GROUP DIFFERENCES IN DELINQUENCY: WHAT IS THERE TO EXPLAIN?

Abstract

Race and ethnic difference in delinquency are examined using the National Longitudinal Study of Adolescent Health. It is argued that crime theories that attempt to explain race and ethnic differences imply consistent effects for different offenses and common mediating processes. Analyses suggest some degree of group consistency in delinquent behaviors for Mexican-Americans, Puerto Ricans, Native Americans, and some Asian groups, but not for African-Americans. Black youth have higher rates of violent offenses than white youth, lower rates of substance use, and similar rates of property offending. Some variables are consistent mediators while others are not. Crime theories can account for the low rates of delinquency among Asian Americans while theories of violence and substance use are needed to understand differences between black and white youth. The findings are inconsistent with the idea that group differences among youth are due to the socioeconomic status of their families or neighborhoods. The race patterns are also inconsistent with the stereotype of high crime rates in black communities.

Criminologists have a long-standing interest in race and ethnic differences in crime and delinquency. Their explanations of these differences usually focus on why groups with low status have higher crime rates than groups with high status. Theorists argue, for example, that members of low status groups engage in more crime because their opportunities for legitimate goal attainment are blocked, because they experience general strain, or because they are segregated in socially disorganized neighborhoods (e.g., Agnew 2006; Merton 1968; Shaw and McKay 1942; Sampson and Wilson 1995). Other scholars attribute higher crime rates among subordinate groups to weak informal social controls, ineffective parenting and low self-control, or differential association and membership in deviant subcultures (e.g., Gottfredson and Hirschi 1990; Akers 1998: Cohen 1955; Cloward and Ohlin 1960).

In this research, we use data from the National Longitudinal Study of Adolescent Health (Add Health) to compare the prevalence of delinquent behaviors among African-Americans, Native Americans, and specific Hispanic and Asian groups, to the prevalence of delinquency among the non-Hispanic white majority. We focus on the most common types of delinquency--violent offenses, property offenses, and substance use—and on standard mediating variables.

Our goal is to determine whether a theory of crime or theories of particular types of crime (e.g., violence) can better explain group differences.¹ Thus we are interested in establishing the proper dependent variable. We are not attempting to test specific criminological theories or to make claims about whether one crime theory is better than another. Nor are we attempting to explain individual differences. A crime theory may explain individual differences even if it does not explain a particular group difference.

We begin with a review of research that has examined the issue of group consistency in the commission of different crimes. We then discuss mediation of these group differences and the implications for theory testing and development. After discussing mediation and suppression, we present our empirical analyses. Finally, in our discussion, we consider the implication of our findings for theories of crime and specific types of crime, and for what has been described as “the majority-minority paradigm” (Sakamoto et al. 2009), which rests on the argument that structural disadvantage explains group differences.

Finding the Right Street

Typically, criminologists use theories of crime (i.e., explanations of law-breaking behavior generally) rather than theories of specific types of crime to understand individual-level criminal or delinquent behavior. When researchers test crime theories, they often use delinquency scales or select one type of crime (usually violent crime) as an indicator of the underlying concept. They assume that individual offenders show some degree of versatility, which many of them do (Gottfredson and Hirschi 1990). This approach, however, may not be useful to the extent offenders specialize in particular types of crime. Individual specialization implies that different types of offending have different causes, and that theory and research should focus on specific types of crime.

When searching for a house, one first finds out what street it’s on (Felson 2008). Similarly, when attempting to explain human behavior, one should first determine the category of behavior that requires explanation. This inductive approach follows Stinchcombe’s (1968, p. 41) suggestion in his classic work on theory construction (see also, Hempel and Oppenheim 1948). While not a criminologist, he used the study of delinquency as an example of where researchers may be attempting to explain the wrong outcome; the “explanandum” could be specific types of delinquency. Gottfredson and Hirschi (1990), on the other hand, favored a broader conceptualization of the dependent variable in their General Theory of Crime. For them, the explanandum is crime and analogous behavior, not specific types of crime. Criminal behavior is also the explanandum for social control theories and social disorganization theory, while the explanandum for strain theory is different in different versions of the theory (e.g., Hirschi 1969; Sampson and Wilson 1995; Agnew 2006). Finally, social learning and rational choice theory can accommodate both specialization and versatility, since they are general theories of human behavior. However, one must discuss specific mechanisms to explain why individuals engage in specific types of crime. For example, one can use social learning theory to explain the intergenerational transmission of violence, but the focus is on the modeling of violence, not crime generally.

Similar principles apply when one is attempting to explain group differences in crime. When crime theories are used to explain group differences, they imply that groups with higher rates of one type of crime should also have higher rates of other types of crime. For example, a minority group with relatively high rates of violent crime should also have relatively high rates of property crime and substance abuse, compared to the majority’s crime rates. On the other hand, if a group has a high rate of substance abuse, but not other crime, then a theory of substance abuse may be needed to explain group differences. In sum, consistent effects imply that a crime theory is needed to explain group differences, while inconsistency implies that we may need more specialized theoretical explanations.

A crime theory also has implications for the examination of variables that mediate group differences. It implies that the processes producing group differences should be similar for different types of crime and that they should help explain why some groups have lower rates of crime than others. The mediation process should be similar for different groups.

Research on Group Consistency

Research based on self-report data generally shows racial inconsistency across delinquent behaviors. A study based on Adolescent Health data found that blacks have higher rates of violent crime than whites, but no higher rates of property or drug crime (Felson et al. 2008; see also Raudenbush, Johnson and Sampson 2003; Zimring and Hawkins 1997; Greenberg 2008; Beaver et al., 2013). The evidence also suggested that the race difference in violence did not involve a general tendency for blacks to commit more serious crime. While race effects were stronger for armed violence than unarmed violence, they were no stronger for injurious violence vs. other violence, for selling vs. using drugs, or stealing more than $50 vs. less. Some evidence suggests that blacks are more likely than whites to under-report offenses (e.g., Hindelang et al. 1981; Tomaskovic-Devey et al. 2006). If that is the case, then blacks have a higher rate of crime than whites, but a particularly high rate of violent crime. A theory of crime and a theory of aggression may both be necessary to account for black-white differences.

Other studies find mixed evidence as to whether race differences are only observed for violence. For example, Hindelang et al. (1981) examined self-reported delinquency among adolescents in Seattle in 1978. Similar to Felson et al. (2008), they found that blacks were more likely than whites to engage in violence, but not most property or drug crimes.² However, blacks were also more likely to engage in more serious forms of theft. In recent analyses of the National Youth Survey, Wright and Younts (2009) also found differences in black-white offending across violent, property, and drug offenses. They did not find statistically significant race differences in violence or property crimes, but did find that blacks had significantly lower rates of drug and disorder crimes.

Studies based on official arrest data yield much stronger race differences. Data from the 2007 Uniform Crime Reports suggests that black youth are much more likely to be arrested for violent crime than white youth, and only somewhat more likely to be arrested for property crime and drug offenses (U.S. Department of Justice 2008; see also, Sampson and Lauritsen 1997; Baglivio, Jackowski, Greenwald, and Howell 2014; DeLisi, 2011).³

Only a few studies have examined offending for specific Hispanic and Asian groups. Felson et al. (2008) found that Puerto Ricans have higher rates of violent crimes and minor property crime than non-Hispanic whites, but not serious property crime or drug crimes. The differences between whites and other Hispanic groups (Mexicans, Cubans, and Central/South American) were also not statistically significant. McNulty and Bellair (2003) examined Asians and Native Americans in their analysis of Add Health data, but only examined differences in violence. They found that Native Americans had higher rates of violence than whites, while Asian Americans had lower rates (see also, Kaufman et al. 2005). They did not distinguish between different Asian groups. Finally, Morris et al. (2007) found that Native Americans in Oklahoma had higher rates of violence and vandalism than whites, but not higher rates of property crime.

