Skip to main content

Constellations of youth criminogenic factors associated with young adult violent criminal behavior


This study identified constellations of childhood risk factors associated with violent criminal behavior in early adulthood. Police data were used to sample violent and nonviolent offenders from a population of young adult males with a history of juvenile probation. Risk factors were retrieved from their juvenile probation files. A single classification tree analysis organized these into a decision tree for violent criminal behavior with good predictive accuracy. Two constellations of risk factors were associated with a high risk of violent criminal behavior. The first consisted of juvenile delinquents who had been moderately involved with criminal peers, who had committed offenses under the influence of drugs, and who came from a dysfunctional family. The second was characterized by having been severely involved with criminal peers and having had criminal family members. Presenting with depressive symptoms in childhood was associated with a low risk of violent criminal behavior. These constellations bear clinical importance as they provide targets for personalized interventions.


It is widely acknowledged, specifically in high-income countries, that a minority of young adult offenders are responsible for the large majority of violent crimes (Falk et al. 2014; Farrington et al. 2006; Laub 2004; Snyder and Sickmund 2006). A vast literature has been amassed on the identification of juvenile offenders at risk to become persistent violent offenders (Baglivio et al. 2014; Hein et al. 2017). Early onset of antisocial or criminal behavior (Moffitt 1993), specifically before the age of 12 (DeLisi and Piquero 2011; Loeber and Farrington 2011), involvement in delinquency and drug use (Lipsey and Derzon 1998), negative or antisocial attitude and the accumulation of criminogenic factors, particularly over multiple domains (Baglivio et al. 2014), have been singled out as key determinants of persistence and escalation of criminal behavior into early adulthood.

However, not all juvenile delinquents who match this profile will develop persistent and violent criminal behavior. Even within high-risk groups there are those who desist from crime before or during the transition from adolescence into adulthood (Bushway et al. 2003; Farrington et al. 2009; Ttofi et al. 2016). From a harm reduction and public safety perspective, this constitutes a need to further differentiate within high-risk target groups for the timely identification of juvenile delinquents at highest risk to develop into young adult violent offenders. Early identification of those juvenile offenders at highest risk to persist or escalate after transitioning into early adulthood should allow for more personalized, specific and effective interventions.

Traditionally, criminological research focuses on prediction models and its resulting risk-assessment instruments are based on logistic regression or ordinary least squares analyses. These analyses primarily identify main effects of predictor variables on a particular outcome. However, regression models are less suited to test how to best combine predictors to obtain optimal predictive accuracy (Stalans et al. 2004). This problem becomes more apparent when the number of predictor variables is high, such as in high-risk groups with multiple risk factors at play. As a consequence, many prediction models tend to assume a ‘one size fits all’ approach (Ngo et al. 2015) that ignores individual particularities and disregards the possibility that not all subgroups within high-risk target populations share the same set of central risk factors (Steadman et al. 2000).

A focus on combinations, or constellations, of risk factors may better explain future violence than stand-alone criminogenic risk factors (Ahonen et al. 2016; Berk and Bleich 2014; Bushway 2013; Onifade et al. 2008). Classification tree analysis (CTA) seems suited to identify constellations of predictor variables associated with a particular outcome. To maximize predictive accuracy, CTA branches off a sample of cases/individuals (juvenile delinquents) in a stepwise fashion according to shared characteristics (criminogenic risk factors) that together are associated with a certain probability of meeting the outcome variable of interest (violent offending). It is a non-parametric technique that identifies homogeneous subgroups in a sample. CTA ranks predictors according to the strength of their association with the outcome and prunes predictors with only a marginal, but significant, contribution. Supposedly, CTA based models of risk assessment are on par with real-life clinical reasoning in making risk judgments, in which characteristics are evaluated sequentially according to their expected relevance to violence/recidivism (Steadman et al. 2000). Especially within high-risk target groups, CTA seems a viable means to identify those juveniles at highest risk of persistent and escalating criminal behavior. This identification is particularly useful when dynamic risk factors are taken into account because those can be targeted by interventions.

Although not conclusive, single CTA-techniques (e.g., CART, CHAID, Rpart) have performed well in predicting violence/recidivism in relatively small study samples (see Ngo et al. 2015 for an overview) such as psychiatric patients (Monahan et al. 2000; Steadman et al. 2000; Thomas and Leese 2003), violent offenders on probation (Stalans et al. 2004), juvenile sex offenders (Van der Put et al. 2013) and ex-detainees (Yang et al. 2010). As such, the current study utilizes a single classification tree approach among a sample of former high-risk juvenile delinquents to answer the following research question: From a multitude of youth criminogenic risk factors, which constellations of risk factors are most strongly associated with violent criminal behavior in early adulthood? In answering this question, this study contributes to the literature on prediction methods for violent crime by using a classification tree method that has not been extensively used for violence prediction.


Study design and sampling procedure

A case–control study was performed that was centered around a Dutch cohort of young adult males born between 1985 and 1993 with a known history of juvenile probation in Amsterdam (n = 2300). Within this cohort of former juvenile delinquents, two groups were distinguished: those who persisted and escalated into committing violent offenses and those who committed no violent offenses and as few as possible nonviolent offenses. Childhood characteristics were compared between these groups to identify determinants of violent offending after transitioning into early adulthood.

Childhood characteristics were extracted from information stored in files from their juvenile probation period. These files contain, amongst others, psychological profiles and descriptions of the situation of the child and his context in the years preceding and during the juvenile probation period. From these files, a set of criminogenic risk factors was systematically extracted (see instruments).

Subjects were assigned to one of the two groups based on their offending information as retrieved from the local police registry. Because police data are stored for a limited number of years, the sampling period in early adulthood was limited both in calendar years (2007–2013) and age (18–27 years). For those born in 1985 (the oldest), follow-up was available for the age period 22–27. For those born in 1993 (the youngest), follow-up was available for their ages 18–19. To minimize misclassifications (i.e. incorrectly identifying someone as nonviolent offender and vice versa), the design of the research population maximized observation time while ensuring a minimum follow-up time of at least 2 years in early adulthood, starting before the age of 23.

