Which Risk Factors are Really Predictive?

Abstract

In this study, we examined the (incremental) predictive validity of Andrews and Bonta’s Central Eight risk factors for recidivism in the German youth correctional system. The sample consisted of N = 589 male youth inmates who were incarcerated for the first time. Recidivism during the 78 months’ follow-up was assessed using official data. The Central Eight risk factors predicted recidivism in survival analyses. In a cross-validation, composite scores predicted general (area under curve [AUC] = .65) and violent recidivism (AUC = .66). The Moderate Four risk factors (family, school, leisure/recreation, substance abuse) showed predictive validity incremental to the Big Four risk factors (history of antisocial behavior, antisocial personality pattern, antisocial cognition, antisocial associates). School was the most predictive single risk factor. The results provide evidence for the applicability of the Central Eight as predictors for recidivism in the German youth correctional system. Furthermore, the study adds to the debate on the importance of dynamic risk factors.

Keywords

Central Eight risk factors for recidivism prisoners risk–need–responsivity (RNR) model incremental predictive validity survival analyses risk assessment criminogenic needs Level of Service Inventory

Andrews and Bonta’s (2010a) risk–need–responsivity (RNR) model of offender treatment includes a classification of the Central Eight risk factors for recidivism that is based on general personality models and cognitive social learning perspectives as well as on research results (Andrews et al., 2012). According to the model, the so-called Big Four risk factors (history of antisocial behavior, antisocial personality pattern, antisocial cognition, and antisocial associates) are causal risk factors that strongly predict criminal recidivism (Andrews & Bonta, 2010a). The Moderate Four risk factors (family/marital circumstances, school/work, leisure/recreation, and substance abuse) are conceptualized as having an additional, but weaker, impact on the predictive validity for recidivism (Andrews & Bonta, 2010a). Our study aims to test the predictive validity of these Central Eight risk factors for two forms of recidivism (general and violent) in German offenders who are incarcerated for the first time in youth correctional facilities. In particular, we investigate whether the Moderate Four risk factors provide incremental predictive validity relative to the Big Four risk factors, as predicted by the RNR model.

The Risk–Need–Responsivity Model of Offender Treatment

The RNR model is one of the most important models of offender treatment and has been broadly discussed in the scientific literature for more than 20 years (e.g., Andrews, Bonta, & Wormith, 2011; Ward, Yates, & Willis, 2012). The popularity of the model is also due to its practical utility and applicability, as the model assumes that adherence to the risk, need, and responsivity principles of offender treatment leads to a reduction in recidivism rates (Andrews & Bonta, 2010a). The risk principle states that the level of offender treatment should correspond to each offender’s individual level of risk for recidivism. According to the need principle, treatment should address offenders’ criminogenic needs (dynamic risk factors). The responsivity principle states that the treatment needs to be responsive offenders’ cognitive and emotional styles. Since the first step of the model is to determine the offenders’ level of risk for recidivism—answering the question of who should be treated (Andrews and Bonta, 2010b)—all subsequent steps that finally lead to a reduction of recidivism rates are based on risk assessment. The identification and classification of risk factors could thus be considered to be the core of the model.

The Central Eight Risk Factors

In the model’s classification of risk factors, Andrews and Bonta (2010a) distinguish between the so-called Big Four and Moderate Four risk factors. They emphasize the importance of dynamic risk factors, which can be addressed in treatment. Out of the eight risk factors of Andrews and Bonta’s (2010a) model, only one of the Big Four (history of antisocial behavior) is typically regarded to be static, and all other risk factors are considered dynamic (McGrath & Thompson, 2012).

The classification of risk factors in the RNR model is hierarchical. The Big Four are major causal risk factors that should be addressed in treatment, and the Moderate Four risk factors are environmental factors that influence recidivism rates directly by providing opportunities for criminal behavior and indirectly by interacting with the Big Four (Andrews & Bonta, 2010a). The predictive validity of the Moderate Four risk factors is presumed to be weaker than the predictive validity of the Big Four. However, there is little empirical consensus on the relative importance of risk factors for recidivism (Singh & Fazel, 2010), and to our knowledge, no previous studies have examined the relative or the incremental predictive validity of all four of the Big Four (as a group) compared with all four of the Moderate Four (as a group) in a prospective study. Empirical results do corroborate the assumption that the risk factors are interrelated (McGrath & Thompson, 2012; Van der Put et al., 2011).

Research on the applicability of the Central Eight risk factors for the prediction of recidivism in German offenders from the youth correctional system is also lacking. Considering these issues, three questions arise: (a) How predictive are the Central Eight risk factors for different forms of recidivism? (b) Do the Moderate Four risk factors provide information that significantly increases the predictive validity for recidivism, relative to the Big Four risk factors? (c) Are some of the Central Eight risk factors redundant for the prediction of recidivism in inmates of German youth correctional facilities (i.e., they do not contribute meaningfully to the prediction of recidivism when the remaining risk factors are controlled)? The latter two questions address the incremental predictive validity of risk factors, which has been neglected by previous research (McGrath & Thompson, 2012; Walters, 2011).

State of Research on the Central Eight and Recidivism

The “first structured outline of the Central Eight risk and need factors” (Andrews & Bonta, 2010a, p. 56) is the Level of Service Inventory–Revised (LSI-R). The LSI-R (Andrews & Bonta, 1995) has been translated to multiple languages (e.g., Spanish, French, and Croatian) and has been shown to significantly predict violent and general recidivism across different samples and cultures (Campbell, French, & Gendreau, 2011; Dahle, 2006; Gendreau, Little, & Goggin, 1996; Girard & Wormith, 2004; Hsu, Caputi, & Byrne, 2009; Smith, Cullen, & Latessa, 2009; Vose, Cullen, & Smith, 2008). In addition, changes in the LSI-R scores have been shown to predict recidivism (Vose, Smith, & Cullen, 2013), and a broad body of the literature has shown that Youth Level of Service Inventories (YLS; for example, the Youth Level of Service/Case Management Inventory [YLS/CMI]; Hoge & Andrews, 2002) predict recidivism across different samples of younger offenders (Olver, Stockdale, & Wong, 2012; Olver, Stockdale, & Wormith, 2009; Onifade et al., 2008; Schmidt, Hoge, & Gomes, 2005; Schwalbe, 2007, 2008; Takahashi, Mori, & Kroner, 2013; Thompson & McGrath, 2012).

In an extensive research project, Dahle (2006) investigated the applicability of the LSI-R to the adult German prison system. He found that the LSI-R predicted recidivism, and that the predictive validity of general recidivism decreased with time, while the predictive validity for violent recidivism increased. He conducted several studies and established an adaptation of the LSI-R to the German prison system (Dahle, Harwardt, & Schneider-Njepel, 2012). However, as the validation samples were all recruited from adult prisons, it remains unknown if the LSI-R predicts recidivism among offenders in the German youth correctional system.

Although they overlap substantially, the LSI-R does not assess all of the Central Eight risk factors in the current outline of the RNR model (Andrews and Bonta, 2010a). After reviewing existing research results including studies on the predictive validity of the LSI-R as well as other measures, Andrews and Bonta (2010a) determined that antisocial personality pattern is one of the Big Four. The antisocial personality pattern is partially represented in items from the LSI-R’s emotional/personal domain, which includes items on mental disorders and mental health treatment (Andrews & Bonta, 2010a). However, in contrast to the emotional/personal domain, the antisocial personality pattern in the RNR model is defined in terms of extremes of normal dimensions of personality that are common to everyone, and in terms of a history of early problematic and deviant behavior (Andrews & Bonta, 2010a). Therefore, there is no comprehensive operationalization of the antisocial personality pattern in the LSI-R and hence it is imprudent to draw conclusions about the predictive validity of the Central Eight risk factors from studies using the LSI-R. In addition, the LSI-R includes additional scales that are not included in the Central Eight, for example, financial situation and accommodation (Andrews & Bonta, 2010a). These domains were not found sufficiently predictive and were therefore not included in the Central Eight. As a result of these differences between the current RNR classification of risk factors (i.e., the Central Eight) and the LSI-R, the sum score of the LSI-R is not equivalent to an assessment of the Central Eight risk factors.

