Toward a Classification of Juvenile Offenders: Subgroups of Serious Juvenile Offenders and Severity of Recidivism

Abstract

The aim of this study was to identify subgroups of serious juvenile offenders on the basis of their risk profiles, using a data-driven approach. The sample consists of 1,147 of the top 5% most serious juvenile offenders in the Netherlands. A part of the sample, 728 juvenile offenders who had been released from the institution for at least 2 years, was included in analyses on recidivism and the prediction of recidivism. Six subgroups of serious juvenile offenders were identified with cluster analysis on the basis of their scores on 70 static and dynamic risk factors: Cluster 1, antisocial identity; Cluster 2, frequent offenders; Cluster 3, flat profile; Cluster 4, sexual problems and weak social identity; Cluster 5, sexual problems; and Cluster 6, problematic family background. Clusters 4 and 5 are the most serious offenders before treatment, committing mainly sex offences. However, they have significantly lower rates of recidivism than the other four groups. For each of the six clusters, a unique set of risk factors was found to predict severity of recidivism. The results suggest that intervention should aim at different risk factors for each subgroup.

Keywords

juvenile offenders subgroups risk profile severity recidivism

Introduction

Serious juvenile offenders are a great burden on society and are therefore an important target for intervention. The aim of intervention is to keep these serious offenders from continuing their careers into adulthood, with their behaviour getting a life course–persistent character (Moffitt, Caspi, Harrington, & Milne, 2002). In previous research we found several static and dynamic risk factors that were associated with recidivism and severity of recidivism in a large sample of the top 5% most serious juvenile offenders in the Netherlands (n = 728; Mulder, Brand, Bullens, & van Marle, 2011). The results showed that antisocial behaviour during treatment, lack of problem-solving strategies, family problems (lack of parenting skills, criminal behaviour in the family), and offence history (number of convictions, having one or more unknown victims of past offences) predicted recidivism. These findings are in line with international studies on serious juvenile delinquency in which the following risk factors for persistence of offending were found: early age of onset, violent behaviour in the past, genetic disposition, conduct disorder, ADHD (Clingempeel & Henggeler, 2003; Cottle, Lee, & Heilbrun, 2001; Loeber & Farrington, 2001; Vermeiren, 2003), psychopathic personality features (Booker Loper, Hoffschmidt, & Ash, 2001; Johnson, van Voorhis, & Ritchey, 2007), neurocognitive problems, temperament and behaviour problems, inadequate parenting (Moffitt & Caspi, 2001; Raine, Moffitt, Caspi, Loeber, Stouthamer-Loeber, & Lynam, 2005), low family acceptance and low academic achievement (Vermeiren, Bogaerts, Ruchkin, Deboutte, & Schwab-Stone, 2004), living in a bad neighborhood (Oberwittler, 2004), and substance abuse (Dawkins, 1997; Potter & Jenson, 2003).

According to the What Works Principle, effective interventions should aim at the needs of juvenile offenders (needs principle) and treatment intensity should depend on the level of risk (risk principle) (Andrews & Bonta, 1995). If we know which risk factors predict recidivism and especially severity of recidivism, we know which risk factors should be targeted in the first place during treatment. However, the question is (a) Should we consider very serious juvenile offenders as one homogeneous group and offer the same intervention to all serious juvenile offenders? and (b) Are risk factors for recidivism equal for the whole group? And if not: Do serious juvenile offenders actually consist of distinct subgroups with differences in risk profiles and hence different treatment needs?

Theory on the development of criminal behaviour suggests that different pathways exist toward serious juvenile offending, each with its own characteristics (Loeber, Farrington, Stouthamer-Loeber, & Raskin White, 2008). Moffitt et al. (2002) also identified different types of offending behaviour in juveniles, with different prognoses for adulthood. These theories suggest that different types of serious juvenile offenders can be identified. The results of previous research support this notion: For instance, rates of recidivism differ between subgroups of offenders. In previous research, lower rates for sexual recidivism were found than for nonsexual recidivism, both in juvenile sex offenders and in other types of juvenile offenders (Caldwell, 2002; Nisbet, Wilson, & Smallbone, 2004; Prentky, Harris, Frizzell, & Righthand, 2000; Waite et al., 2005). Several studies found risk factors specific for recidivism in subgroups of offenders. For instance, in young sex offenders, poor social skills, sexual deviancy, prior sexual offences, victimized strangers, having had a younger victim, more than one victim and diverse sexual crimes were identified as risk factors (Långström & Grann, 2000; Miner, 2002; Worling & Curwen, 2000). Several general criminological factors also play a role in sexual recidivism: early onset of offending, total number of prior offences (both sex offences and other types of offences), and an antisocial lifestyle. Dynamic risk factors that are related to sexual reoffending are treatment adherence, problem insight, general psychological problems, and failure to complete treatment (Hendriks & Bijleveld, 2008; Kenny, Keogh, & Seidler, 2001; Worling, 2001).

