Reappraising the Impact of Offending on Victimization: A Propensity Score Matching Approach

Abstract

Existing evidence clearly supports an empirical connection between offending and victimization. Often called the “victim–offender overlap,” this relationship holds for both sexes, across the life course, and across a wide range of countries and cultural environments. In addition, the relationship is sustained regardless of the study sample and statistical methods applied in the analyses of the sample data. However, there has yet to be a study that examines this relationship for violent and property crime using quasi-experimental methods accounting for a wide range of potential confounders including individual differences and cultural contexts. This study subjects the victim–offender relationship to testing through propensity score matching for both violent and property crimes using an international dataset. The results show that previous violent and theft offending increases the odds of victimization when matching on individual and contextual factors. This finding supports previous literature and suggests that delinquent behavior may act as a “switch” that exposes one to subsequent violent and theft victimization.

Keywords

international offending propensity score matching treatment effects victimization

Introduction

There are few relationships in criminology as stable as the one between victimization and offending. A multitude of empirical studies has supported the finding that victimization is related to offending, and vice versa, and that this association holds across various methodological and statistical techniques (Jennings, Piquero, & Reingle, 2012). The link between violent and property offending and victimization is also found to hold across international and cultural contexts and does not appear to be culturally dependent (Jennings et al., 2012; Posick, 2013), although there has been some modest evidence that there may be cultural factors that increase or attenuate the strength of the overlap such as the level of individualism in a country (Posick & Gould, 2015). Finally, the overlap holds in both cross-sectional and longitudinal studies (Daigle, Beaver, & Hartman, 2008; Jennings, Higgins, Tewksbury, Gover, & Piquero, 2010).

Despite the ubiquity of the victim–offender overlap, there has not been significant attention paid to whether or not this relationship holds using statistical techniques that approximate quasi-experimental designs. For exceptions to this, see work by Jennings and colleagues on interpersonal violence and the victim–offender overlap using quasi-experimental methods (Jennings, Richards, Tomsich, & Gover, 2015, 2013; Tomsich et al., 2015). In other words, there are remaining questions, such as “is there a ‘treatment effect’ of prior violent/theft offending on subsequent victimization?” In addition, “would any treatment effect hold across international contexts?” Such a design would hone in on the overall effect of offending on victimization using a very conservative method of estimation.

While it is impossible, and certainly unethical, to randomly assign individuals to groups that would either offend or be victimized (to get at the victim–offender overlap), propensity score matching (PSM) allows researchers to mimic the experimental conditions by matching individuals on their overall likelihood—or propensity—to experience a certain outcome. The present study uses PSM to investigate if prior violent and property offending (the “treatments”) have an effect on subsequent violent and property victimization (the outcomes) after matching individuals on several empirically established confounding variables.

Literature Review and Theoretical Backdrop

Because PSM matches individuals based on an overall score calculated using several variables that are theoretically and empirically associated with the outcome (here victimization), it is important to establish why the variables were chosen and how they are linked with the outcome. For the present study, individual, school, and neighborhood factors were used in calculating the propensity score. Each set of factors is related in some way to victimization and has been implicated in the victim–offender overlap. This section, therefore, provides a review of the literature on the variables used in the matching of cases, as well as the treatment and outcome variables, for the present study while discussing why these variables were chosen from a theoretical perspective.

In the following sections, the justification for the matching variables is discussed along with the coding strategy used for each individual variable. Therefore, the data are discussed briefly here before moving on to the subsequent sections. The data for this study come from the second International Self-Report Delinquency Study (ISRD-2), a school-based sample of youth in Grades 7 to 9 in 30 countries who responded to a self-report questionnaire. Countries were included based on the willingness of the country to participate in the study—therefore, the ISRD-2 is a convenience sample at the country level. ISRD-2 researchers utilized a multistage sampling strategy within each country where each participating nation included small-, medium-, and large-sized cities, with a total of at least 2,600 cases from each country (Enzmann et al., 2010; Junger-Tas et al., 2012). Cities were selected based on their size, degree of urbanization, and economic/demographic information. A listing of all schools in the research cities was created that included all seventh-, eighth-, and ninth-grade classrooms, and classrooms were randomly drawn from the original list of classrooms. Public, private, vocational, technical, and academic schools were included in the sample.

Outcome and Treatment Variables

The outcomes, or dependent variables, are prevalence measures of violent victimization (victim of either robbery or assault or both) and theft victimization (victim of theft/stealing) experienced in the previous 12 months. The treatment variables are previous violent offending (committing a robbery or assault or both) and theft offending (committing a theft from a person, car, or shop or the theft of a bicycle) that occurred within the person’s lifetime but not within the past 12 months. These variables are used to examine the victim–offender overlap in the logistic regression and PSM analyses. Given that the prevalence rates for violence are fairly low in this sample (as it is a relatively young, general population sample), including an analysis of theft has two major advantages: (a) it provides a robustness test for the results, and (b) it makes the additional contribution of considering the theft victim–offender overlap which is understudied but still has empirical backing (see Posick, 2013).

