Explaining Differential Patterns of Self-reported Delinquency

Abstract

A rich line of criminological research has utilized latent class analysis (LCA) to identify different groups or classes of individuals. At the same time, behavioral genetic modeling strategies have offered valuable insight into the etiological origins of many phenotypes, such as aggression and antisocial behavior. For the most part, these two streams of inquiry have operated in isolation of one another. The current study represents an attempt to blend a mainstream criminological method (i.e., LCA) with a mainstream behavioral genetic method (i.e., the ACE model). The results indicated that three classes of offenders are identifiable in the National Longitudinal Study of Adolescent Health and that variance in each class was explained by a combination of genetic (h² ranged between .26 and .65) and nonshared environmental (e² ranged between .35 and .74) influences. Directions for future research are considered.

Keywords

Add Health latent class analysis (LCA)ACE model biosocial criminology

Latent class analysis (LCA) is a statistical tool available to researchers attempting to uncover patterns in a sample. The patterns can reflect a wide variety of phenomena, but generally researchers are interested in determining whether different “types” or “classes” of individuals can be identified. In many respects, LCA is similar to factor analysis—both are used as a data reduction tool. One key difference, however, separates these two techniques. On the one hand, factor analysis reduces a list of variables into a smaller set of factors that can be used in place of the original items. LCA, on the other hand, does not reduce a list of variables, but instead attempts to reduce the number of cases (i.e., respondents) into a smaller set of subgroups or classes. Generally speaking, LCA scours a dataset looking for respondents who exhibit a similar response pattern. Individuals who respond similarly to a list of questions are then grouped together into mutually exclusive classes.

Over the past few decades, LCA has gained increasing momentum among researchers from a variety of disciplines (Formann & Kohlmann, 1998). Indeed, scholars have utilized LCA to uncover different classes of individuals for a wide range of outcomes including eating disorders (Dechartres et al., 2011), risky behaviors such as drunk driving (Laska, Pasch, Lust, Story, & Ehlinger, 2009), academic success (Lanza, Rhoades, Nix, Greenberg, & The Conduct Problems Prevention Research Group, 2010), gambling (Lloyd et al., 2010), sex offending behaviors (Deslauriers-Varin & Beauregard, 2010), intimate partner violence (Klostermann, Mignone, & Chen, 2009), substance use (Agrawal, Lynskey, Madden, Bucholz, & Heath, 2006; Guo, Aveyard, Fielding, & Sutton, 2009; Schwartz, Wetzler, Swanson, & Sung, 2010), victimization (Nylund, Bellmore, Nishina, & Graham, 2007; Reid & Sullivan, 2009), and even criminal/delinquent behavior (Connell, Cook, Aklin, Vanderploeg, & Brex, 2011; DeLisi et al., 2011; D’Unger, Land, & McCall, et al., 2002; Hasking, Scheier, & ben Abdallah, 2011; Sullivan, Childs, & O’Connell, 2010; Vaughn et al., 2011). The latter group of findings—LCA analysis of criminal/delinquent behavior—will be the focus of this article.

Latent Classes of Offenders

Much attention, both theoretical and empirical, has been devoted to identifying the different classes of offenders that exist within the population (Piquero, 2008). Perhaps the most prominent theoretical statement concerning the different classes of offenders is Moffitt’s (1993) developmental taxonomy. Briefly, Moffitt argued that the population of offenders was not made up of one homogenous group. Instead, building on an already large body of evidence, she hypothesized that two different types of offenders existed in the population. The first group is referred to as adolescence-limited (AL) offenders. These offenders, according to Moffitt, restrict their offending to the adolescent years. In other words, AL offenders do not exhibit antisocial behavior in childhood, and they desist from criminal behavior promptly on entering adulthood. AL offenders account for approximately 90% of all offenders and, therefore, are the primary force driving the aggregate age–crime curve that has been so well established in the criminological literature (e.g., Farrington, 1986). Important for the current focus is that AL offenders restrict their offending to relatively minor forms of delinquency, and they do not commit criminal acts at a high rate (Barnes & Beaver, 2010).

The second group identified by Moffitt (1993) is known as life-course-persistent (LCP) offenders. LCP offenders differ from their AL counterparts in four key ways. First, LCP offenders display signs of antisocial behavior in early childhood (i.e., they display problem behaviors that are outside the normal range for children their age). Second, LCPs’ antisocial behavior remains stable throughout adolescence and into adulthood. Third, though LCPs make up only a small portion of the population (~10%), they account for an overwhelming percentage of all crimes committed (>50%) (DeLisi, 2005). Fourth, unlike AL offenders, LCPs are likely to be involved in interpersonal, violent, and serious forms of crime and delinquency. In short, Moffitt’s taxonomy highlighted two groups of offenders that could be identified according to their longitudinal patterns of offending, their relative frequency of criminal behavior, or by the types of crimes they are likely to commit.

Spurred primarily by Moffitt’s (1993) theoretical statements, as well as by the emergence of the criminal careers paradigm (Blumstein, Cohen, Roth, & Visher, 1986), a long line of research has attempted to identify the different offending trajectories that exist in the population (for reviews see Moffitt, 2006; Piquero, Farrington, & Blumstein, 2003). In a recent review, Piquero (2008) discussed the use of the group-based trajectory model (Nagin, 2005) in analyzing longitudinal patterns of offending. This review revealed that more than 80 studies had been conducted in which the longitudinal trajectories of offending were analyzed. In short, the body of evidence surrounding the longitudinal trajectories of offending has revealed that multiple groups of offenders are identifiable in the population and the factors underlying the etiology of one group may not be the same as those underlying a different group.

