The Effects of Police Contact on Trajectories of Violence

Abstract

This study uses a life course framework to investigate how police contacts may serve as a potential turning point in a violent crime trajectory. Drawing on the central ideas from deterrence and labeling theories, we determine whether individuals on different violent offending trajectories increase or decrease their offending following a police contact. Analyzing nine waves of data from the Rochester Youth Development Study, an integrated propensity score matching and latent class growth model was used. First, three violent trajectory groups emerged including high offenders, non-offenders, and low offenders. Second, after accounting for selection bias using propensity score matching procedures, experiencing a police contact increased the likelihood of future violent offending for the entire sample and for those who were on a low violent-offending trajectory specifically. These findings are interpreted as partial support for labeling theory. Limitations of the study and directions for future research are discussed.

Keywords

deterrence labeling life-course criminology interpersonal violence quantitative methods

The criminal justice system has historically operated under the deterrence doctrine, which holds that justice system contact will serve to reduce subsequent criminal behavior. Individuals who experience certain, swift, and adequately severe sanctions are thought to align their future behavior with societal expectations due to the threat of future punishments (Beccaria, 1764/1963; Gibbs, 1975; Zimring & Hawkins, 1973). Despite such efforts to control interpersonal violence, formal intervention can have unintended consequences. Labeling theory suggests a dismal outcome for those who experience justice system contact; official intervention is hypothesized to set in motion a chain of events that actually lead to the worsening of criminal behavior (Becker, 1963; Lemert, 1951, 1967). Specifically, those who are labeled by formal sanctions are likely to be blocked from conventional opportunities, form negative self-identities, and become more involved with criminal others that can embed them into a career of crime (Bernburg & Krohn, 2003; Bernburg, Krohn, & Rivera, 2006; Kirk & Sampson, 2013; Lopes et al., 2012; Paternoster & Iovanni, 1989; see also Link, Cullen, Struening, Shrout, & Dohrenwend, 1989).

It is not only possible but, indeed, probable that official intervention will have different effects on different types of offenders. Paternoster and Iovanni (1989) suggest that “we should not expect labeling effects to be invariant across subgroups” (p. 381). Pogarsky (2002), advocating a deterrence perspective, makes a similar argument. Perhaps, Sherman (1993) puts it best:

Does punishment control crime? This question provokes fierce debates in criminology and public policy. Yet there is ample evidence that it is the wrong question. Widely varying results across a range of sanction studies suggest a far more useful question: under what conditions does each type of criminal sanction reduce, increase, or have no effect on future crimes? Answering that question is central to the future of research in crime and delinquency. (p. 445)

Then, it is possible that both deterrence and labeling theories may be relevant depending on specific characteristics of offenders. Yet, after numerous acknowledgements in the literature of the need to explore contingent labeling and deterrent effects for several decades (Chiricos, Barrick, Bales, & Bontranger, 2007; D. Huizinga & Henry, 2008; Sherman, 1993; Thorsell & Klemke, 1972), there have been few empirical studies that have addressed this question generally (D. Huizinga & Henry, 2008, p. 249) and none have taken a life-course approach with a specific focus on interpersonal violence. This is an unfortunate empirical omission; developmental and life-course theories anticipate that life transitions such as justice system contact may have varying influences on behavioral trajectories (see Elder, 1985; Moffitt, 1993), a possibility that will be discussed later in detail.

The current study examines the different claims made within deterrence and labeling theories about the effect of official intervention but does so by exploring whether effects vary across individuals with divergent criminal histories. We begin with a brief summary of sanction effects research in aggregated samples. We next explore the effects of justice system contact on subsequent behavior in a developmental and life-course context and review the limited empirical research that has examined how sanction effects might vary across groups with differing criminal histories. With this background, we then examine the impact of police contact on subsequent violence for individuals following distinct developmental trajectories using longitudinal data and an integrated propensity score matching and latent class growth model.

Deterrence and Labeling Research in Aggregated Samples

In a comprehensive review of the deterrence literature, Nagin (1998) found that cross-sectional and scenario-based empirical research provides evidence that perceptions do indeed influence criminal offending. Yet, the most methodologically rigorous studies tend to show considerably weaker effects compared with studies that are conducted with cross-sectional data or those that use few or no control variables (Paternoster, 1987). A recent meta-analysis finds that perceptions of certainty of arrest have a statistically significant but relatively weak effect on criminal behavior (Pratt, Cullen, Blevins, Daigle, & Madensen, 2006). Similarly, weak but statistically significant evidence for the specific deterrent effect of arrest on subsequent delinquency is obtained from meta-analyses of the Minneapolis Domestic Violence Experiment and Spouse Abuse Replication Program studies (Maxwell, Garner, & Fagan, 2002).

In comparison, numerous studies have found support, or partial support, for labeling theory (Bernburg & Krohn, 2003; Farrington, 1977; Gold & Williams, 1970; D. H. Huizinga & Esbensen, 1992; D. Huizinga, Weiher, Espiritu, & Esbensen, 2003; Kaplan & Damphouse, 1997; Krohn, Lopes, & Ward, in press; Paternoster, 1978; Paternoster & Piquero, 1995; D. J. Smith, 2006). For example, Bernburg and Krohn (2003) found that having a police contact or being arrested as a juvenile significantly increased serious crime and general delinquency in early adulthood after making adjustments for adolescent delinquency, family poverty, educational capabilities, and race. Still, other research fails to yield support for either perspective (Hemphill, Toumbourou, Herrenkohl, McMorris, & Catalano, 2006; D. Huizinga, Elliott, & Dunford, 1986; Klein, 1986; McAra & McVie, 2007; Thomas, 1977). For instance, McAra and McVie (2007) used propensity score matching (PSM) with the goal of approximating a treatment effect of arrest using observational data. They found that once individuals in the arrest and non-arrest groups were adequately matched, there were no significant official intervention effects on either the prevalence or frequency of serious delinquency.

In short, there is mixed support for labeling or deterrence theory in the aggregate. This is true for different types of sanctions such as arrest and conviction. The inconsistencies in empirical findings regarding the effects of criminal sanctions on subsequent behavior have brought renewed interest in uncovering for “whom” deterrence or labeling may operate (Chiricos et al., 2007; Morris & Piquero, 2013; Piquero, Paternoster, Pogarsky, & Loughran, 2011). Studies that have attempted to answer this question have estimated official intervention effects in disaggregated samples and looked for differences across subgroups defined by criminal history, sex, race, and “stakes in conformity.” Analyses focusing on sex, race, and stakes in conformity have been informative (Ageton & Elliott, 1974; Baumer, 1997; Berk, Campbell, Klap, & Western, 1992; Bernburg & Krohn, 2003; Chiricos et al., 2007; Gendreau, Little, & Goggin, 1996; Harris, 1975) and warrant further empirical attention, but we believe that the consideration of the (criminal) behavioral patterns of offenders may offer more for understanding the varying effects of official intervention and is consistent with the broader theoretical movement to consider offenders within a developmental and life-course context. In addition, the policy implications of this line of research may be more palatable; that is, sanctioning individuals based on what they have done in the past—rather than on their gender or race—is more consistent with the stated procedures of the formal criminal justice system.

Deterrence and Labeling in a Life-Course and Developmental Context

A life-course approach suggests that differences in the developmental history of offending, and not simply the level of offending at a discrete point in time, may be important in determining the effects of official intervention on subsequent violence. From these perspectives, an experience with law enforcement is seen as a potential turning point in the lives of offenders (Sampson & Laub, 2005), which can change one’s criminal trajectory. For instances, state sanctions are viewed as an important institution of social control which is linked to cumulative disadvantage (Sampson & Laub, 1997). “[Labeling] theory specifically suggests a ‘snowball’ effect—that adolescent delinquency and its negative consequences (e.g., arrest, official labeling, incarceration) can ‘mortgage’ one’s future, especially later life chances molded by schooling and employment” (Sampson & Laub, 1997, p. 15). Experiencing an official label can knife off conventional opportunities leaving individuals with fewer prosocial alternatives to offending (Sampson & Laub, 1997). From a deterrence perspective, official intervention can trigger the beginnings of the desistance process (Bushway, Thornberry, & Krohn, 2003) that deflects a trajectory of crime toward conforming behavior.

Moffitt’s taxonomic theory of criminal behavior suggests that there are fundamentally different types of offenders. On the one hand, life-course-persistent offenders commit acts of crime and delinquency relatively consistently over time. The onset of life-course-persistent offending is believed to be caused by the interacting forces of neuropsychological deficits and exposure to criminogenic environments early in life. Continuity in behavior is maintained through proactive and reactive person–environment interactions. On the other hand, adolescence-limited offenders engage in acts of crime and delinquency during the period of adolescence as an expression of independence brought on by a maturity gap between their biological and social ages.

Moffitt (1993) predicts that adolescence-limited offenders can become ensnared by justice system contact, which might interfere with the normal desistance process that characterizes this group (i.e., they are subject to labeling effects). Life-course-persistent offenders, however, might be unlikely to experience labeling effects as their behavioral trajectory is quite problematic already. At the same time, however, life-course-persistent offenders actively participate in the continuity of their behavior, in part, by interpreting and reacting to environmental stimuli in distinct, unhealthy ways (Moffitt, 1993). With respect to justice system contact, life-course-persistent offenders may feel that police are targeting them as opposed to serving justice, which may contribute to maintenance of criminal behavior.