In sum, there is limited research on group consistency among different racial and ethnic groups. The existing survey research suggests that blacks probably have higher rates of violence but not crime generally. The research on other groups is more limited and has provided more inconsistent results. In addition, we are not aware of any research that has examined specific Asian groups. It cannot be assumed that Asians (or Hispanics) with different national origins behave similarly. Finally, estimates of group differences in specific offenses in prior research have not adjusted for the severity of offenses. It may be that what appears to be a group difference in some offenses in fact captures between-offense differences in severity. An adequate test of group consistency must address this issue.

Mediation and Suppression of Group Differences

Evidence of group consistency alone may not be sufficient to support a theory of crime. If groups show consistency in offending, but the variables favored by the crime theory only affects some types of delinquency, then the theory will not adequately explain group differences. We again may need to substitute a more specialized theory.

Different crime theories suggest different mechanisms for explaining group differences in offending. Support for a particular theory requires that: (1) groups differ on that mechanism in the theoretically expected directions; (2) the mechanism has similar effects on different offenses; and (3) the mechanism explains a portion of the group differences in offending. If these three conditions are met, the proposed mechanism and theory contribute to our understanding of race/ethnic differences across offenses.

Mediation is clearest when groups are consistent in their offending and a proposed mediator has a strong effect in the same direction across offenses. If the groups then differ on the mediator, we expect to see at least some mediation of the group differences. However, observing group inconsistency does not necessarily mean that theories of crime are irrelevant. It may be that an unobserved variable suppresses group differences for a particular crime. For example, variables related to opportunity may explain why group differences are not observed for a particular type of crime (Gottfredson and Hirschi 1990). Motivated offenders may not commit a particular type of offense if they cannot find worthwhile targets or if guardianship is effective for that specific offense (Felson and Boba 2010). In other words, offenders would commit the crime if they could, but they lack the opportunity. Such a pattern would imply that we need to supplement, not substitute for, the crime theory in question.

Simultaneously studying consistency in group differences and mediating processes may help us understand what accounts for group differences and thus inform theory development. For example, suppose we find evidence that blacks are less likely than whites to have a strong school commitment and that school commitment is negatively associated with both violent and property offenses. So far, it sounds like school commitment is mediating race differences. However, suppose we also find that race has a total effect on violent offending but not property offending. The same routes are ending in different destinations. The pattern suggests that we need to look for a suppressor effect for property offending, i.e., some unobserved factor that offsets the tendency of blacks to engage in property crime because of their weak school commitment. On the other hand, if we find that a potential mediator is affecting one outcome but not another, then a specialized theory seems more adequate. For example, suppose our theory argues that school commitment inhibits delinquency, but evidence suggests school commitment affects violent offending but not property offending. In that case, we may need a theory of violence rather than a theory of crime to explain the group differences. The routes and the destination are different.

Crime theories should also be able to explain why some ethnic groups have lower rates of delinquency than whites. For example, research shows that Asian Americans have higher levels of education than whites and have achieved parity in the labor market (e.g., Sakamoto et al. 2009). That success has been attributed to structural and cultural factors that lead Asian Americans to work hard, to have strong family ties, and to emphasize their children’s education (e.g., Kitano and Sue 1973; Kao et al. 1996; Hao and Bonstead-Bruns 1998; Steinberg et al. 1992; Kao 1995; Pong et al. 2005). If Asian Americans have strong bonds to school and family and relatively high socioeconomic status, then they should have relatively low rates of all types of delinquency. Criminology has paid little attention to Asian Americans and even less attention to specific Asian groups. We seek to address these omissions with the current analyses. It is just as theoretically important to explain why some groups have low rates of delinquency as it is to explain why some groups have high rates.

Most studies of racial group mediation focus solely on group differences in violence (McNulty and Bellair 2003; Haynie and Payne 2006; Sampson et al. 2005; Kaufman et al. 2005; Farrington, et al. 2003).⁴ An exception is Wright and Younts (2009), who studied both positive and negative mediators (or suppressors) of race differences in delinquency using data from the National Youth Survey. They argued that various beliefs and institutions have developed within African American communities that decrease criminal behavior, and that it is important to look at both positive and negative mediators of race differences. They found that some factors, such as single-parent families and lower educational attainment, were associated with an increase in delinquency among blacks relative to whites, while other factors, such as increased religiosity and strong family ties, decreased black delinquency relative to whites. The direct effects of race were reduced when positive mediators were introduced into equations and increased when negative mediators were introduced. In other words, the former produced race differences and the latter suppressed them. They did not examine whether the variables that mediate group differences are similar for different types of delinquency and for different groups.

Current Study

We examine whether minority groups differ from non-Hispanic whites in delinquency generally or only in particular types of delinquency. The patterns will address the issue of whether theories of crime can account for group differences. We begin by examining race and ethnic differences in violence and drunkenness using Add Health’s in-school sample. This sample has over 90,000 students from a nationally representative sample of schools, providing adequate statistical power to compare delinquency rates of a large number of Hispanic and Asian groups to non-Hispanic whites. The delinquency rates of youth whose ancestry is Japanese, Korean, Asian Indian, Vietnamese, and Central/ South Americans have, to our knowledge, never been studied. We then examine race and ethnic differences using the in-home sample. An advantage of this dataset is that it includes more delinquency items, which permits a better analysis of group consistency. The in-home survey includes questions on violent offenses, property offenses, and various forms of substance use. A limitation of this dataset, however, is that only a limited number of ethnic groups have sample sizes sufficient for statistical analyses. We examine two Asian groups (Chinese and Filipino), three Hispanic groups (Mexicans, Puerto Ricans, and Cubans), as well as Native-Americans and African-Americans.

We base our mediation analysis on the in-home sample due to the limited number of measures in the in-school survey. We examine some standard mediating variables, i.e., academic performance, attachment to parents, intact family, religiosity, depression, and school and family measures of socioeconomic status. Researchers have associated these variables with multiple theories of crime, so our results cannot be viewed as a test of any particular one. For example, the effects of academic performance can be explained by control theory, strain theory, the general theory of crime, as well as other theories. For our purposes, it does not matter to which theory a variable is most aligned, since we are only interested in whether these variables play a similar role in explaining group differences for different types of crime. Our focus is on the consistency of the effects of standard mediators not whether they are the best set of mediators.

Because some of the Asian and Hispanic groups include recent immigrants, we include two measures of cultural assimilation as mediating variables: whether the adolescent was foreign born and whether English was the primary language spoken at home. Greenman and Xie (2008), using Add Health data, found that immigration status is an important predictor of delinquency for Chinese, Filipino, Mexican, Puerto Rican, and Cuban youth. For each of these groups, delinquency levels increased between the first and second generations (see also Sampson et al. 2005; Morenoff and Astor 2007). These studies suggest that assimilation leads ethnic group delinquency to converge with the white majority.