Though much research has utilized LCA (or similar models such as the semiparametric group-based trajectory model [Nagin, 2005]) to uncover longitudinal offending trajectories (Piquero, 2008), other scholars have utilized LCA to uncover different classes of offenders by grouping individuals according to the types of offenses they are most likely to commit or by the frequency with which they commit certain acts (Connell et al., 2011; DeLisi et al., 2011; D’Unger et al., 2002; Hasking et al., 2011; Sullivan et al., 2010). In other words, cross-sectional data can also be used with LCA. Rather than estimating offending trajectories (a term typically reserved for an estimation of something that unfolds over time; see generally Laub & Sampson, 2003), LCA can be used to estimate offense patterns at discrete points in the life course. Moffitt (1993) notes that LCPs will be more likely to commit interpersonal crimes and that they may be expected to do so at higher rates than ALs. Cross-sectional LCA models can inform this line of inquiry.

Vaughn and his colleagues (2011) analyzed cross-sectional data drawn from the National Epidemiologic Survey on Alcohol and Related Conditions and utilized LCA to group individuals based on their involvement in 34 different behaviors—ranging from minor substance use to interpersonal violence. The LCA indicated that a four-group model best fit the structure of the data. The first group, which comprised the largest portion of the sample (66%), revealed a pattern of behavior that is typical of most adolescents; these respondents admitted to several minor offenses, such as skipping school and getting pulled over for reckless driving, but showed little-to-no involvement in serious/violent behaviors. Two groups displayed moderate involvement in antisocial behavior. Finally, a fourth group of offenders displayed a high frequency of involvement in all 34 behaviors. This final group consisted of less than 10% of the sample. In addition to identifying the four different classes of offenders, Vaughn et al. reported that a range of factors predicted group membership. For example, respondents classified into the high-rate group were more likely to be a racial minority, made less money, and were likely to be unmarried as compared to the normative group of offenders (i.e., Group 1).

Other scholars have reported similar findings using cross-sectional data (Connell et al., 2011; Hasking et al., 2011). Sullivan et al. (2010) identified four groups of respondents who displayed varying involvement in a range of risky behaviors such as engaging in unprotected sex and taking part in drug sales. Like Vaughn et al. (2011), Sullivan and colleagues examined whether a range of influences predicted group membership. Their statistical models revealed that high-rate offenders were more likely to be male, were monitored less frequently by their parents, and received less social support from their parents, teachers, and/or friends as compared to abstainers.

In summary, there is a growing body of research that has utilized LCA to uncover the different patterns or groups of offenders that exist in the population. Although there is some variation, most studies have identified either three or four groups of offenders. Interestingly, these studies share two common threads. First, nearly all studies identify one group of respondents that is relatively inactive in antisocial behavior. Second, one group of offenders—typically made up of a small portion of the sample—exhibits antisocial behavior at a much higher rate as compared to the other group(s). A number of environmental (i.e., social) influences have been tied to group membership, but these factors only accounted for a small portion of the variance. Equally important, however, is that practically every study mentioned above has overlooked the impact of heritable (i.e., genetic) factors on group membership. The current study will demonstrate how LCA can be blended with contemporary behavioral genetic modeling strategies.

Genetic Influences and Class Membership

Criminology has a rich tradition of utilizing LCA to identify and define subgroups of individuals within a given sample. Much is now known about the environmental factors that impart risk on being in one class or another (Farrington, Ttofi, & Coid, 2009; Hasking et al., 2011; Sullivan et al., 2010; Vaughn et al., 2011). What has yet to be considered by criminologists, however, is whether genetic factors may underlie the etiology of the different classes identified by LCA. This oversight is somewhat surprising given the burgeoning body of evidence tying genetic factors to the development of antisocial behavior (e.g., Ferguson, 2010). Indeed, behavioral geneticists have long noted that genetic factors are implicated in the onset of criminal/antisocial behavior (Raine, 1993), and some have even hypothesized that genetics underlie the different patterns of offending (Moffitt, 1993). Yet, only one study has examined this possibility directly (Muthén, Asparouhov, & Rebollo, 2006).

Muthén and colleagues (2006) combined LCA with contemporary behavioral genetic modeling strategies. In their analysis, Muthén et al. estimated an LCA on 11 items, measured cross-sectionally, which tapped alcohol dependence/abuse. For example, respondents were asked to indicate their tolerance for alcohol, whether they had experienced withdrawal symptoms, and whether their drinking had interfered with their lives. The LCA analysis indicated that a two-group solution best fit the data.¹ The next step was to examine the genetic and environmental influences on group membership.² This step utilized a common behavioral genetic modeling strategy known as the ACE model (see below for more information about the ACE model). Importantly, this portion of the analysis indicated that genetic influences may impact group membership.³ Heritability (h²) estimates ranged from .22 for the high-rate group to .86 for the low-rate group. This study is important for the current purposes because it suggests that genetic factors may underlie the different patterns of behavior that can be identified with LCA. The current study will directly examine this possibility using a sample of twin pairs drawn from a nationally representative study of American adolescents.