When investigating the incremental predictive validity of the YLS, many studies have focused on the comparison with other risk assessment measures for juveniles (e.g., Catchpole & Gretton, 2003; Schmidt, Campbell, & Houlding, 2011; Welsh, Schmidt, McKinnon, Chattha, & Meyers, 2008), and only a few studies have investigated the relative incremental predictive validity of the risk domains (e.g., Marshall, Egan, English, & Jones, 2006; McGrath & Thompson, 2012). Using a large sample of 3,568 Australian young persons under justice supervision in the community, McGrath and Thompson (2012) recently analyzed the incremental predictive validity of the Australian Adaptation of YLS/CMI (YLS/CMI-AA; Hoge & Andrews, 1995) in a 1-year follow-up. They found that the combination of all Central Eight risk factors predicted recidivism better than only the dynamic risk factors or only the static risk factor (prior and current offenses). Their results showed that the predictive validity of a combination of four dynamic risk factors (education/employment, peer relations, substance abuse, and attitudes/beliefs) and the static risk factor (prior and current offenses) was not significantly different from the predictive validity of all eight risk domains together. Thus, the three risk domains—family and living, personality/behavior, and leisure/recreation—did not show incremental predictive validity. The authors assumed that the weak performance of the family and living and personality/behavior domains in multivariate analyses was due to high intercorrelations with the other dynamic risk factors (McGrath & Thompson, 2012).

Present Study

Our study aims to add to the branch of research investigating relative and incremental predictive validity of risk factors for recidivism in young offenders. In contrast to other studies, we investigate the incremental predictive validity of two groups of risk factors—the Big Four and the Moderate Four—as two groups and as eight individual predictors. To the best of our knowledge, our study is the first to address the predictive validity of the Central Eight using a sample of German youth correctional inmates.

Consistent with previous research results (Girard & Wormith, 2004; Olver et al., 2009), we expected the Central Eight risk factors to predict general recidivism and violent recidivism. We expected the Big Four and the Moderate Four to show incremental predictive validity relative to each other. Furthermore, we explored which variables were redundant for the prediction of recidivism to identify the best ratio of assessment effort and model fit. That is, we sought to identify a combination of variables out of the Central Eight that requires the lowest level of resources but predicts recidivism best. In addition to conventional statistical procedures like correlational and receiver operating characteristic (ROC) analyses, we used survival analyses to examine the predictive validity for recidivism.

Method

Sample and Procedure

This study was part of an extensive prospective research project (2004-2012) examining the effects of prison sentences on offenders who were incarcerated in youth correctional facilities for the first time. Participants were recruited from six youth correctional facilities in Northern Germany. The prisons were located in several German counties and varied in terms of security level. All male, German youth offenders who were serving a custodial sentence for the first time and were referred to one of the participating youth correctional facilities between 1998 and 2001 were asked if they wanted to participate in the study. According to the prison staff, about 89% of the youth offenders agreed to participate in the study. The only exclusion criteria were insufficient German language skills and a non-German passport. Our sample (N = 2,405) can be considered to be a high-risk sample since only 6.7% of all convictions in the German juvenile penal system lead to the incarceration of an offender (Walkenhorst, 2010). Thus, offenders who are incarcerated in a youth correctional facility either committed very serious crimes or have a long history of less serious crimes.

To reduce participant fatigue in completing the large number of questionnaires that were administered, we split the first interview in two parts, which took place M = 3.3 months (SD = 3.4) and M = 5.8 months (SD = 3.7) after correctional facility admission. We scheduled a second interview shortly before each individual’s planned release from prison (M = 17.2 months after admission; SD = 9.5) to reassess some of the dynamic questionnaires and to assess additional relevant information at release. All three interview sessions were limited to 90 min each. However, due to this design, some participants could not be reassessed because their release date was changed on short notice.

The original sample size was reduced for several reasons. First, 299 youth offenders were part of the cross-sectional portion of the study. Therefore, they were interviewed only once in prison and information on recidivism was not gathered. Second, we excluded youth offenders from the sample if they had a previous incarceration documented in their prison files (n = 41) to have a sample that was more homogeneous and specific. Finally, some participants could not be reassessed after the first interview for a variety of reasons, such as transfer to another correctional facility, to a hospital, or to a forensic mental institution. Some participants were suicidal and could therefore not be interviewed, and others had already been released from prison before all interviews were completed.

For this study, we included only participants who had completed all interviews. Sufficient data for all risk factors were available for a subsample of 589 youth offenders. The participants’ ages at incarceration ranged from 14.5 to 25.4 years (M = 19.4, SD = 1.9). This range in age is consistent with German legislation, which allows offenders to be committed to youth correctional facilities between the ages of 14 and 24 years, although the actual average age at which offenders are imprisoned in youth correctional facilities in Germany is 20 years (Dünkel & Geng, 2007). On average, the participants were incarcerated for 17.8 months (SD = 9.5, range = 3.3-65.7 months). They had committed the following crimes (index crimes): property offenses (38.1%), bodily injury (32.5%), grievous bodily harm (13.4%), drug offenses (5.3%), sex offenses (1.4%), homicide and manslaughter (.5%), and other crimes (8.8%). During their prison term, 6.6% participated in social therapy, drug therapy, or antiaggression training.

Participation in the study was voluntary and we assured confidentiality according to the criteria of informed consent. The youth offenders were informed that they could refuse to take part without any negative repercussions. They received 10 Euros (approximately US$9 at the time the study was conducted) as compensation. The study was approved by the review boards of the prison administrations, the Ministries of Justice of the involved federal states, and the National Data Protection Authorities. The advisory board of the Criminological Research Institute of Lower Saxony provided ethics review and approval of this study.

Measures

Three different sources provided information on the Central Eight risk factors: (a) interviews with participants, (b) official criminal records, and (c) prison files. Thirteen research assistants conducted the standardized interviews and read the questions out loud to participants. Participants had the chance to ask for clarification if they did not understand a question and gave their responses to each item out loud. Official criminal records were requested from the German Federal Office of Justice (Bundesamt für Justiz), which administers the German Federal Central Register (Bundeszentralregister), a database of all legal decisions. Information in this register constitutes the official criminal record, capturing all detected incidents of criminal conduct since the age of criminal responsibility (14 years).

To provide for comparability with previous studies, we operationalized the risk factors according to the LSI-R (see Table 1). Deviations from the LSI-R are briefly discussed below.

Table 1:

