Better than clairvoyance? Exploratory results on predictors of regular release, substance use and re-delinquency after forensic addiction treatment

Abstract

Forensic addiction treatment (FAT) aims to reduce the risk of offences originating in substance use. Hence, reoffending is a key outcome criterion besides therapy completion and substance use. Previous research on success prognoses resulted in heterogeneous sets of predictors. This is the first study to exploratorily investigate three types of predictors and to compare their predictive values on three outcome criteria by using a coherent, multi-centric database. Data from 4,201 patients released from German FAT were used, testing 182 variables. Stepwise logistic regression analyses were computed. Resulting predictor sets were widely heterogeneous. While regular release could satisfactorily be predicted by historical variables only, variance explanation was much lower for substance use and re-delinquency. Including treatment-related and post-treatment variables raised variance explanation up to R² = .46 (p < .001). Success should be seen as a dynamic process that is strongly influenced by alterable variables taking effect during and after FAT.

Introduction

Substance use is related to delinquent behaviour in complex and reciprocal ways (e.g. (1–3)). Criminological research has identified substance use as an important risk factor for delinquent behaviour (i.e. any form of criminal conduct, regardless of whether it is on record, charged or convicted; for a recent meta-review, see (4)). The development of a substance use disorder (SUD) mediates and boosts this impact of substance use on delinquency, especially on offences committed to obtain drugs or to finance purchasing (5). In turn, among incarcerated offender populations, the prevalence of substance use and related disorders is elevated (6,7) and many offences leading to imprisonment can be attributed to the influence of psychoactive substances (e.g. in Canada, up to 42% – see (8)). Therefore, the societal burden and criminological importance of SUD are relevant topics for both the legal and forensic treatment systems (9).

The German Criminal Code (Strafgesetzbuch, StGB) provides one of the few jurisdictional sanctions specifically designed for addicted offenders (10). Section 64 StGB allows courts to order inpatient forensic addiction treatment (FAT). See Tomlin and colleagues (16) for a comprehensive description of the German legal regulations concerning FAT. If a person commits an offence as a result of a substance use disorder, courts shall place the person in a specialised forensic hospital. One of the central prerequisites for the application of this time- and cost-intensive legal measure is a positive treatment prognosis on an individual level. In the wording of the law:

Such order is only to be made if there is a sufficiently reasonable prospect that the person can be cured […] by way of placement in an addiction treatment facility or that a relapse into addictive behaviour and the commission of serious unlawful acts […] can be prevented for a substantial period of time. (11)

In 2023, this legal text was reworded and the requirements tightened. However, all of the patients examined in this study were accommodated and released under the hereby mentioned previous version.

In criminal proceedings, the court, advised by expert witnesses, must therefore determine whether such a prospect of successful therapy exists or not.

Outcome predictors of forensic addiction treatment

Against this background, it is understandable why, particularly in Germany, various studies on predictors of success of FAT can be found (12–15). The aim of these studies was usually to identify individual or sets of background (i.e. variables that indicate the clinical or legal history), sociodemographic or sometimes setting variables that could be used to predict FAT success. Only recently was a comprehensive review published with the aim of making the research findings from Germany accessible to an English-speaking readership by Tomlin et al. (16).

Predicting the release mode

In most studies, treatment success was defined by the mode of release (i.e. whether the therapy is completed regularly or discontinued prematurely). This can be determined directly and easily from routinely collected data. However, this criterion says nothing about whether the superordinate purpose of the treatment was achieved, namely the long-term improvement of the SUD and the prevention of new offences. As a result of over 25 years of research, the following risk factors for premature termination can be summarised (13): comorbid personality disorder (especially psychopathy), criminal history, early or very late onset of SUD, violent offences, low level of education and lack of occupational training or experience. Inconsistent findings pointed to the variables of early onset of delinquent behaviour, adverse childhood experiences and familiar background as well as most psychological constructs (such as the Big Five personality traits). Multivariate studies could explain 17–21% of outcome variance with various sets of patient-related predictors (17–20). If data on the treatment process were integrated (e.g. incidences during treatment), variance explanation rose to 26% (21). Furthermore, if setting variables were considered (e.g. in which hospital the treatment took place), 32% of the variance could be explained (18). Scheuschner et al. (15) recently published a review on such treatment-related predictors.

In addition (and from outside the forensic context), SUD severity and posttraumatic stress disorder (22), bad mood, stress and comorbid psychiatric problems (23), a history of psychiatric treatments in the past (24) and a lower income (25) turned out to be relevant risk factors for premature termination of addiction treatment.

Predicting re-delinquency

Fewer studies focused on reoffending (i.e. committing offences that become a matter of formal record) or re-delinquency as a wider concept (i.e. any form of delinquent behaviour regardless of whether it became a matter of formal record, indictment or conviction) after FAT and tried to identify predictors linked to this distal outcome criteria. To summarise German research, motivational, functional and social resources were related to courses without illegal behaviour, while criminal history and specificities of the index offence made re-delinquency more likely (14). Two multivariate studies used a survival analytic design to identify sets of person- and setting-related predictors (26,27). While setting indicators turned out to lack predictive value, the analyses resulted in widely heterogeneous sets of patient-related risk factors. Walde and Lungwitz (28) and Tomlin et al. (16) underline the importance of re-delinquency related risk and protective factors that refer to the follow-up period (i.e. variables that emerged only after release from the forensic hospital), especially continued substance use after treatment (e.g. (29)).

