Portal de Periódicos da CAPES

Still Searching for a True Race? Reply to Kramer et al. and Alba et al.

2016; University of Chicago Press; Volume: 122; Issue: 1 Linguagem: Inglês

10.1086/687806

1537-5390

Aliya Saperstein, Andrew M. Penner,

School Choice and Performance

Previous articleNext article FreeStill Searching for a True Race? Reply to Kramer et al. and Alba et al.1Aliya Saperstein and Andrew M. PennerAliya SapersteinStanford University Search for more articles by this author and Andrew M. PennerUniversity of California, Irvine Search for more articles by this author Corrections to this articleErratum for “Still Searching for a True Race? Reply to Kramer et al. and Alba et al.”PDFPDF PLUSFull Text 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.We wrote “Racial Fluidity and Inequality in the United States” (Saperstein and Penner 2012) with the aim of jump-starting a conversation about how race is best conceptualized in studies of stratification. Does assuming that people have static, often mutually exclusive, “races” help us understand disparities in the contemporary United States? Or are there inequality-sustaining mechanisms we might have missed by assuming a process of consensus through which racial categorizations are ascribed at birth and effectively fixed? That two teams of scholars took time to engage with our work is a positive sign that a much-needed conversation is happening. It is, of course, disheartening that Kramer et al. (2016) (hereafter KDH) and Alba et al. (2016) (hereafter AIL) come to the conclusion that we, at worst, misinterpreted our data and, at best, overstated our case. Nevertheless, we are grateful to have this opportunity to clarify our claims and intentions and to offer new evidence that returning to assumptions of rigid racial ascription is not the way forward.Our reply addresses the three main points of empirical critique across the two comments: (1) that a relationship between social status and racial categorization of similar direction and magnitude could be produced by measurement error; (2) that our findings are neither as common nor as generalizable as we claimed; and (3) that we did not adequately demonstrate that stereotypes, operating through what the interviewer does or does not hear about the respondent, are a key causal mechanism. We either provide evidence that directly refutes each point, explain why it results from a misreading of our argument, or both. In the process, we underscore our earlier findings with additional evidence of how social factors shape categorization: through selective processes of “ethnic attrition” as well as what interviewers knew about the respondents’ use of crack cocaine.First, a clarification on the scope of this debate: the details of our empirical analysis are not the only aspect of our article being subjected to scrutiny; also in question for both commenters is whether or not race and ethnicity should be conceptualized as stable individual characteristics. Given space constraints, we focus on the specific empirical critiques, but it is important not to lose sight of their theoretical foundations and implications. Although AIL and KDH give us credit for advancing a “sophisticated” and “provocative” thesis, we are not the first to conclude that recording these characteristics, and assigning individuals to categories, is a more complex task than racial commonsense and received practice would indicate. Thus, we encourage readers to weigh not only other empirical research that demonstrates the relationship between social status and racial categorization can run in both directions (Saperstein and Gullickson 2013; Saperstein, Penner and Kizer 2014; Young, Sanchez, and Wilton 2015; Simonovitz and Kezdi 2016) but also the broader body of theoretical work on the social construction of race and ethnicity.Can the Results Be Explained by Measurement Error?Both AIL and KDH raise concerns about how much of the racial fluidity we reported is “real” and how much is due to measurement error. AIL highlight this issue by dividing their descriptive statistics to show which respondents experienced one or two changes in racial classification compared to three or more. KDH claim that our original estimates for the effect of social status on racial categorization could result from introducing random variation alone.To counter these critiques, we begin with our preferred approach to potential coding mistakes in the data: we show that when we remove cases with classification discrepancies that might be considered “errors,” we get results similar to those we originally reported. We then consider the results presented by KDH (in their Table 1) and demonstrate that there is nothing in their approach that suggests our core finding can be attributed solely to measurement error. However, we also note that most discussions of measurement error assume the existence