Content uploaded by Emil O. W. Kirkegaard
Author content
All content in this area was uploaded by Emil O. W. Kirkegaard on Jan 06, 2024
Content may be subject to copyright.
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
ISSN: 0025-2344
Systemic Racism Does Not Explain Variation in Race
Gaps on Cognitive Tests
Emil O. W. Kirkegaard∗
Abstract
Systemic racism theory predicts that counties where there are more White people and where people are more racist
against non-Whites should have larger race gaps on cognitive measures. We used county-level data from the United
States to test these predictions of the systemic racism model. We used cognitive test results from state scholastic tests
from the Stanford Education Data Archive (SEDA) 4.1, which provided data for Black-White and Hispanic-White
gaps from 1,473 and 1,750 counties, respectively. Contrary to predictions from the systemic racism model, we find
that cognitive race gaps are smaller in counties with more Republicans: r’s with %Republican are -.54 and -.59 for
Black-White and Hispanic-White gaps, respectively. Gaps also tend to be smaller where there are more White people,
with correlations of %White with Black-White and Hispanic-White gaps of r = -.30 and -.38 (all results p < .001).
We furthermore used data from Project Implicit as a measure of latent racism against Blacks. However, these also
tended to have the wrong direction of results: Higher implicit anti-Black racism was associated with smaller cognitive
gaps. Regression modeling reduced the effect sizes, but not the general pattern of directions. The same pattern
was also seen for social status gaps as the outcome variable. Results were entirely contrary to the predictions of the
systemic racism model.
Keywords: Systemic racism, Institutional racism, Intelligence, Cognitive ability, Scholastic tests, Stanford
Education Data Archive, Implicit Association Test, Republicans, Democrats, Black-White gap, Hispanic-
White gap
1 Introduction
The United States is a systemically racist country according to many contemporary writers (Feagin, 2006;
Kendi, 2019; Taylor, 2019). The model of society is that institutions are systemically set up so as to
overtly or indirectly favor White people. They were set up as self-serving institutions by White people for
White people. This includes essentially every aspect of society: education, housing, government, banking,
insurance, policing, the judicial system, and so on. Thus, from a scientific perspective, it is a model based
on the explicit and implicit attitudes of White people, who are according to this model the most powerful
group in the country, such that their attitudes determine the overall direction of the country. They use their
influence to keep other groups down using whatever means necessary, such as designing laws that indirectly
target other groups, implement ostensibly objective, but actually racially biased testing that controls access
to higher education and prestigious occupations. The current views are similar to historical ones that have
been promoted since the early 1900s by various social scientists (McClelland, 1973; Myrdal, 1944).
Considering the United States as a country, the distribution of populations and political attitudes is
starkly diverse, or segregated, in the country. Some areas harbor many Republicans, some many Democrats,
some many White people, others only a few. These differences largely reflect historical settlement patterns
and later migrations (Fischer, 1989; JayMan, 2017; Woodard, 2012), but also residential discrimination.
Thus, based on the White systemic racism model, we can derive certain predictions for scientific testing:
∗Email: the.dfx@gmail.com
370
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
1.
US counties with more Republican voters should have larger racial gaps, as the Republican share is
believed to be indicative of anti-Black sentiments in the population, and these are considered to be
causing lower Black performance via stereotype threat, micro-aggressions, and various other kinds of
overt and covert discrimination. For the same reason, there should be a negative association between
the Black average test score and the Republican voting share.
2.
Specifically, counties with higher anti-Black racism scores on the Implicit Association Test should
have larger racial gaps.
3.
Counties with a larger White % of the population should have larger Black-White gaps, as Whites are
the ones responsible for the oppression of non-Whites. When they are more numerically powerful,
they are better able to implement their policies that favor Whites over other groups, especially Blacks.
In this study the focus is on the test score gaps which represent cognitive ability in a broad sense. This
is done because many decades of research have shown that cognitive ability is causal for later socioeconomic
outcomes (Belsky et al., 2016, 2018; Hegelund et al., 2019; Herrnstein & Murray, 1994; Marioni et al., 2014;
Strenze, 2007, 2015). Thus, racial gaps in socioeconomic outcomes are largely cognitive ability gaps, or
“skill gaps” as some economists prefer to call them (Neal & Johnson, 1996). Studies that include measures
of general intelligence, whether in the form of scholastic tests or intelligence tests, find that controls for
these measures starkly reduce racial gaps in social status outcomes (Fryer, 2011; Herrnstein & Murray, 1994;
Kirkegaard, 2017; Murray, 2021; Neal & Johnson, 1996; Nordin & Rooth, 2007; Nyborg & Jensen, 2001).