The following exclusion criteria were applied. Subjects were excluded if they left Amsterdam within the follow-up period according to the municipal records database (n = 414). Participants were excluded if they could not be matched with the police registry (n = 413). Because the majority of young adults who were not found in the police registry could not be matched due to registration errors in their names and/or dates of birth, absent matches did not indicate those former juvenile delinquents who desisted from crime in early adulthood.

With the end of the observation period as a reference, applying this design yielded a population of 1473 young adult offenders with a known history of juvenile probation and a mean age of 22.5 years (SD = 2.38). The distinction between violent and nonviolent offenders was made between those with (n = 733) and without (n = 740) one or more young adult violent offenses in the police registry. No differences between the groups were observed for the number of years of follow-up and the average starting age of follow-up. From this group of 1473 young adult offenders, a study population was sampled. Their juvenile probation files were quantified. For violent offenders, sampling was random. For nonviolent offenders, those with the least nonviolent offenses were prioritized. Not all files were available due to logistic reasons or other reasons (n = 358). Files could be unavailable due to youth care involvement with younger siblings, ongoing juvenile justice procedures and school attendance officers who could consult with the juvenile probation agency up to age 23.


The Juvenile Forensic Profile (FPJ) was used to extract information from the files (Brand and Van Heerde 2010). This instrument was originally developed for forensic research using files of juveniles with a PIJ order (Placement in an Institution for Juveniles for mandatory treatment), the most severe measure in the Dutch juvenile justice system. The FPJ has good psychometric properties with an overall inter-rater reliability of r = 0.73 and Kappa = 0.61. Its convergent validity with the SAVRY (structured assessment of violence risk in youth; Meyers and Schmidt (2008), which aims to measure the same construct as the FPJ, is good with Kappa = 0.61. The FPJ also has good predictive validity (AUC = 0.803) (Brand 2005a, b; Van Heerde et al. 2004; Van Heerde and Mulder 2005).

The FPJ consists of around 70 risk factors, arranged in seven domains: history of criminal behavior, upbringing and environment, offense situation and drugs, psychological functions, psychopathology, social and relational, and behavior during stay in the institution. The majority of risk factors are measured on a three point scale with 0 = no risk, 1 = moderate risk and 2 = severe risk. Types of offenses committed during youth (property, violent and sexual offenses), out of home placement and committing under the influence of alcohol/drugs were yes/no measures. The FPJ domain behavior during stay in institution and the items number of (non)violent offenses were omitted due to insufficient information (juvenile probation files are not police or juvenile justice records). Some items were added to score information about the household composition, if offenses during youth were mainly committed alone or in a group and parental somatic problems. Table 2 presents all risk factors included in the study.

Scoring of childhood criminogenic risk factors

Two raters were trained in extracting information on risk factors from juvenile probation files. With these two trained raters supervising, a set of practice files was scored by six raters independently. After the threshold of 80% consensus in administered scores on these practice files was reached, files were allocated to single raters and every tenth file was scored by multiple raters. In a 1 year data-collection period as many files as possible were scored. Scoring of a typical file took 4 h on average, raters could score two files in a day.


In a 1-year data collection period, 387 from 1115 available files were quantified (34.7%). This sample consisted of 146 violent offenders (37.7%; base rate Pviolence = .38) and 241 nonviolent offenders (62.3%). On average, violent offenders had as young adults committed 16.4 (SD = 14.0) offenses of which 3.3 (SD = 2.00) were violent. Nonviolent offenders had on average committed 2.0 (SD = 1.3) exclusively nonviolent offenses (see Table 1).

Table 1 Overview of the mean number of (violent) offenses committed during early adulthood and specific type of offenses committed at least once by young adult violent and nonviolent offenders

Inter-rater reliability

Based on 42 multiple scored files that yielded a pool of 148 score forms, inter-rater reliability (IRR) measures were calculated. IRR was satisfactory with R = 0.79 (p < .001), ICC = 0.77 (CI95% 0.75–0.80, p < .001) and Kappa = 0.67 (p < .001). The proportion of raw agreement was 81.9%, of deviating scores between raters by one point was 16.5% and of deviation by two points was 1.6% (i.e. no vs. severe risk). The raw correlation (R) and intra-class correlation coefficients (ICC) were very strong (Cicchetti 1994). Kappa was substantial according to the Landis and Koch classification (Landis and Koch 1977).


Analyses were performed with the SPSS 21.0 statistical package. Chi square tests and one-way ANOVAs identified univariate differences between violent and nonviolent offenders on sociodemographic characteristics and all FPJ items. Chi square Automatic Interaction Detection (CHAID) analysis was performed to identify constellations of risk factors associated with violent offending. CHAID analysis is a classification tree analysis (CTA) to identify the most important explanatory variables for an outcome measure.

The analysis starts with the total sample in a “root node”. Continuous variables are transformed into categorical variables. Except for dichotomous predictors, the algorithm explores which pair of scores (categories) is least significant with respect to the outcome. This pair is merged. This step is repeated to find the next pair of categories. For dichotomous outcome variables, CHAID uses the Chi square test to identify the predictor variable that, with respect to the outcome, most significantly splits the root node into two or more homogeneous groups according to this predictor’s categories. The same steps are then applied for the subgroups, creating more branches within the sample. This recursive partitioning ends when no further splits improve within-partition homogeneity, or when a predetermined maximum tree-depth is reached. The final partitions are referred to as leaves, end nodes or terminal nodes, in which cases with the same scores on risk factors are grouped together. These terminal nodes represent higher-order interactions between predictor variables unlikely to have been discovered with regression analyses (Berk and Bleich 2014).

All statistically significant univariable differences (p < .05) between the two groups were included as predictor variables. The minimum number of cases in terminal-nodes was 20 (5% of the sample) and the significance level of p < .05 for splitting nodes was applied. Missing values were treated as valid scores and are represented in the final tree. A misclassification risk of the final tree was estimated using cross validation with tenfolds.

Pviolence indicates the proportion of violent offenders within a node. Each endnode’s Pviolence was set against the base rate of violent offending in the overall sample to distinguish between low and high risk of violent offending. A Pviolence twice the base rate was defined as high risk and a Pviolence twice as low the base rate was defined as low risk.


