Study on Gamma Bias Paul Connor
Study on Gamma Bias Paul Connor
Study on Gamma Bias Paul Connor
Manuscript version of
Funded by:
• Haas Business School, XLab
• National Science Foundation
© 2022, American Psychological Association. This manuscript is not the copy of record and may not exactly
replicate the final, authoritative version of the article. Please do not copy or cite without authors’ permission.
The final version of record is available via its DOI: https://dx.doi.org/10.1037/pspa0000314
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Running head: INTERSECTIONAL IMPLICT BIAS
Intersectional Implicit Bias: Evidence for Asymmetrically Compounding Bias and The
Abstract
Little is known about implicit evaluations of complex, multiply categorizable social targets.
Across five studies (N = 5,204), we investigated implicit evaluations of targets varying in race,
gender, social class, and age. Overall, the largest and most consistent evaluative bias was pro-
lower-class biases. By contrast, we observed less consistent effects of targets‘ race, no effects of
targets‘ age, and no consistent interactions between target-level categories. An integrative data
analysis highlighted a number of moderating factors, but a stable pro-women/anti- men and pro-
upper-class/anti- lower-class bias across demographic groups. Overall, these results suggest that
implicit biases compound across multiple categories asymmetrically, with a dominant category
(here, gender) largely driving evaluations, and ancillary categories (here, social class and race)
exerting relatively smaller additional effects. We discuss potential implications of this work for
Keywords: implicit bias, intersectionality, social class, person perception, social cognition
3
INTERSECTIONAL IMPLICIT BIAS
Intersectional Implicit Bias: Evidence for Asymmetrically Compounding Bias and The
categories, including race, gender, social class, and age (Greenwald & Lai, 2020; Nosek, 2005).
These biases may have weighty social consequences, influencing decision making in contexts
including employment, medicine, and voting (e.g., Greenwald, Banaji, & Nosek, 2015; Jost et
al., 2009).
such as race, gender, social class, and age. Yet within the empirical literature on implicit bias,
biases regarding such categories have typically been studied in isolation from each other, and
most measures of implicit bias have been designed to isolate and measure biases regarding a
single binary categorical preference at a time. For example, Nosek (2005) employed Implicit
Association Tests (IATs; Greenwald, McGhee, & Schwartz, 1998) to demonstrate that US
participants display implicit evaluative biases favouring White targets over Black targets, women
over men, the rich over the poor, the young over older adults, and many others. However, IATs
measure only a single categorical preference at a time, and do not speak to how multiple
identities jointly contribute to implicit bias. Does a White, rich, young woman prompt implicit
evaluations four times more positive than a Black, poor, old man? Are some social categories
more influential than others? Do the categories interact with each other, such that, for example,
implicit gender bias operates differently depending on the race, social class, age, weight, or
To date, psychologists have produced few answers to these questions, despite the rising
prominence of a intersectional approaches within psychological science (e.g., Cole, 2009; Goff,
& Kahn, 2013; Kang & Bodenhausen, 2015). There is, however, considerable evidence that
implicit evaluations are sensitive to multiple aspects of target stimuli. Wittenbrink, Judd, and
Park (2001) found implicit racial bias to be moderated by the visual contexts in which targets
were presented. When Black and White targets were depicted on a street corner, participants
displayed greater anti-Black bias compared to when targets were depicted inside a church.
Similarly, Barden, Maddux, Petty, and Brewer (2004) found moderation of implicit bias by
visual context and targets‘ clothing. When Black and White targets were depicted inside a jail,
4
INTERSECTIONAL IMPLICIT BIAS
participants displayed pro-White bias when targets were shown in prison clothes, but pro-Black
bias when targets were shown in suits and ties. In keeping with this theme of moderation,
participants showed greater implicit bias against Black targets with more racially prototypical
features (Livingston & Brewer, 2002), and toward Black targets with neutral facial expressions
compared to smiling Black targets (Steele, George, Cease, Fabri, & Schlosser, 2018). Each of
these findings suggests that implicit evaluative biases respond to multiple variables within target
stimuli. By implication, when targets are multiply categorizable––as in most everyday social
interactions––it follows that implicit evaluations will likely be shaped by multiple dimensions of
social categorization.
Several schools of thought have considered how intergroup biases respond when multiply
social categories are displayed by social targets (for recent reviews, see Nicolas, de la Fuente, &
Fiske, 2017, and Petsko & Bodenhausen, 2019). Here, we consider in detail select treatments,
focusing upon those most relevant to the present work and results.
One thesis is that negative and positive biases compound when multiple social identities
are displayed simultaneously. In early work, Brown and Turner (1979) relied on Tajfel and
Turner‘s (1979) social identity theory to predict that separate intergroup biases would combine
additively in the presence of multiple dimensions of social categorization. Their reasoning held
that intergroup bias will increase in a linear fashion according to the number of dimensions on
which a social target is perceived to be an out-group member, and decrease according to the
number of dimensions on which they are perceived as an in-group member. A similar thesis is
the averaging model of Singh, Yeoh, Lim, and Lim (1997), which proposes that intergroup bias
is a function of the number of perceived out-group memberships divided by the total number of
Other scholars have suggested that biases may compound across categories in interactive
ways. Grounded in the writings of Black feminist activist Frances Beale (1970), Ransford (1980)
proposed the multiple jeopardy/advantage hypothesis, which posits that individuals belonging to
multiple stigmatized social categories are vulnerable to ‗multiple jeopardy:‘ a negative bias that
exceeds the sum of the negative biases associated with each category. By contrast, individuals
belonging to multiple positively-valued social categories may benefit from ‗multiple advantage:‘
5
INTERSECTIONAL IMPLICIT BIAS
a positive bias that exceeds the sum of the positive biases associated with each category (see also
Almquist, 1975; King, 1988; Landrine, Klonoff, Alcaraz, Scott, & Wilkins, 1995). In her widely
multiple jeopardy in the US legal system: despite General Motors hiring disproportionately fewer
Black women, the company was exculpated of both race and gender discrimination due to
employing sufficient numbers of (White) women and (male) Black people (DeGraffenreid v.
women in the USA; Cooper, 2015). Within this literature, however, it has not always been clear
social categories, or simply that individuals with multiple marginalized social identities suffer
intersectionality have at times been divided on the question of whether the concept can or should
be reduced to these kinds of quantitative predictions (e.g., Cole, 2009; Bowleg, 2008).
evidence has been most consistent with multiple additive main effects on intergroup bias
compounding across different social categorizations (e.g., Crisp, Hewstone, & Rubin, 2001,
Study 1; Hewstone, Islam, & Judd, 1993; Islam & Hewstone, 1993, Study 2; Singh, Yeoh, Lim,
& Lim, 1997; Vanbeselaere, 1991; van Oudenhoven, Judd, & Hewstone, 2000). At other times,
evidence has been consistent with multiplicative disadvantages stemming from combined
stigmatized social identities (e.g., Brown & Turner, 1979; Diehl, 1990; Marcus-Newhall, Miller,
Holz, & Brewer, 1993; Vanbeselaere, 1991), or with multiplicative advantages stemming from
combined positively- valued social identities (Brewer, Ho, Lee, & Miller, 1987; Eurich-Fulcher,
Thus, despite some ambiguity regarding the presence and pattern of interaction effects,
theories of compounding bias make clear predictions with regard to the specific sub-groups of
multiply categorizable targets that should evoke the most positive or negative implicit
evaluations. In the case of implicit bias, for example, prior evidence suggests that Americans‘
implicit evaluative biases typically favour White over Black targets (Nosek, Banaji, &
6
INTERSECTIONAL IMPLICIT BIAS
Greenwald, 2002), women over men (Richeson & Ambady, 2001, Rudman & Goodwin, 2004),
the upper class over the lower class (Horwitz & Dovidio, 2017; Rudman, Feinberg, & Fairchild,
2002), and the young over older adults (Nosek, 2005). Theories of compounding bias therefore
predict that among targets varying in race, gender, social class, and age, the most negative
implicit evaluative biases should be displayed toward lower-class, older Black men, whereas the
most positive biases should be displayed toward upper-class, younger White women.
Category Dominance
Other researchers have challenged the claim that separate biases will necessarily
compound in additive or interactive ways toward multiply categorizable targets.1 One alternate
view is the category dominance model (Macrae, Bodenhausen, & Milne, 1995), which is
premised on the notion that due to the complexity of social stimuli, humans must by necessity act
as ‗cognitive misers‘ (Fiske & Taylor, 1991). When facing multiply categorizable targets, this
view holds, people will often rely on a single social category to guide social perception. Which
specific category becomes dominant depends on many factors, such as the situational or chronic
salience of different categories, the goals of perceivers, and/or perceivers‘ prejudices. Once the
dominant category is activated, it will inhibit the activation of competing categories. In support
of this, Macrae and colleagues showed that when participants were primed with a specific social
category (i.e., Asian or woman) and observed a multiply categorizable target (i.e., an Asian
woman), concepts associated with the primed category became more cognitively accessible,
while concepts associated with the non-primed category became less cognitively accessible (see
The category dominance model therefore predicts that in evaluations of targets varying in
race, gender, social class, and age, a single dominant categorization will drive bias. Importantly,
the model does not necessarily predict what the dominant category will be––if no specific
category is primed by researchers, the dominant category will depend upon the perceivers‘
1
Other perspectives that challenge the notion of compounding bias include Urada, Stenstrom,
and Miller‘s (2007) threshold-based feature detection model, and Kang and Chasteen‘s (2009)
category salience-based selective inhibition model. For the sake of brevity, we do not discuss
these theories in the present manuscript, though our data is arguably relevant to, and fails to
Select studies have investigated implicit bias toward multiply categorizable targets.
Thiem, Neel, Simpson and Todd (2019) used a weapon identification task (Payne, 2001) and
sequential priming tasks to measure automatic associations between weapons and headshots of
targets varying in race (Black and White), gender, and age. Consistent with compounding bias
accounts, each social category influenced responses, with participants displaying a greater
tendency to associate Black, male, and adult targets with weapons compared to White, female,
and child targets. Additionally, there was some evidence of a multiplicative multiple-jeopardy
effect, with Black male targets appearing to evoke stronger associations with threat than could be
explained by main effects of race and gender alone. Similarly, Perszyk, Lei, Bodenhausen,
Richeson, and Waxman (2019) used the Affective Misattribution Procedure (AMP; Payne,
Cheng, Govorun, & Stewart, 2005) to measure children‘s implicit evaluations of headshots of
child targets varying in race (White and Black) and gender. In this study a race × gender
interaction emerged, with Black boys eliciting more negative evaluations than could be
Other work has considered the intersecting effects of race and class. Moore‐Berg,
Karpinski, and Plant (2017) presented images of the upper bodies of targets varying in race
(Black and White) and social class (signalled via targets‘ wearing either t-shirts or suits) within a
‗shoot/don‘t-shoot‘ task (Correll, Park, Judd, & Wittenbrink, 2002). Similarly, Mattan, Kubota,
Li, Venezia, and Cloutier (2019) used an Evaluative Priming Task (EPT; Fazio, Sanbonmatsu,
Powell, & Kardes, 1986) to measure implicit evaluations of headshots of targets varying in race
(Black and White) and background color (red and blue), with participants trained to associate
background colors with higher or lower social status. The results of these studies varied, with
five unique patterns of results emerging from five separate experiments. However, one consistent
result was that in each experiment, upper-class White targets were relatively favored by
responses (though not always more so than lower-class White targets or upper-class Black
targets). These studies can therefore also be considered broadly consistent with compounding
bias models, with upper-class Whites appearing to be the sub-group most favored by displayed
biases.
By contrast, other studies have yielded results more consistent with the category
dominance model. Mitchell, Nosek, and Banaji (2003) presented Black athletes and White
8
INTERSECTIONAL IMPLICIT BIAS
politicians as stimuli within an IAT, but had participants categorize targets either via profession
(Athlete vs. Politician) or race (Black vs. White). When targets were categorized by profession,
biases favoured Black athletes, but when targets were categorized by race, biases favoured the
White politicians. The same authors also presented Black female and White male targets within a
Go/No-Go Association Test (Nosek & Banaji, 2001), and manipulated the relative salience of
targets‘ race and gender. Results indicated that when race was salient, participants evaluated
White males more positively than Black females, but when gender was salient, participants
evaluated Black females more positively than White males. Similarly, Yamaguchi and Beattie
(2019) found that when Black and White female and male targets were categorized according to
race within IATs, participants displayed substantial anti-Black/pro-White implicit racial bias, but
little implicit gender bias. But when targets were categorized according to gender, participants
displayed pro-female/anti- male implicit gender bias, but little implicit racial bias.
