Abstract
An abundance of scholarship has examined the racial invariance thesis positing that the causes of violence, especially markers of disadvantage, are similar across racial/ethnic groups. More recently, research has adopted “yardsticks” to provide more meaningful assessments of the thesis, including incorporating Latinos into analyses and using statistical tests to compare disadvantages’ effects across groups. Less attention, however, has been given to the measure of violence the thesis applies to. Although intended to explain offending, criminologists commonly substitute measures of race-/ethnic-specific arrest and victimization. Using 2010–2014 National Incident-Based Reporting System data for 453 census places, we examine whether the relationship between structural disadvantage and race-/ethnic-specific violence varies across measures of offending, arrest, and victimization. Consistent with “lenient interpretations” of the thesis, we find that disadvantage is generally associated with higher rates of violence among Whites, Blacks, and Latinos, regardless of the measure of violence. However, at odds with “strict interpretations” of the thesis, there are significant differences in the magnitude of disadvantages’ effects across groups and these differences are conditioned somewhat by the measure of violence examined. Implications of these findings and directions for future research are considered.
In his seminal work, the Truly Disadvantaged, Wilson (1987) argued that racial inequalities in violence are rooted in the divergent structural circumstances of the communities of Whites and Blacks. The argument that race differences in crime and violence are the product of differential exposure of Blacks to criminogenic structural conditions has come to form the core of the “racial invariance hypothesis” (Sampson & Wilson, 1995). In short, this thesis predicts that macrostructural characteristics, particularly markers of structural disadvantage, are associated with crime and violence in similar ways across racial/ethnic groups (Ousey, 1999). Although much scholarship finds that disadvantage elevates violence among all racial/ethnic groups, some caution “the assumption of racial similarity in the fundamental causes of violence or crime is far from settled” (Steffensmeier, Ulmer, Feldmeyer, & Harris, 2010, p. 1134; also see Ousey, 1999).
More recently, Steffensmeier, Ulmer, Feldmeyer, and Harris (2010, p. 1160) argued that ambiguity in the scope and conceptual conditions used to evaluate the racial invariance thesis have left it a “moving target” making “definitive empirical conclusions about the viability of the hypothesis difficult.” Since its publication, a number of scholars have adopted their suggested “yardsticks,” including incorporating Latinos into comparisons, examining multiple types of violence (e.g., homicide, index violence, robbery), and using statistical tests to compare disadvantages’ effects across groups (Bethelot, Brown, Thomas, & Burgason, 2016; Hernandez, Velez, & Lyons, 2018; Painter-Davis & Harris, 2016; Wright, Turanovic, & Rodriguez, 2016).
Despite these advances, there remains ambiguity regarding the appropriate measure of violence for which tests of the racial invariance thesis should be applied. On the one hand, some scholars have argued that the invariance thesis was originally intended to explain patterns of offending (Messner, Beaulieu, Isles, & Mitchell, 2014). Yet estimating race-specific and ethnic-specific offending rates presents a “vexing methodological challenge” because such data are uncommon and often limited to specific types of crime and/or to particular locales (Messner et al., 2014, p. 2; also see Sampson, 1987). For example, self-report and victimization surveys that include markers of offending often omit the more serious crimes some argue are most applicable to the thesis (e.g., homicide and robbery) and/or are often unavailable for a wide range of macrosocial units. On the other hand, sources of official crime data that resolve some of these issues, such as the Uniform Crime Report’s (UCR) Supplemental Homicide Reports, only code the offender’s race for homicide and inconsistently code ethnicity. In turn, comparisons are often limited to the Black–White divide, failing to include Latinos, now the largest minority group in the United States (Pew Research Center, 2011).
These limitations have led scholars to use data on violent arrests (e.g., Light & Harris, 2012; Sampson, 1987; Shaw & McKay, 1942; Steffensmeier et al., 2010) or homicide victimization (e.g., Krivo & Peterson, 1996; Light & Ulmer, 2016) as proxies of offending in assessments of the racial invariance thesis. 1 Criticisms of these data choices center on their underlying assumptions (a) that race-/ethnic-specific differences in violent arrests capture differences in behavior rather than social control and (b) that violent victimizations are typically intraracial, between people of the same race/ethnicity. In turn, these practices also presuppose that the predictors and processes that lead to violent offending among specific racial/ethnic groups are the same for violent arrest (Steffensmeier et al., 2010) and victimization (Krivo & Peterson, 1996). While the literature is filled with a variety of studies leveraging offending, arrest, or victimization data for the purposes of operationalizing race-specific violence, there remains little in the way of systematic empirical inquiry examining how such choices affect conclusions regarding the racial invariance thesis.
As we spell out below, there are competing perspectives on whether offending, arrest, and victimization data can be used interchangeably in tests of the racial invariance thesis. Different measures of violence may provide “lenient” or “stricter” tests and differing conclusions regarding the degree of racial similarity in the structural sources of violence. Sorting out the appropriate dependent variables in tests of the racial invariance thesis is crucial for advancing our understanding of race and crime across communities, as well as for setting the stage for meaningful tests of racial invariance claims. While it is beyond the scope of this article or the capacity of our data to sort out every aspect of measuring the appropriate dependent variable in what is a complex debate, our goal is to unpack some of these unresolved issues surrounding alternative measures of crime and to provide a template for theory and future research.
