853842

research-article2019
PSPXXX10.1177/0146167219853842Personality and Social Psychology BulletinRivers et al.

Article

Personality and Social

On the Roles of Stereotype Activation and

Psychology Bulletin
2020, Vol. 46(3) 349­–364
© 2019 by the Society for Personality
Application in Diminishing Implicit Bias and Social Psychology, Inc
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/0146167219853842
https://doi.org/10.1177/0146167219853842
journals.sagepub.com/home/pspb

Andrew M. Rivers1, Jeffrey W. Sherman1, Heather R. Rees1,

Regina Reichardt2, and Karl C. Klauer3

Abstract
Stereotypes can influence social perception in undesirable ways. However, activated stereotypes are not always applied in
judgments. The present research investigated how stereotype activation and application processes impact social judgments
as a function of available resources for control over stereotypes. Specifically, we varied the time available to intervene in
the stereotyping process and used multinomial modeling to independently estimate stereotype activation and application. As
expected, social judgments were less stereotypic when participants had more time to intervene. In terms of mechanisms,
stereotype application, and not stereotype activation, corresponded with reductions in stereotypic biases. With increasing
time, stereotype application was reduced, reflecting the fact that controlling application is time-dependent. In contrast,
stereotype activation increased with increasing time, apparently due to increased engagement with stereotypic material.
Stereotype activation was highest when judgments were least stereotypical, and thus, reduced stereotyping may coincide
with increased stereotype activation if stereotype application is simultaneously decreased.

Keywords
prejudice/stereotyping, self-regulation, implicit cognition, multinomial modeling

Received January 11, 2018; revision accepted April 28, 2019

A vast literature in psychological science demonstrates that application is critical to understanding when stereotypes will
stereotypes influence social judgment (e.g., Hamilton & or will not bias judgments and to understanding how to effec-
Sherman, 1994). In part, this influence reflects the fact that tively intervene to influence stereotyping.
stereotypes are readily activated, or made accessible, upon
perceiving members of distinct social groups (e.g., Kunda,
Operating Conditions of Stereotype
Davis, Adams, & Spencer, 2002; Kunda & Spencer, 2003).
Indeed, stereotype activation frequently exhibits characteris- Activation and Application
tics of automaticity. For example, perceiving a stereotypic Prominent models of stereotyping propose distinct roles for
image can lead to stereotype activation even when the image stereotype activation and application. For example, Devine’s
is incidental to current goals or is perceived outside of aware- (1989) dual-process model proposed that stereotype activa-
ness entirely (Kawakami, Dovidio, & Dijksterhuis, 2003; tion is inescapable among individuals with knowledge of
Lepore & Brown, 1997). Considerable research has addition- those stereotypes. However, the influence of activated ste-
ally shown that stereotypes are more likely to influence judg- reotypes on judgment can be modulated depending on the
ments when processing resources are scarce, supporting the perceiver’s concurrent processing goals and their ability to
view that stereotyping is an efficient process (Sherman,
Macrae, & Bodenhausen, 2000). 1
University of California, Davis, USA
Although stereotypic biases are pervasive, in some cir- 2
University of Regensburg, Germany
cumstances, people are able to correct for their biases (e.g., 3
University of Freiburg, Freiburg im Breisgau, Germany
Monteith, Ashburn-Nardo, Voils, & Czopp, 2002; Monteith,
Corresponding Authors:
Sherman, & Devine, 1998). Thus, there is a conceptual dis- Andrew M. Rivers, Department of Psychology, University of California,
tinction between the activation of stereotypes and their appli- Davis, 1 Shields Avenue, Davis, CA 95616, USA.
cation to judgment (e.g., Gilbert & Hixon, 1991). Activated Email: [email protected]
stereotypes are not always applied and, sometimes, judg- Jeffrey W. Sherman, Department of Psychology, University of California,
ments are contrasted or corrected away from activated ste- Davis, 1 Shields Avenue, Davis, CA 95616, USA.
reotypes. The distinction between stereotype activation and Email: [email protected]
350 Personality and Social Psychology Bulletin 46(3)

correct for the influence of active stereotypes. Fazio, Jackson, of stereotypes need not be deliberate and resource-intensive,
Dunton, and Williams (1995) also emphasize the role of cor- and that people can rapidly shift the likelihood that they
rective processes following stereotype activation, positing apply activated stereotypes to judgments. These corrective
that stereotypes are automatically activated upon encounter- shifts have been observed even on implicit tasks that are
ing members of another group, but that corrective processes typically thought to preclude intentional, resource-dependent
can intervene to modulate their influence. Motivations and processes (Glaser & Kihlstrom, 2006; Glaser & Knowles,
situational conditions conducive to the operation of correc- 2008; Sherman, 2006; Sherman et al., 2008).
tive processes are critical (Fazio, 1990; Fazio et al., 1995). Providing additional support for efficient stereotype cor-
When people are motivated to correct against biasing influ- rection, Moskowitz and colleagues (Moskowitz, Gollwitzer,
ences and conditions are sufficient for the operation of cor- Wasel, & Schall, 1999; Moskowitz & Li, 2011) have shown
rective processes, then judgments will deviate from activated that stereotype activation can be preconsciously inhibited
stereotypic knowledge. when people chronically pursue egalitarian goals or when
The models proposed by Devine and Fazio are reflective egalitarian goals are made salient. These inhibitory mecha-
of a broad consensus that stereotype activation is relatively nisms can operate under conditions thought to preclude the
more automatic (i.e., unintentional, resistant to interfer- operation of strategic processes. Specifically, those who pur-
ence, outside of awareness, and efficient) than corrective sue egalitarian goals appear to inhibit stereotype activation,
processes that prevent the application of stereotypes to even on sequential priming measures using brief stimulus
judgment (e.g., Hamilton & Sherman, 1994). An important onset asynchronies (SOAs), which significantly limit the
implication is that the conditions that permit stereotype time available to correct for stereotyping. Altogether, these
activation are assumed to be less restrictive than are the results indicate that the characterizations of stereotype acti-
conditions that permit correction. Thus, factors that selec- vation as automatic and stereotype application (and correc-
tively interfere with stereotype correction should increase tion) as controlled may be overly simplified.
stereotyping by increasing the likelihood that stereotypes
will be applied in judgments. Conversely, factors that Measuring Stereotype Activation and
increase the ability to intervene should reduce stereotyping
via reduced application.
Application
Progress in understanding the impact and nature of stereo-
type activation and application has been hampered by
Resource-Dependent Stereotype Activation methodological limitations. In particular, the presumption
Although the broad characterization of stereotype activa- that stereotype activation is automatic whereas stereotype
tion as automatic and stereotype application as controlled application or its converse, stereotype correction, is nonau-
has been the consensual view of stereotyping for some tomatic has formed the theoretical rationale for measuring
time, there are hints that it may be oversimplified. For activation and application with different experimental
example, work by Gilbert and Hixon (1991) showed that tasks. Because it is assumed to be an automatic process
the availability of cognitive resources influenced both ste- (unintentional, resource independent, lacking awareness,
reotype activation and application. In particular, the pres- spontaneous), stereotype activation has generally been
ence of a cognitive load while encountering stereotypic assessed with implicit measures (e.g., Implicit Association
content reduced the extent to which stereotypes were acti- Test [IAT], sequential priming), which are commonly
vated. In addition, stereotypic judgments were particularly understood to measure the activation of mental associations
likely if (a) participants were not cognitively busy during a and preclude the influence of controlled cognitive pro-
stereotype activation phase (and, therefore, had stereotypes cesses. In contrast, because stereotype correction is assumed
activated) and (b) were busy during a stereotype applica- to be a nonautomatic process (intentional, resource depen-
tion phase (and, therefore, could not correct for stereotypic dent, involving awareness, strategic), it has generally been
influence). Completing a simultaneous task appeared to assessed with explicit judgment tasks, which are presumed
influence both stereotype activation and stereotype applica- to reflect controlled cognitive processes while minimizing
tion, challenging the prevailing notion that stereotype acti- the influence of automatic processing.
vation is categorically automatic. An implicit assumption of this “task dissociation”
approach (for a review, see Sherman, Krieglmeyer, &
Calanchini, 2014) is that the tasks used to measure stereo-
“Automatic” Corrective Processes type activation and stereotype application are process-
Other research has challenged the idea that stereotype-cor- pure—that is, that they uniquely measure the process in
recting processes are necessarily resource-intensive. Glaser question without contamination from other mental pro-
and Kihlstrom’s (2006) Compensatory Automaticity model cesses. In other words, the task-dissociation approach
asserts that habitual correction for stereotypic biases routin- assumes that measures of activation (e.g., sequential prim-
izes corrective processes such that preventing the application ing task) reflect only differences in activation and not
Rivers et al. 351

