Elsevier

NeuroImage

Volume 43, Issue 4, December 2008, Pages 775-783
NeuroImage

The Quadruple Process model approach to examining the neural underpinnings of prejudice

https://doi.org/10.1016/j.neuroimage.2008.08.033Get rights and content

Abstract

In order to investigate the systems underlying the automatic and controlled processes that support social attitudes, we conducted an fMRI study that combined an implicit measure of race attitudes with the Quadruple Process model (Quad model). A number of previous neural investigations have adopted the Implicit Association Test (IAT) to examine the automatic processes that contribute to social attitudes. Application of the Quad model builds on this previous research by permitting measures of distinct automatic and controlled processes that contribute to performance on the IAT. The present research found that prejudiced attitudes of ingroup favoritism were associated with amygdala, medial and right lateral orbitofrontal cortex. In contrast, prejudiced attitudes of outgroup negativity were associated with caudate and left lateral orbitofrontal cortex. Frontal regions found in previous neural research on the IAT, such as anterior cingulate, dorsolateral prefrontal cortex and operculum were associated with detecting appropriate responses in situations in which they conflict with automatic associations. Insula activity was associated with attitudes towards ingroup and outgroup members, as well as detecting appropriate behavior.

Introduction

The most complete understanding of neural systems underlying social attitudes must include systems underlying both automatic and controlled processes. For example, in the case of prejudice, a person may have an automatic tendency to judge an outgroup member in a negative manner or an ingroup member in a positive manner but control the expression of those automatic tendencies for a variety of reasons (Monteith et al., 1998, Sherman, in press). In other words, automatic associations may not be expressed because of controlled processing. In order to address the problematic behavioral measurement of automatic aspects of attitudes, a small amount of neural research interested in the neural systems underlying prejudice has drawn on the Implicit Association Test (IAT: Greenwald et al., 1998) (Chee et al., 2000, Cunningham et al., 2004, Knutson et al., 2006, Knutson et al., 2007, Luo et al., 2006, Phelps et al., 2003, Phelps et al., 2000, Richeson et al., 2003). Although the IAT is one of the most popular behavioral measures of automatic (or implicit) attitudes, particularly for prejudice (Fazio and Olson, 2003), behavioral and modeling research suggests that IAT performance reflects both automatic and controlled processes (Conrey et al., 2005, Sherman et al., 2008). Previous fMRI studies involving the IAT are consistent with this view; significant activation has been found in regions associated with automatic processing, such as the amygdala (Cunningham et al., 2004; Phelps et al., (2000) although see Phelps et al., (2003) for evidence that amygdala is not critical for implicit racial bias) as well as controlled processing such as the dorsolateral prefrontal cortex (Richeson et al., 2003), middle frontal gyrus (Knutson et al., 2007), ventrolateral prefrontal cortex (Luo et al., 2006), and cingulate gyrus (Luo et al., 2006). The proposed research builds on this work by applying the Quadruple Process model (Quad model; Conrey et al., 2005, Sherman et al., 2008) to the interpretation of IAT performance in an fMRI environment, in order to examine the neural correlates of specific automatic and controlled processes that contribute to prejudice.

The IAT measures attitudes by examining differential abilities to associate targets and attributes (e.g., targets of race: Black or White; attributes: Pleasant or Unpleasant). Participants are presented with stimuli and must categorize them using response keys that are associated with both a target and an attribute. In the Congruent condition of the race version of the task, the response keys represent pairings of targets and attributes that reflect negatively biased associations towards Black targets (e.g., Black/Unpleasant for one response key; White/Pleasant for the other response). In the Incongruent condition of this task, response keys represent pairings reflecting negative associations towards White targets (e.g., Black/Pleasant for one response key, White/Unpleasant for the other response key). Implicit bias is indicated by significantly slower response times in the Incongruent condition compared to the Congruent condition. In other words, if participants are slower to categorize stimuli when target and attribute pairings reflect negative Black/positive White associations in comparison to pairings that reflect positive Black/negative White associations, then they are considered to have negative implicit attitudes towards Black targets and/or positive implicit attitudes towards White targets.