Overview of the Operationalization of the Central Eight Risk Factors

Risk Factors/Indicators	Source	Description of the Risk Factor by Andrews and Bonta (2010a)/Calculation of the Indices	Time of Assessment	M (SD)
History of antisocial behavior		“Early involvement in a number and variety of antisocial activities in a variety of settings (…). Major indicators include being arrested at a young age, a large number of prior offenses, and rule violations while on conditional release.” (p. 58)		.42 (.19)
Any prior convictions?	C. R.	The participant was given a score of 1 if he had any prior convictions for other offenses than the index offense(s)		.79 (.41)
Two or more prior convictions?	C. R.	The participant was given a score of 1 if he had 2 or more prior convictions		.68 (.47)
Three or more prior convictions?	C. R.	The participant was given a score of 1 if he had 3 or more prior convictions		.58 (.49)
Three or more current convictions?	P. F.	The participant was given a score of 1 if he had committed 3 or more index crimes		.08 (.27)
Convictions before age 16?	C. R.	The participant was given a score of 1 if he had committed a crime that led to a conviction before the age of 16		.45 (.50)
Any juvenile detention before incarceration?	I.	“Have you ever been in juvenile detention prior to incarceration?” (0 = no, 1 = yes)	1a	.46 (.50)
Escape history from the correctional institution?	P. F.	The participant was given a score of 1 if he had escaped or tried to escape from prison during incarceration		.06 (.24)
Ever punished for institutional misconduct?	P. F.	Mean score of (a) disciplinary sanctions during imprisonment (0 = no, 1 = yes) and (b) noncompliance with instructions from the prison staff (0 = no, 1 = yes)		.50 (.39)
Charge laid on probation/parole suspended?	P. F.	If the participant was incarcerated because of revocation of probation according to the criminal file, or if it was noted in his criminal file that his probation had been suspended, he was given a score of 1	1a	.02 (.15)
Official record of assault/violence	C. R.	The participant was given a score of 1 if he had a prior conviction for a violent offense in his criminal files and/or if his index crime was a violent one		.59 (.49)
Antisocial personality pattern		“Defined according to the Multidimensional Personality Questionnaire. . . . Weak Constraint (low on traditionalism, or endorsing high moral standards; low on harm avoidance, or low on an avoiding excitement and danger; low on self-control; low on being reflective and planful). Negative Emotionality (aggression, or causes discomfort in others; alienation and feels mistreated; stress reaction dominated by anger and irritability).” (p. 58)		.48 (.13)
		“We underscore the word ‘pattern’ because it includes both the personality facets . . . and a pattern of law-violating and problematic behaviors, often evidenced early in life.” (p. 218)
Early history of law-violating and problematic behavior^a	I.	The average score of the SCID-II screening questionnaire assessing symptoms of conduct disorder (Wittchen, Schramm, Zaudig, Mombour, & Unland, 1993) was used to assess this domain	1b	.44 (.20)
Low constraint	I.	The mean score of the scales traditionalism, control, and harm avoidance was calculated. Subsequently, the mean score was subtracted from 1, to obtain a measure of low constraint	1b	.51 (.12)
High negative emotionality	I.	A mean score of the scales aggressivity, alienation, and stress reactivity was calculated	1b	.53 (.17)
Antisocial cognition		“This set of variables includes attitudes, values, beliefs, rationalizations and a personality identity that is favorable to crime. . . . Specific indicators would include identification with criminals, negative attitudes toward the law and justice system, a belief that crime will yield rewards, and rationalizations that specify a broad range of conditions under which crime is justified. . . .” (p. 59)		.58 (.12)
Supportive of crime	I.	The author-created scale^b consisted of 8 statements. Each item was rated on a 4-point Likert scale (“To what extent do you agree to the following statements?”: 1 = not at all, 2 = somewhat, 3 = highly, 4 = extremely).The mean score was divided by 4	2	.43 (.14)
Unfavorable toward convention	I.	16 items from the scale “Attitudes towards the legitimacy of the law” by Ortmann (1987) and 4 items from the scale “Orientation towards social norms and values” by Seitz (1983) were all rated on a 4-point Likert scale (“How accurate is the following statement?”: 1 = not at all, 2 = somewhat, 3 = highly, 4 = extremely). The mean score was divided by 4	2	.58 (.13)
Poor attitudes toward sentence	I.	10 items from the scale “Restriction of autonomy” by Ortmann (1987) were rated on a 4-point Likert scale (“How accurate is the following statement?”: 1 = not at all, 2 = somewhat, 3 = highly, 4 = extremely). The mean score was divided by 4	2	.66 (.16)
Poor attitude toward supervision	I.	7 items from the scale “Expectation of hostility by the staff of the correctional institution” and 6 items from the scale “Sensitive intolerance to the treatment by the penal system” by Seitz (1983) were rated on a 4-point Likert scale (“How accurate is the following statement?”: 1 = not at all, 2 = somewhat, 3 = highly, 4 = extremely). The mean score was divided by 4	2	.66 (.17)
Antisocial associates		“This risk/need factor includes both association with procriminal others and a relative isolation from anticriminal others.” (p. 59)		.68 (.27)
Some criminal friends^a	I.	4 items from Scale 4 of the “Questionnaire of imprisonment” by Ortmann (1987; translation by the authors) were used: “My friends outside will always come into contact with the law,” “My friends outside are just a bunch of criminals,” “My friends outside are just bad and they will never change,” and “Eventually, most of my friends outside will end up in jail.” Items were rated on a 4-point Likert scale (“How accurate is the following statement?”: 0 = not at all, 1 = somewhat, 2 = highly, 3 = extremely). Participants with an average of more than 2 were scored with 1	1b	.57 (.50)
Few noncriminal friends^a	I.	2 items from Scale 4 of the “Questionnaire of imprisonment” by Ortmann (1987; translation by the authors) were used: “My friends outside stick to the law” and “None of my friends outside has a criminal record.” Items were rated on a 4-point Likert scale (“How accurate is the following statement?”: 0 = extremely, 1 = highly, 2 = somewhat, 3 = not at all). Participants with an average of more than 2 were scored with 1	1b	.93 (.25)
Social isolation	I.	We used a short version of the UCLA Loneliness Scale (Russell, Peplau, & Cutrona, 1980). Items were rated on a 4-point Likert-type scale (“How accurate is the following statement?”: 0 = not at all, 1 = somewhat, 2 = rather correct, 3 = extremely correct). Participants with an average of >2 on this scale were rated with 1	2	.38 (.48)
Family		“The key to assessing both family of origin for young people and marital circumstances for older people is the quality of interpersonal relationships with regard to antisocial behavior, including monitoring, supervision, and disciplinary approaches. . . . In the assessment of youths, the two key parenting variables are nurturance/caring and monitoring supervision.” (p. 59)		.40 (.30)
Nonrewarding parental	I.	“Have you ever lived in foster care?” (0 = no, 1 = yes)	1a	.34 (.47)
Nonrewarding parental	I.	“Have you ever been supervised by the child protection services?” (0 = no, 1 = yes)	1a	.66 (.47)
Nonrewarding parental	I.	“Did one of your parents have drinking problems/drug problems in your childhood?” (0 = no, 1 = yes)	1a	.35 (.48)
Criminal parents I.		“Does one of your parents have a criminal record?” (0 = no, 1 = yes)	1a	.23 (.42)
School		“Generally, the risk/need factors are low levels of performance and involvement and low levels of rewards and satisfactions.” (p. 59)		.47 (.27)
Less than regular Grade 10?	I./P. F.	Participants who indicated that they studied at a “Förderschule” (remedial school) or who dropped out of school before imprisonment were given a score of 1. Participants who visited a school with a higher level than Förderschule during imprisonment or who had not dropped out of Hauptschule (secondary modern school) or any higher education school before imprisonment were scored 0	1a	.51 (.50)
Less than regular Grade 12?	I./P. F.	Participants who indicated that they studied at a Förderschule or “Hauptschule” (secondary modern school) or who had dropped out of school before imprisonment were given a score of 1. Participants who visited a school with a higher level than Hauptschule during imprisonment or who had not dropped out of “Realschule” (junior high) or any higher education school before imprisonment were scored 0	1a	.89 (.32)
Suspended or expelled at least once?	I.	“Have you ever been expelled from school?” (0 = no, 1 = yes)	1a	.36 (.48)
Participation/performance	P. F.	“Does the inmate have any kind of learning disability?” (0 = no, 1 = yes)		.21 (.40)

Participation/performance	I.	“Did you drop out of school?” (0 = no, 1 = yes)	1a	.53 (.50)
Participation/performance	I.	“How many times did you have to repeat a class?” (0 = not at all or once; 1 = more than once)	1a	.27 (.45)
Leisure/recreation		“Low levels of involvement and satisfactions in anticriminal leisure pursuits.” (p. 59)		.28 (.45)
Absence of recent participation in an original organized activity	P. F.	A participant given a score of 0 if he had participated on a regular basis in at least 2 of the following leisure/recreation activities: (a) attending religious services, (b) playing soccer, (c) playing table tennis, (d) reading, (e) working out in the gym, (f) playing other ball games; all other participants were rated with 1		.28 (.45)
Substance abuse		“The risk/need factor is problems with alcohol and/or other drugs (tobacco excluded).” (p. 60)		.48 (.26)
Alcohol problem ever?	P. F./I.	A participant was given a score of 1 if he stated that he was alcohol dependent (0 = not dependent, 1 = dependent), or if alcohol dependence was noted in his prison file	1a	.14 (.34)
Drug problem ever?	P. F./I.	A participant was given a score of 1 if (a) he either indicated using drugs (heroin, cocaine, surrogate drugs) at least once in a while, (b) he indicated using cannabis and/or LSD on a regular basis, (c) the prison file included a note that the participant had abused any of the substances before imprisonment, or if (d) he indicated that he had been in treatment for drug abuse before	1a	.70 (.46)
Drug problem currently?	P. F./I.	A participant was given a score of 1 if (a) he indicated having consumed one of the aforementioned drugs during incarceration, or (b) if it was noted in the prison file that the participant had abused drugs during incarceration, or (c) if he had started drug treatment in prison and dropped out of it	2	.34 (.48)
Law violations under drugs/alcohol?		A participant was given a score of 1 if he affirmed that he committed his offense under either alcohol or drug influence	1a	.72 (.45)

Note. UCLA = University of California, Los Angeles; LSD = lysergic acid diethylamide; SCID-II = Structured Clinical Interview for DSM-III-R Axis II; DSM-III-R = Diagnostic and Statistical Manual of Mental Disorders (3rd ed.; rev.); I. = interview, C. R. = criminal record, P. F. = prison file; RNR = risk–need–responsivity; LSI-R = Level of Service Inventory–Revised. The time of the assessment for the interview is coded as follows: 1a = first part of the first interview, 1b = second part of the first interview, 2 = second interview;

These factors were included twice in the average of the respective risk factor (in correspondence with the weighting of the RNR model/the LSI-R model).

The items (translated by the authors) were as follows: “Committing crime pays, because (a) ‘the others also make money wherever they can,’ (b) ‘the risk of getting caught is quite small,’ (c) ‘it’s fun,’ (d) ‘the benefits are higher than the punishment,’ (e) ‘you don’t have a chance to get things in a different way anyways,’ (f) ‘sometimes you just have to get your way by force,’ (g) ‘you can show the authorities that you are smarter than they are,’ and (h) ‘you can prove to your friends that you are brave and cool.’”