While it was possible to ensure that the offending occurred before the victimization given that the questions ask about a lifetime and 12-month time frame for offending, the attitudinal data are cross-sectional. Therefore, because the respondents were asked about victimization experiences and the matching variables (to be discussed next) at the same time, it is possible that these variables have been affected by the prior offending. While a limitation, it is likely that this allows for an even more conservative estimate of the treatment effect as the treated and control groups are potentially matched not only on predictors of offending but also on some consequences of offending.

The relationship between offending and victimization has garnered much recent attention, and a variety of crime types have been investigated using rigorous methods and statistical techniques. For instance, Mulford and colleagues (2016) uncover a strong relationship between victimization and offending in felony convictions. Over time, they see that those who continue to offend violently are also repeat victims and that desisting from offending is much more common for those who are not currently being victimized. Recent findings reveal that those individuals who are bullied are at much greater risk of offending later in life—even into adulthood (DeCamp & Newby, 2015). A substantial literature also indicates that those who are involved with dating conflicts are often exposed to violence as both victims and offenders (Ybarra, Espelage, Langhinrichsen-Rohling, & Korchmaros, 2016) and that those involved with intimate partner violence are much more likely to be both victims and offenders (Muftić, Finn, & Marsh, 2015). In sum, the current status of the victim–offender overlap is supported across a large subset of crime types. Still, this examination must be accounted for across contexts using conservative estimation of effects.

Individual Variables

Demographic characteristics are important when investigating the victim–offender overlap as the common offender reflects the common victim—they are more often than not young, Black, unmarried, and male (Lauritsen & Laub, 2007; Lauritsen, Sampson, & Laub, 1991). These demographics also reflect the common repeat victim (Chang, Chen, & Brownson, 2003). Accounting for these confounding variables is essential for the PSM analysis. Sex, grade, and immigrant status are demographic characteristics included in the matching analysis. Sex is measured by the variable Male which is a dichotomous measure where 0 = female, 1 = male. Grade is a proxy for age and is often used as opposed to age in school-based samples because it more adequately reflects “social” age (Junger-Tas et al., 2012). Grade 7 and Grade 8 are included in the logistic regression models with Grade 9 left out as the reference category. Nonnative is a dichotomous variable (0 = native; 1 = nonnative) indicating whether or not the respondent is native born or a first or second generation immigrant in their respective country.

Characteristics about one’s family can lend insight into offending and victimization risk. Hirschi’s (1969, 2004) seminal work on social control has provided strong evidence that family bonding is a protective factor for offending. Family togetherness, or social capital, exerts a strong protective factor on juvenile delinquency even above that provided by the school (Dufur, Hoffmann, Braudt, Parcel, & Spence, 2015). Elements of family bonding such as lack of family involvement, parental monitoring, and attachment to parents are also predicative of victimization (Esbensen, Huizinga, & Menard, 1999) while positive family climate can insulate individuals from victimization (Schreck & Fisher, 2004). Family bonding has been associated with offending and victimization in international (Posick & Rocque, 2015) and longitudinal studies (Piquero, MacDonald, Dobrin, Daigle, & Cullen, 2005) making the association quite robust. Family bonding is measured here using four items indicating the respondent’s (a) closeness to their mother and (b) father; (c) leisure time with their family; and (d) time spent having dinner with their family (α = .55).

It should be noted that the alpha of .55 is quite low for this measure. While not ideal, keeping the measure in the analyses provides a better theoretical rationale for the model, and the low alpha, in this case, makes the measure conservative in its effect on the outcome variable. Dropping the variable entirely does not allow for any matching on this measure, and it does not allow for controlling its effect in the regression analyses. As the measure is used for matching (in the primary analysis) and not estimation, it is a little less problematic for the current study.

Empirical work connecting delinquent peer association with offending and delinquency has been plentiful (see Warr, 2002). Antisocial behavior is largely a group activity and those who associate with others who exhibit antisocial behavior are more likely to be antisocial themselves—this is particularly true when considering escalation of and desistence from antisocial behavior (Monahan, Rhew, Hawkins, & Brown, 2014). Likewise, studies investigating the link between exposure to victimization and association with delinquent peers reveal a positive relationship (Schreck & Fisher, 2004; Schreck, Fisher, & Miller, 2004). Belonging to a delinquent group, such as a gang, substantially increases one’s chances of offending and victimization likely because of the cycle of violence inherent in gang-related retaliation (Taylor, Peterson, Esbensen, & Freng, 2007; Thornberry, Krohn, Lizotte, Smith, & Tobin, 2003). Delinquent peer association is assessed by asking participants about the number of delinquent activities engaged in by their friends which include five different behaviors: (a) drug use, (b) stealing, (c) burglary, (d) assault, and (e) robbery. These dichotomous items were coded as 0 = no and 1 = yes. These items were summed to create a scale ranging from 0 to 5 where higher scores indicate a greater number of delinquent acts by friends (α = .70).

Self-control is one of the strongest predictors of crime and continues to dominate explanations of criminal behavior across time, gender, and geography (Pratt & Cullen, 2000; Rocque, Posick, & Piquero, 2016). Schreck (1999) extended the general theory of crime to explain victimization and concluded that low self-control exposes individuals to victimization because those who are impulsive and neglect long-term consequences of behavior are more likely than those who do have self-control to engage in risky behaviors and put themselves in dangerous environments. Since Schreck’s seminal piece, meta-analysis has confirmed the association between low self-control and victimization (Pratt, Turanovic, Fox, & Wright, 2014). Self-control is measured using a modified Grasmick, Tittle, Bursik, and Arneklev (1993) scale which includes 12 items from the original scale covering temperament, self-centeredness, risk-seeking, and impulsivity (α = .83).