Although the IAT is often used to measure automatic aspects of bias, performance of the task also recruits a variety of controlled processes (Conrey et al., 2005, Sherman et al., 2008). Consider the Stroop Task (Stroop, 1935), which is highly similar in structure to the IAT. A young child who knows colors but does not know how to read and a fully literate adult may make an equally small number of errors on the task. However, very different processes are at work for the adult and the child. On incongruent trials (e.g., the word “Blue” written in red ink), the adult must overcome a habit to read the word in order to name the color of the ink correctly. In contrast, the child has no habit to overcome; s/he simply responds to the color of the ink.

The same principle applies to the IAT (and many other implicit measures of attitudes), which has a Stroop-like structure of Congruent (e.g., pairing Black faces with negative words and White faces with positive words) and Incongruent (e.g., pairing Black faces with positive words and White faces with negative words) trials. The same behavioral outcome may reflect very different underlying processes. Whereas some people may exhibit weak bias because they successfully overcome their strong automatic evaluative associations, others may exhibit weak bias because they do not hold biased attitudes. The measure itself cannot distinguish between the two cases.

The joint contribution of automatic and controlled processes to IAT performance has been reflected in fMRI studies of this topic (for lesion studies using the IAT see Milne and Grafman (2001); Phelps et al., 2003). Previous neural research that combines the IAT and fMRI usually takes one of two approaches. In one approach, individual differences in IAT performance from outside the scanner are examined in relation to neural activity from a separate task in the scanner. Studies using this approach have focused on racial bias and have found significant amygdala activity (Cunningham et al., 2004, Phelps et al., 2000) and dorsolateral prefrontal cortex activity (Richeson et al., 2003) when viewing unfamiliar Black faces in comparison to White faces. These studies suggest that racial bias predicts the engagement of neural regions associated with both automatic processing (e.g., amygdala) and controlled processing (e.g., dorsolateral prefrontal cortex) when viewing outgroup faces in relation to ingroup faces.

In a second approach, neural activity is examined while participants perform the IAT in the scanner (attitudes toward objects in the natural world: Chee et al., 2000; moral issues: Luo et al., 2006; political issues: Knutson et al., 2006; gender and race: Knutson et al., 2007). As mentioned above, the behavioral measure of bias in an IAT paradigm rests on the discrepancy between reactions in the Incongruent and Congruent conditions. Therefore, fMRI studies of IAT performance typically compare neural activity in the Incongruent condition to the Congruent condition. These studies have found neural regions associated with controlled processes to be more active in Incongruent than Congruent conditions (middle frontal gyrus: Knutson et al., 2007; ventrolateral prefrontal cortex and anterior cingulate: Luo et al., 2006; left inferior frontal gyrus: Knutson et al., 2006, Chee et al., 2000 did not design their study to permit a direct comparison between the Incongruent and Congruent conditions). However, although the discrepancy in reaction time between the Incongruent and Congruent conditions is the basis for behaviorally measuring implicit bias, it is unlikely that neural activity associated with this comparison measures automatic components of attitudes in an fMRI environment. From an fMRI perspective, the comparison of the Incongruent and Congruent conditions represents the difference between neural systems that are engaged for a condition that may involve response competition and a condition with less response competition (rather than the comparison of the presence versus absence of an automatic association). In other words, the main contrast used in behavioral research to measure automatic bias translates into an analysis of controlled processes when conducted in the fMRI environment.