We first establish that there are race and ethnic differences in our potential mediating variables. Mediation and/or suppression can only occur if groups differ on these variables. Then we estimate equations in which we use group membership to predict each of our three offense indices. Potential mediating variables are then added to the equations.⁵ If they have similar associations with the offending outcomes, then it suggests that crime theories have utility in explaining crime even when we do not observe group consistency. We also indicate the extent to which the entire set of variables mediates group differences. In an appendix, we show how much group effects decline with the introduction of specific variables or blocks of variables.

It was suggested earlier that groups may differ in the seriousness rather than the type of crimes they commit. For example, a finding that blacks commit more violence than whites, but similar levels of substance use, would be consistent with offense inconsistency, but may reflect that violence is more serious than substance use and that blacks only commit more serious delinquency than whites. One could then argue that crime theories do not apply to minor delinquency (e.g., underage drinking) since those behaviors are normative among adolescents. Moreover, as most self-reported delinquent behaviors in adolescent surveys are minor, a delinquency scale on such surveys may not reveal group differences.

To adjust our race/ethnic and mediator coefficients for offense severity, we estimate multivariate, multilevel Rasch models of violent, property, and substance use items (Raudenbush, Johnson, and Sampson 2003). These models assume that less frequently occurring criminal behaviors are more serious. They apply item response theory (IRT) to scales of dichotomous delinquency items and thus account for varying levels of item seriousness (Osgood, McMorris, and Potenza 2002). They give greater weight to more serious behaviors based on the argument that serious behaviors better capture an individual’s propensity to commit a particular type of crime. Moreover, these models allow us to compare race, ethnic, and mediator coefficients both within and across offending categories, while also adjusting for individuals embedded in similar social contexts (i.e., schools).

By using a Rasch model, we go beyond previous research on group differences in specific types of crime. Whereas Felson et al. (2008) assumed differences in seriousness of offenses, we explicitly model these differences and control for them. We also extend their analyses by examining mediation, by examining Asian groups, and by including an analysis of the large in-school sample.

Data and Methods

We explore our hypotheses using data from the National Longitudinal Study of Adolescent Health (Add Health). Add Health is a school-based, nationally representative survey of American adolescents enrolled in grades 7 through 12 in 1993–1994. The core sample was drawn from 80 high schools stratified by region, urbanicity, size, type, and ethnic composition. For schools not covering grades 7 through 12, feeder middle schools were also selected, bringing the total number of sampled schools to 132. The in-school survey was administered to all attending students (N=90,118) in one class period in the Fall of 1994. For analyses using the in-school sample, we exclude cases not part of the nationally representative sample (8 percent missing). Of the remaining cases, 14 percent were missing data on race, ethnicity, or our delinquency outcomes, resulting in a final sample of 70,750 students.⁶

In Spring of the same academic year as the in-school survey, approximately 200 students, stratified by grade and gender, were sampled from each of the schools and administered detailed in-home questionnaires (N~20,000). This survey repeated most of the in-school items and added more sensitive questions on delinquency, substance use, and other health-related behaviors. Sensitive items were asked using laptop computers and audio-CASI technology. Add Health’s use of computers for eliciting more sensitive information yields higher frequencies of self-reported delinquency than the usual methods (Turner et al. 1998). The in-home survey also included many more measures that can be used in our mediation analyses. For analyses using the in-home sample (wave 1), we exclude cases that were not part of the nationally representative sample or that were missing data on race or ethnicity and any of the delinquency items, resulting in a final sample of 18,060 students. To maintain statistical power and reduce bias resulting from non-random missingness, we impute values for missing data on any remaining independent variables into five datasets using the ICE commands in Stata 11 with all covariates included in the imputation equations. The amount of covariate missingness was modest, and the variable with the most missing values (school grades) accounted for less than five percent of the final sample.

Dependent Variables

The dependent variables for our analyses are self-reported delinquency items taken from the in-school and the in-home surveys. In the in-school survey, we focused our attention on fighting and drunkenness, as measures of property offending and more serious drug use were not available. Fighting is based on the question, “In the past year, how often have you gotten into a physical fight?” with possible responses “never” (0), “1 or 2 times” (1), “3 to 5 times” (2), “6 or 7 times” (3), or “more than 7 times” (4). Drunkenness is based on the question, “During the past twelve months, how often did you get drunk?” with possible responses “never” (0), “once or twice” (1), “once a month or less” (2), “2 or 3 days a month” (3), “once or twice a week” (4), “3 to 5 days a week” (5), and “nearly everyday” (6). Responses to both items were right-skewed, with fifty-four percent of respondents reporting never having fought and sixty-nine percent reporting they had never gotten drunk. We analyzed three types of delinquency from the in-home survey. Violence is based on seven items: “In the past 12 months, how often did you (1) get into a serious physical fight, (2) shoot or stab someone? (3) use or threaten to use a weapon to get something from someone, (4) pulled a knife or gun on someone, (5) hurt someone badly enough to need bandages or care from a doctor or nurse, and (6) take part in a fight where a group of your friends was against another group?” Property offenses are also measured from six items: “In the past 12 months, how often did you (1) take something from a store without paying for it, (2) go into a house or building to steal something, (3) steal something worth more than $50, (4) paint graffiti or signs on someone else’s property or in a public place, (5) deliberately damage property that didn’t belong to you, and (6) steal something worth less than $50?” Finally, we measure substance use with five items: “In the past month, how many times did you (1) drink alcohol, (2) use cocaine, (3) use inhalants, (4) use marijuana, and (5) use other drugs (LSD, PCP, ecstasy, mushrooms, speed, ice, heroin, prescription medications)?”

As the in-home offending items were not measured on the same metrics (e.g., some were dichotomous, some were ordinal, and some were counts), all items were dichotomized based on whether the respondent reported engaging in the behavior or not. The Kuder-Richardson 20 (KR20) scale reliability coefficient is .68 for the violence items, .74 for the property items, and .58 for the substance use items.⁷

Race/Ethnicity

Add Health is useful for examining group differences because it is a large and nationally-representative sample, and because it includes detailed information on race and ethnicity. For the in-school survey, respondents were first asked their Hispanic origin, followed by their racial background. They could answer “yes” to multiple racial and ethnic categories. To classify individuals into mutually-exclusive categories, we first classified them as Hispanic or non-Hispanic. Those who affirmed Hispanic descent were further categorized as Mexican/Chicano, Cuban, Puerto Rican, Central/South American, or other Hispanic. Non-Hispanic respondents were identified as white, black, Asian, Native American, or other race. Asian respondents were further sub-divided into Chinese, Filipino, Japanese, Asian Indian, Korean, Vietnamese, or other Asian categories. Respondents who chose multiple races or unlisted racial/ethnic backgrounds were placed in “mixed race/ethnicity” and “other race/ethnicity” categories, respectively.

Race/ethnic identification in the in-home survey follows a similar process, but with minor modification. As in the in-school questionnaire, respondents were first asked their Hispanic background and a follow-on question on specific Hispanic origin. They were then asked their racial background.⁸ However, following the race questions, respondents were also asked, “Which [category] best describes your racial background?” This allows us to classify mixed-race respondents into the corresponding subjectively-defined and exclusive category. Unfortunately, Asians and Hispanics with mixed descent (e.g., Chinese-Korean or Mexican-Cuban) could not be identified. We therefore classify these individuals (N=261) as “other/mixed” to keep the race and ethnic categories mutually exclusive. Many of the ethnic categories were underrepresented (<1% of the sample), making statistical inference for these groups difficult. We thus focus our efforts on those categories with sufficient size for statistical tests: non-Hispanic white, black, Chinese, Filipino, Native American, Mexican, Puerto Rican, and Cuban. We also include an “other/mixed race or ethnicity” category that includes respondents who self-identified as other Asian, other Hispanic, Central/South American, or unidentified mixed race/ethnicity.