2 Data
We used data from several sources. First, we used scholastic test data from Stanford Education Data Archive
(SEDA;
https://edopportunity.org/about/
) version 4.1. These are county-level average performance
on mandatory state scholastic tests. These have been transformed to a common metric using data from the
federal National Assessment of Educational Progress (NAEP) database. The Black-White test score gap
was 0.65 units weighed by the square root of the population size across all units, while the Hispanic-White
gap was 0.47 units. The scale is intended to be “standardized to this nationally averaged reference cohort
within subject, grade, and year”. The database also provides precomputed racial gap sizes on a composite
measure of socioeconomic status
1
, as well as some covariates of interest such as sample sizes and school
factors. The SES composite is scaled to have “an enrollment-weighted mean of 0 and standard deviation of
1 across all geographic school districts in 2008-2012”. The Black-White and Hispanic-White gaps are 2.45
and 1.47. The SEDA documentation provides details of the construction of the dataset (Fahle et al., 2021).
Figures A1-A7 show maps of the race gaps in test scores and SES, as well as the means.
The maps show that large parts of the country have missing data for gaps. This is due to the relative
absence of Blacks from these areas. As such, the analyses are based on the regions that do harbor sufficient
numbers of Blacks to estimate their mean performance and thus the gap size to Whites.
Second, we used election data from the MIT Election Lab (
https://electionlab.mit.edu/
). This
provides election data at the county level from 2000 to 2020. The datafile provides the vote count for each
presidential election. We calculated the vote shares for the total number of votes in each election. Thus,
the Republican and Democrat vote shares do not exactly sum to 1.00, as some third parties received votes
as well. However, they are usually close: The mean sum for the 2020 election is .982 with a minimum value
of .870.
Third, we used spatial and population count data from the US government census (
https://
www.census.gov/data/datasets/time-series/demo/popest/2010s-counties-total.html#par
_textimage_70769902
,
https://catalog.data.gov/dataset/tiger-line-shapefile-2017-nation
1
“For districts, counties, metropolitan areas and states we use data from the ACS to construct measures of median family
income, proportion of adults with a bachelor’s degree or higher, proportion of adults that are unemployed, the household
poverty rate, the proportion of households receiving SNAP benefits, and the proportion of households with children that are
headed by a single mother. We also combine these measures to construct a single socioeconomic status composite.“ The
composite is calculated based on principal components analysis.
371
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
-u-s-current-county-and-equivalent-national-shapefile
). These were needed to compute pop-
ulation density, which we desired to include in the models, as well as spatial data for creation of the
maps.
Fourth, we used data from Project Implicit (
https://implicit.harvard.edu/implicit/blog
.html
). In a 2018 blogpost (Redford, 2018), they provided a dataset and method for computing county-level
average race implicit association test (IAT) results. Specifically, these are designed to measure people’s
association between positive outcomes and White Americans compared to Black/African Americans. We
used their data and code, which was based on a sample size of a million test takers. Figure 1 shows a map
of the mean race implicit association test scores by county.2
Figure 1: Mean race implicit association test score by county. Figure reproduced from the Project Implicit blog
(https://implicit.harvard.edu/implicit/blog.html).
Unless otherwise noted, all analyses used the square root of the population size or the test population
size as weights. A prior study has found this to be a reasonable compromise between unit weights which
gives large effects to uncertain data from small samples, and sample size-weighting, which results in large
units completely dominating the relationships (Fuerst & Kirkegaard, 2016). This approach results in weights
that are linear transformations of the inverse of the standard error of the mean, as shown in Equation 1
where SE = standard error, SD = standard deviation, and n = sample size.
SEmean =
SD
√n(1)
The supplementary materials contain the R notebook, data, high-quality figures, maps for all the
primary variables, and other materials for the project (
https://osf.io/m8gs5/
). The R notebook can
also be found at https://rpubs.com/EmilOWK/systemic_racism.
3 Results
Before presenting the main results, we present summary statistics for the primary variables in Table 1.
Similarly, we present the correlation matrix between the variables in Table 2. These are intended to provide
the reader with a summary of the bivariate associations.
The correlation matrix shows pervasive associations with variables in various directions. Most of the
correlations have small p values given the large sample size. Some noteworthy relationships are the strong
positive correlations between the Black-White and Hispanic-White gap sizes (about r = .70), which are
furthermore strongly linked to the SES gaps in the same counties (r’s about .50). The race gap sizes are
negatively related to Republican voting shares (r’s -.12 to -.60), to the White population share (r’s -.14 to
2
We downloaded a newer version of the dataset from the website. However, we found the documentation too poor for using
the data and used the older 2018 dataset instead.
372
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Table 1: Summary statistics for the primary variables. SD = standard deviation, MAD = median absolute deviation
(robust SD analog), IAT = Implicit Association Test.