differences in application, whereas measures of application briefly exposed to photographs of Black or White male prime
(e.g., explicit judgment task) reflect only application. faces. In past research, exposure to these Black or White
Returning to Gilbert and Hixon’s (1991) work investigating primes influenced how threatening the target faces were
the automaticity of stereotype activation and application, we judged to be. To examine the influence of processing
find an example of this approach. Participants in Gilbert and resources on stereotyping, stereotype activation, and stereo-
Hixon’s (1991) experiments completed a word-fragment type application, we varied the time separating onset of
completion task in which each fragment could be completed prime and target images, or SOA. Previous research has
in either a stereotype-relevant or stereotype-irrelevant way. shown that the influence of primes on target judgments is
The number of stereotypic word completions was inter- reduced with longer SOAs (Fazio et al., 1995). As such, we
preted as an index of stereotype activation, uncontaminated expected that, as SOA increased, the influence of racial
by stereotype application or the prevention thereof. On a primes on judgments of target images would be reduced.
subsequent task, participants judged the degree to which Most importantly, we apply a statistical model that allows us
their Asian or White interaction partner possessed stereo- to derive independent estimates of stereotype activation and
typic traits. This task was assumed to measure the degree to application to shed light on how these processes relate to ste-
which participants would apply or prevent the application of reotyping, and how they are influenced by SOA. The influ-
stereotypes to their judgments. However, there is now a ence of stereotypes may be reduced as SOA increases
great deal of evidence that measures of stereotype activa- because stereotype activation is diminished, either due to
tion, even implicit measures (e.g., sequential priming, IAT), inhibition of the stereotype (Monteith et al., 1998; Moskowitz
also reflect the influence of stereotype application and cor- & Li, 2011) or to passive decay of activated stereotypic
rective processes (Krieglmeyer & Sherman, 2012; Sherman knowledge (Kunda et al., 2002). But increases in SOA might
et al., 2008). Similarly, explicit judgment tasks used to mea- also allow people to better prevent the application of acti-
sure stereotype application are necessarily influenced by the vated stereotypes to their judgments (e.g., Glaser & Knowles,
extent of stereotype activation. Thus, it is difficult to deter- 2008; Sherman et al., 2008). In this case, it is not the extent
mine the extent to which performance on either kind of task of stereotype activation that is critical, but the degree to
reflects stereotype activation, stereotype application, or which people apply active stereotypes. Application of the
mixtures of both processes. The task-dissociation approach SMT model allows us to independently test how SOA affects
precludes strong conclusions about the conditions under each of these mechanisms and the likelihood that they con-
which these processes do or do not occur. tribute to reductions in stereotypic judgments.

The Present Research Experiment 1

The present research is the first investigation into the operat-
ing conditions of stereotype activation and application that
Participants
does not rely on the task-dissociation approach. In this line of In Experiment 1, we sought to collect data from at least 80
work, we use the Stereotype Misperception Task (SMT; participants. In total, 92 undergraduate students1 at the
Krieglmeyer & Sherman, 2012), a sequential priming proce- University of California, Davis participated in the experi-
dure designed specifically to disentangle the joint influences ment for partial course credit (71.1% female, Mage = 18.9
of stereotype activation and application processes. A mathe- years; 58% Asian, 21% Caucasian, 20% Latino/a).
matical model of SMT task performance independently Participants completed the experiment as the second of three
assesses the extents of stereotype activation and application unrelated tasks in an hour-long experimental session. Based
within the same task and under identical conditions, avoiding on previously reported effect sizes from Krieglmeyer and
the inherent problematic assumptions of task dissociation. In Sherman (2012; d = 1.04), Experiment 1 was powered well
this line of work, we investigate the roles of stereotype acti- above .95 to detect stereotypic bias in the SMT procedure.
vation and application under conditions that vary the time However, we were specifically interested in changes in bias
available to correct for the influence of stereotypes, provid- resulting from our within-subjects manipulation of SOA.
ing a direct test of the extent to which the processes of stereo- After choosing our desired sample size, sensitivity analysis
type activation and application operate efficiently and the in G*Power 3.1 indicated that 80 participants would allow us
extent to which they are dependent on the opportunity for to detect an effect size of dz = .317, corresponding to a small
intentional processing. We also examine how stereotype acti- to medium effect, with power set at 1 − β = .80 (Faul,
vation and application relate to biases in judgment, and how Erdfelder, Buchner, & Lang, 2009). Power analysis in subse-
those relationships may be affected by the opportunity for quent experiments relies on observed effects of the SOA
intentional processing. manipulation. Two participants were excluded from analyses
In the current variant of the SMT, participants judged a for using a single key response for all trials.2 Including these
series of pixelated target faces according to their level of data in analyses does not change the direction or statistical
threat. Prior to viewing these target faces, participants were interpretation of the reported results.
352 Personality and Social Psychology Bulletin 46(3)

Stimuli was presented for 150 ms following the fixation cross. On

half of the trials, a blank screen was then displayed for 175
Prime stimuli consisted of photographs of 24 Black and 24 ms, followed by onset of a target image. On the other half of
White males, each approximately 20 to 30 years old. Each the trials, target images were presented immediately after
face was cropped at the base of the neck and superimposed offset of prime images. This created trials with one of two
on a plain gray background (see Phills, Kawakami, Tabi, within-subjects levels of SOA: 150 ms or 325 ms. Target
Nadolny, & Inzlicht, 2011). In addition to these, we included images were always shown for 100 ms and were then
a set of “neutral” prime images that contained no racial cues. replaced by a static visual mask that remained on the screen
These consisted of the outline of a face superimposed on a until the participant rendered a judgment. Participants indi-
gray background (see Krieglmeyer & Sherman, 2012). cated whether they judged the target to be more or less
Target stimuli consisted of 48 computer-generated face threatening than average by pressing either the “D” or “K”
morphs created by Oosterhof and Todorov (2008) to vary keys on a standard computer keyboard. The following trial
systematically in perceived threat. Each of the computer- began 500 ms after the previous judgment.
generated images depicts a male face with morphed facial
features that were either two standard deviations above or
below the neutral point of threat. These images were dis- Results
torted with a pixilation filter using photo-editing software Analytic plan. The full design for Experiment 1 was 3 (prime
to increase their ambiguity. Target images differ objectively type: Black vs. White vs. neutral) × 2 (target type: high
along the dimension of evaluation to apply the multinomial threat vs. low threat) × 2 (SOA: 150 vs. 325 ms), with all
processing tree model to the data (see “Multinomial factors manipulated within subjects. Our analysis is com-
Modeling Analyses” section below). The choice of ±2 prised of two stages. In addition to quantifying the effect of
standard deviations was made to be consistent with the SOA on mean levels of racial bias, we use the SMT multino-
original validation of the SMT procedure reported in mial processing tree model to disentangle the contribution of
Krieglmeyer and Sherman (2012).3 multiple component processes (see “Multinomial Processing
Analyses” section below for a detailed description).
In the first analytic stage, we investigate whether partici-
Procedure
pants exhibited stereotypic biases in their judgments about
In each session, one to four participants completed experi- target stimuli. The extent of bias on the SMT, called the SMT
mental procedures individually in separate computer cubi- Effect, is determined by comparing responses on trials that
cles. After providing informed consent, participants learned include Black versus White prime faces (see Krieglmeyer &
that the experiment was concerned with rapid impression Sherman, 2012, for validation of this index). We describe the
formation. Participants then viewed SMT task instructions traditional full factorial ANOVA analyses in the supplemen-
and completed 12 practice trials of the SMT procedure. Two tal appendix because they are not the focus of the present
test blocks of the SMT procedure followed the practice trials. findings. For the SMT, the design consists of two prime lev-
After completing SMT test blocks, participants completed els (Black vs. White) and two SOA levels (150 vs. 325 ms).
several exploratory measures, a brief demographic question- This analytic stage quantifies (a) the magnitude of racial bias
naire (age, sex, ethnicity), and an experimental debriefing.4 in participants’ judgments and (b) the degree to which SOA
moderates racial bias in participants’ judgments.
SMT. The SMT developed by Krieglmeyer and Sherman In the second analytic stage, we investigate how SOA
(2012) served as the dependent measure. The structure of the influences the component processes of stereotype activation
SMT is similar to commonly used sequential priming mea- and application. Thus, in the second analytic stage, we apply
sures (e.g., Affect Misattribution Procedure; Payne, Cheng, the SMT multinomial processing tree model, which has been
Govorun, & Stewart, 2005). In the present experiment, par- developed and validated for exactly this purpose—to investi-
ticipants were instructed to “respond as quickly as possible,” gate patterns of responding across the SMT procedure
judging each target face morph as either “more threatening” (Krieglmeyer & Sherman, 2012; also see Payne, Hall,
or “less threatening” than the average target. Each target was Cameron, & Bishara, 2010, for a similar approach with the
preceded by one of the three prime types (Black face, White Affect Misattribution Procedure). To accomplish this, the
face, or neutral face outline). Participants were explicitly SMT model analysis includes responses to all three prime
informed to not respond to the prime faces but to attend to and both target types.
them for later questions. Because of these explicit instruc-
tions, any influence of the prime faces on threat judgments is SMT effects. Our first research question pertained to the mag-
assumed to be unintentional. nitude of stereotypic biases in threat judgments as a function
The two SMT test blocks consisted of 72 trials each and of SOA. The SMT effect, an index of stereotypic bias, is cal-
included a self-paced break between the two blocks. Each culated by subtracting the proportion of “more threatening”
trial began with a fixation cross for 500 ms. A prime image judgments on White prime trials from the proportion of “more
Rivers et al. 353

Table 1. Proportion “More Threatening” Responses as a Function of Prime (White vs. Neutral vs. Black), Target (Low vs. High Threat),
and Experimental Level for Experiments 1 to 4.