Although we may hypothesize that subgroups do exist and that there are differences in risk factors, we do not know what the best classification of juvenile offenders is and what the differences in risk profiles actually are. In previous research, offenders have usually been classified on the basis of the type of offence they committed, for instance, property offenders, violent offenders, or sex offenders, sometimes in combination with another variable, such as substance use (Dembo & Schmeidler, 2003). However, we may find that a classification based on specific combinations of risk factors might be actually more adequate than a classification on the basis of the type of offences committed. Sex offenders, for instance, appear to be a heterogeneous group based on both offence characteristics and risk factors (Långström, Grann, & Lindblad, 2000; Parks & Bard, 2006; van Wijk, Mali, Bullens, & Vermeiren, 2007). This approach of classification based on risk factors is interesting, because it focuses on the combination of aspects that underlie problematic behaviour instead of the behaviour itself (symptomatic approach). There have been some studies in which subgroups were found with distinct combinations of risk factors, for instance, on the basis of neuropsychological characteristics or personality typologies (Stefurak, Calhoun, & Glaser, 2004; Teichner et al., 2000), but there has not been much research on risk factors for recidivism in these subgroups. Furthermore, there has been little research on comparing subgroups of offenders or on comparing their risk profiles (Onifade et al., 2008).

The aim of this study was to find an optimal classification of serious juvenile offenders on the basis of specific combinations of risk factors. Cluster analysis was used to search for subgroups in a sample of the top 5% most serious offenders in the Netherlands. We expect to find that serious juvenile offenders are a heterogeneous group, consisting of several subgroups with clearly distinct patterns of risk factors. Next we looked at differences in recidivism rates and at differences in risk factors that predict severity of recidivism. We are not only interested in reducing the rate of recidivism, but also in reducing the severity of recidivism (harm reduction; Marshall & McGuire, 2003; Laws, 1996). The outcome variable therefore is severity of recidivism. We expect that subgroups will differ in recidivism rates and risk factors that predict severity of recidivism, with severity being defined by the amount of harm, the type of offence, and the maximum penalty.

Method

Participants

Participants in this study were male adolescents aged 12 to 23 years sentenced to placement in a Dutch juvenile institution for mandatory treatment (PIJ order, Placement in an Institution for Juveniles; Lodewijks, Doreleijers, & de Ruiter, 2008) between 1995 and 2004 (n = 1,147). A mandatory treatment order can last from 2 to 6 years and is the most severe measure in the Dutch juvenile justice system, which is applied to juveniles between 12 and 18 years old (van der Linden, Ten Siethoff, & Zeijlstra-Rijpstra, 2003; Stevens & van Marle, 2003). Data on these serious juvenile offenders was used for cluster analysis to come to a classification in subgroups.

The criminal records of the participants were collected in June 2008. These records were used to register recidivism and severity of recidivism. Overall recidivism among all Dutch juvenile offenders in juvenile judicial institutions (both prisons and treatment facilities) is 70% within 4 years (Wartna, Harbachi, & van der Laan, 2005). Internationally, recidivism rates in serious juvenile offenders were found up to 80% (Trulson, Marquart, Mullings, & Caeti, 2005). Previous research shows that within 1 year after release most recidivists have already reoffended, but recidivism continues to rise in the first 5 to 8 years after release (Wartna et al., 2005). We started to register recidivism (dependent variable) from the moment the PIJ order officially was ended and we included only those juvenile offenders who had been released for at least 2 years. Eventually 728 participants were included in the analyses. The mean age at release from the treatment facility was 20 years (SD = 1.63). Time until first reoffence ranged from 0 to 8.08 years (M = 1.2, SD = 1.5). The minimum time at risk was 2 years and the mean time at risk in our study was 5.83 years (SD = 2.39, Mdn = 5.58 years, range = 2-11.17 years).

Instruments

Juvenile Forensic Profile (FPJ)

A list of 70 risk factors were assessed with the FPJ (Brand & van Heerde, 2004), an instrument that was especially developed for forensic research based on file data. This instrument was derived from existing internationally and nationally validated instruments for risk assessment and for measuring problem behaviour (e.g., Child Behaviour Check List, Structured Assessment of Violence Risk in Youth, Psychopathy Check List: Youth Version, Juvenile-Sex Offender Assessment Protocol, HCR-20 Violence Risk Assessment Scheme, Forensic Profiles–40). The list contains risk factors concerning seven domains: history of criminal behaviour, family and environment, offence-related risk factors and substance use, psychological factors, psychopathology, social behaviour/interpersonal relationships, and behaviour during stay in the institution (see appendix). Each risk factor is measured on a 3-point scale, with 0 = no problems, 1 = some problems, and 2 = severe problems. Previous research on the FPJ list showed that the available file information was thorough and complete enough to be able to score the instrument (van ’t Hoff, Brand, van Parijs, & van Heerde, 2002). The interrater reliability was tested (double scoring of 80 files, r = .73, K = 0.61) and a high convergent validity of the FPJ list with the SAVRY was found (van Heerde, Brand, van ’t Hoff, & Mulder, 2004). The predictive validity of the instrument was tested in the first sample of 102 boys (area under the curve [AUC] of .803, with a sum score of nine risk factors; Brand, 2005a). In sum, the psychometric qualities of the instrument were found to be satisfactory (Brand, 2005b; van Heerde & Mulder, 2005).

Classification of Recidivism

To measure recidivism, all convictions starting at release from the institution were registered, together with the date and type of the offence committed. The choice for official reconviction data is straightforward, but it does have some limitations. The most important limitation is that an unknown number of offences will get lost, as only those that have led to a conviction are counted. Using self-report would probably lead to higher recidivism rates than official records. Another limitation is the possible influence on reconviction of changes in policy (Friendship, Beech, & Browne, 2002). Despite these limitations, official reconviction data were used because they provide an objective and clear measure for recidivism (Heilbrun et al., 2000).