Variable n Mean ±SD Median MAD Min Max Skew Kurtosis
White-Black test gap 1473 0.59 ±0.20 0.58 0.17 -0.04 1.61 0.62 1.41
White-Hispanic test gap 1750 0.43 ±0.21 0.41 0.19 -1.02 1.36 0.40 2.08
White-Black SES gap 1274 2.49 ±0.74 2.51 0.72 0.12 4.62 -0.05 -0.14
White-Hispanic SES gap 1469 1.38 ±0.46 1.34 0.37 0.18 4.39 1.33 4.87
Republican % 2020 3152 0.65 ±0.16 0.68 0.15 0.09 0.96 -0.79 0.14
mean IAT racism 3064 0.34 ±0.15 0.35 0.09 -0.49 1.19 -0.52 4.84
mean White test score 3016 0.11 ±0.21 0.11 0.20 -0.81 1.00 0.13 0.67
mean Black test score 1511 -0.47 ±0.20 -0.48 0.20 -1.14 0.31 0.21 0.13
mean Hispanic test score 1774 -0.27 ±0.21 -0.28 0.20 -1.56 0.76 0.23 1.27
White population % 3214 0.70 ±0.27 0.79 0.24 0.00 1.00 -1.00 0.01
population density log10 3142 -4.77 ±0.78 -4.76 0.63 -7.86 -1.56 0.01 0.61
urban 3214 0.07 ±0.19 0.00 0.00 0.00 1.00 2.90 8.13
suburb 3214 0.11 ±0.24 0.00 0.00 0.00 1.00 2.20 3.88
Table 2: Correlation matrix for the primary variables. Correlations above the diagonal are weighted by the square root
of the population size, correlations below are unweighted. p values are not indicated because almost all associations
have p < .001, only trivial sized associations, |r| < .10, have p > .01 for some variable pairs. Sample sizes vary by the
pairs (pairwise missing data handling). SES = socioeconomic status composite.
1 2 3 4 5 6 7 8 9 10 11 12 13 14
1. White-Black
test gap 0.72 0.56 0.36 -0.53 -0.53 -0.04 0.61 -0.40 -0.20 -0.30 0.52 0.40 0.24
2. White-Hispanic
test gap 0.66 0.28 0.59 -0.59 -0.60 -0.09 0.62 -0.09 -0.47 -0.38 0.47 0.39 0.27
3. White-Black
SES gap 0.45 0.24 0.40 -0.12 -0.15 -0.15 -0.01 -0.62 -0.30 -0.14 0.03 0.21 -0.21
4. White-Hispanic
SES gap 0.35 0.54 0.38 -0.38 -0.40 -0.08 0.29 -0.11 -0.36 -0.31 0.34 0.30 0.07
5. Republican %
2020 -0.43 -0.45 -0.13 -0.28 0.99 0.28 -0.53 0.01 0.12 0.60 -0.66 -0.54 -0.45
6. Republican %
2016 -0.43 -0.46 -0.16 -0.30 0.98 0.27 -0.50 0.04 0.16 0.62 -0.63 -0.53 -0.40
7. Mean IAT 0.05 -0.02 -0.23 -0.04 0.24 0.24 0.02 0.21 0.10 0.38 -0.09 -0.13 0.00
8. White
mean score 0.59 0.53 -0.11 0.25 -0.36 -0.33 0.07 0.48 0.39 -0.22 0.57 0.32 0.49
9. Black
mean score -0.34 -0.04 -0.59 -0.12 0.01 0.05 0.24 0.54 0.62 0.20 0.17 -0.13 0.29
10. Hispanic
mean score -0.14 -0.48 -0.32 -0.31 0.03 0.08 0.05 0.46 0.59 0.28 0.09 -0.18 0.18
11. White
population % -0.18 -0.26 -0.28 -0.27 0.52 0.53 0.31 -0.05 0.29 0.28 -0.38 -0.46 -0.19
12. population
density log10 0.45 0.28 -0.12 0.18 -0.52 -0.47 -0.04 0.33 0.29 0.20 -0.13 0.57 0.63
13. Urban 0.37 0.31 0.10 0.22 -0.43 -0.41 -0.05 0.27 -0.01 -0.06 -0.24 0.50 0.06
14. Suburb 0.24 0.22 -0.24 0.06 -0.39 -0.35 0.00 0.39 0.32 0.19 -0.26 0.59 0.16
-.38), and the implicit association racism measures (about -.10). Associations between Republican voting
share and non-White mean test scores are near zero. These results are contradictory to the predictions
373
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
of the systemic racism model, which predicts positive associations of gap sizes with voting share, white
population share and implicit racism, and negative correlations of % Republican with non-White test scores.
Figures A8-A13 show scatterplots between racial gaps and expected predictors.
The scatterplots support the correlations yet reveal some nonlinearities, which necessitate handling in
regression models. Most notable was the non-monotonic relationship with the IAT score and the gap sizes,
which suggested a maximal gap size at an intermediate IAT score. This finding does not seem to make any
sense from the perspective of systemic racism theory, or the theory behind IATs.
Tables 3 and 4 show a series of regression models that attempt to clarify possible lines of causation
for the race gaps.
Table 3: Regression models for explaining the Black-White test score gap. * p < .01, ** p < .005, *** p < .001.
Values in parentheses are standard errors. Nonlinear terms are handled using natural splines.