White prime Neutral prime Black prime

Low threat High threat Low threat High threat Low threat High threat
Experiment 1
150-ms SOA .23 (.21) .29 (.24) .27 (.27) .42 (.30) .49 (.30) .54 (.27)
325-ms SOA .36 (.24) .40 (.24) .28 (.28) .41 (.31) .54 (.29) .55 (.26)
Experiment 2
150-ms SOA .28 (.20) .31 (.20) .25 (.28) .40 (.29) .49 (.27) .51 (.25)
200-ms SOA .33 (.21) .36 (.22) .27 (.29) .41 (.30) .51 (.27) .52 (.30)
325-ms SOA .38 (.24) .42 (.23) .29 (.29) .39 (.29) .49 (.25) .54 (.26)
Experiment 3
150-ms SOA .36 (.27) .36 (.22) .24 (.30) .31 (.32) .33 (.30) .38 (.28)
325-ms SOA .38 (.26) .42 (.29) .23 (.27) .34 (.33) .31 (.31) .42 (.27)
Experiment 4
175-ms SOA with mask .33 (.26) .38 (.26) .26 (.28) .31 (.32) .49 (.30) .48 (.31)
175-ms SOA no mask .31 (.27) .29 (.27) .25 (.34) .32 (.34) .52 (.30) .52 (.27)
350-ms SOA with mask .44 (.27) .46 (.25) .29 (.35) .38 (.36) .52 (.29) .50 (.30)
350-ms SOA no mask .47 (.22) .45 (.27) .28 (.34) .36 (.35) .44 (.32) .42 (.30)

Note. Standard deviations appear in parentheses. SOA = stimulus onset asynchrony.

Table 2. Proportion “More Threatening” Response Difference threat stereotypes are not activated (1 − SAC), a parameter
Score (Black–White Prime) and SMT Effect Size Estimates in representing detection (D) captures the ability to accurately
Experiment 1 by SOA. detect target threat level and respond accordingly. Finally,
Experiment 1 Difference score SMT effect dz when threat stereotypes are not activated (1 − SAC) and
target characteristics are not detected (1 − D), a guessing
150-ms SOA .25 [.18, .33] .737 parameter (G) captures general tendencies to respond with
325-ms SOA .16 [.09, .23] .494
high- or low-threat judgments.
Note. Values given in brackets denote 95% confidence intervals. To further explicate the model, consider the case in which
SMT = Stereotype Misperception Task; SOA = stimulus onset asynchrony. a participant is responding to a trial in which the prime face
is Black and the target image is low in threat. In this case, the
threatening” judgments on Black prime trials (Krieglmeyer & Black prime activates threat-related stereotypic content with
Sherman, 2012).5 We calculated the SMT effect for both SOA the probability of SAC. If the stereotype is activated and it is
conditions. The data confirmed our primary prediction; the applied to the judgment, the participant will render a “more
SMT effect was stronger when SOA between prime and target threat” judgment with the probability of SAC × SAP.
images was 150 ms versus 325 ms, t(89) = 4.074, p < .001; However, the participant may correct their judgment away
Hedges gav = .268; 95% confidence interval (CIdifference) = from the activated stereotype, rendering a “low threat” judg-
[.047, .137] (see Tables 1 and 2). To understand the ment with the probability of SAC × (1 − SAP). If the Black
mechanism(s) underlying this difference in bias, we employed prime does not activate the stereotype, participants may cor-
multinomial modeling in the second analytic stage. rectly detect the target image, rendering the “low threat”
judgment with probability (1 − SAC) × D. If the stereotype
Multinomial modeling analyses. We generated a set of equa- is not activated and the extent of threat in the target image is
tions representing the SMT process model that was devel- not accurately detected, then the participant may guess “high
oped and previously validated (see Krieglmeyer & Sherman, threat” with probability (1 − SAC) × (1 − D) × G or, alter-
2012). The SMT process model estimates four latent param- natively, may guess “low threat” with probability (1 − SAC)
eters (see Figure 1). At the initial branch of the model tree, × (1 − D) × (1 − G). Thus, the overall probability of a “high
an activation (SAC) parameter captures activation of racial threat” response on this trial is SAC × SAP + (1 − SAC) ×
stereotypes. That is, to what extent do primes activate ste- (1 − D) × G. The probability of a “low threat” response is
reotypic schema (e.g., the stronger association of threat SAC × (1 − SAP) + (1 − SAC) × D + (1 − SAC) × (1 − D)
with Black than White men)? When stereotypes are active, × (1 − G). According to the SMT, model estimates for SAP,
a parameter representing application (SAP) captures D, and G are conditional probabilities, in that each parame-
whether those stereotypes are applied to judgment or ter’s influence is dependent on the activity (or inactivity) of
whether judgments are corrected away from them. When another parameter (e.g., G is conditional on the inactivity of
354 Personality and Social Psychology Bulletin 46(3)

Figure 1. Multinomial processing tree of the SMT model.

Note. The top tree for trials with black and white primes and the bottom tree for neutral primes. The table on the right depicts responses as a function
of prime and target. The response “more threat” is represented by a “+” sign, and the response “less threat” is represented by a “–” sign. SMT =
Stereotype Misperception Task; SAC = stereotype activation; 1 − SAC = absence of stereotype activation; SAP = stereotype application; 1 − SAP =
stereotype correction; D = detection of target trait; 1 − D = failure to detect target trait; G = tendency to guess more threat; 1 − G = tendency to
guess less threat.

stereotype activation and detection). In contrast, the estimate effect size of model misspecification after controlling for
of SAC represents the unconditional probability of stereo- power (see Cressie, Pardo, & Pardo, 2003). The resulting
type activation. estimate of w = .022 indicated that the magnitude of model
Multinomial modeling analyses were conducted using the misfit was small after controlling for power.6
freely available multiTree computer package (Moshagen, Our primary questions of interest center on whether model
2010). This package implements a maximum likelihood parameters reliably respond to changes in prime-target SOA.
framework to test the goodness of model fit and to estimate To compare model parameters across the two SOA condi-
parameter values. Each of the four parameters is manipulated tions, we first fit a baseline model in which all parameters
in an iterative process until the model’s expected frequencies from the two SOA conditions were permitted to freely vary.
most closely approximate the observed response frequencies. By constraining corresponding parameters across SOA lev-
Parameter estimates vary between 0 and 1, and represent the els, we created nested models to test against the baseline
probability of process involvement. The magnitude of dis- model. The addition of any constraint necessarily reduces fit,
crepancy between model expectations and observed frequen- increasing G2. Large reductions in model fit result in higher
cies is expressed in the G2 statistic and corresponding p value. ΔG2 from baseline to nested models. Statistically significant
A nonsignificant result indicates that any discrepancy between p values extracted from ΔG2 indicate that the constrained
the expected data and the observed data were not detectable. model should be rejected in favor of the baseline model. In
other words, significant p values suggest that parameters dif-
Modeling results. Frequency counts of more and less fer between the levels being compared and should, therefore,
threatening responses were aggregated for each of the SMT be independently estimated (as in the baseline model). It is
trial types. We fit the SMT process model equations to aggre- important to point out that the SMT multinomial model esti-
gated counts from trials in which the SOA was 150 ms and mates the cognitive processes that are thought to underlie
325 ms. When fit to the data, the model G2 statistic suggested stereotyping, but this modeling approach cannot separately
that the fit of the SMT model was acceptable, G2(4) = 6.461, estimate processes that are specific to particular stimuli.
p = .167. To further quantify the magnitude of misfit, we Thus, our investigation tests how SOA influences these
calculated the w coefficient, which can be thought of as the domain-general stereotyping mechanisms.
Rivers et al. 355

Table 3. SMT Multinomial Model Parameter Estimates by Experimental Level in Experiments 1 to 4.