Mediators

As part of our in-home analyses, we determine whether theoretically relevant individual and contextual characteristics explain observed racial or ethnic differences in offending. We examine the following classes of mediators: cultural assimilation, relationship to parents, socioeconomic status, academic performance, religiosity, and depression. Column 1 in Appendix A presents individual-level descriptive statistics for our entire in-home sample.

Our measures of assimilation are based on questions as to whether the respondent was foreign born and whether a foreign language was spoken at home. Note that one could view these as measures of acculturation as well as assimilation. Respondents were coded one if they answered no when asked: “Were you born in the United States?” and zero if they answered yes. They were coded one if they mentioned a foreign language when asked “What language is usually spoken at home?” and zero if they answered “English.”

We have two measures associated with family relationships. Intact family indicates whether or not respondents report living with both biological parents. Parental attachment is the sum of five items referencing relationships with mothers and three items referencing relationships with fathers (α=.86) where students with single parents had only those items for that parent. The five items included: (1) your mother/father is warm and loving toward you; (2) Your mother encourages you to be independent; (3) When you do something wrong that is important; your mother talks about it with you and helps you understand why it is wrong; (4) You are satisfied with the way your mother/father and you communicate with each other; and (5) Overall, you are satisfied with your relationship with your mother/father. Responses ranged from 1, strongly disagree, to 5, strongly agree.

Measures of socioeconomic status include family and school economic characteristics. Family socioeconomic status is based on parents’ education and income. Parents’ education is constructed from students’ self-reports and is based on the parent with the highest educational attainment. Values range from 0, when neither parent achieved above an 8^th grade education, to 5, when at least one parent achieved a post-graduate education. Parents’ income is a parent-reported measure of yearly family income, in dollars. This value was logged to reduce extreme right skew. We standardized the measures using a z-score transformation and averaged their values to create a SES scale. To create a school-level SES measure, we averaged the z-scores for all respondents in the same school.

Our other measures include academic performance, religiosity, and depression. Academic performance is based on the students’ self-reported average grades in the most recent grading period for four courses (English, Math, Science, History). Responses ranged from 1, D or lower, to 4, A. Religiosity is an index of three items (α =.86): (1) “In the past months, how often did you attend religious services?” (2) “How important is religion to you?” and (3) “How often do you pray?” As the scales for these items differ, they were z-transformed and averaged to create the religiosity index. Depression was measured by a depression index of nineteen items (α =.86) based on the Center for Epidemiological Studies Depression Scale (CES-D) (see Meier 2007). Individual items were recorded on a 0 to 3 ordinal scale and summed, with a sample mean of approximately 11.

Finally, we also include covariates for individual age and gender. Although we suspect these exogenous variables will not differ by race or ethnic category, and therefore will not serve as mediators, they are strongly correlated with offending and included in our analyses to ensure proper model specification and coefficient estimation.

Appendix A presents differences in our independent variables across racial and ethnic groups. For each variable, we indicate whether the group is significantly different from non-Hispanic whites. As expected, the groups do not differ significantly by age or gender composition (except that Native Americans are more likely than whites to be male), suggesting that the survey is equally representative of minority and majority boys and girls of secondary school age. Group differences in socioeconomic status are also in the expected directions. Blacks, Native Americans, and Hispanics are more likely than whites and Asians to have parents with lower educational attainment and income. Group differences in school orientation are also generally as expected. Chinese youth have higher school grades than whites and Filipinos, who in turn have higher grades than the other groups.

We also observe differences in parenting. Chinese youth are more likely than white youth to be from intact families, while blacks, Native Americans, and Puerto Ricans are less likely. Blacks, Puerto Ricans, and Mexicans also report less parental closeness than whites. The parental attachment of Asians and whites are generally similar.

Not surprisingly, the results show group differences in immigration. Asians are more likely to be foreign born than Hispanic groups (all mean differences significant at p<.05, except Chinese-Cuban). In addition, Chinese, Mexicans, and Cubans are more likely than whites to speak a foreign language at home. With regard to religiosity, blacks, Filipinos and Mexicans are more religious than whites, whereas Native Americans are less religious. Finally, blacks, Filipinos, Mexicans, and Puerto Ricans report higher depression than whites, Chinese, and Cubans.

Analyses

We conduct two sets of analyses. For the in-school sample, we estimate ordered logistic regression models of our ordinal outcomes in the Stata software package, using mim commands to analyze multiple imputation datasets. We also use SVY commands to adjust our regression models for Add Health’s complex survey design. The survey-adjusted models address observational dependence occurring in Add Health due to error correlation between respondents attending the same school. We also apply post-stratification weights to regain nationally representative estimates by correcting for Add Health’s oversampling of special populations (e.g., genetic, disabled, and ethnic samples) and survey attrition between the waves (Chen and Chantala 2014). Post-stratification weighting is particularly important when examining race differences in Add Health, as blacks from educated families, Cubans, Puerto Ricans, and Chinese adolescents were all oversampled in the original sample.

To analyze violence, property offending, and substance use in the Add Health in-home sample, we employ a multivariate, multi-level Rasch model. Rasch models are a class of IRT models that apply to dichotomous items (Osgood, McMorris, and Potenza 2002). As with other IRT methods, the Rasch model benefits by combining a set of set of survey items into a reliable composite score that makes good use of information from all item responses, regardless of their frequencies. Our Rasch model is formally equivalent to a three-level hierarchical logistic regression (Raudenbush, Johnson, and Sampson 2003). At the first level are the specific offense items and individual offending “propensities” for the three delinquency scales (i.e. violence, property offending, and substance use). Individuals’ delinquency propensities correspond with their placements on linear, interval rulers for each offending category as determined by their category-specific survey responses (Conrad et al. 2010). These propensities may then be expressed as a function of contextually and individually varying covariates (Raudenbush, Johnson, and Sampson 2003). In other words, the multivariate Rasch model enables us to discern if between-individual characteristics (e.g., racial categories and mediating variables at level-two) or between-context characteristics (e.g., school characteristics at level-three) differentially predict involvement in the three delinquency types, each of which is measured on a latent continuum based on responses to type-specific items.

Our Rasch models adjust for varying levels of severity between our delinquency items under the assumption of equal discrimination of each item. Predictions of the specific delinquency items depend on the relative frequency, or severity, of the act in question. By adjusting each act’s log-odds of a “yes” response by its severity, we are able to locate individuals’ offending propensities along a continuum with similar metrics across offense types. A standardized metric also allows for correlation estimations between the offense categories and between-person or between-school coefficient comparisons at the second and third levels, respectively. For a formal presentation of the three-level Rasch equation, see Raudenbush, Johnson, and Sampson (2003:194–5). Our equation differs from theirs only in (1) the number of survey items per offense type, (2) the number of offense types, and (3) school, instead of neighborhood cluster, as the third-level social context. We estimated the models in HLM v.6.0, which also accommodates standard error adjustments for our five multiple imputation datasets.