In 2013, at the end of the observation period, the mean age of all sampled offenders was 22.1 (SD = 2.1) and the mean observation time was 3.86 years (SD = 1.43). Although the design of the cohort and sample allowed for differences in ages at the start/end of the observation time, no differences between violent and nonviolent offenders regarding age and available observation time were observed. Ethnic composition of both groups differed (χ2 = 29.7, p < .001). In comparison with nonviolent offenders, nearly twice as many violent offenders had a North African migrant background (44.7% vs. 21.9%) and fewer had a native Dutch background (9.9% vs. 27.8%). Also, fewer violent offenders grew up with a foster parent (26.6% vs. 42.9%, χ2 = 8.0, p < .01), violent offenders had on average more siblings (3.32 vs. 2.32, t = − 4.41, p < .001) and more violent offenders had an older brother(s) than nonviolent offenders (53.8% vs. 36.3%, χ2 = 6.68, p < .01) (not in table). With respect to other FPJ items,Footnote 1 violent offenders were higher on risk on 28 of 63 FPJ items. This concerned, amongst others, 5 items from the ‘history of criminal behavior’ domain. For example, 98.6% of violent offenders had already as a juvenile committed a violent crime and 74.2% of the nonviolent offenders (Table 2). Nonviolent offenders were higher on risk on the FPJ item presenting with symptoms of depression only (see Table 2).

Table 2 Prevalence and severity of youth criminogenic risk factors among violent offenders and nonviolent offenders

CHAID analysis

The CHAID analysis yielded a decision tree with three levels (Fig. 1 and Table 3). From 5 sociodemographic characteristics and 29 FPJ-items that were found significantly different between the two groups, 6 factors were identified as (most) significant segmenting variables: involvement with criminal peers, offenses under the influence of drugs, offenses under the influence of alcohol, criminal family members, (symptoms of) depressive disorder and dysfunctional family. Cross validation of the model yielded a misclassification risk estimate of .24. The analysis correctly predicted 78% of all cases. The model had a sensitivity of 93% and a specificity of 55%. The AUC was .86 (CI95% .82–.90).

Fig. 1
figure 1

Graphical display of the results from the CHAID analysis with Pviolence (violent criminal behavior) as dependent variable. Pviolence estimates associated with the end nodes only have value as prevalence estimates of violent criminal behavior in comparison with each other

Table 3 Results from the CHAID analysis with violent criminal behavior as dependent variable

Chi squared tests identified involvement with criminal peers as most significant segmenting variable (i.e. χ2 = 107.3, p < .001) that divided the root node into three child nodes according to the level of severity of the risk factor (no/missing vs. moderate vs. severe). These nodes further branched into 8 endnodes (see Fig. 1 and Table 3). Additionally to Fig. 1, Table 3 also presents the distribution of missing values within each of the endnodes.

Nodes associated with a high probability of violent offending were endnodes 9 and 13. As juveniles, offenders in node 9 were severely involved with criminal peers and had moderate/severe criminal family members. Node 9 consisted of 69 offenders (18% of sample) and was associated with Pviolence= .80. Offenders in node 13 were moderately involved with criminal peers, had committed offenses under the influence of drugs, or this was unknown, and had a moderate/severe dysfunctional family. Node 13 consisted of 29 offenders (8% of sample) and was associated with Pviolence= .86.

Endnodes 4 and 11 were associated with a low risk of violent offending. Endnode 4 consisted of 96 offenders (25% of sample) who as juveniles were not involved with criminal peers and had not committed offenses under the influence of alcohol. Two offenders in this node were violent offenders during early adulthood (Pviolence= 0.02). Endnode 11 consisted of 38 offenders (10% of total sample) who as juveniles were moderately involved with criminal peers, had not committed offenses under the influence of drugs and presented with moderate/severe (symptoms of) depressive disorder. One offender in this node was a violent offender during early adulthood (Pviolence= 0.04).


This study aimed to identify constellations of childhood risk factors most strongly associated with violent criminal behavior in early adulthood in a sample of former high-risk male juvenile delinquents. Eight constellations of childhood risk factors were recognized, of which two were associated with the highest risk of violent offending. The first consisted of offenders who as juveniles had been moderately involved with criminal peers, who had committed offenses under the influence of drugs, and who had grown up in a dysfunctional family. The second was characterized by severe involvement with criminal peers and having criminal family members in childhood. Having presented with (symptoms of) depressive disorder as a juvenile was associated with a lower risk of violent offending.

The sample was drawn from a larger population of young adult male offenders with a history of juvenile probation. Within an average observation time of nearly 4 years during early adulthood, half of all offenders were violent offenders. They were recognized in the police registry on account of at least one violent crime committed in early adulthood. Besides the dichotomy of violent and nonviolent offending, violent offenders had on average also committed eight times more offenses than nonviolent offenders.

The sample was segmented into eight constellations marked by the presence and/or severity of specific childhood risk factors. These constellations represent higher-order interactions between predictor variables that most likely would not have been detected with logistic regression analysis. Together, the constellations constituted a model that predicted violent offending well. From 34 childhood risk factors that differed significantly in terms of prevalence and/or severity between violent and nonviolent offenders, six were in one or the other combination most strongly associated with violent offending. Five converge with the ‘central eight’ criminogenic needs known from the literature (Andrews and Bonta 2010). These were involvement with criminal/antisocial peers (Fergusson et al. 2007; Harder et al. 2015), having committed under the influence of alcohol/drugs (Bonta and Andrews 2007), having criminal family members (Geller et al. 2009; Murray and Farrington 2008) and growing up in a dysfunctional family (Chambers et al. 2001; Gorman-Smith et al. 1998).