A number of interesting possibilities emerge when considering the effects of justice system contact across different criminal trajectories. “Incorrigibles” (see Pogarsky, 2002) or those on a chronic offending trajectory such as “life-course-persistent offenders” (see Moffitt, 1993) might be unresponsive to the deterrent effects of justice system contact, and, the labeling process may also be irrelevant for these individuals or have already run its course. Individuals who have lower levels of offending over time may have the naturally occurring desistance process hastened by justice system contact (i.e., the formal costs of crime may give the individual an added reason to stop offending) or reversed (i.e., the labeling process may be invoked, “snaring” individuals as Moffitt [1993, 2006] might anticipate). In short, official intervention may serve as a turning point (deterrent or labeling) for individuals following along a particular trajectory but may have no influence on offending behavior for individuals on another trajectory.

Very few studies have explicitly placed deterrence and labeling within a developmental context, but there are a number of studies that have examined the influence of justice system contact on subsequent behavior across criminal history. In one of the earliest studies examining deterrent and labeling effects across offending subgroups, Cameron (1964) found that arresting novice shoplifters typically resulted in a reduction of their delinquent behavior. D. A. Smith and Gartin (1989) found that the experience of being arrested was more likely to terminate the criminal career of novice offenders, whereas it only reduced offending rates among more experienced offenders.¹ While some evidence points to deterrent effects among novice offenders, DeJong (1997) found that first-time offenders who were incarcerated for misdemeanors were more likely to be rearrested than individuals who were not incarcerated for their offenses. Research yielding the same general conclusions has been reported in other studies (Horwitz & Wasserman, 1979; Taxman & Piquero, 1998). In contrast to studies that show stronger sanction effects for novice offenders, some research demonstrates that sanction effects emerge more strongly for experienced offenders. For example, Klein (1974) collected recidivism data on 100 juveniles from 13 police agencies that he categorized as either high- or low-diversion agencies. Arrest and system processing decreased the rate of subsequent offending; however, this was more likely for more experienced offenders compared with first-time offenders.

Matsueda, Kreager, and Huizinga (2006) linked subsequent offending behavior to experiential risk of apprehension, which they defined as the number of arrests divided by the number of self-reported offenses. They found that as experiential risk of apprehension increased, the perceived risk of apprehension also increased leading to a reduction in subsequent offending behavior. These results are consistent with specific deterrence theory. What is important is that experienced offenders seemed to have lower experiential risk of apprehension values than novice offenders, meaning that experienced offenders were less deterred by arrest.

A recent felony convictions study that aimed to redirect attention to the contingent nature of labeling effects found that those who had adjudication withheld were less likely to recidivate compared with those who had been adjudicated (Chiricos et al., 2007). Yet, this study did not find evidence that the labeling effect was conditioned by “prior record.” In one of the few studies to place sanction effects in a developmental framework, Bhati (2007; see also Bhati & Piquero, 2008) classified individuals based on differences between their actual post-release trajectories and counterfactual trajectories (i.e., the trajectories expected of the same individuals if they were not incarcerated, which is based on their criminal histories) to explore whether labeling, deterrent, or null (incapacitation) effects occurred. Analyses revealed that approximately 56% were unaffected by incarceration (i.e., only incapacitation effect), 40% experienced a deterrent effect, and only 4% experienced a labeling effect. Furthermore, results indicated that those with a greater number of arrests were less likely to experience deterrent effects.

Using data from the National Youth Survey and using an integrated propensity score matching and semiparametric group-based modeling strategy, Morris and Piquero (2013) examined the extent to which arrest influenced subsequent behavior across individuals in different offending trajectories. They identified three trajectory groups and then matched on propensity scores to balance several covariates including age, race, gender, socioeconomic status, grade-point average, index of social disorganization, negative peer influence, prior delinquency, and urbanicity. Results indicated a statistically significant labeling effect for the moderate trajectory group but no effect for the low or chronic trajectory groups—though these effects were in a direction consistent with labeling. In supplemental analyses, the authors estimated inverse probability-to-treatment weighted regression models and found contradictory results. Specifically, they found statistically significant labeling effects for all three trajectory groups. The authors interpreted these findings to indicate null effects for the low-offending group, relatively strong labeling effects for the chronic offending group, and the presence of attenuated labeling effects for the moderate group.

In sum, the empirical evidence is mixed and inconclusive. As Smith and Gartin (1989) have put it, “Perhaps the only conclusion we can draw at this point is that the relationship between punishment and future offending among both novice and experienced offenders is problematic” (p. 97). Unfortunately, some 20 years later, this observation still largely appears to be true (see Huizinga & Henry, 2008). Despite the lack of a clear answer, advances in theory and available methodologies offer promise for our ability to better understand the relationship between justice system contact and subsequent delinquency and violence. We now review the recent advances in statistical methodology that guide the current study and permit a detailed empirical investigation using observational data into the effects of justice system contact on subsequent interpersonal violence across individuals following distinct offending trajectories.

Statistical Methods

To examine whether police contact serves as a turning point in the criminal trajectories of offenders, a classification procedure for different types of offenders is necessary. Previous research has used subjective classification methods to identify groups of individuals (e.g., see Huizinga et al., 1986). There are, however, several notable problems with subjective classification procedures. Subjective interpretations of random variation may result in the identification of an artificial group. At the same time, unusual subgroups that may be quite important can be overlooked by subjective classification procedures. In addition, statistical tests examining differences across subgroups may be undermined by uncertainties about the reliability of group assignment (see Nagin, 2005). Finally, subjective classification schemes may fail to consider the development of offending behavior over time, which can limit their ability to assess whether official intervention serves as a turning point in certain criminal trajectories. Latent class growth analysis (LCGA) helps to overcome these limitations by more objectively classifying individuals into different developmental trajectories and allows for the modeling of intra-individual change in criminal behavior over time.

Methodological limitations of previous studies have also resulted in the inability to discern between labeling and deterrent effects with any definitiveness. A key problem has been the use of imprecise designs that result in an inability to disentangle “selection artifacts” from official intervention effects (D. A. Smith & Paternoster, 1990). With some exceptions (e.g., Chiricos et al., 2007; McAra & McVie, 2007; Morris & Piquero, 2013; D. A. Smith & Paternoster, 1990), research designs have failed to address the inherent problems of making causal inferences with non-experimental data. Propensity score matching analysis can result in covariate balance on observed variables, although it cannot balance unobserved variables; thus, the procedure is most helpful in cases where all the relevant covariates are observed. Importantly, the relative success of the matching procedure can be explicitly examined, and average treatment effects can be estimated that better approximate the causal effect as compared with other statistical approaches that make use of non-experimental data.

Haviland and her colleagues (Haviland & Nagin, 2005; Haviland, Nagin, & Rosenbaum, 2007; Haviland, Nagin, Rosenbaum, & Tremblay, 2008) recently noted the unique benefits of integrating LCGA and PSM to address important questions that emerge from the life-course criminological literature. The method is designed to address whether there are trajectory group-specific effects such as those of justice system contact on subsequent violent crime.

The LCGA-PSM integrated method is essentially the sequential use of LCGA followed by PSM within trajectory groups. Haviland and colleagues (2007) write,

The integration of group-based trajectory modeling and propensity scores is composed of a three-stage analysis. The first stage involves estimating a group-based trajectory model for the outcome and participants of interest . . . In the second stage, each treated individual is matched with one or more untreated individuals . . . We then check the degree of success of the matching strategy in achieving balance . . . In the third stage of the analysis, the treatment effect of the event of interest . . . is analyzed . . . within and across trajectory groups. (p. 249)

In short, two highly useful analytic approaches in their own right are brought to bear on longitudinal data in a systematic and straightforward manner, which ultimately provides an ideal way to determine whether an official intervention experience serves as a turning point for one or more distinct violent-offending trajectory groups.

The analysis proceeds as follows. First, group-based trajectory models (see Nagin, 2005) are estimated using a self-reported violent delinquency index. Specifically, a Poisson model is estimated and the number of violent offending trajectory groups is determined using the Bayesian Information Criteria (BIC) score and other relevant selection criteria. Next, posterior probabilities of group assignment and other common post-estimation statistics are examined to determine the relative precision of the best fitting model. The number of trajectory groups, the proportion of individuals within each trajectory group, and the shapes of the offending trajectories are explored.

With the best fitting LCGA model, individuals within each trajectory group are split into two groups: (1) those having experienced police contact and (2) those having not experienced police contact. Covariate balance is assessed prior to matching using the absolute standardized difference in covariate means. Next, logistic regression models are used to estimate propensity scores for each individual. Optimal full matching procedures are executed (see Hansen, 2004) and then covariate balance is reassessed for each subgroup to test the successfulness of the matching procedures. Specifically, the absolute standardized difference in covariate means is again examined along with the percentage bias reduction statistic. In addition to carrying out PSM within each trajectory group, it is also done for the sample as a whole.