of an objectively verifiable value, as is the case for individual attributes like height or weight. To make an analogous assumption for race is problematic.Table 1. Predicting Racial Classification with and without Classification “Blips” BlackWhite Includes BlipsExcludes BlipsIncludes BlipsExcludes BlipsUnemployed…….34***.41***−.29***−.30***Impoverished…….35***.25*−.36***−.39***Incarcerated…….29+.27−.25**−.31**Received welfare…….15*.20−.13***−.15**Note. Data are from the 1979 NLSY. Each column represents a different model with specifications following from Saperstein and Penner (2012, Table 4), except that these models are limited to the years in which it is possible to identify blips in both current and lagged classification (1982–94).+ . P < .10.*. P < .05.**. P < .01.***. P < .001.View Table ImageBlips versus Shifts in ClassificationAs we noted in “Racial Fluidity and Inequality” (p. 689), roughly one-third of the respondents in the 1979 National Longitudinal Survey of Youth (NLSY) who experience a change in their racial classification have only one discrepant classification over the course of the survey.2 These seemingly idiosyncratic “blips” can be contrasted with racial classification trajectories that represent shifts from one classification to another (e.g., WWWWWOOOOOOOOOOOO) or classification trajectories that repeatedly oscillate, back and forth, between two or more categories (e.g., BBBWBBBBWWBWBWBWB).3 To the degree that blips in classification represent actual errors (e.g., introduced during transcription), and not unconscious racial biases on the part of the interviewers, we might wish to ensure that our findings are not driven by such cases.4For this reanalysis, we operationalize “blips” by considering strings of five adjacent classifications and identifying cases where the respondent has the same race in both the previous two classifications and the subsequent two classifications, but a different race in between. Thus, in the string WWOWW, the third classification would be considered a blip, but strings like WWOOO, WWOWO, or BBOBW would not be considered blips.5 According to this definition, blips account for about 22% of the cases where current and prior survey year racial classifications differ.Having identified these blips, we then reestimate the results using only the nonblip racial classifications. Given that we can only identify blips when we have two previous and two subsequent racial classifications, we also reestimate our original models (from Table 4, p. 699) using only the relevant set of survey years.6 The results from models including and excluding blips are strikingly similar (see Table 1), suggesting that our findings are not driven by one-off classification changes or the types of measurement error that might produce them.Measurement Error in Independent and Dependent VariablesIn evaluating KDH’s critique that the relationship between social status and racial classification presented in “Racial Fluidity and Inequality” is simply a product of measurement error, it is crucial to remember that measurement error in dependent and independent variables can have very different effects.7 Much of the conventional wisdom about the effects of measurement error comes from studies of continuous independent and dependent variables in bivariate ordinary least squares regression models (OLS). In the bivariate OLS case with continuous variables and independent errors, it is well established that measurement error in an independent variable introduces bias, while measurement error in a dependent variable results instead in a loss of efficiency, but less is known about how one should expect measurement error to work in other contexts (see Bound et al. 2001 for a review).With this understanding of measurement error in mind, we engage KDH’s critique and approach to testing our results on its own terms. The logic behind their approach is that each individual has a true race and that fluidity is observed because measures of race vary randomly from this true value. According to this logic, every measure of race should include some measurement error, but KDH’s procedure only makes changes to the race measure being used as the independent variable.8 To better capture the logic that all measures of race are measured with error, we introduce changes to both the independent and dependent variables.9 We estimate 1,000 replications of both KDH’s original procedure and our modified version of their procedure and report the average coefficients from these models in Table 2.10 Table 2 presents estimates from baseline models, models using KDH’s approach (“IV only”), and models with our modification (“IV&DV”). Finally, to provide a sense of where the baseline model results fall in the distribution of the IV&DV results, we report the proportion of the IV&DV coefficients with a magnitude larger than the baseline coefficient. We