Results

Table 1 presents coefficients and standard errors for group differences in violence and drunkenness using the large in-school sample. The results show that blacks are inconsistent: they have a higher rate of violence than non-Hispanic whites but a lower rate of intoxication. Native Americans are also inconsistent: they have a higher rate of violence than whites, but similar rates of drunkenness. Some Asian American groups, on the other hand, show evidence of consistency. The coefficients for Chinese and Vietnamese Americans are negative, indicating that they have lower rates of both violence and intoxication than whites. The Filipino, Indian, and Korean Americans are not consistent: the Filipinos engage in less violence but similar drunkenness to whites, while the Indian and Koreans show the reverse pattern. Finally, Hispanic groups are generally similar to whites in their rates of both types of delinquency. The only exception is Puerto Ricans, who have significantly higher rates of violence than whites. We will be better able to examine the issue of consistency in the in-home sample, where we examine effects on three types of delinquency and adjust for offense severity.

Table 1.

Logistic Regressions of Delinquency by Race/Ethnicity, In-School Survey (Survey-Adjusted)

		Fight		Drunk
Race/Ethnicity	N	Coef	(SE)	Coef.	(SE)
Non-Hispanic White (reference)	40,502
Black	10,057	.25 **	(.07)	− 42 ***	(.08)
Chinese	543	−.38 **	(13)	−.86 ***	(19)
Filipino	993	−.24 *	(10)	−.13	(13)
Japanese	243	−.39	(24)	−.02	(.26)
Asian Indian	255	−.10	(12)	−.44 *	(17)
Korean	431	−.07	(12)	−.64 *	(.26)
Vietnamese	315	−.49 *	(19)	−.60 **	(.22)
Other Asian	398	.11	(15)	−.14	(15)
Native American	693	.44 ***	(06)	.003	(06)
Mexican	2,200	.16	(09)	.09	(08)
Puerto Rican	533	.40 **	(15)	.003	(11)
Cuban	693	.23	(34)	−.13	(33)
Central/South American	1,040	−.05	(09)	−.26	(14)
Other Hispanic	504	.13	(09)	−.08	(12)
Other Race/Ethnicity	1,269	.12 *	(.05)	−.30 ***	(.07)
Mixed Race/Ethnicity	10,081	.11 *	(.05)	31 ***	(06)
Threshold Parameter 0–1		.27 ***	(03)	72 ***	(.07)
Threshold Parameter 1–2		1.70 ***	(03)	1.48 ***	(.07)
Threshold Parameter 2–3		2.48 ***	(04)	1.94 ***	(06)
Threshold Parameter 3–4		2.82 ***	(.04)	2.57 ***	(06)
Threshold Parameter 4–5				3.59 ***	(06)
Threshold Parameter 5–6				4.18 ***	(.06)

^***p<.001,

^**p<.01,

^*p<.05 (two-tailed)

Before proceeding to person-level coefficients of our Rasch model using the in-home sample, it is informative to examine the severity estimates for each offense type at level- one. Table 2 presents item coefficients for our delinquency categories in a reduced-form model (i.e., there are no other variables in the equation). A negative coefficient indicates that the behavior is less frequent than the reference behavior, which approximates item severity (Raudenbush, Johnson, and Sampson 2003). The results are what one would expect. For the violence scale, we see that armed assault, robbery, and threatening with a weapon are far less frequent, and therefore more serious, than assault or fighting with a group. For the property scale, burglary and major theft are more serious than property damage, shoplifting, and minor theft. For the substance use items, cocaine and inhalant use are more serious than marijuana and alcohol use. Finally, the evidence shows that violent offenses are not necessarily more serious than property offenses or substance use. For example, unarmed violence is less serious than burglary and the use of serious drugs.

Table 2.

Rasch Model Item Severities

Item	Coefficient	SE
Violent Offenses
Intercepta	−.90	(.05)
Shot or stabbed someone	−4.05	(.07)
Use or threaten to use a weapon to get something
from someone	−3.09	(06)
Pulled a knife or gun on someone	−2.92	(06)
Hurt someone badly enough to need bandages or
care from a doctor or nurse	−.99	(03)
Take part in a fight where a group of your friends
was against another group	−.87	(04)
Property Offenses
Interceptb	−1.51	(.05)
Go into a house or building to steal something	−2.39	(06)
Steal something worth more than $50	−2.31	(05)
Paint graffiti or signs on someone else’s property
or in a public place	−1.57	(06)
Deliberately damage property that didn’t belong
to you	−.54	(06)
Steal something worth less than $50	−.33	(.03)
Substance Use Offenses
Interceptc	−2.00	(.07)
Use cocaine past 30 days	−3.34	(10)
Use inhalants past 30 days	−3.17	(10)
Use other drugs (LSD, PCP, ecstasy, mushrooms, speed, ice, heroin, prescription) past 30 days	−1.99	(.08)
Use marijuana past 30 days	−.31	(.06)

^aReference item “Get into a serious physical fight”

^bReference item “Take something from a store without paying for it”

^cReference item ‘Drink alcohol at least monthly”

We are also able to determine correlations between the latent scales and the amount of school-level variation in the reduced-form model. The scale correlations are (1) r_{violence-property}=.71, (2) r_{violence-substances}=.60, and (3) r_{property-substances}=.65. Although strongly related, the correlations are not so high as to make the scales indistinguishable. At the school-level, the intraclass correlations (ICC) are (1) r_violence=.07, (2) r_property=.08, and (3) r_substances =.16. Substance use appears to vary more between schools than violence and property offenses.

We now include between-person race and ethnic covariates in the level-two equation. This allows us to test whether race and ethnic differences exist across the delinquency categories. Table 3 and Figure 1 show these group differences in offending. The figure helps to visualize comparisons between non-white groups by showing the odds ratios and 95% confidence intervals for each group for all three offending categories, where whites are assigned an odds ratio of 1.

Table 3.

Rasch HLM of Deviant Behavior by Race/Ethnicity, Add Health In-Home Survey (N= 18060 students, 132 schools)

	Violence		Property		Substance Use
Fixed Effects	Coef.	(SE)	Coef.	(SE)	Coef.	(SE)
Individual-Level
Non-Hispanic White (reference)
Black	.62 ***	(.07)	−.10	(.08)	− 44 ***	(09)
Chinese	− 44 **	(16)	−.16	(14)	−1.48 ***	(31)
Filipino	−.01	(.23)	−.03	(.13)	−1.02 ***	(.26)
Native American	84 ***	(.20)	.48 *	(.22)	.26	(.26)
Mexican	.68 ***	(10)	.33 **	(12)	.04	(12)
Puerto Rican	71 ***	(13)	.32 *	(14)	−.07	(16)
Cuban	.12	(08)	.14	(15)	−.01	(13)
Other/Mixed	.18 *	(09)	.11	(11)	−.38 **	(12)
School-Level
Intercept	−1.11 ***	(05)	−1.53 ***	(.06)	−2.21 ***	(.08)
Variance Components
Intra-school variance	2.48 ***		2.76 ***		2.23
Inter-school variance	13 ***		.21 ***		41 ***

^***p<.001,

^**p<.01,

^*p<.05 (two-tailed)

Figure 1.

Odds Ratios of Delinquency by Offending Type and Race/Ethnic Category

The evidence regarding group consistency depends on the group. Similar to our results in the larger sample, we do not observe consistent effects for African Americans. Compared to whites, blacks have higher rates of violence, lower rates of substance use, and similar rates of property offending. These differences are also significant within the black category. Blacks are significantly more likely to commit violence than property offenses (chi-square=78.3, p<.001), violence than substance use (chi-square=164.5, p<.001), and property offending than substance use (chi-square=12.7, p<.001). We do observe consistent effects for Native Americans, who have higher rates of violence, property offenses, and substance use than whites (although substance use differences fail to reach significance). Mexicans and Puerto Ricans show some consistency, but their substance use is similar to whites.

