Reduction of Prejudice?

Although many scholars have observed a general trend of traditional explicit measures of racial prejudice toward African Americans declining up through the turn of the century (Bobo, Reference Bobo, Smelser, Wilson and Mitchell2001; Krysan, Reference Krysan2011; Schuman et al., Reference Schuman, Steeh and Bobo1997), a more recent account paints a more complex picture: some explicit measures have shown a continuing decline, while others have shown stagnation or even slight increases (Moberg et al., Reference Moberg, Krysan and Christianson2019). This pattern of evidence is certainly consistent with the conclusion that some forms of explicit prejudice have declined, even if not all have. And numerous events that have occurred in the United States since 2016 have highlighted continued and vigorous prejudice.

Skeptics have raised the possibility that rather than attitudes and beliefs truly changing, what may have changed are societal norms for what one can comfortably acknowledge. As Americans have increasingly endorsed egalitarian values, many may now be unwilling to openly admit harboring inegalitarian attitudes. As a consequence, people who have negative feelings toward African Americans or who ascribe negative traits to African Americans may no longer report them honestly (Judd et al., Reference Judd, Park and Ryan1995; Tarman & Sears, Reference Tarman and Sears2005).

This possibility is consistent with a huge literature in psychology on self-presentation. In general, people strive to present themselves favorably when interacting with others, even at the expense of honesty (Goffman, Reference Goffman1959). That is why some people sometimes avoid reporting potentially embarrassing attitudes and behaviors, a tendency referred to as impression management social desirability bias (Paulhus, Reference Paulhus1984, Reference Paulhus, Angleitner and Wiggins1986, Reference Paulhus, Braun, Jackson and Wiley2002; Paulhus & Reid, Reference Paulhus and Reid1991). This social desirability bias is thought to be particularly likely in a survey interview where respondents have to openly admit potentially embarrassing attitudes to an interviewer who might pass judgment based on a reported response (Paulhus, Reference Paulhus, Braun, Jackson and Wiley2002). In line with these concerns, some research suggests that reports of racial prejudice are influenced by the race of the interviewer, such that respondents report more favorable evaluations of Black people when speaking with Black interviewers than when speaking with White interviewers (e.g., Anderson et al., Reference Anderson, Silver and Abramson1988a, Reference Anderson, Silver and Abramson1988b; Finkel et al., Reference Finkel, Guterbock and Borg1991; Hatchett & Schuman, Reference Hatchett and Schuman1975). The reactivity of explicit racial prejudice questions in survey interviews has led scholars to ask “whether or not it is worth all the trouble to administer such surveys in the first place” (Corstange, Reference Corstange2009, p. 46).

Race-of-Interviewer Effects

One type of evidence that has been used to discourage the use of traditional explicit prejudice questions and “new” racism measures is the effects of the race of the survey interviewer on the level of prejudice reported. More than 100 studies have been conducted on the effect of the race of a survey interviewer, and these race-of-interviewer effects are particularly strong on questions that explicitly mention race (Anderson et al., Reference Anderson, Silver and Abramson1988b; Davis, Reference Davis1997; Schuman & Converse, Reference Schuman and Converse1971). The most prominent example of such race-of-interviewer effects comes from the 1971 Detroit Area Study. In that study, White respondents were randomly assigned to be interviewed by either a Black or a White interviewer. Only 26 percent of respondents interviewed by a White interviewer said that they would not “mind if a relative married a Negro.” In contrast, a whopping 72 percent of respondents who were interviewed by a Black interviewer said that they would not mind (Hatchett & Schuman, Reference Hatchett and Schuman1975). This can be viewed as evidence of powerful social desirability bias, a difference of forty-six percentage points.

However, that same study also examined responses to other measures of racial attitudes and documented notably weaker effects: thirty-five percentage points on whether “Negro” and White children should attend the same school, thirty-one percentage points on the question of whether respondents would be disturbed if a “Negro with the same income and education as you moved into your block,” and nine percentage points on the question of whether “Negro” and White children should be allowed to play together. One cannot deny that these numbers are quite different from one another, and if only one such result had been discovered, 9 percent is much smaller than 46 percent.

Nonetheless, these sorts of findings have often been put forward as evidence of social desirability bias in prejudice questions: respondents were not willing to honestly reveal their racial attitudes because they thought a Black interviewer might react negatively to unfavorable opinions about Black people (e.g., Anderson et al., Reference Anderson, Silver and Abramson1988a, Reference Anderson, Silver and Abramson1988b; Finkel et al., Reference Finkel, Guterbock and Borg1991). And it is interesting that researchers have uncritically accepted this evidence as proving that explicit measures underestimate the level of prejudice in the population due to social desirability bias. In fact, there is no evidence from Hatchett and Schuman (Reference Hatchett and Schuman1975) or any other of the studies in this literature about which reports are more valid: the reports to White interviewers or the reports to Black interviewers.

Perhaps we are so inclined to accept the social desirability hypothesis that we assume the reports to White interviewers are the more valid ones, despite no evidence of that. But in fact, it is possible that reports made to White interviewers are the dishonest ones, intended to avoid appearing unprejudiced to people who might harbor anti-Black attitudes. Whether or not this seems plausible, studies of interviewer effects have not routinely ruled it out. So it is possible that reports to Black interviewers might have been more honest, and as a result, White interviewers caused the survey to overestimate true levels of prejudice.

Another important problem with many of the race-of-interviewer studies involving face-to-face interviews is that they did not randomly assign interviewers to respondents (e.g., Anderson et al., Reference Anderson, Silver and Abramson1988b; Dohrenwend et al., Reference Dohrenwend, Colombotos and Dohrenwend1968; Schaeffer, Reference Schaeffer1980). Random assignment rarely happens in face-to-face interviews, particularly with representative samples, because of the cost of sending interviewers across the country. Accordingly, interviewers typically interview people who live near them.

Therefore, interviewers who interview respondents of a different race are more likely to live in racially diverse areas than are interviewers who interview only respondents of their own racial group. Research has established that racial diversity facilitates intergroup contact, which in turn can reduce prejudice (Allport, Reference Allport1954; Pettigrew & Tropp, Reference Pettigrew and Tropp2006). Hence, an alternative explanation for the race-of-interviewer effects seemingly observed in these studies is that people who are interviewed by an interviewer of another race actually have more favorable attitudes toward other racial groups. Thus, what looks like a race-of-interviewer effect is actually due to the higher likelihood of Black interviewers to interview White people with positive attitudes toward African Americans. Again, lack of direct evidence of which reports are more accurate leaves this ambiguity open.