Parameter estimate D G SAC SAP

Experiment 1
150-ms SOA .11 [.08, .14] .32 [.29, .34] .35 [.24, .45] .87 [.75, .99]
325-ms SOA .13 [.09, .17] .32 [.30, .34] .77 [.68, .86] .60 [.58, .63]
Experiment 2
150-ms SOA .10 [.06, .15] .30 [.26, .33] .43 [.30, .55] .73 [.65, .81]
200-ms SOA .11 [.07, .16] .32 [.28, .35] .55 [.42, .68] .65 [.60, .69]
325-ms SOA .11 [.06, .16] .32 [.29, .36] .67 [.54, .80] .58 [.55, .61]
Experiment 3
150-ms SOA .06 [.01, .11] .26 [.22, .30] .37 [.23, .52] .49 [.43 .56]
325-ms SOA .10 [.04, .16] .26 [.22, .30] .59 [.46, .73] .51 [.47, .56]
Experiment 4
175-ms SOA with mask .05 [–.02, .12] .28 [.23, .32] .61 [.48, .74] .60 [.56, .64]
175-ms SOA without mask .01 [–.05, .07] .30 [.25, .34] .55 [.40, .70] .71 [.64, .77]
350-ms SOA with mask .12 [.04, .21] .29 [.24, .34] .94 [.82, 1.0] .52 [.50, .55]
350-ms SOA without mask .04 [–.03, .12] .31 [.27, .36] .66 [.50, .81] .49 [.45, .52]

Note. Values given in brackets denote 95% confidence intervals. SMT = Stereotype Misperception Task; SAC = stereotype activation; SAP = stereotype
application; SOA = stimulus onset asynchrony.

The D parameter could be collapsed across the two SOA Discussion

levels without a statistical loss of model fit, ΔG2(1) =
0.457, p = .499, w = .004. This indicates that D did not Data from Experiment 1 supported our primary prediction
detectably differ between the two SOA levels. Consistent that the biasing impact of primes on judgments would
with previously reported SMT data, the point estimate for D decrease as the time separating prime and target images
(see Table 3) was relatively low at both SOA levels, 150 ms increased. In addition, modeling analyses showed that
(D = 0.114; 95% CI = [.083, .145]) and at 325 ms (D = reduced stereotype application rather than reduced stereo-
0.131; 95% CI = [.093, .169]). However, that CIs did not type activation corresponded with reductions in bias. As time
overlap with 0 indicates that participants reliably discrimi- between prime and target increased, participants were less
nated between high- and low-threat targets. likely to apply prime-activated threat stereotypes. In addi-
Similarly, the G parameter did not differ between SOA tion, an unexpected finding emerged from Experiment 1; ste-
conditions, ΔG2(1) = 0.040, p = .841, w < .001. When no reotype activation was higher when SOA was longer versus
target or prime information was available to inform judg- shorter. This means that stereotype activation was highest
ments, 95% CIs of the G parameter did not overlap with 0.5, when judgments showed the least amount of racial bias.
indicating that participants tended to guess low threat at both
150-ms (G = 0.317; 95% CI = [.294, .341]) and 325-ms Experiment 2
SOA (G = 0.321; 95% CI = [.297, .344]).7
Constraining the SAP parameter across the two SOA lev-
Rationale
els led to a significant loss of fit, ΔG2(1) = 56.950, p < .001, The primary goal of Experiment 2 was to closely replicate
w = .066. This indicates that SAP differed significantly the aforementioned exploratory findings from Experiment 1.
between SOA conditions. The tendency to apply active ste- To be explicit, we now expected that estimates of stereotype
reotypes was lower at 325-ms SOA (SAP = .606; 95% CI = activation would again be higher at longer versus shorter
[.584, .629]) than at 150-ms SOA (SAP = .870; 95% CI = SOA. A secondary goal was to more deeply investigate the
[.749, .990]). CIs for the SAP parameter did not overlap with efficiency of stereotyping processes. In Experiment 1, we
0.5 at either SOA, indicating that active stereotypes tended to sought to understand the relationships among stereotype
be applied (vs. corrected against) at both levels. activation, stereotype application, and stereotypic judgment
The SAC parameter, likewise, differed between the two at the presumed boundaries of strategic responding. Early
SOA levels, ΔG2(1) = 34.619, p < .001, w = .052. semantic priming research suggested that SOAs below 400-
Unexpectedly, stereotype activation was higher at 325-ms ms eliminated such processes (Neely, 1977). However, more
SOA (SAC = .767; 95% CI = [.676, .858]) compared with recent work suggests that a strict threshold of 400-ms SOA is
150-ms SOA (SAC = .345; 95% CI = [.235, .455]). CIs for likely untenable (Hutchison, 2007). In fact, Hutchison (2007)
SAC did not overlap with 0, indicating that stereotypes were found that the use of strategic response strategies could be
active at both SOAs. seen at SOAs as brief as 250-ms. In addition, controlled
356 Personality and Social Psychology Bulletin 46(3)

faking of responses to evaluative priming tasks is possible, 1

even with a relatively brief SOA (i.e., 280-ms) and brief 0.9
D
G
response windows (i.e., 600-ms; Teige-Mocigemba & 0.8 SAC
Klauer, 2013; also see Teige-Mocigemba & Klauer, 2008). In 0.7 SAP

Experiment 1, we showed that stereotyping processes were

Probability Estimate
0.6
diminished with a 350-ms SOA compared with a 150-ms 0.5
SOA. To further examine the efficiency profile of stereotype 0.4
control processes, in Experiment 2, we tested whether greater 0.3
control would be observed with an SOA of 200-ms versus an 0.2
SOA of 150-ms. Although we did not have strong a priori 0.1
predictions for the 200-ms condition, we felt such a condi- 0
tion would be informative. If stereotypic bias is reduced by 150-ms 200-ms 325-ms
Stimulus Onset Asynchrony
an additional 50-ms (from 150-ms to 200-ms SOA), this
would provide evidence that control is possible even at 200-
ms, well below Neely’s (1977) initial boundary. Figure 2. Probability estimates of the parameters SAC
(stereotype activation), SAP (stereotype application), D (target
To accomplish these goals, we modified Experiment 1’s
detection), and G (guessing) by SOA level (150 vs. 200 vs. 325
procedure by adding a third 200-ms SOA level. This addi- ms) in Experiment 2.
tional 200-ms SOA level exactly corresponds to conditions Note. Error bars represent 95% confidence intervals. SAC = stereotype
used in the SMT’s initial validation (see Krieglmeyer & activation; SAP = stereotype application; SOA = stimulus onset
Sherman, 2012). Thus, each of the 144 trials was assigned to asynchrony.
150-ms, 200-ms, or 325-ms SOA.
stronger at 150-ms SOA than 325-ms SOA, F(1, 69) =
Participants 11.16, p = .001, gav = .291; 95% CIdifference = [.038, .150].
However, the SMT effect did not statistically differ between
Seventy-five undergraduate students at the University of 150-ms and 200-ms SOA, F(1, 69) = 2.52, p = .117, gav =
California, Davis participated in Experiment 2 for partial .125; 95% CIdifference = [–.011, .095], or between 200-ms and
course credit (82.2% Female, Mage = 19.4 years; 52% 325-ms SOA, F(1, 69) = 2.18, p = .144, gav = .150; 95%
Asian, 26% Caucasian, 19% Latino/a, 3% Black). Five par- CIdifference = [–.018, .122].
ticipants were excluded from analyses according to our a
priori standards. Including all data in analyses does not Multinomial modeling analyses. When fit to the data, the
change the direction or statistical conclusions of the model G2 statistic suggested that the fit of the SMT model
reported results. Based on effect size estimates from was acceptable, and that the magnitude of misfit controlling
Experiment 1 (dz = .430), 45 participants were necessary to for power was small, G2(6) = 10.519, p = .104, w = .032.
detect a similar effect at 1 − β = .80 power. However, we The effect of SOA on the SAP parameter replicated,
expected that adding an additional within-subjects level of ΔG2(2) = 17.431, p < .001. Active stereotypes were mar-
the SOA manipulation would reduced the number of data ginally more likely to be applied at 150-ms SOA than 200-
points informing estimate at each level of SOA. Given this ms SOA, ΔG2(1) = 3.777, p = .052, w = .019, and reliably
additional uncertainty, we sought to obtain a similar num- more likely to be applied 325-ms SOA, ΔG2(1) = 17.348, p
ber of participants to Experiment 1. Using G*Power and < .001, w = .040 (see Figure 2). Furthermore, the SAP
multiTree power analyses, we estimated that the final sam- estimate at 200-ms SOA was higher than at 325-ms, ΔG2(1)
ple of 70 provided power greater than 1 − β = .95 to detect = 4.949, p = .026, w = .022. Thus, participants were less
effects of the 150-ms versus 325-ms SOA manipulation on likely to apply activated stereotypes at longer versus shorter
the SMT effect, SAP, and SAC at levels similar to SOA.
Experiment 1 (Faul et al., 2009; Moshagen, 2010). To con- The impact of SOA on the SAC parameter also replicated
sider experimental power comparing 150-ms to 200-ms, we the results from Experiment 1. SAC parameters could not be
conducted sensitivity analyses that showed a sample of 70 constrained across the three levels, ΔG2(2) = 6.823, p =
would provide .80 power to detect an effect size as small as .033, w = .025. The SAC estimate at 150-ms SOA did not
dz = .300. detectably differ from the estimate at 200-ms, ΔG2(1) =
1.680, p = .195, w = .012, but was reliably lower than the
estimate at 325-ms, ΔG2(1) = 6.700, p = .010, w = .025.
Results SAC estimates at 200-ms SOA and 325-ms SOA did not dif-
SMT effects. Repeated-measures ANOVA indicated that the fer, ΔG2(1) = 1.641, p = .200, w = .013. Replicating the
SMT effect again differed across SOA, F(1.81, 138)8 = exploratory SAC result from Experiment 1 increases our
4.868, p = .011, ωp = .052.9 Replicating Experiment 1,
2
confidence that stereotype activation is higher at longer ver-
simple comparisons revealed that the SMT effect was sus shorter SOA.
Rivers et al. 357