Empirically, we use 2010–2014 National Incident–Based Reporting System (hereafter, NIBRS) data to systematically examine whether conclusions regarding racial invariance vary depending on the utilization of offending, arrest, or victimization measures. NIBRS is uniquely suited for this study because, unlike other data sources, it provides information on the race and ethnicity of arrestees and victims, as well as the race of offenders, for a broad range of both lethal and nonlethal violent offenses (e.g., homicide, robbery, assault) across a broad range of communities. 2 Critically, by carefully addressing issues of missingness on Latino identifiers, we are able to include what is now the single largest U.S. minority group, Latinos, in our analyses of arrest and victimization. We therefore go beyond prior research in comparing across racial and ethnic groups (White, Black, and Latinos) for multiple measures and types of violence (vs. only a single measure, like arrest or victimization, or a single offense type, like homicide) to determine whether such methodological choices—and the assumptions underlying them—lead to different conclusions regarding racial invariance.
The inclusion of Latinos in our analysis is especially noteworthy. Only a handful of studies have examined whether relationships between structural factors and race-specific violence for Blacks and Whites vary depending on the measure of violence (Chilton & Regoeczi, 2007; Messner et al., 2014). However, to our knowledge, no study has examined these dynamics for Latinos as compared to Blacks and Whites across multiple measures (e.g., arrest and victimization) and types of violence. This oversight is particularly concerning given growing debates regarding the contours of the “Latino paradox” whereby structural disadvantage has weaker effects on Latino violence as compared to Blacks and Whites (Feldmeyer & Steffensmeier, 2009; Painter-Davis & Harris, 2016). At the center of these debates is emerging evidence suggesting that the relationship between disadvantage and Latino violence (and conclusions regarding the Latino paradox and invariance) may vary depending on the measure of violence used (Painter-Davis & Harris, 2016).
Remaining Scope and Conceptual Issues of the Racial Invariance Thesis
Grounded in the work of Shaw and Mckay (1942) who found that rates of delinquency remained high in certain disadvantaged Chicago neighborhoods despite dramatic changes in their racial/ethnic composition, the racial invariance thesis generally posits that the structural causes of crime are similar across racial and ethnic groups. The thesis parallels themes from social disorganization and structural strain/anomie perspectives that propose (a) structural disadvantage is associated with higher crime rates and (b) the disadvantage-crime association exists regardless of the demographic subgroup in question (Ousey, 1999).
Comprehensive reviews of the literature support such predictions. Regarding the former, structural disadvantage is among the most robust correlates of violence (Land, McCall, & Cohen, 1990; Pratt & Cullen, 2005). As for the latter, much scholarship finds that disadvantage elevates violence among all racial/ethnic groups, though there remains disagreement about the magnitude of its effect for different groups (i.e., the degree of support for the racial invariance thesis specifically). Some scholars argue that evidence supports the racial invariance thesis (e.g., Krivo & Peterson, 2000; Peterson & Krivo, 2005), while others view the debate as unsettled (Steffensmeier et al., 2010; also see Ousey, 1999; Parker, 2008; Unnever, Barnes, & Cullen, 2016).
Measuring Violence in Tests of the Racial Invariance Thesis
Measurement of crime remains a central issue around which the invariance debate centers. Many have argued that the racial invariance thesis was originally intended to explain racial/ethnic differences in offending (Messner et al., 2014). As foundational research in this area, the work of Shaw and Mckay (1942) focused on race differences in “delinquency-particularly group delinquency, which constitutes a preponderance of all officially recorded offenses committed by boys and young men…” (p. 96). Likewise, Sampson (1985, p. 649) argues that most criminological theories are designed to explain patterns of crime among “different groups of offenders.” As such, much criminological research that is devoted to examining macrolevel racial disparities in crime focuses on differences in offending (Messner et al., 2014). However, due to issues of data availability outlined earlier, other scholars continue to use race-specific arrest and victimization data as proxies for offending. As we discuss below, there are competing positions on the appropriateness of these practices.
Violent victimization as a proxy for violent offending?
One view is that race-specific victimization data can be used effectively as proxies for race-specific offending in tests of the racial invariance thesis. This perspective argues that the overwhelming majority of violence, especially homicides, are intraracial (i.e., the victim and offender are of the same racial/ethnicity) such that victim’s race should be a strong substitute for offender’s race. This rationale has been used in studies that examine the influence of structural disadvantage on White and Black homicide (Krivo & Peterson, 1996) and, more recently, in studies of Latino homicide (Light and Ulmer, 2016). In adopting this approach, researchers have argued that theoretical processes (e.g., social disorganization, structural strain) hypothesized to affect violent offending should have similar effects on a group’s vulnerability for victimization (Krivo & Peterson, 1996; Wilson, 1987).
Yet there are strong reasons to believe that race-/ethnic-specific violent victimization data may not be a satisfactory proxy for race-/ethnic-specific violent offending. First, the assumption that victims and offenders are of the same race or ethnicity may not be as strongly supported today as it was in the past. This assumption is largely grounded in research that focused on violence among Whites and Blacks (O’Brien, 1987). Major demographic shifts over the past three decades, including heightened immigration, and subsequent increases in diversity have increased intergroup contact, including opportunities for intergroup violence (Harris, Gruenewald, & Painter-Davis, 2015). 3 Latinos have recently surpassed Blacks as the largest minority group in the United States and research suggests that, in comparison to Whites and Blacks, Latino victimization is more likely to involve an offender of a different race/ethnicity. For example, analyses of recent NCVS data (2012–2015) suggest that whereas roughly 60% of Black and White violent victimizations are intraracial, only about 40% of Latino victimizations involve a Latino offender (Morgan, 2017). Second, the assumption that most crime is intraracial may hold more for homicide and aggravated assault, but less for other violent offenses such as robbery relevant to testing racial invariance (Harris et al., 2015). As an example, an analysis of NIBRS data for 2014 suggests that whereas over 70% of homicides and aggravated assaults are intraracial, only about a half of robberies involve victims and offenders of the same race/ethnicity (also see Morgan, 2017).
