.426 0.019 0.131
Sex: male 0.065*** 0.158 0.031 0.005
Household income 2 0.145
Color -0.134*** -0.068 0.017 0.010
g0.323*** 0.353 0.032 0.092
SP_Black 0.036 0.095 0.089 0.001
Household income 3 0.143
Color -0.104*** -0.053 0.013 0.011
g0.321*** 0.351 0.032 0.091
Figure 4: Meta-analysis of the impact of self-perceived race on personal income, controlling for interviewer perception
of race.
Figure 5: Impact of being perceived as black (IP) on personal income (standardized beta), with controls for gand
for self-perception of being black (SP).
123
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Figure 6: Impact of Black (IP where available) on personal income (standardized beta), gis controlled, but there
are no controls for self-perception (SP).
Figure 7: The impact of skin color on personal income (standardized beta), controlled for g, but no controls for SP.
Figure 8: The impact of skin color on personal income (standardized beta), with controls for gand SP.
124
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
4 Discussion
The findings of this study provide robust support for the human capital hypothesis, positing that disparities
in cognitive ability may entirely account for the Black-White earnings gap in the US labor market. We
employed structural equation modeling in our analysis, controlling exclusively for sex, age, and measured
intelligence, thereby addressing concerns related to introducing measures of human capital that could be
influenced by discrimination in the labor market (Gaddis, 2013).
The share of the variance in permanent personal income not explained by other variables but explained
by g(partial
r2
) is consistently above 10 %. This is approximately in line with the total
R2
of 13–17 %
between intelligence and income found in previous research (Dalliard, 2016; Marks, 2022). It is greater than
the
R2
of 6–7 % between permanent personal income and years of education + parental education (Marks,
2022).
Across all datasets except for NLSY97, our results did not reveal a significant negative impact of
being Black on personal income, net of g. The meta-analysis strengthens this conclusion, as both the fixed
and random effects models demonstrate a mean impact of zero for being Black on personal income. All of
our datasets support the finding of a robust positive association between gand income.
In all datasets, we observed a negative association between being perceived as Black and household
income. This finding implies that racial differences in household dynamics unexplained by cognitive ability
of one member of the household may exist. Such disparities could encompass decisions on whom to marry
and the number of working household members. Future research could explore household income differences
when broken down into household types, in particular whether these earnings gaps could be accounted for
by racial differences in the proportion of single and dual-earning households.
Our research also attempts to disentangle the relative importance of self-perception and others’
perception of one’s race. As employers base employment decisions on their perception of an individual’s race,
individuals who identify with one race but are perceived as another offer a critical test of discrimination. We
found no evidence in any of our models suggesting that others’ perceptions of an individual’s race influenced
income once self-perceived race was controlled for.
Overall, these datasets present strong evidence supporting the human capital hypothesis, indicating
that differences in cognitive ability primarily account for the Black-White earnings gap in the US labor
market.
Declarations of interest
None
Note
The replication code and Supplementary Material is available at: https://osf.io/4mpdb/
References
ANES. (2012). ANES > 2012 Time Series Study > Skin Color Scale. Retrieved from
https://archive.is/
2uzqUhttps://doi.org/10.1037/bul0000425 (Archive.Is)
ANES. (2014). User’s guide and codebook for the ANES 2012 time series study. Univ. of Michigan and
Stanford Univ., Ann Arbor, MI and Palo Alto, CA. Retrieved from
https://electionstudies.org/
data-center/2012-time-series-study/ (ANES | American National Election Studies)
Arcidiacono, P., Lovenheim, M., & Zhu, M. (2015). Affirmative action in undergraduate education. Annual
Review of Economics,7(1), 487-518. doi:
https://doi.org/10.1146/annurev-economics-080614
-115445
125
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
BLS. (2015). Aptitude, achievement & intelligence scores | National Longitudinal Surveys. Re-
trieved from
https://web.archive.org/web/20230130221458/https://nlsinfo.org/content/
cohorts/nlsy79/topical-guide/education/aptitude-achievement-intelligence-scores
BLS. (2016). National Longitudinal Survey of Youth 1979 cohort, 1979-2016 (rounds 1-27). Retrieved from
https://www.nlsinfo.org/investigator/pages/login?p=timeout
(Bureau of Labor Statistics,
U.S. Department of Labor. Produced and distributed by the Center for Human Resource Research (CHRR),
Ohio State Univ., Columbus OH: 2019)
BLS. (2017). National Longitudinal Survey of Youth 1997 cohort, 1997-2017 (rounds 1-18). Retrieved from
https://www.nlsinfo.org/investigator/pages/login?p=timeout
(Bureau of Labor Statistics,
U.S. Department of Labor. Produced and distributed by the Center for Human Resource Research (CHRR),
Ohio State Univ., Columbus OH: 2019)
Breit, M., Scherrer, V., Tucker-Drob, E. M., & Preckel, F. (2024). The stability of cognitive abilities: A
meta-analytic review of longitudinal studies. Psychological Bulletin,150 (4), 399-439. doi:
https://
doi.org/10.1037/bul0000425
Caemmerer, J. M., Maddocks, D. L. S., Keith, T. Z., & Reynolds, M. R. (2018). Effects of cognitive
abilities on child and youth academic achievement: Evidence from the WISC-V and WIAT-III. Intelligence,
68, 6-20. doi: https://doi.org/10.1016/j.intell.2018.02.005
Dalliard. (2016). IQ and permanent income: Sizing up the “IQ paradox”. Retrieved
from
https://humanvarieties.org/2016/01/31/iq-and-permanent-income-sizing-up-the-iq
-paradox/ (Human Varieties, January 31)
Deary, I. J. (2014). The stability of intelligence from childhood to old age. Current Directions in Psychological
Science,23 (4), 239-245.