Predictor/Model Basic White mean Urbanicity % White IAT State Interaction Nonlinear
Intercept 1.03***
(0.016)
0.78***
(0.020)
0.70***
(0.026)
0.71***
(0.026)
0.71***
(0.027)
0.80***
(0.035)
0.61***
(0.046)
0.71***
(0.043)
Republican %
2020
-0.71***
(0.029)
-0.41***
(0.030)
-0.31***
(0.037)
-0.23***
(0.042)
-0.25***
(0.043)
-0.22***
(0.055)
0.18
(0.085) nonlinear
mean White
test score
0.41***
(0.021)
0.42***
(0.024)
0.4***
(0.025)
0.44***
(0.025)
0.50***
(0.026)
0.48***
(0.026)
0.47***
(0.025)
urban 0.13***
(0.022)
0.12***
(0.022)
0.12***
(0.022)
0.11***
(0.023)
0.09***
(0.023)
0.11***
(0.022)
suburb -0.02
(0.022)
-0.01
(0.022)
-0.01
(0.022)
-0.01
(0.021)
-0.05
(0.022)
-0.02
(0.022)
population
density log10
0.00
(0.010)
0.00
(0.010)
0.00
(0.010)
0.00
(0.010)
0.02
(0.011)
-0.01
(0.010)
White popul. % -0.09***
(0.023)
-0.10***
(0.025)
-0.15***
(0.035)
0.21**
(0.067) nonlinear
implicit racism 0.06
(0.061)
0.09
(0.058)
0.08
(0.058) nonlinear
State fixed effects no no no no no yes yes yes
Republican% *
White% interaction
-0.70***
(0.113)
R2adj. 0.283 0.427 0.450 0.456 0.458 0.569 0.580 0.612
N 1469 1469 1469 1469 1466 1466 1466 1466
The regression results indicate that the directional effect of Republican voting share is consistently
negative across model specifications. In no case does it turn positive. The models with nonlinear effects have
somewhat better adjusted R
2
s than their linear comparison models. However, because the natural splines
cannot be easily interpreted, it is necessary to plot the model predictions to understand them. Accordingly,
Figure A14-A19 shows the model predictions of the three nonlinear predictors for both race gaps.
The plots reveal little support for the systemic racism model even allowing for nonlinear effects.
Republican vote share seems to be largely associated with smaller race gaps, and only after about 60 % of
the vote share does this relationship turn marginally positive. The relationship between IAT scores and race
gaps is near zero, with no notable nonlinearity. Thus, the nonlinearity seen previously (in Figures A10 and
A13) was mostly accounted for by the inclusion of the covariates. White population share continues to have
a slight negative association with cognitive gap sizes, which is somewhat stronger for the values closer to
100 %.
Exploratory modeling indicated a notable interaction effect for Republican vote share and White
population share for the Black-White cognitive gap. However, this did not replicate for the Hispanic gap.
Model predictions from the interaction models are shown in Figures 2 and 3.
The interaction that exists for the Black-White gap is remarkable in that it is opposite to systemic
racism model prediction: the gap shrinks faster when there is both an increasing White population share
374
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Table 4: Regression models for explaining the Hispanic-White test score gap. * p < .01, ** p < .005, *** p < .001.
Values in parentheses are standard errors and p values. Nonlinear terms are handled using natural splines.
Predictor/Model Basic White mean Urbanicity % White IAT State Interact Nonlinear
Intercept 0.92***
(0.015)
0.68***
(0.018)
0.66***
(0.025)
0.70***
(0.025)
0.66***
(0.027)
0.68***
(0.037)
0.64***
(0.046)
0.67***
(0.048)
Republican %
2020
-0.77***
(0.026)
-0.51***
(0.027)
-0.49***
(0.034)
-0.39***
(0.036)
-0.43***
(0.037)
-0.13**
(0.044)
-0.04
(0.072) nonlinear
mean White
test score
0.40***
(0.021)
0.44***
(0.023)
0.46***
(0.023)
0.45***
(0.023)
0.56***
(0.023)
0.56***
(0.023)
0.53***
(0.023)
urban 0.13***
(0.020)
0.09***
(0.021)
0.09***
(0.021)
0.02
(0.020)
0.02
(0.0206)
0.02
(0.019)
suburb 0.02
(0.020)
0.02
(0.020)
0.01
(0.020)
-0.02
(0.018)
-0.03
(0.019)
-0.03
(0.018)
population
density log10
-0.03***
(0.008)
-0.03***
(0.008)
-0.03***
(0.008)
0.01
(0.008)
0.01
(0.009)
0.00
(0.008)
White
popul. %
-0.14***
(0.019)
-0.16***
(0.020)
-0.29***
(0.025)
-0.21***
(0.055) nonlinear
implicit racism 0.21***
(0.054)
0.14**
(0.048)
0.14**
(0.048) nonlinear
State fixed effects no no no no no yes yes yes
Republican% *
White% interaction
-0.15
(0.095)
R2adj. 0.344 0.453 0.468 0.484 0.488 0.631 0.632 0.665
N 1743 1743 1743 1743 1736 1736 1736 1736
Figure 2: Model prediction of the Black-White test score gap from the interaction models. The lines show the linear
fit. The ribbon indicates the 95 % confidence interval. Model predictions for nonlinear terms in the regression models.
Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021). Model predictions are
based on holding the other variables constant at their mean or mode values for numerical and categorical predictors,
respectively. The ribbon indicates the 95 % confidence interval.