Third, victimization may not be an appropriate proxy for offending as recent theorizing and empirical inquiry suggests that, especially for Latinos, the structural sources of victimization may differ somewhat from those of offending (Harris et al., 2015; Shihadeh & Barranco, 2010). As evidence, Shihadeh and Barranco (2010) find that structural disadvantage is unassociated with Latino homicide victimization in some localities, whereas Painter-Davis and Harris (2016) observe that disadvantage elevates Latino homicide arrests. Such a discrepancy implies structural and cultural characteristics of Latino communities may enhance social organization in ways that decrease Latino violence but may be more protective against Latino offending than victimization (Painter-Davis & Harris, 2016). A key reason that disadvantaged Latino populations may be victimized at rates that exceed their rates of offending is because Latinos, especially immigrants, may be viewed as attractive targets for crime because they are less likely to engage in the formal economy and, thus, more likely to carry cash on hand. Moreover, Latinos (especially immigrants) may fear deportation and harbor distrust of authorities, making them reluctant to contact the police or to carry firearms for protection (Barranco & Shihadeh, 2015).
Violent arrest as a proxy for violent offending?
There are also opposing views on whether race-/ethnic-specific violent arrest data can be used as a proxy for violent offending. One view is that violent arrests can substitute for violent offenses because typical critiques of arrest data (i.e., that they underestimate true levels of offending and racial/ethnic differences in arrest) are less problematic for serious forms of violence such as homicide and, to a lesser extent, robbery. Research has shown a high correlation between race-specific arrest rates and race-specific offending rates measured via survey data, at least for some offenses (Hindelang, 1978). Likewise, other studies suggest that crime seriousness is the strongest predictor of arrest and that there is little racial bias in arrest decisions for robbery and, especially, homicide (D’Allesio & Stolzenberg, 2003).
Central to our purposes here, scholars holding this position point to research finding that structural factors are associated with measures of violent arrest and offending in similar ways (Chilton & Regoeczi, 2007). Messner, Beaulieu, Isles, and Mitchell (2014), for example, argue that Black homicide arrests can effectively represent Black homicide offending because, even though they are only moderately correlated with one another, they have similar relationships with structural conditions, particularly disadvantage. Thus, these measures are valid indicators of each other according to Lazarsfeld and Thielens’s (1972) standard of the “interchangeability of indices,” a foundation of social research that postulates that even when two variables are imperfectly related to one another, they can be interchanged if they have similar relationships with a third outside variable.
An alternative view is that arrest data for violent offenses should not be used as a proxy for offending. Scholars here point to the lack of strong correlations between race-specific arrest and offending data, implying that they are not indicators of the same construct (Neopolitan, 2005). Drawing on conflict theory, for example, some scholars argue that arrests are markers of social control that disproportionately drive minority crime rates higher because they are less able to resist the imposition of punishment and because “authorities share common stereotypes linking them with crime” (Liksa & Chamlin, 1984, p. 384). In essence, race/ethnic differences in arrests for serious forms of violence—including homicide, robbery, and violent index offenses—reflect to some extent racial/ethnic differences in social control (e.g., Liksa & Chamlin, 1984). By extension, the “interchangeability of indices” issue that proponents of arrest data cite (Messner et al., 2014) is seen as an open empirical debate that warrants careful analysis.
Type of crime, strict and lenient tests, and measures of violence
Notably, the arguments of the above positions depend in part on scope and conceptual issues that are central to the racial invariance hypothesis itself, including the type (or severity) of crime under examination. Steffensmeier et al. (2010, p. 1138) note that “the criminal landscape is vast, and obviously a considerable stretch exists in terms of what is meant by ‘crime.’” Such sentiment intersects with arguments regarding what measure of violence should be used in tests of invariance. On the one hand, arrests for more serious forms of violence, like homicide and robbery, are more likely to be reported and, thus, may proxy offending to a large extent. On the other hand, arrests may not be suitable substitutes for offending for some other forms of violence, like aggravated assault, for which arrests represent a smaller fraction of all offending. At the same time, different types of violence vary in the extent to which they are intra- or interracial in ways that impact the measurement of race-specific violence (Harris et al., 2015). Even if homicides remain largely intraracial for most groups (and, therefore, effectively proxy offending data), the same may not be true of other forms of violence like robbery.
More broadly, the issue of whether to utilize a strict or lenient test of the racial invariance thesis matters, as well. A strict test would require disadvantage to be positively associated with violence among each racial group in statistically similar ways (i.e., same magnitude) and across different types of violence (e.g., homicide and index violence; Steffensmeier et al., 2010). However, these strict benchmarks may be unrealistic for assessing racial invariance claims given (a) measurement error in crime data, especially by race, (b) methodological challenges for tests of the racial invariance thesis, like restricted distributions (McNulty, 2001), 4 and (c) problems with significance testing (i.e., the cult of significance; see Ziliak & McCloskey, 2008). Lenient tests, by contrast, may provide more flexibility to deal with methodological constraints posed by the thesis, while still allowing scholars to assess its core hypothesis—whether particular structural characteristics are universal sources of crime and violence by race and ethnicity.
In short, the scope and conceptual issues that pervade the racial invariance literature broadly also intersect with the strategic choices made in measuring the dependent variable of violence specifically. While a number of studies have assessed the racial invariance thesis using a single measure of violence, offending, arrest, or victimization, few have examined these measures simultaneously to systematically evaluate how measurement of the dependent variable affects conclusions regarding the racial invariance thesis. We turn now to the parameters of the current study that seeks to address this gap.