Further evidence suggests that the direct manipulation of category salience is not always
necessary for a single category to dominate responses to multiply categorizable targets. Jones
and Fazio (2010) used a weapon identification task to measure participants‘ tendency to perceive
objects as guns versus tools while exposed to images of primes varying in race (Black and
White), gender, and occupational status (high or low, e.g., professor, sanitation worker). In this
study, participants instructed to attend to primes‘ race displayed an implicit racial bias were
relatively more likely to perceive guns/tools while exposed to Black/White primes, but showed
little gender- or occupation-based bias. However, when participants were not instructed to attend
to any specific social category, the only bias displayed was gender-based, with participants
Finally, other researchers have argued that category dominance in implicit evaluation
tasks also depends on the task employed. Gawronski, Cunningham, LeBel, and Deutsch (2010)
measured implicit evaluations of targets varying in race (Black and White) and age via EPTs and
AMPs, while instructing participants to attend either to targets‘ race or age. Results suggested
that non-attended categories affected evaluations on the AMP but not the EPT, leading the
authors to argue that tasks structured to induce response interference––such as the EPT––may be
especially conducive to category dominance, whereas other tasks such as the AMP are not.
understanding how implicit evaluative bias operates toward multiply categorizable targets is
likely to be critical to understanding how it operates in everyday life. However, current evidence
concerning implicit bias and multiply categorizable targets is inconclusive. Whereas some work
supports theories of compounding bias, and suggests that implicit biases tend to compound
across multiple social categories, other work aligns better with the category dominance
perspective, and suggests that implicit evaluations are often driven by a single dominant
categorical dimension.
full-body target photographs of males varying in race (Black or White) and social class status. In
Study 2, we extended on this approach, and incorporated target images varying in race, gender,
and social class, as well as a data-driven approach to determine the primary dimensions of
perceived target-level variation and their respective influence on implicit evaluations. In Study 3,
we again measured implicit evaluations of targets varying in race, gender, social class and age,
but presented targets via full-body or upper-body photographs, and tightened experimental
control over potential confounds by shuffling targets‘ faces and bodies. In Study 4, we tested the
generalizability of our results by obtaining data from two nationally representative samples of
US adults, and by comparing results across different measurement methods. Finally, in Study 5
we conducted an integrative data analysis of the data from Studies 2-4 to test the extent to which
patterns of results differed among different sub-groups of respondents, and to better elucidate the
The present research offers theoretical, empirical, and methodological advances for the
study of intersectional implicit bias. At the theoretical level, this work presents a novel
compounding bias and category dominance. At the empirical level, the present work is to our
knowledge the first to measure implicit evaluations of targets systematically varying in the
variables of race, gender, social class, and age, each of which tend to be simultaneously
perceptible among the majority of real-world social targets. And at the methodological level, the
present work is to our knowledge the first to focus specifically upon measuring and modelling
10
INTERSECTIONAL IMPLICIT BIAS
implicit evaluations of multiply categorizable targets at the individual target level, which we
argue carries multiple advantages over previous approaches. All data and code used in the
current project are accessible via the Open Science Framework (https://osf.io/sbpna/).
Study 1
targets varying in race (Black or White) and social class. Theories of compounding bias predict
that pro-White/anti-Black biases and pro-upper-class/anti- lower-class biases should both occur,
resulting in lower-class Black targets being evaluated most negatively, and upper-class White
targets being evaluated most positively. They also suggest possible interaction effects, with
either lower-class Black targets producing especially negative responses (multiplicative multiple
multiple advantage). Conversely, the category dominance model suggests that either race or
social class will emerge as the dominant category driving implicit bias.
We gathered 130 full-body color photographs of Black and White adults (60 Black, 70
White). Targets appeared on plain white backgrounds facing forward with neutral expressions.
Photographs were presented to 1788 U.S. adults recruited via MTurk, who rated the photographs
on perceived yearly income (ICC = 0.43), perceived age (ICC = 0.70), and whether they
perceived targets to be Black (ICC = 0.88) or White (ICC = 0.95). Raters offered judgments of
an average of 29.73 (SD = 13.61) randomly selected photographs, and each photo was rated on
race (Black and White) and income, but matched in age (see Figure 1). In each study, targets‘
mean perceived income varied significantly across class categories (all p < .001) but not race
categories (all p > 0.69), whereas targets‘ mean perceived race varied significantly across race
categories (all p < .001) but not class categories (all p > .08).2 Additionally, there were no
2
The p value of 0.08 referred to resulted from a t-test comparing Study 1b‘s 16 lower-class and
16 upper-class targets on their mean categorizations as White (see the bottom-left bar plot in
Figure 1). Although not ideal, this result is un-problematic for interpreting Study 1b‘s results. As
shown in Figure 2, Study 1b‘s Black targets (who were categorized as White 3% of the time)
produced more positive evaluations than Study 1b‘s White targets (who were categorized as
11
INTERSECTIONAL IMPLICIT BIAS
significant interactions between race and class categories in predicting perceived income or race
(all p > 0.19), and no significant main effects or interactions of race and class categories in
Figure 1. Target groups used in Studies 1a and 1b, and figures displaying raters‘ judgments of
perceived income, age, and race ratings of each group. Bars indicate 95% confidence intervals.
Participants for Study 1a (N = 307, 196 women, 100 men, 11 missing gender data, Mage =
20.3, SDage = 1.9, 129 Asian,3 125 White, 28 Latino, 9 Black, 5 other race, 11 missing race data)
and Study 1b (N = 533, 340 women, 170 men, 1 non-binary, 22 missing gender data, Mage = 20.5,
SDage = 2.63, 268 Asian, 173 White, 54 Latino, 10 Other race, 6 Black, 22 missing race data)
White 91% of the time). It is therefore highly unlikely that participants responded more
positively to the upper-class targets (who were categorized as White 50% of the time) than the
lower-class targets (who were categorized as White 43% of the time) due to a race confound.
3
Our demographic survey did not delineate between sub-categories of Asian-identifying
were undergraduates who participated for course credit. Study 1a used a within-subjects design,
with participants‘ responding to all six of the target groups in a randomized order, whereas Study
1b used a between-subjects design, with participants randomly assigned to respond to one of the
Single-Target IATs
In both studies we measured implicit evaluations of target groups via evaluative Single
Target IATs (ST-IATs; Bluemke & Friese, 2008; Wigboldus, Holland, & van Knippenberg,
2004).4 Each ST-IAT began with a practice block, in which the labels ―Good‖ and ―Bad‖
appeared at the top left and right of participants‘ computer screens. Across 20 trials participants
then classified words appearing on their screens as either good (e.g., Beautiful) or bad (e.g.,
Agony) as quickly as possible via timed computer key presses. Following this, the word
―Person‖ also appeared at either the top left of screens (in ‗compatible‘ blocks), or the top right
of screens (in ‗incompatible‘ blocks), and participants categorized words as ―Good‖ or ―Bad‖
and targets as a ―Person.‖ Participants were randomly assigned either to complete two
compatible blocks (of 20 then 40 trials) followed by two incompatible blocks (of 20 then 40
target groups and the concepts of wealth and poverty. In this measure the labels ―Good‖ and
―Bad‖ were replaced with ―Wealth‖ and ―Poverty,‖ and the positively and negatively valanced
words were replaced with words evoking wealth (e.g., Rich, Wealth, Affluent) and poverty (e.g.,
Table 1
Single Target IAT procedure
Block Task description Left key (E) Right key (I) Trials
1 Practice block Positivea/Wealth words c Negativeb /Poverty d words 20
2 Compatible block 1 Positive/Wealth words + target images Negative/Poverty words 20
3 Compatible block 2 Positive/Wealth words + target images Negative/Poverty words 40
4 Incompatible block 1 Positive/Wealth words Negative/Poverty words + target images 20
5 Incompatible block 2 Positive/Wealth words Negative/Poverty words + target images 40
4
ST-IATs are highly similar to the Single-Category IAT (SC-IAT) introduced by Karpinski and
Steinman (2006). We follow Bluemke and Friese (2008) in distinguishing between the tasks on
the basis that the SC-IAT uses an in-task response maximum latency window while the ST-IAT
does not. In the present manuscript, we did not use a limited response latency window, so
(Greenwald, Nosek, & Banaji, 2003). On this measure, higher/lower scores indicate greater
automatic associations between target groups and positive/negative concepts in evaluative ST-
IATs, and between target groups and wealth/poverty in wealth ST-IATs. D Scores from ST-IATs
display comparable psychometric properties to the more commonly used two-category IAT
(Greenwald & Lai, 2020). We estimated the average split-half reliability of the valence and
wealth ST-IATs to be 0.665 and 0.68, respectively (the valence ST-IAT figure combines data
from Studies 1a and 1b). All implicit measures in the present manuscript were administered
Demographics
In both studies demographic information (age, gender, race, and political orientation) was
Results
For Study 1a we fitted a 2 (target race: Black, White) × 3 (target class: low, middle, high)
repeated measures ANOVA predicting participants‘ D scores on the evaluative ST-IAT. For
Study 1b we fitted separate 2 (target race: Black, White) × 2 (target class: low, high) independent
samples Analyses of Variance (ANOVA) predicting D scores on both the evaluative and wealth
ST-IATs. All analyses were conducted in R version 3.6.1 (R Core Team, 2019).
Evaluative ST-IATs
In both studies there was a significant main effect of targets‘ social class, Study 1a:
F(2,598) = 18.93, p < .001 , = 0.02, Study 1b: F(1,516) = 5.27, p = 0.02, = 0.01, with
participants responding more positively to upper-class targets than lower-class targets. In Study
1a, participants responded more positively to upper-class targets than middle-class targets and to
middle-class targets than lower-class targets, although this latter difference did not reach
5
These figures (and all split-half reliability figures reported in this paper) are based on average
split-half correlations from 100 random splits of the ST-IAT data corrected according to the
statistical significance (see Figure 2). By contrast, there were no significant main effects of race
in either study: Study 1a, F(1,299) = 2.07, p = 0.15, = 0.001, Study 1b, F(1,516) = 2.47, p =
0.12, = 0.005, nor any significant race × class interactions: Study 1a, F(2,598) = 0.28, p =
Wealth ST-IAT
In the wealth ST-IAT in Study 1b, there was again a main effect of target class, F(1,518)
= 23.72, p < 0.001 , = 0.04, with upper-class targets producing stronger relative associations
with wealth than lower-class targets (see Figure 2). There was no significant effect of target race,
F(1,518) = 0.0008, p = 0.98, < 0.001, and no significant race × class interaction, F(1,518) =
Figure 2. Mean IAT D scores by target group for Studies 1a and 1b. Bars indicate 95%
confidence intervals. Cohens‘ d and statistical significance of t tests between social class groups
collapsing across races are also reported (NS = not significant, * = p < 0.05, *** = p < 0.001).
(ANOVAs) in both studies were well-powered to detect small main and interaction effects. Study
1a achieved 80% power to detect smaller effects ( = 0.005) than Study 1b ( = 0.015). See
Discussion
In Studies 1a and 1b, participants evaluated targets with higher perceived incomes more
positively than targets of lower perceived incomes. By contrast, evaluations were not
significantly affected by target groups‘ race, nor did we observe any significant race × class
interaction effects. These results are most consistent with the category dominance model, and
diverge from previous findings regarding the effects of race and class on implicit bias (Mattan et
15
INTERSECTIONAL IMPLICIT BIAS
al., 2019; Moore-Berg et al., 2017; though Mattan and colleagues observed a similar result in
their third study). Those past studies, it is worth noting, did not hold perceived social class
constant across races. By contrast, our Black and White target groups were pre-matched on
explicit ratings of perceived incomes, and our wealth ST-IAT in Study 1b verified that automatic
associations between target groups and wealth did not differ significantly across races. Our use
of full-body target photographs may also have been a factor, elevating the influence of targets‘
bodies––a primary source of social class cues (e.g., Becker, Kraus, & Rheinschmidt-Same, 2017;
Gillath, Bahns, Ge & Crandall, 2012; Schmid-Mast & Hall, 2004)––relative to the influence of
targets‘ faces––a source of race cues––on evaluations. Both previous studies used stimuli which
devoted a more equal share of visual space to cues of race and class.
Study 2
race (we incorporated Asian as well as Black and White targets6 ), as well as on social class,
gender, and age. We also addressed whether the lack of pro-White/anti-Black implicit racial bias
observed in Study 1 might have occurred simply due to our specific sampling population
possessing little pro-White/anti-Black implicit bias (Studies 1a‘s and 1b‘s samples were largely
evaluations of target groups displaying three different races (e.g., Asian, Black, and White), two
genders (female vs. male), two levels of social class (high vs. low), and two levels of age (old vs.
young) using the methods of Study 1 would require 24 separate experimental conditions. This
target groups.
the level of individual targets via Target D Scores. This measure relies on a similar logic to a
6
The choice to include Asian rather than another race of targets was partly pragmatic, due to
their availability within our photograph database, but was also informed by an interest in the
potential for our majority Asian student samples to show greater racial bias if their racial ingroup
standard ST-IAT D Score, but rather than measuring an individual participant‘s response to a
target group in compatible vs. incompatible trials, Target D Scores measure an entire sample‘s
response to an individual target in compatible vs. incompatible trials. This allows researchers to
study systematic variation in implicit evaluations both between and within groups of targets, and
thereby to more efficiently model the simultaneous effects of multiple simultaneously varying
target-level variables.
social class, and age, as these categories are perceptible in most social interactions, and have
been the focus of much of the previous work into implicit evaluative bias. However, we did not
wish to presume in advance how participants would spontaneously perceive and categorize such
complex targets. In recent work, Koch, Imhoff, Dotsch, Unkelbach, and Alves (2016) studied the
content of social perceptions in a data-driven way. Rather than rating targets on pre-chosen traits,
participants provided ratings of the similarity/dissimilarity of pairs of targets,, which were then
subjected to Multidimensional Scaling (MDS, for a review, see Borg & Groenen, 2005) to
identify the primary dimensions underlying participants‘ judgments. We used this method to
ascertain whether indeed race, class, gender and age spontaneously shape implicit bias. Study 2
Target Photographs
We selected 54 images (18 Asian8 , 18 Black, and 18 White targets) from a large database
of 726 full-body target images (54 Asian female, 63 Asian Male, 115 Black female, 154 Black
7
We deviated from this pre-registration by predicting Target D Scores calculated
according to the algorithm described below rather than logged response times between 300ms
and 10,000ms. This deviation reflects our evolving understanding of how best to model and
analyze ST-IAT data at the individual target level, and had only a minor impact on conclusions
8
All Asian targets used in the present manuscript appear subjectively to be of prototypically East
Asian appearance, though it is a limitation of the present manuscript that neither our data nor the
Chicago Face Database norming data relied upon for the Studies 3 & 4 targets distinguishes
male, 140 White female, 200 White male). In addition to the images, the database contains
490,359 explicit ratings of the targets made by 3,311 US adults (1,875 women, 1031 men, 24
non-binary, 381 missing gender data, Mage = 23.8, SDage = 8.6, 1,116 Asian, 1,089 White, 414
Latino, 117 Black, 575 other race or unreported) on 24 different personality and demographic
traits selected as central to person perception. Traits measured were: warm (ICC = 0.23),
competent (ICC = 0.31), honest/moral (ICC = 0.13), dominant (ICC = 0.16), submissive (ICC =
0.14), anxious/easily upset (ICC = 0.08), open to new experiences/complex (ICC = 0.15),
conventional/uncreative (ICC = 0.09), attractive (ICC = 0.33), income (ICC = 0.39), education
(ICC = 0.27), occupational prestige (ICC = 0.39), subjective socioeconomic status (ICC = 0.43),
age (ICC = 0.72), political orientation (ICC = 0.26), and race (measured via a multiple choice
categorical response; ICCs for dummies indicating Asian, Black, and White categorizations =
For each race (Asian, Black, and White), we selected 9 female and 9 male targets
varying in social class and age. There was some minor non-orthogonality between target-level
variables (maximum r = 0.15, see Table 2). However, we were able to control for such
imbalances by estimating effects of targets‘ race while controlling for their precise levels of
perceived social class, and vice versa, as is done in conjoint experimental designs with
Table 2
Descriptive statistics of targets chosen for Study 2
Correlations 1. 2. 3. 4. 5. 6.