School Variables

The school context plays an important role in the level and nature of violence within its walls. Positive school climate (e.g., how students feel about their school, classes, teachers, and activities) reduces potential for violence in the school (Steffgen, Recchia, & Viechtbauer, 2013). The school context also has been found to shape the victim–offender overlap explicitly—weakening or strengthening the overlap depending on the type (urban vs. rural) of school (Posick & Zimmerman, 2015). School Climate is measured using four survey items asking the student about his or her school including (a) if they would miss their school, (b) if they have attentive teachers, (c) if they like their school, and (d) if there are other activities at the school to participate in (α = .65).

On the flipside of positive school climate, disorganized schools, just like disorganized communities, increase violence. Violence in schools and school disorganization have reciprocal relationships whereby those schools that are the most disorganized have the most crime, but school crime can also influence perception of school disorganization (Leadbeater, Sukhawathanakul, Smith, & Bowen, 2015). Furthermore, bullying in school, which is more likely in disorganized schools, has been shown to have a substantial overlap between those who are bully victims and offenders (Chan & Wong, 2015; DeCamp & Newby, 2015). School disorganization is captured using four questions on the survey asking respondents if there is (a) stealing, (b) fighting, (c) vandalism, and (d) drug use in their school (α = .75).

Neighborhood Variables

Living in a disorganized community not only promotes delinquent behavior but also exposes individuals to a host of violent situations (Butcher, Galanek, Kretschmar, & Flannery, 2015; Fox, Lane, & Akers, 2010). Crime is most heavily concentrated within small portions of cities that are highly disorganized, exposing residents to both opportunities to offend and be victimized (Braga & Clarke, 2014). Perceptions of neighborhood disorganization are measured by five questions asking the respondent if they perceive high levels of (a) crime, (b) fighting, (c) drug selling, (d) empty buildings, and (e) graffiti in their neighborhood. Higher scores indicate higher levels of perceived neighborhood disorganization (α = .82).

Neighborhoods that are perceived to be cohesive and have neighbors that trust and rely on one another are thought to be safer and experience fewer crimes (Hipp, 2016). Collective efficacy, the ability of neighbors to come together to address problems in the neighborhood, leads to lower crime in the community (Sampson, Raudenbush, & Earls, 1997) and, therefore, less victimization (Browning, 2002). When residents come to each other’s aid and intervene when delinquency is observed, much crime and violence is prevented or de-escalated. To account for the association between neighborhood collective efficacy and violence, neighborhood efficacy is measured using four items from the survey including if the respondent thought their neighbors were (a) helpful, (b) trustworthy, (c) close-knit, and (d) easy to get along with. Higher scores correspond to greater levels of neighborhood efficacy (α = .75).

The PSM analysis also matches on country of origin—that is, matches were forced to be made within each country. The victim–offender overlap has been shown to exist across the international context but be weaker or stronger depending on country-level characteristics (Posick & Gould, 2015) indicating the need to account for this phenomenon. In addition, the victim–offender overlap has yet to be tested using PSM models accounting for cross national/cultural variation. This is an attractive feature of the subsequent analysis that will put to task the victim–offender overlap using a very rigorous and conservative framework.

The continuous measures used in the matching are converted to percentage of maximum possible scores using the method recommended by Cohen, Cohen, Aiken, and West (1999). This converts scales to a score between 0 and 100 for easier comparison of descriptive statistics. As it is a linear transformation, all analytic techniques are identical to those using additive or mean scales.

The Present Study: The PSM Methodology

The PSM approach allows the researcher to simulate a true experiment by creating two groups of individuals who are matched on several relevant characteristics (in this study—those described above) in an effort to “equalize” the two groups (McCartney, Bub, & Burchinal, 2006). This is very helpful for theoretical work in the social sciences, in particular, because often, the “treatment” cannot be randomly assigned because it is unethical to randomly assign some individuals to get the so-called treatment. Here that would mean randomly assigning individuals to offend against someone and others to not. In addition, individuals self-select into the treatment groups which should be accounted for in the statistical analysis, and one of the best ways to do so is creating a propensity score for individuals to be matched (Apel & Sweeten, 2010; see also Wermink, Blokland, Nieuwbeerta, Nagin, & Tollenaar, 2010).

One of the pressing matters, of much debate in the current social science literature, is how the PSM method achieves a greater estimation of effects beyond regression analysis. There are two major advantages to the PSM in this particular case. First, by creating a propensity score from several relevant covariates, the model preserves degrees of freedom (Shadish, 2013). Second, in regression analysis, all covariates have an effect on the outcome, and the coefficients are estimated using an additive (linear) model which makes it difficult to isolate the treatment effect. Relatedly, PSM allows for a test of the counterfactual. In other words, it is possible to estimate what would have happened if those who did not offend did offend and vice versa. The treatment effect is the residual of the estimate and observed values of the counterfactuals. The average of all the residuals is the estimated treatment effect (Apel & Sweeten, 2010).