Discussion Participants
Experiment 2 provided additional support for the hypothesis Forty-nine undergraduate students at the University of
that increasing the time between racial primes and target California, Davis participated in Experiment 3 for partial course
images mitigates the magnitude of racial bias in judgment. credit (84.4% Female, Mage = 20.4 years; 49% Asian, 18%
Furthermore, stereotype application (SAP) was lower with Caucasian, 25% Latino/a, 6% Black). We sought to collect a
the 325-ms SOA than the 150-ms SOA. As in Experiment 1, sample of at least 48 participants according to our preregistered
differences in stereotype application rather than activation plan on the Open Science Framework11 to achieve .90 power to
corresponded with the lower levels of racial bias observed at detect an effect of dz = .430. Seven participants were excluded
longer SOA. In fact, when stereotype activation was highest, according to our preregistered criteria, resulting in a final sam-
judgments showed the least amount of racial bias. ple of 42. Including all data in analyses does not change the
Although the comparison between 150-ms and 200-ms direction or statistical conclusions of the reported results.
was not significant for the SMT effect or SAC, a general
trend of decreased bias and SAC was observed. Moreover,
Design
SAP was marginally lower at 200-ms SOA versus 150-ms
SOA. Thus, it appears that stereotype correction is possible, The design of Experiment 3 was similar to Experiment 1,
even at 200-ms.10 with one modification: The SMT consisted of a single block
The increase in stereotype activation at longer SOA of 96 trials in which participants encountered 16 prime
appears to be at odds with aspects of existing theory and images of each type (Black, White, neutral). This modifica-
research. There are several reasons to expect that stereotype tion permitted us to retain 12 unseen prime images that were
activation should decrease as SOA increases. First, priming used as lures on the recognition test (see below). Each of the
research suggests that the activation of concepts fades over primes was randomly chosen to be presented either at the
time (e.g., Kunda et al., 2002). Second, previous research 150-ms or the 325-ms SOA, and was presented once with a
suggests that stereotypes are relied upon to a greater extent high threat target and once with a low threat target.
when controlled processing is constrained because they are
needed as efficient social judgment tools/heuristics (Macrae, Prime-recognition task. Directly after completing the SMT
Milne, & Bodenhausen, 1994). As such, when processing procedure, participants completed the recognition memory
resources, such as time, are restricted, stereotype activation measure. Twenty-four previously shown “old” prime images
should increase to promote efficient decision making. and eight “new” lure images were presented in a random
Finally, other work suggests that stereotype activation itself order, and participants were asked to judge whether or not
can be inhibited. If this process is time-dependent, then ste- each image had appeared in the previous task.
reotype activation would be reduced at longer SOA.
However, the data did not corroborate any of these accounts.
Results
Instead, stereotype activation increased as SOA increased.
Experiment 3 directly tested one plausible reason why ste- SMT effects. Surprisingly, a paired-samples t test indicated
reotype activation might increase—increased processing of that the SMT effect did not differ across SOA, t(41) = −.671,
priming imagery. p = .506, gav = −.059, 95% CIdifference = [–.084, .042].12
Moreover, we did not observe significant racial bias at either
SOA level, ps > .7. We reflect on this further in the discus-
Experiment 3 sion below.
Rationale
Multinomial modeling analyses. The fit of the SMT model
Experiments 1 and 2 found and replicated the finding that appeared acceptable and the magnitude of misfit controlling
stereotype activation is higher as SOA increases. for power was small, G2(4) = 2.133, p = .711, w = .023.
Experiments 3 and 4 sought to further characterize the Importantly, longer SOA again led to an increase in
nature of this increase. One possibility is that participants SAC, ΔG2(2) = 6.828, p = .030, w = .041. In contrast,
might continually process the prime images during the lon- SOA had no detectable effect on SAP, ΔG2(2) = .246, p =
ger interstimulus interval. Greater time and effort spent pro- .620, w = .008.
cessing the primes would increase stereotype activation. In
Experiment 3, participants completed a simple recognition Recognition memory. We computed an index of recognition
memory test to determine whether increases in stereotype accuracy for each participant by first calculating the proportion
activation corresponded with better memory for prime of prime images correctly identified as old (Hits) and sub-
images. If primes receive additional processing at longer tracted from that the proportion of lures incorrectly identified
SOA, then they should be recognized at higher rates in the as old (False Alarms). Accuracy was above chance perfor-
memory test (Bower & Karlin, 1974). mance at both 150-ms and 325-ms SOA (ps < .001). Critically,
358 Personality and Social Psychology Bulletin 46(3)

and consistent with our preregistered predictions, recognition Female, Mage = 20.6 years, 59% Asian, 19% Caucasian, 21%
memory was higher at 325-ms than at 150-ms SOA, t(41) = Latino/a, 2% Black). We sought a sample of at least 52 to set
7.717, p < .001; gav = .886, 95% CIdifference = [.146, .250]. power at .80 to detect an effect of dz = .4. According to our a
priori criteria, two participants were excluded from analyses.
Discussion Including all data in analyses does not change the direction
or statistical conclusions of the reported results.
Experiment 3 again found evidence that increases in SOA led
to increases in stereotype activation. Consistent with the
hypothesis that primes received additional processing when Procedure
SOA was long, recognition memory was better for primes Experiment 4’s procedure was similar to that of Experiment
when SOA was 325-ms versus 150-ms. Despite high statisti- 1, with the following modifications. First, on half of the SMT
cal power for the fully within-subject design, the impact of trials we backward masked prime stimuli for 25-ms. The
SOA on the SMT effect and on SAP did not replicate the mask was sized to the same dimensions as prime stimuli and
relationships found in Experiments 1 and 2. Neither did this consisted of a visual black and white static pattern. An equiv-
study find evidence of racial bias in people’s judgments. alent 25-ms of blank screen appeared on trials on which
These results were unexpected given the robust stereotypic prime stimuli were not masked. Thus, the shorter SOA was
biases and effects of SOA observed in Experiments 1, 2, and 175-ms (150-ms prime presentation plus either 25-ms mask
4. It is unclear whether the absence of an effect is a statistical or 25-ms blank screen) and the longer SOA was 350-ms
anomaly or due to an unidentified moderating variable.13 To (150-ms prime presentation plus either 25-ms mask or 25-ms
be clear, there were no changes in the paradigm that we blank screen plus 175-ms blank screen). The mask factor was
expected to change the otherwise robust effect of SOA. manipulated orthogonally, resulting in a 3 (prime type: Black
vs. White vs. neutral) × 2 (target type: high vs. low) × 2
Experiment 4 (SOA: 175- vs. 350-ms) × 2 (prime mask: masked vs.
unmasked) fully within-subjects design. As the SOA manip-
Rationale ulation selectively influenced only the SAC and SAP model
parameters, we sought to increase the precision of estimates
Experiment 3 demonstrated that longer relative to shorter
for these parameters by increasing the proportion of trials
SOA produced greater stereotype activation and led to better
with race primes relative to trials with the neutral prime.
memory for prime images. Both of these results are consistent
Experiment 4 included 64 Black, 64 White, and 32 neutral
with the idea that participants were continuing to process the
prime trials for a total of 160 trials per participant.
primes during the interstimulus interval. Experiment 4 sought
to test whether this additional processing was perceptual or
conceptual in nature. It is possible that iconic memory Results
increases the length of time that prime images are available in SMT effects. SMT effect estimates were compiled and
visual working memory on trials with longer versus shorter entered into a 2 (SOA: 175- vs. 350-ms) by 2 (mask: masked
SOA. Each of the previous experiments presented prime primes vs. unmasked primes) repeated-measures ANOVA
images for 150-ms each, but Sperling’s (1960) partial report model. Replicating Experiments 1 and 2, there was a main
paradigm demonstrates that iconic representations can persist effect of SOA on the SMT effect, F(1, 55) = 40.113,
after offset of visual stimuli for up to 1,000-ms in visual p < .001, ω2p = .407 (see Figure 3).14 SMT effects were
working memory. An iconic memory interpretation would stronger at the shorter 175-ms SOA than the longer 350-ms
suggest that stereotype activation was highest at longer SOA SOA, t(55) = 6.333, p <.001, gav = .468. Although there
levels because prime images were accessible in visual work- was no main effect of the mask manipulation, F(1, 55) =
ing memory for a longer period of time. Orthogonally manip- 0.550, p = .461, ω2p < .001, an interaction between SOA
ulating the presence of a backward-visual mask while holding and mask emerged, F(1, 55) = 7.565, p = .008, ω2p = .103.
SOA constant is a straightforward test of this possible expla- When primes were not masked, there was a strong effect of
nation for increasing stereotype activation. An iconic repre- SOA on the SMT effect, t(55) = 5.407, p < .001, gav =
sentation explanation predicts that stereotype activation will .622. When primes were masked, the effect of SOA on the
increase at higher SOA only when primes are not masked. In SMT effect was still significant, but marketedly smaller in
contrast, if increased processing of the conceptual meaning magnitude, t(55) = 2.462, p = .017, gav = .221. Thus, mask-
and associations of the primes is responsible, then the increase ing primes dampened the influence of SOA on the SMT
in activation should be observed regardless of masking. effect. We then examined simple effects looking within
each SOA level. In contrast to the predictions from the
Participants iconic memory account, the effect of masking primes
changed SMT effects at short SOA, t(55) = 2.659, p = .010,
Fifty-eight undergraduate students at the University of gav = .265, and had no detectable impact at long SOA,
California, Davis participated for partial course credit (70.7% t(55) = −1.545, p = .128, gav = −.169.
Rivers et al. 359