Studies of race-of-interviewer effects in telephone surveys were often able to prevent this problem by randomly assigning interviewers to telephone numbers (Cotter et al., Reference Cotter, Cohen and Coulter1982; Davis, Reference Davis1997; Finkel et al., Reference Finkel, Guterbock and Borg1991). These studies therefore seem to provide stronger evidence of social desirability bias. However, Kim and colleagues (Reference Kim, Krosnick and Lelkes2019) recently put forward an alternative – and largely overlooked – explanation for the seeming race-of-interviewer effect in telephone surveys. Kim et al. (Reference Kim, Krosnick and Lelkes2019) analyzed telephone surveys conducted before the 2008 United States presidential election. In line with previous research, what appeared to be a race-of-interviewer effect was observed: Black interviewers recorded more respondents saying they would vote for Barack Obama than did White interviewers. This suggests that the more interviews that are conducted by Black interviewers in a study, the larger would be the apparent proportion of Americans voting for Mr. Obama.

However, what looked like a race-of-interviewer effect turned out to be a consequence of differential respondent recruitment. Most people can identify the race of a person on the phone (Baugh, Reference Baugh2000), and the similarity-attraction hypothesis from social psychology suggests that an interviewer will be especially successful at recruiting participation from a potential respondent of the same race. That is, Black interviewers may be more likely than White interviewers to successfully recruit Black respondents. And indeed, this pattern was observed by Kim et al. (Reference Kim, Krosnick and Lelkes2019). Furthermore, White respondents who harbor anti-Black prejudice might be more reluctant to be interviewed by a Black person than a White person.

Taken together, this reasoning suggests that the proportions of respondents who were either Black or who were nonprejudiced Whites might have been greater among people interviewed by Black interviewers than among people interviewed by White interviewers. And not surprisingly, the proportion of Black respondents who said they would vote for Mr. Obama was greater than the proportion of White respondents who said they would vote for him. After controlling for the race of the respondent, there was no longer an effect of the race of the interviewer on reported voting intentions. Thus, although differences in results obtained by Black and White interviewers might have been the result of racial prejudice, they were not necessarily the result of intentional misreporting during the interviews.

To be sure, we are not claiming that there are no race-of-interviewer effects. We are merely pointing out that the basis to cite these effects as evidence for social desirability bias is less strong than has sometimes been assumed and that alternative explanations may account for observed evidence.

Public and Private Response Modes

Another widely cited body of evidence for distortion of reports due to social desirability bias and against the use of explicit measures of prejudice comes from studies that compared responses in survey modes with an interviewer present to responses in modes without an interviewer (e.g., Krysan, Reference Krysan1998; Tourangeau & Smith, Reference Tourangeau and Smith1996). Because respondents can reasonably expect interviewers to pass judgment on them based on their answers to survey questions, the social desirability argument suggests that people should be more willing to honestly report embarrassing or controversial opinions and behaviors when no interviewer is present.

In line with this notion, research has found more reports of controversial behaviors (e.g., the number of sexual partners or drug use) being provided when respondents completed a questionnaire anonymously on a computer than when respondents provided responses to questions orally to interviewers (Tourangeau et al., Reference Tourangeau, Jobe and Smith1995; Tourangeau & Smith, Reference Tourangeau and Smith1996). An early meta-analysis comparing computer-administered questionnaires to face-to-face interviews documented more reports of sensitive information in the former (Richman et al., Reference Richman, Weisband and Kiesler1999). Later studies have also documented higher frequencies of controversial or undesirable behaviors and opinions being reported in online surveys than in interviewer-administered modes, such as telephone surveys (Chang & Krosnick, Reference Chang and Krosnick2009, Reference Chang and Krosnick2010; Holbrook & Krosnick, Reference Holbrook and Krosnick2010; Kreuter et al., Reference Kreuter, Presser and Tourangeau2008). Again, it is tempting to conclude that the presence of an interviewer causes reluctance to admit embarrassing things about oneself, so the reports made on computers are presumed to be more valid than the reports made to interviewers.

But there is an alternative view of this finding as well (see Holbrook & Krosnick, Reference Holbrook and Krosnick2010; Kreuter et al., Reference Kreuter, Presser and Tourangeau2008). Specifically, reports of more undesirable attitudes and behavior given under complete privacy do not necessarily reveal more accuracy. This is possible because past studies did not typically assess accuracy and instead assumed that more reports are indicators of more honesty. Instead, interviewers may create a greater sense of accountability when respondents answer questions than when questionnaires are completed privately. And this accountability may inspire respondents to devote more cognitive effort to answering questions and to answer more accurately.

With regard to prejudice measurement, offering privacy to respondents may result in sloppy answering behavior that reduces the validity of prejudice measures. In line with this, Lelkes et al. (Reference Lelkes, Krosnick and Marx2012) found in three experiments that guaranteeing participants complete anonymity when completing a questionnaire sometimes increased socially undesirable answers but consistently decreased the accuracy of the responses. We should therefore not be too quick to equate findings of more prejudice in private answering modes with better measurement due to less social desirability bias, as some scholars have done (e.g., Gnambs & Kaspar, Reference Gnambs and Kaspar2014; Piston, Reference Piston2010; Richman et al., Reference Richman, Weisband and Kiesler1999; Tourangeau et al., Reference Tourangeau, Jobe and Smith1995).

Stark et al. (Reference Stark, Van Maaren and Krosnick2022) explored this alternative explanation using data from the 2008 American National Election Study, which contained an experiment that explored the effect of answering prejudice questions with and without potential social desirability concerns. Before and after the 2008 presidential election, the same representative national sample of 1,009 White non-Hispanic Americans participated in two computer-assisted interviews conducted by an interviewer in the respondents’ homes. During both interviews, respondents answered two explicit prejudice questions about how (1) hard-working and (2) intelligent they thought White people and Black people are. During the second interview, these questions were asked and answered orally. But during the first interview, these questions were asked with Audio Computer-Assisted Self-Interviewing (ACASI): the interviewer handed over a laptop to the respondent, who heard the questions being read aloud on headphones and saw the questions and answer choices on the laptop screen. Answers were typed confidentially on the keyboard without the interviewer seeing or hearing the questions or the answers. ACASI is thought to eliminate impression management concerns.