The integration of these two analytic techniques culminates with the estimation of a treatment effect. With two statistically similar groups, statistical significance of the effect of police contact on subsequent interpersonal violence is estimated for the total sample and for each subpopulation using the Hodges and Lehmann (1962) aligned rank test. In addition, the absolute standardized difference in outcome means is reported as a measure of effect size.

Data and Measures

Sample

This study analyzes data from the Rochester Youth Development Study (RYDS), which features a total of 14 waves of data, collected in three phases. We use data from the entirety of Phase 1, which consists of nine waves of data collected at 6-month intervals beginning in 1988 and ending in 1992 (for a complete description, see Thornberry, Krohn, Lizotte, Smith, & Tobin, 2003). Individuals included in the analysis sample had to be interviewed during every wave between Waves 6 and 9 and had to have valid police intervention records during Waves 7 and 8. In addition, individuals must have been interviewed at least twice between Waves 2 through 5 to be included in the analysis sample. The analysis sample was selected with the rationale that subjects had to be interviewed such that confounding variables used to predict the probability of receiving official intervention were available and data on self-reported police contacts and information on violent offending were observed. Given the ability of LCGA to accommodate missing data in its estimation of trajectories, some missing data were permitted from Waves 2 through 5. Nevertheless, missing interviews were not a serious concern during this time period; 93.4% of the final analysis sample was interviewed during each of Waves 2 through 5 and no individual in the analysis sample had his violent offending trajectory estimated from less than three out of five of the time points. The final analysis sample that fits the above criteria consists of approximately 82% of the original RYDS males (N = 595).² Multiple imputation (MI) was used to address missing data among covariates in the analysis sample.³ It should be noted that the average means and standard deviations of the five MI data sets were substantively similar to those reported for the valid cases (results available on request). Descriptive statistics can be found in the appendix.

Measures

Behavioral trajectory and outcome

Violent crime

The behavioral trajectory and outcome variable of interest is self-reported violent crime. Violent crime is a variety index measuring the total number of six different acts of violence an individual committed in a given wave. The six acts include (1) attacking someone with a weapon, (2) participating in a gang fight, (3) throwing things at people, (4) committing robbery, (5) raping someone, and (6) engaging in other assaults.⁴ Possible scores on this measure range from zero to six. Data from Waves 2 through 6 are used to estimate the LCGA models. With respect to the outcome variable, Wave 9 is used to examine the effects of police contact on subsequent interpersonal violence.

Treatment

Police contact

Police contact is a self-report measure of whether an individual experienced one or more police contacts (i.e., being picked up and formally questioned by police for suspected involvement in a crime) during Waves 7 or 8. Two waves (one year) are used to ensure that a sufficient number of individuals actually experienced justice system contact so that the effects of the treatment can be estimated across different violent crime trajectories.

Covariates

The treatment model consists of 40 covariates from several waves of data. Some time-stable measures were captured at Wave 1 (e.g., race), whereas other measures that vary across time are available at various waves. In most cases, these measures are taken from Wave 6, which is the wave immediately prior to the designated period for treatment. Wave 6 is selected as these measures are temporally close, yet precede in time, the justice system contact. Relevant covariates are organized into the following categories: demographics, neighborhood characteristics, family dynamics, school factors, peer associations, values and mental states, prior delinquent behavior, and prior justice system contact. Measurement details of the covariates can be found in the appendix along with the descriptive statistics.

Results

Trajectory Analysis

Table 1 contains information on model-selection criteria for five different quadratic latent class growth models with latent classes ranging from one to five. The three class model has the lowest BIC, and the metric for model comparison (see Nagin, 2005; see also Kass & Wasserman, 1995; Schwarz, 1978) suggests that there is nearly a 72% chance that the three class model is the correct model. Table 2 contains the mean posterior probabilities for group assignment to the three violent crime trajectories (column) by the assigned group based on maximum probability assignment (row). The conventional cutoff for average posterior probabilities for group assignment is 0.7, with values above this indicating adequate model precision (Nagin, 2005). As shown in the diagonal of Table 2, the mean posterior probabilities for group assignment are 0.81 or greater indicating good model precision. Alternative statistics such as the odds of correct classification similarly indicate that individuals have been assigned to violent crime trajectory groups with adequate precision (results available on request).

Table 1.

Model Selection Criteria for Five Different Poisson LCGA Models Estimating Violent Crime Trajectories.

Latent Class	Log Likelihood	BIC	Prob. Model Correct
1	−2730.044	5,479.254	.000
2	−2387.141	4,819.002	.000
3^a	−2342.381	4,755.037	.720
4	−2330.550	4,756.928	.280
5	−2328.991	4,779.364	.000

Note. LCGA = latent class growth analysis; BIC = Bayesian Information Criteria ; Prob.= Probability.

Indicates selected model.

Table 2.

Mean Posterior Probabilities for Trajectory Group Assignment for the Selected Three-Group Poisson LCGA Model.

	Group
Assigned Group	1	2	3
1	0.891	0.000	0.109
2	0.000	0.899	0.101
3	0.058	0.126	0.816

Note. LCGA = latent class growth analysis.

Figure 1 illustrates the estimated violent-offending trajectory groups from Waves 2 through 6. Just under half of the sample is classified as low offenders and nearly 40% are classified as non-offenders. The high-offending trajectory group is smaller, comprising approximately 11% of the sample. Supplemental model details including parameter estimates, standard errors, t-statistics, and p values for the intercept and growth factors for the three latent classes are available on request.

Figure 1.

Estimated trajectories of violent offending during adolescence for the three latent classes.

Table 3 reports the violent crime mean outcomes and the bivariate results summarizing the effect of police contact on subsequent violence. For the full sample, the mean difference in violent offending between those experiencing a police contact and those not is .302 (p ≤ .01). As for specific trajectory groups, police contacts are significantly related to violent crime in the short-run for all three of the latent classes. On average, high offenders who experience a police contact commit .412 times more types of violent crimes than their counterparts (p ≤ .05). For the low-offending trajectory group, the average difference is .285 (p ≤ .01). The mean difference for non-offenders is the smallest of the three groups, but it is still statistically significant and equal to .084 (p ≤ .05).

Table 3.

Bivariate Associations Between Police Contact and Violent Crime for the Total Sample and by Trajectory Group.

	Police Contact	Violence	Sum of Scores	Expected Under Ho	Mean Score	Z
Total sample (N = 595)
No police contact	68.40%	0.113	113,527.0	121,286.0	278.94	6.43**
Police contact	31.60%	0.415	63,783.0	56,024.0	339.27	—
Trajectory groups
High offenders (n = 68)
No police contact	50.00%	0.265	1,002.5	1,173.0	29.48	2.51*
Police contact	50.00%	0.677	1,343.5	1,173.0	39.51	—
Non-offenders (n = 234)
No police contact	80.77%	0.027	21,939.5	22,207.5	116.08	1.97*
Police contact	19.23%	0.111	5,555.5	5,287.5	123.45	—
Low offenders (n = 293)
No police contact	62.80%	0.174	25,256.5	27,048.0	137.26	3.71**
Police contact	37.20%	0.459	17,814.5	16,023.0	163.44	—

Note. Wilcoxon-Mann-Whitney test of statistical significance; Ho = null hypothesis.

p ≤ .05. **p ≤ .01

Covariate Balance Assessments

Total sample

Table 4 reports standardized bias statistics for treatment and control groups before and after matching for each of the 40 individual covariates. In addition, global measures of covariate balance are reported. A given covariate is substantially out of balance if its standardized bias statistic is greater than 20 (Haviland et al., 2007; Haviland et al., 2008); while values below 20 are deemed acceptable, the lower the standardized bias statistic the better. Well over half of the variables exhibit substantial imbalance prior to matching. In addition to 24 imbalanced covariates, there are eight additional covariates that have standardized bias statistics greater than 10. Overall, the typical covariate is out of balance by approximately 26% of a standard deviation and the logit of the propensity score is imbalanced by more than 90% of a standard deviation. In short, the two groups are not equivalent with respect to a large number of important variables that could confound the observed relationship between police contact and violent crime.

Table 4.

Treatment and Control Group Covariate Bias Statistics Before and After PSM for the Police Contact Treatment.