Severity of recidivism was operationalized by classifying the seriousness of the offences in 12 categories. Severity of offending was determined depending on the Dutch laws’ increasing maximum sentence, the amount of harm, and the amount of violence during the offence. The classification of severity was evaluated by Dutch clinicians and law professionals (van Kordelaar, 2002). Table 1 shows the operationalization of severity of offending behaviour (before treatment). The 12 categories of severity are mutually exclusive.

Table 1.

Operationalization of Offending Before Treatment, n = 728^a

Severity of offending (12 categories)	Valid percentage (per type of offence)
0 = No conviction	0.1
1 = Misdemeanor	11.7
2 = Drug offence	2.0
3 = Vandalism (property)	15.1
4 = Property offence	65.6
5 = Moderate violent offence	50.5
6 = Violent property offence	52.9
7 = Serious violent offence	15.3
8 = Sex offence, same age	16.8
9 = Pedosexual offence	7.3
10 = Manslaughter	11.3
11 = Arson	3.7
12 = Murder	3.7

In the second column, percentages are greater than 100% because most serious juvenile offenders are generalists and commit more than one type of offence.

Procedure

The study was approved by the Medical Ethical Commission of Erasmus University Medical Center in Rotterdam, the Netherlands. The information obtained on the participants was anonymous. After 1 year of treatment, the files were scored with the FPJ-list. We measured after this period to be able to include risk factors during treatment, such as treatment adherence and motivation. All files (n = 1,147) were read and scored with the FPJ list by master’s-prepared students (in psychology or criminology) who were in their last year before graduation. Before scoring the files, the students received 3 weeks of training, and after 3 weeks, the reliability of the scoring was tested. Reconviction data of juvenile offenders were delivered by the official registration center of the Ministry of Justice. The recidivism data include the details on all court appearances, the date and type of offence, and the date of conviction or acquittal. All convictions dated after release from the judicial juvenile institution were counted as recidivism.

Statistics

All statistics were calculated with SPSS 15.0 for Windows. In previous research, exploratory factor analysis was used to find a meaningful classification of risk items of the FPJ list (Mulder, Brand, Bullens & van Marle, 2010). This revealed a nine-factor structure in the data. Next, these nine factors were used as input for hierarchical cluster analysis, which was used to identify subgroups of offenders with specific combinations of risk factors. Cluster analysis is an exploratory multivariate procedure for detecting groupings in data that may be used with dichotomous or interval-level data. It seeks to classify cases in a way that maximizes differences between groups. Because of the large number of risk factors and the fact that our data were not longitudinal (recidivism data were indeed longitudinal, but these we used as an outcome measure) we decided to choose cluster analysis for our approach. In this study we clustered cases to find out how serious juvenile offenders may be classified in subgroups based on common patterns of risk factors. After the hierarchical cluster analysis, an iterative K-means clustering technique was used to identify cases. Split-half analyses were performed to validate the cluster solution we found. Post hoc comparisons between clusters were conducted for a final six-cluster model using analyses of variance (ANOVAs). Before performing a cluster analysis, multivariate outliers were removed, depending on Cook’s distance (>.0050) and Mahalanobis distance (>25.0). After removing outliers, 1,107 (of 1,147) subjects were included in the cluster analysis.

For recidivism, descriptive statistics and frequencies were calculated first. Because the mean score on most risk factors of the FPJ list in our sample was larger than 1, our data were skewed to the right. Therefore, we used the nonparametric Mann-Whitney U test to study the differences between juveniles who recidivated and those who did not; a p value <.05 was considered to be significant. We also used this test to study differences between violent and nonviolent recidivism. We used nonparametric correlation (Spearman’s rho) to study the relation between risk factors and seriousness of recidivism measured in 12 categories. In our analyses, we corrected for multiple comparisons with Bonferroni correction, because of the large number of risk factors (n = 70), which was used in analyses on recidivism. Linear regression analysis was used to test which risk factors predict severity of recidivism. Missing Values Analysis was performed to check if missing values were missing at random, which was the case. Missing values did not significantly influence the outcome of regression analysis.

Results

Cluster analysis

For cluster analysis, we included the total sample of 1,147 participants; after we removed outliers, 1,107 participants were included in the analyses.¹

Based on the scree plot of the fusion coefficients (Mojena’s rule; Aldenderfer & Blashfield, 1984), AIC and BIC criteria, outcome measures of the TSC procedure in SPSS and interpretability, we chose a six-cluster solution, which is shown in Table 2. These are criteria that have been suggested by previous research (Aldenderfer & Blashfield, 1984; Milligan & Cooper, 1985). The six largest clusters were selected and were used as input for iterative K-means clustering. Split-half analyses were used to test the validity of the cluster solution we found. The same pattern that are shown in Table 2 were found in the results of split-half analyses. We studied the differences between the results of the two solutions of cluster analyses with ANOVA. Although the patterns within each cluster were the same for all solutions, the height of the mean factor score differed somewhat between both split-half solutions. After correction for the high number of cases, the differences between factor scores were no longer significant. However, we found one inconsistency: In the first cluster, there was a significant difference between the mean score of the two split-half solutions on Factor 4: Axis I psychopathology. This difference might be explained by the low base rate of psychopathology. Next, after testing the robustness of the six-cluster solution, we tested the consistency between the six clusters. ANOVA showed that the clusters are indeed significantly different on the nine different factors. A division in six clusters provides us with the best-fitting model, both statistically and with respect to clinical interpretability. The six-cluster solution and the mean score on the nine factors of each cluster are shown in Table 2.