Feinkohl, I., Kozma, P., Borchers, F., van Montfort, S. J. T., Kruppa, J., Winterer, G., .. . Pischon, T. (2021).
Contribution of IQ in young adulthood to the associations of education and occupation with cognitive ability
in older age. BMC Geriatrics ,21(1), 346. doi: https://doi.org/10.1186/s12877-021-02290-y
Frisby, C. L., & Beaujean, A. A. (2015). Testing Spearman’s hypotheses using a bi-factor model with
WAIS-IV/WMS-IV standardization data. Intelligence,51 , 79-97. doi:
https://doi.org/10.1016/
j.intell.2015.04.007
Fuerst, J. G. R., Hu, M., & Connor, G. (2021). Genetic ancestry and general cognitive ability in a
sample of American youths. Mankind Quarterly,62 (1), 186-216. doi:
https://doi.org/10.46469/
mq.2021.62.1.11
Gaddis, S. M. (2013). A matter of degrees: Educational credentials and race and gender discrimination in
the labor market (Unpublished doctoral dissertation). Univ. of North Carolina at Chapel Hill.
Harris, K. M., & Udry, R. J. (2015a). National Longitudinal Study of Adolescent to Adult Health (Add
Health) Wave I, 1994-1995 [dataset]. UNC Dataverse. doi: https://doi.org/10.15139/S3/11900
Harris, K. M., & Udry, R. J. (2015b). National Longitudinal Study of Adolescent to Adult Health (Add
Health) Wave II, 1996 [dataset]. UNC Dataverse. doi: https://doi.org/10.15139/S3/11917
Harris, K. M., & Udry, R. J. (2015c). National Longitudinal Study of Adolescent to Adult Health (Add
Health) Wave IV, 2008 [dataset]. UNC Dataverse. doi: https://doi.org/10.15139/S3/11920
Harris, K. M., & Udry, R. J. (2018). National Longitudinal Study of Adolescent to Adult Health (Add
Health), 1994-2018 [public use]: Version 25 (v25) [dataset]. ICPSR - Inter University Consortium for
Political and Social Research. doi: https://doi.org/10.3886/ICPSR21600.V25
126
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Harris, K. M., & Udry, R. J. (2020). National Longitudinal Study of Adolescent to Adult Health (Add
Health) Wave III, 2001-2002 [dataset]. UNC Dataverse. doi:
https://doi.org/10.15139/S3/11918
Hegelund, E. R., Flensborg-Madsen, T., Dammeyer, J., Mortensen, L. H., & Mortensen, E. L. (2019). The
influence of familial factors on the association between IQ and educational and occupational achievement:
A sibling approach. Personality and Individual Differences,149 , 100-107. doi:
https://doi.org/
10.1016/j.paid.2019.05.045
Herrnstein, R. J., & Murray, C. A. (1994). The Bell Curve: Intelligence and Class Structure in American
Life. Free Press.
Hu, M., Lasker, J., Kirkegaard, E., & Fuerst, J. (2019). Filling in the gaps: The association between
intelligence and both color and parent-reported ancestry in the National Longitudinal Survey of Youth
1997. Psych,1(1), 240-261. doi: https://doi.org/10.3390/psych1010017
Hunter, M. (2007). The persistent problem of colorism: Skin tone, status, and inequality. Sociology
Compass,1(1), 237-254. doi: https://doi.org/10.1111/j.1751-9020.2007.00006.x
Jensen, A. R. (1998). The g Factor: The Science of Mental Ability. Praeger.
Kan, K.-J., Wicherts, J. M., Dolan, C. V., & van der Maas, H. L. J. (2013). On the nature and nurture of
intelligence and specific cognitive abilities: The more heritable, the more culture dependent. Psychological
Science,24 (12), 2420-2428. doi: https://doi.org/10.1177/0956797613493292
Kanazawa, S. (2005). The myth of racial discrimination in pay in the United States. Managerial and
Decision Economics,26(5), 285-294. doi: https://doi.org/10.1002/mde.1229
Keith, V. M., & Herring, C. (1991). Skin tone and stratification in the black community. American Journal
of Sociology ,97 (3), 760-778. doi: https://doi.org/10.1086/229819
Kirkegaard, E. O. W. (2015). Why do persons from higher mean IQ populations do better even after selection?