Results

Before getting to the treatment effects, it is important in PSM to meet several criteria which are outlined by Apel and Sweeten (2010). These criteria include (1) using a clear definition of treatment, (2) using thoughtful model specification, (3) including temporally prior confounders, (4) demonstrating common support, (5) demonstrating covariate balance, and (6) employing sensitivity analyses. In this study, I address each of these suggestions to the extent the data allow. As mentioned in the prior paragraphs, the definition of treatment is prior offending which is a prevalence measure coded as 0 = no and 1 = yes. In other words, those in the treated group have participated in violent and/or theft offending that occurred before the measured subsequent victimization. In this case, temporal ordering is preserved. However, pertaining to Criterion 3, the treatment (offending) is the only variable that I was able to ensure was temporally prior to the outcome (victimization) variable. In regard to Criterion 2, all included variables are related to the dependent variable both theoretically as evidenced by the literature review and in preliminary analyses in the current study and are thus specified in the model.

To establish the need for PSM, a multilevel model was run regressing violent and theft offending on victimization. If no relationship exists in this analysis, the PSM would not be strongly supported. The result of the regression analysis are presented in Table 1. Odds ratios from the logistic regression indicate that involvement in violent offending increases the odds of victimization by 22% and that the same is true of theft offending and theft victimization. In addition, almost all of the covariates (which are used as matching variables in the PSM analysis) are significantly related to victimization. However, given the large sample size, many of the significant results are not substantial. Regardless, this model reaffirms the specification of the PSM model and provides further support for meeting Criterion 2 listed earlier.

Table 1.

Multilevel Logistic Models Regressing Violent and Theft Victimization on Study Variables (Individual n = 57,033; Country n = 30).

	Violent victimization		Theft victimization
	OR (SE)	95% CI	OR (SE)	95% CI
Previous offending
Violent offending	1.22 (0.11)*	[1.02, 1.45]	—	—
Theft offending	—	—	1.22 (0.04)***	[1.15, 1.30]
Individual characteristics
Sex (male)	2.10 (0.08)***	[1.96, 2.26]	1.08 (0.02)***	[1.04, 1.13]
Grade 7	1.27 (0.06)***	[1.16, 1.38]	1.18 (0.03)***	[1.11, 1.24]
Grade 8	1.15 (0.05)**	[1.06, 1.26]	1.12 (0.03)***	[1.06, 1.18]
Nonnative	1.23 (0.08)**	[1.09, 1.39]	1.27 (0.05)***	[1.18, 1.38]
Family bonding	0.99 (0.00)***	[0.99, 0.99]	0.99 (0.00)***	[0.99, 1.00]
Delinquent peers	1.26 (0.02)***	[1.22, 1.29]	1.21 (0.01)***	[1.19, 1.23]
Self-control	0.99 (0.00)***	[1.00, 1.01]	1.00 (0.00)***	[0.99, 1.00]
School variables
School climate	1.00 (0.00)	[1.00, 1.00]	1.00 (0.00)*	[1.00, 1.00]
School disorganization	1.01 (0.00)***	[1.00, 1.01]	1.01 (0.00)***	[1.01, 1.01]
Neighborhood variables
Neighborhood disorganization	1.01 (0.00)***	[1.01, 1.01]	1.00 (0.00)*	[1.00, 1.00]
Neighborhood bonding	1.00 (0.00)***	[1.00, 1.00]	1.00 (0.00)***	[1.00, 1.00]
Intercept	0.07 (0.01)***	[0.06, 0.10]	0.24 (0.02)***	[0.20, 0.29]
Loglikelihood	−12,863.29		−26,531.32

Note. OR = odds ratio; 95% CI = 95% confidence interval.

p < .05. **p < .01. ***p < .001.

Common support is found in the matching conducted in this study as evidenced in Figure 1. Common support indicates that a match was found for each case in the treatment group (with a match that did not receive the treatment). This is important because, logically, those who receive a “treatment” like participating in violence are going to be different than those who do not participate in violence on the confounders being matched (e.g., self-control, neighborhood disorganization, school climate). In other words, those who receive treatment are more likely to have low self-control and live in areas with more disorganization than those who do not receive the treatment making it potentially difficult to find a matched control. Table 2 shows that common support was accomplished and each treatment case found a match. This fulfills the fourth recommendation by Apel and Sweeten (2010). As expected, the treated and untreated differed on almost every study variable. The treated had more or higher levels of each risk factor and fewer or lower levels of the protective factors.

Figure 1.

Balance plot for violent victimization.

Table 2.

Descriptive Statistics of Study Variables for Treated and Untreated Groups.