Current Study

Criminological research has been guided and informed by the empirical results generated from LCA and these results have been used to test some of the most potent criminological theories. In addition, current LCA studies have also examined the correlates to group membership. The current study builds upon previous research in two important ways. First, we provide an LCA of a nationally representative sample of American youths: The National Longitudinal Study of Adolescent Health (Add Health). To our knowledge, this is the first study to ever use these data to conduct an LCA. Our results will thus help to provide additional evidence to the existing pool of LCA studies on antisocial behaviors. Second, we also estimate genetic influences on group membership. If genetic factors are influential, then previous research may be misspecified and future research will need to incorporate controls for genetic effects.

Method

Data

Data for this study were drawn from the Add Health (Udry, 1998). The Add Health is a longitudinal and nationally representative sample of youths who were enrolled in grades 7 through 12 in the 1994-1995 academic year. The general focus of the Add Health study was to assess the health and development of American adolescents. As such, researchers developed an extensive questionnaire that inquired about the respondent’s thoughts, their feelings, their daily activities, and their behaviors (including their delinquent behaviors).

To gain information from these adolescents a sample of high schools was selected by employing stratified random sampling techniques. In short, high schools acted as the primary sampling unit. Initially, 80 eligible high schools were included in the sample frame. Using this list, schools were stratified by region, urbanicity, school type, ethnic mix, and size. After stratification, schools were selected with a probability proportional to their size. This sampling procedure ultimately resulted in 132 schools being included in the study (Harris et al., 2009; Kelley & Peterson, 1997). All students attending these 132 schools were asked to complete a self-report questionnaire, netting information from more than 90,000 respondents (i.e., the in-school survey).

Starting in April 1995 and continuing through December 1995, the Add Health research team collected more detailed information from a subsample of the students that completed the in-school surveys. For example, information was gathered about the respondent’s personality traits, about their social relationships, and about their behaviors. The surveys were conducted using computer-assisted interviewing techniques via a laptop. For most questions, the interviewer read the question aloud to the respondent who provided verbal responses. This procedure, however, could lead to biased responses on sensitive topics such as sexual experiences and delinquent behavior. To mitigate this possibility, audio computer-assisted interviewing procedures were used when sensitive topics became the focus. The respondent was provided earphones and questions were read aloud by the computer. The respondent then entered his/her responses directly into the computer.

This round of interviews—referred to as the wave 1 survey—was carried out in the youth’s home. Not all 90,000 students who completed in-school surveys also completed the in-home interview. Instead, students listed on each school’s roster provided a sample frame from which respondents were chosen using stratified random sampling procedures. Overall, in-home interviews were conducted with 20,745 adolescents. Respondents ranged between 11 and 21 years of age (μ = 16) at wave 1.⁴

Nested within the Add Health data is a subsample of sibling pairs who were residing in the same household during wave 1 interviews (Rowe & Jacobson, 1998). Included in this subsample were monozygotic (MZ) twins, dizygotic (DZ) twins, full siblings, half-siblings, and unrelated youth living together (e.g., adopted siblings and step siblings). The sibling subsample was compiled in a fashion that allowed more than one pair of siblings to be included per household. This meant that more than two siblings per home were sometimes interviewed. In order to remove any possible biases, the current study restricted the sample to two children per household. This process of excluding certain siblings from the sample was carried out using a semistructured selection method. All MZ and DZ twins were selected with certainty and the other sibling pairs were selected at a rate that would ensure sample sizes large enough to perform statistical analyses. The current study utilized the sibling subsample but restricted the final sample to MZ twins, DZ twins, and full siblings to reduce the confounding influences of assortative mating effects (Boutwell & Beaver, 2010; Krueger, Moffitt, Caspi, Bleske, & Silva, 1998; Rhule-Louie & McMahon, 2007).

Self-reported Delinquency

During wave 1 interviews, respondents were asked to report on their involvement in 17 different delinquent activities. Respondents were asked to indicate how often over the past 12 months they had painted graffiti, damaged property, lied to their parents, stolen from a store, gotten into a serious fight, hurt someone badly enough to require medical attention, run away from home, stolen a car, stolen something worth more than US$50, broken into a house, committed an armed robbery, sold drugs, stolen something worth less than US$50, taken part in a group fight, acted loud or unruly in a public place, carried a weapon to school, and used a weapon in a fight. Responses to most of the questions were coded as follows: 0 (i.e., never), 1 (i.e., one or two times), 2 (i.e., three or four times), and 3 (i.e., five or more times). Two questions (i.e., carried a weapon to school and used a weapon in a fight) were coded dichotomously where 1 indicated that the event had occurred and a 0 indicated that the event had not occurred.

Analysis Plan

The analysis unfolded in three interlocking steps. The first step was to examine the descriptive statistics for each of the 17 different delinquent acts. This step primarily involved observing the frequency of response choices for each of the individual items. These analyses allowed for the observation and the initial identification of response patterns that may exist within the data.

The second step of the analysis utilized a statistical technique known as LCA. Generally, LCA is a statistical tool that can be used to identify patterns within a dataset. In this respect, LCA is analogous to factor analysis because the goal is to reduce all cases (i.e., respondents) appearing in the dataset into a finite number of groups based on their response patterns (Formann & Kohlmann, 1998). The current study employed LCA to identify the different groups or subpopulations of adolescents that responded to the 17 delinquency items in similar ways.