Set against the base rate of violent offending in early adulthood (38%), that should be interpreted as a sampling result, two of eight constellations of childhood risk factors were associated with a high risk of violent offending. Together, these comprised 25.3% of the total sample. The first, characterized by moderate involvement with criminal peers and having committed offenses under the influence of drugs (including unknown) and growing up in a dysfunctional family, had a 86% probability of violent offending. Using the FPJ, the risk factor dysfunctional family was assessed based on specific indicators mentioned in their juvenile probation files, among which: severe conflicts between siblings, chaotic household, bad/unhealthy nutrition, parentification, disrupted child-parent hierarchy and stranger in the role of head of the family. Based on these indicators, the risk factor dysfunctional family corresponds with the ‘adverse childhood experiences’ construct (ACE; Felitti et al. 1998), specifically to its constituent categories parental mental health problems, parental substance abuse, and household violence. Concerning the offenders in this constellation, it could be hypothesized that their ability to form secure attachments to others and the attribution of benevolent motives to others had corroded (Murphy et al. 2014). Drug use, at least occasionally, may have lowered their threshold to engage in criminal behavior (Andrews and Bonta 2010) and the use of violence possibly serves as a means of self-preservation (Polaschek et al. 2009). For juvenile delinquents who do not strongly affiliate with criminal peers, growing up in a dysfunctional family and offending under the influence of drugs were distinctive for developing young adult violent criminal behavior. For diversion efforts to be effective, the early lives of these offenders should not be ignored. Purely ‘offense specific’ interventions seem unsuited (Perez et al. 2018; Reavis et al. 2013).

The second high-risk constellation, marked by severe involvement with antisocial/criminal peers and having criminal family members, had a 80% probability of violent offending. Besides the influence of criminal members in their social network, no individual risk-factors were present in this high-risk constellation. This suggests that this constellation reflects social learning mechanisms that are well known from the criminological literature, particularly from research into the intergenerational transference of criminal behavior (Farrington 2002) and to a lesser extent also of violent crime (Besemer et al. 2017; Eichelsheim and van de Weijer 2018). From a clinical perspective this is a particularly tough target group because diversion by influencing social networks of marginalized delinquents is notoriously difficult.

Although not one of the central eight criminogenic risk factors, presenting with symptoms of depressive disorder was prominent in our model given its presence in two constellations with a low risk of violent offending. One of those constellations is noteworthy. It consisted of young adults who, as juveniles, had been moderately involved with criminal peers, who had not committed offenses under the influence of drugs and who had presented with depressive symptoms. Strikingly, violent offending in this constellation was much less prevalent than in the neighboring constellation (Fig. 1, node 10) that shared all risk factors except depressive disorder.

Although this effect of depressive (i.e. mood) disorders has previously been observed (Arseneault et al. 2000; Colins et al. 2010; Vermeiren et al. 2002), its exact relation with violent crime remains unclear. For example, a history of depression has been associated with less subsequent antisocial behavior, but only among boys (Oakley et al. 2009). Mood disorders in adolescence have been demonstrated to lower the probability of violent offending specifically (Oakley et al. 2009). This seems in line with our results, albeit that this effect has also been observed for other types of offending (Bevc et al. 2003). Besides observations, explanations have been proposed also. It has been theorized that depressive disorder should be more common among adolescence-limited than life-course persistent offenders, because depressive symptoms are a manifestation of having the inner potential to more strongly experience emotions such as guilt and shame (Vermeiren et al. 2002). Another explanation is that offenders with symptoms of depressive disorder are less prone to violence due to physical inactivity and low energy/fatigue associated with depression (American Psychiatric Association 2000). However, the exact relation between depressive disorder and physical inactivity, especially among adolescents, is unclear (Stavrakakis et al. 2012). Nevertheless, depressive disorder in childhood is associated with other negative outcomes in later life, among which psychopathology (McLeod et al. 2016; Thapar et al. 2012), and social problems (Copeland et al. 2015). As such, active, timely and responsive transference of juveniles with symptoms of depressive disorder and who age out of the juvenile justice/youth care system to adult services is warranted.

Other salient risk factors that were not part of the final model but differed between violent and nonviolent offenders also offer clues to the prediction of violent offending and may guide additional research initiatives. Although it is beyond scope to discuss all of these here, lack of empathy, impaired conscience development, development towards a personality disorder type B (i.e. antisocial) and negative affect were prominently associated with violent offending. These risk factors correspond with callous-unemotional traits (lack of guilt, absence of empathy and callous use of others), that mark subgroups of juvenile offenders with more severe and stable patterns of criminal behavior (Frick and White 2008) and which are important in the development of persistent and escalating criminal behavior (Carlson et al. 2015; Frick et al. 2014).

Strengths and weaknesses

Based on historical information stored in juvenile probation files we systematically retrieved the presence and severity of a comprehensive set of childhood risk factors. A high inter-rater reliability was attained by using a reliable and valid instrument and a group of well-trained raters. Objective police registry data offered a time window of nearly 4 years during early adulthood within which criminal behavior could be observed.

In composing the population from which violent and nonviolent offenders were sampled, a trade-off was made between utilizing as much observation time as possible (to detect violent offenses), minimizing variety in observation time and including as many participants as possible. Nevertheless, differences in observation times between participants could not be avoided. In a case–control design, this increases the likelihood of incorrectly classifying true cases as controls. However, a mean observation time of nearly 4 years was achieved that greatly exceeded the minimum observation time of 2 years. Importantly, no differences in observation time and starting age at follow-up were observed between sampled violent and nonviolent offenders. Also, these groups contrasted sharply with respect to both young adult offending behavior and a multitude of risk factors which suggests validity in our approach.

In our sample, violence went hand in hand with prolificness. As such, apart from the violence element, differences between both groups might also be driven by differences in prolificness. Controlling for overall differences in levels of criminal activity was not possible because separating prolificness from violence would yield too few ‘non-prolific but violent’ offenders to be able to perform our analyses. Put differently, correcting for prolificness would imply the elimination of violence as the primary outcome.

With a decision tree approach we identified two constellations of risk factors associated with a high risk of violent offending embedded in a model with good predictive accuracy. An important benefit of decision tree models over multiple regression models is that it provides a visualization of characteristics associated with an outcome that is easy to explain and intuitive to understand. Additionally, decision trees can reveal interactions between risk factors that are easily overlooked (Thomas and Leese 2003). As such, it provides a useful means to generate hypotheses about interacting variables associated with violent criminal behavior (Steadman et al. 2000). We consider our model’s good predictive accuracy a promising result, given the combination of a relatively small sample and many predictor variables included in the analysis. However, its specificity (55%) was less convincing. An implication of applying our model would be that a group of juvenile delinquents may receive more diversion efforts than needed regarding the prevention of violent offending. This might seem a lesser problem than the opposite situation where too few preventive efforts would be deployed for those who actually do need it. However, from the risk-need-responsivity model (Andrews and Bonta 2010) we know that any mismatch between the intensity of an intervention and the risk of violence/recidivism can yield opposite effects (Andrews 1995; Lowenkamp and Latessa 2005).