1. Asian categorization
2. Black categorization -0.49
3. White categorization -0.52 -0.47
4. Female a -0.01 -0.01 0.03
5. Age -0.02 -0.01 0.03 -0.04
6. SESb 0.15 -0.15 -0.01 -0.002 -0.02
Descriptives
M(SD) Overall 0.33(0.47) 0.31(0.45) 0.32(0.42) 0.5(0.5) 43.6(12.93) 0(1)
M(SD) Asian Females 0.97(0.03) 0.01(0.03) 0.02(0.05) 1(0) 40.59(11.34) 0.18(0.67)
M(SD) Asian Males 0.99(0.03) 0(0) 0.01(0.02) 0(0) 46.05(13.52) 0.24(0.87)
M(SD) Black Females 0.01(0.02) 0.91(0.15) 0.08(0.08) 1(0) 44.87(13.42) -0.21(1.08)
M(SD) Black Males 0.01(0.03) 0.95(0.05) 0.01(0.02) 0(0) 41.6(14.35) -0.15(1.20)
M(SD) White Females 0(0) 0.01(0.02) 0.89(0.15) 1(0) 43.84(13.38) 0.02(1.01)
M(SD) White Males 0(0) 0.01(0.02) 0.9(0.1) 0(0) 44.64(14.37) -0.08(1.27)
a
Female is a manually coded dummy (1 = Female, 0 = Male)
b
SES is a z-scored average of z-scored ratings on income, education, occupational prestige, and subjective
SES
18
INTERSECTIONAL IMPLICIT BIAS
Participants were 371 undergraduate students who participated for course credit (281
women, 66 men, 1 non-binary, 23 missing gender data, Mage = 20.44, SDage = 2.5, 194 Asian, 93
ST-IATs
described above. The three ST-IATs used as target stimuli the 18 Asian, 18 Black, and 18 White
Race IAT
Participants also completed a two-category Race IAT using black-and-white partial face
images of Black and White targets as stimuli.9 This involved a similar procedure to the ST-IAT,
except that in test trials the labels ―White American‖ and ―Black American‖ appeared on
opposite sides of participants‘ screens, alongside the labels ―Good‖ and ―Bad.‖ Participants
categorized positive words or White faces via a single computer key and negative words or
Black faces via an alternative key (compatible trials), or categorized positive words or Black
faces via a single computer key, and negative words or White faces via an alternative key
algorithm, with higher D scores (split-half reliability = 0.75) indicating anti-Black implicit bias.
The order of the ST-IATs and the Race IAT was randomly counter-balanced.
Difference Ratings
Following the implicit measurement tasks, participants were presented with 60 randomly
selected pairs of the 54 targets and asked to indicate ―how different or similar are these people‖
on 0-100 sliders ranging from ―Very Similar‖ to ―Very Different.‖ This resulted in an average of
14.8 ratings (SD = 3.57) each of 1,431 possible target pairs (ICC = 0.29).
Demographics
measured via the MacArthur ladder measure (Adler, Epel, Castellazzo, & Ickovics, 2000).
Results
Multi-Dimensional Scaling
9
We used the ―Racism IAT‖ available from Millisecond.com
https://www.millisecond.com/download/library/iat/raceiat/)
19
INTERSECTIONAL IMPLICIT BIAS
We computed the mean perceived difference between each of the 1,431 unique target
pairs and subjected the resulting distance matrix to MDS using the majorization approach
assuming an interval scale (SMACOF; De Leeuw & Mair, 2009). A five-dimension solution
proved to be the most parsimonious solution providing good fit (scaling stress of 0.116 and r2 of
0.79; stress of 0.15 or less is generally considered acceptable, Dugard, Todman, & Staines, 2010;
We calculated correlations between targets‘ scores on each MDS dimension and the
explicit trait ratings of each target (Table 3). The first dimension correlated strongly with targets‘
subjective SES (r = 0.91),10 the second with both categorization as Asian (r = -0.81) and
categorization as Black (r = 0.79),11 the third with categorization as White (r = 0.78), the fourth
with categorization as Female (r = 0.81), and the fifth with targets‘ age (r = 0.91). These results
suggested that targets were spontaneously perceived as varying based on core demographic
Table 3
Target-level correlations between targets’ MDS-derived dimension scores and mean explicit
trait ratings. Correlations weaker than 0.2 are not displayed.
MDS Dimensions
1 2 3 4 5
Subjective SES 0.91
Occupational Prestige 0.89
Education 0.85
Income 0.81 0.22 -0.24 0.24
Attractiveness 0.8 -0.31
Competence 0.79
Disorganized/Careless -0.74 -0.33
Dominant 0.73 0.25
Dependable/Self disciplined 0.67 -0.21 0.21
Calm/Emotionally stable 0.61 -0.22
Submissive -0.6 -0.33 -0.23
Hard working 0.55 -0.26 -0.31 0.33
Extraverted/Enthusiastic 0.52 0.33 0.28 -0.3
Reserved/Quiet -0.51 -0.27 -0.34 0.24
Asiana 0.2 -0.81 -0.34 -0.3
Blacka 0.79 -0.47
Liberal 0.62 -0.27 -0.37
Conventional/Uncreative -0.33 -0.38 0.29
10
In the original MDS solution Dimension 1 correlated negatively with measures of social class.
We have reversed its scores throughout the manuscript for ease of interpretation. This has no
targets, and the other (Dimension 3) separating White from Asian and Black targets––is sensible
given that two linear dimensions are necessary to separate the three racial groups represented.
20
INTERSECTIONAL IMPLICIT BIAS
Figure 3. Study 2 targets arranged according to their scores on each of the 5 spontaneously
To identify the optimal scoring algorithm for Target D Scores, we compared different
algorithms with regard to both their internal reliability, as indexed by split-half reliability
estimates, and their convergent validity, as indexed by the strength of their relationships with
associated with implicit evaluations (see Supplementary Materials for more details). The scoring
21
INTERSECTIONAL IMPLICIT BIAS
algorithm for Target D Scores producing the optimal results12 involved (a) identifying all raw
response times toward a specific target in ST-IATs trials, including error trials, (b) eliminating
response times below 100 milliseconds and above 4000 milliseconds (12% and 0.02% of trials,
respectively), (c) penalizing error trials, in which the wrong computer key was pressed in
response to a target (6.5% of all trials) by replacing their latency with participants‘ individual
mean response latency in compatible/incompatible trials plus 600ms, (d) taking the natural log of
each of the remaining response times, (e) computing a difference score for each target
representing the mean logged response time in incompatible trials minus the mean logged
response time in compatible trials. To aid interpretability, these difference scores were then
divided by the overall standard deviation of all logged response times between 100 and 4000
To test the utility of modelling implicit evaluations at the target level, we calculated
Target D Scores for each of the 69 unique targets used in Study 1 (Study 1a split-half reliability
= 0.57, Study 1b split-half reliability = 0.66). Not only was there was a significant positive raw
correlation between target‘s mean income ratings and Target D Scores, r(67) = 0.35, p = .003,
controlling for targets‘ group membership, β = 0.91(SE = 0.36), t(58) = 2.58, p = 0.013, =
0.10.13 Thus, even within target groups, targets judged to have higher incomes produced higher
Target D Scores. This systematic variation had previously been obscured within Study 1‘s target
group-level analyses.14
12
This algorithm also produced the highest internal reliability, so would have been chosen if
both z-scored. Target group membership was entered into the model as a categorical predictor.
14
Target-level variation in implicit evaluations can also be studied via more complex models
predicting raw or logged response times (e.g., Thiem et al., 2019; Mattan et al., 2019). We
discuss Target D Scores‘ advantages over these methods in our general discussion.
22
INTERSECTIONAL IMPLICIT BIAS
To assess the relationship between each MDS dimension and implicit bias, we fit
multiple regression models predicting the Target D Scores (split-half reliability = 0.71) of each
of the 54 Study 2 targets from each of the multi-dimensional scaling dimensions. Results (Table
4) revealed significant associations between Target D Scores and Dimension 1 (Social class),
̂ ̂ = 0.06(0.02), t(48) = 4.07, p < .001, = 0.26, with bias favouring higher class over
0.04(0.02), t(48) = -2.71, p = .01, = 0.13, with bias favouring Asian and Black targets over
White targets, and Dimension 4 (Gender), ̂ ̂ = 0.06(0.02), t(48) = 3.89, p < .001, =
significantly improved model fit, F(9,39) = 3.43, p = 0.003. Main effects of Dimensions 1
(Social class), 3 (Race), and 4 (Gender) each remained significant (see Table 4), but the effects
t(39) = 4.29, p < .001, = 0.32, with the positive interaction slope suggesting a stronger effect
of the social class dimension among female targets (higher scores on Dimension 4 = female
targets). Including three-way interactions between dimensions did not improve model fit, F(7,32)
= 0.48, p = 0.84.
achieved 80% power to detect main effects of approximately = 0.10 and two-way interaction
Table 4
Study 2 results of multiple regressions predicting Target D Scores
Multi-Dimensional Scaling dimensions
Model 1 Model 2
̂ ̂ p r2 ̂ ̂ p r2
(Intercept) 0.019(0.015) 0.216 0.019(0.012) 0.139 NA
Dimension 1 (Social class a) 0.061(0.015) <.001 0.257 0.062(0.013) <.001 0.364
Dimension 2 (Race b ) 0.002(0.015) 0.871 0.001 -0.001(0.014) 0.929 0.005
Dimension 3 (Race c) -0.041(0.015) 0.009 0.132 -0.037(0.014) 0.009 0.163
Dimension 4 (Genderd ) 0.059(0.015) <.001 0.24 0.059(0.013) <.001 0.334
Dimension 5 (Age) -0.008(0.015) 0.602 0.006 -0.013(0.013) 0.342 0.005
Dimension 1 × Dimension 2 -0.023(0.017) 0.171 0.047
Dimension 1 × Dimension 3 0.01(0.015) 0.526 0.01
Dimension 1 × Dimension 4 0.063(0.015) <.001 0.321
Dimension 1 × Dimension 5 -0.024(0.018) 0.173 0.047
Dimension 2 × Dimension 4 -0.015(0.015) 0.335 0.024
Dimension 2 × Dimension 5 0.014(0.013) 0.279 0.03
Dimension 3 × Dimension 4 0.002(0.02) 0.928 <.001
Dimension 3 × Dimension 5 -0.025(0.013) 0.056 0.09
23
INTERSECTIONAL IMPLICIT BIAS
Next, we predicted Target D Scores directly from explicit ratings of target‘s social class
(the average of z-scored mean ratings of subjective SES, occupational prestige, education, and
income, Cohen‘s = 0.98), binary indicators of Asian race, White race, and female gender15 , and
z-scored mean ratings of targets‘ age. We observed significant effects of targets‘ perceived social
class, ̂ ̂ = 0.04(0.02), t(48) = 2.45, p = .02, = 0.11, with bias favouring higher class
over lower class targets, and targets‘ gender, ̂ ̂ = 0.15(0.03), t(48) = 4.96, p < .001, =
0.34, with bias favouring female over male targets. In contrast to MDS dimensions, there were
no significant effects of target race, suggesting that the previously observed effect of Dimension
3 may have occurred due to its overlap with (see Table 4).
Next, we included each two-way interaction between predictors (except between the two
race indicators), which again significantly improved model fit, F(9,39) = 2.46, p = 0.02. Target
gender remained a significant predictor, but was qualified by a significant two-way interaction
with target social class, ̂ ̂ = 0.10(0.03), t(39) = 3.37, p = .002, = 0.23. The pattern of
this interaction suggested a strong effect of social class with regard to female targets, with upper-
15
Targets were coded as Asian, Black, and White if they were categorized as such by raters >
90% of the time. Gender was manually coded by the lead author.