Since the delinquency items were operationalized into ordinal categories, it was necessary to estimate the LCA with a technique that accounted for the limited variation of the indicators. To account for this issue, the LCA model was estimated by allowing the thresholds—as opposed to the means or the intercepts—of the 17 indicators to vary across the different classes of respondents.⁵ Thresholds are more appropriate in this case because intercepts and means (parameters that are modeled by LCA for continuous outcomes) are more likely to be skewed for variables with limited variation. The LCA was estimated using the software package, Mplus (version 6.1).

The LCA unfolded in a stepwise fashion: a two-group model was estimated first, followed by successively more complex solutions. This process was carried out until the analysis converged on the “best-fitting model.” In order to identify the best-fitting solution, several model-fit statistics, along with theoretical considerations, were evaluated after each model was estimated (Brame, Paternoster, & Piquero, 2011). Scholars recommend the use of the Bayesian Information Criterion (BIC) for evaluating model fit (Kass & Raftery 1995; Kreuter & Muthén 2008; Nagin, 2005). The Mplus method for calculating the BIC typically results in a positive statistic, meaning that BICs closer to zero reflect a better fitting model. In addition to evaluating the BIC, it is prudent to consider the estimated group proportions and the estimated posterior probabilities when determining the best-fitting model. Posterior probabilities reflect the likelihood that each case (i.e., respondent) belongs to a specific group. In general, the best-fitting model should generate posterior probabilities that are near unity for each group.

The third and final step of the analysis involved estimating the genetic and environmental influences on being classified in the different groups (as estimated by the LCA) (see generally, Muthén et al., 2006). The ACE model (Neale & Maes, 2004) was estimated to determine the relative impact of heritable factors (h²), shared environmental factors (c²), and nonshared environmental factors (e²) on group membership (see Figure 1). Note that the A corresponds to h² (genetic factors), the C corresponds to c² (shared environmental factors), and the E corresponds to e² (nonshared environmental factors and measurement error).

Figure 1.

Balloon Diagram of the ACE Model

Several elements of the ACE model must be discussed. First, the model contains two observed indicators, one for each sibling. There are two types of observed indicators that will be analyzed: the posterior probabilities of group membership as estimated by the LCA model and a dichotomous variable that identifies group assignment (1 = the respondent was assigned to the group of interest. A threshold ACE model was estimated when the outcome was the group assignment variable.). Second, the three latent factors (i.e., A, C, and E) cumulatively explain 100% of the variance in the observed indicators. Third, the path coefficients leading from the latent factors (i.e., a1, c1, e1, a2, c2, and e2) to the observed variables provide estimates of h², c², and e². Fourth, note that the A factors have a correlation (i.e., the double-headed arrow) that can vary between .50 and 1.00. Since the A factors tap heritable influences, the value for the correlation between the A factors corresponds to the level of genetic relatedness between the two siblings providing data. MZ twins share 100% of their DNA, thus the correlation is set to 1.00 when MZ twins are being analyzed. DZ twins and full siblings share, on average, 50% of their distinguishing DNA. As a result, the correlation is set to .50 when DZ twins and full siblings are analyzed. Fifth, the C factors are also correlated, but their correlation is always set to 1.00. The C factor captures the variance that is due to shared environmental influences. Since, by definition, shared environments are always identical between two siblings, the correlation is set to 1.00 for all sibling pairs. Finally, the E factors are left free to vary. By definition, nonshared environmental factors are not shared between two siblings, thus there is no correlation between two siblings on nonshared environmental influences.

To summarize, the analysis examined the different response patterns that are observable among the Add Health participants. Identifying different groups of adolescents based on their responses to the 17 delinquency items was the first goal of the analysis. After identifying different groups of respondents, the next goal of the analysis was to determine the relative impact of genetic (heritable), shared environmental, and nonshared environmental factors on being classified into each group. The results from the analysis are presented in the next section.

Findings

The analysis began by examining the distribution of responses for each of the 17 items. Table 1 displays the mean value for each of the 17 items along with the proportion of respondents who answered with a 0, 1, 2, or 3. As can be seen, the mean values ranged between .05 (weapon in a fight) and .90 (lie to parents). Generally speaking, observation of the mean values revealed that relatively minor behaviors (e.g., lie to parents, rowdy in public, and steal less than US$50) were more prevalent as compared to more serious acts (e.g., break into house, steal more than US$50, and weapon in a fight).

Table 1.

Descriptive Statistics and Response Frequencies for Add Health Delinquency Variables

	Response Category
Item	Mean	0	1	2	3
Graffiti	.12	.92	.05	.01	.01
Damage property	.24	.82	.14	.02	.02
Lie to parents	.90	.46	.32	.10	.12
Steal from store	.38	.76	.16	.04	.05
Serious fight	.43	.70	.22	.05	.04
Hurt someone	.23	.83	.13	.02	.02
Run away	.08	.94	.05	.01	.01
Steal car	.15	.90	.08	.01	.02
Steal more than US$50	.08	.95	.04	.01	.01
Break into house	.08	.95	.03	.01	.01
Use weapon	.06	.96	.03	.01	.01
Sell drugs	.12	.93	.04	.01	.02
Steal less than US$50	.33	.80	.12	.03	.05
Group fight	.25	.81	.14	.03	.02
Rowdy in public	.70	.53	.32	.08	.08
Weapon to school	.08	.92	.08	—	—
Weapon in fight	.05	.95	.05	—	—

In addition to the mean values for each of the delinquency items, Table 1 presents the proportion of respondents who selected the different response categories for each item. A close examination of the table reveals that the majority of respondents selected never (i.e., 0) for all but one of the items. The lone exception was for the item asking about the frequency with which the respondent lied to their parents. Given the nature of this behavior (i.e., its relatively minor status) it is not surprising that the majority of respondents selected a response category other than never.