Figure 3. SMT effect by SOA (175- and 350-ms) and backward mask (present and absent) in Experiment 4.
Note. Error bars represent 95% confidence intervals. SMT = Stereotype Misperception Task; SOA = stimulus onset asynchrony.

Multinomial modeling analyses. For each SMT model param- Recognition memory. Recognition accuracy was indexed
eter, we generated two models—one that permitted the free using the hits minus false alarms index used in Experiment
interaction of the SOA and mask factors and one that allowed 3.15 We entered this into a repeated-measures ANOVA test-
only main effects. The restricted main effects model was ing the effects of SOA and backward mask. Replicating
then compared against the interaction model, with significant Experiment 3, recognition for prime images was better at
ΔG2 indicating the presence of an interaction between the 200-ms SOA compared with 25-ms SOA, F(1, 55) =
two factors. The SMT model provided a good approximation 129.155, p < .001, ω2p = .692. Recognition was directionally
to the data, G2(8) = 5.559, p = .696, w = .026. reduced when primes were backward masked, but this effect
Once again, increasing SOA from 25-ms to 200-ms led to did not approach significance, F(1, 55) = 1.552, p = .218,
higher SAC, ΔG2(1) = 9.316, p = .002, w = .032 (see Table ω2p = .010. There also was no interaction between SOA and
3). As a main effect, masking prime stimuli produced a small the backward mask, F(1, 55) = 1.450, p = .234, ω2p = .008.
but detectable increase in SAC versus not masking primes,
ΔG2(1) = 4.260, p = .039, w = .019. This result is the oppo-
Discussion
site of what would be predicted by the iconic memory
account. The critical prediction from the iconic memory The primary goal of Experiment 4 was to test whether
interpretation is an interaction between SOA and mask, such increases in SAC at longer SOA were the result of additional
that masking primes reduces SAC, but only when SOA is perceptual processing of prime images in iconic memory. If
long. There was no evidence of the critical SOA × Mask true, masking prime stimuli would interact with the SOA fac-
interaction, ΔG2(1) = 1.274, p = .259. tor. We found no evidence for this critical interaction. Instead,
Replicating Experiments 1 and 2, SAP was higher when we found that backward masking prime stimuli, if anything,
SOA was shorter versus longer, ΔG2(1) = 55.251, p < appeared to modestly increase stereotype activation. In con-
.001, w = .078. Masking prime stimuli decreased SAP ver- cert with the recognition memory results from Experiment 3,
sus not masking primes, ΔG2(1) = 35.506, p < .001, w = this pattern of results suggests that prime images receive
.061. There was an interaction between SOA and mask on greater conceptual, rather than perceptual, processing when
the SAP parameter, ΔG2(1) = 11.010, p < .001. Simple SOA is longer.
comparisons revealed that SOA had a strong effect on SAP
when primes were not masked, ΔG2(1) = 21.049, p <
Within-Paper Meta-Analysis
.001, w = .049. SOA had a significant but attenuated effect
on SAP when primes were masked, ΔG2(1) = 9.721, p = We sought to quantify the effect of SOA on our three vari-
.002, w = .033. ables of interest across the four reported experiments. To do
360 Personality and Social Psychology Bulletin 46(3)

Figure 4. Forest plot depicting effect of SOA on SMT effect by experiment.

Note. Error bars represent 95% confidence intervals. SMT = Stereotype Misperception Task; SOA = stimulus onset asynchrony.

Figure 5. Forest plot depicting effect of SOA on SAC by experiment.

Note. Error bars represent 95% confidence intervals. SOA = stimulus onset asynchrony; SAC = stereotype activation.

this, we conducted mixed-effects meta-analytic tests using should reject the null hypothesis that the present experi-
the “metafor” package in the open source R platform (R ments were examining the same effect of SOA on the SMT
Development Core Team, 2010; Viechtbauer, 2010). We effect (Higgins, Thompson, Deeks, & Altman, 2003). In
aggregated data from short (150- to 175-ms) and long (325- other words, detecting heterogeneity implies that the effect
to 350-ms) SOA conditions and converted each effect size of SOA on the SMT effect was moderated by a third vari-
statistic to the r correlation coefficient (Lakens, 2013). able (Figure 4).16
The meta-analytic estimate for the effect of SOA on the The effect of SOA on the SAC model parameter was sig-
SMT effect was significant, Z = 2.639, p = .008, r = .387; nificant, Z = 3.128, p = .002, r = .190; 95% CI = [.072,
95% CI = [.105, .612]. In addition, there was detectable .304]. There was no evidence for heterogeneity, Q(3) = .225,
heterogeneity, Q(3) = 20.294, p < .001, indicating that we p = .974 (Figure 5).
Rivers et al. 361

Figure 6. Forest plot depicting effect of SOA on SAP by experiment.

Note. Error bars represent 95% confidence intervals. SOA = stimulus onset asynchrony; SAP = stereotype application.

The effect of SOA on the SAP model parameter was sig- stereotypic biases in peoples’ judgments decreased. This
nificant, Z = 3.437, p < .001, r = .208; 95% CI = [.091, means that stereotypes were most active when stereotype-
.320]. There was no evidence for heterogeneity, Q(3) = congruent biases in judgment were weakest. The present
4.223, p = .238 (Figure 6). work demonstrates that stereotype activation might not nec-
Taken together, these within-paper meta-analyses indicate essarily always result in biased judgment, and that prevent-
that increasing SOA had a reliable impact on reducing both ing the application of activated stereotypes can be an effective
the SMT effect and stereotype application (SAP), as well as strategy to reduce bias.
on increasing stereotype activation (SAC).
Stereotyping: Automaticity, Control, and
General Discussion Operating Conditions
The present research sought to understand the roles of basic The present results shed light on the operating conditions of
mechanisms thought to underlie stereotypic biases in social the basic mechanisms of stereotype activation and applica-
judgment—stereotype activation and stereotype application. tion. We tested our hypotheses using an indirect measure of
We manipulated the amount of time separating stereotypic stereotyping that reflects unintentional racial biases and
primes and social targets, finding that increases in time manipulated SOA at levels assumed to preclude deliberative
reduced the magnitude of stereotypic bias. Two mechanistic processing. Under these conditions, we found evidence for
accounts could have explained this reduction in bias. First, the modulation of both stereotype activation and applica-
reductions in stereotyping could have corresponded with tion, indicating that these processes are more dynamic over
decays or suppression of stereotype activation. Second, even short periods of time than previously thought (but see
reductions in stereotyping could have corresponded instead Cunningham, Zelazo, Packer, & Van Bavel, 2007). In terms
with reductions in stereotype application. We found strong of stereotype application, this finding is consistent with both
evidence for the second possibility. When the two basic ste- the Compensatory Automaticity model of stereotyping
reotyping mechanisms were independently estimated with (Glaser & Kihlstrom, 2006), which proposes that stereo-
the SMT processing tree model, shifts in stereotype applica- type-corrective processes themselves can proceed rapidly
tion corresponded with stereotypic biases in peoples’ social (i.e., at SOAs assumed to preclude controlled processes;
judgments. In contrast, stereotype activation appeared to Neely, 1977), and the suggestion from the Quadruple pro-
play a subordinate role to stereotype application. In fact, ste- cess model of implicit social cognition (Sherman et al.,
reotype activation consistently increased with time, even as 2008) that associative biases can be overcome relatively
362 Personality and Social Psychology Bulletin 46(3)

