Racial Rigidity in the United States: Comment on Saperstein and Penner

Previous articleNext article FreeCommentary and DebateRacial Rigidity in the United States: Comment on Saperstein and Penner1Rory Kramer, Robert DeFina, and Lance HannonRory KramerVillanova University Search for more articles by this author , Robert DeFinaVillanova University Search for more articles by this author , and Lance HannonVillanova University Search for more articles by this author PDFPDF PLUSFull TextSupplemental Material Add to favoritesDownload CitationTrack CitationsPermissionsReprints Share onFacebookTwitterLinked InRedditEmailQR Code SectionsMoreTo conserve space for the publication of original contributions to scholarship, the comments in this section must be limited to brief critiques; author replies must be concise as well. Comments are expected to address specific substantive errors or flaws in articles published in AJS. They are subject to editorial board approval and peer review. Only succinct and substantive commentary will be considered; longer or less focused papers should be submitted as articles in their own right. AJS does not publish rebuttals to author replies.As many sociologists have pointed out, the system of racial categorization in the United States is historically unique (Davis 2001; Morning 2008). One oft-discussed example of U.S. exceptionalism is the rigidity surrounding the infamous "one-drop rule" and the notion that all persons with any African ancestry are black. In countries without that historical legacy there is much less formality in racial categorization and thus a greater potential for individuals to change racial classifications as they move through the life course.2In "Racial Fluidity and Inequality in the United States" (American Journal of Sociology 118 [3]: 676–727) Saperstein and Penner (2012; henceforward S&P) call into question the assumed high level of rigidity in racial categorization in the United States. They assert instead that the racial rigidity that social science has long documented and analyzed is mistaken (or has changed) and that, regardless of phenotype, the probability that an individual will identify or be seen as white or black can be significantly altered by undergoing raced status experiences such as going to prison, falling into poverty, or utilizing public assistance.One implication of S&P's findings is that white privilege in the United States is not as durable as commonly believed, since an individual's likelihood of being seen as white, and of enjoying all of the privileges that go along with that racial status, can be significantly harmed by one's missteps and misfortunes. Conversely, members of racial minority groups can increase the likelihood that they will be classified and treated as white by avoiding such things as involvement with the criminal justice or public welfare systems.While for most of sociology's history there has been clear agreement that race in the United States is best thought of as an ascribed master status, S&P challenge this common wisdom and claim that fluidity is "an integral part of the 'social invention' … that is race in the United States" (p. 680). Moreover, S&P suggest that their results are not "driven by the small minority of respondents for whom racial perceptions are thought to be more fluid and complex" (p. 686) but rather can be generalized to the average person.We appreciate S&P's basic point that racial categorization is a social construction and thus theoretically subject to change. However, S&P's article presents a provocative thesis regarding widespread racial fluidity across the life course, but on close inspection it does not provide evidence that racial change is a common experience across all racial and ethnic groups nor that social status experiences cause individual racial change. Contrary to S&P's claim that their modeling strategies "demonstrate consistent evidence of the reciprocal relationship between racial fluidity and inequality in the United States" (p. 687), we argue that none of their modeling strategies disentangle the well-documented effect of race on social status from their newly proposed effect of social status on racial categorization. While S&P argue that the robust correlations between race and social status variables reflect how racial groups are stereotyped, these relationships could simply reflect the enduring structural reality of racism's impact on life chances. For example, the observed statistical associations between the status variables and black self-identification do not necessarily tell us that people are more likely to see themselves as black if they experience a disfavored social status. They only reveal what sociologists have long known—that black people are more likely to suffer poverty, incarceration, unemployment, and other negative social outcomes.S&P incorrectly assume that random error in the survey would make it "more difficult, perhaps even impossible, to find evidence of the expected relationship between social position and racial classification" (p. 689). In fact, that error is exactly what enables their results, as a flawed control variable can lead to unfounded conclusions regarding the theoretically important independent variables in the model. In this case, even slight error in the measurement of prior racial classification/identification can lead to unwarranted conclusions about the causal impact of social position on racial categorization. After demonstrating how their findings could be driven by measurement error, we more directly test S&P's causal mechanism by exploiting variability in an interviewer's knowledge about a respondent and find no evidence of a reciprocal relationship between an individual's social status and an observer's assessment of that individual's race. In the United States, race continues to be best conceived of as an ascribed master status.Results Despite or Because of Measurement ErrorS&P present two alternative readings of the observed discrepancy in racial identification over time—either a social constructivist view in which fluidity is part of the unstable equilibrium of U.S. racial divisions or a "primordialist" view in which fluctuation is "generated by poor question wording and limited answer options or is an issue of comprehension. … This [primordialist view] implies that clearer questions or better categories would eliminate the fluidity we observe and that eliminating these inconsistencies