All of the coefficients for the Chinese youth are negative, suggesting some consistency. However, the coefficient for property offenses is weaker than the coefficients for violence and substance use and statistically non-significant. In addition, Chinese substance use is particularly low. The Chinese coefficient for substance use is significantly larger than the Chinese coefficient for violence (chi-square=12.38, p<.001) and property offending (chi-square=16.02, p<.001). Filipinos, on the other hand, are inconsistent: they differ from whites only in substance use. Finally, we do not observe significant differences between whites and Cubans. A consequence of this is that mediation analyses will have no import for the Cuban group.

Table 4 presents estimates from the equation including all of our covariates. Many of the variables are predictive of different types of delinquency in the expected direction: Net of other covariates, gender, assimilation, parental attachment, academic performance, religiosity, and depression are related to all three types of offenses in the same direction. These results suggest that the theoretical mechanisms associated with these variables generalize across behaviors.

Table 4.

Rasch HLM of Deviant Behavior by Race/Ethnicity, Add Health In-Home Survey (N= 18060 students, 132 schools)

	Violence		Property		Substance Use
Fixed Effects	Coef	(SE)	Coef.	(SE)	Coef.	(SE)
Individual-Level
Non-Hispanic White (reference)
Black	.44 ***	(06)	−.21 **	(.08)	−.62 ***	(.08)
Chinese	−.17	(15)	.16	(12)	−1.00 **	(30)
Filipino	.30	(.20)	.34 **	(12)	−.55 *	(24)
Native American	.51 **	(.20)	.30	(.20)	.04	(.25)
Mexican	.50 ***	(09)	.42 ***	(11)	.19 *	(09)
Puerto Rican	.50 ***	(12)	.25	(16)	−.16	(16)
Cuban	.22 **	(08)	.38 *	(16)	.15	(12)
Other/Mixed	.33 ***	(09)	.37 ***	(10)	−.10	(11)
Female	−1.09 ***	(04)	− 72 ***	(04)	− 25 ***	(.05)
Age	−.12 ***	(01)	−.15 ***	(.02)	.24 ***	(.02)
Foreign Born	− 40 ***	(08)	−.39 ***	(.07)	−.59 ***	(13)
English Not Primary at Home	−.15 *	(.07)	−.35 ***	(09)	−.59 ***	(10)
Parental Attachment	− 02 ***	(00)	− 04 ***	(00)	− 04 ***	(.00)
Intact Family	−.06	(05)	.23 ***	(06)	.11	(.06)
SES	−.12 ***	(03)	.05	(03)	.08 *	(.03)
Grades	− 45 ***	(03)	−.30 ***	(03)	− 41 ***	(04)
Religiosity	−.10 ***	(03)	−.15 ***	(.02)	−22 ***	(03)
Depression	.05 ***	(00)	.05 ***	(00)	.05 ***	(00)
School-Level
School SES	−.19	(10)	.58 ***	(15)	.68 ***	(14)
Intercept	2.46 ***	(26)	2.10 ***	(30)	−4.67 ***	(33)
Variance Components
Intra-school variance	2.06 ***		2.39 ***		1.92 ***
Inter-school variance	.05 ***		14 ***		.09 ***

^***p<.001,

^**p<.01,

^*p<.05 (two-tailed)

The measures of socioeconomic status, on the other hand, are not consistent predictors of delinquency. This is largely due to unexpected effects on non-violent offending. At both the parent and school level, SES is negatively related to violence, but positively related to property offending and substance use. In addition, the school SES variable explains over 50% of the school-level variance in violence and substance use, and close to this amount of school-level property offending.⁹

Age and intact family are also inconsistent predictors of delinquency. Net of other covariates, age is negatively correlated with violence and property offending, but positively correlated with substance use. Being from an intact family is only related to property offenses, and this association is in the unexpected positive direction.

Finally, we consider the extent to which we can explain group differences with the entire set of mediating variables and with each set of mediators entered separately into the equation (Appendix B). For a global mediation analysis, we compare the group effects with all controls in Table 4 to the group effects without controls in Table 3. Results (not shown) demonstrate that, for the significant main effects in Table 3, our mediators explain up to 40% of the Native American effect on violence, 40% of the Puerto Rican effect on property offending, and 50% of the Chinese effect on substance use. Examining each set of mediators individually (Appendix B), we see that grades are the most effective mediator across all offense/group combinations (e.g., 77% of Chinese violence and 55% of Puerto Rican property offending). Note, however, that the grades coefficient mediates the race effect on violence while suppressing the race effect on substance use (Wright and Younts 2009).

Discussion

We used a large, nationally representative sample to examine group differences in different types of delinquency. Our empirical contribution is to examine group differences in specific offenses and the role of potential mediating variables using a sophisticated method that controls for the seriousness of offenses. Our theoretical contribution is to show the implication of consistent and inconsistent effects for different theoretical explanations of group differences in crime. We are not attempting to use the mediation analysis to test specific theories of crime. Rather, our purpose is to determine whether explanations of crime generally, or explanations of specific crimes are required to explain group differences. In other words, we want to establish the proper dependent variable.

We argued that the pattern of group differences has important implications for whether a crime theory can explain those differences. Crime theories, i.e., theoretical explanations of why people break the law, imply that similar group differences and similar mediating variables should be observed for different types of crime. Exceptions could mean that motivated offenders would commit the crime if they could, but that they lack the opportunity. On the other hand, if effects are not observed for multiple types of crime, we should be skeptical that crime theories can provide an adequate explanation for a group difference. Skepticism should be even greater if we observe sign reversals, i.e., effects that are in the opposite direction predicted by the crime theory. Sign reversals are difficult to attribute to measurement problems or differences in opportunity. Inconsistent group effects suggest that we must be more specific in our thinking about the dependent variable--our explanandum. We should either use more specialized theoretical explanations of particular types of crime or propose suppressing mechanisms. We should not ignore exceptions, particularly when there are so many and when the signs of coefficients are in different directions.

Black/White Differences

We do not observe consistent effects of race for different offenses. While black youth are more likely than white youth to engage in violence, they are equally likely to commit property offenses, and they are less likely to use illegal substances.¹⁰ The inconsistent pattern challenges the stereotypical image of the criminality of black communities. It is also a challenge to the idea that crime theories can explain race differences. In addition, our results suggest that the race pattern is not due to any tendency for African-Americans to commit more serious offenses. The Rasch model allowed us to examine race differences controlling for the seriousness of offenses. We acknowledge that use of infrequency as a measure of seriousness is indirect and that it could be affected by opportunity factors as well. However, Felson et al. (2008) found a similar race pattern when making pair-wise comparisons of crimes that clearly differed in their seriousness. For example, they compared injurious violence to violence, and selling drugs to using drugs.

It is possible that the patterns we observe are affected by a tendency for blacks to underreport offenses (e.g., Tomaskovic-Devey et al. 2006). It seems unlikely that such a bias would affect race differences in reporting some offenses but not others, however. Why should blacks report violence but not drinking or theft? While reporting bias could be stronger for more serious offenses since the social desirability factor could be stronger, the Rasch method controls for seriousness. On the other hand, perhaps a reporting bias offsets the tendency for blacks to have higher rates of property offenses and completely reverses effects on substance use. We might then conclude that theories of crime as well as theories of aggression are necessary to explain race differences, but that theories of aggression are more important.