In line with social desirability bias, the mean ratings of Black people as hard-working and intelligent were more pro-Black in the face-to-face interviews than in the ACASI interviews. On scales that were coded to range from 0 to 1 (1 meaning most hard-working), the mean rating of Black people decreased from 0.527 in the oral mode to 0.490 in the ACASI mode (p < .001). Also, significantly more participants said in the private mode that White people are more hard-working than Black people (oral: 44.97 percent, ACASI: 49.34 percent, $p=.018$) and more intelligent than Black people (oral: 39.74 percent, ACASI: 43.96 percent, $p=.018$). This fits the narrative of many proponents of implicit prejudice measures, as these results suggest that openly asking explicit prejudice questions in a survey interview leads to an underestimation of the real level of anti-Black prejudice (Piston, Reference Piston2010).

However, not in line with this reasoning, significantly more people said that Whites are less hard-working and less intelligent than Black people in the private answer mode than in the oral administration mode (hard-working oral: 3.21 percent, ACASI: 8.46 percent, p < .001; intelligent oral: 1.75 percent, ACASI: 3.38 percent, $p=.037$). As a consequence, there was hardly any difference between modes in the differential measures comparing White and Black people’s hard-working nature (Δ = 0.007, $p=.119$) and intelligence (Δ = 0.012, $p=.001$) on scales ranging from 0 to 1. The social desirability bias discussion typically centers around social norms that prevent the open expression of negative attitudes toward other social groups. Evidence of a substantive number of respondents seemingly hiding their negative attitudes toward their own group (Whites) does not fit with this narrative.

There was also no evidence that the privately measured attitudes were more valid than the orally reported attitudes when predictive validity was assessed by using the stereotypes measure to predict other explicitly measured attitudes and implicit measures. If oral reports of stereotypes suffer from distortion due to social desirability response bias, then the privately assessed attitudes should be more strongly related to known correlates of prejudice. This particular study examined the correlation with attitudes toward homosexuals that were measured in ACASI mode (and presumably undistorted by social desirability response bias), and this correlation did not vary by administration mode. Moreover, orally reported stereotypes of Black people were more strongly associated with implicit prejudice toward Black people measured with the Affect Misattribution Procedure (AMP) than were the stereotypes of Black people reported privately in ACASI mode. If implicit measures of prejudice are not affected by social desirability response bias, then one would expect a stronger association of the implicit measure with privately expressed prejudices than with openly expressed prejudices, but the opposite was observed.

We are not claiming that social desirability bias does not exist nor that it does not distort responses to explicit prejudice questions. Instead, we are claiming that the literature cited to support this conclusion using studies comparing administration modes is not as informative as it might seem. There is convincing evidence that offering privacy to respondents so that they do not have to fear judgment for their opinions leads to more reports of prejudicial attitudes. However, few studies were able to compare the reports in the private answer mode to a benchmark that allowed conclusions about the validity of these reports (for exceptions, though not with regard to prejudice, see Holbrook & Krosnick, Reference Chang and Krosnick2010; Kreuter et al., Reference Kreuter, Presser and Tourangeau2008; Tourangeau & Smith, Reference Tourangeau and Smith1996). Most existing research relies on the “more-is-better” assumption (Krumpal, Reference Krumpal2013) and equates reports of more controversial attitudes and behavior to more honest answers. Stark et al.’s (Reference Stark, Van Maaren and Krosnick2022) study suggests that this conclusion might be premature, as reports of “more” prejudice do not necessarily imply more accurate measurement.

Randomized Response Technique

Another prominent form of evidence for distortion of explicit prejudice measures by social desirability bias comes from studies implementing techniques that guarantee absolute privacy to respondents. One such approach is the randomized response technique (RRT). This procedure makes it impossible for the interviewer to know whether a respondent gave a prejudiced answer or not.

Various versions of RRTs have been developed that can be classified into forced-response variants and forced-question variants (Krumpal, Reference Krumpal2013). All variants involve some form of randomizing (e.g., a coin toss or dice throw) that the respondent conducts without the interviewer being able to see the outcome. In forced-response RRTs, the outcome of the randomizing determines whether the respondent will answer a sensitive question truthfully (e.g., answer “yes” or “no” when a coin toss shows heads) or to simply respond with “yes” (when the coin shows tails). In forced-question RRTs, the outcome of the randomizing determines which of two questions the respondent will answer, only one of which is a sensitive question, and the other of which has a known expected distribution of responses. The interviewer conducting the RRT cannot know which question the respondent answered, and it is possible to implement calculations to derive the proportion of respondents who reported an embarrassing attribute.

Many studies have found more socially undesirable responses with the RRT than when the same question is asked directly of respondents. Because RRTs are typically limited to yes/no responses, the majority of studies concerned socially undesirable behaviors such as stealing (Wimbush & Dalton, Reference Wimbush and Dalton1997) or cheating (Scheers & Dayton, Reference Scheers and Dayton1987) that people have either committed or not (for overviews, see Holbrook & Krosnick, Reference Chang and Krosnick2010; Lensvelt-Mulders et al., Reference Lensvelt-Mulders, Hox and Van Der Heijden2005). But there are also studies that found more reports of socially undesirable attitudes such as prejudice with the RRT than in direct questions (e.g., anti-Semitism, Krumpal, Reference Krumpal2012). For instance, a study in Germany found that about 25 percent more people said that they would mind if their hypothetical daughter had a relationship with a dark-skinned Nigerian immigrant with an RRT than when they were asked directly (Ostapczuk et al., Reference Ostapczuk, Musch and Moshagen2009). This result is in line with the notion that social desirability response bias distorts results of explicit prejudice questions.