	Total Sample			High Offenders			Non-Offenders			Low Offenders
	d_X	d_Xm	%BR	d_X	d_Xm	%BR	d_X	d_Xm	%BR	d_X	d_Xm	%BR
Demographics
Race
Black	0.74	4.84	−5.54	48.95	0.00	1.00	18.93	7.10	0.63	11.67	3.21	0.73
Hispanic	4.41	1.82	0.59	24.14	4.62	0.81	13.22	9.93	0.25	13.04	4.11	0.68
Age	23.65	4.89	0.79	12.97	24.36	−0.88	42.04	6.43	0.85	13.20	4.18	0.68
Family poverty	2.37	5.52	−1.33	26.40	51.50	−0.95	5.76	7.11	−0.23	5.69	5.69	0.00
Neighborhood characteristics
Proportion African American	4.43	5.23	−0.18	11.57	6.44	0.44	18.08	7.88	0.56	15.37	5.55	0.64
Proportion in poverty	17.09	4.42	0.74	27.61	33.13	−0.20	4.44	5.75	−0.30	18.51	6.03	0.67
Neighborhood disorganization	1.27	3.70	−1.92	13.71	40.93	−1.99	3.41	7.80	−1.28	5.18	1.75	0.66
Neighborhood integration	7.58	2.84	0.63	12.55	8.78	0.30	11.14	7.22	0.35	7.38	5.82	0.21
Neighborhood satisfaction	11.62	3.21	0.72	8.74	11.33	−0.30	3.07	3.44	−0.12	17.38	2.13	0.88
Neighborhood arrest rate	15.42	5.21	0.66	52.66	9.00	0.83	13.66	5.08	0.63	0.40	4.59	−10.35
Family
Family structure	5.38	3.05	0.43	28.01	11.42	0.59	4.99	5.90	−0.18	2.35	6.51	−1.77
Parental supervision	23.63	2.88	0.88	70.74	42.04	0.41	8.96	10.74	−0.20	15.78	7.91	0.50
Attachment to parent	16.12	4.07	0.75	11.89	23.59	−0.98	13.02	3.86	0.70	22.90	5.97	0.74
Primary caregiver’s educational expectations of child	41.70	4.18	0.90	33.08	48.19	−0.46	39.49	5.93	0.85	30.02	2.23	0.93
Attachment to child	40.83	2.83	0.93	21.65	16.54	0.24	31.16	3.27	0.89	35.29	3.88	0.89
Consistency in discipline	15.15	6.35	0.58	7.39	17.62	−1.38	20.83	6.71	0.68	4.17	3.30	0.21
School factors
Commitment to school	40.35	4.10	0.90	45.89	36.52	0.20	48.77	10.05	0.79	20.29	2.90	0.86
Aspiration for college	33.57	3.96	0.88	36.47	46.53	−0.28	35.89	5.96	0.83	18.65	3.38	0.82
Prosocial activities	15.13	4.68	0.69	37.78	5.55	0.85	6.90	5.59	0.19	6.69	2.97	0.56
Educational expectations	28.85	3.19	0.89	7.13	19.92	−1.79	37.06	9.35	0.75	22.73	4.27	0.81
School clubs with friends	12.74	2.18	0.83	27.13	2.48	0.91	15.22	5.70	0.63	18.41	3.37	0.82
Peer associations
Peer delinquency	45.95	2.86	0.94	40.48	44.73	−0.10	31.31	9.18	0.71	24.90	5.42	0.78
Peer drug use	44.47	2.38	0.95	27.27	5.01	0.82	20.53	7.00	0.66	29.91	4.79	0.84
Peer delinquent values	32.69	2.86	0.91	17.06	26.65	−0.56	31.50	4.40	0.86	11.19	5.23	0.53
Dating	34.91	1.88	0.95	20.48	12.74	0.38	12.69	6.18	0.51	30.51	5.30	0.83
Sexual activity under 15	3.04	7.25	−1.38	17.91	0.12	0.99	17.67	5.62	0.68	12.48	2.21	0.82
Risky time with friends	40.27	4.44	0.89	30.40	9.06	0.70	22.01	9.97	0.55	23.52	4.26	0.82
Gang involvement	41.62	2.61	0.94	18.33	48.97	−1.67	3.75	5.06	−0.35	30.67	4.49	0.85
Values
Self-image	27.94	4.45	0.84	25.01	13.49	0.46	22.41	6.68	0.70	29.29	8.85	0.70
Depression	19.17	5.02	0.74	3.82	20.59	−4.39	15.88	7.23	0.54	12.44	9.69	0.22
Delinquent values	47.50	1.22	0.97	82.16	27.07	0.67	16.45	5.77	0.65	30.91	8.48	0.73
Prior criminal behavior
Violent crime (Time 1)	26.27	3.75	0.86	1.38	39.39	−27.47	13.41	14.19	−0.06	4.25	6.97	−0.64
Violent crime (Time 2)	24.36	4.68	0.81	13.85	14.11	−0.02	11.69	22.94	−0.96	7.80	4.20	0.46
Violent crime (Time 3)	31.67	3.13	0.90	33.72	13.21	0.61	6.72	2.27	0.66	19.37	6.06	0.69
Violent crime (Time 4)	39.93	1.96	0.95	19.84	26.56	−0.34	6.17	21.39	−2.46	20.92	2.80	0.87
Violent crime (Time 5)	37.73	1.70	0.95	30.23	50.05	−0.66	0.00	0.00	0.00	16.42	3.74	0.77
General delinquency	61.95	2.53	0.96	39.08	32.37	0.17	30.82	4.52	0.85	63.50	3.96	0.94
Drug use	35.11	3.78	0.89	36.22	23.59	0.35	17.97	5.43	0.70	27.78	4.59	0.83
Aggression	45.46	2.16	0.95	35.81	1.28	0.96	30.01	5.77	0.81	37.29	5.46	0.85
Prior justice system contact
Previous contact or arrest	35.71	2.70	0.92	48.51	9.22	0.81	39.94	7.12	0.82	11.82	4.56	0.61
Global covariate balance statistics
M	25.94	3.61	0.46	27.70	21.97	−0.78	18.67	7.29	0.34	18.75	4.77	0.32
Median	27.10	3.46	0.86	26.76	18.77	0.24	16.17	6.31	0.63	17.90	4.53	0.73
Maximum	61.95	7.25	0.97	82.16	51.50	1.00	48.77	22.94	0.89	63.50	9.69	0.94
Logit propensity score	92.37	0.23	1.00	184.93	28.59	0.85	132.55	5.21	0.96	101.15	0.71	0.99

Note. PSM = propensity score matching; BR = bias reduction.

After matching, every covariate exhibits excellent balance. Specifically, no covariate mean difference exceeds 7.25% of a standard deviation. The bias reduction statistics indicate that all except four of the variables underwent a bias reduction as denoted by the positive percentage bias reduction values. The average percentage bias reduction was 46% and the median percentage bias reduction was 86%. The average covariate is now imbalanced by less than 4% of a standard deviation and, moreover, the group means of the logit of the propensity scores are essentially identical. For the sample as a whole, the full optimal matching procedure has created two groups that are comparable on all observed covariates.

Trajectory groups

Table 4 also contains the pre- and post-matching standardized bias statistics along with percentage bias reduction values for the high offenders, non-offenders, and low offenders, respectively. The high offenders have 25 covariates that have standardized bias statistics greater than 20 and an additional 10 covariates that have standardized bias statistics greater than 10. The average covariate is out of balance by approximately 28% of a standard deviation. The means of the logit of the propensity score are different by nearly two standard deviations. Following the matching procedure, the treatment and control groups are not sufficiently balanced. The median bias reduction for a covariate is only 24%. The differences in the logit of the propensity scores across the treated and control groups were reduced by 85%, but this global summary balance measure is still out of balance by nearly 29% of a standard deviation. After matching, the average covariate still differs by 22% of a standard deviation. In short, covariate balance was not achieved for the high-offending group.

Non-offenders have 16 covariates that are substantially out of balance and another 14 with standardized bias statistics greater than 10. Overall, the covariate means for a variable differ by approximately 19% of a standard deviation. And, the means for the logit of the propensity score diverge by 1.33 standard deviations. After matching, all but 2 covariates have acceptable levels of covariate balance.⁵ The average covariate post-matching is imbalanced by only about 7% of a standard deviation, and the logit of the propensity score is out of balance by only approximately 5% of a standard deviation. These values are down from 19% and 132%, respectively.

Finally, the far right side of Table 5 contains the pre- and post-matching covariate balance statistics for the low-offending trajectory group. The covariate means for a total of 15 of the measures are substantially different across the two groups, and another 15 variables have standardized bias statistics that are greater than 10. Overall, the average covariate is imbalanced by about 19% of a standard deviation, and the logit of the propensity score is out of balance by slightly more than one standard deviation. After matching, all 40 covariates have standardized bias statistics below 10. The average covariate is now imbalanced by less than 5% of a standard deviation. The average bias reduction was 32% and the median bias reduction for a covariate was 73%. The logit of the propensity score underwent a 99% bias reduction leaving the covariate means to differ by less than 1% of a standard deviation.

Table 5.

Post-Matching Treatment Effects of Police Contact for the Total Sample and by Trajectory Group.

	Est.	SE	t	d
Total sample	5696.27	2078.11	2.74**	0.30
High offenders	—	—	—	—
Non-offenders	378.50	312.55	1.21	0.18
Low offenders	1722.42	743.64	2.32*	0.35

Note. Est.= Estimate.

p ≤ .05. **p ≤ .01.

In sum, the optimal full matching procedure has created a treatment and control group for the total sample, low-offending trajectory group, and non-offending trajectory group in which the treatment effects of police contact on subsequent violent offending can be estimated with much confidence. Perhaps unsurprisingly, covariate balance was unachievable for the high-offending trajectory group, as this result has occurred in other analyses using similar methodology (Haviland et al., 2007; Haviland et al., 2008). Treatment effect estimates for the high-offending group are invalid and, therefore, not reported.