quickly and efficiently (Calanchini & Sherman, 2013). At of these basic mechanisms will provide the basis for
the same time, we found that even short increases of time ­developing effective interventions. For example, consider
between racial primes and target images affected the likeli- accounts of stereotyping that suggest that stereotype activa-
hood of applying stereotypic information in judgments. tion is the primary driver of biased judgments (e.g., ‘cogni-
Participants were most likely to apply active stereotypes tive monster,’ Bargh, 1999). In these accounts, when
when brief periods of time separated prime and target stereotypic knowledge is accessible, judgments are influ-
images. Experiment 2 found that stereotype application was enced in an assimilative fashion, inevitably producing down-
diminished at each stepwise increase in time separating stream bias, particularly under conditions thought to interfere
racial primes and target images. Thus, stereotype applica- with control (e.g., lack of time or cognitive resources;
tion and control over it, though relatively quick, is not an Devine, 1989; Fazio et al., 1995). Indeed, consistent with
entirely automatic process. this proposal, we found that people generally tended to apply
Interestingly and unexpectedly, we also found that stereo- active stereotypes across each of the experiments.
type activation was greater when there was more time However, challenging the activation-dominant perspective,
between primes and targets. We initially hypothesized that the extent of stereotype activation could not account for the
additional time would reduce rather than increase stereotype magnitude of peoples’ biases in the current research. The mag-
activation due to enhanced suppression or passive decay. nitude of racial bias in judgments and stereotype activation
However, there was no evidence in our experiments support- were dissociated. This is difficult to reconcile with theorizing
ing the hypothesis that stereotype activation decayed or was that posits a direct and inevitable pathway from stereotype
suppressed as time increased. It is possible that our activa- activation to stereotypic bias (e.g., Bargh, 1999). Even when
tion results reflect the fact that stereotype activation is not stereotype activation was strong, participants were able to pre-
entirely efficient, and requires time to unfold (Gilbert & vent the application of stereotypes under suboptimal condi-
Hixon, 1991). We think that it is more likely that these results tions that are presumed to interfere with strategic processes.
reflect the fact that the extent of even highly efficient pro- This does not necessarily mean that interventions aimed at
cesses can be increased by additional processing. Indeed, we reducing stereotype activation via suppression or other means,
found evidence from a measure of recognition accuracy that or by changing stereotypic knowledge outright will be unsuc-
primes were attended to and processed more thoroughly at cessful. There is no question that such reductions in activation
longer versus shorter SOAs. Evidence from the backward can reduce bias. At the very least, reducing activation bypasses
masking manipulation suggested that this additional process- the need to inhibit the application of stereotypes. Nevertheless,
ing is conceptual, rather than perceptual, in nature. the present results suggest that the relationship between ste-
An important caveat is that our conclusions about the reotype activation and stereotyping is more nuanced than is
operating conditions of stereotype activation and application sometimes described. Even under conditions that make con-
were based on a specific manipulation of processing trol difficult, the present results showed that stereotyping was
resources: time. Other manipulations, such as a cognitive more dependent on the extent of stereotype application than
load (Gilbert & Hixon, 1991), may yield different conclu- activation. Practically, this suggests that interventions to
sions. An alternative explanation regarding stereotype acti- reduce bias should emphasize the role of practice in correcting
vation draws on the “goal looms larger” effect (e.g., Goschke for the influence of active stereotypes (e.g., Calanchini,
& Kuhl, 1993). In contrast to semantic priming where we Gonsalkorale, Sherman, & Klauer, 2013; Kawakami, Dovidio,
might expect rapid decay in concept activation, goal priming Moll, Hermsen, & Russin, 2000).
has been shown to lead to increased accessibility over time
until the goal is fulfilled. Based on research on intergroup Declaration of Conflicting Interests
emotions (e.g., Cottrell & Neuberg, 2005; Mackie & Smith, The author(s) declared no potential conflicts of interest with respect
2002), it seems likely that our prime pictures may have acti- to the research, authorship, and/or publication of this article.
vated an emotional-motivational state of threat, fear, and cor-
responding goals for protection. As a consequence, the Funding
accessibility of threat representations may have increased
The author(s) disclosed receipt of the following financial support
over time because they were relevant to goals for protection. for the research, authorship, and/or publication of this article:
Further research is needed to further delineate the conditions Preparation of this article was partially supported by an Anneliese
under which our observations hold. Maier Research Award from the Alexander von Humboldt
Foundation to J.W.S. and K.C.K.
Self-Regulating Implicit Biases: Cognitive
Notes
Mechanisms
1. Data from an additional 12 participants were collected because
Results from this study underscore the importance of data collection proceeded more quickly than expected.
researching the basic mechanisms underlying the ability to Analyses were only performed on the full dataset and at no
mitigate stereotypic biases. A comprehensive understanding prior time. Sensitivity power analysis indicated that the final
Rivers et al. 363

sample of 90 participants allowed us to detect an effect size of moderation by SOA level). There is also a significant three-way
dz = .299 at 1 − β = .80. interaction between prime, SOA, and mask that corresponds to
2. We set two a priori criteria for exclusion of data for each exper- the interaction between SOA and mask reported here.
iment. First, participants who used a single key for every trial 15. For full transparency, we wish to make clear that the recogni-
were excluded (e.g., Krieglmeyer & Sherman, 2012; Payne, tion memory measure was not preregistered as in Experiment
Cheng, Govorun, & Stewart, 2005). Second, we excluded par- 3. Nevertheless, our predictions were the same; longer SOA
ticipants whose proportion of “more threatening” responses should be expected to correspond with higher SAC (stereotype
fell 2.5 standard deviations outside the sample distribution (see activation) as well as higher recognition accuracy.
Krieglmeyer & Sherman, 2012). 16. We conducted follow-up meta-analytic tests to determine if any
3. Readers can view and download all target images used in this variables statistically accounted for heterogeneity of SOA on
study at: osf.io/pqbhf/ the SMT effect. No clear conclusions emerged from these tests
4. See Supplemental Appendix Table A.2 for all demographic and that could account for the heterogeneity observed.
exploratory measures by experiment.
5. ANOVA analyses are detailed in the supplemental appendix.
The pattern of ANOVA results (i.e., significant racial bias and
Supplemental Material
decreased bias at longer stimulus onset asynchrony [SOA]) is Supplemental material is available online with this article.
not changed when target type is included as a factor and neutral
is added as a third level within the prime type factor. As stated
earlier, the target factor and neutral prime type are included
References
in the experimental design because they are necessary for the Bargh, J. A. (1999). The cognitive monster. In S. Chaiken & Y.
Stereotype Misperception Task (SMT) multinomial model Trope (Eds.), Dual-process theories in social psychology
(also see Krieglmeyer & Sherman, 2012). (pp. 361-382). New York, NY: Guilford Press.
6. In addition, we compared the performance of the SMT model Bower, G. H., & Karlin, M. B. (1974). Depth of processing pictures
with other frequently used process dissociation models (see of faces and recognition memory. Journal of Experimental
Supplemental Appendix Table A.1). The SMT model unambig- Psychology, 103, 751-757.
uously outperformed competitor models on Akaike information Calanchini, J., Gonsalkorale, K., Sherman, J. W., & Klauer, K.
criterion (AIC)/Bayesian information criterion (BIC)/minimum C. (2013). Counter-prejudicial training reduces activation
description length (MDL) criteria in all experiments, with the of biased associations and enhances response monitoring.
exception that the Affect Misattribution Procedure (AMP) European Journal of Social Psychology, 43, 321-325.
model maximized information on AIC criterion for Experiment Calanchini, J., & Sherman, J. W. (2013). Implicit attitudes reflect
4 only. associative, non-associative, and non-attitudinal processes.
7. The detection and guessing parameters were not meaningfully Social and Personality Psychology Compass, 7, 654-667.
impacted by experimental manipulations in any of the present Cottrell, C. A., & Neuberg, S. L. (2005). Different emotional
experiments. For simplicity, these parameters are not discussed reactions to different groups: A sociofunctional threat-based
further. approach to “prejudice.” Journal of Personality and Social
8. Huynh–Feldt correction applied for violation of sphericity. Psychology, 88, 770-789.
9. As in Experiment 1, including the target type factor and the Cressie, N., Pardo, L., & Pardo, M. C. (2003). Size and power con-
neutral level of the prime type factor does not change the pat- siderations for testing loglinear models using φ-divergence test
tern of results (i.e., significant racial bias and moderation by statistics. Statistica Sinica, 13, 555-570.
SOA level). Cunningham, W. A., Zelazo, P. D., Packer, D. J., & Van Bavel,
10. A sample of 101 participants would provide 80% power to J. J. (2007). The iterative reprocessing model: A multilevel
detect a similar impact of SOA (150 ms vs. 200 ms) on SAP framework for attitudes and evaluation. Social Cognition, 25,
(where w = .019). A sample of 173 would be required to pro- 736-760.
vide 80% power to detect a similar impact of SOA (150 ms vs. Devine, P. G. (1989). Stereotypes and prejudice: Their automatic
200 ms) on the SMT effect (where dz = .190). and controlled components. Journal of Personality and Social
11. Preregistration available at https://osf.io/2xws8/ Psychology, 56, 5-18.
12. Including the target type factor and neutral level within the Faul, F., Erdfelder, E., Buchner, A., & Lang, A. (2009). Statistical
prime type factor does not change this pattern of results. power analysis using G*Power 3.1: Tests for correlation and
13. One highly speculative possibility, which we do not have regression analyses. Behavior Research Methods, 41, 1149-
adequate data to test, is that the heavily publicized shooting 1160.
of concertgoers in Las Vegas occurred the week prior to data Fazio, R. H. (1990). Multiple processes by which attitudes guide
collection. Having such a salient exemplar of a White male behavior: The MODE model as an integrative framework.
committing a violent act could increase the accessibility of asso- Advances in Experimental Social Psychology, 23, 75-109.
ciations between White males and threat, eliminating the other- Fazio, R. H., Jackson, J. R., Dunton, B. C., & Williams, C. J. (1995).
wise robust stereotypic biases we observed in each of our other Variability in automatic activation as an unobtrusive measure
experiments. A priori, we expected Experiment 3 to produce the of racial attitudes: A bona fide pipeline? Journal of Personality
same effects of SOA as observed in Experiments 1, 2, and 4. and Social Psychology, 69, 1013-1027.
14. As in each of the previous Experiments, including the target Gilbert, D. T., & Hixon, J. G. (1991). The trouble of thinking:
type factor and the neutral level of the prime type factor does Activation and application of stereotypic beliefs. Journal of
not change the pattern of results (i.e., significant racial bias and Personality and Social Psychology, 60, 509-517.
364 Personality and Social Psychology Bulletin 46(3)