Therefore, the neural systems that support automatic attitudes in fMRI studies are often estimated by conducting other analyses within the IAT task structure. For example, implicit moral attitudes were examined by investigating the neural systems activated in relation to arousing moral target stimuli compared to non-arousing moral stimuli. Although arousal is equated with automaticity, it is possible that greater control is engaged in relation to arousing stimuli and, therefore, any results may reflect a combination of automatic and controlled processing. This analysis found significant activity in regions more typically associated with automatic associations for the arousing moral stimuli (amygdala, ventromedial prefrontal cortex: Luo et al., 2006). Another study examined implicit preference for political figures by regressing individual differences in explicit ratings of preference for the political figures on the Congruent condition map (in relation to a control condition). Although preferences were inferred as implicit, they were measured using explicit ratings, raising the possibility that both automatic and controlled processing contribute to the results. This analysis found significant activity in regions associated with automatic associations and regions associated with controlled processing (e.g., left superior frontal gyrus (BA 10), medial frontal gyrus (BA 11), right precentral gyrus (BA 6) and middle frontal gyrus (BA 8): Knutson et al., 2007). These studies illustrate the difficulty in using the IAT as a measure of automatic associations in an fMRI environment. Although automatic associations may contribute to arousal or explicit preferences, these measures may also reflect controlled processes.

In summary, previous fMRI studies of social attitudes using the IAT have found significant activation in neural systems associated with automatic processing and neural systems associated with controlled processing. One way of building on this research is to take an approach that permits researchers to relate this neural activity to specific automatic and controlled psychological processes. The present research illustrates one way to achieve this approach by applying the Quad model to the analysis of an IAT performed inside the scanner.

The Quad model was developed by Conrey et al., 2005, Sherman et al., 2008) to measure the joint contribution of automatic and controlled processes to performance on implicit measures of cognition. The Quad model is a multinomial model (Batchelder and Riefer, 1999) that measures the independent influences of four qualitatively different processes on implicit task performance by estimating a parameter value for each: automatic activation of an association with the stimulus (AC), the ability to detect an appropriate response (D), the success at overcoming automatically activated biased associations (OB), and the influence of any response bias that may guide overt responses when other guides to response are absent (G). The Activation parameter (AC) refers to the degree to which biased associations are automatically activated when responding to a stimulus. All else being equal, the stronger the associations, the more likely they are to be activated and to influence behavior. The Detection parameter (D) reflects a relatively controlled process that detects appropriate and inappropriate responses. Sometimes, the activated associations conflict with the detected correct response. For example, on incompatible trials of the Stroop Task (Stroop, 1935) or incompatible trials of implicit attitude measures (e.g., pairing Black faces with positive words), automatic associations or habits conflict with detected correct responses. In such cases, the Quad model proposes that an Overcoming Automatically Activated Biased Associations process resolves the conflict. As such, the Overcoming Biased Associations parameter (OB) refers to self-regulatory efforts that prevent automatically activated associations from influencing behavior when they conflict with detected correct responses. Finally, the Guessing parameter (G) reflects general response tendencies that may occur when individuals have no associations that direct behavior, and they are unable to detect the appropriate response. Guessing can be random, but it may also reflect a systematic tendency to prefer a particular response. For example, incorrectly categorizing a target face stimulus as “unpleasant” in the IAT could be considered a socially undesirable response. To avoid that possibility, participants may adopt a conscious guessing strategy to respond with the positive rather than the negative key. Thus, guessing can be relatively automatic or controlled.