Of course, CHAID has some weaknesses. A disadvantage of classification trees in general is the volatility of their outcomes. Small changes in input information may lead to large changes in the tree it constructs. Also, chance interactions can take up a prominent place in the decision tree it produces and not all possible trees are examined. Together, these weaknesses of CHAID influence the generalizability of its results and, consequently, the utility of the models it finds (Thomas and Leese 2003). Random forests (Breiman 2001), an elaboration on CTA, was developed to stabilize decision trees. Random forests (RF) is an algorithm that randomly selects sets of predictor variables from a larger pool of variables to build hundreds of individual decision trees that are averaged in a single outcome tree (Ritter 2013). RF has been shown to outperform logistic regression in risk prediction (Berk and Bleich 2014; Bushway 2013) when applied to large actuarial databases. Nonetheless, RF has the disadvantage of being a ‘black box’ to the end user (Zeng et al. 2017). Although its input and output can be seen, their users are blind to how one’s risk estimate is derived. Additionally, RF is difficult to implement well (Bushway 2013) and there are ethical dilemma’s around the use of computer algorithms such as RF as predictive tools to guide criminal justice decisions (Chan and Bennett Moses 2016).

From a variety of single classification tree approaches, the Rpart algorithm in R for recursive partitioning (Therneau and Atkinson 2006) is a more contemporary technique than CHAID. Two features of Rpart made it less suited to our data, however. The first is that Rpart allows for binary splits only, whereas most of our predictor variables had three levels. Second, for cases with a missing value on a splitting variable, Rpart uses the ‘best’ surrogate variable (e.g., bodyweight for body length) to be able to classify cases to the next node. In contrast, CHAID treats missing values as a discrete category that may or may be not merged with other score categories, which more closely resembles daily practice. Beyond the decision tree approach, latent class analysis (LCA) was an alternative analytical strategy to identify combinations of risk factors associated with violent criminal behavior. LCA was unsuited to our data, however, given the high number of variables with missing values (Table 2) in combination with the default option of LCA to eliminate cases with 1 or more missing values. Imputation of missing values was unsuited as a fallback strategy, because other variables provided insufficient substantive support with respect to variables that required imputation. A traditional classification tree as we used is well suited for dealing with missing values, as was frequently the case in our data.

Our results should be validated before making claims about its generalizability. Important to note is that the Pviolence estimates of the constellations in our model are not to be generalized to juvenile offender populations. They only have a relative value as prevalence estimates of violent offending in comparison with each other, because the base rate of violent offending was a result of our sampling procedure. To validate our findings it would be preferable to perform a decision tree approach in a prospective cohort study. Nonetheless, five of the six risk factors in our model trace back to the central eight criminogenic needs known from the literature (Andrews and Bonta 2010). Noteworthy, our model resembles that of Ortega-Campos et al. (2016), who used CHAID analysis to predict recidivist sanctionable antisocial behavior among Spanish juveniles charged in a court case. As did we, Ortega-Campos et al. (2016) found involvement with antisocial peers to be a key segmenting factor, experiencing mental health problems to be associated with a lower, and having criminal family members with a higher probability of recidivism.

Finally, despite being one of CHAID’s benefits, missing information adds complexity to the interpretation of our model. The CHAID analysis treated missing scores as valid scores. For each risk factor involved, the analysis explored all possible arrangements to determine under which particular grouping of scores (missing/none/moderate/severe) the strongest predictive power regarding violent offending was observed. That information was used to merge categories and, ultimately, to construct the model. To illustrate, the predictive power of depressive disorder was strongest when those scoring no or missing were grouped together and contrasted with those scoring moderate/severe. Again, the model does reflect reality in which the presence or absence of a risk factor may also be unknown. For certain risk factors, missing information is equally likely to indicate that a problem was simply not observed or that a juvenile probation officer was unable to ascertain its presence at that time. As such, our model mimics reality because juvenile probation officers on occasion can neither affirm nor exclude the presence of a particular risk factor. In this respect, a benefit of decision trees over regression based models is that, for example, juvenile probation workers are not hindered by missing information in estimating one’s risk.

Nevertheless, we performed sensitivity analyses with alternating coding strategies for missing values of the two most important risk factors, as based on their associated χ2’s (i.e. involvement with criminal peers and offenses under the influence of drugs) (see Fig. 1). In the original model, missing values on involvement with criminal peers were merged with the no category. We first placed missing values in the moderate or severe involvement with criminal peers category, which remained the most important segmenting variable and χ2 only slightly decreased. Minor changes were only observed when missing values were considered moderate. In another sensitivity analysis, we recoded missing values on offenses under the influence of drugs and offenses under the influence of alcohol as no, given frequent missing values on these variables. This resulted in a marginal decline in their χ2. Taking these observations into consideration, coupled with the resulting segmentation of our model suggests robustness in our approach.


Among a group of former high risk juvenile delinquents, a large set of stand-alone childhood risk factors was reduced in number using CHAID analysis to a more compact and manageable set that constituted a model with sufficient predictive accuracy regarding young adult violent offending. We identified eight constellations of risk factors associated with a certain probability of violent offending. The model comprised, among others, two constellations depicting juvenile delinquents at highest risk of violent offending in early adulthood. Juvenile probation officers should be vigilant on youngsters’ involvement with criminal peers, whether or not they commit offenses under the influence of drugs and alcohol, the presence of criminal family members and any indication for growing up in a dysfunctional family. Presenting with symptoms of depressive disorder was associated with a lower probability of violent offending. Our results may be helpful in the selection and allocation of prevention and diversion programs for juvenile probation youth.

Availability of data

Data will not be shared. Many variables in combination with a relatively small research population may expose the identity of participants.