Treatment Effect Estimates of Police Contact on Violence

Table 5 contains the Hodges-Lehmann treatment effect estimates as well as Cohen’s d values of effect size. Beginning with the full sample, the aligned rank test shows that individuals self-reporting a police contact during the treatment period are likely to engage in more acts of violent crime in Wave 9. This treatment effect is statistically significant (p < .01), yet the effect is small (d = .30). While the bivariate relationship between police contact and violent crime during Wave 9 was statistically significant for the non-offending trajectory group, the treatment effect estimate failed to achieve statistical significance. However, the treatment effect for the low-offending trajectory group is statistically significant (p ≤ .05) and is small to moderate (d = .35). This is not the case for non-offenders (p > .05), and, unfortunately, the data do not allow us to discern whether this is true for high offenders.

Discussion and Conclusions

Using an integrated propensity score matching and latent class growth model, the current study examined the extent to which police contact influenced subsequent violent crime across different offending subpopulations of males from the RYDS. Findings revealed evidence for small and statistically significant treatment effects of police contact on violent crime for the total sample. When examining the sample as a whole, the results are consistent with the central ideas of labeling theory’s deviance amplification hypothesis. While aggregate analysis is informative, prior research and existing theory both point to the importance of investigating how individuals with distinct offending histories might respond differently to the experience of police contact—that is, aggregate analyses may obscure important differences in the treatment effects across divergent violent offending subgroups. Indeed, this looks to be the case. While treatment effect estimates were unavailable for the high-offending trajectory group, results indicated no significant treatment effect of police contact on subsequent interpersonal violence for the non-offending group (p > .05) but a significant treatment effect for the low-offending group (p < .05). In short, when individuals are successfully matched on 40 covariates, there is empirical evidence for a short-run labeling effect of the police contact treatment for the low-offending trajectory group only.

Consistent with Moffitt’s (1993, 2006) predictions, those following along a low-offending trajectory appear to be worse off following a police contact relative to similar others that were not the subject of police investigation. The results of this study charge the police with a most difficult task. When offenses become known to the police, it is their duty to respond to them and uphold the law. In doing so, however, police intervention may unintentionally make the offending problem of the individual worse in the short run. The findings partially support the non-intervention logic that society should “leave kids alone whenever possible” (Schur, 1973, p. 155). Yet, some of the behaviors that comprise the violent crime measure in this study are very serious offenses including such things as attacking someone with a weapon and rape. These types of offenses certainly warrant a response from the criminal justice system. As such, understanding the specific processes that unfold following the experience of justice system contact which lead to greater interpersonal violence will be of the utmost importance to reducing labeling effects on violent behavior.

A similar adverse treatment effect of police contact was not found to exist for non-offenders. Importantly, intervening in the lives of these youth does not appear to foster interpersonal violence. Supplemental analysis of variance (ANOVA) analyses examining differences between trajectory groups in covariate means suggest that there are statistically significant differences between non-offenders and low offenders in several different domains (e.g., family, school, peer associations) (results available on request). Compared with low offenders, non-offenders have stronger parental attachment, greater commitment to school, fewer delinquent peers, lower depression, and so on. Collectively, these systematic differences in risk and protective factors may give non-offenders more social capital to avoid the adverse consequences justice system contact might otherwise trigger.

Although bivariate results indicated initial evidence for a labeling effect among the high-offending group, unbiased treatment effects were unattainable and we cannot therefore draw any trustworthy conclusions for this group. The fact that the high-offending trajectory group resulted in unsuccessful matching is not surprising as prior research using the same methodology has also been unsuccessful in matching (e.g., gang joiners and non-gang joiners) within a high-rate offending trajectory group (see Haviland et al., 2007; Haviland et al., 2008). Future research should seek to use the current methodology using alternative, larger data sets to increase the chances of successfully matching in the high-offending trajectory group, while keeping in mind the importance of using data that has detailed information on covariates spanning key developmental domains.

While the current study has attempted to address an important theoretical question using longitudinal data and sophisticated statistical analyses, it is certainly not without limitations. The current study has addressed a first-order issue: Does the effect of official intervention vary across violent offending subgroups? While there is some support for the deviance amplification hypothesis, lacking from this research is an assessment of the specific mechanisms by which justice system contact may lead to increases in subsequent violent behavior. Labeling theory is explicit in the identification of mediating pathways that include blocked access to conventional opportunity, alteration of self-identity, and acquisition of deviant peers (Paternoster & Iovanni, 1989). Without focusing on these pathways in particular, research can only identify whether there is support for the basic idea of deviance amplification and cannot speak to whether the data truly support a labeling process as specified by theory. Future research should seek to investigate these mediating processes across different behavioral trajectories, which may provide important insights into exactly why one particular trajectory experiences deviance amplification and another does not.

For many, justice system contact may not simply be a one-time event. Using the current data, restricting the analysis to first timers was not possible due to the need to maintain large trajectory groups to maximize the likelihood of successful matching. While our approach properly accounted and adjusted for any police contacts occurring prior to the treatment period, future research should look at exploring the influence of initial contacts specifically as well as the cumulative effect of experiencing multiple contacts. Relatedly, an alternative and probably more realistic conceptualization of justice system contact is to consider it as a repeatable treatment and incorporate available modeling strategies to account for the fluid nature of justice system contact.

Notwithstanding these limitations, the current study has taken a step toward further understanding the relationship between justice system contact and subsequent behavior that has been the subject of much debate in the deterrence and labeling literatures for decades. The life-course perspective with its focus on trajectories and transitions offers an important organizing framework that should continue to be used in future research seeking to uncover how prior behavior may influence the way in which an individual responds to justice system contact. Relatedly, studies should also consider exploring the effects of official intervention on prevalence, variety, and frequency of offending, and ideally all within the same study. Using these different behavioral measures would allow one to assess to what extent justice system contact influences the termination of a criminal career, the number of distinct acts that an individual engages in, and the total number of offenses an individual commits, respectively.

Footnotes

Appendix

Descriptive Statistics and Measurement Details (N = 595)