The structure of the Quad model is depicted as a processing tree in Fig. 1. In the tree, each path represents a likelihood. Processing parameters with lines leading to them are conditional upon all preceding parameters. For instance, Overcoming Biased Associations (OB) is conditional upon both Activation of Associations (AC) and Detection (D). If no automatic association exists, participants may still be able to detect an appropriate response (D) using information other than an automatic association, but OB cannot be calculated because there is no automatic association to overcome. Similarly, Guessing (G) is conditional upon the lack of Activation of Associations (1  AC) and the lack of Detection (1  D). The conditional relationships described by the model form a system of equations that predict the number of correct and incorrect responses in different conditions (e.g., compatible and incompatible trials). For example, a Black face stimulus in an incompatible block of a Black–White IAT will be assigned to the correct side of the screen with the probability: AC × D × OB + (1  AC) × D + (1  AC) × (1  D) × G. This equation sums the three possible paths by which a correct answer can be returned in this case. The first part of the equation, AC × D × OB, is the likelihood that the association is activated and that the correct answer can be detected and that the association is overcome in favor of the detected response. The second part of the equation, (1  AC) × D, is the likelihood that the association is not activated and that the correct response can be detected. Finally, (1  AC) × (1  D) × G, is the likelihood that the association is not activated and the correct answer cannot be detected and that the participant guesses by pressing the positive (“pleasant”) key. Because the “pleasant” and “Black” categories share the same response key in the incompatible block, pressing the positive key in response to a Black face stimulus will return the correct answer. The respective equations for each item category (e.g., Black faces, White faces, positive words, and negative words in both compatible and incompatible blocks) are then used to predict the observed proportion of errors in a given data set. The model's predictions are then compared to the actual data to determine the model's ability to account for the data. A χ2-estimate is computed for the difference between the predicted and observed errors. In order to best approximate the model to the data, the four parameter values are changed through maximum likelihood estimation until they produce a minimum possible value of the χ2. The final parameter values that result from this process are interpreted as relative levels of the four processes. For a complete description of data analysis within the Quad model, see Conrey et al., (2005).

To date, the Quad model has been applied to and has been shown to accurately predict behavior on a variety of priming tasks, including semantic priming tasks (Gawronski and Bodenhausen, 2005, Sherman et al., 2008) and the weapon identification task (Conrey et al., 2005, Payne, 2001, Sherman et al., 2008). The model also has been applied extensively to the IAT; (Conrey et al., 2005, Greenwald et al., 1998, Sherman et al., 2008) and the Go/No-Go Association Task (GNAT; Nosek and Banaji, 2001, Gonsalkorale et al., in press, Sherman et al., 2008).

The viability of the Quad model depends on four critical elements: model fit (i.e., does the model adequately approximate behavioral data?), stochastic validity of the parameters (i.e., can the model's parameters be influenced independently?), construct validity of the parameters (i.e., do the parameters signify the processes claimed by the model?), and predictive validity of the parameters (i.e., do the parameters predict meaningful behaviors?). The Quad model has succeeded on all fronts. As described above, the model has shown its ability to accurately predict performance on a variety of priming tasks, IATs, and the GNAT, demonstrating good model fit for these tasks (Conrey et al., 2005, Gonsalkorale et al., in press, Sherman et al., 2008).

The stochastic validity of the model has been established in a number of ways (Conrey et al., 2005, Sherman et al., 2008). For example, implementing a response deadline in an IAT designed to assess implicit attitudes about flowers and insects reduced Detection (D) and Overcoming Biased Associations (OB), but left Association Activation (AC) and Guessing (G) unaffected. Manipulating the base rate of left-hand versus right-hand responses in the same task affected Guessing (G), but none of the other three parameters (AC, D, OB). The expectation that one's performance on the weapon identification task would be observed by others decreased participants' ability to detect the appropriate response (D), but increased success at Overcoming Biased Associations (OB). These results indicate that the four parameters of the Quad model can vary independently, providing clear evidence for the stochastic validity of the model.

The construct validity of the model parameters also has been established by a number of findings (Conrey et al., 2005, Sherman et al., 2008). The fact that Detection (D) and Overcoming Biased Associations (OB) were reduced by a response deadline supports the claim that the two parameters reflect controlled processes that require cognitive capacity. In contrast, the finding that activation (AC) and Guessing (G) were unaffected by the response deadline is consistent with their depiction as relatively automatic processes that do not require significant cognitive capacity. The validity of OB as a measure of self-regulation was further established by demonstrations that it is impaired by alcohol consumption and decreases with age (Sherman et al., 2008). Extensive research has shown both alcohol use (e.g., Easdon and Vogel-Sprott, 2000) and aging (e.g., Hasher and Zacks, 1988) to be associated with impairments in self-regulation. The fact that altering the base rate of left-hand and right-hand responses influenced G corroborates the portrayal of that parameter as a general response bias.

