Scholars of race have stressed the importance of thinking about race as a multidimensional construct, yet research on racial inequality does not routinely take this multidimensionality into account. We draw on data from the U.S. National Longitudinal Study of Adolescent Health to disentangle the effects of self-identifying as black and being classified by others as black on subsequently being arrested. Results reveal that the odds of arrest are nearly three times higher for people who were classified by others as black, even if they did not identify themselves as black. By contrast, we find no effect of self-identifying as black among people who were not seen by others as black. These results suggest that racial perceptions play an important role in racial disparities in arrest rates and provide a useful analytical approach for disentangling the effects of race on other outcomes.
Most Americans think about racial difference the way U.S. Supreme Court Justice Potter Stewart thought about pornography: we know it when we see it. Similarly, we talk about racial discrimination as if the process and its likely targets were obvious. However, operationalizing these concepts in research presents an intriguing dilemma: Just who is at risk of experiencing racial discrimination in the United States? On its face, the question seems almost silly; the history of oppression and racial domination of African Americans, Native Americans, Asian Americans, and Latinos is well known and well documented (e.g., Takaki 1994). The current legacy of covert racism, stereotyping, and implicit prejudice has also received significant attention (e.g., Bonilla-Silva 2010; Brubaker et al. 2004; Greenwald et al. 1998). Yet, research has also demonstrated that race is a multidimensional construct that is not nearly as straightforward to measure as most people assume (e.g., Bailey et al. 2014; Goldstein and Morning 2000; Roth 2010; Telles and Lim 1998).
Demographic research typically measures race in one of two ways. Historically, racial data in the United States were collected by survey interviewers or census enumerators, who classified respondents during face-to-face interactions (Snipp 2003). More recently, and coincident with changes toward mail-back, telephone, and Internet surveys, research has tended to focus on self-reports of racial group membership. Rather than trying to determine whether one method or the other yields a better measure of race, we argue that self-identification and classification by others represent different dimensions of race that might also have different consequences for the life outcomes of individuals (for further discussion, see Saperstein 2012). Although many Americans racially identify in ways that are congruent with how they are classified by others, understanding how different measures of race are related to inequality is important because it offers insight into the processes that produce racial disparities. For example, if racial inequality is primarily a function of how individuals identify themselves, we might expect factors like oppositional culture to be driving inequality (cf. Fordham 2008); if inequality is primarily a function of how individuals are classified by others, factors such as contemporary discrimination are likely to play a larger role (cf. Pager 2003).
To illustrate, we take a novel analytical approach to the standard demographic study of racial disparities that seeks to disentangle the effects of racial self-identification and classification by others. In particular, we extend research on racial disparities in arrest rates (Kirk 2008), and the role of bias in producing these differences (Kochel et al. 2011), by examining how having been classified as black affects arrest rates among individuals who did not self-identify as black. We then compare this estimate with the effect of self-identifying as black among individuals who were not classified as black. Our approach sheds light on the relative importance of these two dimensions of race for arrest and speaks to whether racial disparities in arrest rates are related to how individuals are seen by others, how they identify themselves, or some combination of both.
Data and Methods
Disentangling how later-life outcomes are related to how individuals self-identify and are racially classified by others requires longitudinal data that include both measures of race. We use data from Waves 3 and 4 of the National Longitudinal Study of Adolescent Health (Add Health), one of the few national surveys to collect data on the respondent’s racial self-identification and the interviewer’s racial classification of the respondent at the same point in time. Add Health follows a nationally representative sample of Americans who were enrolled in grades 7–12 in 1994–1995 (Wave 1); Wave 3 data were collected in 2001–2002, when respondents were ages 18–26, and Wave 4 data were collected six years later when respondents were ages 24–32. The longitudinal aspect of these data allows us to observe how one’s racial classification or identification at one point in time is related to the likelihood of reporting a subsequent arrest. This time ordering is also important to consider because previous research suggests that contact with the criminal justice system shapes subsequent racial classification and self-identification (Saperstein and Penner 2010; Saperstein et al. 2014). We thus restrict our analyses to respondents who did not report an arrest in Wave 3, and examine the 10,733 cases that had complete information on Wave 3 race and Wave 4 arrest. Of these respondents, 18 % reported an arrest in Wave 4.
We predict the odds of respondents reporting their first arrest between Wave 3 and Wave 4, focusing on differences between respondents’ racial self-identification and their racial classification by the interviewer, as recorded in Wave 3. Interviewer classification of the respondent’s race was recorded at the end of the Wave 3 interview, and interviewers were instructed to code the respondents’ race from their own observation (Harris 2009). We use the measure as a proxy for how respondents are likely to be racially classified more generally. Interviewers selected one option from the following categories: “white,” “black or African American,” “American Indian or Alaska Native,” and “Asian or Pacific Islander.” Respondents used the same categories to self-identify in Wave 3, but they were allowed to give multiple responses; respondents who chose more than one category were also asked to choose which category best described them. We use this “best race” self-identification measure to better match the forced-choice format of the interviewer’s classification, although supplementary analyses using an indicator for giving any black response provide similar results.