take a similar approach to extend the analysis to self-identification as white, and interviewer classification as both black and white.11Table 2. Reexamining KDH’s Approach to Measurement Error BlackWhite BaselineIV OnlyIV & DVP(B I)Self-identification model: Unemployed….76.85.37.02−.41−.59−.28.04 Impoverished….83.86.42.03−.59−.75−.35.00 Incarcerated….27.42.22.44−.55−.61−.30.05 Received welfare….25.53.24.48−.07−.41−.20.94Interviewer classification model: Unemployed….32 .33.54−.29 −.26.02 Impoverished….35 .36.61−.37 −.32.00 Incarcerated….32 .30.43−.33 −.28.07 Received welfare….18 .11.14−.12 −.11.26Note. Models include other status variables and controls from Saperstein and Penner (2012, Table 4). Cols. P(B I) report the proportion of the 1,000 replicates of the IV & DV models with coefficients of greater magnitude than the baseline results (i.e., how often does eq. [2] produce coefficients as large as the baseline model). In the interviewer classification models, we do not replicate the “IV only” (KDH) approach because the racial classification from a given survey year can be both an independent variable and a dependent variable.View Table ImageOur small change to their procedure has large consequences. Like KDH, when we introduce changes only to the independent race variable, we find results that are similar in magnitude or larger than the baseline coefficients. However, when we perturb both the independent and dependent race variables, this is no longer the case. The pattern of results in the baseline models is infrequently produced by the IV&DV model; for example, the IV&DV coefficient for unemployment is as large as the baseline coefficient just 4% of the time when predicting self-identification as white, and 2% of the time when predicting self-identification as black.12 Put simply, we show that their results can only be obtained if one makes the unfounded assumption that measurement error only affects race when it is an independent variable. Thus, although we cannot condone KDH’s approach to evaluating measurement error, we conclude that there is nothing in KDH’s critique to suggest measurement error is the only reason the relationship between social status and racial categorization appears to run in both directions.Who Is Really “At Risk” of Racial Fluidity?The critique that we overstated the case for racial fluidity raises questions about (1) how common such fluidity is likely to be and (2) how general is the relationship between status and racial categorization. We address each of these issues in turn. Both AIL and KDH assert that meaningful changes are less common than we suggested and, when they do occur, such changes are limited to people who are racially “ambiguous.” We do not disagree that categorical fluidity is observed more frequently in particular subpopulations, and we stated as much in “Racial Fluidity and Inequality” (p. 707). However, it is problematic to assume that either racial ambiguity or categorization as “Hispanic” (or “Latino”) is a static characteristic. Further, it is important to clarify it is not the level of categorical racial fluidity but rather the process through which social status influences racial perceptions that we propose to be more general.Significance of Fluidity not Primarily about FrequencyIn “Racial Fluidity and Inequality” we highlighted that more than 20% of the 12,686 NLSY respondents experienced at least one change in their racial classification. AIL get a lower estimate because they apply survey weights. Although we were careful to discuss the level of fluidity relative only to the sample, we acknowledge that we did not explicitly state that we were not using the NLSY to establish the level of racial fluidity as a population parameter. From our perspective, a “true” estimate of racial fluidity is about as meaningful a concept as a “true” measure of race. Even gauging whether there is a lot of racial fluidity, or only a little, depends on how stable one expects race to be.To us, the significance of racial fluidity stems not from its common-ness but from its utility for understanding processes of racial categorization. Fluid cases provide leverage that studying racially stable people cannot: they allow us to ask, what predicts being assigned to, or being removed from, a particular racial category? For this purpose, what matters most is whether a given sample provides enough variation to study the correlates of category assignment. That said, we recognize that establishing the magnitude and scope of racial fluidity is likely to interest many researchers and address these issues below.To weight or not to weightAs our aim was to exploit the repeated racial classifications to better understand predictors of categorization, the issue of weighting is most relevant for our multivariate analyses. We estimate coefficients for our key status variables using both the 1979 sample weight