Variable	Description	M	SD	Wave
Violent crime outcome	See measures section.
Short-run		0.21	0.59	9
Official intervention	See measures section.
Self-reported police contact		0.32	—	7-8
Demographics
Race	Coded as a series of dummy variables with Whites as the reference group.
Black		0.63	—	1
Hispanic		0.18	—	1
Age	Age is a continuous measure of the respondent’s age during Wave 6.	16.47	0.81	6
Family poverty	Family poverty is a dichotomous variable that serves to distinguish between respondents who come from families living above federally defined poverty level coded as 0 from those families below it coded as 1. To create this measure, the number of individuals in G1’s house was determined and the before-tax income was calculated. These two values were used and compared against the U.S. Government’s federal poverty chart.	0.30	—	4
Neighborhood characteristics
Proportion African American	The measure of the proportion of African Americans residing in the youth’s neighborhood comes from U.S. Census data for Monroe County during the 1990 data-collection period. The proportion of Blacks in the neighborhood is calculated by taking the number of Blacks (not of Hispanic origin) and dividing them by the total number of individuals residing in the corresponding census tract.	0.53	0.27	1990C
Proportion in poverty	The measure of the proportion of families in poverty residing in the youth’s neighborhood also comes from 1990 U.S. Census data. This measure was calculated by taking the number of families in poverty and dividing this value by the number of families residing in the corresponding census tract.	0.33	0.14	1990C
Neighborhood disorganization	Tapping the level of conflict and violence, drug use, and general disorder in the adolescent’s neighborhood as perceived by the primary caregiver (PC; G1), this variable is a measured with a 17-item scale. The adolescent’s PC was asked to report the extent to which a number of things were a “big problem,” “sort of a problem,” or “not a problem.” Specifically, they were asked about how much of a problem the following things were in their neighborhood: different racial or cultural groups who do not get along with each other; vandalism, building, and personal belongings broken and torn up; little respect for rules, law, and authority; high unemployment; winos and junkies; prostitution; abandoned houses or buildings; sexual assaults or rapes; gambling; burglaries and thefts; rundown and poorly kept buildings and yards; assaults and muggings; street gangs or delinquent gangs; syndicate, mafia, or organized crime; buying or selling stolen goods; drug use or drug dealing in the open; and, homeless street people. Higher scores indicate greater neighborhood disorganization.	1.64	0.64	6
Neighborhood integration	Neighborhood integration is a 7-item measure that assesses the extent to which neighbors know one another and whether the people who live in the neighborhood can be relied on. Using a 4-point Likert type scale, the adolescent’s PC (G1) was asked, “how many people live in your neighborhood do you know by sight; how many people live in your neighborhood do you talk to on a regular basis; how often do you and other people who live in your neighborhood borrow things like tools or recipes from each other; how often do you and other people who live in your neighborhood ask each other to drive or take your children somewhere?” Response options included “never,” “seldom,” “sometimes,” and “often.” Higher scores on this measure are indicative of higher neighborhood integration.	2.19	0.68	5
Neighborhood satisfaction	This variable is a 3-item measure tapping into the PC’s satisfaction with the neighborhood. On a 4-point Likert type scale, respondents were asked to state how satisfied they were with your relationship with people in this neighborhood; the way people in this neighborhood take care of where they live; and the neighborhood as a good place to bring up children. Response options ranged from “very dissatisfied” to “very satisfied,” and higher scores on this measure are indicative of higher neighborhood satisfaction.	2.88	0.66	5
Neighborhood arrest rate	Neighborhood arrest rate is a measure of the resident arrest rate for the respondent’s census tract. Using official records, the total number of individuals arrested within the youth’s census tract was divided by the population and standardized by 1,000. Higher values indicate a greater neighborhood arrest rate.	4.01	1.98	A
Family
Family structure	Family structure is a dichotomous variable indicating whether a respondent lives at home with both parents. This measure is mostly reported from the respondent’s PC (G1), but some adolescents (G2) did provide data on the family structure when G1 interviews were unavailable. Responses were coded such that 0 = “does not live with both parents” and 1 = “lives with both parents.”	0.33	—	6
Parental supervision	Parental supervision is a 4-item scale measuring the quality of supervision that the youth self-reports. On a 4-point Likert type scale, respondents were asked to rate the extent to which the PC knows where the child is and knows who they are with when they are away from home, as well as how important these two supervisory tasks are to the PC. Higher values designate greater parental supervision of the respondent by the PC—whether this is one of the biological parents or some other male or female.	3.52	0.46	6
Attachment to parent	Parental attachment is an 11-item scale tapping into how attached the respondent is to the PC (e.g., mom, dad, other male, or other female). Respondents were asked a number of questions including, “do you get along well with your PC; do you feel that you can really trust your PC; does your PC not understand you (reverse coded); is your PC is too demanding (reverse coded); do you really enjoy your PC; does your PC interfere with your activities (reverse-coded); do you think your PC is terrific; do you feel very angry toward your PC (reverse-coded); do you have a lot of respect for your PC; do you feel very proud of your PC; and, do you feel violent toward your PC (reverse coded).” Higher scores indicate a greater level of parental attachment to the PC.	3.39	0.43	6
PC’s educational expectation of child	This variable is a 2-item measure tapping the PC’s (G1) expectancies regarding whether the youth will succeed in educational pursuits. The PC was asked whether they believed that G2 would graduate from high school and whether he would go on to college. Response options to these two questions were “no,” “depends,” and “yes.” Higher scores indicate greater expectations for the child to further his or her education.	2.33	0.83	6
Attachment to child	Attachment to child is an 11-item scale measuring how attached the PC (G1) (e.g., mom, dad, other male, or other female) is to the adolescent (G2). This is a parallel scale to the parental-attachment scale above meaning that the items are similar, but this measure instead asks about how attached the PC is to the child. Higher scores indicate a greater level of attachment to the adolescent.	3.48	0.46	6
Consistency in discipline	Consistency in discipline measures the youth’s perception of their PC’s (G1) patterns of discipline. Respondents were asked the following questions: “How often do you get away with things; once PC decides a punishment, how often can you get out of it; how often do you get punished sometimes, but not other times, for doing the same thing; how often does PC have to ask you to do the same thing more than once; when you get punished, how much does the kind of punishment you get depend on PC’s mood?” Response options were “often,” “sometimes,” “seldom,” and “never.” Higher scores are coded such that they indicate lower consistency in discipline.	2.36	0.52	6
School factors
Commitment to school	Commitment to school is a 10-item scale measuring one’s level of agreement on a 4-point Likert scale to statements pertaining to school commitment. For instances, respondents were asked whether school is boring to you (reverse coded); you like school a lot; you don’t really belong at school (reverse coded); you usually finish your homework; you try hard at school; and, getting good grades is very important to you. Response options ranged from “strongly agree” to “strongly disagree.” Higher scores on this measure designate greater commitment to school.	3.08	0.38	6
Aspiration for college	Aspirations for college is a 2-item scale indicating how important it is for one to, first, graduate from high school and, second, to go to college. On a 4-point Likert type scale, response options ranged from “not important at all” to “very important.” Higher scores indicate greater aspirations for college.	3.34	0.86	6
Prosocial activities	This variable is a 5-item scale measuring one’s involvement in prosocial activities many of which take place in the school environment. On a 4-point Likert type scale with response options of “never,” “seldom,” “sometimes,” and “often,” respondents were asked how often they took part in school sports such as intramurals and varsity sports; took part in school activities such as clubs or special events such as a school play; took part in organized sports or teams outside school; took part in any organized musical or singing group, including in school; and, took part in other organized groups such as the “Y,” Boys and Girls Clubs, or scouts. Higher scores on this measure indicate greater involvement in prosocial activities.	1.68	0.66	6
Educational expectations	Educational expectations is a parallel measure to the parent’s education expectations of child scale. While the items are the same, this measure taps the youth’s (G2) own expectancies regarding whether he or she will succeed in his or her educational pursuits. As with the other measure, higher scores indicate greater expectations for the individual to further his or her education.	2.60	0.73	6
School clubs with friends	The last of the school factor variables taps the extent to which individuals engage in school activities such as clubs or special events such as a school play with their friends. Response options on this measure include “never,” “seldom,” “sometimes,” and “often.” Higher scores on this measure indicate greater involvement in school clubs with friends.	1.23	0.57	6
Peer associations
Peer delinquency	Peer delinquency is an 8-item scale measuring the proportion of one’s peers who engage in delinquent acts. Respondents were asked to indicate whether “none,” “a few,” “some,” or “most” of their friends have done delinquent things, which, for example, include stole something worth more than $100; attacked someone with a weapon with the idea of seriously hurting them; took a car or motorcycle for a ride or drove without the owner’s permission; and damaged or destroyed someone else’s property on purpose. Higher scores indicate greater peer delinquency.	1.32	0.51	6
Peer drug use	Peer drug use is a 4-item scale measuring the proportion of one’s peers who engage in drug use. Respondents were asked to indicate whether “none,” “a few,” “some,” or “most” of their friends have used marijuana, crack, alcohol, or other drugs (e.g., LSD, heroin, acid). Higher scores indicate greater peer drug use.	1.41	0.53	6
Peer delinquent values	This variable is a 6-item scale that measures whether one’s peer group would say that it was “wrong,” “not say anything,” or “say it was okay,” if the respondent engaged in certain delinquent acts including using a weapon or force to get money or things from people; hitting somebody with the idea of hurting them; stealing something worth $50; damaging or destroying someone else’s property on purpose, taking a car or motorcycle for a ride or drive without the owner’s permission; and skipping classes without an excuse. Higher scores on this measure reflect higher peer delinquent values.	1.31	0.42	6
Dating	Dating is a single-item dichotomous variable measuring whether an individual has a special “boyfriend” or “girlfriend.” Responses were coded such that 0 = no boyfriend/girlfriend or 1 = boyfriend/girlfriend.	0.46	—	6
Sexual activity under 15	This variable is a dichotomous indicator measuring whether the respondent had engaged in sexual intercourse prior to the age of fifteen. Responses were coded such that 0 = no sexual activity under fifteen and 1 = sexual activity under fifteen.	0.34	—	—
Risky time with friends	Risky time with friends measures the average amount of time spent with the respondent’s three closest friends. For each friend, respondents were asked how many times a week they get together with their friends without adults, drive around in a car with no special place to go, and get together where someone is using or selling drugs or alcohol. On a 5-point scale, response options ranged from “never” to “every day.” Higher scores indicate more risky time with friends.	2.16	0.71	6
Gang involvement	Gang involvement is a dichotomous measure of whether an individual was involved in a gang at any time between Wave 2 and Wave 6. Beginning in Wave 2, respondents were asked during the time period since the last interview whether they were a member of a street gang or “posse” (the term used by Rochester adolescents). Responses to the questions of gang participation during the five waves on interest were recoded such that 1 = gang involvement (one or more waves) and 0 = no gang involvement.	0.27	—	2-6
Values and mental states
Self-image	Self-esteem is a 9-item scale measuring the extent to which an individual agrees or disagrees, on a 4-point Likert scale, with statements about oneself. Examples of statements tapping self-esteem include, “at times you think you are no good at all (reverse coded); you feel that you have a number of good qualities; you feel you do not have much to be proud of (reverse coded); or, you feel you are at least as good as other people.” Higher scores indicate greater self-esteem.	3.21	0.42	6
Depression	This variable is a 14-item scale tapping how frequently individuals felt depression symptoms such as feeling hopeful about the future; feeling depressed or very sad; not feeling like eating because of being upset about something; or, thinking seriously about suicide. On a 4-point scale, response options ranged from “never” to “often”; thus, higher scores indicate greater depression symptoms.	1.98	0.47	6
Delinquent values	Delinquent values is an 11-item scale measuring how wrong one thinks it is to engage in a variety of delinquent behaviors. For example, respondents were asked about acts of delinquency such as using hard drugs, drinking alcohol, stealing something worth $50, taking a car or motorcycle for a ride without the owner’s permission, damaging or destroying someone’s property on purpose, or hitting someone with the idea of hurting them. Response options included “not wrong at all,” “a little bit wrong,” “wrong,” and “very wrong.” Higher scores indicate greater approval of delinquent behaviors.	1.40	0.43	6
Prior criminal behavior
Violent crime (Time 1)	Violence is a self-reported variety index tapping into the total number of six distinct acts of violence an individual committed in a given wave. The violent acts include attacking someone with a weapon; engaging in a gang fight; throwing things at people; committing robbery; raping someone; and, engaging in other assaults. Scores on this measure range from zero to six and higher scores indicate greater variety of violent offending. This is the same measure that is used for the behavioral trajectories and outcomes. To be clear, a violence variable for each wave is included in the treatment model. In other words, five different violent crime variables (Waves 2 through 6) are used as covariates to ensure balance in behavior at each time point.	0.56	0.91	2
Violent crime (Time 2)	—	0.48	0.87	3
Violent crime (Time 3)	—	0.47	0.85	4
Violent crime (Time 4)	—	0.37	0.72	5
Violent crime (Time 5)	—	0.36	0.73	6
General delinquency	General delinquency is a self-reported variety index tapping into the total number of 32 distinct acts of general delinquency committed in Wave 6. The general delinquency acts cover a wide range of criminal and delinquent behaviors and include things such as: truancy; carrying a weapon; property damage; public rowdiness; theft of several different amounts (e.g., less than $5, $5-50, $50-100, $100+); auto-theft; assault; robbery; obscene phone calls; fraud; marijuana sales; hard drug sales; and so forth. Scores on these measures range from zero to 32, and higher values reflect a greater involvement in various acts of general delinquency.	1.38	2.32	6
Drug use	This variable is a self-reported variety index tapping into the total number of ten different drugs that were used in Wave 6. The drugs include marijuana; inhalants; hallucinogens; cocaine; crack; heroin; pcp; tranquilizers; downers; and, uppers. Scores on this measure range from 0 to 10, with higher scores indicative of higher drug-use variety during Wave 6.	0.15	0.49	6
Aggression	Aggression is measured with eleven items from the Child Behavior Checklist, which was administered to the parents during Wave 6 (Achenbach, 1991). This measure assesses general aggressive tendencies in youth but also includes some items that reflect hyperactivity and impulsiveness. PCs of subjects reported whether G2 had done any of the following been unable to sit still or been restless or hyperactive; been cruel to animals; been cruel, bullying, or mean to others; demanded a lot of attention; felt that others were out to get him or her; been impulsive or acted without thinking; screamed a lot; been stubborn, sullen or irritable; had sudden changes in mood or feelings; had temper tantrums or a hot temper; and, threatened people. Response options included “never,” “sometimes,” and “often.” Higher scores on this measure indicate greater aggression.	0.38	0.35	6
Official intervention	The variable is measured using official records data from the Rochester Police Department and County Wide Registration and is a dichotomous indicator of whether an individual experienced one or more police contacts or arrests during Waves 1 through 6 (0 = no prior justice system contact or 1 = previous justice system contact).
Prev. justice system contact		0.34	—	1-6