Data and Methods
To examine these research questions, we combine data on violence from the NIBRS for the years 2010–2014 with information on social structural characteristics of communities throughout the United States as drawn from the Census and the American Community Survey 5-year estimates (2010–2014). All Census variables were downloaded from the National Historical Geographic Information System (Minnesota Population Center, 2016). NIBRS has gained prominence in recent years as a “seminational” source of official crime data, providing full or partial coverage of crime in more than 30 states, covering about 30% of the population and over one fourth of the nation’s crime. Once a state becomes NIBRS-certified, individual police agencies report detailed incident-level information on crimes reported to the police to the Federal Bureau of Investigation (Justice Research and Statistical Association, 2018). In contrast to the limited racial demographic information included in the UCR (e.g., no reliable markers of ethnicity; only codes race of victim, offender, and arrestee for homicide; race data for offenses other than homicide is included only for arrestees), NIBRS provides information on the race and ethnicity of arrestees and victims and the race of offenders for a broad range of crimes.
Unit of Analysis
The unit of analysis is the incorporated census place, which includes nonoverlapping geographic units, ranging from small towns, villages, and boroughs housing several thousand residents up to the largest metropolitan statistical areas (Feldmeyer & Steffensmeier, 2009). This unit of analysis is advantageous because census places vary considerably in size, racial/ethnic composition, structural characteristics, and violence, and are large enough to provide sizable numbers of each racial/ethnic group and adequate counts of violence to allow for meaningful statistical analyses. Census places were included only if (a) they had a total population in 2010 of at least 5,000, and have at least 500 Whites, 500 Blacks, and 500 Latinos, and (b) there were at least 2 years of valid NIBRS data between 2010 and 2014, in which all agencies within the census place reported to NIBRS. Our final sample size is 453 census places.
Dependent Variables
Our dependent variables include race-/ethnic-specific census place measures of violent offending, arrest, and victimization. We include counts of homicide (see below) and rates (per 100,000 people) of robbery and violent index offenses (sum of robberies, rapes, aggravated assaults, and homicides), forms of violence that have been central to tests of the racial invariance thesis (see review in Steffensmeier et al., 2010). We construct measures of arrest and victimization for all three racial/ethnic groups (Whites, Blacks, and Latinos) and measures of offending for Whites and Blacks.
Confounded White offense rates
A limitation of our study is that a measure of offending is not included for Latinos because a Latino identifier for offenders is not consistently collected by NIBRS. 5 As such, we are unable to estimate the association of disadvantage with Latino offending and our race-specific offender counts are confounded by the inclusion of Latinos. This is particularly problematic for the White offending counts, which are inflated when Latino offenders (most of whom are coded as White) are subsequently included in the White crime counts within official crime data (see Steffensmeier et al., 2011).
To conduct more accurate tests of racial invariance, we estimate non-Latino White offending counts using an adjustment procedure similar to that used in prior research (Steffensmeier et al., 2011). The three-step method involves employing correction factors for each offense that are used to remove “estimated Latino offending counts” from the original confounded White (Latino and non-Latino) offending counts in NIBRS. First, using Latino arrest data as a proxy for Latino offending, we create correction factors to account for the likelihood of arrest (because only a portion of offenders are arrested) and to upward adjust Latino arrest counts so that they better approximate Latino offending patterns. Such a process mirrors prior studies that have transformed White and Black arrest rates into “offending rates” by multiplying race-specific arrest counts by the overall (i.e., not race-specific) ratio of crime specific offenses to arrests to account for differences across jurisdictions in the likelihood of arrest (e.g., Chilton & Regoeczi, 2007; Sampson, 1987). Because nearly all Latino arrestees are coded as White, we modify this procedure and limit our correction factor—the ratio of offenses to arrests—to Whites, 6 whereby the White offense count, including both Whites and most Latinos, is divided by the sum of White and Latino arrests. As a second step, we then estimate Latino offense counts by multiplying the Latino arrest counts by the correction factor. Third, we subtract these estimated Latino offense counts from the White offense counts to generate unconfounded White, non-Latino offense counts. These counts are used to create rates of non-Latino White offending (see, e.g., Appendix A). The overall result of this process is that our unconfounded non-Latino White offending rates are appreciably lower than the confounded White offending rates and produce similar results as other methods of correcting for the confounding of Latinos with Whites in official crime data (see Steffensmeier et al., 2011).
Missing data on ethnicity
While NIBRS is strategic for examining our research questions, missing data for race and ethnicity are still problematic. 7 Following prior research (e.g., Harris et al., 2015), values for missing data were imputed using imputation by chained equations (Royston, 2004). Using this imputation procedure, 10 imputed data sets were created at the incident level for each year under analysis (2010–2014). The 10 imputed data sets were then averaged, for each year, to account for the uncertainty associated with the imputations. After averaging the imputations, the data were aggregated to the police agency level using originating agency identifiers and a “months reported correction” factor was applied to account for the number of months each police agency provided data to NIBRS (see Schwartz & Gertseva, 2010). Finally, offending, arrest, and victimization counts were aggregated to the census place level to create yearly, census place measures of violence (for additional details on procedure, see Harris et al., 2015; Painter-Davis, 2013). These counts were pooled across years (2010–2014) to add stability to the estimates and to ensure adequate counts. The counts were adjusted to account for the number of years the census place reported crime data (Peterson & Krivo, 2010). The homicide data were retained as counts, while the robbery and violent index offenses were used to compute rates per 100,000. The rates are square root transformed to normalize their distribution and to account for any nonlinearity in relationships.
Independent Variables
Following prior racial invariance research (see review in Steffensmeier et al., 2010), our primary independent variable is a structural disadvantage index that we construct separately for Whites, Blacks, and Latinos. Principal component analysis is used to construct the index, combining poverty (percentage below the poverty line), unemployment (percentage of the civilian population aged 16–64 who are unemployed), female headship (percentage of all families headed by a single female with children under 18 present), and educational disadvantage (percentage of the population who are 25 years or older without a high school degree or its equivalent). This index captures the overlap of multiple forms of disadvantage and effectively deals with collinearity issues that arise among disadvantage indicators (Land et al., 1990).