However, a shortcoming of most RRT studies is their reliance on the “more-is-better” assumption that also limits research comparing public and private survey modes. That is, without being able to compare the results to an objective benchmark, researchers can only assume that more undesirable responses in the private condition reflect more accurate answers (Krumpal, Reference Krumpal2012).Footnote ⁴

Moreover, plenty of studies have not found more undesirable reports using RRTs than using direct questions (e.g., Akers et al., Reference Akers, Massey and Clarke1983; Danermark & Swensson, Reference Danermark and Swensson1987), and some even reported fewer such reports in RRTs than in direct questions (e.g., Brewer, Reference Brewer1981; Holbrook & Krosnick, Reference Chang and Krosnick2010; Williams & Suen, Reference Williams and Suen1994). This suggests that the technique does not always work as intended. In fact, research has found that many respondents disobeyed the RRT instruction (Boeije & Lensvelt-Mulders, Reference Boeije and Lensvelt-Mulders2002; Ostapczuk et al., Reference Ostapczuk, Musch and Moshagen2009). This may be caused by the complex and unexpected instructions of many RRTs (why do you need to toss a coin during an interview?) and the difficulty of understanding how the randomization guarantees respondents’ privacy. As a consequence, some respondents may not trust that their privacy is protected. For instance, Holbrook and Krosnick (Reference Chang and Krosnick2010) found that the coins of respondents should have come up “heads” in 147 percent of the tosses to produce the results they found. Although new versions of RRTs have been developed that can account for such response biases to some extent (Coutts & Jann, Reference Coutts and Jann2011; Cruyff et al., Reference Cruyff, Van Den Hout and Van Der Heijden2007), the evidence in favor of social desirability bias produced by existing RRT studies is less unambiguous than many think.

Item Count Technique

Another method to collect information on sensitive attitudes and behavior is the item count technique (ICT, Miller, Reference Miller1984), which is also sometimes called the list technique (Kuklinski, Sniderman, et al., Reference Kuklinski, Cobb and Gilens1997) or the unmatched count technique (Coutts & Jann, Reference Coutts and Jann2011). This approach has a clear advantage over the RRT as it can be implemented without requiring respondents to conduct a randomization themselves. In the ICT, respondents might be shown a list of statements and asked how many of these statements they find upsetting. Kuklinski, Sniderman and colleagues (Reference Kuklinski, Sniderman and Knight1997) used this list:

• – The federal government increasing the tax on gasoline

• – Professional athletes getting million-dollar-plus salaries

• – Large corporations polluting the environment

To ensure privacy, respondents are not asked which statements are upsetting, only how many. Respondents are randomly assigned to see the list of these three statements or to see a list with one additional and sensitive statement, such as “A Black family moving next door to you.” The proportion of respondents who are upset by a Black family moving in next door can be inferred by comparing the mean number of statements reported in the condition with the short list to the mean number reported in the condition with the added statement.

Just like the randomized response technique, the ICT has been criticized for only reporting population averages that do not allow measurements of individual opinions or behaviors. However, statistical techniques have been developed that enable researchers to gauge correlations between variables (Blair & Imai, Reference Blair and Imai2012; Holbrook & Krosnick, Reference Chang and Krosnick2010; Imai, Reference Imai2011). And the double list variant, the LISITT approach, and the item sum technique overcome issues of inefficient estimators and low statistical power (Corstange, Reference Corstange2009; Glynn, Reference Glynn2013; Trappmann et al., Reference Trappmann, Krumpal and Kirchner2014).

Many of the existing ICT studies on prejudice are made ambiguous by the possibility of floor and ceiling effects (Kuklinski, Cobb, & Gilens, Reference Kuklinski, Cobb and Gilens1997; Martinez & Craig, Reference Martinez and Craig2010; Redlawsk et al., Reference Redlawsk, Tolbert and Franko2010). Such effects occur when respondents believe that all or none of the non-sensitive statements are upsetting, for example. That is, a respondent might be upset by increasing the tax on gasoline, million-dollar-plus salaries of athletes, and pollution by large corporations. This respondent cannot admit being upset by a Black family moving in next door, because the answer “four” would explicitly reveal this bias. Similarly, reporting that none of the four concepts are upsetting would reveal a respondents’ opinion about a Black family moving in next door. Such floor and ceiling effects have often been mentioned (Kuklinski, Cobb, & Gilens, Reference Kuklinski, Cobb and Gilens1997; Martinez & Craig, Reference Martinez and Craig2010; Redlawsk et al., Reference Redlawsk, Tolbert and Franko2010), but statistical techniques to detect and adjust for them have only been proposed fairly recently (Blair & Imai, Reference Blair and Imai2012) and implemented rarely. To minimize this risk, researchers should select foil statements that are negatively correlated with one another (e.g., making it legal for two men to marry and teaching intelligent design in school) to decrease the likelihood that respondents will be upset by all or none of the statements (Glynn, Reference Glynn2013; Janus, Reference Janus2010). But this has not always been done.

Another potential problem with the ICT is that respondents’ evaluations of the foil statements might be changed by offering the sensitive statement, rendering responses to the short and long lists incomparable (Blair & Imai, Reference Blair and Imai2012). Much research in psychology has shown that perceptions of objects change when perceived in relation to specific other objects (e.g. Higgins & Lurie, Reference Higgins and Lurie1983; Stevens, Reference Stevens1957). Such perceptual contrast effects can take two forms in an ICT. First, by adding a statement that people find not at all upsetting, the upsetting nature of the other statements may change. For instance, compared to a Black family moving in next door, pollution by a large corporation may seem even more upsetting. This would lead to an overestimation of prejudice in the ICT.

Second, the evaluation of the additional statement may change when it is presented in a list as compared to when it stands on its own. Knoll (Reference Knoll2013), for instance, found that 20 percent fewer Americans reported feeling that the American culture and way of life are threatened by foreign influence when asked in an ICT compared to an open question. Knoll (Reference Knoll2013) argued that this is because some groups feel the social pressure to overreport feelings of threat, but it could just as well have been that the foreign influence felt much less threatening in comparison to the pollution of large corporations.

Furthermore, for an ICT to effectively measure racial attitudes, the sensitive item must be an unambiguous measure of prejudice that cannot be rejected due to principled, non-race-related stances (Sniderman & Hagendoorn, Reference Sniderman and Hagendoorn2007). For instance, finding a government-implemented affirmative action policy upsetting could be due to prejudice against African Americans or due to a principled opposition to governmental intervention (Kuklinski, Sniderman, et al., Reference Kuklinski, Cobb and Gilens1997). Research on Islamophobia showed that negative attitudes toward Muslims are closely related to the rejection of Muslim practices such as wearing headscarves or Islamic education (Van der Noll, Reference Van der Noll2014). However, adding these practices to a list experiment could lead to an overestimation of negative attitudes toward Muslims because some groups of people reject such practices based on principles such as gender equality or separation of church and state without harboring prejudice toward Muslims (Adelman & Verkuyten, Reference Adelman and Verkuyten2020; Dangubić et al., Reference Dangubić, Verkuyten and Stark2020; Sleijpen et al., Reference Sleijpen, Verkuyten and Adelman2020).

