Science Briefs

Automatic and Controlled Components of Implicit Stereotyping and Prejudice

In our research, we have attempted to separate the automatic and controlled components of responses within individual implicit measures of stereotyping and prejudice.

By Jeff Sherman, PhD

People may be unaware of important underlying beliefs and attitudes that affect their behavior. Even when they are aware of these beliefs and attitudes, they may be reluctant to report them veridically. This is especially true in the domain of intergroup perception, where people may face strong social sanctions for expressing negative attitudes about social groups. These so-called "willing and able" problems are significant impediments to studying stereotyping and prejudice.

In recent years, an increasingly popular response to these problems has been the use of implicit measures of stereotyping and prejudice (for a review, see Fazio & Olson, 2003). These measures aim to circumvent the "willing and able" obstacles by measuring attitudes and beliefs without participants' awareness that they are being measured. Many proponents of these measures further argue that, even if made aware of the nature of the task, people are unable to control their responses. Thus, these measures are seen as reflecting the unintended, stimulus-driven, automatic activation of information in memory, whose expression largely cannot be altered or inhibited (e.g., Devine, 1989; Fazio, Jackson, Dunton, & Williams, 1995; Greenwald, McGhee, & Schwartz, 1998). When these measures are taken in combination with explicit measures (e.g., questionnaires), researchers aim to compare and contrast automatic and controlled facets of stereotyping and prejudice.

Implicit Measures are not Process-Pure
However, there are two significant drawbacks to this approach. First, it confounds the processes of interest (automatic vs. controlled) with the particular measurement tasks. Because the tasks may differ in a number of ways beyond the extent to which they tap automatic versus controlled processes, there is a danger of misinterpreting dissociations in task performance. For example, many observed dissociations between implicit and explicit memory tasks may be reinterpreted as dissociations between tasks that tap perceptual versus conceptual processes (e.g., Roediger, 1990).

A second drawback is that no task is process pure. Undoubtedly, implicit measures of stereotyping and prejudice are less susceptible to the influence of intention and controlled processes than are explicit measures. Nevertheless, any behavioral task that requires an observable response (e.g., a button press) likely involves an ongoing interplay between simultaneously occurring automatic and controlled processes. As such, the behavioral response, in and of itself, is incapable of specifying the nature of the underlying processes that produced the response.

Consider the Stroop Task, for example (Stroop, 1935). A fully literate adult and a young child who knows colors but does not know how to read 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 incompatible 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 implicit measures of stereotyping and prejudice, many of which have a Stroop-like structure of compatible (e.g., Black faces/negative words; White faces/positive words) and incompatible (e.g., Black faces/positive words; White faces/negative words) trials. The performance of two people who appear to have equally strong implicit biases may reflect very different underlying processes. Whereas one person may have strong implicit associations that are successfully overcome, the other may have weaker associations that are not overcome so well. Thus, behavioral outcomes on implicit measures of stereotyping and prejudice may not reflect differences in underlying attitudes, per se.

Separating Multiple Automatic and Controlled Components of Implicit Measures: The Quad Model
In our research (Conrey, Sherman, Gawronski, Hugenberg, & Groom, in press), we have attempted to separate the automatic and controlled components of responses within individual implicit measures of stereotyping and prejudice. In taking this approach, we avoid the task/process confound that is problematic for many investigations of automatic and controlled processes. This approach also allows us to examine the simultaneous operation and interaction of multiple processes in implicit task performance.

We base our analysis on the Process Dissociation Procedure (PD) pioneered by Jacoby and his colleagues (e.g., Jacoby, 1991) to separate different processing components within a single task. However, our research extends the basic PD model in important ways. Whereas basic PD analyses produce a single estimate of automatic and controlled processing within a given task, we believe it is critical to distinguish between two distinct automatic processes and two distinct controlled processes. To assess each of these processes, we have proposed the Quadruple Process Model of implicit task performance (Conrey et al., in press).