	Violent victimization		Theft victimization
	Untreated (n = 55,937)	Treated (n = 1,069)	Untreated (n = 49,858)	Treated (n = 7,100)
Individual characteristics
Sex (male)	47.56***	69.71	46.97***	55.38
Grade 7	32.62***	23.54	33.55***	24.66
Grade 8	32.99	34.85	33.06	33.07
Grade 9	34.39***	41.61	33.39***	42.27
Nonnative	7.13***	11.95	6.95***	9.00
Family bonding	81.48 (0.07)***	73.57 (0.63)	82.19 (0.07)***	75.16 (0.22)
Delinquent peers	0.75 (0.00)***	2.09 (0.05)	0.64 (0.00)***	1.73 (0.02)
Self-control	61.96 (0.08)***	44.84 (0.59)	62.86 (0.09)***	53.16 (0.23)
School variables
School climate	69.49 (0.09)***	61.32 (0.74)	70.20 (0.10)***	63.71 (0.27)
School disorganization	36.81 (0.10)***	51.54 (0.78)	36.01 (0.11)***	44.85 (0.30)
Neighborhood variables
Neighborhood disorganization	20.14 (0.10)***	35.56 (0.88)	19.45 (0.10)***	27.43 (0.31)
Neighborhood bonding	74.97 (0.12)***	70.36 (1.00)	75.37 (0.13)***	71.44 (0.37)
Common support	100%	100%	100%	100%

Note. The p values for significant differences determined using two-sided phi tests for dichotomous variables and t tests for continuous variables.

p < .05. **p < .01. ***p < .001.

To ensure that each case is similar enough to its matches, covariate balance, Criterion 5, must be met. The PSM successfully balanced covariates ensuring the conditional independence assumption. In PSM, the researcher must compare the overlap of the propensity score matches between the control and treatment groups (Apel & Sweeten, 2010). Figures 1 (balance for the violence PSM model) and 2 (balance for the theft PSM model) show that the raw data are not quite balanced, which is to be expected as that is the job of the PSM model. After matching, balance is achieved. The curves on the right-hand side of the figures are nearly identical suggesting that the covariates (i.e., confounders) are properly balanced.

Figure 2.

Balance plot for theft victimization.

Bias was explored in Table 3. Bias in the matching of treatment and control groups would indicate that the lack of matching might impact the results. While no matching method using any dataset will be perfect, it is essential to check for potential mismatches that could influence the results. By comparing the means of each of the study variables, one can calculate the percent bias. When the percent bias exceeds -3.0 or 3.0, there may be bias in the matching that would influence the results. Rubin’s B is a statistic that indicates whether the bias observed in the mean differences between control and treatment groups is problematic (Rubin, 2001). Rubin’s B was not statistically significant in any of the cases in the current study suggesting that any bias is nonproblematic and does not impact the results.

Table 3.

Means of Study Variables for Treated and Control Groups.

	Violent victimization			Theft victimization
	Treated	Control	% bias	Treated	Control	% bias
Individual characteristics
Sex (male)	0.70	0.71	–3.0	0.54	0.55	–1.1
Grade 7	0.24	0.23	1.6	0.25	0.24	1.8
Grade 8	0.35	0.36	–2.7	0.33	0.33	0.6
Nonnative	0.12	0.11	1.9	0.09	0.08	2.5
Family bonding	73.57	73.41	0.9	75.16	75.23	–0.4
Delinquent peers	2.09	2.04	3.3	1.73	1.67	4.3
Self-control	44.84	46.34	–7.6	53.16	53.17	0
School variables
School climate	61.32	61.52	–0.9	63.71	64.05	–1.5
School disorganization	51.54	51.69	–0.6	44.85	44.57	1.2
Neighborhood variables
Neighborhood disorganization	35.56	34.49	4.1	27.43	27.47	–0.2
Neighborhood bonding	70.36	67.70	8.5	71.44	70.79	2.2

Note. Bias is determined to be nonproblematic after examining Rubin’s B (Rubin, 2001).

The results of the treatment effect analysis (which uses a one-to-one nearest neighbor approach with no specified caliper) are depicted in Figure 3 and show that for both violent and theft offending, there was a significant and modest treatment effect on victimization. The average treatment effect (ATE), or the difference in mean (average) outcomes between individuals assigned to the treatment and individuals assigned to the control at the population level (from untreated to treated), indicates that those individuals who are involved in offending increase their probability of experiencing victimization by 5.6% for violence and 5.8% for theft. The average treatment effect on the treated (ATT), the average effect of treatment on those subjects who ultimately received the treatment, is 2.5% for violence and 4.1% for theft. This suggests that, on average, offenders—given all of their characteristics—increase their probability of victimization by 2.5% for violence and 4.1% for theft by offending rather than if they abstain from offending behavior. Finally, the average treatment effect on the untreated (ATU), or the average effect of treatment on those subjects who ultimately did not receive the treatment, is 5.6% for violence and 6.0% for theft. In other words, if those individuals in the untreated group who do not offend hypothetically do offend, they increase their probability of violent victimization by 5.6% and theft victimization by 6.0%.

Figure 3.

Treatment effects.

Sensitivity Analyses

In PSM, it is always a good idea to conduct sensitivity analyses to ensure that the model presented is not sensitive to any one specification. As PSM is fairly new in the social sciences, the “best practices” for model specification are only recently being identified (see, for example, Peters, Hochstetler, DeLisi, & Kuo, 2015). A table of supplemental results is presented in the appendix. Sensitivity analyses were carried out which varied the “bandwidth” or distance between matched cases (also known as calipers) on study variables (calipers of .001, .01, .02, and .03 were used). In addition, various propensity score and nearest neighbor marching methods were used and an alternative weighting technique was applied (inverse probability weighting). The results of the supplemental analyses showed no substantial differences for multiple nearest neighbors within various calipers for the different methods.