  1. Information based on which FPJ-items were scored was extracted from offenders’ juvenile probation files. On average, offenders were 13.98 years old (SD = 2.71; range 1.44–17.04 years) when they first came in contact with the youth care system. Those who had been < 12 years old during their initial contact with the youth care system had received youth protection first which, in our sample, was always followed by juvenile probation. Offenders had been under guidance of youth care/juvenile probation for a total time of 3.66 years (SD = 3.17), on average. At entry in the youth care system, (family) anamnesis is a standard intake procedure. Additionally, many offenders had been subjected to psychiatric/psychological inquiry and screenings, in which extensive (family) anamnesis is also standard. These anamneses provided information about circumstances and (adverse) events in early childhood, also for those juveniles who had entered the youth care system after 12 years of age. To summarize, a cumulative measure of the risk factors that were present somewhere during their childhood years is provided at the average age of closing the file (13.98 + 3.66 = 17.64 years).



analysis of variance


area under the curve


chi square automatic interaction detection


classification tree analysis


juvenile forensic profile


intraclass correlation coefficient


inter-rater reliability

Pviolence :

probability of violent criminal behavior in early adulthood


order placement in an institution for juveniles for mandatory treatment


random forests


  • Ahonen, L., Loeber, R., & Pardini, D. (2016). The prediction of young homicide and violent offenders. Justice Quarterly,33(7), 1265–1291.

    Article  Google Scholar 

  • American Psychiatric Association. (2000). DSM-IV-TR: Diagnostic and statistical manual of mental disorders-IV-text revision (4th ed., Vol. 75). Washington, DC: American Psychiatric Association.

    Google Scholar 

  • Andrews, D. A. (1995). The psychology of criminal conduct and effective treatment. In J. McGuire (Ed.), What works: Reducing offending–guidelines from research and practice (pp. 35–62). New York: Wiley.

    Google Scholar 

  • Andrews, D. A., & Bonta, J. (2010). The psychology of criminal conduct (5th ed.). New Providence: Routledge.

    Google Scholar 

  • Arseneault, L., Moffitt, T. E., Caspi, A., Taylor, P. J., & Silva, P. A. (2000). Mental disorders and violence in a total birth cohort: Results from the Dunedin Study. Archives of General Psychiatry,57(10), 979–986.

    Article  Google Scholar 

  • Baglivio, M. T., Jackowski, K., Greenwald, M. A., & Howell, J. C. (2014). Serious, violent, and chronic juvenile offenders. Criminology & Public Policy,13(1), 83–116.

    Article  Google Scholar 

  • Berk, R., & Bleich, J. (2014). Forecasts of violence to inform sentencing decisions. Journal of Quantitative Criminology,30(1), 79–96.

    Article  Google Scholar 

  • Besemer, S., Ahmad, S. I., Hinshaw, S. P., & Farrington, D. P. (2017). A systematic review and meta-analysis of the intergenerational transmission of criminal behavior. Aggression and Violent Behavior,37, 161–178.

    Article  Google Scholar 

  • Bevc, I., Duchesne, T., Rosenthal, J., Rossman, L., Theodor, F., & Sowa, E. (2003). Young offenders’ diagnoses as predictors of subsequent adult criminal behaviour. Paper presented at the the 111th Convention of the American Psychological Association, Toronto, Canada.

  • Bonta, J., & Andrews, D. A. (2007). Risk-need-responsivity model for offender assessment and rehabilitation. Rehabilitation,6(1), 1–22.

    Google Scholar 

  • Brand, E. F. (2005a). Onderzoeksrapport PIJ-dossiers 2003C: Predictieve validiteit van de FPJ-lijst. The Hague: Department of Safety and Justice, National Agency of Correctional Institutions.

    Google Scholar 

  • Brand, E. F. (2005b). Een maat voor de kwaliteit van instrumenten voor risicotaxatie. In M. C. v. V. Sjerps, JA (Ed.), Het onzekere bewijs. Gebruik van statistiek en kansrekening in het strafrecht. Deventer: Kluwer.

  • Brand, E. F., & Van Heerde, W. K. (2010). Handleiding Forensisch Profiel justitiële Jeugdigen (FPJ-lijst). The Hague: Department of Safety and Justice, National Agency of Correctional Institutions.

    Google Scholar 

  • Breiman, L. (2001). Random forests. Machine Learning,45(1), 5–32.

    Article  Google Scholar 

  • Bushway, S. D. (2013). Is there any logic to using logit. Criminology & Public Policy,12(3), 563–567.

    Article  Google Scholar 

  • 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(2), 129–153.

    Article  Google Scholar 

  • Carlson, M., Oshri, A., & Kwon, J. (2015). Child maltreatment and risk behaviors: The roles of callous/unemotional traits and conscientiousness. Child Abuse and Neglect,50, 234–243.

    Article  Google Scholar 

  • Chambers, J., Power, K., Loucks, N., & Swanson, V. (2001). The interaction of perceived maternal and paternal parenting styles and their relation with the psychological distress and offending characteristics of incarcerated young offenders. Journal of Adolescence,24(2), 209–227.

    Article  Google Scholar 

  • Chan, J., & Bennett Moses, L. (2016). Is Big Data Challenging Criminology? Theoretical Criminology,20(1), 21–39.

    Article  Google Scholar 

  • Cicchetti, D. V. (1994). Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment,6(4), 284–290.

    Article  Google Scholar 

  • Colins, O., Vermeiren, R., Vreugdenhil, C., Van den Brink, W., Doreleijers, T., & Broekaert, E. (2010). Psychiatric disorders in detained male adolescents: A systematic literature review. Canadian Journal of Psychiatry. Revue Canadienne de Psychiatrie,55(4), 255–263.

    Article  Google Scholar 

  • Copeland, W. E., Wolke, D., Shanahan, L., & Costello, E. J. (2015). Adult functional outcomes of common childhood psychiatric problems: A prospective, longitudinal study. JAMA psychiatry,72(9), 892–899.

    Article  Google Scholar 

  • DeLisi, M., & Piquero, A. R. (2011). New frontiers in criminal careers research, 2000–2011: A state-of-the-art review. Journal of Criminal Justice,39(4), 289–301.

    Article  Google Scholar 

  • Eichelsheim, V. I., & van de Weijer, S. G. (2018). Intergenerational continuity of criminal and antisocial behaviour: An international overview of studies. Abingdon: Routledge.