Antisocial Personality Pattern

Our operationalization of antisocial personality pattern included (a) the personality facets low constraint and high negative emotionality and (b) a pattern of early law-violating and problematic behavior. To assess the personality facets, we used the German version of the Multidimensional Personality Questionnaire (MPQ; Tellegen, 1982). Consistent with the Dunedin Multidisciplinary Health and Development Study (Caspi et al., 1997; Krueger et al., 1994), we chose a short, 144-item version of the German MPQ (F. M. Spinath & A. Angleiter, personal communication).

Furthermore, we operationalized the behavioral component of the antisocial personality pattern by assessing symptoms of conduct disorder before the 14th birthday (vs. 15th birthday in the original questionnaire, to avoid a ceiling effect). Therefore, we used the German version of the screening questionnaire for antisocial personality disorder from the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders (3rd ed.; rev.; DSM-III-R; American Psychiatric Association, 1987) Axis II (SCID-II; Wittchen, Schramm, Zaudig, Mombour, & Unland, 1993). The 12-item self-report instrument consists of statements corresponding to symptoms of conduct disorder (0 = no, 1 = yes) according to the diagnostic criteria of the DSM-III-R.

Family/Marital Circumstances

For this risk domain, the young age of our participants required that we adapt the LSI-R operationalization. As Andrews and Bonta (2010a) suggested focusing on the family domain and not on the marital circumstances in the case of young offenders, the marital situation of our participants was not considered in our operationalization (cf. Andrews & Bonta, 2010a).

School/Work

Due to the young age of the participants and the fact that many participants dropped out of school, they also did not participate in the labor market. Therefore, we included only LSI-R items on education and omitted items on the work situation. As the German school system differs from North American school systems, we had to adapt this domain to the German school system (cf. Dahle et al., 2012).

Recidivism

We used official data from the German Central Criminal Register as recidivism criteria. Any new custodial sentence after release from prison was included in the general recidivism criterion. The violent recidivism criterion included custodial sentences for bodily injury, grievous bodily harm, robbery, or homicide/manslaughter. The date of recidivism was defined as the date of the new offense. For each individual, the follow-up period (time at risk) was defined as number of months between the date of release from prison and the date when the criminal files were requested (December 31, 2007). Due to the variation in the participants’ length of imprisonment and different assessment interview dates (over the course of the 3-year assessment period), the times at risk differed substantially among participants, ranging from 15.7 to 114.4 months (M = 77.2; SD = 15.0).

Statistical Analyses

In most prospective studies, it is not possible to avoid variations in the times at risk (Allison, 2010). These differences in time at risk make it improper to measure recidivism with merely bivariate outcomes (e.g., correlation coefficients, logistic regressions). Survival analyses are an appropriate statistical procedure to analyze these kinds of data sets (Allison, 2010; Brown, Amand, & Zamble, 2009; Grieger & Hosser, 2012), as they have the advantage of simultaneously taking into account whether an event (recidivism) occurred and the timing of the event (if it did occur). Furthermore, survival analyses can deal with unequal observation periods, which lead to randomly censored data. In our study, the variation in the time at risk was caused by a variety of factors such as the timing of the interviews (our assessment period amounted to 3 years) and the duration of imprisonment. While the timing of the assessment is not associated with the likelihood of an event (reoffense) to occur, causing noninformative censoring, the duration of imprisonment may be associated with risk of recidivism. This is because in Germany, in addition to the sentence received, the fact that inmates are found to be no longer a threat to themselves or the public (partially deduced from conduct in prison) influences the actual length of incarceration. If inmates are found to be no longer dangerous, the rest of their sentence can be placed on probation after they have served at least two thirds of their sentence. Thus, differences in the length of imprisonment may lead to informative censoring, which may bias results. As a solution to this problem, we followed Allison’s (2010) approach and included the duration of imprisonment as a covariate in the survival analyses. However, Allison (2010) states that “it can lead to computational difficulties if a large proportion of the observations is censored” (p. 13). Hence we chose the average time at risk (78 months) as a cut-off for the follow-up period. Thus, any reoffense that occurred 78 months or more after release was not examined in this study. As a consequence, the number of randomly censored cases was limited to 35.7% (violent recidivism) or 14.3% (general recidivism).

Predictive Validity for Recidivism

To ensure consistency with previous studies, we examined correlations of the Central Eight and recidivism (see Table 2). In addition, we calculated ROC analyses for the single scales (see Table 2). For reasons mentioned above, we conducted Cox regression analyses (a survival analytical statistical procedure) to examine the predictive validity of the Central Eight for recidivism more closely. We chose hierarchical Cox regressions entering the duration of imprisonment as a covariate in the first step to control for the effect of censoring.

Table 2:

ROC Analyses and Pearson Correlation Coefficients Between the Central Eight Risk Factors and Recidivism

	1	2	3	4	5	6	7	8	AUC Violent	AUC General
1. History of antisocial behavior									.59***	.59**
2. Antisocial personality pattern	.08*								.56*	.56*
3. Antisocial cognition	.05	.43***							.59***	.56*
4. Antisocial associates	.09*	.26***	.24***						.56**	.58**
5. Family	.12**	.23***	.07†	.12**					.55*	.53
6. School	.16***	.29***	.15***	.10*	.20***				.63***	.60***
7. Leisure/recreation	.11**	.13**	.03	.07	.07†	−.01			.55†	.51
8. Substance abuse	.12**	.27***	.21***	.17***	.14**	.00	.11**		.56*	.60***
Violent recidivism	.16***	.11**	.14***	.12**	.09*	.22***	.10*	.11*
General recidivism	.13**	.09	.10*	.14***	.06	.16***	.03	.16***

Note. ROC = receiver operating characteristic; AUC = area under curve. The correlations were all calculated for the whole sample (N = 589).

†

p ≤ .1. *p ≤ .05. **p ≤ .01. ***p ≤ .001.

Incremental Predictive Validity

To investigate the incremental predictive validity of the Big Four and Moderate Four risk factors for recidivism, we calculated three different models of Cox regression analyses for both recidivism criteria. In Model 1, we added all risk factors simultaneously in the second step, thus testing the predictive validity of a linear combination of all Central Eight risk factors after controlling for the duration of imprisonment. In Model 2, we included only the Big Four risk factors in the second step, and in Model 3 only the Moderate Four risk factors in the second step. We used likelihood ratio tests to examine the incremental predictive validity of the Big Four and Moderate Four risk factors (Field, 2009). The likelihood ratio statistic is defined by the subtraction of the −2 log-likelihood coefficient of the full model (Model 1) from the reduced model (Model 2). This difference is approximately chi-square distributed. Finally, we conducted backward stepwise Cox regression analyses to identify statistically redundant predictors for both recidivism criteria. Then, we calculated Model 4, which included only the statistically nonredundant predictors for each recidivism criterion and analogously conducted −2 log-likelihood tests.

Cross-Validation

To avoid an overfit of the model to the data, and to increase the generalizability of our results, we cross-validated the data, splitting the sample randomly into two equally sized subsamples (cf. Walters, 2011). One subsample served as the derivation sample (n = 244) and the other subsample (n = 245) was the cross-validation sample. Using the derivation sample, we conducted Cox regression analyses to detect the relative contributions of the risk factors for the prediction of recidivism, indicated by the beta weights. Then, we used these beta weights to form weighted sum scores for both recidivism criteria (cf. Walters, 2011). Analogously, we calculated sum scores with only the nonredundant variables that we had identified in backward stepwise Cox regression analyses for both forms of recidivism using the derivation sample. Then, we calculated the predictive accuracy of the weighted sum scores using the area under curve (AUC) statistic and Pearson correlation coefficients for the cross-validation sample. As the common risk assessment measures classify the offenders into low-, medium-, and high-risk offenders, we calculated cut-off values for the weighted sum scores that split the derivation sample into thirds. Then we applied the formula for the weighted sum score and the classification of risk level to the cross-validation sample and the entire sample. We conducted Kaplan–Meier survival analyses and log-rank tests to compare the recidivism rates of the low-, medium-, and high-risk offender groups.

Results

Within the 78-month observation period, 74.9% of the participants received another custodial sentence for any kind of crime (general recidivism) and 40.7% committed another violent offense that prompted a new custodial sentence (violent recidivism). The intercorrelations between the Central Eight risk factors, their correlations with recidivism, and ROC values for the single scales are indicated in Table 2. All risk factors correlated significantly with violent recidivism, and only family and leisure/recreation did not correlate significantly with general recidivism. The Central Eight risk factors were strongly intercorrelated. Leisure/recreation showed rather low intercorrelations with other variables, and the antisocial personality pattern showed the strongest correlations with the other risk factors.