24
INTERSECTIONAL IMPLICIT BIAS
class female targets eliciting positive evaluations, but little effect of social class for male targets
Figure 4. The top panel displays targets ordered by their Target D Scores (the row above) and
arranged according to their exact Target D Scores (the row below). The bottom panel displays
the interaction between targets‘ gender and perceived social class (a z-scored composite of
targets‘ perceived income, subjective SES, occupational prestige, and education) in predicting
Target D Scores.
A single sample t-test on participants‘ D Scores suggested that the sample exhibited
significant anti-Black/pro-White bias from the traditional two-category Race IAT (M = 0.30, SD
Discussion
25
INTERSECTIONAL IMPLICIT BIAS
between targets‘ gender and social class, with upper-class female targets eliciting especially
from MDS dimension scores or from explicit ratings of targets. By contrast, target race yielded
more equivocal effects, with an apparent anti-White bias emerging from MDS Dimension scores,
but failing to emerge when Target D Scores were predicted from targets‘ explicit race
These results do not align neatly with theories of compounding bias or the category
dominance model. Theories of compounding bias are consistent with especially positive
evaluations of upper-class female targets, but offer little explanation as to why we observed little
evidence of anti-Black bias in our ST-IATs (if anything, we observed weak evidence of anti-
White bias). Meanwhile, the category dominance model can make sense of equivocal or absent
race and age effects, as well as the relatively large effect of target gender. However, it does not
provide an easy explanation of interaction effects, which require at least some participants to be
IATs, our sample displayed a robust pro-White/anti-Black bias on the traditional Race IAT. This
suggests that the ST-IAT results cannot be explained as being simply a function of sampling
bias.
Study 3
tighter experimental control over our target stimuli, swapping the same target faces onto multiple
target bodies, thus holding constant body shape and clothing across target race categories, and
holding constant facial features exactly constant across social class categories. Second, all racial
groups were presented together within ST-IAT tasks. In Studies 1 and 2, targets of different races
were presented within separate ST-IAT tasks, raising the possibility that participants may have
used recoding strategies that suppressed implicit racial biases (e.g., Meissner & Rothermund,
16
If each participant‘s responses were dominated by a single category, gender-biased
participants should produce a main effect of gender, and class-biased participants should produce
a main effect of class. Such participants could collectively display main effects of both class and
gender, but should not, in theory, display an interaction between the two categories
26
INTERSECTIONAL IMPLICIT BIAS
2013). Third, we investigated whether the use of full-body targets in Studies 1 and 2 had elevated
the influence of targets‘ bodies––a primary source of social class cues (e.g., Becker et al., 2017;
Gillath et al., 2012; Schmid-Mast & Hall, 2004)––relative to the influence of targets‘ faces––
likely the primary source of race cues––due to targets‘ bodies dominating the visual space of
stimuli. To probe this, in Study 3 we presented targets both as upper-body images from the waist
Stimuli Development
Faces
We selected 24 unique faces from the Chicago Face Database (CFD; Ma, Correll, &
Wittenbrink) varying in race (8 Asian, 8 Black, 8 White), gender (12 male, 12 female), and age
(12 old, 12 young), with two faces chosen to represent each race/age/gender subgroup. Based on
CFD norming data, there were no significant differences among the chosen faces in perceived
attractiveness or racial prototypicality between race, age, or gender groups (all F < 1.27 , all p >
0.27)., nor differences in female or male categorization between race or age groups (all F < 0.002
, all p > 0.98), nor significant differences in Asian, Black, or White categorization between
gender or age groups (all F < 0.02 , all p > 0.89), and no significant differences in perceived age
between race or gender groups (all F < 0.03 , all p > 0.97).
Bodies
The 24 bodies we selected varied in terms of gender (12 male, 12 female), age (12 old, 12
young), and perceived socioeconomic status (12 high-SES, 12 low-SES), with three bodies
chosen to represent each gender/age/SES subgroup. Based on explicit ratings17 in which each
body was rated by an average of 84.1 raters (SD = 111.0), there were no significant differences in
perceived attractiveness between race, age, or gender groups (all F < 2.80, all p > 0.10), no
significant differences in perceived age between gender or SES groups (all F < 2.14 , all p >
0.15), and no significant differences in perceived SES or income between gender or age groups
(all F < 0.64 , all p > 0.43). Unavoidably, due to the strong correlation between ratings of
attractiveness and subjective SES in the data (r = 0.53), there was a significant difference in
perceived attractiveness between SES groups, with the high-SES bodies (M = 53.9, SD = 10.4)
17
Ratings of each body were made with different faces attached to each body, rendering these
rated significantly more attractive than the low-SES bodies (M = 30.6, SD = 7.6), F(1,22) = 39.3,
p < 0.001.
We used Adobe Photoshop software to attach each of the 6 faces to each of the 6 bodies
within each age/gender subgroup. This resulted in 144 total stimuli, which were then assembled
into six target groups. Each group contained 8 Asian, 8 Black, and 8 White targets, 12 female
and 12 male targets, 12 young and 12 old targets, and 12 high-SES and 12 low-SES targets (see
Figure 5. The 24 faces and 24 bodies combined to create 144 unique targets arranged into six
groups in which each face and body appears once. Both upper-body presentation (Study 3a) and
Participants for Study 3a (N = 871, 591 women, 223 men, 11 non-binary, 46 missing
gender data, Mage = 23.0, SDage = 8.0, 411 Asian, 253 White, 77 Latino, 26 Black, 30 other race,
39 missing race data) and Study 3b (N = 656, 489 women, 149 men, 7 non-binary, 11 missing
28
INTERSECTIONAL IMPLICIT BIAS
gender data, Mage = 20.83, SDage = 2.8, 364 Asian, 145 White, 84 Latino, 10 Black, 36 Other
race, 17 missing race data) were undergraduate students who participated for course credit. We
excluded ST-IAT data from five participants in Study 3b who experienced technical issues
during the ST-IAT task resulting in mean response times that were unreasonably large (>
registered at https://aspredicted.org/qz5yu.pdf.19
After providing informed consent, participants were randomly assigned to one of the six
target groups, and completed two consecutive ST-IATs containing their target group as stimuli
following the procedures described above. 20 In Study 3a participants viewed targets in upper-
Difference Ratings
confirm that targets‘ race, gender, social class, and age would again emerge as the primary
spontaneous dimensions underlying such judgments. Following Study 3a‘s initial data collection
(see footnote 15), we considered this to be sufficiently established, and omitted the difference
ratings from the additional data collected for Study 3a and from Study 3b (see Supplementary
18
After the original planned sample size was reached in Study 3a (N = 379), the split-half
reliability of the Target D Scores was so low (0.37) that we decided to collect additional data,
some minor changes to the study design, omitting similarity/difference ratings of pairs of targets
and the Symbolic Racism Scale, and adding explicit ratings scales of targets‘ attractiveness,
competence, political orientation, and photo blurriness. These changes had minor effects on the
understanding of how best to model and present our results. See Supplementary Materials for
more details.
20
We included two ST-IATs because in Study 3 there were 24 targets per ST-IAT, compared
with 8 and 18 targets per ST-IAT in Studies 1 and 2. We therefore wanted to increase the number
Participants also rated their 24 targets via 0-100 sliders on perceived gender (ICCs =
0.89, 0.87 in Studies 3a and 3b, respectively), race (three separate sliders measuring perceptions
of targets as Asian, ICCs = 0.87, 0.86, Black, ICCs = 0.91, 0.89, and White, ICCs = 0.85, 0.84)
social class (ICCs = 0.55, 0.59), and age (ICCs = 0.61, 0.58). We also measured perceptions of
targets‘ warmth (ICCs = 0.22, 0.21), extroversion (ICCs = 0.11, 0.14), attractiveness (ICCs =
0.20, 0.22), competence (ICCs = 0.30, 0.31), political orientation (ICCs = 0.26, 0.27), and photo
evaluations.
Demographics
Results
Manipulation Checks
ratings of the targets and our a priori categorizations of targets as male, Asian, Black, White,
high-SES, and older/younger. Correlations indicated that each variable was manipulated as
intended (see bolded correlations in Table 5). There was also relatively little non-orthogonality
between these variables, with the exception of a correlation between SES and age ratings (Study
3a: r = 0.15, Study 3b: r = 0.12). To control for this non-orthogonality, we again used target-
level analyses modelling targets‘ social class and age as continuous variables.
Table 5
Correlations between our a priori categorizations and participants‘ subjective ratings of targets
Female ratings Asian ratings Black ratings White ratings SES ratings Age ratings
Study 3a
Asian ratings 0.01
Black ratings 0.004 -0.489
White ratings -0.017 -0.464 -0.545
SES ratings -0.028 0.074 -0.028 -0.034
Age ratings -0.035 0.078 0.012 -0.096 0.151
Female categorization 0.998 0.01 -0.004 -0.009 -0.025 -0.032
Asian categorization 0 0.998 -0.495 -0.456 0.071 0.073
Black categorization 0.007 -0.493 0.999 -0.54 -0.031 0.004
White categorization -0.008 -0.505 -0.504 0.996 -0.039 -0.077
SES categorization -0.003 0.001 0.002 0.005 0.911 0.039
Age categorization 0 0.004 0.005 -0.018 0.127 0.947
Study 3b
Asian ratings 0.021
Black ratings -0.002 -0.493
White ratings -0.028 -0.472 -0.533
30
INTERSECTIONAL IMPLICIT BIAS
Because the same faces and bodies were shared by multiple targets, we fitted cross-
classified hierarchical linear models (HLMs) predicting Target D Scores (Study 3a split-half
reliability = 0.54, Study 3b split-half reliability = 0.59), and included in each model random
intercepts for the 24 unique target faces and 24 unique target bodies (see Table 6). For all HLMs
we used the R packages lme4 (Bates, Maechler, Bolker, & Walker, 2015) and lmerTest
Study 3a. We first predicted Target D Scores from z-scored mean ratings of targets‘
subjective SES, z-scored mean ratings of targets‘ age, and dummy variables indicating Asian
race, White race, and female gender. We observed significant effects of target race, with both
Asian targets, ̂ ̂ = 0.10(0.02), t(18.85) = 4.30, p < .001, = 0.1321 , and White targets,
̂ ̂ = 0.09(0.02), t(18.69) = 4.07, p < .001, = 0.12, evaluated more positively than Black
targets (for the simultaneous addition of both race dummies r2 22 = 0.07). There was no
significant difference between evaluations of Asian and White targets, t(18.97) = -0.24, p =
0.81. Female targets were also evaluated more positively than male targets, ̂ ̂ = 0.20(0.02),
t(13.36) = 8.71, p < .001, = 0.49. Neither targets‘ social class nor age exhibited significant
21
refers to semi-partial r2 values (Edwards, Muller, Wolfinger, Qaqish, & Schabenberger,
2008) computed using the standardized generalized variance approach with the r2glmm R
generalized variance approach with the r2glmm R package (Jaeger, 2017) between full models
Doing so did not significantly improve model fit, (9) = 7.53, p = 0.58, so we relegate these
results to Supplementary Materials. Finally, in a third model, we tested if the effects observed in
our initial model were robust to controlling for targets‘ z-scored mean ratings on perceived
6.23, p < .001, = 0.31, but all other target level variables were non-significant (See Table 6).
Table 6
Results from hierarchical linear models in Study 3a and Study 3b
Study 3a (upper-body targets)
Model 1 Model 3
̂ ̂ p a
SD ̂ ̂ p a
SD
Fixed effects
(Intercept) -0.129(0.021) <.001 -0.107(0.032) 0.002
Social class 0.007(0.011) 0.569 0.004 -0.026(0.045) 0.563 0.003
Asian 0.096(0.022) <.001 0.127 0.075(0.038) 0.059 0.032
White 0.091(0.022) <.001 0.115 0.041(0.062) 0.514 0.004
Female 0.2(0.023) <.001 0.488 0.203(0.033) <.001 0.306
Age 0.006(0.011) 0.598 0.003 0(0.017) 0.995 <.001
Warmth -0.004(0.022) 0.851 <.001
Extroversion 0.003(0.018) 0.869 <.001
Attractiveness 0.023(0.024) 0.334 0.009
Competence 0.016(0.049) 0.739 <.001
Liberal -0.043(0.029) 0.143 0.019
Blurry 0.016(0.013) 0.249 0.019
0.534 0.536
Random effects
Face 0.007 0.015
Body 0.034 0.042
Residual 0.107 0.106
Study 3b (full-body targets)
Model 1 Model 3
̂ ̂ p a
SD ̂ ̂ p a
SD
Fixed effects
(Intercept) -0.152(0.028) <.001 -0.123(0.029) <.001
Social class 0.044(0.016) 0.01 0.12 0.022(0.042) 0.594 0.002
Asian 0.101(0.027) 0.001 0.108 0.06(0.038) 0.117 0.021
White 0.092(0.026) 0.003 0.091 0.037(0.061) 0.547 0.003
Female 0.232(0.033) <.001 0.491 0.237(0.033) <.001 0.349
Age -0.009(0.016) 0.585 0.005 -0.006(0.017) 0.74 0.001
Warmth -0.016(0.022) 0.487 0.004
Extroversion -0.016(0.015) 0.297 0.011
Attractiveness 0.035(0.026) 0.186 0.016
Competence -0.012(0.043) 0.772 <.001
Liberal -0.022(0.028) 0.425 0.006
Blurry -0.044(0.012) <.001 0.131
0.507 0.617
Random effects
Face 0.026 0.012
Body 0.061 0.031
Residual 0.113 0.117
Note: Black is the reference category for race contrasts
a
refers to semi-partial r2 statistics, except the bottom-most values, which indicates r2 for the full
model.