Table 2.

Six-Cluster Solution, n = 1,107

Nine factor scores	Cluster 1, n = 233	Cluster 2, n = 128	Cluster 3, n = 336	Cluster 4, n = 113	Cluster 5, n = 86	Cluster 6, n = 211
F1: Antisocial behaviour during treatment	1.052	0.195	-0.370	0.163	-0.595	-0.542
F2: Sexual problems	-0.194	-0.481	-0.470	1.592	1.405	-0.319
F3: Family problems	0.401	-0.197	-0.519	0.011	-0.610	0.779
F4: Axis 1 Psychopathology	0.380	-0.459	-0.438	0.303	0.261	0.025
F5: Offence history	-0.024	1.260	-0.318	-0.135	-0.477	-0.149
F6: Conscience and empathy	0.601	0.221	0.126	0.199	-0.980	-0.685
F7: Social skills and cognitive abilities	0.454	-0.175	-0.368	0.721	-0.543	0.085
F8: Social network	0.348	0.135	-0.282	0.217	-0.301	-0.027
F9: Substance abuse	0.532	0.584	-0.062	-0.753	-1.049	-0.016

Each cluster was given an interpretive label on the basis of the factor scores. Serious juvenile offenders score high on most risk factors (Mulder, Brand, Bullens & van Marle, 2010). Their belonging to one specific cluster means that juvenile offenders in the cluster score higher on one or two of the factors than juvenile offenders in the other clusters. Cluster 1 consist of juvenile offenders who are characterized by antisocial behaviour during treatment, lack of conscience and empathy, and substance abuse—the antisocial identity offenders. Cluster 2 are frequent offenders (high score on the factor offence history) who also have problems with substance abuse. Cluster 3 do not score higher than the other clusters on one of the factors and can be labelled as juvenile offenders with a flat profile. Cluster 4 are juvenile offenders with both sexual problems and lack of social skills and cognitive abilities, sexual problem group with a weak social identity. Cluster 5 are offenders with sexual problems only. The last cluster, Cluster 6, consists of juvenile offenders with a problematic family background.

Subgroups, Pattern of Offending Prior to the Mandatory Treatment Order

For the analyses of offending behaviour and recidivism we included only those subjects who had a time at risk of at least 2 years (n = 728). The pattern of offending of each cluster before treatment was studied. The results in Table 3 show that Clusters 4 and 5 have committed mainly sex offences before treatment (80% and 72%, respectively). These two clusters committed far fewer property offences and violent offences before treatment than the other four clusters: 35% to 47% for Clusters 4 and 5, against 88% to 96% for Clusters 1, 2, 3, and 6. If we look at the three most serious violent offences only, this pattern seems to be somewhat different. Serious assault was still more prevalent in Clusters 1, 2, 3, and 6. But manslaughter was most prevalent in Clusters 1, 3, 5, and 6 and murder was most prevalent in Clusters 3, 5, and 6. However, the differences in these three categories were not significant with ANOVA. The differences on property offending (F = 43.29, p < .001), violent offending (F = 48.52, p < .001), and sexual offending (F = 77.83, p < .001) were significant with Bonferroni correction.

Table 3.

Subgroups, Patterns of Offending, n = 728

Cluster	Property offending	Violent offending	Sexual offending	Serious assault	Manslaughter	Murder
1. Antisocial identity, n = 150	91	95	8	29	12	2
2. Frequent offenders, n = 91	96	96	11	29	5	1
3. Flat profile, n = 203	92	95	12	23	12	5
4. Sexual problems and weak social identity, n = 59	47**	47**	80**	14	5	0
5. Sexual problems, n = 46	35**	46**	71**	9	15	7
6. Problematic family background, n = 160	88	92	8	24	15	5

Note: Values are percentages.

Significant differences with the other four clusters, p ≤ .05 (Bonferroni correction).

Recidivism, Prevalence

Next, recidivism rates were calculated for each subgroup. With ANOVA, differences between clusters were analyzed. Clusters 1, 2, 3, and 6 do not appear to differ significantly from each other considering recidivism. But Clusters 4 and 5 score significantly lower than all of the other clusters on overall recidivism, violent recidivism, and severity of recidivism. This in contrast with the situation before treatment: Clusters 4 and 5 scored significantly higher than all of the other clusters on severity of offending before treatment. This is to be expected because offenders in these two clusters mainly committed sex offences before treatment, which score high on the severity scale. However, this accentuates the discrepancy we find after treatment, when these two clusters commit significantly less severe offences than all the other clusters. Recidivism rates are shown in Table 4.

Table 4.

Differences in Recidivism and Severity Between Subgroups

	Recidivism (%)	Violent recidivism (%)	Sexual recidivism (%)	Severity of past offences	Severity of recidivism
1. Antisocial identity, n = 150	87	75	5	6.97	5.56
2. Frequent offenders, n = 91	84	69	3	6.57	5.16
3. Flat profile, n = 203	85	65	5	7.03	4.87
4. Sexual problems and weak social identity, n = 59	58**	39**	10	8.20**	3.10**
5. Sexual problems, n = 46	50**	28**	7	8.52**	2.24**
6. Problematic family background, n = 160	85	66	5	7.09	5.16

Significant differences with the other four clusters, p ≤ .05 (Bonferroni correction).