In the next step, we tested the predictive validity of the whole set of risk factors for recidivism using Cox regression analyses. Table 3 provides an overview of the results of the Cox regression analyses. Both recidivism criteria could be significantly predicted by the Central Eight risk factors (Model 1). The reduced models included only the Big Four (Model 2) or the Moderate Four (Model 3) risk factors, yet both showed a significant model fit for both of the recidivism criteria. Descriptively, Model 3 showed a better data fit than Model 2, indicating greater predictive validity of the Moderate Four risk factors in comparison with the Big Four. Descriptively, the model fits were better for the violent recidivism criterion than for the general recidivism criterion as indicated by the higher chi-square and the lower −2 log-likelihood values. Across all models and both recidivism criteria, the “school” domain showed the highest predictive validity, such that offender groups that differ on this factor by a score of 1 were between 3.8 and 4.5 times more likely to reoffend with a violent offense in the follow-up period, and between 2.3 and 2.4 times more likely to be reincarcerated for any kind of crime committed after release (if the respective other risk factors are accounted for).

Table 3:

Results of Cox Regression Analyses Predicting Violent and General Recidivism

	Violent Recidivism						General Recidivism
	Model 1		Models 2/3		Model 4		Model 1		Models 2/3		Model 4
	β	HR	β	HR	β	HR	HR	β	β	HR	β	HR
Length of imprisonment	0.01	1.01†	0.01	1.01†	0.01	1.01†	0.00	1.00	0.00	1.00	0.00	1.00
Big Four
History of antisocial behavior	0.82	2.28*	1.18	3.27***	0.94	2.56**	0.48	1.62†	0.70	2.02**	0.48	1.62†
Antisocial personality	−0.83	0.44	0.24	1.27			−0.37	0.69	0.26	1.30
Antisocial cognition	1.41	4.09*	1.54	4.65*	1.40	4.05*	0.83	2.94†	0.91	2.48†
Antisocial associates	0.42	1.52	0.46	1.58†			0.42	1.52*	0.45	1.58*	0.46	1.58*
Moderate Four
Family	0.12	1.13	0.16	1.17			−0.08	0.92	−0.03	0.97
School	1.42	4.12***	1.49	4.45***	1.33	3.77***	0.82	2.27***	0.89	2.43***	0.79	2.27***
Leisure/recreation	0.25	1.28†	0.30	1.35*	0.25	1.28†	0.09	1.09	0.12	1.12
Substance abuse	0.39	1.47	0.55	1.73*			0.43	1.53*	0.58	1.79**	0.45	1.58*
Δχ²	60.88***		29.77*/46.47*		54.98***		46.42***		26.78*/33.52*		42.74***
−2 Log-likelihood^a	2883.61		2914.72/2898.02		2889.52		5098.29		5117.92/5111.18		5101.97

Note. β = beta weights; HR = hazard ratio. Δχ² is the change in the model fit after adding the variables in the second step. Length of imprisonment was entered in the first step of the hierarchical Cox regression analyses. In the second step, one of the models was entered into the Cox regression. Model 1 included all Central Eight risk factors (df = 8). Model 2 included the Big Four risk factors (df = 4). Model 3 included the Moderate Four risk factors (df = 4). Model 4 included the violent/general recidivism German Big Four risk factors (violent recidivism: school, history of antisocial behavior, antisocial cognition, and leisure/recreation; general recidivism: school, antisocial associates, substance abuse, and history of antisocial behavior).

Model fit index for the Cox regression analyses.

†

p ≤ .10. *p ≤ .05. **p ≤ .01. ***p < .000.

The backward stepwise Cox regressions identified two new constellations of the most important predictors for both forms of recidivism, the violent recidivism German Big Four and the general recidivism German Big Four. Both of the reduced models included school and history of antisocial behavior. In addition, the violent recidivism German Big Four included the factors antisocial cognition and leisure/recreation, and the general recidivism German Big Four included the factors antisocial associates and substance abuse. The results of Cox regression analyses conducted with these violent/general recidivism German Big Fours (Model 4) are indicated in Table 3.

To test our hypothesis that the combination of the Big Four and Moderate Four (i.e., all Central Eight risk factors) shows incremental predictive validity compared to the Big Four risk factors alone, we calculated likelihood ratio tests between Models 1 and 2. The results of the test revealed that the full model predicted recidivism significantly better than the reduced model (violent recidivism: Δ−2 log-likelihood = 31.11, df = 4, p < .001; general recidivism: Δ−2 log-likelihood = 19.63, df = 4, p < .001). This indicates that for both forms of recidivism, the Moderate Four risk factors showed incremental predictive validity.

Analogous likelihood ratio tests were conducted to analyze the incremental predictive validity of the risk factors that were not included in the violent/general recidivism German Big Fours, comparing Models 1 and 4 for the respective forms of recidivism. For both forms of recidivism, the full model (Model 1) did not predict recidivism significantly better than the reduced models (Model 4) with only the violent/general recidivism German Big Fours (violent recidivism: Δ−2 log-likelihood = 5.91, df = 4, p = .21; general recidivism: Δ−2 log-likelihood = 3.68, df = 4, p = .45). That is, risk factors that were not part of the violent/general recidivism German Big Fours did not show incremental predictive validity.

Finally, to test the validity of our findings, we conducted a cross-validation of the predictive validity. We randomly split our sample into two subsamples and calculated weighted sum scores of the Central Eight, the Big Four, the Moderate Four, and the violent/general recidivism German Big Fours using the derivation subsample. Subsequently, we calculated correlation coefficients and ROC curve analyses with the cross-validation sample. The results of the cross-validation are denoted in Table 4.

Table 4:

Cross-Validation of the Big Four, Moderate Four, the Central Eight, and Both of the Violent/General Recidivism German Big Four Risk Factors Using ROC Analyses and Correlation Coefficients

	Violent Recidivism						General Recidivism
	Entire Sample		Derivation Sample		Cross-Validation Sample		Entire Sample		Derivation Sample		Cross-Validation Sample
	r	AUC (95% CI)	r	AUC (95% CI)	r	AUC (95% CI)	r	AUC (95% CI)	r	AUC (95% CI)	r	AUC (95% CI)
Big Four	.21***	.62*** [.57, .67]	.24***	.63*** [.57, .70]	.18**	.60** [.54, .67]	.18***	.62*** [.57, .67]	.21***	.64*** [.57, .71]	.17**	.61** [.53, .69]
Moderate Four	.25***	.64*** [.60, .69]	.23***	.63*** [.57, .70]	.26***	.65*** [.58, .71]	.21***	.64*** [.58, .69]	.20**	.63*** [.56, .70]	.24***	.65*** [.57, .73]
Central Eight	.29***	.67*** [.63, 71]	.30***	.68*** [.62, .74]	.27***	.66*** [.60, .72]	.24***	.65*** [.60, .70]	.24***	.66*** [.59, .72]	.25***	.65*** [.57, .72]
Violent/general recidivism German Big Four	.29***	.67*** [.63, .72]	.29***	.68*** [.61, .74]	.28***	.67*** [.60, .73]	.24***	.65*** [.60, .70]	.24***	.66*** [.59, 73]	.25***	.64** [.57, .72]

Note. CI = confidence interval; ROC = receiver operating characteristic; AUC = area under curve. The violent recidivism German Big Four included school, antisocial cognition, leisure/recreation, and history of antisocial behavior. The general recidivism German Big Four included school, antisocial associates, substance abuse, and history of antisocial behavior. These risk factors were identified to have incremental predictive validity for their respective forms of recidivism. Weighted sum scores of the Central Eight, Big Four, Moderate Four, and violent/general recidivism German Big Fours were calculated using the algorithm from the derivation sample. Sample sizes were n = 294 (derivation sample), n = 295 (cross-validation sample), and N = 589 (entire sample).

p ≤ .01. ***p ≤ .001.

The composite score of the Central Eight showed moderate predictive validity for both forms of recidivism in the cross-validation sample (violent recidivism: r = .27, p < .001; AUC = .66, p < .001; general recidivism: r = .25, p < .001; AUC = .65, p < .001). The reduced models identified by the backward stepwise Cox regression analyses (violent/general recidivism German Big Fours) showed similar predictive validities (violent recidivism: r = .28, p < .001; AUC = .67, p < .001; general recidivism: r = .25, p < .001; AUC = .64, p < .001). Using Kaplan–Meier survival analyses, we compared the recidivism rates for the three offender risk groups (cf. Figure 1). The algorithm that classified offenders as low-, medium-, and high-risk in the derivation sample was utilized to classify offender types in the cross-validation sample. For both forms of recidivism, the offender types showed significantly different recidivism curves in the cross-validation subsample (violent recidivism: log-rank χ² = 17.31, df = 2, p < .001; general recidivism: log-rank χ² = 12.33, df = 2, p = .002). Again, the results of the reduced models using the violent/general recidivism German Big Fours showed comparable results (violent recidivism: log-rank χ² = 24.07, df = 2, p < .001; general recidivism: log-rank χ² = 12.38, df = 2, p < .001). The results of the Kaplan–Meier survival analyses depicting general and violent reoffending curves by each offender group for the entire sample are presented in Figure 1.