32
INTERSECTIONAL IMPLICIT BIAS
Study 3b. We fitted the same series of cross-classified HLMs predicting Target D Scores
for the Study 3b full-body targets. Again, we observed a significant effect of target race, with
both Asian targets, ̂ ̂ = 0.10(0.03), t(18.44) = 3.80, p = 0.001, = 0.05, and White
than Black targets (for the simultaneous addition of both race dummies r2 = 0.06), but no
significant differences between Asian and White targets, t(19.17) = -0.35, p = 0.73. We also
observed significant effects of target gender, with female targets evaluated more positively than
males, ̂ ̂ = 0.23(0.03), t(19.79) = 7.06, p < .001, = 0.38, and of target social class, with
t(21.59) = 2.83, p = .01, = 0.05. Targets‘ age did not significantly affect implicit evaluations
As in Study 3a, adding two-way interactions did not significantly improve model fit,
(9) = 11.99, p = 0.21 (see Supplementary Materials), and target gender was the only
manipulated factor that remained a significant predictor over and above the control variables,
̂ ̂ = 0.24(0.03), t(23.46) = 7.31, p < .001, = 0.34. In this model we also observed a
significant effect of photo blurriness, with more blurry photos eliciting more negative
Figure 6. The effects of target race and gender in Study 3a and 3b visualized by showing each
unique face and body arranged according to their mean Target D Scores (lower rows) and rank-
Simulation-based power sensitivity analyses suggested that due to the package lmerTest‘s
(Kuznetsova et al., 2017) use of the Satterthwaite degrees of freedom method, statistical power
varied between effects. Both Study 3a and 3b achieved 80% power to detect main effects
between approximately = 0.10 and = 0.15 and interaction effects between approximately
Discussion
social class, and age. Across both methods we observed a dominant effect of target gender,
which exerted effects many times larger than any other target-level factor. This result is most
34
INTERSECTIONAL IMPLICIT BIAS
consistent with the category dominance model, which posits that responses to multiply
categorizable targets will be driven by single dominant categories. This theory is agnostic to
which category will dominate when participants are not primed or manipulated in specific ways,
and our finding that gender emerged as the dominant category in the present context is notable.
However, this dominance of gender was not absolute. We also observed effects of
targets‘ race, with Asian and White targets evaluated more positively than Black targets in both
studies, and social class, with upper-class targets evaluated more positively than lower-class
targets in Study 3b. These results therefore also provide some level of support for the notion of
compounding bias, as they suggest that implicit biases do combine additively, at least to some
Three further results of Study 3 were noteworthy. First, the presence of anti-Black bias in
both studies was consistent with the idea that such biases may have been suppressed in Studies 1
and 2, perhaps as a result of recoding strategies (Meissner & Rothermund, 2013). Second, the
observation of a significant effect of social class only for the full-body targets in Study 3b
aligned with the idea that full-body target images may increase the relative salience of social
class. Third, Study 3, with its more tightly controlled design, did not replicate the interaction
between target gender and social class observed in Study 2, calling into question the
Study 4
multiply categorizable social targets varying in race, gender, social class, and age. Two issues,
though, animated our last studies. First, Study 3 used non-representative samples of university
students (71% and 75% female and 49% and 55% Asian in Studies 3a and 3b, respectively).
Second, Study 3 relied solely on ST-IATs to measure implicit evaluations. As discussed above,
Gawronski and colleagues (2010) have argued that different measurement procedures might
produce different patterns of implicit biases toward multiply categorizable targets. In Study 4 we
sought to address these issues by (a) recruiting a nationally representative sample of U.S. adults,
and (b) measuring implicit evaluations via three different methods: ST-IATs, EPT (Fazio et al.,
age, and race via Prolific (Study 4a: N = 1620, 803 women, 790 men, 20 non-binary, 7 missing
gender data, Mage = 38.6, SDage = 14.2, 1167 White, 155 Black, 140 Asian, 103 Latino, 38 other
race, 17 missing race data; Study 4b: N = 846, 423 women, 415 men, 4 non-binary, 4 missing
gender data, Mage = 44.5, SDage = 20.8, 620 White, 117 Black, 58 Asian, 34 Latino, 8 other race,
All participants were randomly assigned to evaluate one of the six target groups used in
Study 3, which they viewed in in either full-body or upper-body presentation (Study 4a) or
upper-body presentation only (Study 4b). In Study 4a participants completed two ST-IATs and
one EPT, with the tasks randomly ordered. In Study 4b participants completed one AMP.
9 participants and EPT data from 6 participants for having mean response times greater than
3000ms.23 In Study 4b we excluded 38 participants for uniform responses on the AMP, and 9
ST-IATs
EPTs in Study 4a began with 10 practice trials in which the symbols ―***‖ were
presented in the center of participants‘ screens for 200ms, followed by an interstimulus gap of
100ms, and then one of 24 positive words or 24 negative target words (e.g., ―honor‖, ―lucky‖,
―evil‖, ―cancer‖, Draine & Greenwald, 1998). Participants were tasked with categorizing the
target words as either ―Good‖ or ―Bad‖ as quickly as possible via E or I computer key presses,
with the assignment of valences to keys randomised between participants. Following this,
participants performed 96 test trials (4 per target) in which the multiply categorizable target
images were presented as primes in place of the ―***‖ symbols. Each multiply categorizable
target image was presented prior to two positive and two negative target words, and there was a
23
We deviated slightly from our pre-registration due to our evolving understanding of the
optimal algorithm for computing ST-IAT Target D Scores by using response time cut-offs of
100ms and 4000ms instead of 100ms and 6000ms, and by penalizing error trials. As reported in
2500ms gap between the presentation of each prime/target pairing. Participants took breaks after
In the AMP in Study 4b, the words ‗Unpleasant‘/‗Pleasant‘ appeared at the top left/right
of participants‘ screens. In each trial a multiply categorizable target was displayed as a prime for
75ms, followed by an inter-stimulus gap of 125ms, followed by one of 200 Chinese characters
displayed for 100ms, followed by a pattern mask. Participants were tasked with categorizing the
Chinese characters as either less pleasant than average or more pleasant than average via their E
and I keys, respectively. Participants completed 10 practice trials, followed by 150 test trials,
Participants in Study 4a rated each of the 24 targets in their assigned target group via 0-
100 sliders on perceived gender (ICC = 0.91), race (three separate sliders measuring perceptions
of targets as Asian, ICC = 0.88, Black, ICC = 0.92, and White, ICC = 0.84) social class (ICC =
0.53), age (ICC = 0.59), attractiveness (ICC = 0.18), and photo blurriness (ICC = 0.48).
Demographics
Results
For the ST-IAT data, we calculated Target D Scores for each of the 288 unique target
images (144 targets presented in both full- and upper-body formats) according to the algorithm
described above (split-half reliability = 0.40). For the EPT and AMP data, we again undertook a
data-driven process to determine which scoring algorithm would produce the highest combined
internal reliability and convergent validity. This process suggested that both EPT and AMP data
require different scoring algorithms to optimize measurement. This was especially the case for
the EPT: applying the ST-IAT algorithm to the EPT data resulted in virtually zero internal
24
. We chose 96 trials to obtain a roughly equivalent amounts of potentially useable trials per
participant for the ST-IAT and EPT measures (in total, two ST-IATs provide approximately 80
For the EPT, the method providing the best measurement involved (a) identifying all raw
response times toward a specific target in EPT trials, (b) eliminating response times below 175
milliseconds and above 1000 milliseconds (0.006% and 0.097% of trials, respectively), (c) taking
the natural log of the remaining response times, (e) computing a difference score for each target
representing the mean logged response time to the target in incompatible trials minus the mean
logged response time to the target in compatible trials. For interpretability, we again divided
these differences by the overall standard deviation of all logged EPT response times between 175
and 1000 milliseconds. This procedure yielded an estimated split-half reliability for the EPT
For the AMP, the method providing the best measurement involved (a) identifying all
responses following each specific target prime, (b) eliminating responses faster than 75
milliseconds or slower than 4500 milliseconds (0.006% and 0.013% of trials, respectively), (c)
computing the proportion of the Chinese characters judged more positive than average following
each target prime (M = 0.62, SD = 0.03, range = 0.53-0.70). This procedure yielded an estimated
For each Target D Score (ST-IAT, EPT, and AMP), we fitted a separate series of cross-
classified HLMs. To test for differences between full-body and upper-body presentation in Study
4a, separate full-body and upper-body Target D Scores were computed for each target, and both
were included in each model. For Study 4b, a single Target D Score was computed for each
target.
An initial model predicted Target D Scores from fixed effects of z-scored mean ratings of
targets‘ subjective SES, dummy variables indicating Asian race, White race, and female gender,
and z-scored mean ratings of targets‘ age. As in Study 3, we included in each model random
intercepts for targets‘ faces and bodies. A second model added a dummy variable indicating
body), and a third model added two-way interactions between each target-level factor and the
full-body indicator to test whether the effect of targets‘ social class, race, gender and age were
25
This result is similar to that of Gawronski and colleagues (2010), who also found the AMP to
provide a much more reliable measurement tool for measuring evaluations of multiply
moderated by presentation format. If these interaction terms failed to significantly improve fit
compared to the second model, they were removed. A fourth model added two-way interactions
between each target-level factor. Again, if these interaction terms failed to significantly improve
fit compared to the previous model, they were removed. A fifth and final model added z-scored
ST-IAT Target D Scores. For ST-IAT Target D Scores, in the initial model we observed
significant effects of target social class, with higher-class targets evaluated more positively than
lower-class targets, ̂ ̂ = 0.03(0.01), t(23.12) = 5.1, p < .001, = 0.13. We also observed
significant effects of target gender, with female targets evaluated more positively than male
targets, ̂ ̂ = 0.14(0.01), t(20.15) = 11.49, p < .001, = 0.43, and target race, with both
Asian targets, ̂ ̂ = 0.06(0.01), t(266.43) = 3.91, p < .001, = 0.07, and White targets,
̂ ̂ = 0.05(0.01), t(263.94) = 3.78, p < .001, = 0.07, evaluated more positively than
Black targets (for the simultaneous addition of both race dummies r2 = 0.04). There was no
significant difference between evaluations of Asian and White targets, t(273.32) = -0.13, p =
0.89. Targets‘ age had no significant effect on implicit evaluations (see Table 7). In the second
model, we observed a significant effect of the full-body target indicator, with full-body targets
evaluated more negatively than upper-body targets, ̂ ̂ = -0.05(0.01), t(261.03) = -4.52, p <
.001, = 0.09. Model fit was not significantly improved by adding two-way interactions
between the full-body target indicator and each of the target-level factors, (5) = 4.25, p = 0.51,
or by adding two-way interactions between each of the target-level factors, (9) = 4.98, p =
0.84. Fixed effects estimates remained virtually unchanged after controlling for attractiveness
and photo blurriness (results of Models 1 and 5 are reported in Table 7; for full results of all
EPT Target D Scores. For EPT Target D Scores in Study 4a, in the initial model we
observed significant effects of target social class, with higher-class targets evaluated more
also observed significant effects of target gender, with female targets evaluated more positively
than male targets, ̂ ̂ = 0.05(0.01), t(20.01) = 4.05, p < .001, = 0.09, and target race,
with Asian targets evaluated more positively than both Black targets, ̂ ̂ = 0.03(0.02),
39
INTERSECTIONAL IMPLICIT BIAS
2.34, p = .02, = 0.03 (for the simultaneous addition of both race dummies r2 = 0.03). There
was no significant difference between evaluations of White and Black targets, t(263.87) = -0.24,
p = 0.81. Targets‘ age also had no significant effect on implicit evaluations (see Table 7). In the
second model, there was no significant effect of the full-body target indicator, t(260.88) =- 0.19,
p = 0.85. Model fit was not significantly improved by adding two-way interactions between the
full-body target indicator and each of the target-level factors, (5) = 5.52, p = 0.36, or by
adding two-way interactions between each of the target-level factors, (9) = 5.31, p = 0.81.
After controlling for attractiveness and photo blurriness, the gender and pro-Asian/anti-Black
biases remained significant, but the effect of social class and the difference between Asian and
AMP Target D Scores. For the AMP Target D Scores in Study 4b, in the initial model
we observed significant effects of target social class, with higher-class targets evaluated more
positively than lower-class targets, ̂ ̂ = 0.01(0.002), t(138) = 3.99, p < .001, = 0.12. We
also observed significant effects of target gender, with female targets evaluated more positively
than male targets, ̂ ̂ = 0.02(0.005), t(138) = 5.01, p < .001, = 0.18, and target race, with
both Asian targets, ̂ ̂ = 0.02(0.01), t(138) = 3.68, p < .001, = 0.10, and Black targets,
̂ ̂ = 0.02(0.01), t(138) = 2.98, p = 0.003, = 0.07, evaluated more positively than White
targets (for the simultaneous addition of both race dummies r2 = 0.08). There was no significant
difference between evaluations of Asian and Black targets, t(138) = 0.74, p = 0.46. Targets‘ age
also had no significant effect on implicit evaluations (see Table 7). Model fit was not
significantly improved by adding two-way interactions between each of the target-level factors,
(9) = 5.75, p = 0.76. Only target gender remained significant after controlling for
power to detect main effects of between approximately = 0.05 and = 0.075, and
interaction effects between approximately = 0.025 and = 0.075, while Study 4b achieved
26
The Asian-White result refers to a model fit with Asian set as the reference level for the race
variable.