Though most respondents selected never for each of the items, there was considerable variation across the different questions. For instance, when asked about their involvement in serious fights, approximately 70% of respondents selected never, but roughly 22% indicated that they had been involved in at least one or two serious fights (i.e., they selected 1). Many of the other items showed a similar variation. Also, it is worth noting that a small portion of respondents indicated a fairly high level of involvement for each of the 17 items. This finding suggests that a small portion of respondents may be disproportionately engaging in delinquency as compared against the mean level of involvement. To further explore this possibility, we turn to the results from the LCA.

When estimating the LCA, it was first necessary to settle on the best-fitting model. In order to determine when the best-fitting model had been reached, the BIC, the group proportions, and the posterior probabilities for each model were observed. These statistics can be found in Table 2. The table presents model-fit statistics for three different LCA models. The first analysis estimated a two-group solution.⁶ As can be seen, the group proportions were high (26% in Group 1 and 74% in Group 2) and the posterior probabilities were above .90 for both groups. Combined, these statistics suggest that the two-group model provided a good fit to the data. The two-group model, however, needed to be compared to another solution before it could be deemed the best-fitting model.

Table 2.

Model-Fit Statistics, Group Proportions, and Posterior Probabilities for Latent Class Analysis

	Two-Group Model	Three-Group Model	Four-Group Model
Log likelihood	−28,420.04	−27,596.64	−27,227.27
Bayesian Information Criterion	57,617.05	56,362.83	56,016.66
Group proportions
Group 1	942 (26%)	328 (9%)	477 (13%)
Group 2	2,622 (74%)	1,203 (34%)	244 (7%)
Group 3	—	2,033 (57%)	823 (23%)
Group 4	—	—	2,020 (57%)
Posterior probabilities
Group 1	.94	.92	.90
Group 2	.97	.91	.93
Group 3	—	.94	.85
Group 4	—	—	.93

A three-group model was estimated next and the results from this analysis can be found in the second column of Table 2. A quick glance at these results reveals that the three-group model provided a better fit to the data as compared to the two-group solution. Specifically, the three-group model provided a smaller BIC statistic, the group proportions remained high, and the posterior probabilities were above .90 for all three groups. Moving to the four-group solution, we see that the BIC was smaller than the three-group model, suggesting a superior fit. However, observation of the group proportions suggested that the four-group model may not provide a better fit to the data. To be specific, the proportion of respondents falling into Group 2 was relatively small. Also, the posterior probabilities dipped below .90 for Group 3 in the four-group model. Based on the full range of evidence, it was determined that the three-group solution was the best-fitting model.

In an effort to fully grasp and visualize the results from the three-group model, the predicted probability of answering each of the 17 items with a nonzero response (i.e., indicating involvement in the act) was graphed according to group status. As displayed in Figure 2, there are three clearly delineated response patterns. Individuals in Group 1, which comprised approximately 10% of the sample, were more likely to give a nonzero response to all 17 items as compared to individuals in each of the other two groups. In other words, individuals in Group 1 indicated that they were involved in many different activities. Adolescents falling into Group 2 showed a more moderate trend; these respondents not only indicated relatively high rates of minor acts such as lying to parents and acting rowdy in public but also indicated relatively low rates of serious criminal acts such as breaking into a house and using a weapon to take something from someone. Finally, individuals classified into Group 3 revealed almost no involvement in any of the 17 delinquent acts. Indeed, these youth showed only modest probabilities of admitting that they had lied to their parents or acted rowdy in public over the last 12 months—two of the most common behaviors.

Figure 2.

Graphical Depiction of the Three-group Latent Class Analysis

The final step of the analysis was to estimate the ACE model for each of the three groups identified by the LCA. Two ACE models were estimated per group. One ACE model analyzed the posterior probability of group membership and another ACE model analyzed a variable that identified group assignment (i.e., a dichotomous variable where 1 = respondent was assigned to the group of interest). The estimates garnered from these ACE model analyses can be found in Table 3. Three findings are worthy of mention. First, genetic factors influenced each of the different offending patterns. Genetic influences were largest for individuals classified into Groups 1 (i.e., high-rate offenders) and 3 (i.e., low-rate offenders) and were smallest for individuals classified into Group 2 (i.e., moderate-rate offenders). Second, the shared environment had no discernible effect on being classified into any of the three response patterns. Third, the nonshared environment explained a significant portion of the variance for each of the different groups. Nonshared environmental factors explained a larger portion of the variance for individuals classified into Group 2 as compared to Groups 1 and 3.

Table 3.

ACE Model Parameter Estimates

	Heritability (A)	Shared Environment (C)	Nonshared Environment (E)
Group 1
Posterior probability	.33*	.00	.67*
Group assignment	.65*	.00	.35*
Group 2
Posterior probability	.26*	.00	.74*
Group assignment	.36*	.00	.64*
Group 3
Posterior probability	.46*	.00	.54*
Group assignment	.60*	.00	.40*

Note: *p<.05.