Figure 1:

Kaplan–Meier Analyses Predicting General and Violent Recidivism Using the Central Eight Versus Each Set of Violent/General Recidivism German Big Four Risk Factors

Discussion

Our prospective study investigated the predictive validity of the Central Eight risk factors for two forms of recidivism in a sample of German inmates of youth correctional facilities who were incarcerated for the first time. In particular, we were interested in the incremental predictive validity of the Moderate Four risk factors. Consistent with our hypothesis, the Central Eight risk factors predicted both forms of recidivism in survival analyses. Furthermore, our results indicate that the Moderate Four risk factors provide incremental predictive validity. The predictive validity of the weighted overall score in the cross-validation sample was comparable with the results reported by previous studies of juvenile offenders (Olver et al., 2009; Schmidt et al., 2011; Schmidt et al., 2005; Schwalbe, 2007). Thus, our study indicates that the risk factors identified by Andrews and Bonta (2010a) are also valid predictors for recidivism in German inmates of youth correctional facilities.

Another objective of our study was to examine the predictive validity of the single risk factors relative to each other in incremental predictive validity analyses. Inductively, we identified a set of the most predictive risk factors using backward stepwise Cox regression analyses, resulting in what we referred to as the violent recidivism German Big Four (school, history of antisocial behavior, antisocial cognition, and leisure/recreation) and the general recidivism German Big Four (school, antisocial associates, substance abuse, and history of antisocial behavior). We found that the inclusion of the whole set of the Central Eight risk factors did not improve the prediction of recidivism over inclusion of only the violent/general recidivism German Big Four risk factor constellations.

In interpreting these results, we have to take into consideration the high intercorrelations of the risk factors, which were anticipated by the authors of the model (Andrews & Bonta, 2010a). The fact that we did not find an incremental predictive effect of the risk factors that were not part of the violent/general Recidivism German Big Fours does not necessarily imply that they do not predict recidivism (cf. Table 2). However, the shared variance with the other, more predictive risk factors may have attenuated their predictive validity in multivariate analyses. A similar assumption was made by McGrath and Thompson (2012), who also found that especially the non-incremental “personality” and “family” domains were highly intercorrelated with the other risk factors. The fact that the antisocial personality pattern was positively related to recidivism on a bivariate level (cf. Table 2), but negatively skewed in Model 1 in the multivariate analyses (cf. Table 3), supports this assumption, as it may hint at suppression effects in the multivariate Cox regression analyses.

Interestingly, the four risk factors that were identified as the most predictive and incremental in McGrath and Thompson’s (2012) study were all included in either the violent recidivism German Big Four or the general recidivism German Big Four (or in the case of school, in both German Big Fours). The only factor that was incremental in our study but not in McGrath and Thompson’s study was leisure/recreation. Despite the low AUC value and bivariate correlation with recidivism (cf. Table 2), this domain was included in the violent recidivism German Big Four. Again, the intercorrelations with the other risk factors merit consideration. The fact that this domain is statistically not as strongly intercorrelated with the other domains may explain why it provided incremental predictive validity.

Contrary to the predictions of the RNR model, the Moderate Four risk factors predicted recidivism better than the Big Four (cf. Tables 3 and 4). To understand this finding, we have to take a closer look at the predictive validities of the single scales. It may be assumed that the school domain in particular explains a large portion of the variance in the dependent variable (cf. Tables 2 and 3). The fact that this risk factor also showed the highest predictive validity in the violent and general recidivism German Big Fours (cf. Table 3) underscores this assumption. Lookwood, Nally, Ho, and Knutson (2012) also showed that a low level of education is a strong predictor for recidivism. Given that “school” is one of the Moderate Four, this finding might explain why the Moderate Four predicted recidivism better than the Big Four risk factors in our sample of youth correctional facility inmates.

Another explanation for the weaker predictive validity of the Big Four is the relatively weak predictive validity of the “history of antisocial behavior” domain (cf. Table 2), which was found to be the best predictor of recidivism in other studies (e.g., Girard & Wormith, 2004). The weaker finding in the present study may be due to range restrictions in the history of antisocial behavior domain, which are caused by the selection of participants. As our sample consisted only of offenders who were incarcerated for the first time, they did not differ substantially in some of the variables assessed by the history of antisocial behavior domain (e.g., number of prior crimes). The homogeneity of the sample in terms of history of antisocial behavior may have led to an attenuation of the predictive validity of this domain. Therefore, future studies with similar samples should find an operationalization that is tailored to this offender group.

Another noteworthy finding of our study was that, in contrast to other studies (Andrews, Bonta, & Wormith, 2006; Girard & Wormith, 2004; Olver et al., 2009), violent recidivism was descriptively better predicted than general recidivism in our sample. One possible cause of this finding is a variance restriction in the dependent variable (general recidivism) due to the high general recidivism rate in our study. The high recidivism rates in our sample correspond to results from a German representative sample (Jehle, Albrecht, Hohmann-Fricke, & Tetal, 2010), and they are also comparable with recidivism rates found in other European high-risk samples (e.g., Plattner et al., 2009; Wartner, el Harbachi, & van der Laan, 2005). Only 25.1% of the participants in our study did not recidivate with any kind of crime. Therefore, the variance left to explain was restricted compared with the violent recidivism criterion (59.3% nonrecidivists vs. 40.7% recidivists). On average, studies reporting a higher predictive validity for general recidivism than for violent recidivism had shorter observation periods (around 30 months); the variance restriction in these studies (Girard & Wormith, 2004; Olver et al., 2009) with restricted observation periods may have impacted violent recidivism but not general recidivism. For example, using a sample of 630 youth offenders, Girard and Wormith (2004) found a better predictive validity for general than for violent recidivism in the 31-month follow-up. Furthermore, Olver et al. (2009) reported in their meta-analysis (average observation period of the studies included was 29.1 months) that predictive validity was better for general recidivism than for violent recidivism. However, a recent meta-regression of different risk assessment tools confirms that, in general, violent recidivism can be better predicted than general recidivism (Singh, Grann, & Fazel, 2011). Further underscoring our assumption, Dahle (2006) found in a German sample that the predictive validity of the LSI-R for violent recidivism (low base rate at the beginning of the study) increased with time, while the predictive validity for general recidivism decreased with time.

Strengths and Limitations

Our study has several strengths. We address the question of incremental predictive validity of risk factors for recidivism, which is considered to be an important issue that has been neglected in forensic research (Walters, 2011). Our study had a prospective design, and our observation period was more than twice as long as the average for studies in young adult offenders (Olver et al., 2009; Schwalbe, 2008). In addition, our study is the first to address the question of incremental predictive validity of the Central Eight risk factors in a sample of German inmates of youth correctional facilities. For the prediction of recidivism, we used survival analyses, which are considered to be more accurate and adequate than conventional methods (Allison, 2010). Furthermore, we cross-validated the overall prediction, and our sample was a high-risk institutional sample. Both factors typically diminish the predictive validity of risk assessment instruments (Schwalbe, 2007, 2008).

However, our study has several limitations. Our sample consisted of German offenders who were incarcerated in a youth correctional facility for the first time, which represents only a subgroup of the whole offender population. Whether results might generalize to other populations must be carefully considered (Takahashi et al., 2013), because a number of issues make generalization imprudent. For instance, the age group in our sample was diverse, ranging from 14.5 to 25.4 years, because we selected the offenders according to their institutional history (incarcerated for the first time) rather than variables like age or offender group, which are popular selection criteria for other studies. Considering this wide age range, developmental differences within our sample may have had a moderating effect on the relationship between the offenders’ criminogenic needs and recidivism rates.

Furthermore, as discussed, our approach might reduce the significance of the “history of antisocial behavior” domain, effectively enhancing the predictive validity of the dynamic risk factors in comparison. Future studies should investigate which of the two approaches yields better results for the prediction of recidivism inmates of youth correctional facilities: classifying the youth offenders according to their age group or according to their imprisonment history. The selection of redundant variables might yield different results in other samples, and hence the review and selection of potentially redundant variables should be conducted in different samples.