is desirable" (pp. 681–82). Such framing presents a false dichotomy. Acknowledging that the NLSY survey includes flawed measures of racial identity that inaccurately map to common understandings of racial categories does not imply a belief in primordialism; rather, it rightly recognizes that question wording matters. Social constructivism does not negate the need for clear racial categories on surveys, nor does everything social scientists observe have deep theoretical implications. Sometimes, observed data variation is just measurement error.A closer inspection of the NLSY data raises important issues about its suitability for S&P's theoretical interests. They claim that fluidity is surprisingly widespread because 20% of the NLSY sample experiences at least one change in white/black/other racial classification over two decades. We have serious concerns about presenting an analytic model based on those survey options because "other" is not a racial identity; it is the absence of measured identity, and it is disproportionately involved in the classification changes observed by S&P.3 While the vast majority of those initially classified as white or black were consistently classified that way across the full 17 survey years, only 1% of the respondents who were classified as "other" in 1979 consistently received that designation. Furthermore, the 20% overall rate of change identified by S&P is more accurately described as over 85% for the Hispanic population and 8% for the non-Hispanic population. We believe this is not due to racial fluidity, per se, but rather a flawed classification scheme given to interviewers with inadequate directions on its use. In sum, we believe that the bulk of racial change identified by S&P is driven by unclear survey categories ("other") or alternative forms of measurement error (about two-thirds of the non-Hispanic changers experience only one change across the 17 survey years).S&P could argue that their appendix table A2 (p. 718) demonstrates that the effects of the intervening social status variables on racial classification were significant even when filtering out respondents that self-identified as Hispanic, Native American, or multiracial in 1979 ("populations with high, but theoretically fixed, propensities toward ambiguity"; p. 707).4 Moreover, it could be that data noise, whether due to categorical ambiguity or human error, would only serve to make it harder for them to find statistically significant status effects were the noise randomly distributed (p. 689). We argue that while it is true that noise in the status variables decreases the likelihood that the status effects will be statistically significant, noise in the key control variable, the racial category selected on the prior survey, does the opposite.To illustrate this point, the first column in Table 1 provides estimates from a logistic regression analysis of black self-identification of the type used by S&P. We only include the social status variables that we were able to accurately match with the values in S&P's descriptive statistics table (p. 694). The point of our analyses is not to exactly replicate S&P's coefficients but rather to illustrate how the modeling strategy of controlling for prior racial categorization cannot distinguish between random error and status driven fluidity. This is true for all of S&P's models in the discussion under scrutiny here, regardless of the specific status variables in the model, the use of demographic controls, and adjustments for missing data.Table 1. Logistic Regression of Status Effects for Black Self-identification in 2002 Under Different Assumptions About Fluidity versus Random ErrorVariableEstimates controlling for the original black in 1979 measure used by S&P to model status-driven fluidityEstimates controlling for a created black in 1979 measure that is different from black in 2002 only via induced random errorEver incarcerated…..28.33(.35)(.27)Ever unemployed>4 months…..75***.77***(.20)(.15)Ever below poverty line…..95***.91***(.24)(.17)Ever received welfare…..29.30(.21)(.16)Ever graduated college….−.27−.20(.25)(.18)Ever married as a teen….−1.07**−1.46***(.35)(.26)Ever teen parent…..89***.97***(.29)(.23)Lives in an inner city….1.00***1.04***(.27)(.21)Lives in a suburb…..074.047(.26)(.19)Black Self-ID, 1979….8.15***6.73***(.22)(.15)N….7,7187,718Note. Numbers in parentheses are SEs. In both the original and the simulation data the same percentage of the sample (1.6%) has a discrepancy in black racial identification for 1979 and 2002. All variables follow S&P's coding instructions and match the approximate means S&P provide in their Table 1 (p. 694).*. P < .05.**. P < .01.***. P < .001.Ed. note. AJS received updated estimates for the second column of table 1 after the contributions to this exchange were finalized for print. In the interests of completeness and accuracy, this updated information is available as an appendix to this comment; these estimates do not represent a substantive change in the authors' analysis but do alter the magnitude of the output slightly.View Table ImageConsistent with S&P's results (p. 699), we find highly statistically significant independent effects for poverty, long-term unemployment, ever married as a teen, ever a teen parent, and currently living in an inner city on racial identity (Table 1, col. 1). However, these results do not demonstrate the existence of theoretically meaningful fluidity where social status is systematically related to changes in racial identity. Instead, as long as there is a small amount of random error in the 1979 racial measurement that differs from the random error in the 2002 racial measurement, the social status variables will simply exhibit their familiar structural relationships with race (e.g., white people are less likely to experience incarceration) that S&P interpret in the reverse direction (incarcerated individuals are less likely to be seen as white).To show that S&P's results can be reproduced without theoretically meaningful fluidity we present the results of a simulation analysis in the second column of Table 1. We first assume that there is no individual racial change and set each individual's racial identity in 1979 equal to the corresponding 2002 self-identity. At this point, by definition, if we were to regress the 2002 identity on the 1979 identity it would explain 100% of the variation. We then randomly change the 1979 black/non-black status of 