Recidivism, Prediction

For each subgroup, a unique risk profile was found that predicts severity of recidivism. In Table 5, the results of stepwise linear regression analysis are shown for each of the six subgroups.

Table 5.

Linear Regression Analysis, Stepwise: Predicting Severity of Recidivism

Step	Variable entered	R ²	ΔR²	p	Final step: Standardized β coefficient
Subgroup 1: Antisocial identity (n = 150)
1	Parents’ parenting skills	.092	.092	.001	.269
2	Psychotic disorder	.145	.053	.003	-.233
Subgroup 2: Frequent offenders (n = 91)
1	Criminal behaviour in the family	.100	.100	.019	.233
2	Gambling	.164	.064	.004	.271
3	Large social network	.226	.062	.008	-.250
4	Drug abuse	.276	.050	.016	-.237
Subgroup 3: Flat profile (n = 203)
1	Being acquainted to ones victim(s)	.024	.024	.060	-.130
2	Alcohol abuse	.044	.020	.026	-.154
3	Feelings of hostility before treatment	.062	.018	.031	-.149
4	Lack of empathy	.080	.018	.049	.136
Subgroup 4: Sexual problems and a weak social identity (n = 59)
1	Involvement with antisocial peers	.206	.206	.001	.376
2	Incidents in the institution	.323	.117	.005	.318
3	Authority problems	.379	.056	.030	.240
Subgroup 5: Sexual problems (n = 46)
1	Academic achievement	.115	.115	.032	-.340
Subgroup 6: Problematic family background (n = 160)
1	Accessibility of the parents	.042	.042	.008	-.205
2	Incidents in the institution	.077	.035	.016	.187

The results show that in Subgroup 1 (antisocial identity), severity of recidivism can be predicted by antisocial behaviour during treatment and the absence of a psychotic disorder (R² = .145). In the second subgroup (frequent offenders), criminal behaviour in the family, a small social network, gambling problems, and the absence of drug abuse predict severity of recidivism (R² = .276). In the flat profile subgroup, Subgroup 3, severity of recidivism can be predicted by having one or more unknown victims of past offences, lack of empathy, not reporting feeling of hostility before treatment, and the absence of alcohol abuse (R² = .080). In Subgroup 4, with sexual problems and a weak social identity, involvement with antisocial peers, aggressive incidents in the treatment facility, and authority problems predict severity of recidivism (R² = .379). The risk factor that predicts recidivism in Subgroup 5, with sexual problems only, is the absence of low academic achievement (R² = .115). Finally, in the last subgroup that can be characterized by a problematic family background, the results show that accessibility of the parents and the occurrence of aggressive incidents in the treatment facility are predictive for severity of recidivism (R² = .077).

Discussion

The aim of this study was to find distinct subgroups within a sample of serious juvenile offenders, on the basis of the risk factors they have in common. Using cluster analysis, we succeeded to find six subgroups, each with its own characteristics. The first subgroup consists of antisocial juvenile offenders: They are characterized with a lack of conscience and empathy, substance abuse, and with problematic behaviour during treatment. This subgroup seems to be the most serious one of the six, with the highest rates for recidivism and severity of recidivism. The second subgroup is composed of frequent offenders with substance abuse problems. We found one subgroup that does not show a peak on either one of the risk domains. They have been labelled as the “flat profile” group: They do not score higher on any of the factors than the other clusters of juvenile offenders under a mandatory treatment order. The sixth subgroup consists of juvenile offenders with problems in the family and during childhood, such as lack of parenting skills, domestic violence, and neglect. Finally, we found two groups, the fourth and the fifth cluster, which are characterized by sexual problems: one with a lack of social and cognitive skills and one with sexual problems only. These two groups commit mainly sex offences before treatment. The two groups with sexual problems have the lowest recidivism rates of all six groups. The differences between the two groups with sexual problems and the other four groups were significant on almost all offending and reoffending variables. The sexual problems commit the most serious offences before treatment, sex offences scoring high on the severity scale (Category 8 and 9). After treatment, however, the sexual problem groups score by far the lowest on both the rate of recidivism and on severity of recidivism. The two sexual problem groups do not differ significantly from each other on either offending before treatment or recidivism; neither do the other four clusters or subgroups.

Next, the results of regression analysis show that each cluster or subgroup has its own unique set of risk factors that significantly predicts severity of recidivism. The strength of the predictive value of the six models differs somewhat: Severity of recidivism is harder to predict in the nonspecific subgroup and the subgroup with family problems. The explained variance in these groups is lower than 10% (R² = .077/.08, r = .27/.28), which stands for a AUC of about .66, with .50 being the same as chance (according to the conversion table presented by Rice & Harris, 2005). But although these rates are quite low for actuarial risk assessment, they are more useful for risk assessment in clinical practice. And they give important information about risk factors that are (relatively) of most importance and that should be targeted during treatment. In the subgroup with family problems, juvenile offenders who have experienced peer rejection and had low academic achievement in the past show less severe recidivism. In the flat profile subgroup, severity of recidivism is predicted by having one or more unknown victims in the past, lack of victim empathy, the absence of alcohol abuse, and the absence of feelings of hostility before treatment.