In sum, the ICT is a promising approach to avoiding social desirability response bias if the technique is implemented well. However, the existing evidence for such social desirability bias with regard to prejudice should be taken with a pinch of salt due to floor and ceiling effects (Kuklinski, Cobb, & Gilens, Reference Kuklinski, Cobb and Gilens1997; Martinez & Craig, Reference Martinez and Craig2010; Redlawsk et al., Reference Redlawsk, Tolbert and Franko2010) and perceptual contrast effects (Blair & Imai, Reference Blair and Imai2012).

Bogus Pipeline

Another method used to eliminate social desirability bias is the bogus pipeline technique, wherein the researcher pretends to measure people’s “true” attitudes via their physical behavior (Jones & Sigall, Reference Jones and Sigall1971; Sigall & Page, Reference Sigall and Page1971). For example, in one study, some undergraduates were randomly assigned to answer questions orally, and other undergraduates answered the same questions under “bogus pipeline” conditions, meaning that they were attached to what they believed to be a lie detector (Jones & Sigall, Reference Jones and Sigall1971). The idea is that embarrassment of being caught lying weighs more strongly than the potential embarrassment of holding an undesirable attitude. In line with this reasoning, White participants said that various derogatory attributes were truer of “Negroes” under the latter condition than under the former (Jones & Sigall, Reference Jones and Sigall1971). A meta-analysis of thirty-one studies confirmed that the bogus pipeline technique indeed elicits more reports of undesirable attitudes and behavior (Roese & Jamieson, Reference Roese and Jamieson1993). This also applies to racial prejudice (e.g., Carver et al., Reference Carver, Glass and Katz1978; Plant et al., Reference Plant, Devine and Brazy2003). However, because of the necessity of a lab setup wherein people must be attached to a bogus lie detector, traditional bogus pipeline approaches are not well suited to measuring prejudice in representative samples of the general public.

More recent applications of the bogus pipeline technique circumvent the necessity of a lab setup. Instead of pretending to measure people’s true attitudes through their physical behavior, participants have simply been told that the study contains “sophisticated methods developed by psychologists” to detect deception, so lying is futile (Cohen et al., Reference Cohen, Jussim and Harber2009, p. 293). For example, participants in one study were told the following information about a question regarding helping behavior (“How often do you stop for stranded motorists? (never, rarely, sometimes, usually, always)”): “This question might appear innocent enough, but, in fact, it is one of many tools psychologists use to detect people who lie to create a positive impression of themselves. With the possible exception of policemen on patrol, NO ONE ‘usually’ or ‘always’ stops for stranded motorists. People who say they do are most likely lying.” This simple instruction led to a significant increase in antisemitic responses (Cohen et al., Reference Cohen, Jussim and Harber2009). Although not widely used and ethically problematic because of its deception, this approach could be used in national surveys. In fact, some researchers have already implemented it to reduce socially desirable responses in online surveys (e.g. Beattie, Reference Beattie2017).

Like the other techniques reviewed in this section, however, bogus pipeline research has uniformly relied on the “more-is-better” assumption in order to claim that it yields more accurate measurements, even though more reports of undesirable attitudes do not necessarily imply more accurate reports (Lelkes et al., Reference Lelkes, Krosnick and Marx2012). In fact, bogus pipeline techniques have been criticized for eliciting more undesirable responses only because of situational demands put on the participants (Arkin, Reference Arkin and Tedeschi1981; Ostrom, Reference Ostrom1973). Thus, it remains unclear if reports of more prejudice in a bogus-pipeline setup reflects more accurate reports.

Conclusions about Social Desirability

We conclude this discussion by returning to the question with which we began: Is social desirability response bias the explanation for any discrepancy between traditional explicit measures of prejudice and implicit measures? The research we have reviewed provides some evidence that socially desirable response bias can lead people to misrepresent their attitudes when they answer an explicit prejudice question. However, we have also identified a series of problems in the evidence that make socially desirable bias in prejudice measures less convincing than is generally assumed.

We have discussed alternative explanations for race-of-interviewer effects in studies that did not use random assignment of interviewers. Moreover, some of the race-of-interviewer effects in studies with random assignment can be explained by differential non-response, not social desirability bias. We have also shown evidence that higher reports of prejudice in survey modes that provided absolute privacy may be partially or largely caused by less accurate answering. Moreover, results of studies making use of the randomized response technique and the item count technique are not as diagnostic as one would hope. The fundamental problem here is that researchers have relied on the “more is better” rule rather than directly assessing the accuracy of reports collected by various different methods. Therefore, it seems prudent to conclude that the literature on social desirability response bias contaminating explicit reports of prejudice is suggestive but not conclusive. Clearly, more work is needed. But because the evidence reported here that explicit prejudice against African Americans does not seem notably less common than implicit prejudice suggests that social desirability response bias is not an obvious source of error in explicit reports.

Issues of Validity and Reliability

Some observers have expressed concern about the near zero correlations often observed between explicit and implicit measures of the same construct. Of course, such correlations are not expected to be 1.0, because that would imply that the measures are entirely redundant and that the implicit measures provide no information not already conveyed by the explicit measures. But most attitude theories suggest that explicit acknowledgments of prejudice are likely to have roots in implicit attitudes, so there should be non-negligible correspondence between the two.

In the four surveys we analyzed earlier, correlations between the implicit measures (the AMP and IAT) and the explicit measures (disliking, stereotypes, symbolic racism) varied between $r=.06$ and $r=.30$ (with most being about $r=.20$). These numbers are similar to correlations between explicit and implicit measures in meta-analyses, which varied between .20 and .24 (Cameron et al., Reference Cameron, Brown-Iannuzzi and Payne2012; Hofmann et al., Reference Hofmann, Gawronski and Gschwendner2005). Data from the Project Implicit website yielded a correlation of $r=.27$ with regard to race-related attitudes (Nosek, Smyth, et al., Reference Nosek, Greenwald, Banaji and Bargh2007). In contrast, correlations between explicit measures of prejudice tend to be notably stronger. For instance, the surveys we analyzed earlier yielded correlations mostly above $r=.55$.

The weak correlations between explicit and implicit measures have been explained by some observers as being due to social desirability bias contaminating the explicit measures (e.g., Nosek & Smyth, Reference Nosek and Smyth2007). In line with this argument, Nier (Reference Nier2005) found that race IAT scores were more strongly associated with answers to the explicit Modern Racism scale in a bogus pipeline condition than in a condition without the bogus pipeline. However, Phillips and Olson (Reference Phillips and Olson2014) showed that the weak correlation is not necessarily caused by social desirability bias in response to explicit prejudice questions. Their research suggests that the weak correlation might occur because people tend to deliberate about their responses to explicit prejudice questions, whereas implicit measures pick up people’s gut feelings.