Due to the large amount of information we collected, we conducted several interviews. A limitation of our study concerns the fact that the information on the Central Eight risk factors was collected at different times during imprisonment. While for some items (e.g., “Have you ever been supervised by the child protection services?”) the answers may not have been different at another point of time, some changes might be expected for other questions or information (e.g., the factor antisocial associates).

We assessed recidivism using official data. Even though this approach has several advantages (e.g., no systematic dropout due to missing values on the recidivism criteria), it should be noted that there might be a large number of undetected crimes that were not captured by our recidivism criteria. Another limitation concerns the use of self-report data. Most of the information on the Central Eight risk factors was taken from interviews. Thus, the participants might have responded in a socially desirable way.

To provide for a consistent operationalization of the variables, we omitted some information assessed by the LSI-R (e.g., marital status, information on the work situation) that did not apply to the younger offenders. However, this information might have proven fruitful in the assessment of older participants. Another limitation of operationalization in this study concerns the family domain. Some of the criteria we assessed rather represent a history of extremely poor supervision and relationship with the parents (e.g., living in foster care, supervision by the child protection services). Thus, the (rather poor) results of the predictive validity of this domain have to be interpreted with caution. Future studies should focus more on the quality of relationship with family members, parenting style, and/or monitoring and supervision.

Conclusion

Overall, our results provide support for the importance of most of the risk factors in Andrews and Bonta’s (2010a) RNR model and its predictive validity for the German youth correctional system. On a bivariate level, all risk factors predicted at least violent recidivism and thus our study suggests that they are all relevant for clinical use. However, the results of our incremental predictive validity analyses show that the hierarchy in the predictive validity of the risk factors was different in our sample: The Moderate Four risk factors were descriptively more predictive of recidivism than the Big Four. In fact, we identified more predictive constellations of risk factors, the “violent recidivism German Big Four” (school, history of antisocial behavior, antisocial cognition, and leisure/recreation) and the “general recidivism German Big Four” (school, antisocial associates, substance abuse, and history of antisocial behavior). Given the limited treatment possibilities in prison (due to short sentences and/or shortage of prison staff), the results of our analyses may hint at risk factors that are the most promising to achieve a reduction in the recidivism rates in German youth correctional facility inmates. In addition, considering the high intercorrelations of the risk factors (in particular, the antisocial personality pattern, cf. Table 2), it may be expected that changes in one risk domain also transfer change onto other risk factors. This should be investigated by future studies.

As McGrath and Thompson (2012) pointed out, studies on the relative predictive validity of single scales can help to clarify and, if necessary, adapt the weighting of the single domains in the sum score of a measure for specific samples. As research in Germany has started to investigate the predictive validity of the LSI-R (Dahle, 2006) only recently, the results of our study may be considered when weighing the scales (e.g., by including more items for more predictive domains) in the process of adapting the youth version of the LSI to the German youth correctional system.

Despite the assumptions of the RNR model that classifies school as one of the Moderate Four, school was the most predictive risk factor across all models and analyses. This strong result stresses the need of schooling programs in prison and the need/importance of including participation in schooling programs as part of the conditions of parole.

Furthermore, since dynamic risk factors (school, antisocial cognition, antisocial associates, substance abuse, and leisure/recreation) were found to have incremental predictive validity, our findings corroborate the importance of treatment programs addressing the criminogenic needs of the offenders. Finally, the results of this study could be unique to this sample of German youth, and so there remains a need for more research on the incremental predictive validity of risk factors for recidivism in different samples.

Footnotes

Acknowledgements

The authors would like to thank Amanda C. Jones for her help refining the English in this article.

This research was supported by grants from the German Research Foundation (Grant HO 2553/1-3).

Lena Grieger received her diploma in psychology from the University of Bonn, Germany. Currently, she is a PhD candidate at the Technical University of Braunschweig, Germany. Her research interests focus on risk factors for recidivism and on offender treatment.

Daniela Hosser, PhD, is the head of the Department of Developmental, Personality and Forensic Psychology at the Technical University of Braunschweig, and former chair of the psychology and law division of the German Psychological Society. Until 2009, she worked as the assistant director of the Criminological Research Institute of Lower Saxony. Her research interests include life course crime, prison research, offender treatment, and victimology.

References

Allison

P. D.

(2010). Survival analysis using SAS: A practical guide (2nd ed.). Cary, NC: SAS Institute.

American Psychiatric Association. (1987). Diagnostic and statistical manual of mental disorders (3rd ed. rev.). Washington, DC: Author.

Andrews

D. A.

Bonta

(1995). Level of Service Inventory–Revised. Ontario, Canada. Multi-Health Systems.

Andrews

D. A.

Bonta

(2010a). The psychology of criminal conduct (5th ed.). New Providence, NJ: LexisNexis.

Andrews

D. A.

Bonta

(2010b). Rehabilitating criminal justice policy. Psychology, Public Policy, and Law, 16, 39-55. doi:10.1037/a0018362

Andrews

D. A.

Bonta

Wormith

J. S.

(2006). The recent past and near future of risk and/or need assessment. Crime & Delinquency, 52, 7-27. doi:0.1177/0011128705281756

Andrews

D. A.

Bonta

Wormith

J. S.

(2011). The risk-need-responsivity (RNR) model: Does adding the good lives model contribute to effective crime prevention? Criminal Justice and Behavior, 38, 735-755. doi:10.1177/0093854811406356

Andrews

D. A.

Guzzo

Raynor

Rowe

R. C.

Rettinger

L. J.

Brews

Wormith

J. S.

(2012). Are the major risk/need factors predictive of both female and male reoffending? A test with the eight domains of the Level of Service/Case Management Inventory. International Journal of Offender Therapy and Comparative Criminology, 56, 113-133. doi:10.1177/0306624X10395716

Brown

S. L.

Amand

M. D.

St. Zamble

(2009). The dynamic prediction of criminal recidivism: A three-wave prospective study. Law and Human Behavior, 33, 25-45. doi:10.1007/s10979-008-9139-7

10.

Campbell

M. A.

French

Gendreau

(2011). The prediction of violence in adult offenders: A meta-analytic comparison of instruments and methods of assessment. Criminal Justice and Behavior, 36, 567-590. doi:10.1177/0093854809333610

11.

Caspi

Begg

Dickson

Harrington

Langley

Moffitt

T. E.

Silva

P. A.

(1997). Personality differences predict health-risk behaviors in young adulthood: Evidence from a longitudinal study. Journal of Personality and Social Psychology, 73, 1052-1063. doi:10.1037/0022-3514.73.5.1052

12.

Catchpole

R. E. H.

Gretton

H. M.

(2003). The predictive validity of risk assessment with violent young offenders: A 1-year examination of criminal outcome. Criminal Justice and Behavior, 30, 688-708. doi:10.1177/0093854803256455

13.

Dahle

K.-P.

(2006). Strengths and limitations of actuarial prediction of criminal reoffence in a German prison sample: A comparative study of LSI-R, HCR-20 and PCL-R. International Journal of Law and Psychiatry, 29, 431-442. doi:10.1016/j.ijlp.2006.03.001

14.

Dahle

K.-P.

Harwardt

Schneider-Njepel

(2012). Inventar zur Einschätzung des Rückfallrisikos und des Betreuungs- und Behandlungsbedarfs von Straftätern: LSI-R, deutsche Version des Level of Service Inventory–Revised nach Don Andrews und James Bonta [Assessment inventory to estimate risk of recidivism and supervision and treatment needs of offenders: LSI-R, German version of the Level of Service Inventory–Revised by Don Andrews and James Bonta]. Göttigen, Germany: Hogrefe.

15.

Dünkel

Geng

(2007). Rechtstatsächliche Befunde zum Jugendstrafvollzug in Deutschland [Legal facts and research results on the German juvenile penal system]. Forum Strafvollzug, 2, 65-80.

16.

Field

(2009). Discovering statistics using SPSS (3rd ed.). London, UK: SAGE.

17.

Gendreau

Little

Goggin

(1996). A meta-analysis of the predictors of adult offender recidivism: What works! Criminology, 34, 575-604. doi:10.1111/j.1745-9125.1996.tb01220.x

18.

Girard

Wormith

J. S.

(2004). The predictive validity of the Level of Service Inventory–Ontario Revision on general and violent recidivism among various offender groups. Criminal Justice and Behavior, 31, 150-181. doi:10.1177/0093854803261335

19.

Grieger

Hosser

(2012). Attention deficit hyperactivity disorder does not predict criminal recidivism in young adult offenders: Results from a perspective study. International Journal of Law and Psychiatry, 35, 27-34. doi:10.1016/j.ijlp.2011.11.005

20.

Hoge

R. D.

Andrews

D. A.

(1995). Australian adaptation of the Youth Level of Service/Case Management Inventory. North Tonowanda, NY: Multi-Health Systems.

21.