Note. 1990C = 1990 U.S. Census for Monroe County; PC = primary caregiver.

Authors’ Note

Official arrest data were provided electronically by the New York State Division of Criminal Justice Services. Points of view, conclusions, and methodological strategies in this document are those of the authors and do not necessarily represent the official position or policies of the funding agencies or data sources.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Research for this project was supported by the Harry Frank Guggenheim Foundation Dissertation Fellowships program. Support for the Rochester Youth Development Study has been provided by the Office of Juvenile Justice and Delinquency Prevention (86-JN-CX-0007), the National Institute on Drug Abuse (R01DA005512), the National Science Foundation (SBR-9123299), and the National Institute of Mental Health (R01MH63386). Work on this project was also aided by grants to the Center for Social and Demographic Analysis at the University at Albany from NICHD (P30HD32041) and NSF (SBR-9512290).

Notes

Author Biographies

Jeffrey T. Ward is an assistant professor in the Department of Criminal Justice at the University of Texas at San Antonio. His research interests include developmental and life-course criminology, sanction effects, gangs, and quantitative methodology. His work has appeared in Crime & Delinquency, Criminal Justice and Behavior, Criminology, and Journal of Criminal Justice.

Marvin D. Krohn is currently a professor in the Department of Sociology and Criminology & Law at the University of Florida. He has a long-standing interest in the etiology of delinquency and drug use, focusing primarily on social-psychological and life-course approaches. For the past 26 years he has worked on the Rochester Youth Development Study, a three-generational longitudinal panel study targeting those at high risk for serious crime and delinquency. His book (with co-authors Terence P. Thornberry, Alan J. Lizotte, Carolyn A. Smith, and Kimberly Tobin), Gangs and Delinquency in Developmental Perspective, was the American Society of Criminology’s recipient of the 2003 Michael J. Hindelang Award for Outstanding Scholarship. He also co-authored Delinquent Behavior (with Don C. Gibbons) and Researching Theories of Crime and Delinquency (with Charis E. Kubrin and Thomas D. Stucky) and has co-edited four compendiums on crime and delinquency. In addition, he has contributed to numerous research articles and book chapters. He was recently named a Fellow of the American Society of Criminology.

Chris L. Gibson is a Research Foundation Professor at the University of Florida and an associate professor in the Department of Sociology and Criminology. He also holds a courtesy faculty appointment in the Levin College of Law’s Criminal Justice Center. He was the 2009 National Institute of Justice W. E. B. Du Bois Fellow and the 2013 Tory Caeti young scholar of the Academy of Criminal Justice Sciences. His current research uses an interdisciplinary approach to examine how traits and other individual characteristics influence antisocial behavior and victimization at a given time and through the life course, while integrating neighborhood and other contextual predictors as moderating or direct influences. His book (w/Marv Krohn) titled Handbook of Life-Course Criminology: Emerging Trends and Directions for Future Research was published in 2013 by Springer-Verlag.

References

Ageton

S. S.

Elliott

(1974). The effects of legal processing on delinquent orientations. Social Problems, 22, 87-100.

Alison

(2001). Missing data. Thousand Oaks, CA: Sage.

Baumer

(1997). Levels and predictors of recidivism: The Malta experience. Criminology, 35, 601-628.

Beccaria

(1963). On crimes and punishment. New York, NY: MacMillan. (Original work published 1764)

Becker

H. S.

(1963). Outsiders: Studies in the sociology of deviance. New York, NY: Free Press.

Belson

W. A.

(1968). The extent of stealing by London boys and some of its origins. Advancement of Science, 25, 171-184.

Berk

R. A.

Campbell

Klap

Western

(1992). The deterrent effect of arrest in incidents of domestic violence: A Bayesian analysis of four field experiments. American Sociological Review, 57, 698-708.

Bernburg

J. G.

Krohn

M. D.

(2003). Labeling, life chances, and adult crime: The direct and indirect effects of official intervention in adolescence on crime in early adulthood. Criminology, 41, 1287-1318.

Bernburg

J. G.

Krohn

M. D.

Rivera

C. J.

(2006). Official labeling, criminal embeddedness, and subsequent delinquency: A longitudinal test of labeling theory. Journal of Research in Crime & Delinquency, 43, 67-88.

10.

Bhati

A. S.

(2007). Studying the effects of incarceration on offending trajectories: An information-theoretic approach (Technical Report). Washington, DC: Justice Policy Center, The Urban Institute.

11.

Bhati

A. S.

Piquero

A. R.

(2008). Estimating the impact of incarceration on subsequent offending trajectories: Deterrent, criminogenic, or null effect? Journal of Criminal Law and Criminology, 98, 207-253.

12.

Bushway

S. D.

Thornberry

T. P.

Krohn

M. D.

(2003). Desistance as a developmental process: A comparison of static and dynamic approaches. Journal of Quantitative Criminology, 19, 129-153.

13.

Cameron

M. O.

(1964). The booster and the snitch. New York, NY: Free Press.

14.

Chiricos

Barrick

Bales

Bontranger

(2007). The labeling of convicted felons and its consequences for recidivism. Criminology, 45, 547-581.

15.

DeJong

(1997). Specific deterrence and survival analysis: Integrating theoretical and empirical models of recidivism. Criminology, 35, 561-576.

16.

Elder

G. H.

(1985). Life-course dynamics: Trajectories and transitions, 1968-1980. Ithaca, NY: Cornell University Press.

17.

Farrington

D. P.

(1977). The effects of public labeling. British Journal of Criminology, 17, 112-125.

18.

Gendreau

Little

Goggin

(1996). A meta-analysis of the predictors of adult offender recidivism: What works! Criminology, 34, 575-607.

19.

Gibbs

J. P.

(1975). Crime, punishment, and deterrence. New York, NY: Elsevier.

20.

Gold

Williams

(1970). National study of the aftermath of apprehension. Prospectus, 3, 3-12.

21.

Hansen

B. B.

(2004). Full matching in an observational study of coaching for the SAT. Journal of the American Statistical Association, 99, 609-618.

22.

Harris

(1975). Imprisonment and the expected value of criminal choice: A specification and test of aspects of the labeling perspective. American Sociological Review, 40, 71-87.

23.

Haviland

Nagin

D. S.

(2005). Causal inferences with group-based trajectory models. Psychometrika, 70, 557-578.

24.

Haviland

Nagin

D. S.

Rosenbaum

P. R.

(2007). Combining propensity score-matching and group-based trajectory analysis in an observational study. Psychological Methods, 12, 247-267.

25.

Haviland

Nagin

D. S.

Rosenbaum

P. R.

Tremblay

R. E.

(2008). Combining group-based trajectory modeling and propensity score matching for causal inference in nonexperimental longitudinal data. Developmental Psychology, 44, 422-436.