Discussion

The use of LCA has been instrumental in shedding light on the existence of subgroups nested within larger populations. LCA techniques have illuminated certain classes of individuals for a host of outcomes ranging from drug abuse and victimization (Reid & Sullivan, 2009; Vaughn et al., 2011) to aggressive, violent, and criminal behaviors (Connell et al., 2011). The goal of the current study was to expand on the methodology of LCA by blending it with behavioral genetic techniques. Our research, moreover, is intended to provide a template for examining the etiological underpinnings of latent class membership in the population. Analysis of the Add Health data revealed two important findings.

First, LCA models detected the presence of three distinct patterns of responses based on 17 items tapping a variety of delinquent behaviors. This first finding is of interest because it is consistent with Moffitt’s (1993) developmental taxonomy of life-course persistent, adolescent limited, and abstaining behavioral typologies. In accordance with Moffitt’s description of LCP offending, roughly 10% of the sample had a high probability of being involved in all types of delinquency, including serious acts such as stealing a car and getting into group fights. The second group that was identified engaged in a range of relatively minor and age normative misconduct (e.g., acting rowdy in public). Although members of Group 2 did report some forms of criminal activity (e.g., breaking and entering), their offending largely mirrored what Moffitt (1993) described as an AL offender. Finally, a third group emerged that reported almost no delinquent involvement. In line with Moffitt’s (1993) suggestion, this group generally abstained from criminal activity, engaging instead in only the most modest forms of misbehavior (e.g., lying to parents).

The second important finding in this study centers on the use of the ACE model to examine the genetic and environmental underpinnings of class membership. The results of these model-fitting analyses revealed that genetic factors were implicated in the etiologies of all three classes described above. Interestingly, the heritability of Groups 1 and 3 were the highest in the sample. These findings are of note because they fall in line with existing research that has detected genetic involvement in the prediction of both LCP offending and abstaining behaviors (Barnes, Beaver, & Boutwell, 2011; Boutwell & Beaver, 2008). This body of work suggests that the etiology of high-rate offending and the etiology of abstaining behaviors are more influenced by genetic factors than by environmental factors (Barnes et al., 2011).

Equally important, however, was the finding that the nonshared environment accounted for the remaining variance across all three groups. Consequently, the shared environment accounted for zero variation in class membership. These findings deserve additional attention because research in the field of criminology has only recently begun to differentiate the two environmental influences empirically and theoretically (see, for example, Beaver, 2008). Our results, while providing evidence of environmental influences on criminal behavior, suggest that the nonshared environment exerts a stronger influence on class membership than the shared environment (Ferguson, 2010). As a result, criminologists may need to reconsider the emphasis placed on shared environmental effects when attempting to explain variation in antisocial behavior (Rhee & Waldman, 2002; Turkheimer & Waldron, 2000).

The current study was not without limitations and it is important to pause momentarily to mention them in more detail. First, the current findings were based on identifying latent class membership in a sample of sibling pairs drawn from the Add Health Study. Because these results are constrained to the use of sibling data, it remains possible that our findings may not generalize to a broader population of singletons. Despite this likelihood, it is worth pointing out that prior researchers have compared the Add Health sibling subsample with the full sample of participants that was designed to be nationally representative (Jacobson & Rowe, 1998). The results of these comparisons found no systematic differences across samples (Beaver, 2008; Jacobson & Rowe, 1998). The second weakness concerns our inability to identify specific alleles (i.e., genes) that correspond to different offending patterns. Given our finding that genetic influences were significantly involved in each of the offending patterns, future research should be able to establish which measured genes correlate with membership into different offending classes.

Our research is intended to provide a template for examining the biosocial underpinnings of latent class membership in the population. The use of ACE models in conjunction with mainstream criminological methods such as LCA modeling techniques represents a valuable tool for researchers seeking a broader understanding of criminological phenomena. Ultimately, the incorporation of behavioral genetic techniques with more traditional criminological methodologies will continue to push the field toward a clearer understanding of the emergence and maintenance of criminogenic behaviors across the life course.

Footnotes

Acknowledgements

The authors wish to thank Kevin M. Beaver for his helpful comments on an earlier draft. This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due to Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health data files is available on the Add Health website (). No direct support was received from grant P01-HD31921 for this analysis.

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 received no financial support for the research, authorship, and/or publication of this article.

Notes

Bios

J.C. Barnes is an assistant professor in the Department of Criminology at The University of Texas at Dallas. His research seeks to understand how genetic and environmental factors combine to impact criminological phenomena. Recent works have attempted to reconcile behavioral genetic findings with theoretical developments in criminology. His work can be found in journals such as Aggressive Behavior, Criminology, Intelligence, Journal of Marriage and Family, Justice Quarterly, and Physiology & Behavior.

Brian B. Boutwell is currently an assistant professor in the College of Criminal Justice at Sam Houston State University. His research interests cut across multiple academic fields including behavior genetics, life-course criminology, and evolutionary psychology. His published works have appeared in a variety of academic journals including Criminology, Physiology and Behavior, Justice Quarterly, Journal of Research in Crime and Delinquency, Health Education and Behavior, Journal of Criminal Justice, International Journal of Environmental Health Research, Criminal Justice and Behavior, and Criminal Behaviour and Mental Health.