1.6% of the sample, the same percentage of individuals who actually changed in the NLSY. We then replicate this procedure 100 times, estimating the baseline logistic regression model in Table 1 after each random assignment of 1979 racial self-identities.The results of the simulation are striking: averaging the 100 sets of estimated coefficients and standard errors for each of the social status variables reveals that we can reproduce S&P's basic pattern of results (shown in Table 1) even though by construction the changes or nonchanges in race between 1979 and 2002 are random. As such, S&P's methodology is fundamentally incapable of distinguishing between the impact of racial categorization on social status and the reverse. Their inclusion of prior racial categorization as a key control variable is, in practice, irrelevant for their theoretical concerns, and the residual correlations tell us nothing about their central causal claim. In sum, S&P's results do not depend on the existence of theoretically meaningful fluidity; they can be achieved simply by introducing a sufficient amount of random noise in their primary control variable (here, the respondents' self-reported race when they were 14–22 years old).5We do not contend that the difference between the two racial measurements is always the product of random error.6 In fact, we expect that nonrandom measurement error plays the same role. The point of the simulation analysis is that, consistent with the logic behind statistical significance testing, it is important to first rule out random error. Because S&P do not, the models merely report well-established correlations between race and social status with the variables flipped from one side of the equation to the other. For example, when we simply move the ever below the poverty line variable to the other side of the equation and use the race the respondent self reported as an adult as the predictor, we unsurprisingly find the same significant association (see Table 2). In interpreting this significant relationship, it is important to keep in mind that just because an individual has his or her racial identity officially recorded after an event has occurred does not mean that the person lacked that racial identity in the years leading up to the event; race existed and influenced these respondents before they entered the survey and anyone asked them about it.Table 2. Logistic Regressions Under Different Assumptions about Causal DirectionVariableBlack Self-ID 2002Ever in Poverty 1979–2002Ever in poverty, 1979–2002….1.42***…(.23)Black Self-ID, 2002….…1.42***(.23)Black Self-ID, 1979….8.22***−.07(.21)(.23)N….7,7187,718Note. Numbers in parentheses are SEs. All variables follow S&P's coding instructions and match the approximate means S&P provide in their Table 1 (p. 694).*. P < .05.**. P < .01.***. P < .001.View Table ImagePresumably to buttress their causal claim, S&P report the results of Granger causality estimates in their appendix (table A6).7 However, it is widely accepted that Granger causality tests cannot reliably reveal true structural causality.8 For example, they cannot overcome a basic concern in gauging causality, whereby a third factor could be driving both the dependent variable and the variable in question. In this case, a likely third factor is the racial identity and raced experiences that an individual had before he or she was ever contacted by the NLSY. More fundamentally, S&P delineate a very particular causal mechanism: social status alters racial self-perception and the perceptions of others via internalized racial stereotypes. To establish causality, S&P must provide evidence directly related to that specific causal mechanism. In the following section, we present two direct tests of their causal mechanism using NLSY data.Possible Ways to Test S&P's Causal MechanismShowing that random error can create similar findings to S&P and that Granger causality is very different from actual causality does not mean that S&P's thesis is untestable with the NLSY data. Building on an idea S&P present in a footnote, we offer below two novel approaches to testing S&P's causal argument directly. For the specific variables we examined with these approaches (illicit drug use, illicit drug sales, and prison record), none of the analyses support S&P's thesis.S&P provide only one direct test of their argument that an interviewer's racial classification could be "colored by the respondent's answers" (p. 688). They point out that interviewers heard the respondent racially self-identify in 1979 before classifying that respondent as white/black/other. To test whether hearing the respondent's answers to the racial identification question leads to bias in an interviewer's classification S&P compare the level of consistency between racial identification and classification in 1979 with the level of consistency between 1979 racial identification and 1980 racial classification. The logic underlying this test is that in 1980 the racial self-identity question was not asked, and, presumably, the interviewer would not know or remember the youth's responses from the previous year. Noting that statistical tests fail to find any difference, S&P concluded that "the interviewers' hearing the respondents' self-identification in 1979 did not significantly influence their classification" (p. 688). An interesting aspect of this conclusion is that it implies that interviewers' racial classifications will not be influenced by hearing a respondent specify a black identity, yet will be influenced by a respondent mentioning a status indirectly associated with black identity, such as having one's first child as a teenager.9 For the sake of this comment, the most important aspect of this conclusion is that it points to the possibility of better tests of S&P's causal mechanism that exploit myriad variation in the data concerning the interviewer's level of knowledge about the respondent's answers.Question Variation in Different Survey WavesIn 1980 (but not in 1981) the NLSY asks a series of questions about the respondent's criminal behavior. Thus, extending the logic of S&P's test of the influence of the 1979 racial identification question, one could examine whether survey items about a respondent's criminal behavior matter for interviewer racial classification, relative to years when the questions were not asked. Here, we would be comparing the impact of reported criminal behavior in 1980 on racial