used by AIL and a customized panel weight. If anything, the evidence for our claims is stronger when we use the survey weights (see Table 3).13Table 3. Predicting Racial Categorization with and without Weights BlackWhite Unweighted1979 WeightsPanel WeightsUnweighted1979 WeightsPanel WeightsSelf-identification model: Unemployed….76***.79***.79***−.41***−.42***−.42*** Impoverished….83***.93***.93***−.59***−.65***−.65*** Incarcerated….27.37.33−.55*−.69**−.69** Received welfare….25.35.35−.07−.23+−.23+Interviewer classification model: Unemployed….32***.31***.31***−.29***−.27***−.27*** Impoverished….35***.51***.51***−.37***−.49***−.49*** Incarcerated….32*.39*.39*−.33***−.42***−.42*** Received welfare….18**.28***.28***−.12***−.21***−.21***Note. Data are from the 1979 NLSY. Model specifications follow Saperstein and Penner (2012, Table 4). Models predicting self-identification and classification use different panel weights, corresponding to the different survey years on which they draw.+ . P < .10.*. P < .05.**. P < .01.***. P < .001.View Table ImageHowever, in considering whether or not our estimates should have been weighted, it is important to recognize that the survey’s weighting schemes treat racial categories as fixed strata. Respondents are weighted differently depending on how they were classified by NLSY in 1978—which is not necessarily consistent with how respondents later identified themselves, were perceived by interviewers, or would have been recorded in the 1970 census (to which the weighted population distributions were pegged). So, if one qualified as “Hispanic” based on the survey screener, one would be assigned separately calculated “Hispanic” weights throughout.14 Yet, respondents also could self-report Hispanic origins in 1979, or answer “yes” that they were “Hispanic, Latino, or of Spanish origin” in 2002. The complexity of who counts more or less when the weights are employed is highlighted by the fact that just 467 (32%) of the 1,437 respondents who have “Hispanic” weights are consistently “Hispanic” across all three measures. Given this, and the fact that weighting did not affect our multivariate analyses, we present unweighted frequencies and model estimates throughout our reply.Fluidity and ambiguityAIL and KDH emphasize that the vast majority of Hispanics (as defined by the 1979 measure) have fluid racial classifications, and that these “ambiguous” cases account for a sizeable proportion of the overall fluidity in the sample. They also imply that much of the racial fluidity would be eliminated with a more appropriate set of ethnoracial categories. We disagree that a better measure of race would eliminate either fluidity or ambiguity; some cases will fit better in a given classification scheme than others, and this is true for all types of classification (Zerubavel 1991). Further, rather than thinking of racial ambiguity as a relatively fixed characteristic of the person being categorized, as AIL and KDH seem to, we think of racial ambiguity as the result of a confluence of factors from the individual’s own characteristics to the circumstances in which the categorization takes place. Neither people nor populations are always racially ambiguous; ambiguity (or lack thereof) is socially constructed and entwined with the classification scheme, such that different people will be “ambiguous” in different schemes.We illustrate that fluidity and the expectation of ambiguity do not always go hand-in-hand by comparing levels of racial fluidity in the NLSY across a range of individual characteristics likely to be associated with categorical ambiguity (or difficulty fitting particular individuals into U.S. racial classification schemes). This point can be seen most clearly in Table 4 among NLSY respondents who selected “origin or descent” categories in 1979 that might indicate multiracial heritage (e.g., black and Asian Indian); if anything, as a group, they are less likely to have fluid racial classifications over the course of the survey than those who report a single origin (or whose multiple origins do not cross racial boundaries).15 Racial fluidity is more common not only among people who report Hispanic origin in 1979 but also among people who were not born in the United States,16 and people whom we classified as changing their self-identification between 1979 and 2002. Yet some fluidity is present even among respondents who are non-Hispanic, U.S. born, and stably self-identified. Thus, racial fluidity and racial ambiguity should be treated as distinct analytical concepts, regardless of how related they might seem.17Table 4. Individual Ambiguity and Racial Fluidity % With at Least One Discrepant Racial ClassificationNReported multiracial origins in 1979…191,380Did not report multiracial origins in 1979…2111,116Reported a Hispanic origin in 1979…871,976Did not report a Hispanic origin in 1979…810,520Not born in