Additionally, we control for a variety of structural characteristics that research has shown to be correlated with violence across census places, including segregation with separate Black–White and Latino–White indices of dissimilarity across block groups within a census place, immigration (percentage of the total population that is foreign born), the relative size of the young male population (percentage of the population that are male and aged 15–24), residential instability (percentage of households that experienced turnover in residents in the past 5 years), entropy as a measure of racial heterogeneity, 8 population density (logged), total population (logged), police per capita, and Census region (Midwest, South, West, and Northeast as the reference).
Analytic Procedures
Two main multivariate methods are employed. First, seemingly unrelated regression (SUR) is used for models estimating violent index and robbery rates. SUR is more appropriate than ordinary least squares for comparing effects across multiple groups from a single sample (e.g., same unit of analysis) because it accounts for the correlated errors associated with shared unmeasured predictors across groups. In doing so, SUR provides more robust standard errors for comparing the relationships between covariates and dependent variables as they differ (or not) across groups (see, e.g., Steffensmeier et al., 2010). Second, we use negative binomial regression when predicting homicide counts to account for the highly skewed distribution of this offense along with the presence of many zero values. The natural logarithm of the race-/ethnic-specific population at risk is included as an offset variable in the negative binomial models and its coefficient is fixed to one, so that the model can be interpreted as per capita rates. We employ the suest command via Stata (StataCorp, 2007) to combine the estimates and covariance matrices of the separate negative binomial regression models (White, Black, and Latino) into one covariance matrix. Like the traditional SUR approach, the suest command accounts for the correlated errors associated with shared unmeasured predictors across groups and provides more robust standard errors.
After estimating both types of regression models, we implement statistical tests to better clarify whether there are important differences in the relationships between disadvantage and race-specific violence depending on the measure of violence used. Specifically, F tests are employed to assess whether (a) the effects of disadvantage on violence for a particular group vary depending on the measure of violence used (offense, arrest, and victimization) and (b) whether the effects of disadvantage on violence vary across racial/ethnic groups depending on the measure of violence. These F tests address two interrelated questions: First, are measures of violence interchangeable for specific racial/ethnic groups? Second, do conclusions regarding racial invariance in the causes of crime vary depending on the measure of violence? 9
Results
Table 1 displays patterns of violence and structural disadvantage for our sample of 453 census places. Panel A displays rates across multiple measures of race-/ethnic-specific violence and within-race correlations of these measures, while panel B displays markers of structural disadvantage by race/ethnicity. We note five key findings. First, regardless of the type of violence (homicide, robbery, or violent index) and its measure (offending, arrest, and victimization), Black rates are the highest, followed by Latinos and then Whites. These racial/ethnic differences in violence are consistent with separate bodies of literature examining racial/ethnic patterns of offending (Felson & Kreager, 2015), arrests (Steffensmeier et al., 2010; 2011), and victimization (Lauritsen & Heimer, 2010). Second, the correlations between race-/ethnic-specific measures of violence (offense, arrest, and victimization) in general range from moderate to very strong in size.
Violence and Structural Disadvantage by Race and Ethnicity Across Census Places.
Note. N = 453.
a Homicides are estimated as counts in regression models but are displayed as rates here. bWhite non-Latino offense rates are estimated by subtracting estimated Latino offense counts from white (Latino and non-Latino) offense counts. cRace-specific measure. dSame value for each group.
However, third, the strength of these correlations vary widely depending on the racial/ethnic group, as well as the type of violence (homicide and robbery) and its measure (arrest, offending, and victimization). Notably, within-race correlations across the measures of violence tend to be the strongest for index violence and weakest for robbery and homicide. For example, the correlation between homicide victimization and arrest rates among Latinos is only moderate (.42) but is strong for index violence (.75). Among Blacks, the correlations between robbery offending and victimization rates (.78) and between offending and arrest rates (.74) are both quite stronger, but notably weaker between victimization and arrest rates (.55).
Fourth, turning to panel B, in general, Blacks and Latinos experience levels of disadvantage that are higher than Whites. Black and Latino levels of poverty are similar with nearly 30% below the poverty line, compared with roughly 13% of White residents. Likewise, while Latinos have the highest level of educational disadvantage (33%), Blacks are far more likely to be headed by a female with children under age 18 (30%). Relatedly, fifth, while there are striking racial differences in disadvantage, there is some important overlap in the distribution of disadvantage across racial/ethnic groups. Notably, as indicated by racially disaggregated standard deviations in panel B, compared to White communities, there is wider variation among Blacks and Latinos across each indicator of disadvantage. This greater degree of variation creates some overlap in the distributions of disadvantage for Blacks, Whites, and Latinos, and is particularly strong for Latinos, whose disadvantage levels on some dimensions intersect closely with those of Blacks (poverty, unemployment) and Whites (unemployment). Because of this overlap, our analysis of racial invariance may be less vulnerable to problems of restricted distributions (see Note 4; Steffensmeier et al., 2010), an issue we return to in our discussion.
Turning to our primary multivariate models, Figures 1 and 2 graph the coefficients for the relationships between the disadvantage index and race-/ethnic-specific violence. All models include a full set of controls as described above (model statistics, coefficients, and standard errors are provided in Appendix B, Table B1). Figure 1 provides an assessment of the “interchangeability of indices” by comparing the unstandardized coefficients of the effects of structural disadvantage on markers of violence (offending, arrest, and victimization) separately for each racial/ethnic group. Significance tests are used to examine whether the effect of structural disadvantage on a particular group varies significantly depending on the measure of violence. Per our discussion above, lenient tests only require that disadvantage be significantly associated with violence for each measure, whereas strict tests require that disadvantage be related to violence in statistically similar ways (i.e., no differences in coefficients).