It is also possible that race differences in the tendency of offenders to drop out of school can explain the patterns. Perhaps school authorities are more likely to expel black students than white students for engaging in violence, but they do not discriminate for nonviolent crimes. Note that the opposite argument could be made just as easily.

Overall, the findings suggest that a theory of aggression is needed to explain the most prominent racial pattern. We have neither evidence nor an interest in making claims as to what particular theory should be used. Perhaps blacks respond to discrimination with anger and this anger only leads to crimes involving aggression, as suggested by frustration-aggression approaches (e.g., Berkowitz 1989; Unniver et al., 2009). Perhaps race differences in violence can be attributed to a subculture of violence, an honor culture, or a “code of the streets” in African-American communities (Anderson 1999). Perhaps the spread of violence in African-American communities is due to residential segregation and the contagiousness of violence (Felson et al. 2008; Fagan, Wilkinson and Davies 2007). We leave it to future research to decide which theory of violence works best.

Our results also suggest that we need to explain why white youth are more likely than black youth (and other races and ethnicities) to use alcohol and drugs. A theory to explain these patterns is, to our knowledge, not currently available.¹¹ One possibility is that white youth participate in a party subculture that involves extensive alcohol and drug use. Whatever the process involved, the results suggest that future research on race differences in delinquency should emphasize theories of aggression and substance use.

Ethnic Differences

These findings highlight the need to think about subgroup heterogeneity when examining racial and ethnic crime differences. There is no reason to expect that all Asian or all Hispanic groups should have the same rates. The experiences and cultures of different ethnic groups are different, even when their ancestry comes from the same continent.

We see clear evidence for consistency among Chinese Americans. The coefficients were all in the same direction and the results were consistent for both samples. However, Chinese youth were particularly unlikely to use illegal substances, so theories of substance use may also help explain their behavior. Vietnamese Americans also showed consistency, but we did not analyze their property offenses. Filipino, Indian and Korean American youth appear to be narrower in their offending. Filipinos have low rates of violence while Indians and Koreans have low rates of substance use. The results suggest that theories of crime should be useful for explaining group differences, but that there are also exceptions that require explanation. In addition, the fact that we are able to demonstrate consistency for some groups suggests that its absence for other groups (i.e., blacks) is real and not likely the result of methodological limitations.

We observe mixed evidence regarding group consistency among Puerto Ricans and Mexicans. These groups are more likely than whites to engage in violence and property offenses, but their rates of substance use are similar. Perhaps crime theories can explain their delinquency, but their tendency to use illegal substances is offset by some other process. The other Hispanic groups are similar to whites in their offending. Finally, we see evidence of group consistency among Native Americans. However, their rates of drunkenness are similar to whites in the larger sample.

We acknowledge that we had only mixed success in determining ethnic group consistency for some groups (e.g., Filipinos). The patterns were not always consistent. Perhaps for some groups the pattern was there but we lacked the cell size to detect significant differences. This issue is visible in our confidence intervals in Figure 1. More research is needed to identify exactly how the delinquency of Native Americans and specific Asian and Hispanics groups differs from the delinquency of non-Hispanic whites.

Mediation

We also addressed the consistency issue by examining mediation processes. We examined whether the role of standard variables associated with crime theories is similar for different types of crime. No one, to our knowledge, has made these comparisons. Our analyses revealed that four group-related variables had consistent effects on different types of delinquency: academic performance, religiosity, depression, and assimilation. Further analyses showed that in some cases these variables were sufficiently strong mediators to reduce race and ethnic coefficients to non-significance (see Appendix B).¹² The results suggest that these variables are useful in explaining group differences in crime. In some cases they may act as suppressors. For example, the greater tendency for whites to steal and use drugs would be even stronger if it were not for the fact that other factors (e.g., academic performance) lead them in the opposite direction.

Family and school socioeconomic status effects differ from the above pattern. They are inconsistent predictors and therefore are not likely to be useful in explaining group differences in delinquency. They were negatively related to violence, but positively related to non-violent delinquency. They may, however, be useful in predicting group differences in violence.

It is possible that the relationship between some of our mediators and delinquency are spurious, given that our data are cross-sectional. That should not affect inferences about mediation. In addition, it is difficult to imagine why spuriousness would produce different correlations for different types of delinquency. For example, it seems unlikely that some unmeasured variable can account for why socioeconomic variables are negatively related to violence, but positively related to non-violent delinquency. Note that cross-sectional data are appropriate for examining group effects since they are clearly exogenous. A longitudinal design with lagged variables underestimates group differences because it only examines group effects on changes in delinquency during the time period between administrations. This may explain why McNulty and Bellair’s (2003) results were different from ours.

Future research is needed to examine whether relatively high rates of drinking and drug use among white youth are due to a party culture. In addition, it is possible that the relatively high rates of violence among black youth would be even higher were it not for their relatively high levels of sobriety. In analyses not presented, we included drunkenness as a control variable in a logistic regression equation predicting violence in the in-school sample. The race difference in violence became even stronger.

The Majority-Minority Paradigm

The emphasis on economic disadvantage and inequality in understanding criminal behavior reflects the general approach to group differences in the social sciences. According to what Sakamoto et al. (2009) term “the majority-minority paradigm,” racial and ethnic minorities have lower socioeconomic status than whites because they have historically been subjected to discrimination or disadvantaged in some other way. In criminology, the majority-minority paradigm implies that racial and ethnic minorities have higher crime rates than whites because they have lower social status or lower economic status: they lack legitimate economic opportunity, experience strain, or are segregated in socially disorganized neighborhoods (e.g., Merton 1968; Sampson and Wilson 1995; Anderson 1999). These circumstances cause crime, or produce poverty or subcultural responses which then lead to crime. In one way or another, the studies cited earlier include structural disadvantage as part of their explanation of race differences.

We think that our evidence, along with evidence from earlier studies, suggests that this paradigm requires revision if it is to explain group differences in delinquency. Our results suggest that the link between lower social and economic status and juvenile delinquency is weak at best. African-Americans—the most disadvantaged group—do not have high rates of delinquency. They only have high rates of violence. Asian Americans—a group with a history of discrimination—generally have low rates of delinquency. Recent immigrants—who are also often disadvantaged—also have low rates of delinquency (Vaughn et al. 2013). While these groups have not suffered as much discrimination as African-Americans, the majority-minority paradigm implies that all minority groups who have suffered discrimination should have higher delinquency rates than non-Hispanic whites.

Our evidence regarding the effects of economic status also challenges current versions of the majority-minority thesis. We do not see support for the idea that youth from disadvantaged families are more likely to engage in delinquency than youth from families with better resources; the finding is consistent with a literature dating back to the 1970s (e.g., Tittle, Villemez and Smith 1978). In fact, adolescents from poor families are less likely to drink and use drugs. Perhaps they are less able to afford them. However, adolescents from poor families do not commit more property offenses, and the fact that the sign of the coefficient for SES is positive (b= .05; see Table 4) argues against measurement error. Poverty appears to be criminogenic only for violent delinquency. Theories that emphasize economic disadvantage cannot easily explain why youth from impoverished families do not show consistency.