In the other four subgroups, the models for prediction of severity of recidivism are quite stronger (with AUC ranging from .70 to .87). Subgroup 1 is characterized by antisocial behaviour, and the presence of antisocial behaviour also predicts severity of recidivism after treatment. Psychiatric problems are most prevalent in this subgroup, but those juvenile offenders who had psychotic symptoms show less severe recidivism. Possibly for some of the juvenile offenders in this subgroup, the psychotic disorder preceded or even caused criminal behaviour in the past, and treatment was applied to these symptoms, which possibly caused a reduction in problematic behaviour.

In the second subgroup, problems with parents and peers (criminal behaviour in the family, an antisocial network), gambling problems, and the absence of drug abuse predict severity of recidivism. This last fact might be caused by the fact that drug abuse was tackled during treatment. The absence of drug abuse might have caused a reduction in severity of recidivism.

Finally, in the two groups with sexual problems we found two different sets of risk factors that predict severity of recidivism. In the subgroup with lack of social and cognitive skills involvement with antisocial peers, aggressive incidents in the treatment facility and authority problems predict severity of recidivism. The absence of low academic achievement predicts recidivism in the subgroup with sexual problems only.

Although this study produced interesting findings, it does have several limitations. First, not all of the subgroups were of an acceptable size. The sexual problem subgroups were quite small, which is in line with previous studies (n = 59, 8% of the sample, and n = 46, 6% of the sample, respectively) and so was the frequent offender and substance abuse subgroup (n = 91). However, the time of follow-up was considerable and the number of risk factors was large. A strong point in this study was that the results were significant after correction for the large number of risk factors that were studied. Another limitation is that we based the results on file information. Also, a repeated measures design would give more information about the development of risk factors over time. In future research, repeated measures with a larger sample is recommended.

With respect to the What Works principle (Andrews & Bonta, 1995), the results of this study provide information on the specific characteristic of serious juvenile offenders. We found differences in risk between clusters of serious juvenile offenders, which has implications for the level of intensity of intervention according to the risk principle. Each cluster was found to have a different set of risk factors that predicts severity of recidivism, which indicates that the needs of juvenile offenders are different for each cluster. According to the needs principle, interventions should aim at the specific needs of offenders.

Conclusion

In this study, we found six distinct subgroups on the basis of specific combinations of risk factors. Each subgroup appears to have its own unique set of risk factors that predicts severity of recidivism after treatment.

We can conclude that by dividing serious juvenile offenders into six clusters or subgroups we gain information on the specific needs of serious juvenile offenders, which differ according to the subgroup they belong to. The predictive validity differs somewhat between subgroups, which means that there is still a lot to improve with respect to risk assessment. But nonetheless, risk assessment in three of the six distinct subgroups, the two sexual problem groups and the frequent offender group, appears to be a lot better than in the group as a whole. Of course we should look into this further in future research. Another important finding is that each subgroup has its own unique set of problems that predict severity of recidivism and that should be targeted during treatment. Recidivism and severity of recidivism are very high in serious juvenile offenders. If we can reduce recidivism and severity of recidivism, the burden on society may diminish considerably.

Footnotes

Appendix

The Forensic Profile for Juvenile Offenders (FPJ)

1: FAMILY AND ENVIRONMENT	2: HISTORY OF CRIMINAL BEHAVIOUR	SOCIAL BEHAVIOUR/INTERPERSONAL RELATIONSHIPS
Young age of onset	Violent behaviour in the past	Antisocial behaviour in the institution
Accessibility of the parents	Criminal nonviolent behaviour in the past	Network, emotional support
Parenting skills	Number of past offences	Network, quantity
Authority problems	Severity of past offending
Involvement with criminal peers		Intimate relationships
Criminal behaviour in the family	PSYCHOLOGICAL FACTORS	Prosocial leisure activities
Physical abuse	Empathy	Social skills
Neglect	Lack of conscience
Sexual abuse	Amendable
Previous contact with mental health services	Impulse control	BEHAVIOUR DURING STAY IN THE INSTITUTION
Substance abuse by parents	Problem insight	Avoidant coping style
Mental health problems, parents	Sexual problems	Negative (aggressive) coping style
Peer rejection	Intelligence/ IQ	Positive (support seeking) coping style
Academic achievement		Therapeutic relationship
Truancy		Lack of treatment adherence
OFFENCE-RELATED RISK FACTORS AND SUBSTANCE USE		Incidents, aggression in the institution
Number of solo offences	PSYCHOPATHOLOGY	Treatment motivation
Number of group offences	Anxiety disorder	Self-care
Offence during medication stop	Depressive disorder (last year)	Positive commitment to school/work
Substance abuse during/ preceding the offence	Neurological problems	Escape, absconding
Alcohol abuse	Conduct disorder
Drug abuse	Feelings of hostility
Gambling	Autism spectrum disorder
Familiarity with the victim	Psychotic disorder
Pedosexuality	Sadism
Searching for a victim, planning the offence

Declaration of Conflicting Interests

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

Funding

The author(s) received no financial support for the research and/or authorship of this article.

Notes

References

Aldenderfer

M. S.

Blashfield

R. K.

(1984). Cluster analysis. Beverly Hills, CA: Sage.