Scholars have also argued that the dissimilarity of the tasks that participants perform when answering an explicit question and when completing implicit measures might explain the weak correlations between implicit and explicit measurements (Hofmann et al., Reference Hofmann, Gawronski and Gschwendner2005). For instance, implicit measures such as the IAT or the AMP are measures of a preference for one group over another (e.g., a preference for White faces over Black faces). However, studies often compared such implicit measures to explicit measures that did not assess a preference and instead tapped attitudes toward individual objects (e.g., Cunningham et al., Reference Cunningham, Nezlek and Banaji2004; Nier, Reference Nier2005). Another popular explanation considers implicit and explicit attitudes to be “related but distinct” constructs that explain different portions of variance in people’s decisions and behaviors (Bar-Anan & Vianello, Reference Bar-Anan and Vianello2018; Blair et al., Reference Blair, Dasgupta, Glaser, Mikulincer, Shaver, Borgida and Bargh2015; Nosek & Smyth, Reference Nosek and Smyth2007). From this perspective, the low correlation is to be expected and is not a reason to question validity.

A more problematic finding that raises serious questions about the validity of implicit measures of prejudice is the fact that different measures of implicit bias correlate extremely weakly with one another (Blanton et al., Reference Blair, Ma and Lenton2015; Bosson et al., Reference Bosson, Swann and Pennebaker2000; Brauer et al., Reference Brauer, Wasel and Niedenthal2000; Olson & Fazio, Reference Olson and Fazio2003; Rudolph et al., Reference Rudolph, Schröder-Abé and Schütz2008). In the 2008 ANES data, the correlation between the AMP and IAT was only $r=.18$. If different implicit measures tap the same underlying evaluation of social groups, why are the measures essentially unrelated to one another?

One common explanation for such weak correlations between implicit measures is random measurement error (Nosek, Greenwald, & Banaji, Reference Nosek, Greenwald, Banaji and Bargh2007). Such error seems inevitable, and, if random, will weaken correlations between measures (Cunningham et al., Reference Cunningham, Preacher and Banaji2001). One of the most common solutions to this problem is the application of multiple different measures to the same person. Scores on these multiple measures can be averaged so that random measurement errors cancel out, or the multiple measures can be used in latent variable structural equation models, which do an even better job of eliminating attenuation of relations due to random error. But one of the great strengths of the popular implicit attitude measures is that they each combine large sets of measurements to yield total scores. Thus, random measurement error in individual assessments seems unlikely to explain the strikingly weak correlations between different implicit measures.

The weak correlation between implicit measures is particularly concerning given that implicit bias has long been presented as an automatic and relatively uncontrollable characteristic that is resistant to change (Greenwald & Banaji, Reference Greenwald and Banaji1995). Studies in the early 2000s challenged this conceptualization of implicit bias, showing that scores on implicit measures such as the IAT are reactive to small external changes (e.g. Barden et al., Reference Barden, Maddux and Petty2004; Blair et al., Reference Blair, Ma and Lenton2001; Boysen et al., Reference Boysen, Vogel and Madon2006; Dasgupta & Greenwald, Reference Dasgupta and Greenwald2001; Park et al., Reference Park, Felix and Lee2007; Wittenbrink et al., Reference Wittenbrink, Judd and Park2001). For instance, White college students manifested lower implicit racial bias on a race IAT in the presence of a Black experimenter than in the presence of a White experimenter (Lowery et al., Reference Lowery, Hardin and Sinclair2001). This context-dependency has led to a reconceptualization of implicit bias as a person’s attitude that is, just like attitudes in general, influenced by the context the person is in (Greenwald & Banaji, Reference Greenwald and Banaji2013; Jost, Reference Jost2019).

Although contextual influence may, at least partially, explain the low correlation between implicit measures, this raises a different question about the validity of such measures. The fact that the experimenter’s race leads to lower implicit bias scores (Lowery et al., Reference Lowery, Hardin and Sinclair2001) seems surprisingly similar to the race-of-interviewer effects discussed above (Anderson et al., Reference Anderson, Silver and Abramson1988b; Davis, Reference Davis1997; Schuman & Converse, Reference Schuman and Converse1971). Does this imply that implicit measures of prejudice are susceptible to social desirability response bias? If so, this would undermine one of the motivations to dismiss explicit measures of prejudice in favor of implicit measures.

Concerns have also been voiced about the unsatisfactory reliability of some measures of implicit bias. Ideally, a measurement instrument should lead to very similar conclusions if it is applied to the same people. However, research found low reliability of evaluative or affective priming, particularly when pictures of outgroup members were used (Banse, Reference Banse2001). Reliability was much better when the method used unambiguous group labels instead of pictures (De Houwer, Reference De Houwer, Petty, Fazio and Brinol2009). The reason for this low reliability might lie in the fact that priming methods do not make salient the category that is being primed (e.g., African Americans) and tap attitudes toward the used stimuli instead of the category (De Houwer, Reference De Houwer, Petty, Fazio and Brinol2009; Olson & Fazio, Reference Olson and Fazio2003).

Research has also shown that the most widely used implicit measure, the IAT, has only weak to modest test-retest reliability (Bar-Anan & Nosek, Reference Bar-Anan and Nosek2014; Blanton et al., Reference Blanton, Burrows and Jaccard2016; Devine et al., Reference Devine, Forscher and Austin2012). Over time, IAT scores have been found to be considerably less stable than scores on corresponding explicit measures (Gawronski et al., Reference Gawronski, Morrison and Phills2017). Similarly, the AMP also shows only modest test-retest reliability (Cooley & Payne, Reference Cooley and Payne2017). Whereas some consider this low reliability a central concern (Tetlock & Mitchell, Reference Tetlock and Mitchell2009), others have pointed out that this concern is based on the no longer widely endorsed perception of implicit bias as a stable individual trait (Jost, Reference Jost2019).