Hoge

R. D.

Andrews

D. A.

(2002). Youth Level of Service/Case Management Inventory (YLS/CMI). Toronto, Canada: Multi-Health Systems.

22.

Hsu

C.-I.

Caputi

Byrne

M. K.

(2009). The Level of Service Inventory–Revised. A useful risk assessment measure for Australian offenders? Criminal Justice and Behavior, 36, 728-740. doi:10.1177/0093854809335409

23.

Jehle

J.-M.

Albrecht

H.-J.

Hohmann-Fricke

Tetal

(2010). Legalbewährung nach strafrechtlichen Sanktionen. Eine bundesweite Rückfalluntersuchung 2004 bis 2007 [Legal probation following criminal sanctions. A national survey of recidivism from 2004 to 2007]. Berlin, Germany: Bundesministerium der Justiz.

24.

Krueger

R. F.

Schmutte

P. S.

Caspi

Moffitt

T. E.

Campbell

Silva

P. A.

(1994). Personality traits are linked to crime among men and women: Evidence from a birth cohort. Journal of Abnormal Psychology, 103, 328-338. doi:10.1037/0021-843X.103.2.328

25.

Lookwood

Nally

J. M.

Knutson

(2012). The effect of correctional education on post-release employment and recidivism: A 5-year follow-up study in the state of Indiana. Crime & Delinquency, 58, 380-396. doi:10.1177/0011128712441695

26.

Marshall

Egan

English

Jones

R. M.

(2006). Relative validity of psychopathy versus risk/needs-based assessments in the prediction of adolescent offending behavior. Legal and Criminological Psychology, 11, 197-210. doi:10.1348/135532505X68719

27.

McGrath

Thompson

A. P.

(2012). The relative predictive validity of the static and dynamic domain scores in risk-need assessment of juvenile offenders. Criminal Justice and Behavior, 39, 250-263. doi:10.1177/0093854811431917

28.

Olver

M. E.

Stockdale

K. C.

Wong

S. C. P.

(2012). Short and long-term prediction of recidivism using the Youth Level of Service/Case Management Inventory in a sample of serious young offenders. Law and Human Behavior, 36, 331-334. doi:10.1037/h0093927

29.

Olver

M. E.

Stockdale

K. C.

Wormith

J. S.

(2009). Risk assessment with young offenders. A meta-analysis of three assessment measures. Criminal Justice and Behavior, 36, 329-353. doi:10.1177/0093854809331457

30.

Onifade

Davidson

Campbell

Turke

Malinowski

Turner

(2008). Predicting recidivism in probationers with the Youth Level of Service Case Management Inventory (YLS/CMI). Criminal Justice and Behavior, 35, 474-483. doi:10.1177/0093854807313427

31.

Ortmann

(1987). Resozialisierung im Strafvollzug [Resocialization in the penal system]. Freiburg im Breisgau, Germany: Max-Planck-Institut für ausländisches und internationales Strafrecht.

32.

Plattner

Steiner

The

S. S. L.

Kraemer

H. C.

Bauer

S. M.

Kindler

. . . Feucht

(2009). Sex-specific predictors of criminal recidivism in a representative sample of incarcerated youth. Comprehensive Psychiatry, 50, 400-407. doi:10.1016/j.comppsych.2008.09.014

33.

Russell

D. W.

Peplau

L. A.

Cutrona

C. E.

(1980). The revised UCLA Loneliness Scale: Concurrent and discriminant validity evidence. Journal of Personality and Social Psychology, 39, 472-480.

34.

Schmidt

Campbell

M. A.

Houlding

(2011). Comparative analysis of the YLS/CMI, SAVRY, and PCL:YV in adolescent offenders: A 10-year follow-up into adulthood. Youth Violence and Juvenile Justice, 9, 23-42. doi:10.1177/1541204010371793

35.

Schmidt

Hoge

R. D.

Gomes

(2005). Reliability and validity analysis of the Youth Level of Service/Case Management Inventory. Criminal Justice and Behavior, 32, 329-344. doi:10.1177/0093854804274373

36.

Schwalbe

C. S.

(2007). Risk assessment for juvenile justice: A meta-analysis. Law and Human Behavior, 31, 449-462. doi:10.1007/s10979-006-9071-7

37.

Schwalbe

C. S.

(2008). A meta-analysis of juvenile justice risk assessment instruments: Predictive validity by gender. Criminal Justice and Behavior, 35, 1367-1381. doi:10.1177/0093854808324377

38.

Seitz

(1983). Zur Struktur und Erfassung der Persönlichkeit von Inhaftierten- am Beispiel eines inhaftierungsadäquaten Persönlichkeitsfragebogens [On the structure and assessment of personality in offenders—using the example of an adequate personality questionnaire for the prison setting]. Zeitschrift für Differentielle und Diagnostische Psychologie, 4, 261-281.

39.

Singh

J. P.

Fazel

(2010). Forensic risk assessment: A metareview. Criminal Justice and Behavior, 37, 965-988. doi:10.1177/0093854810374274

40.

Singh

J. P.

Grann

Fazel

(2011). A comparative study of risk assessment tools: A systematic review and metaregression analysis of 68 studies involving 25,980 participants. Clinical Psychology Review, 31, 499-513. doi:10.1016/j.cpr.2010.11.009

41.

Smith

Cullen

F. T.

Latessa

(2009). Can 14,737 women be wrong? A meta-analysis of the LSI-R and recidivism for female offenders. Criminology & Public Policy, 8, 183-208. doi:10.1111/j.1745-9133.2009.00551.x

42.

Takahashi

Mori

Kroner

D. G.

(2013, June 17). A cross-validation of the Youth Level of Service/Case Management Inventory (YLS/CMI) among Japanese juvenile offenders. Law and Human Behavior. Advance online publication. doi:10.1037/lhb0000029

43.

Tellegen

(1982). Brief manual for the Multidimensional Personality Questionnaire. Minneapolis: University of Minnesota Press.

44.

Thompson

A. P.

McGrath

(2012). Subgroup differences and implications for contemporary risk-need assessment with juvenile offenders. Law and Human Behavior, 36, 345-355. doi:10.1037/h0093930

45.

Van der Put

C. E.

Stams

G. J. J. M.

Hoeve

Dekovic

Spanjaard

H. J. M.

van der Laan

P. H.

Barnoski

R. P.

(2011). Changes in the relative importance of dynamic risk factors for recidivism during adolescence. International Journal of Offender Therapy and Comparative Criminology, 56, 296-316. doi:10.1177/0306624X11398462

46.

Vose

Cullen

F. T.

Smith

(2008). The empirical status of the Level of Service Inventory. Federal Probation, 72, 22-29.

47.

Vose

Smith

Cullen

F. T.

(2013). Predictive validity and the impact of change in total LSI-R score on recidivism. Criminal Justice and Behavior, 40, 1383-1396.

48.

Vose

Lowenkamp

C. T.

Smith

Cullen

F. T.

(2009). Gender and the predictive validity of the LSI-R. A study of parolees and probationers. Journal of Contemporary Criminal Justice, 25, 469-471. doi:10.1177/1043986209344797

49.

Walkenhorst

(2010). Jugendstrafvollzug [The Juvenile penal system]. Politik und Zeitgeschichte, 7, 15-21.

50.

Walters

G. D.

(2011). Taking the next step: Combining incrementally valid indicators to improve recidivism prediction. Assessment, 18, 227-233. doi:10.1177/1073191110397484

51.

Ward

Yates

P. M.

Willis

G. M.

(2012). The good lives model and the risk need responsivity model. A critical response to Andrews, Bonta, and Wormith (2011). Criminal Justice and Behavior, 39, 94-110. doi:10.1177/0093854811426085

52.

Wartner

B. S. J.

el Harbachi

van der Laan

A. M.

(2005). Jong vast: Een cijfermatig overzicht van de strafrechtelijke recidive van ex-pupillen van justitiële jeugdinrichtingen [Youth in detention: Prevalence of recidivism in juvenile justice institutions]. The Hague, Netherlands: Wetenschappelijk Onderzoek-en Documentatiecentrum.

53.

Welsh

J. L.

Schmidt

McKinnon

Chattha

H. K.

Meyers

J. R.

(2008). A comparative study of adolescent risk assessment instruments: Predictive and incremental validity. Assessment, 15, 104-115. doi:10.1177/1073191107307966

54.

Wittchen

H.-U.

Schramm

Zaudig

Mombour

Unland

(1993). SKID-II. Strukturiertes Klinisches Interview für DSM-III-R Achse II (Persönlichkeitsstörungen) [SCID-II. Structured clinical interview for DSM-III-R Axis II (personality disorders)]. Weinheim, Germany: Beltz Test.