375
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure 3: Model prediction of the Hispanic-White test score gap from the interaction models. For explanations, see
Figure 2.
and an increasing Republican vote share. It is this combination where one would expect the more pernicious
effects of racism to influence the results, based on systemic racism theory and intersectionality more broadly.
As we saw in Table 2, the gap sizes are strongly related across test scores and SES. This is shown in
the scatterplot in Figure 4.
Figure 4: Scatterplot of test score gaps and SES gaps. The orange line shows the linear fit and the brown line
shows a nonlinear smoothing fit using a generalized additive model as implemented in geom_smooth() in the ggplot2
package (Wickham et al., 2021). The ribbon indicates the 95 % confidence interval.
The plots show that the relationship is mostly linear. In fact, one might fruitfully ignore the gap
aspect of the issue and focus instead on the mean scores. Doing so reveals the overall pattern very clearly,
376
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
shown in Figure 5.
Figure 5: Scatterplot of average test score and socioeconomic status by race. The red, green, and blue lines show
the linear fits within groups, the orange line shows the overall linear fit, and the brown line shows the overall nonlinear
fit based on the generalized additive model as implemented in geom_smooth() in the ggplot2 package (Wickham et
al., 2021). The ribbon indicates the 95 % confidence interval.
Thus, we see that county race populations higher in test scores are also higher in SES, and this is
true both within groups (the red, green, and blue lines), and overall (orange and brown lines). The very
strong association (r = .87) reveals the strong meritocratic element in social inequality in the United States.
As a robustness test, we also fit regression models for the SES race gaps. The results were mostly
similar to the ones for test scores in terms of the direction of effects. Republican and White fractions were
negatively related to gap sizes in most specifications and otherwise around zero. The only oddity was that
in one specification, IAT predicted the Hispanic gap size, but not when adjusted for state fixed effects.
These model results are shown in full in the technical output.
4 Discussion
The present study had many findings of note. First, racial gaps on scholastic tests and in SES were related
to many other variables. However, the associations were not as expected by the systemic racism model. One
would expect that when more people in a given polity (in this case, a county) have hostile or less favorable
attitudes to Blacks and Hispanics, gaps in scholastic test scores (cognitive ability) and socioeconomic status
would be larger. However, we found that these gaps are actually larger in counties with fewer Republican
voters, and with a lower White population share. This is unexpected because of the well-studied associations
between left-wing attitudes and pro-non-European attitudes (‘antiracist’ in the terminology of systemic
racism writers). These unexpected associations did not disappear entirely when we statistically adjusted for
potential confounders such as population density, urbanicity, and state fixed effects. The clear failure of
these predictions thus throws considerable doubt on the systemic racism model.
Second, we found that race gaps were consistent across variables, i.e., that counties with larger test
score gaps also had larger SES gaps. This is a prediction from the competing meritocratic theory. In
this model, SES outcomes such as income, educational attainment and crime are substantially caused by
differences in intelligence at the individual level, and group-level differences are merely the sum of individual
differences. In support of this, when we combined datasets of average test scores and SES, there was
377
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
a very strong relationship between them, which was also true when we combined data across races in a
single model. Our findings of test score x SES gaps replicate findings in a prior study of US states with
disaggregated cognitive ability and SES by race (Kirkegaard & Fuerst, 2016). The overall finding of very
strong correlations between subnational polity cognitive ability and social status was also previously found
to hold across the 35 or so independent states of the Americas (Fuerst & Kirkegaard, 2016), as well as for
subnational units of Argentina (Kirkegaard & Fuerst, 2017).
The negative associations of White and Republican fractions of the population with race gaps are
difficult to explain. We suggest this reflects largely self-selection patterns in settlement. Historically, Blacks
were mostly located in the South where they were freed from slavery following the conclusion of the US
Civil War in 1865. Later, large numbers of Blacks migrated north and settled in northern states and cities
such as Detroit (Tolnay, 2003), where they occupied mainly low-cost housing in urban settings. These
Blacks have in general been found to be positively selected for human capital traits. Tolnay (2003) writes:
The PUMS [census] files have proven to be most useful for studying the educational selection
of migrants from the South because the education level of an individual changes relatively little
after a certain age. That evidence shows that early black southern migrants (in 1910 and 1920)
were significantly more likely to be literate than blacks who remained in the South. In later
years (from 1940 to 1970) the migrants had significantly higher levels of educational attainment
(years of schooling) than the sedentary southern black population (Tolnay 1998a; see also,
Hamilton 1959, Lieberson 1978b). In contrast, the migrants were less likely to be literate, or
had lower levels of educational attainment, than the black population that they joined in the
North (Tolnay 1998a).
A different way to see this is that Blacks in more northern states are relatively higher in European
ancestry than those in southern states (Bryc et al., 2015). Thus, insofar as European ancestry (admixture) is
also associated with higher cognitive ability and social status, the northern Blacks are expected to be higher
in these traits (Kirkegaard et al., 2019; Lasker et al., 2019). Still, this kind of self-selection does not explain
why the results persist even when we control for the fixed effects of states as well as a host of other factors.
One would have to posit additional self-selection in migration out of the cities into the rural counties that
were previously near-exclusively populated by Whites. Today, Blacks vote about 90 % Democrat (87 % vs.