Glaser, J., & Kihlstrom, J. F. (2006). Compensatory automaticity: Moshagen, M. (2010). multiTree: A computer program for the anal-
Unconscious volition is not an oxymoron. In R. R. Hassin, J. S. ysis of multinomial processing tree models. Behavior Research
Uleman, & J. A. Bargh (Eds.), The new unconscious (pp. 171- Methods, 42, 42-54.
195). New York, NY: Oxford University Press. Moskowitz, G. B., Gollwitzer, P. M., Wasel, W., & Schall, B.
Glaser, J., & Knowles, E. D. (2008). Implicit motivation to con- (1999). Preconscious control of stereotype activation through
trol prejudice. Journal of Experimental Social Psychology, 44, chronic egalitarian goals. Journal of Personality and Social
164-172. Psychology, 77, 167-184.
Goschke, T., & Kuhl, J. (1993). Representation of intentions: Moskowitz, G. B., & Li, P. (2011). Egalitarian goals trigger stereo-
Persisting activation in memory. Journal of Experimental type inhibition: A proactive form of stereotype control. Journal
Psychology: Learning, Memory, and Cognition, 19, 1211- of Experimental Social Psychology, 47, 103-116.
1226. Neely, J. (1977). Semantic priming and retrieval from lexical
Hamilton, D. L., & Sherman, J. W. (1994). Stereotypes. In R. S. memory: Roles of inhibitionless spreading activation and lim-
Wyer & T. K. Srull (Eds.), Handbook of social cognition (2nd ited-capacity attention. Journal of Experimental Psychology:
ed., Vol. 2, pp. 1-68). Hillsdale, NJ: Lawrence Erlbaum. General, 106, 226-254.
Higgins, J. P. T., Thompson, S. G., Deeks, J. J., & Altman, D. G. Oosterhof, N. N., & Todorov, A. (2008). The functional basis
(2003). Measuring inconsistency in meta-analyses. British of face evaluation. Proceedings of the National Academy of
Medical Journal, 327, 557-560. Sciences, 105, 11087-11092.
Hutchison, K. A. (2007). Attentional control and the relatedness Payne, B. K., Cheng, C. M., Govorun, O., & Stewart, B. D. (2005).
proportion effect in semantic priming. Journal of Experimental An inkblot for attitudes: Affect misattribution as implicit mea-
Psychology: Learning, Memory, and Cognition, 33, 645-662. surement. Journal of Personality and Social Psychology, 89,
Kawakami, K., Dovidio, J. F., & Dijksterhuis, A. (2003). Effect 277-293.
of social category priming on personal attitudes. Psychological Payne, B. K., Hall, D. L., Cameron, C. D., & Bishara, A. J. (2010).
Science, 14, 315-319. A process model of affect misattribution. Personality and
Kawakami, K., Dovidio, J. F., Moll, J., Hermsen, S., & Russin, A. Social Psychology Bulletin, 36, 1397-1408.
(2000). Just say no (to stereotyping): Effects of training in the Phills, C. E., Kawakami, K., Tabi, E., Nadolny, D., & Inzlicht, M.
negation of stereotypic associations on stereotype activation. (2011). Mind the gap: Increasing associations between the self
Journal of Personality and Social Psychology, 78, 871-888. and Blacks with approach behaviors. Journal of Personality
Krieglmeyer, R., & Sherman, J. W. (2012). Disentangling stereo- and Social Psychology, 100, 197-210.
type activation and stereotype application in the Stereotype R Development Core Team. (2010). R: A language and environ-
Misperception Task. Journal Personality and Social ment for statistical computing. Vienna, Austria: R Foundation
Psychology, 103(2), 205-224. for Statistical Computing.
Kunda, Z., Davis, P. G., Adams, B. D., & Spencer, S. J. (2002). Sherman, J. W. (2006). On building a better process model: It’s
The dynamic time course of stereotype activation: Activation, not only how many, but which ones and by which means?
dissipation, and resurrection. Journal of Personality and Social Psychological Inquiry, 17, 173-184.
Psychology, 82, 283-299. Sherman, J. W., Gawronski, B., Gonsalkorale, K., Hugenberg, K.,
Kunda, Z., & Spencer, S. J. (2003). When do stereotypes come to Allen, T. J., & Groom, C. J. (2008). The self-regulation of
mind and when to they color judgment? Psychological Bulletin, automatic associations and behavioral impulses. Psychological
129, 522-544. Review, 115, 314-335.
Lakens, D. (2013). Calculating and reporting effect sizes to facili- Sherman, J. W., Krieglmeyer, R., & Calanchini, J. (2014). Process
tate cumulative science: A practical primer for t-tests and models require process measures. In J. W. Sherman, B.
ANOVAs. Frontiers in Psychology, 4, Article 863. Gawronski, & Y. Trope (Eds.), Dual-process theories of the
Lepore, L., & Brown, R. (1997). Category and stereotype activa- social mind (pp. 121-138). New York, NY: Guilford Press.
tion: Is prejudice inevitable? Journal of Personality and Social Sherman, J. W., Macrae, C. N., & Bodenhausen, G. V. (2000).
Psychology, 72, 275-287. Attention and stereotyping: Cognitive constraints on the con-
Mackie, D. M., & Smith, E. R. (2002). Beyond prejudice: Moving struction of meaningful social impressions. European Review
from positive and negative evaluations to differentiated reac- of Social Psychology, 11, 145-175.
tions to social groups. New York, NY: Psychology Press. Sperling, G. (1960). The information available in brief visual pre-
Macrae, C. N., Milne, A. B., & Bodenhausen, G. (1994). Stereotypes sentations. Psychological Monographs: General and Applied,
as energy-saving devices: A peek inside the cognitive toolbox. 74(11), 1-29.
Journal of Personality and Social Psychology, 66, 37-47. Teige-Mocigemba, S., & Klauer, K. C. (2008). “Automatic”
Monteith, M. J., Ashburn-Nardo, L., Voils, C. I., & Czopp, A. M. evaluation? Strategic effects on affective priming. Journal of
(2002). Putting the brakes on prejudice: On the development Experimental Social Psychology, 44, 1414-1417.
and operation of cues for control. Journal of Personality and Teige-Mocigemba, S., & Klauer, K. C. (2013). On the controllabil-
Social Psychology, 83, 1029-1050. ity of evaluative-priming effects: Some limits that are none.
Monteith, M. J., Sherman, J. W., & Devine, P. G. (1998). Cognition and Emotion, 27, 632-657.
Suppression as a stereotype control strategy. Personality and Viechtbauer, W. (2010). Conducting meta-analyses in R with the
Social Psychology Review, 2, 63-82. metafor package. Journal of Statistical Software, 36(3), 1-48.