Retrieved from
https://www.researchgate.net/publication/277591528_Why_do_persons_from
_higher_mean_IQ_populations_do_better_even_after_selection (The Winnower)
Kirkegaard, E. O. W. (2019). The WORDSUM question. Retrieved from
https://emilkirkegaard.dk/
en/2019/01/the-wordsum-question/ (Clear Language, Clear Mind, January 4)
Kirkegaard, E. O. W. (2022). The intelligence gap between black and white survey workers on the Prolific
platform. Mankind Quarterly,63 , 79-88. doi: https://doi.org/10.46469/mq.2022.63.1.3
Kline, R. B. (2016). Principles and Practice of Structural Equation Modeling (4th edition). Guilford Press.
Koenig, K. A., Frey, M. C., & Detterman, D. K. (2008). ACT and general cognitive ability. Intelligence,
36(2), 153-160. doi: https://doi.org/10.1016/j.intell.2007.03.005
Lasker, J., & Kirkegaard, E. O. W. (2022). The generality of educational effects on cognitive ability: A
replication [Preprint]. PsyArXiv . doi: https://doi.org/10.31234/osf.io/8s2vx
Lasker, J., Pesta, B. J., Fuerst, J. G. R., & Kirkegaard, E. O. W. (2019). Global ancestry and cognitive
ability. Psych,1(1), 431-459. doi: https://doi.org/10.3390/psych1010034
Malloy, J. (2013). The onset and development of B-W ability differences: Early infancy to age 3 (Part 1).
Retrieved from
https://humanvarieties.org/2013/05/26/the-onset-and-development-of-b-w
-ability-differences-early-infancy-to-age-3-part-1/ (Human Varieties, May 26)
Manduca, R. (2018). Income inequality and the persistence of racial economic disparities. Sociological
Science,5, 182-205. doi: https://doi.org/10.15195/v5.a8
127
Mankind Quarterly, 65(1), 111-128 2024 Fall Edition
Marks, G. N. (2022). Cognitive ability has powerful, widespread and robust effects on social stratification:
Evidence from the 1979 and 1997 US National Longitudinal Surveys of Youth. Intelligence,94 , 101686.
doi: https://doi.org/10.1016/j.intell.2022.101686
Murray, C. (2002). IQ and income inequality in a sample of sibling pairs from advantaged family backgrounds.
American Economic Review,92 (2), 339-343. doi:
https://doi.org/10.1257/000282802320191570
Quillian, L., Pager, D., Hexel, O., & Midtbøen, A. H. (2017). Meta-analysis of field experiments shows no
change in racial discrimination in hiring over time. Proceedings of the National Academy of Sciences,
114(41), 10870-10875. doi: https://doi.org/10.1073/pnas.1706255114
Rohde, T. E., & Thompson, L. A. (2007). Predicting academic achievement with cognitive ability.
Intelligence,35 (1), 83-92. doi: https://doi.org/10.1016/j.intell.2006.05.004
Roth, P. L., Bevier, C. A., Bobko, P., Switzer, F. S., & Tyler, P. (2001). Ethnic group differences in
cognitive ability in employment and educational settings: A meta-analysis. Personnel Psychology ,54 (2),
297-330. doi: https://doi.org/10.1111/j.1744-6570.2001.tb00094.x
Rushton, J. P., & Jensen, A. R. (2005). Thirty years of research on race differences in cognitive ability.
Psychology, Public Policy, and Law,11 (2), 235-294.
Schwarzer, G. (2023). meta: General Package for Meta-Analysis (Version 6.2-1) [Computer software].
Retrieved from https://cran.r-project.org/web/packages/meta/index.html
Smith, T., Morgan, S., Freese, J., & Davern, M. (2019). General Social Surveys, 1972-2018 cumulative
codebook. Retrieved from https://gss.norc.org/Documents/codebook/gss_codebook.zip
Strenze, T. (2007). Intelligence and socioeconomic success: A meta-analytic review of longitudinal research.
Intelligence,35 (5), 401-426.
Tomaskovic-Devey, D., Thomas, M., & Johnson, K. (2005). Race and the accumulation of human capital
across the career: A theoretical model and fixed-effects application. American Journal of Sociology ,
111(1), 58-89. doi: https://doi.org/10.1086/431779
Viechtbauer, W. (2023). metafor: Meta-Analysis Package for R (Version 4.0-0) [Computer software].
Retrieved from https://cran.r-project.org/web/packages/metafor/index.html
Woodley of Menie, M. A., Fernandes, H. B. F., Figueredo, A. J., & Meisenberg, G. (2015). By their
words ye shall know them: Evidence of genetic selection against general intelligence and concurrent
environmental enrichment in vocabulary usage since the mid-19th century. Frontiers in Psychology ,6.
doi: https://doi.org/10.3389/fpsyg.2015.00361
Zigerell, L. J. (2017). Small study effects in a meta-analysis of racial discrimination field experiments.
Retrieved from https://www.ljzigerell.com/?p=4698 (Sept. 28)
128