12 % in the 2020 presidential election when Trump did particularly well with non-White voters in his second
run; BBC, 2020). This pattern then tends to result in inner cities’ vote shares being largely in favor of the
Democrats (r = -.66 for population density and Republican %, Table 2), which is also the case in general
even without the presence of Blacks, because city dwellers tend to be more left-wing regardless of race.
However, some Blacks who ‘escape the ghetto’ by way of attaining well-paid jobs leave and settle
in more affluent, more White, and more Republican areas away from the city centers, and away from the
traditional areas of Black living (e.g. Dakotas). These relatively upper-class Blacks then have children
who are also above average in intelligence, and who consequently perform relatively well on school tests.
Thus, the result of this account is that suburban and rural areas will tend to have a small, but relatively
elite Black population while the inner cities will have large, but mainly below average Black populations.
This sketch of a historical model can only partially explain the results. The same sketch can be applied to
the Hispanic-White gaps, though it is complicated as Hispanics were not freed from slavery but gradually
immigrated in the last 100 years or so. However, the regression analyses show that associations to White
% of population and Republican % of voters remain to some extent even adjusting for three measures of
urban living (population density, urban % and suburban %). Thus, this account seems incomplete and is a
topic for future research.
Another possibility is that the effects seen after controls are causal effects, and that despite their poor
reputation among journalists and academics, Republicans do tend to promote more functional institutions
that result in relatively smaller race gaps. This could be due to rejection of Black ghetto culture or purely
meritocratic (achievement focused) policies in schools, which would promote individual effort and thereby
raise scholastic ability and later life outcomes among Blacks relative to Blacks who received an anti-schooling
influence in school (Riley, 2016). Such an account has been suggested by multiple authors. These results
378
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
can be seen as an example of a lower-performing ethnic group adopting the dominant group’s culture and
this being a net benefit to them. This would be the within society version of beneficial cultural colonialism
(Gilley 2017).
The present study does not have a way to quantify this true causal influence of Whites or Republicans
versus more cryptic self-selection effects. The fact that a large proportion of the effects were reduced given
imperfect controls suggest that the true causal effect sizes are even closer to zero, though not necessarily
zero. A future research design that could test these competing models would be to analyze the polygenic
scores for intelligence and SES traits of Blacks living in various areas. If these confirm the patterns in the
study, it may not be necessary to assign a causal effect to White population shares or Republican voting
shares. If, however, the human capital polygenic scores of Blacks living in these areas are not higher than
those in city areas with low White population shares and low Republican vote shares, the causal effects
become much more plausible.
5 Limitations
First, we used all available data for county-level racial gaps, but had data for only about half the counties,
as the remaining had too low numbers of Black and Hispanic kids to allow for gap estimates (fewer than
20 persons, or if the reliability of the estimate was below .70). It is possible these unseen counties would
change the results. However, this is unlikely due to their small population counts, and thus small weights in
the models.
Second, all the analyses were cross-sectional. Such analyses cannot establish causality by themselves.
Thus, the coefficients of the regressions should not be taken as prima facie estimates of the direct causal
effects of each variable. However, cross-sectional analyses can falsify some causal models that make strong
predictions of associations that should exist. If such are found not to exist, doubt is thrown on the model
that produced them (Bayesian modus tollens). This was the case with White systemic racism theory whose
predictions were nearly uniformly contradicted by the data.
References
BBC. (2020). Us election 2020: Results and exit poll in maps and charts. BBC News. Retrieved from
https://www.bbc.com/news/election-us-2020-54783016 (November 13)
Belsky, D. W., et al. (2016). The genetics of success: How single-nucleotide polymorphisms associated
with educational attainment relate to life-course development. Psychological Science,27, 957–972. doi:
10.1177/0956797616643070
Belsky, D. W., et al. (2018). Genetic analysis of social-class mobility in five longitudinal studies. Proceedings
of the National Academy of Sciences USA,115, E7275–E7284. doi: 10.1073/pnas.1801238115
Bryc, K., et al. (2015). The genetic ancestry of african americans, latinos, and european americans across the
united states. American Journal of Human Genetics,96 , 37–53. doi:
10.1016/j.ajhg.2014.11.010
Fahle, E. M., et al. (2021). Stanford education data archive technical documentation, version 4.1 june 2021.
Retrieved from
https://stacks.stanford.edu/file/druid:db586ns4974/seda_documentation
_4.1.pdf
Feagin, J. R. (2006). Systemic racism: A theory of oppression (1st ed.). Routledge.
Fischer, D. H. (1989). Albion’s seed: Four british folkways in america. Oxford Univ. Press. Re-
trieved from
https://www.google.com/books?hl=en&lr=&id=eq5oAgAAQBAJ&oi=fnd&pg=PP1&dq=
albion%27s+seed&ots=O-fe5Ym0ag&sig=cHk3eK-DjbOU3uP4Yew5TIOKSCY
Fryer, R. G. (2011). Racial inequality in the 21st century: The declining significance of discrimination. In
Handbook of labor economics (Vol. 4, pp. 855–971). Elsevier. doi:
10.1016/S0169-7218(11)02408-7
379
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Fuerst, J., & Kirkegaard, E. O. W. (2016). Admixture in the americas: Regional and national differences.