40
INTERSECTIONAL IMPLICIT BIAS
80% power to detect main effects between approximately = 0.075 and = 0.10, and
interaction effects between = 0.05 and = 0.10. (see Supplementary Materials for details).
Table 7
Results from hierarchical linear models in Study 4
ST-IAT Target D Scores (Study 4a)
Model 1 Model 5
a a
̂ ̂ p SD ̂ ̂ p SD
Fixed effects
(Intercept) -0.119(0.012) <.001 -0.09(0.013) <.001
Social class 0.032(0.006) <.001 0.128 0.032(0.011) 0.006 0.04
Asian 0.056(0.014) <.001 0.07 0.056(0.015) <.001 0.068
White 0.054(0.014) <.001 0.066 0.054(0.016) 0.001 0.052
Female 0.144(0.013) <.001 0.428 0.144(0.015) <.001 0.328
Age -0.01(0.006) 0.137 0.013 -0.009(0.007) 0.197 0.009
Full-body target -0.057(0.013) <.001 0.088
Attractiveness -0.002(0.013) 0.887 <.001
Blurry -0.008(0.007) 0.254 0.007
0.493 0.534
Random effects
Face <.001 <.001
Body 0.011 0.01
Residual 0.1 0.096
EPT Target D Scores (Study 4a)
Model 1 Model 5
a a
̂ ̂ p SD ̂ ̂ p SD
Fixed effects
(Intercept) 0.108(0.013) <.001 0.109(0.014) <.001
Social class 0.024(0.007) 0.002 0.067 0.004(0.012) 0.756 <.001
Asian 0.032(0.015) 0.037 0.021 0.041(0.016) 0.011 0.032
White -0.004(0.015) 0.813 <.001 0.014(0.018) 0.441 0.003
Female 0.054(0.013) <.001 0.088 0.035(0.016) 0.039 0.024
Age -0.003(0.007) 0.639 0.001 0.004(0.008) 0.566 0.002
Full-body target 0.0002(0.014) 0.989 <.001
Attractiveness 0.025(0.014) 0.07 0.017
Blurry -0.006(0.007) 0.375 0.004
0.168 0.186
Random effects
Face <.001 <.001
Body 0.013 0.009
Residual 0.105 0.104
AMP Target D Scores (Study 4b)
Model 1 Model 2
a a
̂ ̂ p SD ̂ ̂ p SD
Fixed effects
(Intercept) 0.616(0.005) <.001 0.616(0.005) <.001
Social class 0.01(0.002) <.001 0.119 0.005(0.004) 0.173 0.016
Asian 0.004(0.006) 0.462 0.005 0.007(0.006) 0.256 0.011
White -0.018(0.006) 0.003 0.07 -0.012(0.007) 0.069 0.028
Female 0.024(0.005) <.001 0.175 0.019(0.006) 0.002 0.075
Age -0.002(0.003) 0.361 0.007 0(0.003) 0.983 <.001
Attractiveness 0.008(0.005) 0.075 0.027
Blurry 0.003(0.003) 0.202 0.014
0.328 0.346
Random effects
Face <.001 <.001
Body <.001 <.001
Residual 0.029 0.029
Note: Black is the reference category for race contrasts
a
refers to semi-partial r2 statistics, except the bottom-most values, which indicates r2 for the full
model.
Discussion
class, and age using ST-IATs, EPTs, and AMPs. Target gender again emerged as the most
important predictor of implicit evaluations, with female targets evaluated more positively than
41
INTERSECTIONAL IMPLICIT BIAS
males, and target gender explaining more variation in ST-IAT, EPT, and AMP Target D Scores
than any other factor. We also observed smaller but consistent effects of target social class, with
upper-class targets evaluated more positively than lower-class targets via all three methods. By
contrast, the effects of race were inconsistent, with participants favoring White and Asian over
Black targets in ST-IATs, Asian over White and Black targets in EPTs, and Asian and Black
over White targets in AMPs. We observed no significant effects of target age, no significant
These results suggest that the dominant effect of gender in Study 3 was not due to non-
representative sampling. In a sample of U.S. adults nationally representative with regard to race,
gender, and age, target gender exhibited a similar-sized effect on ST-IAT Target D Scores ( =
0.43) as it had in Study 3 ( = 0.49). However, these results also suggest that the dominance of
target gender in Study 3 may have been amplified by its exclusive reliance on ST-IATs.
Although target gender was the largest effect across all three methods used, its relative
Study 5
We conducted Study 5 to address two final questions. First, we were curious how patterns
of responses to multiply categorizable targets varied for different sub-groups of our respondents.
For example, although we found pro-female/anti- male evaluative biases to be the most important
driver of our results, past work has found that such biases tend to larger in women than in men
(Richeson & Ambady, 2001). In Study 5a we conducted an integrative data analysis (Curran &
Second, observed effects of gender, social class, and race are compatible with multiple
explanations. It was possible that participants had simultaneously attended to and displayed bias
with respect to all three categories: gender, class, and race, but it was also possible that specific
groups of participants had attended and shown bias with respect solely to target gender, social
In Study 5a we tested the extent to which measurement tasks (ST-IAT, EPT, AMP),
sample sources (students, Prolific), participants‘ gender, participants‘ race, participants‘ age,
participants‘ SES, and participants‘ political affiliations moderated the effects of target gender,
42
INTERSECTIONAL IMPLICIT BIAS
race, and social class. To do so, we combined all the raw implicit evaluation data from Studies 2,
3, and 4.27 We then computed implicit evaluation scores for each unique participant/target/task
combination in the data. For example, if participant X was exposed to target Y in a ST-IAT, the
participant X‘s responses to target Y within ST-IATs, and then applying the ST-IAT Target D
Score algorithm to this data to compute an evaluation score specific to the participant/target
pairing. Because ST-IAT and EPT Target D Scores require the calculation of difference scores,
evaluation scores for these tasks were calculated only for participant/target pairs with at least one
To allow comparability across tasks, we z-scored the resulting evaluation scores within
tasks (ST-IAT, EPT, AMP). We also converted targets‘ perceived social class into a binary
predictor via median split, and converted participants‘ age, subjective SES, and political
moderators, as well as insufficiently represented racial or gender categories (see Table 8). This
left a final sample of 103,715 unique participant/target/task evaluation scores, representing 3,659
between full- and upper-body presentations in Study 4, we treated responses to targets across
Table 8
Descriptive statistics of moderators included in integrative data analysis
Moderator Categories included Categories excluded
Task ST-IAT: 2,221
EPT: 679
AMP: 759
Sample Students: 1,418
Prolific: 2,241
Gender Women = 2,186 Non-binary = 43
Men = 1,473
Race Asian = 1,003 Other race = 131
Black = 295
Latino = 283
White = 2,078
Age >50 years = 641
31-50 years = 1,035
18-30 years = 1,983
a
Subjective SES High (8-10) = 522
Medium (5-7) = 2,220
Low (1-4) = 917
Political affiliation Liberal (8-10) = 1,918
27
Study 1 data was not included in the integrative data analysis because Study 1 participants
Results
We first measured the main effects of target gender, target race, and target social class by
fitting a cross-classified hierarchical linear model predicting evaluation scores from fixed effects
of each target-level factor, plus random intercepts for participants and targets. These results are
denoted in Figure 7 as ‗Overall effects,‘ with associated and r2 values representing the model
fit improvement from each factor being added to this initial model.
Following this, we tested how the effects of target gender, race, and social class differed
depending on each moderator. To do so, we fitted a full model including fixed effects of each
moderator and each possible two-way interaction between moderators and target-level factors.
This meant that each interaction was tested while controlling for all other interactions. This was
desirable given high levels of covariation among the moderators (e.g., student samples were
largely Asian and largely women, EPT and AMP samples were on average older than ST-IAT
samples). Each two-way interaction is visualized in Figure 7, with the associated and r2
values representing model comparisons between this full model and models with all interactions
except the focal interaction. Given the number of tests run, we adjusted p values using the
Benjamini-Hochberg procedure (Benjamini & Hochberg, 1995; for full model results see
Supplementary Materials).
44
INTERSECTIONAL IMPLICIT BIAS
Figure 7. Effects of target gender, race, and social class, and how these were moderated by task,
sample, and participant gender, race, age, SES, and political affiliation. ‗Overall effect‘ and
r2 values represent model fit improvements from adding each target-level predictor to a model
containing both other target-level predictors. All other and r2 values represent model fit
improvements from adding each two-way interaction to a model containing all other two-way
interactions between moderators and target-level predictors; *** = p < .001, ** = p < .01, * = p <
.05.
Results showed that the effect of target gender was significantly moderated by task, (2)
= 20.8, p < .001, r2 = 0.0002, and participant gender, (1) = 95.5, p < .001, r2 = 0.0008. Bias
favoring female targets was displayed in all tasks and among all sub-groups of participants, but
the effect was stronger in ST-IATs (gender difference = 0.16, SE = 0.01) than in EPTs (gender
difference = 0.05, SE = 0.01) and in AMPs (gender difference = 0.08, SE = 0.01), and was
stronger among women (gender difference = 0.18, SE = 0.01) than men (gender difference =
0.03, SE = 0.01).
The effect of target race was moderated by task, (4) = 26.6, p < .001, r2 = 0.0002, and
bias in ST-IATs (Asian – Black difference = 0.05, SE = 0.02; all other differences NS), little
45
INTERSECTIONAL IMPLICIT BIAS
racial bias in EPTs (all differences NS), and anti-White bias in the AMP (Asian – White
difference = -0.07, SE = 0.02; Black – White difference = 0.06, SE = 0.02, Asian – Black
difference NS). Asian participants displayed a pro-Asian bias (Asian – Black difference = 0.07,
SE = 0.02; Asian – White difference = 0.05, SE = 0.02, White – Black difference NS), Black
participants displayed an anti-White bias (Black – White difference = 0.08, SE = 0.03; Asian –
White difference = 0.07, SE = 0.03, Black – Asian difference NS), Latino participants displayed
a pro-Asian/anti-Black bias (Asian – Black difference = 0.07, SE = 0.03, all other differences
NS), and White participants displayed little racial bias (all differences NS).
The effect of target social class was significantly moderated by participant gender, (1)
= 7.9, p = .005, r2 = 0.00007, with women showing a greater bias (upper – lower difference =
0.07, SE = 0.02) than men (upper – lower difference = 0.04, SE = 0.01). No other interactions
reached significance.
Discussion
In Study 5a we explored how task type, sample source, and participants‘ gender, race,
age, social class, and political affiliation moderated the effects of targets‘ gender, race, and social
class. Although some notable interactions emerged, there was striking consistency across results.
For example, implicit gender bias was greater among women than men, and greater in ST-IAT
tasks than EPTs and AMPs. However, every sub-group of respondents displayed a pro-
female/anti- male bias. Similarly, implicit social class bias was stronger among women than men,
together, these results suggest that while the relative magnitude of implicit gender and social
class biases may vary across demographic groups, the fundamental directions of these biases are
relatively stable.
By contrast, the effect of race was less consistent, with participants displaying pro-
Asian/anti-Black bias in the ST-IAT, little detectable racial bias in the EPT,28 and anti-White
bias in the AMP. Additionally, Asian participants displayed a clear ingroup bias favoring Asian
over Black and White targets, Black participants favored Asian and Black targets over Whites,
28
Via the Target D Score analysis in Study 4a, a pro-Asian/anti-Black/anti-White bias was
detected via the EPT data. This difference likely reflects the data exclusions and different scoring
Latino participants favored Asian over Black targets, and White participants displayed no
With the exception of the differences in implicit gender bias between women and men
(Richeson & Ambady, 2001), the majority of these interactions involve novel observations. We
are not aware of any prior work that would have predicted the effect of target gender to be
strongest in ST-IATs, the effect of target social class to be stronger among women than men,
White participants to show the least racial bias of any racial group, or a robust anti-White bias to
emerge on AMPs. Each of these findings may warrant further attention and research, yet given
their exploratory and unanticipated nature, each should also be regarded as preliminary and
suggestive only.
categories of target gender, race, and social class, or alternatively, separate sub-groups of
participants may have attended to each category. The noisiness of implicit bias data makes it
difficult to tease these alternate explanations apart, but one way to do so is to assess the
relationship between separate biases at the level of individual participants. In our case, we
focused on the relationship between participants‘ implicit gender bias and participants‘ implicit
social class bias, as these were the two most consistent biases displayed in our data, and could
29
both be easily quantified.
Here, the reasoning is that if our observation of both gender and social class biases came
about via distinct groups of participants attending to either targets‘ gender or to targets‘ social
class, this would be expected to produce a negative correlation between the two biases among
participants. This is due to the expected distributions of each kind of bias among each sub-group
of participants. The gender-focused group would produce a distribution of gender bias scores
centered above zero, and a distribution of social class bias scores centered near zero.30
Meanwhile, the class-focused group would produce a distribution of social class bias scores
29
For this analysis we ignored racial bias due to the inconsistency of racial biases in our data,
and the complexity involved in creating racial bias scores from evaluations of three categories of
targets.
30
We say centered near zero here because the zero point (equivalent response times in
compatible vs. incompatible trials) does not necessarily indicate a lack of bias.