Table 1 reports the cross-tabulation of Wave 3 self-identified and interviewer-classified race. We focus on the black–nonblack divide in our analyses because the number of respondents in many cells of a more detailed cross-tabulation is relatively small (e.g., fewer than 10 respondents self-identified as black but were classified as Asian). Table 1 shows, as expected, that the vast majority of respondents were classified and identified consistently. However, we believe the discrepant cases provide important information not because the number of discrepancies is large, but because the discrepancies provide leverage on whether one dimension of race or the other seems to be driving racial disparities. That is, our contribution hinges not on the magnitude of the effects but on the substantive or theoretical insights our approach reveals. In this case, we aim to highlight the effect of racial classification on arrest by comparing the 33 respondents who were classified as black by the interviewer but did not self-identify as black, and the 45 respondents who self-identified as black but were not classified as black by the interviewer, with their peers who both self-identified and were classified as nonblack.
We first estimate a logistic regression model predicting whether people who were classified as black by the interviewer had higher arrest rates than those who were not classified as black, among respondents who did not self-identify as black in Wave 3.1 We then examine the converse, restricting the sample to respondents who were not classified as black and examining whether people who self-identified as black were more likely to be arrested than those who did not self-identify as black.2 For both approaches, we estimate additional models that add controls for a wide range of information collected in Waves 1 and 3, including whether the respondent reported a Hispanic origin or multiple races in either wave, whether the respondent was classified or identified as “other” in Wave 1,3 as well as the respondent’s nativity, gender, marital status, level of education, age, parental education, and the self-identified race of the Wave 3 interviewer (see Tables 2 and 3 in the appendix). We also include controls for a variety of Wave 3 neighborhood characteristics, including the county logged adult arrest rate (per 100,000); logged median house value at the census block level; and the proportions of census-classified blacks, census-classified Latinos, people living below the poverty line, and adults over age 25 with less than a high school diploma.4 To account for both prior delinquency and any delinquent acts that occurred between Wave 3 (when race was measured) and Wave 4 (when arrest was measured), we used principal components analysis to construct one delinquency index using variables from Wave 3 and one using variables from Wave 4. By controlling for whether the respondent reported engaging in activities such as damaging property, using weapons for stealing, and getting into physical fights, we hope to rule out that the differences we observe in arrest rates might be attributable to differences in the degree to which respondents engaged in behavior that increased the likelihood that they would be arrested.5
Figure 1 depicts estimates from a pair of logistic regression models predicting an arrest between Waves 3 and 4. The first bar in Fig. 1 shows that being seen as black by others is associated with a sizable and statistically significant increase in the odds of experiencing a future arrest (odds ratio = 2.78, p < .01), even among individuals who do not identify themselves black. By contrast, the second bar of Fig. 1 shows that the effect of self-identifying as black is close to null and not significant among respondents who were not classified by the interviewer as black (odds ratio = .97, p = .96). In interpreting this finding, it is important to note that there are more respondents who self-identify as black but are not seen as black than there are respondents who are seen as black but do not self-identify as black (see Table 1). This suggests that the absence of a significant effect on arrest for self-identifying as black, among respondents who are not classified as black, cannot simply be attributed to a lack of statistical power.
We provide insight into one of the mechanisms behind racial disparities in arrests through a novel approach that leverages cases of inconsistent racial self-identification and classification by others. Our findings suggest that racial disparities in arrest rates in the contemporary United States are more closely related to how young adults are perceived racially by others, relative to how they identify their own race. We urge caution in interpreting these results because of the small cell sizes involved, and we note that this pattern of appearance mattering more than self-identification might not hold for other outcomes; determining which dimensions of race matter most in which contexts remains an important task for future research. Nevertheless, our study illustrates the added value of using multiple measures of race simultaneously in the same study (rather than treating one measure or another as the universal gold standard) and joins a growing literature that demonstrates the multidimensionality of race. We encourage demographers—and scholars of social inequality—to think more strategically about how to measure race to better reflect their specific research question and the mechanisms they expect might be creating or maintaining racial disparities.
The authors are grateful to the Russell Sage Foundation, the UC Center for New Racial Studies, and Stanford's Institute for Research in the Social Sciences for support, to Jessica Kizer for research assistance and to Sara Wakefield for useful comments and discussions. A previous version of this article was presented at the annual meeting of the Population Association of America, San Francisco, May 2012.
Given the relatively small numbers of respondents who were racially classified and identified differently, we report additional analyses in Online Resource 1 using a variety of alternative estimation strategies for our standard errors (all of which yield substantively similar results).
We obtained similar results from a model comparing respondents who were neither classified nor identified as black with those who (1) were both classified and identified as black, (2) were classified as black but did not identify as black, and (3) were not classified as black but identified as black.
We add these controls to account for concerns about survey design effects and the consistency of self-identification across waves (see Cheng and Powell 2011).
Models with and without controls yield similar results. We prefer the model without controls out of concern for data sparseness. We do not constrain our sample sizes to be equal across models because of the number of respondents who have missing data on control variables.
Our models do not account for other situational factors that also predict arrest because Add Health does not record detailed information about the circumstances surrounding the arrest (e.g., the evidence on which the arrest was based, the arrestee’s demeanor, and whether the arrestee complied with police commands). Future work examining the relationship between these characteristics and racial perceptions would be of great value for understanding the relationship between race and arrest (cf. Goff and Richardson 2012; Kochel et al. 2011).