26.

Hemphill

S. A.

Toumbourou

J. W.

Herrenkohl

T. I.

McMorris

B. J.

Catalano

R. F.

(2006). The effect of school suspensions and arrests on subsequent adolescent behavior in Australia and the United States. Journal of Adolescent Health, 39, 736-744.

27.

Hindelang

M. J.

Hirschi

Weis

J. G.

(1981). Measuring delinquency. Beverly Hills, CA: Sage.

28.

Hodges

J. L.

Lehmann

E. L.

(1962). Rank methods for combination of independent experiments in the analysis of variance. Annals of Mathematical Statistics, 33, 482-497.

29.

Horwitz

Wasserman

(1979). The effect of social control on delinquent behavior: A longitudinal test. Sociological Focus, 12, 53-70.

30.

Huizinga

Elliott

D. S.

Dunford

F. W.

(1986, March). The effect of arrest on subsequent delinquent behavior. Paper presented at the Annual Meetings of the Academy of Criminal Justice Sciences, Orlando, FL.

31.

Huizinga

Henry

K. L.

(2008). The effect of arrest and justice system sanctions on subsequent behavior: Findings from longitudinal and other studies. In Liberman

A. M.

(Ed.), The long view of crime: A synthesis of longitudinal research (pp. 220-254). New York, NY: Springer.

32.

Huizinga

Weiher

A. W.

Espiritu

Esbensen

(2003). Delinquency and crime: Some highlights from the Denver Youth Survey. In Thornberry

T. P.

Krohn

M. D.

(Eds.), Taking stock of delinquency: An overview of findings from contemporary longitudinal studies (pp. 47-92). New York, NY: Kluwer Academic.

33.

Huizinga

D. H.

Esbensen

(1992). An arresting view of juvenile justice. School Safety, 15-17.

34.

Kaplan

H. B.

Damphouse

K. R.

(1997). Negative social sanctions, self-derogation, and deviant behavior: Main and interactive effects in longitudinal perspective. Deviant Behavior, 18, 1-26.

35.

Kass

R. E.

Wasserman

(1995). A reference Bayesian test for nested hypotheses and its relationship to the Schwarz Criterion. Journal of the American Statistical Association, 90, 928-934.

36.

Kirk

D. S.

Sampson

R. J.

(2013). Juvenile arrest and collateral educational damage in the transition to adulthood. Sociology of Education, 86, 36-62.

37.

Klein

M. W.

(1974). Labeling, deterrence, and recidivism: A study of police dispositions of juvenile offenders. Social Problems, 22, 292-303.

38.

Klein

M. W.

(1986). Labeling theory and delinquency policy: An experimental test. Criminal Justice and Behavior, 13, 47-79.

39.

Krohn

M. D.

Lopes

Ward

J. T.

(in press). Effects of police intervention on later offending. In Farrington

D. P.

Murray

(Eds.), Labeling theory: Empirical tests, advances in criminological theory. New Brunswick, NJ: Transaction Publishers.

40.

Lemert

E. M.

(1951). Social pathology: A systematic approach to the theory of sociopathic behavior. New York, NY: McGraw-Hill.

41.

Lemert

E. M.

(1967). Human deviance, social problems, and social control. Englewoods Cliffs, NJ: Prentice Hall.

42.

Link

B. G.

Cullen

F. T.

Struening

Shrout

P. E.

Dohrenwend

B. P.

(1989). A modified labeling theory approach to mental disorders: An empirical assessment. American Sociological Review, 54, 400-423.

43.

Lopes

Krohn

M. D.

Lizotte

A. J.

Schmidt

N. M.

Vasquez

B. E.

Bernburg

J. G.

(2012). Labeling and cumulative disadvantage: The impact of formal police intervention on life chances and crime during emerging adulthood. Crime & Delinquency, 58, 456-488.

44.

Matsueda

R. L.

Kreager

Huizinga

(2006). Deterring delinquents: A rational choice model of theft and violence. American Sociological Review, 71, 95-122.

45.

Maxwell

C. D.

Garner

J. H.

Fagan

J. A.

(2002). The preventative effects of arrest on intimate partner violence: Research, policy and theory. Criminology & Public Policy, 2, 51-80.

46.

McAra

McVie

(2007). Youth justice? The impact of system contact on patterns of desistance from offending. European Journal of Criminology, 4, 315-345.

47.

Moffitt

T. E.

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

48.

Moffitt

T. E.

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

F. T.

Wright

J. P.

Blevins

K. R.

(Eds.), Taking stock: The status of criminology theory (pp. 277-311). New Brunswick, NJ: Transaction Publishers.

49.

Morris

Piquero

(2013). For whom do sanctions deter and label? Justice Quarterly, 30, 837-868.

50.

Nagin

D. S.

(1998). Criminal deterrence research at the outset of the twenty-first century. Crime and Justice: A Review of Research, 23, 1-42.

51.

Nagin

D. S.

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

52.

Paternoster

(1978). The labeling effect of police apprehension: Identity, exclusion, and secondary deviance (Doctoral dissertation). Florida State University, Tallahassee. (UMI No. 7917065)

53.

Paternoster

(1987). The deterrent effect of the perceived certainty and severity of punishment: A review of the evidence and issues. Justice Quarterly, 4, 173-217.

54.

Paternoster

Iovanni

(1989). The labeling perspective and delinquency: An elaboration of the theory and an assessment of the evidence. Justice Quarterly, 6, 359-394.

55.

Paternoster

Piquero

(1995). Reconceptualizing deterrence: An empirical test of personal and vicarious experiences. Journal of Research in Crime & Delinquency, 32, 251-286.

56.

Piquero

A. R.

Paternoster

Pogarsky

Loughran

(2011). Elaborating the individual difference component of deterrence theory. Annual Review of Law and Social Science, 7, 335-360.

57.

Pogarsky

(2002). Identifying “deterrable” offenders: Implications for research on deterrence. Justice Quarterly, 19, 431-452.

58.

Pratt

Cullen

F. T.

Blevins

Daigle

Madensen

(2006). The empirical status of deterrence theory: A meta-analysis. In Cullen

F. T.

Wright

J. P.

Blevins

K. R.

(Eds.), Taking stock: The status of criminological theory (pp. 367-395). New Brunswick, NJ: Transaction Publishers.

59.

Rubin

D. B.

(1987). Multiple imputation for nonresponse in surveys. New York, NY: John Wiley.

60.

Sampson

R. J.

Laub

J. H.

(1997). A life-course theory of cumulative disadvantage and stability in delinquency. In Thornberry

T. P.

(Ed.), Developmental theories of crime and delinquency (pp. 133-161). New Brunswick, NJ: Transaction Publishers.

61.

Sampson

R. J.

Laub

J. H.

(2005). A life-course view of the development of crime. Annals of the American Academy of Political and Social Sciences, 602, 12-45.

62.

Schur

E. M.

(1973). Radical nonintervention: Rethinking the delinquency problem. Englewood Cliffs, NJ: Prentice Hall.

63.

Schwarz

(1978). Estimating the dimension of a model. Annals of Statistics, 6, 461-464.

64.

Shannon

L. W.

(1980). Juvenile delinquency and adult crime, 1948-1977 (Racine, Wisconsin): Three birth cohorts. Iowa City: Iowa Urban Community Research Center, University of Iowa.

65.

Shannon

L. W.

(1985). A more precise evaluation of the effects of sanctions (Final report to the National Institute of Justice). Iowa City: Iowa Urban Community Research Center, University of Iowa.

66.

Sherman

L. W.

(1993). Defiance, deterrence, and irrelevance: A theory of the criminal sanction. Journal of Research in Crime & Delinquency, 30, 445-473.

67.

Smith

D. A.

Gartin

P. R.

(1989). Specifying specific deterrence: The influence of arrest on future criminal activity. American Sociological Review, 54, 94-105.

68.

Smith

D. A.

Paternoster

(1990). Formal processing and future delinquency: Deviance amplification as selection artifact. Law & Society Review, 24, 1109-1132.

69.

Smith

D. J.

(2006). Social inclusion and early desistance from crime (Report No. 12 of the Edinburgh Study of Youth Transitions and Crime). Edinburgh, Scotland: Center for Law and Society, Old College, University of Edinburgh.

70.

Taxman

F. S.

Piquero

(1998). On preventing drunk driving recidivism: An examination of rehabilitation and punishment approaches. Journal of Criminal Justice, 26, 129-143.

71.

Thomas

C. W.

(1977). The effect of legal sanctions on juvenile delinquency: A comparison of the labeling and deterrence perspectives (Unpublished report). Department of Sociology, Bowling Green State University, Bowling Green, OH.

72.

Thornberry

T. P.

Krohn

M. D.

Lizotte

A. J.

Smith

C. A.

Tobin

(2003). Gang and delinquency in developmental perspective. New York, NY: Cambridge University Press.

73.

Thorsell

B. A.

Klemke

L. W.

(1972). The labeling process: Reinforcement and deterrent. Law & Society Review, 6, 393-403.

74.

Zimring

F. E.

Hawkins

G. J.

(1973). Deterrence. Chicago, IL: University of Chicago Press.