Summary and Conclusion

The results of the present study parallel past research. Even after matching individuals on a host of theoretically relevant variables, violent and theft offending predicted subsequent violent and theft victimization. The treatment effect was found when matching within country of origin suggesting the treatment effect is robust across international and cultural contexts. While the overall treatment effect is modest, it exists, suggesting that prior offending is a powerful switch that exposes one to later victimization.

The PSM approach used in this study provided a unique examination of the victim–offender overlap which should be used for other associations identified in the criminological and victimological literature and to answer other remaining research questions (see, for example, Boutwell & Beaver, 2010; DeLisi, Barnes, Beaver, & Gibson, 2009; Wermink et al., 2010). If science is truly about attempting to disprove claims (in an effort to discover the real relationships in our social world), researchers must expose these social relationships to rigorous analysis. PSM is one way to achieve this goal in the absence of random controlled trials (for a couple of cautionary notes, refer to Shadish, 2013 and Thoemmes & Kim, 2011).

The results have significant theoretical and policy implications. First, the generality of the victim–offender overlap for both violence and theft holds across the international context using a quasi-experimental design suggesting that the effects of offending on victimization hold in different cultural contexts. This substantiates the theoretical underpinnings of the overlap. Programs and policies aimed at addressing the risk factors of those who have been victimized may want to consider the past offending histories of the victims. Second, intervention programs that step in to thwart retaliatory violence can look to both those who have been victimized and those who have offending histories as a starting point as the two are intimately linked, regardless of time and place.

While this study attempted to be as rigorous as possible, limitations exist suggesting that research on the victim–offender relationship is far from complete. First, only the treatment could be assured to be temporally prior to the outcome. PSM is strongest when all confounders are temporally prior to the outcome. Second, offending and victimization are relatively rare events in the dataset used in this study. This limits the generalizability of results despite the PSM modeling approach. Like any experiment, PSM works best when a significant proportion of the sample experiences the treatment and outcome. Finally, not all possible confounders were specified in the model highlighting areas where future research can match on additional variables. Similarly, only certain violent and theft victimization outcomes were explored leaving open likely contingencies in the victim–offender overlap using alternative outcome measures. Others, such as Zimmerman, Farrell, and Posick (2017), find that the overlap varies by the victim–offender relationship which is not measurable in this dataset. Overall, the victim–offender overlap, once again, stands against attempts to disprove it, but the work is not yet done. PSM and other experimental and quasi-experimental approaches to testing this relationship are needed and indicate exciting areas for future research.

Footnotes

Appendix

Supplemental/Sensitivity Analyses.

	ATE	SE	p value
Violence	0.0564059	0.020499	.006
Theft	0.0577706	0.0100231	.000
3 nearest neighbors (violence)	0.0331036	0.0112438	.003
5 nearest neighbors (violence)	0.02976	0.0094147	.002
3 nearest neighbors (theft)	0.0522359	0.0090219	.000
5 nearest neighbors (theft)	0.0537659	0.0086297	.000
IPW (violence)	0.0497571	0.0134092	.000
IPW (theft)	0.0619044	0.0068567	.000

Note. ATE = average treatment effect. IPW = Inverse Probability Weighting.

Acknowledgements

The author wishes to thank Gary Sweeten for helpful comments on this article.

Declaration of Conflicting Interests

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

Funding

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

References

Apel

R. J.

Sweeten

(2010). Propensity score matching in criminology and criminal justice. In Piquero

A. R.

Weisburd

(Eds.), Handbook of quantitative criminology (pp. 543-562). New York, NY: Springer.

Boutwell

B. B.

Beaver

K. M.

(2010). The role of broken homes in the development of self-control: A propensity score matching approach. Journal of Criminal Justice, 38, 489-495.

Braga

A. A.

Clarke

R. V.

(2014). Explaining high-risk concentrations of crime in the city: Social disorganization, crime opportunities, and important next steps. Journal of Research in Crime and Delinquency, 51, 4, 480-498. doi:10.1177/0022427814521217

Browning

C. R.

(2002). The span of collective efficacy: Extending social disorganization theory to partner violence. Journal of Marriage and Family, 64(4), 833-850.

Butcher

Galanek

J. D.

Kretschmar

J. M.

Flannery

D. J.

(2015). The impact of neighborhood disorganization on neighborhood exposure to violence, trauma symptoms, and social relationships among at-risk youth. Social Science & Medicine, 146, 300-306.

Chan

H. C. O.

Wong

D. S.

(2015). The overlap between school bullying perpetration and victimization: Assessing the psychological, familial, and school factors of Chinese adolescents in Hong Kong. Journal of Child and Family Studies, 24, 3224-3234.

Chang

J. J.

Chen

J. J.

Brownson

R. C.

(2003). The role of repeat victimization in adolescent delinquent behaviors and recidivism. Journal of Adolescent Health, 32, 272-280.

Cohen

Aiken

L. S.

West

S. G.

(1999). The problem of units and the circumstance for POMP. Multivariate Behavioral Research, 34, 315-346.