Two studies provide evidence for the predictive validity of the parameters. First, estimates of individual subjects' Association Activation (AC) parameters derived from an evaluative IAT were positively correlated with association-related reaction time impairment in the same task (Conrey et al., 2005). Thus, the higher the AC, the greater the association-based impairment in performance. At the same time, estimates of subjects' Overcoming Biased Associations (OB) parameters were negatively correlated with association-based reaction time impairment. Thus, the higher the OB, the better able were participants to avoid association-based impairments in performance. These findings also bolster the construct validities of the AC and OB parameters.

In another study (Gonsalkorale et al., in press), non-Muslim Caucasian participants interacted with an experimental confederate who appeared to be and was described as Muslim. Following the interaction, the confederate rated how much he liked the participants, while the participants completed a GNAT measuring implicit bias toward Muslims. The confederate's ratings of how much he liked the participants were predicted by an interaction between the AC and OB parameter estimates taken from the GNAT. Specifically, when participants had low AC estimates of negative associations with Muslims, their level of OB was unrelated to how much they were liked by the confederate. In contrast, participants with high AC estimates of negative associations with Muslims were liked to the extent that they had high OB parameter estimates. Thus, the ability to overcome automatic negative associations on the GNAT predicted the quality of the social interaction when those associations were strong.

In sum, the Quad model has shown its ability to accurately describe behavior on a variety of evaluative priming tasks, semantic priming tasks, IATs, and the GNAT. In addition, the stochastic and construct validities of the model's parameters have been supported by numerous findings. Finally, the predictive validity of the AC and OB parameters has been demonstrated.

The present study uses functional magnetic resonance imaging (fMRI) to examine the neural correlates of prejudice by applying the Quad model to a race IAT performed in an MRI scanner. This approach examines neural activity that is directly related to performing the IAT and generates individual difference measures of automatic and controlled processes that can be used as regressors on this neural activity. Additionally, the Quad model permits measurement of the automatic and controlled processing at a level of specificity not possible in previous studies. The relative nature of the IAT measure (i.e., difference score) conceals the different contributions of ingroup favoritism and outgroup hostility to performance. Unlike previous research that has correlated the relative preference for Black and White (Good minus Bad), the Quad model provides separate estimates of positive ingroup associations and negative outgroup associations, which will refine our understanding of the psychological nature of the activations found in previous IAT research (e.g., amygdala, medial frontal lobe, insula: Cunningham et al., 2004, Knutson et al., 2007, Phelps et al., 2000). In other words, this study will be the first to disentangle two distinct processes that contribute to prejudice — ingroup favoritism (positive associations) and outgroup negativity (negative associations). Further, application of the Quad model will also refine our understanding of the frontal lobe activations found in previous research on the IAT (e.g., Chee et al., 2000, Luo et al., 2006, Knutson et al., 2006, Knutson et al., 2007, Richeson et al., 2003) by generating independent estimates of two distinct controlled processes (D, OB) that can be related to neural activity.

Section snippets

Participants

Sixteen right-handed Caucasian participants (8 female; M = 24.3 years, SD = 4.6 years) were recruited in compliance with the human-subjects regulations of the University of California, Davis, and were compensated with ⁎$10/h or course credits for their participation.