The view that implicit attitudes are stable individual characteristics changed in the early 2000s (Banaji, Reference Banaji, Roediger, Nairne, Neath and Surprenant2001; Cunningham et al., Reference Cunningham, Preacher and Banaji2001). However, most scholars continued to focus on differences between participants in terms of implicit attitude measurements and other variables – an approach that assumes that the implicit measures tap stable individual differences. Newer accounts suggest that implicit bias does not reflect chronically stable individual differences and is instead responsive to situational triggers (Payne et al., Reference Payne, Vuletich and Lundberg2017) or that there is a mix of stable and context-dependent components (Dentale et al., Reference Dentale, Vecchione and Ghezzi2019). From these perspectives, low reliability scores are expected.

Low reliability may thus not be a reason for serious concern if implicit bias is considered context-dependent, but it does raise questions about the informativeness of a person’s score on an implicit prejudice measure. For instance, after completing a race IAT on the Project Implicit website, a person may be told that he or she has a preference for White people over Black people. But, importantly, this result may be attributable to the context in which the measurement was made and not generalizable to other contexts. In other words, the person may be misled into believing they have a general disposition when no such thing exists. Some scholars have suggested treating implicit measurements like the results of blood pressure tests, in that multiple measurements made on the same person may produce different but equally valid results describing that person differently in different situations (Greenwald et al., Reference Greenwald, Brendl and Cai2019; Jost, Reference Jost2019). Unfortunately, most research does not follow this logic and continues to use single applications of implicit measures to assess implicit bias of participants.

A final concern about implicit measures comes from the fact that people are quite able to guess their scores on these measures (Hahn et al., Reference Hahn, Judd and Hirsh2014). This suggests that people are not in fact unaware of their implicit attitudes. That again challenges a fundamental assumption of the implicit bias perspective.

Issues with the Metric

Questions have been also raised about the scoring of responses in the IAT that may lead to an overestimation of implicit bias (Blanton & Jaccard, Reference Blanton and Jaccard2006). Because the race IAT is based on a comparison of reaction times to White and Black faces in tasks involving categorizing words as good or bad, it seems natural to equate equally fast reaction times with no preference for Black people over White people or vice versa (Greenwald et al., Reference Greenwald, Brendl and Cai2019). Interestingly, the zero point on the IAT matches the neutral point on explicit measures (Pasek et al., Reference Pasek, Tahk and Lelkes2009). But neutral behavior toward Black and White people corresponds to a point above zero on the IAT, a point normally interpreted as meaning pro-White preference (Blanton et al., Reference Blanton, Jaccard and Strauts2015). This suggests that the proportion of people manifesting anti-Black prejudice on the IAT is smaller than the proportion of people with positive scores.

This finding might lead one to prefer to focus on associations between implicit bias scores and other variables (Jost, Reference Jost2019). But this will not satisfy the interest in knowing how much implicit bias exists and how that quantity compares with the amount of explicit bias that is present. Furthermore, individuals want to know if they are prejudiced, and learning that requires interpreting absolute levels of implicit association test scores. Implicit bias scholars have routinely offered assessments such as that 68 percent of visitors of the Project Implicit website (Nosek, Smyth, et al., Reference Nosek, Greenwald, Banaji and Bargh2007) or 75 percent of White Americans (Greenwald & Banaji, Reference Greenwald and Banaji2013, p. 47) are implicitly biased against Black people. And the Project Implicit website invites people to take the test to learn if they are prejudiced. Thus, the desire for absolute scores seems powerful yet is now called into question. A related problem is the use of labels such as slight, moderate, or strong to characterize levels of implicit bias when those are arbitrary distinctions (Blanton & Jaccard, Reference Blanton and Jaccard2008). This raises the fundamental question of whether it is appropriate to inform study participants about their IAT scores (Blanton & Jaccard, Reference Blanton and Jaccard2006; Mitchell & Tetlock, Reference Mitchell, Tetlock, Lilienfeld and Waldman2017).

Sample Quality and Generalizability

As we have highlighted above, the vast majority of publications on implicit bias are based on data from non-probability samples that do not allow one to draw solid conclusions about the American adult population (or any other population). For example, many studies have explored correlates of implicit bias or interventions targeting implicit bias among college students. Since there are notable differences between students and non-students in terms of their scores on prejudice measures (Henry, Reference Henry2008), it remains unclear how well these findings generalize. Moreover, hundreds of other studies make use of voluntarily completed IATs from people who visited the Project Implicit website to learn about themselves. The collection of millions of measurements does not legitimize generalizations to any populations. Only a very few studies have administered implicit prejudice measures with representative national samples (Ditonto et al., Reference Ditonto, Lau and Sears2013; Kalmoe & Piston, Reference Kalmoe and Piston2013; Pasek et al., Reference Pasek, Tahk and Lelkes2009, Reference Pasek, Stark and Krosnick2014; Payne et al., Reference Payne, Krosnick and Pasek2010), so most of the findings in this literature remain to be tested with representative samples and should be, since there is profound interest in applying prejudice measures to understand the nation.

Links to Behavior

Another line of research has raised the critique that few studies have documented sizable correlations between measures of implicit bias and discriminatory behavior (Oswald et al., Reference Oswald, Mitchell and Blanton2013; Tetlock & Mitchell, Reference Tetlock and Mitchell2009). Many of the existing studies on this matter have relied on data from university students and have found only weak correlations that appeared only under specific conditions (Blommaert et al., Reference Blommaert, van Tubergen and Coenders2012). The number of studies that link implicit bias to real-world discrimination is increasing (Rooth, Reference Rooth2010; Van den Bergh et al., Reference Van den Bergh, Denessen and Hornstra2010), but the difficulty of implementing the measurement of implicit bias outside of the laboratory puts some limitation on their generalizability.

Meta-analyses of the existing literature (focusing on lab studies) have documented only weak to modest correlations $\left(r=.13\mathrm{to}r=.24\right)$ between implicit racial bias measured by the IAT and discriminatory behaviors (Greenwald et al., Reference Greenwald, Poehlman and Uhlmann2009, Reference Greenwald, Banaji and Nosek2015; Oswald et al., Reference Oswald, Mitchell and Blanton2015) and a modest correlation of $r=.28$ between sequential priming measures and behavior (Cameron et al., Reference Cameron, Brown-Iannuzzi and Payne2012). In addition, a meta-analysis concluded that interventions aimed at reducing implicit bias in various domains rarely lead to changes in behavior (Forscher et al., Reference Forscher, Lai and Devine2019). A systematic review found a preference for White over Black people among most physicians but little evidence of a relationship with the physicians’ decision making (Dehon et al., Reference Dehon, Weiss and Jones2017). These kinds of findings have led to criticism (Mitchell & Tetlock, Reference Mitchell, Tetlock, Lilienfeld and Waldman2017) of earlier claims such as “the automatic White preference expressed in the Race IAT is now established as signaling discriminatory behavior” (Greenwald & Banaji, Reference Greenwald and Banaji2013, p. 47).