Assessment of “interchangeability of indices” using unstandardized disadvantage coefficients predicting different measures of violence for Whites, Blacks, and Latinos. Note: All models include a full set of controls and all disadvantage coefficients are statistically significant predictors at p < .05. O(offense), A(arrest), V(victimization) indicates significant differences in the effect of disadvantage as compared to other measure of violence at p < .05 (two-tailed tests).
Assessment of racial invariance using unstandardized disadvantage coefficients predicting different measures of violence for Whites, Blacks, and Latinos. Note: All models include a full set of controls and all disadvantage coefficients are statistically significant predictors at p < .05. W(White), B(Black), L(Latino) indicates significant differences in the effect of disadvantage as compared to other racial/ethnic group at p < .05 (two-tailed tests).
Figure 2 provides an assessment of “racial invariance” by comparing the structural disadvantage coefficients predicting different measures of violence across racial/ethnic groups. Significance tests are used to assess whether the effect of disadvantage on particular markers of violence vary significantly across groups. Lenient tests of the “racial invariance thesis” only require that disadvantage be significantly associated with higher levels of violence for each group, whereas strict tests require that disadvantage has statistically similar violence-generating effects across groups.
Interchangeability of Indices
Beginning with Figure 1 addressing the interchangeability of indices for (A) homicide, (B) robbery, and (C) index violence, the results indicate, first, that disadvantage is significantly (p < .05) associated with higher rates of homicide, robbery, and index violence across each available measure (i.e., offending, arrest, and victimization) and for all racial and ethnic groups. In short, greater disadvantage links to greater prevalence of homicide, robbery, and overall violence for all groups, regardless of whether using offending, arrest, or victimization measures.
Yet, second, F tests indicate that for each racial/ethnic group, the degree of support for the interchangeability of indices depends on the type of violence examined (e.g., homicide, robbery, and index violence). For example, disadvantage predicts homicide offending, arrest, and victimization in statistically similar ways for each group (i.e., the effects reflect an “interchangeability of indices”—see panel A of Figure 1), but disadvantage’s effect varies significantly depending on the measure of robbery used. The relationship between disadvantage and White robbery offending is significantly stronger than its association with White robbery arrest, while the association between disadvantage and Black robbery offending is significantly stronger than its association with arrest or victimization (see panel B). Like robbery, panel C of Figure 1 indicates that for each group, the magnitude of the effect of disadvantage varies significantly across each measure of index violence used—for Blacks, the relationship between disadvantage and offending is stronger as compared to arrest and victimization. The association between disadvantage and White offending is stronger as compared to arrest but weaker as compared to victimization, while the disadvantage–violence link is greater for Latino victimization as compared to arrest.
Racial Invariance
Turning to Figure 2, we examine whether conclusions regarding racial invariance vary depending on the measure of violence (i.e., offending, arrest, and victimization) for the offenses of (A) homicide, (B) robbery, and (C) index violence. As noted in our analysis of the interchangeability of indices, first, disadvantage is significantly (p < .05) associated with higher rates of all three types of violence among Whites, Blacks, and Latinos, regardless of the measure used. However, second, conclusions regarding racial invariance vary somewhat depending on the measure used. For example, while there is strong support for racial invariance thesis for homicide offending and arrest (i.e., structural disadvantage has statistically similar effects on Black and White homicide offending and statistically similar effects on Blacks, Whites, and Latino homicide arrests), disadvantage’s association with Black homicide victimization is significantly stronger than its association for Whites and Latinos (see panel A of Figure 2). Likewise for robbery (panel B of Figure 2), Black–White comparisons indicate that disadvantage has a stronger effect on Blacks when the measure is offending or arrest (the latter compared to Latinos, as well), but statistically similar effects when the measure is victimization. For index violence (panel C), disadvantage has significantly stronger effects on Black compared with White and Latino index violence across each available measure (the Black-White difference for victimization is only significant at p < .1), whereas disadvantage has statistically similar associations with White and Latino index violence across measures of both arrest and victimization.
Supplemental Analyses
In supplemental models, we examined whether the effects of discrete indicators of disadvantage (e.g., poverty,) and, subsequent conclusions regarding racial invariance, varied depending on the measure of violence used (i.e., offending, arrest, and victimization). To this end, we constructed models separately for each of the four indicators of disadvantage comprising our index while simultaneously including all control variables. These supplemental models (available upon request) revealed that discrete indicators of disadvantage are consistently related to higher rates of violence, except for Latinos where the disadvantage index is more robust than its component parts. Additionally, similar to the findings for the disadvantage index, there are varying degrees of strict/lenient support for the invariance thesis when examining the effects of discrete indicators. We also examined whether the association between key control variables and homicide, and subsequent conclusions regarding racial invariance, varied depending on the measure of homicide used. These models (available by request) showed overwhelming consistency in their predictions across groups. As examples, in all (diversity: 7/7 comparisons) or in nearly all circumstances (population density: 6/7 comparisons, immigration: 5/7, residential instability: 5/7), key structural factors had statistically similar associations with homicide across groups regardless of its measure.
Discussion
The current study sought to provide more meaningful treatment of the dependent variable in tests of the racial invariance thesis in order to assess the degree to which measures of violence are interchangeable in race-/ethnic-specific analyses. Specifically, we began by examining whether the relationship between structural disadvantage and violence for specific racial/ethnic groups varied depending on the measure of violence—here, referred to as the “interchangeability of indices.” We then assessed whether the relationship between disadvantage and violence varied across groups depending on the measure of violence used—referred to as the racial invariance thesis. Key findings were, first, that the structural disadvantage index was consistently related to higher rates of violence for each racial/ethnic group for every measure (offense, arrest, and victimization) and for each type of violence (homicide, robbery, and violent index) examined. The finding that disadvantage is a universal source of violence regardless of its measurement provides lenient support that measures of victimization and arrest can be interchanged with measures of offending in race-specific analyses of disadvantage’s effects. Likewise, it also provides lenient support for the racial invariance thesis in that as disadvantage increased, violence was elevated for all racial/ethnic groups.