Robert G. Morris is Associate Professor of Criminology and Director of the Center for Crime and Justice Studies at the University of Texas at Dallas. His interests surround quantitative analysis, life course criminology, and contemporary issues in criminal justice. His research has appeared in numerous scholarly outlets including Justice Quarterly, Journal of Quantitative Criminology, Intelligence, and Crime and Delinquency. His is also the recipient of multiple research and teaching awards

Todd A. Armstrong is an associate professor at Sam Houston State University. He is interested in the biological causes and correlates of crime and antisocial behavior, criminology theory, and research on offense type patterns. His recent research has appeared in the Journal of Research in Crime and Delinquency, Justice Quarterly, and Criminal Justice and Behavior.

References

Agrawal

Lynskey

M. T.

Madden

P. A. F.

Bucholz

K. K.

Heath

A. C.

(2006). A latent class analysis of illicit drug abuse/dependence: Results from the National Epidemiological Survey on Alcohol and Related Conditions. Addiction, 102, 94-104.

Barnes

J. C.

Beaver

K. M.

(2010). An empirical examination of adolescence-limited offending: A direct test of Moffitt’s maturity gap thesis. Journal of Criminal Justice, 38, 1176-1185.

Barnes

J. C.

Beaver

K. M.

Boutwell

B. B.

(2011). Examining the genetic underpinnings to Moffitt’s developmental taxonomy: A behavioral genetic analysis. Criminology, 49, 923-954.

Beaver

K. M.

(2008). Nonshared environmental influences on adolescent delinquent involvement and adult criminal behavior. Criminology, 46, 341-370.

Blumstein

Cohen

Roth

Visher

C. A.

(1986). Criminal careers and “career criminals” (Vol. I). Report of the Panel on Research on Criminal Careers, National Research Council. Washington, D.C.: National Academy of Press.

Boutwell

B. B.

Beaver

K. M.

(2008). A biosocial explanation of delinquency abstention. Criminal Behaviour and Mental Health, 18, 59-74.

Boutwell

B. B.

Beaver

K. M.

(2010). The intergenerational transmission of low self-control. Journal of Research in Crime and Delinquency, 47, 174-209.

Brame

Paternoster

Piquero

A. R.

(2011). Thoughts on the analysis of group-based developmental trajectories in criminology. Justice Quarterly, forthcoming.

Connell

C. M.

Cook

E. C.

Aklin

W. M.

Vanderploeg

J. J.

Brex

R. A.

(2011). Risk and protective factors associated with patterns of antisocial behavior among nonmetropolitan adolescents. Aggressive Behavior, 37, 98-106.

10.

Dechartres

Huas

Godart

Pousset

Pham

Divac

S. M.

. . . Falissard

(2011). Outcomes of empirical eating disorder phenotypes in a clinical female sample: Results from a latent class analysis. Psychopathology, 44, 12-20.

11.

DeLisi

(2005). Career criminals in society. Thousand Oaks, CA: SAGE.

12.

DeLisi

Vaughn

Beaver

K. M.

Wexler

Barth

A. E.

Fletcher

J. M.

(2011). Fledgling psychopathy in the classroom: ADHD subtypes, psychopathy, and reading comprehension in a community sample of adolescents. Youth Violence and Juvenile Justice, 9, 43-58.

13.

Deslauriers-Varin

Beauregard

(2010). Victims’ routine activities and sex offenders’ target selection scripts: A latent class analysis. Sex Abuse, 3, 315-342.

14.

D’Unger

A. V.

Land

K. C.

McCall

P. L.

(2002). Sex differences in age patterns of delinquent/criminal careers: Results from Poisson latent class analyses of the Philadelphia Cohort Study. Journal of Quantitative Criminology, 18, 349-375.

15.

Farrington

D. P.

(1986). Age and crime. Crime and Justice, 7, 189-250.

16.

Farrington

D. P.

Ttofi

M. M.

Coid

J. W.

(2009). Development of adolescence-limited, late-onset, and persistent offenders from age 8 to 48. Aggressive Behavior, 35, 150-163.

17.

Ferguson

C. J.

(2010). Genetic contributions to antisocial personality and behavior: A meta-analytic review from an evolutionary perspective. Journal of Social Psychology, 150, 1-21.

18.

Formann

A. K.

Kohlmann

(1998). Structural latent class models. Sociological Methods & Research, 26, 530-565.

19.

Guo

Aveyard

Fielding

Sutton

(2009). Using latent class and latent transition analysis to examine the transtheoretical model staging algorithm and sequential stage transition in adolescent smoking. Substance Use & Misuse, 44, 2028-2042.

20.

Harris

K. M.

Halpern

C. T.

Whitsel

Hussey

Tabor

Entzel

Udry

J. R.

(2009). The National Longitudinal Study of Adolescent Health: Research Design [WWW document]. Retrieved from http://www.cpc.unc.edu/projects/addhealth/design.

21.

Hasking

P. A.

Scheier

L. M.

ben Abdallah

(2011). The three latent classes of adolescent delinquency and the risk factors for membership in each class. Aggressive Behavior, 37, 19-35.

22.

Jacobson

K. C.

Rowe

D. C.

(1998). Genetic and shared environment influences on adolescent BMI: Interaction with race and sex. Behavior Genetics, 28, 265–278.

23.

Kass

R. E.

Raftery

A. E.

(1995). Bayes factors. Journal of the American Statistical Association, 90, 773-795.

24.

Kelley

M. S.

Peterson

J. L.

(1997). The National Longitudinal Study of Adolescent Health (Add Health), Waves I & II, 1994-1996: A user’s guide to the machine-readable files and documentation (Data Sets 48-50, 98, A1-A3). Los Altos, CA: Sociometrics Corporation, American Family Data Archive.

