The Status of Implicit Measures of Attitudes

Michael Brownstein (John Jay College) - What is the status of research on implicit bias? Criticism is ubiquitous. Recent meta-analytic reviews suggest that the Implicit Association Test is a “poor” predictor of behavior (Oswald et al. 2013) and that changes in scores on implicit measures may not be associated with changes in behavior (Forscher, Lai, et al., ms). Prominent philosophers have questioned the validity of research on implicit social cognition altogether. Edouard Machery (2017), for example, describes an ongoing “rescue mission” within the field, implying that the relevant research is in peril of being discredited. Machery argues that leading methods for studying and theorizing about implicit bias need to be rethought from the ground up, writing that we should not “build theoretical castles on such quicksand.” Headlines in the popular press have been even more pointed. New York Magazine reports, “Psychology’s Favorite Tool for Measuring Racism Isn’t Up to the Job” (Singal 2017); the Chronicle of Higher Education asks, “Can We Really Measure Implicit Bias? Maybe Not” (Bartlett 2017); and most pointedly, the Wall Street Journal describes “The False ‘Science’ of Implicit Bias” (MacDonald 2017). I argue that while there are significant challenges and ample room for improvement, research on the causes, structure, and behavioral effects of implicit bias continues to deserve a role in the sciences of the mind as well as in efforts to understand, and ultimately combat, discrimination and inequality. First, I describe the central issues that have been described as crises, anomalies, or puzzles for the field. To demonstrate that these alleged anomalies are empirical questions on which progress is steadily being made, I place them in the broader historical context of theorizing on the relationship between attitudes and behavior. I respond to potential criticism, and then, finally, point to directions for future research. Along the way, I highlight the importance of these issues for fundamental questions about the architecture of the mind and the metaphysics of action, especially how mental states, attitudes, and dispositions interact with contextual factors to produce behavior. Specifically, I aim to make progress toward a person-by-situation interactionist theory of the mind and action, which requires rethinking the premises underlying enduring philosophical debates about the importance of personal variables (such as beliefs, traits, or even virtues) and situational opportunities and constraints (including social, environmental, and bodily factors).

Calvin Lai (Washington University, St. Louis) - In a meta-analysis (Forscher, Lai, et al., 2018), we synthesized evidence from 492 experiments (87,418 participants) to investigate the effectiveness of procedures to change performance on implicit measures and their effects on explicit measures and behavior. Implicit measures were diverse and included assessment of constructs such as racial attitudes, gender stereotypes, self-esteem, personality, and moral identity. Approaches to change implicit measures were diverse as well, including procedures that taxed mental resources, showed videos of counterstereotypical black individuals, threatened self-esteem, invoked anger or sadness, or invoked egalitarian motivations. Our meta-analysis finds that changes in implicit measures are possible, but those changes do not necessarily translate into changes in explicit measures or behavior. Specifically, we found that implicit measures can be changed, but effects are often relatively weak. Second, many manipulations may have changed non-associative aspects of implicit measures that have little to do with the relationship between associative mental processes and explicit measures/behavior (Calanchini & Sherman, 2013; Calanchini, Sherman, Klauer, & Lai, 2014). We recently used mathematical modeling to distinguish associative and non-associative processes in the effects of 18 interventions to address implicit racial attitudes. We found that contributions of nonassociative processes are rare. Third, it may be that implicit measures are often not correspondent with behavior. This means that implicit measures are not matched well in measurement features to the behavioral measures they seek to predict (Fishbein & Ajzen, 1975). For instance, the same general measure of implicit racial attitudes has been used to predict willingness to date interracially, physiological activation in response to interacting with someone of another race, and likelihood of discriminating in hiring. In studies examining the role of measurement correspondence, we have found that implicit measures better predict behavior when they are better matched in measurement features. Finally, a fourth possibility is that automatic retrieved associations have weak causal influence or no causal influence at all. Recent theories suggest that weak individual effects of automatically retrieved associations “accumulate” into larger group disparities (Greenwald et al., 2015; Payne et al., 2017).

Agnes Moors (KU Leuven) - Recent meta-analyses show low correlations between implicit attitude measures and behavior, suggesting that these measures are weak predictors of behavior. This has led some authors to challenge the validity of implicit measures (Oswald et al., 2013), whereas others argue that low correlations are obtained because theoretical constraints are overlooked (Brownstein, Madva, & Gawronski, 2017). I propose to approach this debate from the perspective of theories of behavior causation. Dual process models of behavior causation distinguish between (a) a stimulus-driven process in which a stimulus activates the association between the representation of stimulus features and the representation of a response (S-R) and (b) a goal-directed process in which the value and expectancy of the outcomes of one or more behavior options are weighed before an action tendency is activated (S:R-O → R). Traditional dual process models have a default-interventionist architecture in which the stimulus-driven process is automatic and therefore the default determinant of behavior whereas goal-directed processes are nonautomatic so that they can only intervene when there is sufficient opportunity and motivation. Many attitude researchers seem to subscribe to a defaultinterventionist version of dual process theory: An attitude, which can be considered as the representation of a stimulus feature (i.e., valence: positive/negative), is directly connected to the representation of a response (e.g., approach/avoidance). This stimulus-driven process can be overruled by a goal-directed process but only when there is enough opportunity and motivation. Implicit measures are supposed to tap into the stimulus-driven process, and explicit measures in the goal-directed process. I propose an alternative dual process model (Moors, Boddez, & De Houwer, 2017) in which stimulus-driven and goal-directed processes can both be automatic, which implies that they often operate in parallel and enter in competition (i.e., a parallel-competitive architecture). The model, moreover, assumes that when both processes do enter in competition, the goal-directed process should win because it is more likely to lead to optimal behavior. On this view, the goal-directed process determines the lion share of behavior whereas the stimulus-driven process determines behavior only in exceptional cases. From the perspective of the traditional dual process model, which assumes that most behavior is determined by a stimulus-driven mechanism, it makes sense to continue measuring attitudes, especially with implicit measures because they allow bypassing the regulatory attempts of goal-directed processes. From the perspective of the alternative dual process model, however, which assumes that most behavior is determined by a goal-directed mechanism, it makes more sense to focus on measuring the values and expectancies of behavior options. Here too, implicit measures are needed, because values and expectancies are not always consciously reportable. In our lab, we have taken the first steps to design implicit measures that can assess these constructs, taking advantage of the IAT (Greenwald, McGhee, & Schwartz, 1998), the IRAP (Barnes-Holmes, Barnes-Holmes, Stewart, & Boles, 2010), and the RRT (De Houwer, Heider, Spruyt, Roets, & Hughes, 2015).