Yet, second, there was less support for claims of the interchangeability of measures of violence and for the racial invariance thesis when using strict tests that require that disadvantage have statistically similar effects on violence. The degree of strict support varied considerably depending on the type of violence examined. The strongest support for both the interchangeability of measures of violence and racial invariance was observed for homicide, where structural disadvantage operated in statistically similar ways across measures of homicide for each racial/ethnic group and, generally, had statistically similar effects across groups. In contrast, there was far less support for strict tests of either the interchangeability of indices or the invariance perspective for robbery and index violence, where the magnitude of the effects of disadvantage varied significantly across measures and racial/ethnic groups.
The finding that there is stronger support for racial invariance across measures of violence for homicide than for robbery or the violent index can be interpreted in several ways. First, structural disadvantage may exhibit a greater degree of invariance for homicide because the assumptions that researchers employ when using arrest or victimization data in place of offending data may be stronger for homicide than for robbery or index violence (which is heavily dominated by assaults). Recall that a primary assumption in the use of arrest data as a proxy for offending is that racial/ethnic differences in arrest reflect differences in behavior rather than social control and or crime reporting. Homicide is more likely to be reported and arrests for homicide are less subject to influences of racial bias, making such an assumption more reasonable than for robbery or index violence. Likewise, a primary assumption of using victimization data as a proxy for offending is that violence is typically intra-racial, an assumption that may be stronger for homicide compared to robbery. As a second interpretation, noted by Steffensmeier and colleagues (2010), the invariance assumption may only apply—or apply most acutely—to homicide. This interpretation is consistent with research showing that structural disadvantage is a stronger predictor of more serious forms of violence, like homicide, than of assaults that constitute the bulk of violence (Schwartz, 2006).
These findings have important implications for tests of the racial invariance thesis. Despite some differences in the magnitude of disadvantage’s effects, it is essential to stress the most important finding that the structural disadvantage index was universally associated with higher levels of violence across racial/ethnic groups regardless of its measurement. The fact that disadvantage is associated with higher rates of violence for each racial/ethnic group regardless of how violence is measured demonstrates the robustness of the disadvantage effect. At the same time, these findings illustrate the importance of clarifying the lenience/strictness of tests of the racial invariance thesis. Although strict tests of racial invariance provide the highest standard or benchmark for assessing racial invariance claims, they are in several ways (see earlier review) constraining. For example, though there is some overlap in the distributions of disadvantage across groups, several minority communities in our sample (especially those of Blacks) experience levels of disadvantage that White communities rarely experience. Strict tests here are constraining in that the appropriate counterfactuals for making strict comparisons across communities rarely exist (Peterson & Krivo, 2010). Such constraints may contribute to findings of significant differences in the effects of disadvantage across groups and to conclusions that the causes of crime are not racially variant. An emphasis on these differences may obscure the more important finding that disadvantage is a universal source of violence among all groups. Lenient tests, by contrast, provide more flexibility to adjust for the methodological constraints and realities posed by the thesis, while still allowing scholars to assess its core hypothesis of whether structural disadvantage is universal source of crime and violence for all racial/ethnic groups.
The results of the study also have important implications for better understanding the Latino paradox and for incorporating Latinos into tests of the racial invariance thesis. Recent research suggests that there may be a Latino paradox whereby structural disadvantage has weaker effects on Latino violence than for other racial/ethnic groups. On one hand, our finding that the structural disadvantage index elevates violence among Latinos regardless of how violence is measured lends more support to the racial invariance perspective than to the Latino paradox. On the other hand, we did find that disadvantage, in general, (a) had significantly stronger effects on Latino robbery and index-violence victimization than on their patterns of offending for these respective crimes, (b) it often had stronger effects on Black and White violence than on Latino violence, and (c) compared to Blacks and Whites, discrete indicators of disadvantage were less consistently related to higher rates of violence among Latinos. These findings all provide some support for the Latino paradox and suggest that there is a pressing need for further research that incorporates Latinos into analyses that examine the unique effects of structural predictors on their levels of violence across multiple measures and types. As recent scholarship by Hawkins, McKean, White, and Martin (2017) demonstrates, it remains important to examine the correlates of violence for specific racial/ethnic groups because many criminogenic processes cannot be uncovered by studying the total population whose experiences in the aggregate diverge from one another.
Last, we note that unlike much prior research, we find that, in several comparisons, disadvantage predicts violence more closely for Blacks than for Whites, particularly for robbery and index violence. Some possible reasons for this include (a) the use of smaller geographic units than the bulk of prior research (which has focused on Metropolitan Statistical Areas (MSAs)), (b) the sample of localities covered by NIBRS, which tend to be heavily dominated by the South, and/or (c) the use of more recent time frame for assessing racial invariance. Prior research suggests that these factors—size of place (Feldmeyer, Steffensmeier, & Ulmer, 2013), space (Parker & Pruitt, 2000; Light & Harris, 2012), and time (Light & Ulmer 2016)—all have important implications for assessing racial invariance. Notably, much of this research has been limited to Black–White comparisons and or homicide victimization. Future research should reassess the implications of these dimensions for conclusions regarding racial invariance while incorporating other racial/ethnic groups and additional measures of violence.