    Book  Google Scholar 

  • Falk, Ö., Wallinius, M., Lundström, S., Frisell, T., Anckarsäter, H., & Kerekes, N. (2014). The 1% of the population accountable for 63% of all violent crime convictions. Social Psychiatry and Psychiatric Epidemiology,49(4), 559–571.

    Article  Google Scholar 

  • Farrington, D. P. (2002). Families and crime. In J. Q. Wilson & J. Petersilia (Eds.), Crime: Public policies for crime control (pp. 129–148). Oakland: Institute for Contemporary Studies Press.

    Google Scholar 

  • Farrington, D. P., Coid, J. W., Harnett, L. M., Jolliffe, D., Soteriou, N., Turner, R. E., et al. (2006). Criminal careers up to age 50 and life success up to age 48: New findings from the Cambridge Study in Delinquent Development (p. 299). London: Home Office Research Study.

    Google Scholar 

  • Farrington, D. P., Ttofi, M. M., & Coid, J. W. (2009). Development of adolescence-limited, late-onset, and persistent offenders from age 8 to age 48. Aggressive Behavior,35(2), 150–163.

    Article  Google Scholar 

  • Felitti, V., Anda, R., Nordenberg, D., Williamson, D., Spitz, A., Edwards, V., et al. (1998). Relationship of childhood abuse and household dysfunction to many of the leading causes of death in adults: The Adverse Childhood Experiences (ACE) Study. American Journal of Preventive Medicine, 14(4), 245–258.

    Article  Google Scholar 

  • Fergusson, D. M., Vitaro, F., Wanner, B., & Brendgen, M. (2007). Protective and compensatory factors mitigating the influence of deviant friends on delinquent behaviours during early adolescence. Journal of Adolescence,30(1), 33–50.

    Article  Google Scholar 

  • Frick, P. J., Ray, J. V., Thornton, L. C., & Kahn, R. E. (2014). Can callous-unemotional traits enhance the understanding, diagnosis, and treatment of serious conduct problems in children and adolescents? A comprehensive review. Psychological Bulletin,140(1), 1–57.

    Article  Google Scholar 

  • Frick, P. J., & White, S. F. (2008). Research review: The importance of callous-unemotional traits for developmental models of aggressive and antisocial behavior. Journal of Child Psychology and Psychiatry,49(4), 359–375.

    Article  Google Scholar 

  • Geller, A., Garfinkel, I., Cooper, C. E., & Mincy, R. B. (2009). Parental incarceration and child well-being: Implications for urban families. Social Science Quarterly,90(5), 1186–1202.

    Article  Google Scholar 

  • Gorman-Smith, D., Tolan, P. H., Loeber, R., & Henry, D. B. (1998). Relation of family problems to patterns of delinquent involvement among urban youth. Journal of Abnormal Child Psychology,26(5), 319–333.

    Article  Google Scholar 

  • Harder, A. T., Knorth, E. J., & Kalverboer, M. E. (2015). Risky or needy? Dynamic risk factors and delinquent behavior of adolescents in secure residential youth care. International Journal of Offender Therapy and Comparative Criminology,59(10), 1047–1065.

    Article  Google Scholar 

  • Hein, S., Barbot, B., Square, A., Chapman, J., Geib, C. F., & Grigorenko, E. L. (2017). Violent offending among juveniles: A 7-year longitudinal study of recidivism, desistance, and associations with mental health. Law and Human Behavior,41(3), 273.

    Article  Google Scholar 

  • Landis, R. J., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics,33(1), 159–174.

    Article  Google Scholar 

  • Laub, J. H. (2004). The life course of criminology in the United States: The American Society of Criminology 2003 presidential address. Criminology,42(1), 1–26.

    Article  Google Scholar 

  • Lipsey, M. W., & Derzon, J. H. (1998). Predictors of violent or serious delinquency in adolescence and early adulthood: A synthesis of longitudinal research.

  • Loeber, R., & Farrington, D. P. (2011). Young homicide and victims: Development, risk factors and prediction from childhood. Boston: Springer.

    Book  Google Scholar 

  • Lowenkamp, C. T., & Latessa, E. J. (2005). Increasing the effectiveness of correctional programming through the risk principle: Identifying offenders for residential placement. Criminology & Public Policy,4(2), 263–290.

    Article  Google Scholar 

  • McLeod, G. F., Horwood, L. J., & Fergusson, D. M. (2016). Adolescent depression, adult mental health and psychosocial outcomes at 30 and 35 years. Psychological Medicine,46(7), 1401–1412.

    Article  Google Scholar 

  • Meyers, J. R., & Schmidt, F. (2008). Predictive validity of the Structured Assessment for Violence Risk in Youth (SAVRY) with juvenile offenders. Criminal Justice and Behavior,35(3), 344–355.

    Article  Google Scholar 

  • Moffitt, T. E. (1993). Adolescence-limited and life-course-persistent antisocial behavior: A developmental taxonomy. Psychological Review,100(4), 674–701.

    Article  Google Scholar 

  • Monahan, J., Steadman, H. J., Robbins, P. C., Silver, E., Appelbaum, P. S., Grisso, T., et al. (2000). Developing a clinically useful actuarial tool for assessing violence risk. British Journal of Psychiatry,176(4), 312–319.

    Article  Google Scholar 

  • Murphy, A., Steele, M., Dube, S. R., Bate, J., Bonuck, K., Meissner, P., et al. (2014). Adverse childhood experiences (ACEs) questionnaire and adult attachment interview (AAI): Implications for parent child relationships. Child Abuse and Neglect,38(2), 224–233.

    Article  Google Scholar 

  • Murray, J., & Farrington, D. P. (2008). The effects of parental imprisonment on children. Crime and Justice,37(1), 133–206.

    Article  Google Scholar 

  • Ngo, F. T., Govindu, R., & Agarwal, A. (2015). Assessing the predictive utility of logistic regression, classification and regression tree, Chi squared automatic interaction detection, and neural network models in predicting inmate misconduct. American Journal of Criminal Justice,40(1), 47–74.

    Article  Google Scholar 

  • Oakley, C., Hynes, F., & Clark, T. (2009). Mood disorders and violence: A new focus. Advances in Psychiatric Treatment,15(4), 263–270.