Mankind Quarterly,56, 255–373.
Gilley, B. (2017). The case for colonialism. Third World Quarterly,38 (10). doi:
10.1080/01436597.2017
.1369037
Hegelund, E. R., et al. (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: 10.1016/j.paid.2019.05.045
Herrnstein, R. J., & Murray, C. A. (1994). The bell curve: Intelligence and class structure in american life
(1st Free Press pbk ed.). Simon & Schuster.
JayMan. (2017). The genetics of the american nations. Retrieved from
https://www.unz.com/jman/
the-genetics-of-the-american-nations/ (February 8)
Kendi, I. X. (2019). How to be an antiracist (1st ed.). One World.
Kirkegaard, E. O. W. (2017). Race, discrimination, cognitive ability and income: Analysis of the add health
dataset. Retrieved from https://osf.io/pwuc4/
Kirkegaard, E. O. W., & Fuerst, J. (2016). Inequality in the united states: Ethnicity, racial admixture and
environmental causes. Mankind Trimestral,56, 580–606.
Kirkegaard, E. O. W., & Fuerst, J. (2017). Admixture in argentina. Mankind Quarterly,57 , 542–580.
Retrieved from http://mankindquarterly.org/archive/issue/57-4/4
Kirkegaard, E. O. W., et al. (2019). Biogeographic ancestry, cognitive ability and socioeconomic outcomes.
Psych,1(1), 1–25. doi: 10.3390/Psychology1010001
Lamstein, A., Johnson, B., Trulia, & Inc. (2020). choroplethr: Simplify the creation of choropleth maps in
r. Retrieved from https://CRAN.R-project.org/package=choroplethr
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: 10.3390/psych1010034
Lüdecke, D., Aust, F., Crawley, S., & Ben-Shachar, M. S. (2021). ggeffects: Create tidy data frames
of marginal effects for “ggplot” from model outputs. Retrieved from
https://CRAN.R-project.org/
package=ggeffects
Marioni, R. E., et al. (2014). Molecular genetic contributions to socioeconomic status and intelligence.
Intelligence,44 , 26-32. doi: 10.1016/j.intell.2014.02.006
McClelland, D. C. (1973). Testing for competence rather than for “intelligence”. American Psychologist ,
28, 1-14. doi: 10.1037/h0034092
Murray, C. A. (2021). Facing reality: Two truths about race in america (1st American ed.). Encounter
Books.
Myrdal, G. (1944). An american dilemma: The negro problem and modern democracy. New York: Harper
& Bros.
Neal, D. A., & Johnson, W. R. (1996). The role of premarket factors in black-white wage differences.
Journal of Political Economy,104, 869-895. doi: 10.1086/262045
Nordin, M., & Rooth, D.-O. (2007). Income gap between natives and second generation immigrants
in sweden: Is skill the explanation? Social Science Research Network. Retrieved from
https://
papers.ssrn.com/abstract=984423
380
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Nyborg, H., & Jensen, A. R. (2001). Occupation and income related to psychometric g. Intelligence,29,
45-55. doi: 10.1016/S0160-2896(00)00042-8
Redford, L. (2018). Mapping county-level geographical variation in implicit racial attitudes. Project Implicit
Blog. Retrieved from https://implicit.harvard.edu/implicit/blog.html (November 4)
Riley, J. L. (2016). Please stop helping us: How liberals make it harder for blacks to succeed (Reprint ed.).
Encounter Books.
Strenze, T. (2007). Intelligence and socioeconomic success: A meta-analytic review of longitudinal research.
Intelligence,35 , 401-426. doi: 10.1016/j.intell.2006.09.004
Strenze, T. (2015). Intelligence and success. In S. Goldstein, D. Princiotta, & J. A. Naglieri (Eds.),
Handbook of intelligence (p. 405-413). Springer. doi: 10.1007/978-1-4939-1562-0_25
Taylor, K.-Y. (2019). Race for profit: How banks and the real estate industry undermined black homeown-
ership. Univ. of North Carolina Press.
Tolnay, S. E. (2003). The african american “great migration” and beyond. Annual Review of Sociology ,
29, 209-232. doi: 10.1146/annurev.soc.29.010202.100009
Wickham, H., et al. (2021). ggplot2: Create elegant data visualisations using the grammar of graphics.
Retrieved from https://CRAN.R-project.org/package=ggplot2
Woodard, C. (2012). American nations: A history of the eleven rival regional cultures of north america
(Reprint ed.). Penguin Books.
Appendix
Figure A1: Hispanic mean scholastic test score. Maps of test scores and SES (socioeconomic stratus composite)
variables. The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test = scholastic test
scores, SES = socioeconomic status composite. Maps made using the choroplethr package (Lamstein et al., 2020).
381
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A2: Black mean scholastic test score. Maps of test scores and SES (socioeconomic stratus composite)
variables. The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test = scholastic test
scores, SES = socioeconomic status composite. Maps made using the choroplethr package (Lamstein et al., 2020).