While we see the current study as making important advances, there remain several limitations that future research should address. First, because we lacked a marker of Latino offending, we were unable to assess whether disadvantage affects Latino arrests and victimizations in similar ways to Latino offending. As reporting of Hispanic identifiers for offenders improves, it will be important for research to replicate the current analysis across measures of offending. Second, our study is restricted to official sources of data and thus misses the “dark figure of crime” not reported to the police. Future research should extend this study by comparing findings from official data sources to self-report and victimization surveys.
The current study drew attention to key theoretical and methodological issues in race–crime research that have received little attention: Whether different measures of violence can interchange for each other in race-/ethnic-specific research, and in turn whether conclusions regarding racial invariance vary depending on the measure of violence used. Our study found that although structural disadvantage is a universal source of violence across racial/ethnic groups regardless of its measurement (victimization, arrest, and offending), there are key differences in the sources of race-/ethnic-specific violence that are linked to its measurement. We hope that our research might serve as a template for future racial invariance studies that seek to integrate and understand the complex ways in which structural disadvantage shapes race-specific violence.
Supplemental Material
Supplemental Material, Supplemental_RJ - Race/Ethnicity and Measures of Violence at the Macro Level: Is Disadvantage Invariant Across Race-/Ethnicity-Specific Arrest, Victimization, and Offending?
Supplemental Material, Supplemental_RJ for Race/Ethnicity and Measures of Violence at the Macro Level: Is Disadvantage Invariant Across Race-/Ethnicity-Specific Arrest, Victimization, and Offending? by Noah Painter-Davis, and Casey T. Harris in Race and Justice
Footnotes
Appendix A
Appendix B
Seemingly Unrelated Regression (SUR) Models of Disadvantage Index Predicting White, Black, and Latino Violence Across Measures of Offending, Arrest, and Victimization.
| Offending (O) | Arrest (A) | Victimization (V) | F Tests Across Measures | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Negative binomial SUR of race-/ethnic-specific homicide, tests of interchangeability of indices and racial invariance | |||||||||||||||
| R2 | b | SE | R2 | b | SE | R2 | b | SE | O-A | O-V | A-V | ||||
| Whites | .05 | .30 | (.04) | * | .07 | .26 | (.05) | * | .08 | .26 | (.04) | * | NS | NS | NS |
| Blacks | .07 | .35 | (.05) | * | .07 | .33 | (.06) | * | .10 | .38 | (.05) | * | NS | NS | NS |
| Latinos | — | — | — | .04 | .21 | (.09) | * | .07 | .19 | (.07) | * | — | — | NS | |
| F Tests Across Groups | W–B | W–L | B–L | W–B | W–L | B–L | W–B | W–L | B–L | ||||||
| NS | — | — | NS | NS | NS | * | NS | * | |||||||
| SUR of race-/ethnic-specific robbery rates, tests of interchangeability of indices (panel A) and racial invariance (panel B) | |||||||||||||||
| Panel A | R2 | b | SE | R2 | b | SE | R2 | b | SE | O–A | O–V | A–V | |||
| Whites | .40 | 1.07 | (.09) | * | .30 | .49 | (.05) | * | .52 | 1.09 | (.11) | * | * | NS | * |
| Blacks | .33 | 3.01 | (.31) | * | .17 | 1.30 | (.17) | * | .41 | 1.54 | (.16) | * | * | * | NS |
| Latinos | — | — | — | .24 | .42 | (.13) | * | .37 | .93 | (.20) | * | — | — | * | |
| Panel B | R2 | b | se | R2 | b | se | R2 | b | se | ||||||
| Whites | .39 | .90 | (.07) | * | .29 | .43 | (.04) | * | .51 | .81 | (.08) | * | |||
| Blacks | .32 | 2.21 | (.25) | * | .15 | .88 | (.14) | * | .38 | .81 | (.12) | * | |||
| Latinos | — | — | — | .24 | .35 | (.13) | * | .36 | .58 | (.17) | * | ||||
| F tests across groups | W-B | W-L | B-L | W-B | W-L | B-L | W-B | W-L | B-L | ||||||
| * | — | — | * | NS | * | NS | NS | NS | |||||||
| SUR of race-/ethnic-specific violent index rates, tests of interchangeability of indices (panel A) and racial invariance (panel B) | |||||||||||||||
| Panel A | R2 | b | SE | R2 | b | SE | R2 | b | SE | O-A | O-V | A-V | |||
| Whites | .36 | 1.79 | (.15) | * | .27 | 1.01 | (.11) | * | .44 | 2.01 | (.17) | * | * | * | * |
| Blacks | .35 | 4.64 | (.44) | * | .22 | 2.20 | (.30) | * | .35 | 3.25 | (.32) | * | * | * | * |
| Latinos | — | — | — | .26 | 1.10 | (.19) | * | .34 | 1.67 | (.27) | * | — | — | * | |
| Panel B | R2 | b | SE | R2 | b | SE | R2 | b | SE | ||||||
| Whites | .35 | 1.45 | (.12) | * | .26 | .79 | (.08) | * | .43 | 1.57 | (.12) | * | |||
| Blacks | .34 | 3.50 | (.34) | * | .21 | 1.47 | (.20) | * | .32 | 1.97 | (.23) | * | |||
| Latinos | — | — | — | .25 | .87 | (.18) | * | .34 | 1.18 | (.24) | * | ||||
| F tests across groups | W-B | W-L | B-L | W-B | W-L | B-L | W-B | W-L | B-L | ||||||
| * | — | — | * | NS | * | † | NS | * | |||||||
Note. N = 453. All models include full set of control variables. For both robbery and index violence, separate SUR models were estimated for tests of the interchangeability of indices and racial invariance.
† p < .10, *p < .05, NS, not significant at p < .10 (two tailed) for both coefficients and F tests.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by National Institute of Justice (2015-R2-CX-0046).
Supplemental Material
Supplemental material for this article is available online.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