Nor do we find support for the idea that attending a school with disadvantaged students leads to delinquency. Youth from disadvantaged schools are less likely to engage in property offending and use illegal substances, and they are similar in their violent offending. One could argue that a contextual measure based on census data might provide a better measure of neighborhood disadvantage. For the reported analyses, we did not use census data because the school-based Add Health data is cross-nested with students’ neighborhoods and many students share the same census blocks. In supplementary analyses, however, we created an individual-level measure of “neighborhood disadvantage” based on factor analyses of block-level census characteristics (see Sampson, Raudenbush, and Earls 1997). When this measure was included in survey-adjusted models, the effects were similar to those we obtained for school disadvantage: adolescents from disadvantaged census blocks were less likely to engage in property offending and substance use.¹³ The positive effect of neighborhood disadvantage on violence was similarly weak and also statistically insignificant. The finding of a null effect of neighborhood socioeconomic status on violence is consistent with some studies and inconsistent with others (e.g., Sampson et al. 2005; Kling, Ludwig, and Katz 2005; McNulty and Bellair 2003). On the other hand, our results regarding substance use are consistent with a recent study of adolescents from schools in the Seattle metropolitan area from 1998 to 2003 (Snedker et al. 2009). That study also found that adolescents living in economically disadvantaged neighborhoods have lower rates of alcohol and marijuana use.

We recognize that discrimination has important consequences for minority groups and that social and economic status play a role in some types of crime. We believe, however, that the majority-minority approach needs to be revised to accommodate the delinquency patterns we observe, and perhaps different mechanisms need to be considered. In the case of race differences, our results suggest that mechanisms leading to violence, particularly gun violence, be examined. For example, we suggested that violence is contagious and spreads in racially segregated communities because of adversary effects and the code of the streets. Such explanations emphasize housing discrimination as a driving force and thus are consistent with the majority-minority paradigm.

Conclusion

Our evidence suggests that crime theories can account for the low rates of delinquency among Asian Americans. We observe consistent group differences for different crimes and consistent mediation. We leave it to future research to determine which crime theories explain low rates of delinquency among Asian-Americans. We think it is an important endeavor since it sheds light on the issue of why other groups have high crime rates. That is a central criminological concern.

Our evidence suggests that theories of crime are likely to be less successful in explaining differences between black and white youth. High rates of violent crime and low rates of drinking and drug use among African-American youth suggest we need to consider theories of aggression and substance use. However, it may be that theories of crime are also relevant but that we were unable to detect small effects.

The results suggest that it is important to pay attention to the behavioral outcomes that require explanation. We recognize that failure to find significant effects for some crimes in some studies is to be expected, given measurement problems. Sign reversals, on the other hand, are a more serious challenge. They suggest that we need to pay careful attention to the dependent variable when examining group differences.

Appendix A. Descriptive Statistics by Race/Ethnicity, Add Health In-Home Survey (Survey-Adjusted)

	Full Sample		Non-HispWhite	Black	Chinese	Filipino	Native American	Mexican	Puerto Rican	Cuban	Other/ Mixed
Variable	Mean (%)	SD	Mean (%)	Mean (%)	Mean (%)	Mean (%)	Mean (%)	Mean (%)	Mean (%)	Mean (%)	Mean (%)
Female	.50		.50	.51	.43	.48	.37 *	.50	.48	.52	.49
Age	15.50	1.82	15.44	15.68	15.44	15.83	15.09	15.48	15.37	15.46	15.78
Foreign Born	.06		.01	.02	.43 *	.55 *	.02	.22 *	.09 *	.33 *	.43 *
English Not Primary at Home	.07		.01	.01	.57 *	.27 *	.03	.45 *	.22 *	.64 *	.47 *
Parental Attachment	28.78	8.32	29.51	25.90 *	30.61	30.32	28.02	28.42 *	26.77 *	28.24	28.55 *
Intact Family	.54		.59	.29 *	.86 *	.68 *	.38 *	.56	.34 *	.54	.58
SES	−.04	.79	.05	−.18 *	.25	.34 *	−.22 *	−.55 *	−.30 *	−.43 *	−.06
Grades	2.80	.78	2.87	2.57 *	3.41 *	2.97	2.48 *	2.63 *	2.46 *	2.54 *	2.83
Religiosity	−.02	.89	−.09	.20 *	−.37	.38 *	−.47 *	.11 *	−.09	−.09	.08 *
Depression	10.94	7.55	10.32	12.16 *	10.68	12.95 *	12.65 *	12.71 *	12.86 *	10.47	11.92 *
N	18060		9674	3831	280	538	144	1496	501	441	1155

^*p<.05 (two-tailed) Survey-Adjusted Wald Test compared to Non-Hispanic Whites

Appendix B. Race/Ethnic Mediation, In-Home Survey

		Violence		Property		Substance Use
Race/Ethnicity		Coef.	% A	Coef.	% A	Coef.	% A
Black
	Without Controls	.62 ***		ns		− 44 ***
	Assimilationa	.62 ***	.00			− 45 ***	.03
	Parentingb	.51 ***	−.18			−.58 ***	.31
	SESc	.59 ***	−.05			− 44 ***	.00
	Grades	.46 ***	−.27			−.59 ***	.34
	Religiosity	.70 ***	.12			− 34 ***	−.24
	Depression	.54 ***	−.13			−.54 ***	.22
Chinese
	Without Controls	−.44 **		ns		−1.48 ***
	Assimilationa	−.17	−.62			−1.00 ***	−.33
	Parentingb	−.38*	−.14			−1.45 ***	−.02
	SESc	−.47 **	.07			−1.48 ***	.00
	Grades	−.10	−.77			−1.23 ***	−.17
	Religiosity	− 49 **	.11			−1.55 ***	.05
	Depression	−.52 **	.19			−1.59 ***	.07
Filipino
	Without Controls	ns		ns		−1.02 ***
	Assimilationa					−.55 *	−.46
	Parentingb					−1.00 ***	−.03
	SESc					−1.00 ***	−.02
	Grades					− 91 ***	−.11
	Religiosity					−.83 **	−.19
	Depression					−1.16 **	.14
Native A	merican
	Without Controls	.84 ***		.48 *		ns
	Assimilationa	.86 ***	.02	.51 *	.06
	Parentingb	.83 ***	−.02	.53 *	.10
	SESc	.80 ***	−.04	.48 *	.01
	Grades	.63 **	−.25	.32	−.34
	Religiosity	.81 ***	−.04	.45 *	−.06
	Depression	71 **	−.16	.34	−.29
Mexicar
	Without Controls	.68 ***		.33 **		ns
	Assimilationa	.83 ***	.22	.58 ***	.75
	Parentingb	.65 ***	−.04	.30 **	−.10
	SESc	.49 ***	−.28	.31 **	−.05
	Grades	.47 ***	−.30	.17	−.49
	Religiosity	.75 ***	.10	.40 **	.20
	Depression	.56 ***	−.18	.21	−.37
Puerto F	ican
	Without Controls	71 ***		.32 *		ns
	Assimilationa	.76 ***	.06	.39 *	.21
	Parentingb	.65 ***	−.09	.30 *	−.08
	SESc	.63 ***	−.11	.32 *	−.01
	Grades	.48 ***	−.32	.15	−.55
	Religiosity	.76 ***	.06	.37 *	.14
	Depression	61 ***	−.14	.22	−.32

^aforeign born and English not primary at home variables

^bparental attachment and intact family variables

^cindividual-level and school-level SES variables

^***p<.001,

^**p<.01,

^*p<.05 (two-tailed)