We conducted an initial investigation of implicit (i.e., automatically revealed) racial phenotypicality bias, focusing specifically on implicit attitudes toward Afrocentric versus Eurocentric hair texture. Across three high-powered studies, the last of which was formally preregistered, we found that although implicit and more commonly used explicit measures of racial phenotypicality bias were positively correlated with each other, this correlation, similar to the relationship between implicit and explicit race attitudes (Nosek et al., 2007), was small in size. Crucially, implicit and explicit phenotypicality bias each uniquely contributed to the prediction of relevant social behavior among White American participants (Studies 2–4). Moreover, the two also differed in mean levels: Whereas White American participants expressed neutrality on explicit measures of attitude (Studies 2–3), implicit measures revealed a large degree of preference in favor of Eurocentric (straight or wavy) over Afrocentric (curly) hair texture.
Strikingly, in Study 4, for which data were collected in June 2021, White American participants even expressed significantly positive views toward Afrocentric (curly) hair texture on an explicit (self-report) measure of attitude. This difference between Studies 2–3, in which self-reports were at neutrality, and Study 4, in which they showed outgroup-favoring tendencies, may have emerged for a number of reasons. Potentially, participants may have been more reluctant to express even neutrality on an absolute scale (with the endpoints labeled “extremely unprofessional” vs. “extremely professional” or “extremely uncaring” vs. “extremely caring”) rather than a relative scale (with the endpoints labeled “much less professional” vs. “much more professional” or “much less caring” vs. “much more caring”).
More importantly, however, data from the Project Implicit website (e.g., Charlesworth & Banaji, 2019) as well as from representative social surveys (e.g., Nosek et al., 2009) suggest that White Americans have become increasingly reluctant to express views that they believe may be seen by others as biased. What counts as biased is, of course, multiply determined, but social norms are known to play a key role in this regard (Tankard & Paluck, 2016). As such, as more jurisdictions follow the 11 states already banning discrimination based on hair texture as of this writing (California, New York, New Jersey, Virginia, Colorado, Washington, Maryland, Connecticut, New Mexico, Delaware, and Nebraska), social norms may shift even more strongly toward making it unacceptable to express negative views of natural hairstyles. If this is, indeed, going to be the case, then we anticipate that the importance of implicit measures, which are less amenable to voluntary control in line with participants’ egalitarian values, will only gain in importance in the study of racial phenotypicality biases in the coming years. Such considerations apply particularly strongly to political liberals, such as the majority of the Project Implicit participants recruited for the present studies, who are known to be especially sensitive to concerns related to equity and fairness (e.g., Graham et al., 2009).
Explicit and implicit phenotypicality bias were also clearly dissociated among African American participants: Whereas the former revealed preference for Afrocentric hair texture, the latter revealed preference for Eurocentric hair texture. This result is reminiscent of other findings among members of historically disadvantaged social groups who often express preference for the ingroup on explicit measures but tend to reveal neutrality or even outgroup-favoring tendencies on implicit measures (Nosek et al., 2007). Such effects may emerge as a result of members of such social groups internalizing prevalent societal attitudes, as a result of personal experiences with discrimination, and a number of other factors. As such, the present project suggests that future studies of racial phenotypicality bias may benefit from the inclusion of implicit measures, such as the Implicit Association Test (Greenwald et al., 1998), in addition to more commonly used self-report measures not only among White American but also among Black American participants.
Notably, more closely investigating the antecedents and correlates of hair texture bias among Black Americans may be especially instructive with regard to understanding the mechanisms underlying such effects, on which the present data are relatively silent. For example, implicit negativity toward natural hair may stem from perceptions of professionalism (e.g., Koval & Rosette, 2020; Opie & Phillips, 2015). Additionally, or alternatively, natural hairstyles may be perceived to be associated with political causes such as the Black liberation movement of the 1960s and its present-day successors, whereas chemically relaxed hairstyles may be perceived to be associated with assimilation into White culture. In this context, it may be particularly instructive to probe how (implicit) self-esteem and (implicit) self-concept are associated with implicit hair attitudes specifically among Black Americans. Notably, predictive validity of the hair attitude IAT among Black Americans may be improved by using more culturally appropriate category labels (e.g., “natural” vs. “relaxed”) instead of “curly” versus “straight”, which were selected primarily with White American participants in mind.