47
INTERSECTIONAL IMPLICIT BIAS
centered above zero, and a distribution of gender bias scores centered near zero. This would
mean that even with substantial amounts of added measurement error, individuals exhibiting
relatively higher gender bias scores would be more likely to belong to the gender-focused group,
and so would be more likely to exhibit relatively lower social class bias scores. Conversely,
individuals exhibiting relatively higher social class bias scores would be more likely to belong to
the class-focused group, and so would be more likely to exhibit relatively lower gender bias
scores. This should produce a negative correlation between the two kinds of bias, which we
Results
To quantify participants‘ gender and social class bias, we used the evaluation scores from
in Study 5a, and for each of 3,657 participants31 computed gender bias scores (participants‘ mean
evaluation scores for female targets minus their mean evaluation scores for male targets; M =
0.13, SD = 0.46), and social class bias scores (participants‘ mean evaluation scores for upper
class targets minus their mean evaluation scores for lower class targets; M = 0.06, SD = 0.46).
These scores displayed a significant positive correlation, r = 0.07, t(3655) = 3.99, p < .001 (see
Figure 8). To assess how unlikely this correlation would be if the data were produced by distinct
groups focused on separate categories, we simulated samples of 3,656 gender and social class
bias scores with means and standard deviations matching our observed data, but manipulated the
data such that half the sample was ‗gender-focused‘ (producing a distribution of gender bias
centered above zero and a distribution of social class bias centered at zero), while the other half
of the sample was ‗class-focused‘ (producing in a distribution of social class bias centered above
zero and a distribution of gender bias centered at zero). For each simulated distribution, we
computed the correlation between the two biases. From 10,000 simulated datasets, 98% of these
correlations fell below zero, and no correlations were higher than r = 0.03 (see the right panel of
31
Two participants were missing data on responses to males or females.
48
INTERSECTIONAL IMPLICIT BIAS
Figure 8).
Figure 8. The observed correlation between participants‘ implicit gender bias and implicit social
class bias scores (left), and the distribution of correlations between simulated bias scores
matching our data in N, means, and standard deviations, but derived from distinct gender-focused
Discussion
gender and social class bias scores. Via simulation, we then demonstrated that such a correlation
would be extremely unlikely if observed gender and social class biases result from distinct
groups of participants attending solely to gender or to social class. This suggests it is highly
unlikely our results were driven by distinct single-category- focused groups of participants.
General Discussion
Implicit bias is central to the study of social cognition. Given that people are multiply
categorizable, understanding the influences of such intersectionality upon implicit bias is likely
to be vital for understanding its effects in everyday social contexts. In the present research, we
examined implicit evaluations of multiply categorizable social targets, testing two competing
theories about intersectional intergroup bias. We also developed and tested the reliability of a
novel method of measuring and modelling implicit bias at the level of individual targets.
In Study 1 we observed implicit evaluations of Black and White males to be driven solely
by targets‘ social class, with bias favoring upper-class over lower-class targets. In Study 2, we
measured implicit evaluations of targets varying in race, gender, social class, and age, and found
49
INTERSECTIONAL IMPLICIT BIAS
results to be primarily driven by a specific positive bias favoring upper-class female targets. In
Study 3, we used similarly intersectional targets, and explored the impact of portraying targets in
full-body versus upper body photographs on implicit evaluations. Here, we observed effects of
targets‘ race, with Asian and White targets evaluated more positively than Black targets, and of
targets‘ social class, with upper-class targets evaluated more positively than lower-class targets
(though only when targets were displayed in full-body presentation). Most striking, however,
was the dominant effect of target gender, with positive/negative evaluations of female/male
samples of US adults, and measuring implicit evaluations not just via ST-IATs, but also via
EPTs and AMPs. Across all measures, we observed target gender to be the largest driver of
implicit evaluations, though its dominance was less pronounced in EPTs and AMPs than in ST-
IATs. We also again observed effects of targets‘ social class and race, though the effect of race
was inconsistent across tasks, with participants displaying anti-Black bias in the ST-IAT, pro-
Asian bias in the EPT, and anti-White bias in the AMP. Finally, in Study 5 we conducted an
integrative data analysis to test a number of potential moderating factors. Results showed that
while all groups of participants displayed pro-female implicit gender bias and pro-upper-class
implicit social class bias, both biases were stronger among women than men. Results also
showed the effect of race varied across racial groups, with Asians displaying a preference for
Asian over White and Black targets, Black participants displaying a preference for Asian and
Black targets over White targets, Latinos displaying a preference for Asian over Black targets,
The present work makes theoretical, empirical, and methodological contributions to the
study of implicit evaluative bias toward multiply categorizable targets. On a theoretical level, we
believe our results are best accounted for by a synthesis of compounding bias and category
dominance approaches. Consistent with the category dominance model (Macrae et al., 1995), we
intersectional targets in each of our studies. In Study 1, social class was dominant. In Studies 3
and 4, target gender was dominant. And even in Study 2, despite its more complex results, target
gender still uniquely accounted for substantially more variation in Target D Scores than any
other target-level predictor. These results are consistent with the notion that when faced with
50
INTERSECTIONAL IMPLICIT BIAS
complex social stimuli, social perceivers act as ‗cognitive misers‘, and make implicit evaluations
that are strongly influenced by specific social categories, and are relatively unaffected by others.
However, our results are also consistent with the notion that implicit biases compound––
at least to some extent––across multiple categories at once. In Studies 3 and 4, which used the
most tightly controlled set of targets, we observed simultaneous effects of targets‘ gender, race,
and social class. And in Study 5b, we found little evidence that these results represented separate
groups of participants attending solely to each factor. So, although we found little consistent
evidence for the kind of multiplicative interaction effects suggested by the multiple
negative implicit evaluations consistently being made of targets displaying multiple intersecting
stigmatized social identities (in this case, lower SES males), and the most positive implicit
evaluations being made of individuals displaying multiple intersecting positively- valued social
The overall picture emerging from the present work is therefore one of theoretical
compromise. Implicit evaluative biases toward complex multiply categorizable targets do appear
to compound across categories, but do so asymmetrically, with a dominant category (here, target
gender) playing a leading role, less dominant categories (here, target race and social class)
exerting relatively small additional effects, and peripheral categories (here, target age) having
This compromise position offers a novel rationale for grappling with intersectionality in
identities, or the idea that specific category intersections give rise to emergent phenomena that
cannot be understood by studying each category in isolation (e.g., Cole, 2009; Ghavami &
Peplau, 2012; Goff, & Kahn, 2013; Kang & Bodenhausen, 2015). The present work does not
invalidate these perspectives, but complements them, by suggesting that there may also be
Importantly, just like emergent intersectional effects, these asymmetries may also only be
discoverable via intersectional research programs. For example, in past research on implicit
51
INTERSECTIONAL IMPLICIT BIAS
evaluative bias, targets‘ race, gender, social class, and age have tended to produce biases of
roughly comparable size (Nosek, 2005). However, this work has rarely used an intersectional
lens, and the present results suggest that such methods may provide little guidance regarding the
relative influence of each category when participants respond to complex, multiply categorizable
targets. Indeed, even our traditional two-category Race IAT used in Study 2 provided a poor
guide to responses to more complex targets, with participants displaying robust anti-Black bias
on the two-category Race IAT, but no evidence of anti-Black bias when responding to multiply
encounters, this suggests that advancing understanding of how implicit bias operates in real-
implicit evaluations of multiply categorizable targets. This is consistent with one previous study,
in which target gender was the sole significant predictor of categorization errors in a weapon
identification task (Jones & Fazio, 2010). However, this prior work involved both a relatively
small and non-representative sample (79 college students), and as a relatively small and
idiosyncratic set of stimuli (8 total stimuli varying in race, gender, and occupation, with
occupations not matched across races or genders, and no reported pre-testing of stimuli). The
present results therefore provide a substantially more robust demonstration of this phenomenon.
It has long been established that individuals display pro-female evaluative biases via
binary implicit measures (Nosek, 2005). However, compared with evaluative biases regarding
race, or implicit associations between genders and specific social roles or abilities (e.g., Carlana,
2019; Levinson & Young, 2010), this phenomenon has attracted relatively little attention.
However, its dominance in the present results suggests the greater attention to gender-based
biases might have an important role to play in building our understanding of the causes and
One possible explanation for this result is that the dominance of gender was mediated by
its overall visual salience. While race was conveyed within our stimuli by targets‘ faces and
exposed skin, and social class was conveyed by targets‘ clothing, gender was conveyed by both
faces and clothing. This may have made gender the most visually salient category, producing its
dominant effect. Notably, however, even if this were the underlying mechanism, this would not
52
INTERSECTIONAL IMPLICIT BIAS
preclude our results from generalizing to real-world interactions, as in most everyday contexts
individuals‘ faces and bodies/clothing both tend to be visible, and to communicate gender.
promising path forward for studying intersectional implicit biases. Previously, researchers in this
area have used one of two approaches. One approach has been to measure and model implicit
attitudes at the level of target groups, either by calculating stand-alone measures of evaluations
of target groups representing intersectional category combinations (e.g., Jones & Fazio, 2010;
Mitchell et al., 2003, Studies 4 & 5; Moore-Berg et al., 2017; Perszyk et al., 2019), or by
calculating multiple binary preferences from responses to targets varying on multiple categories
(e.g., Gawronski et al., 2010; Mitchell et al., 2003, Studies 1-3; Yamaguchi & Beattie, 2019).
However, this approach obscures systematic variation in implicit evaluations within target
groups. By allowing investigators access to such within-target- group variation, Target D Scores
allows for the investigation of the simultaneous influence of a greater number of target-level
approaches such as these require target groups to be orthogonal with respect to both manipulated
variables and potential confounds, which is often not possible. As discussed above, Target D
Scores allow for greater control over non-orthogonalities and confounds by allowing researchers
to estimate effects of target-level predictors while controlling for targets‘ precise levels of other
A second prior approach has been to measure and model responses to multiply
categorizable targets at the level of individual (usually logged) response times (e.g., Mattan et
al., 2019; Thiem et al., 2019). Like Target D Scores, this method allows researchers to study
systematic variation in implicit evaluations within target groups, and to control for target-level
simple measure of samples‘ overall implicit evaluations of individual targets, and allow for the
fitting of more straightforwardly interpretable models compared to raw response time models,
which typically require interaction terms between target-level characteristics and indicators of
allow researchers to assess the reliability of measured evaluations of targets. This allows
53
INTERSECTIONAL IMPLICIT BIAS
distinguishing between ranges of response times that contribute reliable information regarding
implicit evaluations, and ranges of response times that contribute only unhelpful random noise. 32
Some limitations regarding the present research should be noted. The first regards our
restricted ability to detect higher order three-way or four-way interactions between target-level
factors. We were reasonably well-powered to detect two-way interactions, which the multiple
However, there are other possible three- or four-way interaction patterns which do not entail the
presence of two-way interactions, and we did not test for these given our limited number of
stimuli. It is also plausible that there exist interactions which imply two-way interactions but
whose effect sizes fall below levels we were sufficiently powered to detect. Consequently, while
the present results do speak against the idea that certain patterns of interactions––including
evaluations of multiply categorizable targets, they do not speak to the existence of such effects at
effect of target gender in ST-IATs than the EPT and AMP. Previous researchers have argued that
likely to produce category dominance (Gawronski et al., 2010). However, these researchers
theorized that EPTs––which also rely on response interference––would produce greater category
dominance effects than AMPs. By contrast, we observed a more dominant effect of gender in the
AMP than the EPT, suggesting it is unlikely our category dominance results were a function of
32
This was well illustrated in Study 4, where we observed Target D Scores to capture virtually
zero reliable variation when we applied our ST-IAT algorithm directly to the EPT data. If we had
relied on response time-level modelling in the present project, we would not have known that the
EPT data required a different scoring algorithm altogether to obtain some level of internally
reliable measurement.
33
For example, if there were a three-way multiple jeopardy effect resulting in especially negative
evaluations of lower SES Black male targets, tests of two-way interactions should in theory
detect especially negative evaluations of Black male targets, lower SES Black targets, lower SES
response interference tasks alone. One potentially important difference separating the ST-IAT
method from the EPT and AMP methods is its reliance on key presses made directly in response
to the multiply categorizable targets themselves, rather than subsequently displayed words (the
EPT) or Chinese characters (the AMP). Plausibly, there may be a temporal factor involved in the
after stimulus presentation, and thereafter reduced, or a focal effect, whereby tasks requiring
responses directly to targets focus attention on targets‘ dominant categories in a way that other
It is also unclear why the effect of target race varied across measurement tasks. Here, the
most anomalous result was the anti-White bias displayed in the AMP, which runs counter to the
anti-Black evaluative bias typically displayed by U.S. adults (e.g., Nosek, 2005), and previously
demonstrated via AMPs using multiply categorizable targets (Gawronski et al., 2010). This result
also ran counter to the anti-Black bias displayed by our samples in Studies 3 and 4a via ST-IATs.
However, it is not unprecedented to obtain results counter to expectations when using the AMP
to detect implicit prejudice (Teige-Mocigemba, Becker, Sherman, Reichardt, & Klauer, 2017).
Given the number of studies run in the present manuscript, as well as the number of effects
measured in each study, an anomalous result of this nature is perhaps not surprising.