25.

Klostermann

Mignone

Chen

(2009). Subtypes of alcohol and intimate partner violence: A latent class analysis. Violence and Victims, 24, 563-576.

26.

Kreuter

Muthén

(2008). Analyzing criminal trajectory profiles: Bridging multilevel and group-based approaches using growth mixture modeling. Journal of Quantitative Criminology, 24, 1-31.

27.

Krueger

R. F.

Moffitt

T. E.

Caspi

Bleske

Silva

P. A.

(1998). Assortative mating for antisocial behavior: Developmental and methodological implications. Behavior Genetics, 28, 173-186.

28.

Lanza

S. T.

Rhoades

B. L.

Nix

R. L.

Greenberg

M. T.

, & The Conduct Problems Prevention Research Group. (2010). Modeling the interplay of multilevel risk factors for future academic and behavior problems: A person-centered approach. Development and Psychopathology, 22, 313-335.

29.

Laska

M. N.

Pasch

K. E.

Lust

Story

Ehlinger

(2009). Latent class analysis of lifestyle characteristics and health risk behaviors among college youth. Prevention Science, 10, 376-386.

30.

Laub

J. H.

Sampson

R. J.

(2003). Shared beginnings, divergent lives: Delinquent boys to age 70. Cambridge, MA: Harvard University Press.

31.

Lloyd

Doll

Hawton

Dutton

W. H.

Geddes

J. R.

Goodwin

G. M.

Rogers

R. D.

(2010). Internet gamblers: A latent class analysis of their behaviours and health experiences. Journal of Gambling Studies, 26, 387-399.

32.

Moffitt

T. E.

(1993). Adolescence-limited and life-course persistent antisocial behavior: A developmental taxonomy. Psychological Review, 100, 674-701.

33.

Moffitt

T. E.

(2006). A review of research on the taxonomy of life-course persistent versus adolescence-limited antisocial behavior. In Cullen

Francis T.

Wright

John P.

Blevins

Kristie R.

(Eds.), Taking stock: The status of criminological theory. New Brunswick, NJ: Transaction Publishers.

34.

Muthén

Asparouhov

Rebollo

(2006). Advances in behavioral genetics modeling using Mplus: Applications of factor mixture modeling to twin data. Twin Research and Human Genetics, 9, 313-324.

35.

Nagin

D. S.

(2005). Group-based modeling of development. Cambridge, MA: Harvard University Press.

36.

Neale

M. C.

Maes

H. H. M.

(2004). Methodology for genetic studies of twins and families. Dordrecht, The Netherlands: Kluwer Academic Publishers B.V.

37.

Nylund

Bellmore

Nishina

Graham

(2007). Subtypes, severity, and structural stability of peer victimization: What does latent class analysis say? Child Development, 78, 1706-1722.

38.

Piquero

A. R.

(2008). Taking stock of developmental trajectories of criminal activity over the life course. In Liberman

Akiva M.

(Ed.), The long view of crime: A synthesis of longitudinal research. New York: Springer.

39.

Piquero

A. R.

Farrington

D. P.

Blumstein

(2003). The criminal career paradigm. Crime and Justice, 20, 359-506.

40.

Raine

(1993). The psychopathology of crime: Criminal behavior as a clinical disorder. San Diego, CA: Academic Press.

41.

Reid

J. A.

Sullivan

C. J.

(2009). A latent class typology of juvenile victims and exploration of risk factors and outcomes of victimization. Criminal Justice and Behavior, 36, 1001-1024.

42.

Rhee

S. H.

Waldman

I. D.

(2002). Genetic and environmental influences on antisocial behavior: A meta-analysis of twin and adoption studies. Psychological Bulletin, 128, 490-529.

43.

Rhule-Louie

D. M.

McMahon

R. J.

(2007). Problem behavior and romantic relationships: Assortative mating, behavior contagion, and desistance. Clinical Child and Family Psychology Review, 10, 53-100.

44.

Rowe

D. C.

Jacobson

K. C.

(1998). National longitudinal study of adolescent health. Pairs code book. Chapel Hill, NC: Carolina Population Center.

45.

Schwartz

Wetzler

Swanson

Sung

S. C.

(2010). Subtyping of substance use disorders in a high-risk welfare-to-work sample: A latent class analysis. Journal of Substance Abuse Treatment, 38, 366-374.

46.

Sullivan

C. J.

Childs

K. K.

O’Connell

(2010). Adolescent risk behavior subgroups: An empirical assessment. Journal of Youth and Adolescence, 39, 541-562.

47.

Turkheimer

Waldron

(2000). Nonshared environment: A theoretical, methodological, and quantitative review. Psychological Bulletin, 126, 78-108.

48.

Udry

J. R.

(1998). The National Longitudinal Study of Adolescent Health (Add Health), Waves I & II, 1994-1996 (Data Sets 48-50, 98, A1-A3, Kelley, M.S. & Peterson J.L.) [machine-readable data file and documentation]. Chapel Hill, NC: Carolina Population Center, University of North Carolina at Chapel Hill (Producer). Los Altos, CA: Sociometrics Corporation, American Family Data Archive (Producer & Distributor).

49.

Vaughn

M. G.

DeLisi

Gunter

Beaver

K. M.

Perron

B. E.

Howard

M. O.

2011. The severe 5%: A latent class analysis of the externalizing behavior spectrum in the United States. Journal of Criminal Justice, 39, 75-80.