Importantly, in line with existing theoretical perspectives (Maddox, 2004) and empirical findings (Blair et al., 2004), the present data also suggest that implicit phenotypicality bias and implicit race bias, although statistically related to each other, are far from redundant. Specifically, implicit race attitudes accounted for only about 20 percent of the meaningful variance in implicit hair attitudes (Study 1). Moreover, when data were aggregated across studies, both implicit hair attitudes and implicit race attitudes significantly predicted a measure of intergroup behavior, specifically expressions of support toward a Black plaintiff who sued a corporation alleging discrimination on the basis of her natural hairstyle, among White American participants (Studies 2–4). Notably, each IAT did so incrementally, after accounting for the effects of the other IAT and the effects of explicit measures. As such, these data are clearly in conflict with recent perspectives questioning whether implicit and explicit (race) attitudes represent different constructs.
The latter finding regarding predictive validity is both methodologically and theoretically informative. When it comes to the methodological features of the present project, in investigating predictive validity, we sought to implement several recommendations formulated in the recent meta-analysis by Kurdi et al. (2019b). Specifically, the sample size in the present studies was over ten times larger than the median effect size in studies investigating predictive validity of the IAT and related measures. This distinction is made especially relevant by the fact that the mean effect size revealed by the meta-analysis for both implicit (β = 0.14) and explicit predictors (β = 0.11) was small. In addition, also in line with recommendations by Kurdi et al. (2019b) and more general recommendations by Nosek et al., (2018, 2019), we confirmed the exploratory results emerging from Studies 1–3 in a preregistered and highly powered final study (Study 4). Finally, and crucially, we made sure that the to-be-predicted behavioral measure had excellent validity and internal consistency. As is well known from classical test theory, noisy criterion measures, and especially unvalidated one-shot measures of behavior, impose a (sometimes quite severe) upper limit on the predictive validity of attitude measures (e.g., Furr & Bacharach, 2013). On a related note, we explicitly took measurement error into account by relying on a structural equation modeling approach.
When it comes to the theoretical implications of the present work for the attitude–behavior relationship, based on the correspondence principle (Ajzen & Fishbein, 1977) and recent empirical findings by Irving and Smith (2020), we expected that a specific implicit attitude (toward Afrocentric hair texture) would be more predictive of a specific behavior (views about discrimination based on Afrocentric hair texture) than a more general implicit measure of race attitudes. Although some individual studies included in the present paper seemed to provide initial evidence in favor of this idea, in the more highly powered last study, both implicit hair attitudes and implicit race attitudes were equally associated with the relevant criterion behavior. Moreover, this finding was also confirmed in an internal meta-analysis combining data from all relevant studies.
Do these findings imply that the correspondence principle is irrelevant to implicit social cognition? We believe that this conclusion is premature, for a number of reasons. Notably, we used the correspondence principle in making multiple decisions related to design of the hair attitude IAT. For example, we used only images of women and only images of Black individuals on the basis of the correspondence principle. Specifically, we reasoned that views on a legal case involving discrimination against a Black woman would best be predicted by an IAT that included stimuli only of the relevant race and gender, especially given that “male” and “White” are perceived as default categories in present-day American society (e.g., Bosson et al., 2018). Moreover, we may have failed to find direct evidence for the correspondence principle in the present studies because hair may be a relatively central and prototypical feature of Black women. The incremental predictive validity of the more specific IAT may have been higher for a feature that is less centrally associated with the overarching group representation itself.Footnote 8
We hope that future work will more systematically explore the operation of the correspondence principle in the context of the IAT and implicit social cognition more generally. When it comes to the IAT specifically, the correspondence principle may operate either at the level of attributes and attribute stimuli or at the level of categories and category stimuli. For example, when it comes to attribute stimuli, one may reason that specific attributes (such as “professional” vs. “unprofessional”) may be more predictive of the legal judgments in the present studies than general attributes (such as “good” vs. “bad”) given that the legal case centers specifically on the question of whether natural Black hair and professionalism are compatible with each other. However, in this context, considerable added complexity may be created by the fact that stereotype IATs using clearly valenced attributes are usually highly correlated, or even redundant, with attitude IATs (Kurdi et al., 2019a; Phills et al., 2020).