Daigle

L. E.

Beaver

K. M.

Hartman

J. L.

(2008). A life-course approach to the study of victimization and offending behaviors. Victims & Offenders, 3, 365-390.

10.

DeCamp

Newby

(2015). From bullied to deviant: The victim–offender overlap among bullying victims. Youth Violence and Juvenile Justice, 13, 3-17.

11.

DeLisi

Barnes

J. C.

Beaver

K. M.

Gibson

C. L.

(2009). Delinquent gangs and adolescent victimization revisited: A propensity score matching approach. Criminal Justice and Behavior, 36, 808-823.

12.

Dufur

M. J.

Hoffmann

J. P.

Braudt

D. B.

Parcel

T. L.

Spence

K. R.

(2015). Examining the effects of family and school social capital on delinquent behavior. Deviant Behavior, 36, 511-526.

13.

Enzmann

Marshall

I. H.

Killias

Junger-Tas

Steketee

Gruszczynska

(2010). Self-reported youth delinquency in Europe and beyond: First results of the Second International Self-Report Delinquency Study in the context of police and victimization data. European Journal of Criminology, 7, 159-183.

14.

Esbensen

F. A.

Huizinga

Menard

(1999). Family context and criminal victimization in adolescence. Youth & Society, 31, 168-198.

15.

Fox

K. A.

Lane

Akers

R. L.

(2010). Do perceptions of neighborhood disorganization predict crime or victimization? An examination of gang member versus non-gang member jail inmates. Journal of Criminal Justice, 38, 720-729.

16.

Grasmick

H. G.

Tittle

C. R.

Bursik

R. J.

Arneklev

B. J.

(1993). Testing the core empirical implications of Gottfredson and Hirschi’s general theory of crime. Journal of Research in Crime & Delinquency, 30, 5-29.

17.

Hipp

J. R.

(2016). Collective efficacy: How is it conceptualized, how is it measured, and does it really matter for understanding perceived neighborhood crime and disorder? Journal of Criminal Justice, 46, 32-44.

18.

Hirschi

(1969). Causes of delinquency. Berkeley: University of California Press.

19.

Hirschi

(2004). Self-control and crime. In Baumeister

R. F.

Vohs

K. D.

(Eds.), The handbook of self-regulation: Research, theory, and applications (pp. 537-552). New York, NY: Guilford Press.

20.

Jennings

W. G.

Higgins

G. E.

Tewksbury

Gover

A. R.

Piquero

A. R.

(2010). A longitudinal assessment of the victim-offender overlap. Journal of Interpersonal Violence, 25, 2147-2174.

21.

Jennings

W. G.

Piquero

A. R.

Reingle

J. M.

(2012). On the overlap between victimization and offending: A review of the literature. Aggression and Violent Behavior, 17, 16-26.

22.

Jennings

W. G.

Richards

Tomsich

Gover

(2015). Investigating the role of child sexual abuse in intimate partner violence victimization and perpetration in young adulthood from a propensity score matching approach. Journal of Child Sexual Abuse, 24, 659-681.

23.

Jennings

W. G.

Richards

Tomsich

Gover

Powers

(2013). A critical examination of the causal link between child abuse and adult dating violence perpetration and victimization from a propensity score matching approach. Women & Criminal Justice, 23, 167-184.

24.

Junger-Tas

Marshall

I. H.

(2012). Introduction to the International Self-Report Study of Delinquency (ISRD-2). In Junger-Tas Marshall

I.H.

Enzmann

Killias

Steketee

Gruszczynska

. (Eds.), The many faces of youth crime (pp. 3-20). New York, NY: Springer.

25.

Lauritsen

J. L.

Laub

J. H.

(2007). Understanding the link between victimization and offending: New reflections on an old idea. Crime Prevention Studies, 22, 55-76.

26.

Lauritsen

J. L.

Sampson

R. J.

Laub

J. H.

(1991). The link between offending and victimization among adolescents. Criminology, 29, 265-292.

27.

Leadbeater

Sukhawathanakul

Smith

Bowen

(2015). Reciprocal associations between interpersonal and values dimensions of school climate and peer victimization in elementary school children. Journal of Clinical Child & Adolescent Psychology, 44, 480-493.

28.

McCartney

Bub

K. L.

Burchinal

M. R.

(2006). Selection, detection, and reflection. In McCartney

Burchinal

M. R.

Bub

K. L.

(Eds.), Best practices in quantitative methods for developmentalists (Monographs of the Society for Research in Child Development, Vol. 71, No. 3, pp. 105-126). Boston, MA: Blackwell.

29.

Monahan

K. C.

Rhew

I. C.

Hawkins

J. D.

Brown

E. C.

(2014). Adolescent pathways to co-occurring problem behavior: The effects of peer delinquency and peer substance use. Journal of Research on Adolescence, 24, 630-645.

30.

Muftić

L. R.

Finn

M. A.

Marsh

E. A.

(2015). The victim-offender overlap, intimate partner violence, and sex: Assessing differences among victims, offenders, and victim-offenders. Crime & Delinquency, 61, 899-926.

31.