classification in 1980 with the impact of reported criminal behavior in 1980 on racial classification in 1981. While S&P only include variables in their models where racial stereotypes reflect and exaggerate an empirical reality (i.e., black people are more likely to be perceived as having a prison and black people are more likely to have a prison record), an alternative would be to examine stereotypes with no in would the effects of racial stereotypes on an interviewer's classification from the effects of structural respondent's of drug provides an to test the of stereotypes for racial classification. While there are stereotypes of black as more likely white to use survey evidence consistently this widespread belief and while black are as being much more involved in the of the reality is very different in the of the most widely used illicit and logistic regression analyses of the relationship between self-reported drug and interviewer racial classification for 1980 survey when the drug questions were and the drug questions were not interviewers were more likely to an individual as black after hearing of illicit drug use or that would support S&P's and following S&P's if there is no difference between the coefficients for 1980 and one can that hearing a respondent's self-reported drug does not significantly influence an interviewer's results are in Table For each model, the coefficients associated with racial classification are statistically across both In and to respondent use actually the likelihood of white classification, regardless of whether the interviewer was about the In fact, drug use of any type the likelihood of white classification across both years in statistically while use or was associated with a likelihood of black classification. The association between black classification and drug was in both years, but the difference between the two coefficients is not statistically Thus, it that a respondent's answers to the drug questions do not significantly the interviewer's racial of Effects on by the About the or in in other in other in Numbers in parentheses are interviewer SEs. All variables are and all models are estimated with logistic data are using P < .05.**. P < .01.***. P < Table in their analytic S&P of their own such as ever or ever by responses across survey from questions that were only in a particular year. This type of variable construction has for testing S&P's causal mechanism. While in the example regarding racial impact on classification S&P to assume that the interviewer in 1980 would not have to 1979 the to have been for many of S&P's The that, for example, an interviewer in would know about a respondent's poverty status from when it was not on the in of is with both common survey and from NLSY and interviewers did not have any to information about a respondent's previous social experiences from answers given in However, it is that in where an interviewer had the respondent in previous survey interviewers remember the respondent's While such a possibility to us for all of S&P's ever measures (e.g., an interviewer a respondent being for 17 or more years we could see the possibility for the durable of an S&P argue that in decreases the respondent's of being seen as white and the respondent's of being classified as black by the interviewer any (p. that about of the respondents in the NLSY were a interviewer in any given provides yet way to test S&P's S&P are the relationship between ever and racial classification should be significantly when the interviewer is and thus had no to information about years Conversely, following S&P's the effect should be most when interviewers have with a respondent's test S&P's causal we follow the basic of S&P's models and logistic in which a respondent's classification as black or white on the previous racial classification and whether a respondent was ever We these basic models for two one in which the respondent has been by the same interviewer before and one in which the interviewer has the respondent and thus has no information about a respondent's history the items on the in of the In both the sample is to those respondents who were not the from the logistic are in Table As in S&P's the effect of ever being was highly statistically significant and in the expected direction for the sample of respondents that had the same interviewer the effect was present for the sample of respondents that had the interviewer More the effects were statistically between the two Logistic Regression for the of as or Black are from Numbers in parentheses are interviewer SEs. is given in All models include race of respondent and data are using P < .05.**. P < .01.***. P < Table in of an interviewer hearing a respondent's drug or that a respondent was in prison, our results offer no support for the notion that the observed associations with racial classification are due to the causal mechanism described by where an interviewer's is by the respondent's answers" (p. 688). While the of space us from significantly other the data offer more for S&P's For example, in to the full in which the respondent had their race observed at the in the to into the NLSY sample the respondent's race was classified before any questions were asked about and public one could exploit the variability in question between the two to test whether hearing a respondent from welfare for the observer's classification of the respondent as black own analysis that it did results are available on To test about racial fluidity we to use the that interviewers cannot be influenced by knowledge that, they do not to directly test S&P's causal mechanism fail to support their claim that social status variables exhibit their significant associations with race because of the of stereotypes. We that the correlations reflect structural relationships where racial categorization social S&P racial classifications by survey without whether and how those classifications then impact the analytic results 2001; and 2008). S&P an important question by how racial classification is related to status and but they do not the to answer that question. while other sociologists for the rigidity of racial they did not test their own causal mechanism nor the actual and of racial that racial fluidity into in of inequality is less a matter of (p. with this we find robust correlations between the same social status variables and race that can be in causal However, S&P go on to that they (p. and