Alternatively, or in addition, correspondence may be thought of as applying to the category labels or category stimuli used on the IAT. For example, in Studies 1–3 we used a race IAT that had both male and female faces as category stimuli, whereas in Study 4 we used a race IAT that had only female faces as category stimuli. Based on the correspondence principle, it may have been expected that the race IAT in Study 4 would be more predictive of the criterion behavior given that the criterion behavior related specifically to discrimination against a Black woman. However, IAT effects tend to generally be more strongly influenced by category (and attribute) labels than by the specific category (and attribute) stimuli (e.g., Axt et al., 2021; Mitchell et al., 2003). As such, the predictive validity of the race IAT may have been improved had the category labels been changed to “White women” and “Black women” from “White people” and “Black people”. These specific variations, and other variations along similar lines, may be more systematically explored in future work.
Moreover, in Studies 2–3, implicit hair attitudes predicted a behavior (judgments regarding a legal case) that was, to a large extent, under participants’ volitional control. As such, the present project provides evidence in favor of another idea proposed by Kurdi et al. (2019b), namely that, contrary to the assumptions of several dual-process theories, implicit attitudes may be associated not only with relatively automatic but also with relatively controllable behaviors. At the same time, it should be noted that legal judgments were more strongly predicted by explicit judgments than by implicit attitudes. However, to a considerable degree, this effect is likely to have been due to method variance shared between the explicit scale and the to-be-predicted behavior (Campbell & Fiske, 1959). Moreover, third variables, including drive for consistency between attitudes and behavior (Festinger & Carlsmith, 1959) and motivation to control prejudiced responding (Plant & Devine, 1998), may also have played a role.
Given that, to our knowledge, the present paper reports the first peer-reviewed studies demonstrating the existence and downstream consequences of implicit phenotypically bias more generally and implicit bias against Afrocentric hair more specifically, the present studies should be understood as providing an existence proof argument. We hope that research in social cognition will continue to investigate implicit racial phenotypicality biases and their relationship with social behavior and further clarify some issues left unaddressed here.
One such issue concerns stimulus variation. Specifically, although Studies 2–3 and Study 4 yielded converging results despite differences in stimulus materials, future work may further explore other possible variations in the IAT procedure. For example, in the present studies we used only relatively dark-skinned Black targets; in the future, investigators may want to ask whether the effect extends to light-skinned Black targets. Similarly, in the present studies we used only female targets; future work may want to examine generalization to targets of other genders. Moreover, the mean level and predictive validity estimates provided by IATs relying on real facial images may be different from the estimates provided by the IATs in the present studies, which all relied on line drawings. Finally, although the present data suggest that hair texture bias emerges especially strongly in the context of Black (female) targets, future work may want to more systematically explore variation by race both in terms of mean levels and in terms of predictive validity.
Given that the studies reported above rely exclusively on the Implicit Association Test, it is unclear to what extent implicit attitudes toward Eurocentric versus Afrocentric hair texture are automatically activated without specific instruction to categorize stimuli along this dimension. Future work using alternative implicit measures of attitudes, such as the evaluative priming task (Fazio et al., 1995) or the affect misattribution procedure (Payne et al., 2005), may be used to investigate this issue. We view such replications and extensions of the present work as especially important given that performance on different implicit measures of attitude has been suggested to emerge from overlapping but not fully identical cognitive processes (De Houwer & Moors, 2010). Similarly, based on the present data, it is unclear whether and to what extent implicit phenotypicality bias would predict criterion behaviors other than the one chosen for the present studies. Specifically, it may be of interest to move away from tightly controlled but relatively artificial settings toward exploring more ecologically realistic behaviors. However, we note that in doing so, it is important to keep in mind considerations regarding the validity and internal consistency of the to-be-predicted behaviors (e.g., Kurdi et al., 2019b).
Finally, recent work in implicit social cognition has demonstrated striking similarities between response time-based measures of implicit attitudes administered to individual participants and algorithmically derived measures of semantic association obtained from large corpora of online text (e.g., Caliskan et al., 2017; Kurdi et al., 2019a; for a review, see Caliskan & Lewis, 2020) and online images (e.g., Steed & Caliskan, 2021). Moreover, relative to the usually modest relationship found between implicit (and explicit) attitudes and intergroup behavior at the level of individuals, studies investigating the same relationship at the level of geographic units tend to reveal considerably larger effects (Hehman et al., 2019; Payne et al., 2017). We hope that future work will be able to combine these novel approaches with the idea suggested by the present project, namely that focusing on specific features of the attitude object (e.g., biases based on specific phenotypic traits) may improve the prediction of intergroup behavior, to investigate the precursors, magnitude, and downstream consequences of explicit and implicit social evaluations beyond the laboratory.