The Quad Model (see Figure 1) is a multinomial model (see Batchelder & Riefer, 1999) designed to disentangle four qualitatively distinct processes that contribute to performance on implicit measures that rely on the logic of response compatibility (as illustrated above with Stroop Task). The four processes are: The automatic activation of an association (Association Activation, AC), the ability to determine a correct response (Discriminability, D), the success at overcoming automatically activated associations (Overcoming Bias, OB), and the influence of a general response bias that might guide responses in the absence of other available guides to response (Guessing, G). Whereas AC and G are automatic processes (though G need not be), D and OB are controlled processes.

As an example of how the four process operate, consider an evaluative priming task using pictures of Black and White faces as primes and positive and negative words as targets (e.g., Fazio et al., 1995). In such a task, the presentation of a Black face may automatically activate a negative evaluation (AC) that influences responses to a subsequently presented stimulus word. Depending on the trial type, this automatic tendency may be compatible or incompatible with the correct response determined through discrimination (D) of the target word. If the target word is negative, then the response tendency produced by the automatic evaluation and the response determined via discrimination are compatible. In this case, there is no conflict, and there is no need to overcome bias (OB) in order to produce the correct response. However, if the two response tendencies are incongruent (a Black prime followed by a positive target word), whether the automatic association or accurate discrimination drives the response is determined by whether the participant succeeds in overcoming his or her associations. If no association is activated and the correct response cannot be determined, participants must guess (G).

Though I have used an evaluative priming example, the logic is exactly the same with any implicit measure that compares compatible and incompatible trials. Indeed, to date, our results have come primarily from two different tasks, the Implicit Association Test (IAT: Greenwald et al., 1998) and the Weapons Identification Task (e.g., Payne, 2001).

Analyses using the Quad Model are based on error rates occurring on different types of trials. The processing tree presented in Figure 1 illustrates how the model predicts correct and incorrect responses on compatible and incompatible trials as a function of the operations of the four different processes. For example, there are three different ways to arrive at an incorrect response on incompatible trials. Each of these three combinations of processes represents a set of conditional probabilities by which the incorrect response is produced. These sets of conditional probabilities are used to generate model predictions that are compared to actual results to test for model fit, and are used to generate parameter estimates for each of the four processes (for details, see Conrey et al., in press).

At the most basic level, our data demonstrate that performance on both the IAT and the WIT is a function of all four of the proposed processes. If any process is removed from the model, the model fails. Other data showed that forcing participants to respond quickly on an IAT significantly reduced Discrimination and Overcoming Bias, but did not affect Activation and Guessing. This supports our view that D and OB are controlled processes, whereas AC and G are relatively automatic. In another study, we used the parameter estimates of the four processes to predict biases in response latencies on an IAT. The data showed that response time bias was positively correlated with estimates of the AC parameter, supporting the status of AC as a measure of automatic attitudes. In contrast, response time bias was negatively correlated with the OB parameter, confirming that success at overcoming automatic biases results in smaller estimates of implicit prejudice.

In another application of the model, we re-analyzed data collected by Lambert, Payne, Jacoby, Shaffer, Chasteen, and Khan (2003). In their study, they showed that an anticipated public context ironically increased the extent of implicit stereotyping. Based on a standard PD analysis, they concluded that this effect was due to diminished control in the public versus private context rather than to an increase in stereotype activation in the public context (which could be predicted by drive-based models of social facilitation; e.g., Zajonc, 1965). However, re-analysis with the Quad Model showed a very different result. When automatic and controlled processes were decomposed into four separate components, the results showed that, although Discrimination was diminished in the public condition, Overcoming Bias was enhanced in that condition. Thus, one type of controlled process was inhibited by an audience, and another was enhanced by the audience. Moreover, our analysis showed that the Activation parameter did increase in the public condition. In Lambert et al.'s analysis, this effect had been obscured by the simultaneous increase in Overcoming Bias, which was not measured. Together, these results show that an anticipated audience increases bias on an implicit measure because it inhibits people's ability to discriminate the correct response on the task, and because it increases activation of the dominant stereotypic response.