Andrews

D. A.

Bonta

(1995). The psychology of criminal conduct (2nd ed.). Cincinnati, OH: Anderson.

Booker

Loper A.

Hoffschmidt

S. J.

Ash

(2001). Personality features and characteristics of violent events committed by juvenile offenders. Behavioural Sciences and the Law, 19, 81-96.

Brand

E. F. J. M.

(2005). Research study PIJ-files 2003C: Predictive validity of the FPJ-list. The Hague: Department of Justice, National Agency of Correctional Institutions.

Brand

(2005). A measure for the quality of instruments for risk assessment. In Sjerps

M. F.

Coster van Voorhout

J. A.

(Eds.). The uncertain evidence. The use of statistics and probability in criminal law (pp. 429-456). Deventer: Kluwer.

Brand

E. F. J. M.

van Heerde

W. K.

(2004). Scoring Manual FPJ-list. Juvenile forensic profile. The Hague: Department of Justice, National Agency of Correctional Institutions.

Caldwell

M. F.

(2002). What we do not know about juvenile sexual reoffence risk. Child Maltreatment, 7, 291-302.

Clingempeel

W. G.

Henggeler

S. W.

(2003). Aggressive juvenile offenders transitioning into emerging adulthood: Factors discriminating persistors and desistors. American Journal of Orthopsychiatry, 73, 310-323.

Cottle

C. C.

Lee

R. J.

Heilbrun

(2001). The prediction of criminal recidivism in juveniles: A meta-analysis. Criminal Justice and Behaviour, 28, 367-394.

10.

Dawkins

M. P.

(1997). Drug use and violent crime among adolescents. Adolescence, 32, 396-405.

11.

Dembo

Schmeidler

(2003). A classification of high-risk youths. Crime & Delinquency, 49, 201-230.

12.

Friendship

Beech

A. R.

Browne

K. D.

(2002). Reconviction as an outcome measure in research: a methodological note, British journal of criminology, 42, 442-444.

13.

van Heerde

W. K.

Mulder

E. A.

(2005). Research study PIJ-files 2003-B. Clinicalevaluation FPJ-list & SAVRY in Eikensteinand FORA. The Hague: Department of Justice, National Agency of Correctional Institutions.

14.

van Heerde

W. K.

Brand

E. F. J. M.

Hoff

G. I. C. M. van t

Mulder

E. A.

(2004). Inter-rater reliability and convergent validity of the Juvenile Forensic Profile. The Hague: Department of Justice, National Agency of Correctional Institutions.

15.

Heilbrun

Witte

Keeney

Brock

Westendorf

Waite

. . . Shumate

(2000). Risk factors for juvenile criminal recidivism. The postrelease community adjustment of juvenile offenders. Criminal Justice and Behaviour, 27, 275-291.

16.

Hendriks

Bijleveld

(2008). Recidivism among juvenile sex offenders after residential treatment. Journal of Sexual Aggression, 14, 19-32.

17.

Hoff

G. I. C. M. van t

Brand

E. F. J. M.

Parijs

A. J.

van Heerde

W. K. van

(2002). Pilot study PIJ-files. From word to number. The Hague: Department of Justice, National Agency of Correctional Institutions.

18.

Johnson

L. S.

van Voorhis

Ritchey

P. N.

(2007). Personality, criminal behaviour, and risk assessment: Implications for theory and practice. Criminal Justice and Behaviour, 34, 60-75.

19.

Kenny

D. T.

Keogh

Seidler

(2001). Predictors of recidivism in Australian juvenile sex offenders: Implications for treatment. Sexual Abuse, 13, 131-148.

20.

Kordelaar

W. F. J. M.

(2002). Decision support in research on mental ability in criminal law. A forensic psychological study (PhD thesis), Catholic University Brabant. Deventer: Kluwer.

21.

Långström

Grann

(2000). Risk for criminal recidivism among young sex offenders. Journal of Interpersonal Violence, 15, 855-871.

22.

Långström

Grann

Lindblad

(2000). A preliminary typology of young sex offenders. Journal of Adolescence, 23, 319-329.

23.

Lodewijks

H. P. B.

Doreleijers

T. A. H.

de Ruiter

(2008). Savry risk assessment in violent Dutch adolescents: Relation to sentencing and recidivism. Criminal Justice and Behaviour, 35, 696-709.

24.

Loeber

Farrington

D. P.

(2001). The significance of child delinquency, child delinquents: Development, intervention and service needs. Thousand Oaks, CA: Sage.

25.

Loeber

Farrington

D. P.

Stouthamer-Loeber

Raskin White

(2008). Violence and serious theft. Development and prediction from childhood to adulthood. Taylor & Francis group.

26.

Milligan

G. W.

Cooper

M. C.

(1985). An examination of procedures for determining the number of clusters in a data set. Psychometrika, 50, 159-179.

27.

Miner

M. H.

(2002). Factors associated with recidivism in juveniles: An analysis of serious juvenile sex offenders. Journal of Research in Crime and Delinquency, 39, 421-436.

28.

Moffitt

T. E.

Caspi

(2001). Childhood predictors differentiate life-course persistent and adolescence-limited antisocial pathways among males and females. Development and Psychopathology, 13, 355-375.

29.