However, another recent meta-analysis on the link between the IAT and intergroup behavior revealed that the low correlations can partly be explained by methodological shortcomings in the existing literature (Kurdi et al., Reference Kurdi, Seitchik and Axt2019). Many studies were underpowered, did not correct for measurement error, used variants of the IAT instead of the standard IAT, and had weak correspondence between the IAT and the criterion behavior. Among studies without these methodological shortcomings, the correlation between implicit attitudes and behavior was $r=.37$, which is certainly larger than that documented in earlier meta-analysis. However, these correlations were similar in size to correlations between explicit measures of attitudes and behavior (Kurdi et al., Reference Kurdi, Seitchik and Axt2019). Accordingly, the conclusion that implicit measures of prejudice manifest more predictive validity than do explicit measures seems unsustainable at present, as is the claim that implicit attitudes are a primary driver of discriminatory behavior generally. However, this particular meta-analysis has been criticized for applying selection criteria and coding schemes that are insufficiently rigorous (Blanton & Jaccard, Chapter 12, this volume), so more work is needed on this issue.

Based on dual-process theories of cognition (Olson & Fazio, Reference Olson, Fazio, Petty, Fazio and Briñol2008; Petty et al., Reference Petty, Briñol and DeMarree2007), some scholars have argued that implicit prejudice measures are particularly well suited for predicting attitudes and behaviors that are automatic and less so for attitudes and behaviors that are more deliberate (see Blair et al., Reference Blair, Dasgupta, Glaser, Mikulincer, Shaver, Borgida and Bargh2015; Dovidio et al., Reference Dovidio, Kawakami and Gaertner2002). In line with these arguments, a meta-analysis found stronger correlations between implicit priming measures and behavior in studies in which participants had low motivation or low opportunity to deliberate their behavior (Cameron et al., Reference Cameron, Brown-Iannuzzi and Payne2012). Thus, the value of implicit measures for understanding prejudice may be limited to conditions of automatic thinking and action.

Can Implicit Bias Be Changed?

Decades of research in political science, sociology, and psychology have shown how difficult it is to change opinions, and the slow decline of explicit anti-Black attitudes during the past sixty years illustrates how slowly prejudices typically change. If bias toward a social group is unconscious, changing this attitude may be even harder. However, there are a lot of well-meaning efforts around the world devoted to teaching people how to reduce their implicit and explicit bias, even though there is little evidence of how this can be done. Yet recent developments give hope that difficult attitude change may sometimes be attainable. For example, attitudes toward gay marriage have changed from a majority of Americans opposing it to a vast majority being in favor of it. And a recent study showed that attitudes toward transgender people can be lastingly changed through a relatively brief intervention (Broockman & Kalla, Reference Broockman and Kalla2016).

An increasing number of studies have explored how implicit prejudice can be reduced. Research from lab studies showed that exposing people to counter-stereotypical members of a negatively associated social group can weaken both explicit and implicit racial bias (Dasgupta & Greenwald, Reference Dasgupta and Greenwald2001). For instance, exposure to Barack Obama during the 2008 election cycle was associated with lower levels of implicit bias against Black people (Plant et al., Reference Plant, Devine and Cox2009). The problem in the current political climate – at least in the United States – is that people who harbor bias against minorities are more likely to be exposed to reinforcing negative information about members of these groups due to selective use of social media. This may fortify implicit bias, because negative behaviors by members of a social group influence implicit evaluations of a person from that group more strongly than do positive behaviors (Ratliff & Nosek, Reference Ratliff and Nosek2011).

More recent research has examined intervention programs attempting to reduce implicit bias. One study found that a combination of raising awareness of implicit bias and teaching bias reduction strategies for daily life reduced implicit racial bias (Devine et al., Reference Devine, Forscher and Austin2012). An imagined intergroup contact intervention has also been found to reduce implicit bias toward immigrants among school children (Vezzali et al., Reference Vezzali, Capozza and Giovannini2012). Moreover, a program that trained participants to use meaningful negation (“that’s wrong”) instead of simple negation (“no”) reduced implicit racial bias significantly (Johnson et al., Reference Johnson, Kopp and Petty2018). And a diversity training program reduced implicit bias against women in STEM among male (but not female) university faculty (Jackson et al., Reference Jackson, Hillard and Schneider2014).

Despite these promising results, many scholars are cautious about whether implicit attitudes can be changed easily. Many studies found that attempts to reduce implicit bias can be ineffective or even backfire and increase the bias (Dasgupta & Greenwald, Reference Dasgupta and Greenwald2001). Processes that have reduced explicit prejudices, such as increasing propositional knowledge, correction of prior information, and raising awareness of the likely consequences of information, have no effect on implicit bias (e.g. Petty et al., Reference Petty, Tormala and Brinol2006; Teachman et al., Reference Teachman, Gapinski and Brownell2003). Another strategy that has been deemed ineffective and sometimes even increases implicit bias is giving individuals global nonspecific instructions (e.g., “don’t be biased” Frantz et al., Reference Frantz, Cuddy and Burnett2004; Kim, Reference Kim2003). Nonetheless, a highly publicized meta-analysis concluded that interventions can change implicit bias, though the effects are generally weak (Forscher et al., Reference Forscher, Lai and Devine2019). Importantly, changes in implicit bias did not necessarily translate into changes in behavior.

If implicit bias is a reflection of the cultural institutions in a society at a given moment in time (Payne et al., Reference Payne, Vuletich and Lundberg2017), this bias may change as the cultural institutions change (Hardin & Banaji, Reference Greenwald and Banaji2013). Similar to changes in explicit prejudice (Krysan, Reference Krysan2011), research found that implicit biases with regard to race, skin-tone, and sexuality have become more egalitarian during the last decade (Charlesworth & Banaji, Reference Charlesworth and Banaji2019). Similar changes have not been observed with regard to obesity, age, or disability – perhaps because societal debate has centered less on discrimination with regard to these characteristics. Thus, implicit bias may be subject to change not as the result of training programs but rather because of changes in cultural institutions, the same institutions that are thought to give rise to implicit bias in the first place.