Mulford

C. F.

Blachman-Demner

D. R.

Pitzer

Schubert

C. A.

Piquero

A. R.

Mulvey

E. P.

(2016). Victim offender overlap: Dual trajectory examination of victimization and offending among young felony offenders over seven years. Victims & Offenders. Advance online publication. doi:10.1080/15564886.2016.1196283

32.

Peters

D. J.

Hochstetler

DeLisi

Kuo

H. J.

(2015). Parolee recidivism and successful treatment completion: Comparing hazard models across propensity methods. Journal of Quantitative Criminology, 31, 149-181.

33.

Piquero

A. R.

MacDonald

Dobrin

Daigle

L. E.

Cullen

F. T.

(2005). Self-control, violent offending, and homicide victimization: Assessing the general theory of crime. Journal of Quantitative Criminology, 21, 55-71.

34.

Posick

(2013). The overlap between offending and victimization among adolescents results from the second international self-report delinquency study. Journal of Contemporary Criminal Justice, 29, 106-124.

35.

Posick

Gould

L. A.

(2015). On the general relationship between victimization and offending: Examining cultural contingencies. Journal of Criminal Justice, 43, 195-204.

36.

Posick

Rocque

(2015). Family matters: A cross-national examination of family bonding and victimization. European Journal of Criminology, 12(1), 51-69.

37.

Posick

Zimmerman

G. M.

(2015). Person-in-context: Insights on contextual variation in the victim-offender overlap across schools. Journal of Interpersonal Violence, 30, 1432-1455.

38.

Pratt

T. C.

Cullen

F. T.

(2000). The empirical status of Gottfredson and Hirschi’s general theory of crime: A meta-analysis. Criminology, 38, 931-964.

39.

Pratt

T. C.

Turanovic

J. J.

Fox

K. A.

Wright

K. A.

(2014). Self-control and victimization: A meta-analysis. Criminology, 52, 87-116.

40.

Rocque

Posick

Piquero

A. R.

(2016). Self-control and crime: Theory, research, and remaining puzzles. In Vohs

K. D.

Baumeister

R. F.

(Eds.), Handbook of self-regulation: Research, theory, and applications (3rd ed., pp. 514-532). New York, NY: Guilford Press.

41.

Rubin

D. B.

(2001). Using propensity scores to help design observational studies: Application to the tobacco litigation. Health Services and Outcomes Research Methodology, 2, 169-188.

42.

Sampson

R. J.

Raudenbush

S. W.

Earls

(1997). Neighborhoods and violent crime: A multilevel study of collective efficacy. Science, 277, 918-924.

43.

Schreck

C. J.

(1999). Criminal victimization and low self-control: An extension and test of a general theory of crime. Justice Quarterly, 16, 633-654.

44.

Schreck

C. J.

Fisher

B. S.

(2004). Specifying the influence of family and peers on violent victimization extending routine activities and lifestyles theories. Journal of Interpersonal Violence, 19, 1021-1041.

45.

Schreck

C. J.

Fisher

B. S.

Miller

J. M.

(2004). The social context of violent victimization: A study of the delinquent peer effect. Justice Quarterly, 21, 23-47.

46.

Shadish

W. R.

(2013). Propensity score analysis: Promise, reality and irrational exuberance. Journal of Experimental Criminology, 9, 129-144.

47.

Steffgen

Recchia

Viechtbauer

(2013). The link between school climate and violence in school: A meta-analytic review. Aggression and Violent Behavior, 18, 300-309.

48.

Taylor

T. J.

Peterson

Esbensen

F. A.

Freng

(2007). Gang membership as a risk factor for adolescent violent victimization. Journal of Research in Crime & Delinquency, 44, 351-380.

49.

Thoemmes

F. J.

Kim

E. S.

(2011). A systematic review of propensity score methods in the social sciences. Multivariate Behavioral Research, 46, 90-118.

50.

Thornberry

T. P.

Krohn

M. D.

Lizotte

A. J.

Smith

C. A.

Tobin

(2003). Gangs and delinquency in developmental perspective. Cambridge, MA: Cambridge University Press.

51.

Tomsich

Jennings

W. G.

Richards

Gover

Powers

(2015). Childhood physical maltreatment and young adult dating violence: A propensity matching approach. Journal of Interpersonal Violence Violence. Advance online publication. doi:10.1177/0886260515599657

52.

Warr

(2002). Companions in crime: The social aspects of criminal conduct. Cambridge, UK: Cambridge University Press.

53.

Wermink

Blokland

Nieuwbeerta

Nagin

Tollenaar

(2010). Comparing the effects of community service and short-term imprisonment on recidivism: A matched samples approach. Journal of Experimental Criminology, 6, 325-349.

54.

Ybarra

M. L.

Espelage

D. L.

Langhinrichsen-Rohling

Korchmaros

J. D.

(2016). Lifetime prevalence rates and overlap of physical, psychological, and sexual dating abuse perpetration and victimization in a national sample of youth. Archives of Sexual Behavior, 45, 1083-1099.

55.

Zimmerman

G. M.

Farrell

Posick

(2017). Does the strength of the victim-offender overlap depend on the relationship between the victim and perpetrator? Journal of Criminal Justice, 48, 21-29.