the United States…70868Born in the United States…1711,736Self-id changed between 1979 and 2002…451,544Self-id did not change…186,090Note. Data are from the 1979 NLSY. All indicators are coded following Saperstein and Penner (2012).View Table ImageTo address AIL and KDH’s claims that the residual “other” category drives the observed fluidity, we compare the NLSY to Add Health. We find similar levels of change in racial classification between waves 3 and 4 in Add Health as we do year-to-year changes in NLSY (4.5% vs. 6.0%, respectively), despite Add Health’s more specific classification scheme, which included “American Indian or Alaska Native” and “Asian or Pacific Islander” but did not include “other.” Although neither survey provided “Hispanic or Latino” as a category, unlike AIL and KDH, we are skeptical that adding another option for the interviewers’ classifications would yield racial stability. A “Hispanic” category might make some people’s classifications more stable, but could make other people’s classifications less stable than they were when there were fewer options available; the boundaries between white and Hispanic, black and Hispanic, or Asian and Hispanic are far from clear.We agree with AIL and KDH that examining how the level of fluidity differs across subpopulations is necessary to fully understand processes of racial categorization, and the relationship between racial fluidity and ambiguity is important to consider. Nevertheless we chose to focus our efforts instead on the predictors of categorization because the question of whether or not status characteristics change along with changing racial categorizations is what connects the social construction of race to broader issues of inequality.The Role of Status in Racial and Ethnic CategorizationIt might seem counterintuitive that observed categorical racial fluidity could be concentrated in particular subpopulations, while the relationship between social status and racial categorization could be a more general one. This apparent paradox is partially resolved by recognizing that the status factors associated with a change in categorization from nonwhite to white in one year (e.g., graduating from college) can also help to explain why someone might continue being categorized as white in the future. Further, even if status shapes racial perceptions for everyone, some people’s observed racial categorizations can still be more sensitive to changes in social position.To demonstrate this, we briefly discuss the merits of thinking of racial categorization in terms of continuous probabilities. We then respond directly to AIL’s critique by presenting fixed effects models estimating the association between social status and racial categorization across a series of subpopulations. We conclude this section by examining whether status factors also shape who identifies as Hispanic.Observed categorical change vs. continuous probabilities of categorizationInstead of each person in the United States being assigned a race in a static, categorical sense, we find it helpful to think of everyone as having a probability of identifying or being classified in each category that represents the chance that they will identify or be seen as a particular category at a particular point in time (see “Racial Fluidity and Inequality,” pp. 706–8). We argue that status considerations—and changes in social position, in particular—have the potential to nudge these probabilities by a few percentage points, in one direction or the other. Whether a change in a person’s probability of categorization results in observed fluidity depends on a host of other factors including the categories offered, who is doing the categorizing, and where the person being categorized was located in the various probability distributions to begin with. It is in this sense of continuous probabilities of categorization, and how those probabilities are shaped by status cues, that we intended our results to be seen as applicable to Americans in general.The best evidence of this comes from our collaborative work on the psychological process of racial categorization (Freeman et al. 2011). We replicated our basic finding from the NLSY in an experimental setting by showing that the same faces are racially categorized differently depending on whether they are portrayed in high-status (business suit) or low-status (janitorial coveralls) clothing. A novel mousetracking tool allowed us to examine whether or not the status cues had an effect on racial perception when the categorization itself appeared unaffected. In the low-status condition, even when individuals ultimately classified faces as white, their mouse moved significantly closer to the box for making a “black” classification than it did when the same face was presented in the high-status condition. Conversely, in the high-status condition, when a face was classified as “black,” average mouse trajectories veered significantly closer to the “white” category. These results suggest that even when status cues do not change how people are racially classified in categorical terms, they can play an important role in shaping racial perceptions more broadly.General vs. subpopulation specific resultsAIL, in particular, question the claim that the bidirectional relationship between social status and racial categorization is a more general phenomenon. AIL interpreted their evidence of a relationship between social status and classification as “weak” based on models restricted to people who self-identified in 1979 using either Hispanic origin categories or origin categories that crossed racial lines (see AIL, Table 2).18 We revisit their conclusion across a broader range of subpopulations, including only the NLSY cross-sectional sample and only non-Hispanic or non-multiracial respondents. AIL note issues of statistical power in their models, and we too consider most of our models to be underpowered. Nevertheless, the exercise is useful to demonstrate that, overall, estimates for the associations of racial categorization with status factors are consistent in direction and magnitude regardless of subpopulation (see Table 5).Table 5. Comparison of Model Estimates Predicting Racial Categorization With Different Sample Restrictions Fixed Effect ModelsWhole Sample (no FE)Whole SampleCross-sectional SampleNon-Hispanic (1979 and 2002)Hispanic (1979 or 2002)Non-multiracial (1979 and 2002)Multiracial (1979 or 2002)Non-AIAN, Hispanic or MultiracialAIAN, Hispanic or MultiracialSelf-identified as white: Unemployed…−.057***−.012−.008−.016**−.013−.015*−.001−.007−.028 Impoverished…−.069***−.015+−.014+−.005−.029−.013+.005−.003−.048* Incarcerated…−.087***−.047**−.058*−.034**−.101−.038*−.182*−.007−.152** Received welfare…−.043***.004−.009−.013*−.006.012−.054+−.008+−.019Self-identified as black: Unemployed….019***.008**.006*.008**.010.009**−.004.008*.009 Impoverished….021***.003.003.003.002.001.015.002.006 Incarcerated….020*.008.016.009.002.009−.003.007.011 Received welfare….016***.005+.008*.006+.001.004.017.007+.003Classified as white: Unemployed…−.027***−.002.002.000−.006−.002−.001.000−.003 Impoverished…−.032***.003.002−.001.015.003.002.000.014 Incarcerated…−.043***−.008−.013−.007+−.022−.005−.032*−.005−.018 Received welfare…−.029***−.003.002−.001−.015−.001−.016*−.001−.008Classified as black: Unemployed….005***.002*.001.001.005+.001.006.001.005* Impoverished….006***.000.001.000−.003.000.001.000.000 Incarcerated….008***.003.001.005+−.004.003.007.004.003 Received welfare….006***.001.000.001.002.001.001.001.002Note. Data are from the 1979 NLSY. Each coefficient represents a different model predicting racial categorization using the relevant status variable and controls for age, living in the South, year fixed effects, and interviewer characteristics (age, gender, education, and race). The first model also includes a control for prior racial classification/identification (as relevant), and additional controls for whether the respondent reported a Hispanic origin in 1979, multiple races in 1979, or was born outside the United States. Following Saperstein and Penner (2012), SEs in the non-FE models account for the clustering of observations within interviewers; fixed effect models account for the clustering of observations within respondents. AIAN—American Indian or Alaska Native.+ . P < .10.*. P < .05.**. P < .01.***. P < .001.View Table ImageLike AIL, we estimate a series of models with a common set of controls (for considerations such as age) in which our key status factors are entered individually.19 Unlike AIL, our models include respondent fixed effects.20 We also show results for the full sample with and without respondent fixed effects so the direction, magnitude, and statistical significance of the estimates can be compared across the full range of model specifications.21 We do not expect all of the coefficients in Table 5 to be statistically significant, particularly given our original fixed effect model results (see “Racial Fluidity and Inequality,” app. table A3, where status factors were entered simultaneously rather than separately). When we split the sample into smaller subpopulations, we have even less power to distinguish statistically significant differences than in our original analyses. However, there is still information to be gained by comparing estimates from these fixed effects models relative to estimates from models with less conservative controls but more statistical power.The strongest evidence supporting our perspective in Table 5 can be found in the relationship between unemployment and racial categorization—for both self-identification and interviewer classification. Across all four panels of fixed effects models, 28 of 32 estimates for the association between long-term unemployment and racial catego