Nonetheless, it is worth noting that target race in general tended to produce relatively
inconsistent effects compared to target gender and social class, regardless of the measurement
method. In Studies 1a, 1b, and 2, we observed no robust effect of race, and only in Studies 3a,
3b, and 4a did we observe robust anti-Black race effects in-keeping with prior literature. As
discussed above, one explanation for these results may be that because targets of different race
were presented in separate ST-IATs in Studies 1 and 2, participants used recoding strategies
(Meissner & Rothermund, 2013) to suppress anti-Black bias in these studies. Another is that due
to perceived causal effects of race on social class (Pew Research Centre, 2019), and the process
of augmentation (Kelley, 1973), matching target groups on explicit ratings of social class in
Studies 1 and 2 may have led to the Black targets being perceived as higher on other traits
conferring social class status, such as competence or industriousness. But neither of these
explanations accounts for the anti-White bias observed in the AMP task in Study 4b. Given the
consistency with which anti-Black bias is typically displayed in two-category IATs, the
inconsistency of race effects in the present work is itself noteworthy, as it provides further
55
INTERSECTIONAL IMPLICIT BIAS
evidence that we are yet to fully understand implicit bias in the context of complex, multiply
categorizable targets.
Other major challenges for future research include incorporating even greater naturalistic
complexity within target stimuli. In the present research, we focused on target-level variation in
race, gender, social class, and age—four target dimensions that are perceptible in many if not
most social interactions. Of course, real-world social targets vary on far more than just these four
variables; modelling such complexity will require the study of other social variables, including
variation in body shape (Bessenoff & Sherman, 2000; Teachman, Gapinski, Brownell, Rawlins,
& Jeyaram, 2003), sexual orientation (Banse, Seise, & Zerbes, 2001; Steffens & Buchner, 2003),
social and physical contexts (Barden et al., 2004; Wittenbrink et al., 2001), facial expressions
Finally, the present work focused only on identifying basic implicit evaluative biases
therefore be vital to assess how well implicit evaluations of multiply categorizable targets align
with explicit bias measures, and how well each kind of measure predicts discriminatory
behaviors. One key criticism of traditional implicit bias tests has been their relatively low
correlations with discriminatory behavior (e.g., Oswald, Mitchell, Blanton, Jaccard, & Tetlock,
2013; but see Jost et al., 2009; Greenwald, Banaji, & Nosek, 2015). It may be the case that
participants‘ spontaneously displayed implicit biases toward multiply categorizable targets will
better predict behavior in real social contexts than traditional binary measures. This possibility is
References
Almquist, E. M. (1975). Untangling the effects of race and sex: The disadvantaged status of
Adler, N. E., Epel, E. S., Castellazzo, G., & Ickovics, J. R. (2000). Relationship of subjective and
objective social status with psychological and physiological functioning: Preliminary data
Banaji, M.R., & Hardin, C.D. (1996). Automatic stereotyping. Psychological Science, 7, 136–
141.
Banse, R., Seise, J., & Zerbes, N. (2001). Implicit attitudes towards homosexuality: Reliability,
validity, and controllability of the IAT. Zeitschrift für experimentelle Psychologie, 48(2),
145-160.
Barden, J., Maddux, W. W., Petty, R. E., & Brewer, M. B. (2004). Contextual moderation of
racial bias: the impact of social roles on controlled and automatically activated
Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models
Beale, F. (1970). Double jeopardy: To be Black and female. In T. Cade (Ed.), The Black woman
Becker, J., Kraus, M. W., & Rheinschmidt-Same, M. L. (2017). Cultural expressions of social
class and their implications for beliefs and behavior. Journal of Social Issues.
Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and
powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B
Bessenoff, G. R., & Sherman, J. W. (2000). Automatic and controlled components of prejudice
toward fat people: Evaluation versus stereotype activation. Social Cognition, 18(4), 329-
353.
Bluemke, M., & Friese, M. (2008). Reliability and validity of the Single‐Target IAT (ST‐IAT):
assessing automatic affect towards multiple attitude objects. European journal of social
Borg, I., & Groenen, P. J. (2005). Modern multidimensional scaling: Theory and applications.
Bowleg, L. (2008). When Black + lesbian + woman ≠ Black lesbian woman: The methodological
challenges of qualitative and quantitative intersectionality research. Sex roles, 59(5), 312-
325.
Brewer, M. B., Ho, H. K., Lee, J. Y., & Miller, N. (1987). Social identity and social distance
among Hong Kong schoolchildren. Personality and Social Psychology Bulletin, 13(2),
156-165.
Brown, R. J., & Turner, J. C. (1979). The criss-cross categorization in intergroup discrimination.
Carlana, M. (2019). Implicit stereotypes: Evidence from teachers‘ gender bias. The Quarterly
170.
Correll, J., Park, B., Judd, C. M., & Wittenbrink, B. (2002). The police officer's dilemma: Using
Crisp, R. J., Hewstone, M., & Rubin, M. (2001). Does multiple categorization reduce intergroup
Curran, P. J., & Hussong, A. M. (2009). Integrative data analysis: the simultaneous analysis of
DeGraffenreid v. GENERAL MOTORS ASSEMBLY DIV., ETC., 413 F. Supp. 142 (E.D. Mo.
1976). https://law.justia.com/cases/federal/district-courts/FSupp/413/142/1660699/
De Leeuw, J., & Mair, P. (2009). Multidimensional scaling using majorization: SMACOF in
Diehl, M. (1990). The minimal group paradigm: Theoretical explanations and empirical
Dijksterhuis, A., & Van Knippenberg, A. D. (1996). The knife that cuts both ways: Facilitated
Draine, S. C., & Greenwald, A. G. (1998). Replicable unconscious semantic priming. Journal of
Dugard, P., Todman, J., & Staines, H. (2010). Approaching multivariate analysis. A practical
Edwards, L. J., Muller, K. E., Wolfinger, R. D., Qaqish, B. F., & Schabenberger, O. (2008). An
R2 statistic for fixed effects in the linear mixed model. Statistics in medicine, 27(29),
6137-6157.
categorisations: the effect on intergroup bias, Personality and Social Psychology Bulletin,
21, 149–159.
Fazio, R. H., Jackson, J. R., Dunton, B. C., & Williams, C. J. (1995). Variability in automatic
Fazio, R. H., Sanbonmatsu, D. M., Powell, M. C., & Kardes, F. R. (1986). On the automatic
Fiske, S. T, & Taylor, S. E. (1991). Social cognition (2nd ed.). New York: McGraw-Hill.
Gawronski, B., Cunningham, W. A., LeBel, E. P., & Deutsch, R. (2010). Attentional influences
Gillath, O., Bahns, A. J., Ge, F., & Crandall, C. S. (2012). Shoes as a source of first impressions.
https://doi.org/10.1016/j.jrp.2012.04.003
Goff, P. A., & Kahn, K. B. (2013). How psychological science impedes intersectional
Greenwald, A. G., & Lai, C. K. (2020). Implicit social cognition. Annual Review of
Psychology, 71.
Greenwald, A. G., Banaji, M. R., & Nosek, B. A. (2015). Statistically small effects of the
Implicit Association Test can have societally large effects. Journal of Personality and
Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. (1998). Measuring individual differences in
implicit cognition: the implicit association test. Journal of personality and social
Greenwald, A. G., Nosek, B. A., & Banaji, M. R. (2003). Understanding and using the implicit
Hainmueller, J., Hopkins, D. J., & Yamamoto, T. (2014). Causal inference in conjoint analysis:
Hewstone, M., Islam, M. R., & Judd, C. M. (1993). Models of crossed categorization and
Horwitz, S. R., & Dovidio, J. F. (2017). The rich—love them or hate them? Divergent implicit
and explicit attitudes toward the wealthy. Group Processes & Intergroup
Islam, M. R., & Hewstone, M. (1993). Intergroup attributions and affective consequences in
majority and minority groups. Journal of Personality and Social Psychology, 64(6), 936.
Jaeger, B. (2017). r2glmm: Computes R Squared for Mixed (Multilevel) Models. R package
Jones, C. R., & Fazio, R. H. (2010). Person categorization and automatic racial stereotyping
1073-1085.
Jost, J. T., Rudman, L. A., Blair, I. V., Carney, D. R., Dasgupta, N., Glaser, J., & Hardin, C. D.
ideological and methodological objections and executive summary of ten studies that no
Kang, S. K., & Bodenhausen, G. V. (2015). Multiple identities in social perception and
Karpinski, A., & Steinman, R. B. (2006). The single category implic it association test as a
measure of implicit social cognition. Journal of personality and social psychology, 91(1),
16.
Kelley, H. H. (1973). The processes of causal attribution. American psychologist, 28(2), 107.
60
INTERSECTIONAL IMPLICIT BIAS
King, D. K. (1988). Multiple jeopardy, multiple consciousness: The context of a Black feminist
Koch, A., Imhoff, R., Dotsch, R., Unkelbach, C., & Alves, H. (2016). The ABC of stereotypes
Kuznetsova, A., Brockhoff, P, & Christensen, R. (2017). lmerTest package: Tests in linear mixed
Landrine, H., Klonoff, E.A., Alcaraz, R., Scott, J., & Wilkins, P. (1995). Multiple variables in
Levinson, J. D., & Young, D. (2010). Implicit gender bias in the legal profession: An empirical
Livingston, R. W., & Brewer, M. B. (2002). What are we really priming? Cue-based versus
Ma, D. S., Correll, J., & Wittenbrink, B. (2015). The Chicago face database: A free stimulus set
Macrae, C., Bodenhausen, G. V., & Milne, A. B. (1995). The dissection of selection in person
Mattan, B. D., Kubota, J. T., Li, T., Venezia, S. A., & Cloutier, J. (2019). Implicit Evaluative
Biases Toward Targets Varying in Race and Socioeconomic Status. Personality and
Meissner, F., & Rothermund, K. (2013). Estimating the contributions of associations and
recoding in the Implicit Association Test: The ReAL model for the IAT. Journal of
Mitchell, J. P., Nosek, B. A., & Banaji, M. R. (2003). Contextual variations in implicit
Moore‐Berg, S., Karpinski, A., & Plant, E. A. (2017). Quick to the draw: How suspect race and
Nicolas, G., de la Fuente, M., & Fiske, S. T. (2017). Mind the overlap in multiple categorization:
Nosek, B. A., & Banaji, M. R. (2001). The Go/No-go Association Task. Social Cognition, 19,
625–666.
Nosek, B. A., Banaji, M. R., & Greenwald, A. G. (2002). Harvesting implicit group attitudes and
beliefs from a demonstration web site. Group Dynamics: Theory, Research, and
Oswald, F. L., Mitchell, G., Blanton, H., Jaccard, J., & Tetlock, P. E. (2013). Predicting ethnic
Payne, B. K. (2001). Prejudice and perception: The role of automatic and controlled processes in
Payne, B. K., Cheng, C. M., Govorun, O., & Stewart, B. D. (2005). An ink‐ blot for attitudes:
Perszyk, D. R., Lei, R. F., Bodenhausen, G. V., Richeson, J. A., & Waxman, S. R. (2019). Bias
Petsko, C. D., & Bodenhausen, G. V. (2019). Multifarious person perception: How social
https://www.pewsocialtrends.org/wp-content/uploads/sites/3/2019/04/Race-
report_updated-4.29.19.pdf
R Core Team. (2019). R: A language and environment for statistical computing. R Foundation
Ransford, H. E. (1980). The prediction of social behavior and attitudes. In V. Jeffries & H. E.
Richeson, J. A., & Ambady, N. (2001). Who's in charge? Effects of situational roles on
Rudman, L. A., Feinberg, J., & Fairchild, K. (2002). Minority members' implicit attitudes:
Automatic ingroup bias as a function of group status. Social Cognition, 20(4), 294-320.
Rudman, L. A., & Goodwin, S. A. (2004). Gender differences in automatic ingroup bias: Why do
women like women more than men like men? Journal of Personality and Social
Singh, R., Yeoh, B. S., Lim, D. I., & Lim, K. K. (1997). Cross‐categorization effects in
Schmid-Mast, M., & Hall, J. A. (2004). Who is the boss and who is not? Accuracy of judging
https://doi.org/10.1023/b:jonb.0000039647.94190.21
Steele, J. R., George, M., Cease, M. K., Fabri, T. L., & Schlosser, J. (2018). Not always Black
and White: The effect of race and emotional expression on implicit attitudes. Social
Steffens, M. C., & Buchner, A. (2003). Implicit Association Test: separating transsituationally
Tajfel, H., & Turner, J. C. (1979). An integrative theory of intergroup conflict. In W. G. Austin
& S. Worchel (Eds.), The social psychology of intergroup relations (pp. 33–47).
Teachman, B. A., Gapinski, K. D., Brownell, K. D., Rawlins, M., & Jeyaram, S. (2003).
Demonstrations of implicit anti-fat bias: the impact of providing causal information and
Teige-Mocigemba, S., Becker, M., Sherman, J. W., Reichardt, R., & Klauer, K. C. (2017). The
Thiem, K. C., Neel, R., Simpson, A. J., & Todd, A. R. (2019). Are black women and girls
associated with danger? Implicit racial bias at the intersection of target age and
van Oudenhoven, J. P., Judd, C. M., & Hewstone, M. (2000). Additive and interactive models of
Vanbeselaere, N. (1991). The different effects of simple and crossed categorizations: A result of
Wittenbrink, B., Judd, C. M., & Park, B. (2001). Spontaneous prejudice in context: Variability in
815.
Wigboldus, D. H., Holland, R. W., & van Knippenberg, A. (2004). Single target implicit
Xu, K., Nosek, B., & Greenwald, A. (2014). Data from the race implicit association test on the
doi:10.5334/jopd.ac
Yamaguchi, M., & Beattie, G. (2019). The role of explicit categorization in the Implicit