Figure A3: White mean scholastic test score. Maps of test scores and SES (socioeconomic stratus composite)
variables. The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test = scholastic test
scores, SES = socioeconomic status composite. Maps made using the choroplethr package (Lamstein et al., 2020).
Figure A4: Hispanic-White test score gap. Maps of test scores and SES (socioeconomic stratus composite) variables.
The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test = scholastic test scores, SES
= socioeconomic status composite. Maps made using the choroplethr package (Lamstein et al., 2020).
382
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A5: Black-White test score gap. Maps of test scores and SES (socioeconomic stratus composite) variables.
The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test = scholastic test scores, SES
= socioeconomic status composite. Maps made using the choroplethr package (Lamstein et al., 2020).
Figure A6: Hispanic-White gap in socio-economic status. Maps of test scores and SES (socioeconomic stratus
composite) variables. The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test =
scholastic test scores, SES = socioeconomic status composite. Maps made using the choroplethr package (Lamstein
et al., 2020).
Figure A7: Black-White gap in socio-economic status. Maps of test scores and SES (socioeconomic stratus
composite) variables. The black regions indicate missing data. bw = Black-White, hw = Hispanic-White, test =
scholastic test scores, SES = socioeconomic status composite. Maps made using the choroplethr package (Lamstein
et al., 2020).
383
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A8: Hispanic-White test score gap related to % Republican. Scatterplots of racial gaps and predictor
variables. The orange line shows the linear fit and the blue line shows a nonlinear smoothing fit using a generalized
additive model as implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon
indicates the 95 % confidence interval.
Figure A9: Hispanic-White test score gap related to % White population. Scatterplots of racial gaps and predictor
variables. The orange line shows the linear fit and the blue line shows a nonlinear smoothing fit using a generalized
additive model as implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon
indicates the 95 % confidence interval.
384
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A10: Hispanic-White test score gap related to implicit racism. Scatterplots of racial gaps and predictor
variables. The orange line shows the linear fit and the blue line shows a nonlinear smoothing fit using a generalized
additive model as implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon
indicates the 95 % confidence interval.
Figure A11: Black-White test score gap related to % Republican. Scatterplots of racial gaps and predictor variables.
The orange line shows the linear fit and the blue line shows a nonlinear smoothing fit using a generalized additive
model as implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the
95 % confidence interval.
385
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A12: Black-White test score gap related to % White population. Scatterplots of racial gaps and predictor
variables. The orange line shows the linear fit and the blue line shows a nonlinear smoothing fit using a generalized
additive model as implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon
indicates the 95 % confidence interval.
Figure A13: Black-White test score gap related to implicit racism. Scatterplots of racial gaps and predictor variables.
The orange line shows the linear fit and the blue line shows a nonlinear smoothing fit using a generalized additive
model as implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the
95 % confidence interval.
386
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A14: Black-White test score gap related to % Republican. Model predictions for nonlinear terms in the
regression models. Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021).
Model predictions are based on holding the other variables constant at their mean or mode values for numerical and
categorical predictors, respectively. The line shows the nonlinear smoothing fit using a generalized additive model as
implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the 95 %
confidence interval.
Figure A15: Hispanic-White test score gap related to % Republican. Model predictions for nonlinear terms in the
regression models. Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021).
Model predictions are based on holding the other variables constant at their mean or mode values for numerical and
categorical predictors, respectively. The line shows the nonlinear smoothing fit using a generalized additive model as
implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the 95 %
confidence interval.
387
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A16: Black-White test score gap related to % White population. Model predictions for nonlinear terms in
the regression models. Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021).
Model predictions are based on holding the other variables constant at their mean or mode values for numerical and
categorical predictors, respectively. The line shows the nonlinear smoothing fit using a generalized additive model as
implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the 95 %
confidence interval.
Figure A17: Hispanic-White test score gap related to % White population. Model predictions for nonlinear terms in
the regression models. Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021).
Model predictions are based on holding the other variables constant at their mean or mode values for numerical and
categorical predictors, respectively. The line shows the nonlinear smoothing fit using a generalized additive model as
implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the 95 %
confidence interval.
388
Mankind Quarterly, 64(2), 370-389 2023 Winter Edition
Figure A18: Black-White test score gap related to implicit racism score. Model predictions for nonlinear terms in
the regression models. Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021).
Model predictions are based on holding the other variables constant at their mean or mode values for numerical and
categorical predictors, respectively. The line shows the nonlinear smoothing fit using a generalized additive model as
implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the 95 %
confidence interval.
Figure A19: Hispanic-White test score gap related to implicit racism score. Model predictions for nonlinear terms in
the regression models. Model predictions derived using ggpredict() from the ggeffects package (Lüdecke et al., 2021).
Model predictions are based on holding the other variables constant at their mean or mode values for numerical and
categorical predictors, respectively. The line shows the nonlinear smoothing fit using a generalized additive model as
implemented in geom_smooth() in the ggplot2 package (Wickham et al., 2021). The ribbon indicates the 95 %
confidence interval.
389