    Article  Google Scholar 

  • Onifade, E., Davidson, W., Livsey, S., Turke, G., Horton, C., Malinowski, J., et al. (2008). Risk assessment: Identifying patterns of risk in young offenders with the Youth Level of Service/Case Management Inventory. Journal of Criminal Justice,36(2), 165–173.

    Article  Google Scholar 

  • Ortega-Campos, E., García-García, J., Gil-Fenoy, M. J., & Zaldívar-Basurto, F. (2016). Identifying risk and protective factors in recidivist juvenile offenders: A decision tree approach. PLoS ONE,11(9), e0160423.

    Article  Google Scholar 

  • Perez, N. M., Jennings, W. G., & Baglivio, M. T. (2018). A path to serious, violent, chronic delinquency: The harmful aftermath of adverse childhood experiences. Crime & Delinquency,64(1), 3–25.

    Article  Google Scholar 

  • Polaschek, D. L., Calvert, S. W., & Gannon, T. A. (2009). Linking violent thinking: Implicit theory-based research with violent offenders. Journal of Interpersonal Violence,24(1), 75–96.

    Article  Google Scholar 

  • Reavis, J. A., Looman, J., Franco, K. A., & Rojas, B. (2013). Adverse childhood experiences and adult criminality: How long must we live before we possess our own lives? The Permanente Journal,17(2), 44–48.

    Article  Google Scholar 

  • Ritter, N. (2013). Predicting recidivism risk: New tool in Philadelphia shows great promise. National Institute of Justice Journal,271(February), 4–13.

    Google Scholar 

  • Snyder, H. N., & Sickmund, M. (2006). Juvenile offenders and victims: 2006 national report. Washington, DC: Department of Justice.

    Google Scholar 

  • Stalans, L. J., Yarnold, P. R., Seng, M., Olson, D. E., & Repp, M. (2004). Identifying three types of violent offenders and predicting violent recidivism while on probation: A classification tree analysis. Law and Human Behavior,28(3), 253–271.

    Article  Google Scholar 

  • Stavrakakis, N., de Jonge, P., Ormel, J., & Oldehinkel, A. J. (2012). Bidirectional prospective associations between physical activity and depressive symptoms. The TRAILS study. Journal of Adolescent Health,50(5), 503–508.

    Article  Google Scholar 

  • Steadman, H. J., Silver, E., Monahan, J., Appelbaum, P. S., Clark Robbins, P., Mulvey, E. P., et al. (2000). A classification tree approach to the development of actuarial violence risk assessment tools. Law and Human Behavior,24(1), 83–100.

    Article  Google Scholar 

  • Thapar, A., Collishaw, S., Pine, D. S., & Thapar, A. K. (2012). Depression in adolescence. The Lancet,379(9820), 1056–1067.

    Article  Google Scholar 

  • Therneau, T., & Atkinson, B. (2006). R port by Ripley B. 2006. rpart: Recursive partitioning. R package version, 3.1-33.

  • Thomas, S., & Leese, M. (2003). A green-fingered approach can improve the clinical utility of violence risk assessment tools. Criminal Behaviour and Mental Health,13(3), 153–158.

    Article  Google Scholar 

  • Ttofi, M. M., Farrington, D. P., Piquero, A. R., & DeLisi, M. (2016). Protective factors against offending and violence: Results from prospective longitudinal studies. Journal of Criminal Justice,45, 1–3.

    Article  Google Scholar 

  • Van der Put, C. E., Van Vugt, E. S., Stams, G. J. J., Deković, M., & Van der Laan, P. H. (2013). Differences in the prevalence and impact of risk factors for general recidivism between different types of juveniles who have committed sexual offenses (JSOs) and juveniles who have committed nonsexual offenses (NSOs). Sexual Abuse,25(1), 41–68.

    Article  Google Scholar 

  • Van Heerde, W. K., Brand, E. F., Van’t Hoff, G., & Mulder, E. A. (2004). Interrater-reliability and convergent validity of the juvenile forensic profile. The Hague: Department of Safety and Justice, National Agency of Correctional Institutions.

    Google Scholar 

  • van Heerde, W. K., & Mulder, E. A. (2005). Research study PIJ-files 2003-B. Clinical evaluation FPJ-list & SAVRY in Eikensteinand and FORA. The Hague: Department of Justice, National Agency of Correctional Institutions.

    Google Scholar 

  • Vermeiren, R., Schwab-Stone, M., Ruchkin, V., De Clippele, A., & Deboutte, D. (2002). Predicting recidivism in delinquent adolescents from psychological and psychiatric assessment. Comprehensive Psychiatry,43(2), 142–149.

    Article  Google Scholar 

  • Yang, M., Liu, Y., & Coid, J. (2010). Applying neural networks and classification tree models to the classification of serious offenders and the prediction of recidivism. Research Summary, Ministry of Justice, UK. Retrieved from

  • Zeng, J., Ustun, B., & Rudin, C. (2017). Interpretable classification models for recidivism prediction. Journal of the Royal Statistical Society: Series A (Statistics in Society),180(3), 689–722.

    Article  Google Scholar 

Download references


The authors would like to thank Peter van der Ven for his advice in performing the CHAID analyses.


The research was funded by the municipality of Amsterdam. The funder was not involved in the design of the study nor the drafting of the manuscript.

Author information

Authors and Affiliations



MS and TF are the principal investigators and obtained funding for the study. MS and TF coordinated the data-collection. MS, together with MW performed the data-analysis. MS drafted the manuscript with important contributions from TF, MW and AP. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Menno Segeren.

Ethics declarations

Ethics approval and consent to participate

The Medical Ethics Review Committee of the Academic Medical Center Amsterdan ordered that the Medical Research Involving Human Subject Acts (WMO) does not apply to the study and granted a waiver of consent. The Dutch ministry of Justice and the Dutch council of attorneys general from the Dutch public prosecutor’s office approved consultation of the police registry granted permission to retrieve offense information from the registry of the local police on the condition that privacy of participants was guaranteed.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Segeren, M., Fassaert, T., de Wit, M. et al. Constellations of youth criminogenic factors associated with young adult violent criminal behavior. Crime Sci 9, 2 (2020).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI:


  • Violent crime
  • Juvenile offenders
  • Criminogenic risk factors
  • Decision tree analysis
  • Early adulthood