Moffitt

T. E.

Caspi

Harrington

Milne

B. J.

(2002). Males on the life-course-persistent and adolescence-limited antisocial pathways: Follow-up at age 26 years. Development and Psychopathology, 14, 179-207.

30.

Mulder

E. A.

Brand

E. F. J. M.

Bullens

R. A. R.

van Marle

H. J. C.

(2010). A classification of risk factors in serious juvenile offenders and the relation between patterns of risk factors and recidivism. Criminal Behaviour and Mental Health, 20, 23-38.

31.

Mulder

E. A.

Brand

E. F. J. M.

Bullens

R. A. R.

van Marle

H. J. C.

(2011). Risk factors for overall recidivism and severity of recidivism in serious juvenile offenders. International Journal of Offender Therapy and Comparative Criminology, 55, 118-135. doi: 10.1177/0306624X09356683

32.

Nisbet

I. A.

Wilson

P. H.

Smallbone

S. W.

(2004). A prospective longitudinal study of sexual recidivism among adolescent sex offenders. Sexual Abuse, 16, 223-234.

33.

Oberwittler

(2004). A multilevel analysis of neighbourhood contextual effects on serious juvenile offending: The role of subcultural values and social disorganization. European Journal of Criminology, 1, 201-235.

34.

Onifade

Davidson

Livsey

Turke

Horton

Malinowski

. . . Wimberly

(2008). Risk assessment: Identifying patterns of risk in young offenders with the Youth Level of Service/Case Management Inventory. Journal of Criminal Justice, 36, 165-173.

35.

Parks

G. A.

Bard

D. E.

(2006). Risk factors for adolescent sex offender recidivism: Evaluation of predictive factors and comparison of three groups based upon victim type. Sexual Abuse, 18, 319-342.

36.

Potter

C. C.

Jenson

J. M.

(2003). Cluster profiles of multiple problem youth: Mental health problem symptoms, substance use, and delinquent conduct. Criminal Justice and Behaviour, 30, 230-250.

37.

Prentky

Harris

Frizzell

Righthand

(2000). An actuarial procedure for assessing risk with juvenile sex offenders. Sexual Abuse, 12, 71-93.

38.

Raine

Moffitt

T. E.

Caspi

Loeber

Stouthamer-Loeber

Lynam

(2005). Neurocognitive impairments in boys on the life-course persistent antisocial path. Journal of Abnormal Psychology, 114, 38-49.

39.

Rice

M. E.

Harris

G. T.

(2005). Comparing effect sizes in follow-up studies: ROC area, Cohen’s d, and r. Law and Human Behaviour, 29, 615-620.

40.

Stefurak

Calhoun

G. B.

Glaser

B. A.

(2004). Personality typologies of male juvenile offenders using a cluster analysis of the Millon Adolescent Clinical Inventory introduction, International journal of offender therapy and comparative criminology, 48(1), 96-110.

41.

Stevens

van Marle

H. J. C.

(2003). The PIJ-order as a way of reeducation, Sancties, 4, 221-234.

42.

Teichner

Golden

C. J.

Crum

T. A.

Azrin

N. H.

Donohue

van Hasselt

V. B.

(2000). Identification of neuropsychological subtypes in a sample of delinquent adolescents. Journal of Psychiatric Research, 34, 129-132.

43.

Trulson

C. R.

Marquart

J. W.

Mullings

J. L.

Caeti

T. J.

(2005). In between adolescence and adulthood. Recidivism outcomes of a cohort of state delinquents. Youth Violence and Juvenile Justice: An Interdisciplinary Journal, 3, 355-387.

44.

van der Linden

A. P.

ten Siethoff

F. G .A.

Zeijlstra-Rijpstra

A. E. I. J.

(2001). Jeugd en Recht [Youth and law]. AK Houten, Netherlands: Bohn Stafleu Van Loghum.

45.

Van Wijk

A. P.

Mali

B. R. F.

Bullens

R. A. R.

Vermeiren

R. R.

(2007). Criminal profiles of violent juvenile sex and violent juvenile non-sex offenders: An explorative longitudinal study. Journal of Interpersonal Violence, 22, 1340-1355.

46.

Vermeiren

(2003). Psychopathology and delinquency in adolescents: A descriptive and developmental perspective. Clinical Psychology Review, 23, 277-318.

47.

Vermeiren

Bogaerts

Ruchkin

Deboutte

Schwab-Stone

(2004). Subtypes of self-esteem and self-concept in adolescent violent and property offenders. Journal of Child Psychology and Psychiatry, 45, 405-411.

48.

Waite

Keller

McGarvey

E. L.

Wieckowski

Pinkerton

Brown

G. L.

(2005). Juvenile sex offender re-arrest rates for sexual, violent nonsexual and property crimes: A 10-year follow-up. Sexual Abuse, 17, 313-331.

49.

Worling

J. R.

(2001). Personality-based typology of adolescent male sexual offenders: Differences in recidivism rates, victim-selection characteristics, and personal victimization histories. Sexual Abuse, 13, 149-166.

50.

Worling

J. R.

Curwen

(2000). Adolescent sexual offender recidivism: Success of specialized treatment and implications for risk prediction. Child Abuse & Neglect, 24, 965-982.

51.

Wartna

B. S. J.

Harbachi

van der el Laan

A. M.

(2005). Youth in detention: prevalence of recidivism in youth in juvenile justice institutions. WODC, ‘s-Gravenhage.