The map above raises (at least!) two questions. First, what might predict this distribution of IAT scores? And second, given that the differences between scores (though significant with this large sample) are quite small (range = .341 - .456), are they meaningful?
To address these questions, we examined a number of potential predictors of state-level IAT scores. We selected predictors either because they were theoretically plausible (e.g., state level indices of racial segregation, median income, income inequality, the ratio of White:Black state residents, political preferences) or because they are known to share a similar spatial distribution to the observed IAT scores (e.g., ‘social capital’, status as a former slave-holding state). In addition, we examined whether state-level IAT scores could be used to predict actual behaviors – specifically, numbers of racist tweets recorded after the 2012 Presidential election.
What Predicts this
Distribution of IAT Scores?
Correlational analyses revealed several significant
correlates of state-level IAT scores.
Note that where possible we examined variables around the year 2009
because this was the modal year of IAT data collection.
·
We used the ‘Civic Life Index’as
a measure of social capital. This index
captures a composite of volunteering behavior, neighborhood engagement, voter
participation and civic infrastructure.
There was a strong negative correlation between this measure of social
capital and IAT scores (r (49) = -.576, p < .001)
·
State levels of income inequality (GINI) were positively correlated with IAT scores (r (49) = .409, p < .01).
·
The ratio of White to Black state residents (log transformed because it is skewed) was a strong negative predictor of IAT
scores (r (49) = -.733, p < .001) – the more White relative to Black state
residents, the lower the IAT score.
·
We indexed political preferences in terms each
state’s Electoral College vote in the 2008 Presidential election. States that voted for McCain (M = .405, SD =
.030) vs. Obama (M = .398, SD = .027) did not have significantly
different IAT scores (F (1, 49) = 0.79, p > .30).
·
In contrast, status as a former slaveholding
state (in 1860) was a strong predictor of current day IAT scores. Former slave states (M = .426, SD = .022) had
significantly higher IAT scores than non-slaveholding states (M = .390, SD =
.023; F (1, 49) = 27.45, p <. 001).
·
An index of Black/White segregation and state-level median income did not significantly correlate with IAT scores (rs < |.17|, ps > .20)
Of course, many
of these variables were correlated with each other, so we ran a series of regression
analyses that allowed us to identify the unique effects of each potential
predictor. Entering all of the variables
EXCEPT population composition (ratio of White to Black residents) into a
regression equation showed that two predictors remained significant – the Civic
Life Index negatively predicted IAT scores (β = -.449, p < .01), whereas
status as a former slaveholding state positively predicted IAT scores (β =
.439, p < .01).
When, however, the ratio of White to Black state residents
(log transformed) was entered into a regression equation with these two
variables, the contributions of the Civic Life Index (β = -.173, p > .10)
and slaveholding status (β = .214, p > .10) dropped to non-significance. By far the strongest single predictor of
state-level IAT scores was the ratio of White:Black state residents (β = -.470,
p < .01), accounting for a rather astonishing 53.8% of the variance. States where Whites outnumber Blacks substantially
in the population have lower average IAT scores. In contrast, states where Blacks make up
proportionally more of the population have higher average IAT scores.
Figure:
Relationship between White:Black Population Ratio (log) and IAT Scores
Interpretation.
The strong association between population
composition and IAT scores is consistent with the idea that bias increases when
and where there is greater perceived competition between groups. In the political science and sociology literatures this sort of population composition effect is often termed the
“racial threat hypothesis” (Key, 1949). In states where the African American outgroup
comprises a larger proportion of the population, Whites may perceive greater
competition for political, cultural and economic resources (e.g., Tolbert & Grummel, 2003). Further, due to negative stereotypes about
African Americans, they may also perceive greater risk for cross-race crime (e.g., Eitle, D'Alessio & Stolzenberg, 2002).
In contrast, these findings are inconsistent with any simple
hypothesis that contact between members of different racial groups will lead to
reduced bias. (Of course, social psychologists know this anyway!)
Critically, however, this pattern should be interpreted
within its historical context. It is not
a coincidence that bias is greater in former slaveholding states. In these and other states with large African
American communities, for example, emancipation and extension of full voting rights during the
civil rights era posed a significant threat to White’s political power and
historical dominance. Understandings of
racial identity and the nature of intergroup relationships forged in the past
persist today and likely continue to exert influence (see Acharya, Blackwell & Sen, under review; Jackman, 1994). In a different historical context, however,
the relative size of racial groups might not have the same relationship to
bias.
Do State-Level IAT
Scores Predict Anything?
After the 2012 election, geographers at the University of
Kentucky plotted the distribution of racist Tweets geolocated by state. They created a racist tweet location quotient “that
indicates each state's share of election hate speech tweet relative to its
total number of tweets”. We found that the distribution of tweets they
observed can be predicted by state-level IAT scores, such that states with
higher average IAT scores tended to emit more racist tweets (r (49) = .326, p
<. 05).
We further examined whether this relationship varied as a
function of which candidate (Romney or Obama) won the Electoral College vote in
each state. Perhaps unsurprisingly, numbers
of racist tweets were greater in states that voted for Romney (who lost the
election; β = .409, p < .01; R2 = .25). However, there was also a significant
interaction between state vote and IAT scores (β = 6.78, p < .01; ΔR2 =
.14). The positive relationship between
IAT scores and racist tweets was significant among states that voted for Romney
(β
= .48, p < .05), but was non-significant among
states that voted for Obama (β = -.25, p > .20).
Interpretation.
This pattern is consistent with
contemporary theories of racial bias (e.g., aversive racism). Given that social
norms generally prohibit overtly racist behavior, biased attitudes often only
predict behavior when people feel they have an excuse to behave in a biased
fashion or perhaps when something (like frustration or anger about a preferred
candidate losing an election) reduces control over biased responses. Alternately, the normative climate in states
that voted for Romney might be have felt sufficiently accepting of prejudice to some
individuals following the election to sanction the translation of biased
attitudes into behavior.
Conclusion
We started with two initial questions.
What predicts the
spatial distribution of IAT scores in the United States? A cluster of variables are correlated with
state-level IAT scores, including indices of social capital, income inequality
and historical status as a slaveholding state.
Notably, however, the strongest correlate was the ratio of White to
Black state residents, consistent with the idea that intergroup competition and
identity threat contributes to the type of bias indexed by the IAT.
Given that the
differences between states’ IAT scores are quite small, are they
meaningful? The answer to this
second question appears to be yes.
First, the fact that IAT scores are related to variables like social
capital and population composition suggests that they vary in a systematic
fashion. Second, we found that
state-level IAT scores predicted a form of biased behavior – namely the number
of racist tweets produced in each state following the 2012 election of
President Barack Obama. Thus, although
the observed differences between states are indeed small, the very large sample (1.5
million+) from which these estimates of bias are derived means that they are stable, predictable
and predictive.
Finally, it is important to note that the current analyses sample in a limited way from a much larger universe of possible predictors. We believe it was an informed sampling, but there are certain to be other important predictors out there. Further, a more sophisticated approach would examine effects at the county-level. We are currently engaged in some county-level analyses, and hope to report findings in the near future.
Finally, it is important to note that the current analyses sample in a limited way from a much larger universe of possible predictors. We believe it was an informed sampling, but there are certain to be other important predictors out there. Further, a more sophisticated approach would examine effects at the county-level. We are currently engaged in some county-level analyses, and hope to report findings in the near future.
The Data
If you'd like to take a look yourself or have ideas about additional predictors (urban/rural differences being an obvious contender), the data for these analyses can be downloaded here. And many thanks to Brian Nosek and Project Implicit for providing open access to the IAT data!