There are two main conclusions from our research. First, in research on automatic and controlled processes it is useful to move beyond task dissociation paradigms and use process dissociation procedures, instead. Second, it is important to move beyond the simple distinction between automatic and controlled processing, and begin to address important qualitative differences among automatic and controlled processes. As an example, our re-analysis of Lambert et al. (2003) showed that two different controlled processes were affected in opposite ways by the same manipulation. In each of our studies, by assessing all four of the processes in the Quad Model we were able to provide a more comprehensive, nuanced, and accurate description of implicit task performance. To date, the Quad Model has been applied only in the domain of stereotyping and prejudice. However, it should apply more generally to any domain in which automatic impulses are either compatible or incompatible with controlled attempts to overcome those impulses, including research on phobias, addictions, aggression, persuasion, and more. We hope the model will prove to be a useful tool for researchers in many areas of psychology.

Batchelder, W. H., & Riefer, D. M. (1999). Theoretical and empirical review of multinomial process tree modeling. Psychonomic Bulletin & Review, 6, 57-86.

Conrey, F. R., Sherman, J. W., Gawronski, B., Hugenberg, K., & Groom, C. (in press).
Separating multiple processes in implicit social cognition: The Quad-Model of implicit task performance. Journal of Personality and Social Psychology.

Devine, P. G. (1989). Stereotypes and prejudice: Their automatic and controlled components. Journal of Personality and Social Psychology, 56, 5-18.

Fazio, R. H., Jackson, J. R., & Dunton, B. C., Williams, C. J. (1995). Variability in automatic activation as an unobstrusive measure of racial attitudes: A bona fide pipeline? Journal of Personality and Social Psychology, 69, 1013-1027.

Fazio, R. H., & Olson, M. A. (2003). Implicit measures in social cognition research: Their meaning and use. Annual Review of Psychology, 54, 297-327.

Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. K. (1998). Measuring individual differences in implicit cognition: The implicit association test. Journal of Personality and Social Psychology, 74, 1464-1480.

Jacoby, L. L. (1991). A process-dissociation framework: Separating automatic from intentional uses of memory. Journal of Memory & Language, 30, 513-541.

Lambert, A. J., Payne, B. K., Jacoby, L. L., Shaffer, L. M., Chasteen, A. L., & Khan, S. R. (2003). Stereotypes as dominant responses: On the "social facilitation" of prejudice in anticipated public contexts. Journal of Personality and Social Psychology, 84, 277-295.

Payne, B. K. (2001). Prejudice and perception: The role of automatic and controlled processes in misperceiving a weapon. Journal of Personality and Social Psychology, 81, 181-192.

Roediger, H. L. (1990). Implicit memory: Retention without remembering. American Psychologist, 9, 1043-1056.

Stroop. J. R. (1935). Studies on the interference in serial verbal reactions. Journal of Experimental Psychology, 59, 239-245.

Zajonc, R. B. (1965). Social facilitation. Science, 149, 269-274.


Figure 1. The Quadruple Process Model (Quad-Model). Each path represents a likelihood. Parameters with lines leading to them are conditional upon all preceding parameters. The table on the right side of the figure depicts correct (+) and incorrect (-) responses as a function of process pattern and trial type.

About the Author
Jeff Sherman received his PhD in 1994 from University of California, Santa Barbara. From 1994-2004 he held positions of Assistant and Associate Professor at Northwestern University. He is currently a Professor in the Department of Psychology at the University of California, Davis. His research focuses on the cognitive and motivational processes underlying stereotyping and prejudice. This research has been supported by the National Institute of Mental Health since 1996. He is a co-founder of the International Social Cognition Network (ISCON), and currently serves on the Steering Committee. He is Associate Editor of Personality and Social Psychology Bulletin. More information can be found on his faculty page.