Behavioral paradigm

Participants completed a Black–White IAT designed to assess implicit preference for Whites over Blacks. Stimuli for the IAT consisted of 8 Pleasant and 8 Unpleasant pictures (from the International Affective Picture Set (IAPS); Lang et

Behavioral results

In the Quad model, behavioral data analysis focuses on error rates rather than reaction times. The 680 ms response deadline used to increase errors constrains reaction times and, therefore, the typical race bias effect is examined by contrasting error rates in the Incongruent condition with error rates in the Congruent condition. Participants made significantly more errors in the Incongruent than the Congruent condition (F(1,15) = 21.42, p < .05), replicating the typical race bias IAT effect. The

Discussion

The present research investigated the neural systems underlying prejudice by applying the Quad model (Conrey et al., 2005, Sherman et al., 2008) to analyze IAT performance (Greenwald et al., 1998) in an fMRI environment. The addition of the Quad model extended previous investigations of the neural systems underlying social attitudes by examining neural systems in relation to specific automatic and controlled processes that contribute to implicit bias. The present research found that insula

References (41)

  • BaxterM.G. et al.

    The amygdala and reward

    Nat. Rev., Neurosci

    (2002)
  • Brett, M., Anton, J.L., Valabregue, R., Poline, J.B., 2002. Region of interest analysis using an SPM toolbox. In Proc....
  • CheeM.W. et al.

    Dorsolateral prefrontal cortex and the implicit association of concepts and attributes

    NeuroReport

    (2000)
  • ConreyF.R. et al.

    Separating multiple processes in implicit social cognition: the quad model of implicit task performance

    J. Pers. Soc. Psychol

    (2005)
  • CunninghamW.A. et al.

    Separable neural components in the processing of black and white faces

    Psychol. Sci

    (2004)
  • CunninghamW.A. et al.

    Affective flexibility: evaluative processing goals shape amygdala activity

    Psychol. Sci.

    (2008)
  • EasdonC.M. et al.

    Alcohol and behavioral control: impaired response inhibition and flexibility in social drinkers

    Exp. Clin. Psychopharmacol.

    (2000)
  • FazioR.H. et al.

    Implicit measures in social cognition research: their meaning and use

    Annu. Rev. Psychol

    (2003)
  • GawronskiB. et al.

    Accessibility effects on implicit social cognition: the role of knowledge activation and retrieval experiences

    J. Pers. Soc. Psychol.

    (2005)
  • Gonsalkorale, K., von Hippel, W., and Sherman, J.W., in press. Bias and regulation of bias in intergroup interactions:...
  • Cited by (92)

    • Neural Basis of Prejudice and Prejudice Reduction

      2022, Biological Psychiatry: Cognitive Neuroscience and Neuroimaging
    • The neuroscience of intergroup threat and violence

      2021, Neuroscience and Biobehavioral Reviews
    • Ethnicity bias

      2021, Encyclopedia of Behavioral Neuroscience: Second Edition
    • Functional neuroanatomy of racial categorization from visual perception: A meta-analytic study

      2020, NeuroImage
      Citation Excerpt :

      The functional role of the first cluster seems related to bottom-up sensory processing of salient stimuli (Bagnis et al., 2019). In particular, the right INS contributes to subjective feelings and negative visceral reactions that are associated to negative attitudes toward racial out-groups and that are often experienced during prejudice responses (Beer et al., 2008; Harris and Fiske, 2006; Knutson et al., 2007; Lieberman et al., 2005; Liu et al., 2015; Richeson et al., 2003; Ronquillo et al., 2007). Likewise, the right IPL is involved in bottom-up attentional mechanisms in response to relevant stimuli.

    • Toward an integrative science of social vision in intergroup bias

      2019, Neuroscience and Biobehavioral Reviews
    • Brain structure links trait hostile attribution bias and attitudes toward violence

      2019, Neuropsychologia
      Citation Excerpt :

      Additionally, the high risk-seeking attitudes of binge drinkers increases activity in the lateral OFC, and DLPFC (Worbe et al., 2014). Prejudiced negative attitudes toward out-group members is related to increased activation in the lateral OFC, while positive attitudes toward in-group members is associated with the activation of the medial OFC (Beer et al., 2008). In light of these findings, several frontal areas (such as the OFC & DLPFC) may be integral to the determination of both attitudes toward violence and trait hostile attribution bias.

    View all citing articles on Scopus
    View full text