In an international context, Weber and Lynch (30) additionally highlighted the role of adverse childhood experiences on the reoffending rates of incarcerated adolescents with substance use problems, mediated by the severity of their substance use. Correspondingly, Schalast (31) identified experiences of intra-familiar violence as a risk factor in a German FAT sample. In contrast, Hartl (32) reported provisional evidence that abnormalities in childhood/adolescence (such as violence or addiction in the family of origin) could even serve as protective factors.

Predicting substance use

Although the need to take a differentiated view of the success of forensic treatments has increasingly been recognised and people who provide or use forensic treatment services emphasise a broader range of processes and outcomes (33), to our knowledge, little research from within the forensic context has dealt with predictors of substance use after FAT: Pfaff (29) identified statistical trends for higher intelligence, long prison experience and a long time since first offence as protective factors; Bezzel (34) found living in family or with spouse, a medium IQ-value, abstinence during treatment and any prior prison experience as protective factors. In that study, the relationship between substance use and re-delinquency was evident, but lower than expected (34).

Outcome studies from outside the forensic context underline the importance of personality disorders (35), low resilience and lack of employment (36), young age, co-occurring mental disorder, premature therapy termination and setting variables (37) as risk factors of early and numerous relapses. Accordingly, a recent systematic review on alcohol-related disorders lists psychiatric co-morbidity, symptom severity, craving, use of other substances, health and social factors as relevant predictors for relapses (38).

Goal and research questions

The present study adopts the goal of identifying predictors of successful FAT. The question of how to effectively predict the outcome of FAT is not only relevant for the specific German legal framework, but also for various other jurisdictions with sanctions tailored to addressing SUD (e.g. drug treatment courts in the United States) and for forensic treatment in general, as many psychiatric patients are comorbidly diagnosed with SUD, which is then a risk factor for treatment of the main psychiatric diagnosis itself (39).

In addition to most of the previous research investigating treatment completion and reoffending, we add substance use after FAT as a third outcome criteria (we thus follow Tomlin et al. (16) in their recent review). Further, we expand the established field of background (i.e. historical and retrospective) and treatment variables as potential predictors and include follow-up information referring to the period after release. To our knowledge, this is the first study to investigate the role of three groups of potential predictors on three relevant outcome criteria by using a coherent, multi-centric data base. For this reason, we chose an exploratory approach and abstain from explicit hypotheses, although preliminary findings on individual predictors are available.

The following research questions are examined:

1. Can distal FAT outcome criteria also be predicted on the basis of historical and background information?

2. What added value does the inclusion of treatment and follow-up information provide for the prediction of outcome criteria?

3. Do different prediction models for different outcome criteria lead to comparable lists of predictors?

4. Are the final predictor sets suitable for the creation of a prognostic tool for predicting treatment success in FAT?

Method

Data collection and procedure

As part of the forensic quality management of the 14 forensic hospitals situated in the German federal state of Bavaria, basic data as well as parameters of the process and success of therapy have been routinely collected annually since 2010. All forensic inpatients who have been treated for at least six weeks are included. Data collection takes place once at the time of discharge (t₀) and – if the patient has been released on parole after a successful therapy – after one year (t₁ – SD = 1.3 months). The case-specific data set is compiled by trained members of the local hospital staff and is based upon file information in addition to interviews with the patient, and/or the expert knowledge of staff members that are engaged in inpatient treatment of the individual patient (for t₀). At t₁, file information from the forensic aftercare institution is taken into account and – again – an interview with the former patient and/or an interview with the aftercare staff is conducted. Data is collected at both points in time using standardised guidelines and forms (see below).

The patient interviews are conducted by trained local staff members who are not a part of the treatment personnel and begin with a detailed written and verbal explanation of the process and purpose of the data collection. Participation is not remunerated and is voluntary. Only if the patient gives informed and written consent, is the pseudonymised data set transmitted from the clinic to the Institute for Quality Management in the Forensic Treatment System in Bavaria (IfQM), which evaluates and analyses the data centrally. For pseudonymisation, a fictitious number is assigned by the treating hospital in order to be able to assign the different measurement times to the same person.

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2013. All procedures involving human subjects were approved by the Ethics Committee at University of Regensburg (Vote No. 01/082; 2009-10-08).

Sample characteristics

The present analysis is based on 4,201 patients released from FAT between 2010 and 2020 in the federal state of Bavaria (regardless of whether they were released on parole or transferred to prison) who agreed to have their data collected and transferred at t₀. This corresponds to a participation rate of 52%. At t₁, 2,391 data sets were available. The notable difference between t₀ and t₁ is due to 1,369 patients who were transferred to prison after premature termination of forensic treatment (33% of the t₀-sample) and another 441 who refused to participate at t₁ or could not be interviewed because of organisational problems (10%; i.e. they had moved to an unknown address or did not respond when contacted).

At discharge (t₀), patients had been treated for an average of 22.7 months (SD = 10.9) and were 34.9 years old (SD = 9.1). Using ICD-10 (still the standard classification system in Germany), the most common diagnostic groups were multiple substance use disorders (n = 1,743, 42%), followed by alcohol-related (n = 841, 20%), and stimulant-related disorders (n = 699, 17%). Opioids (n = 548, 13%), cannabis (n = 336, 8%) and sedatives (n = 6, 0.1%) played a minor role. Only 28 patients (0.7%) did not have any substance use disorder as their primary diagnosis. This may seem counterintuitive at first sight, but is legally permissible: the required connection between SUD and risk of delinquency can also be present in people who, for example, have a clinically leading personality disorder. Just under half of the patients committed a drug offence (i.e. possession, dealing or trafficking of illegal drugs) as their primary offence (n = 2,067, 49%), followed by violent physical assault (n = 737, 18%), property crimes (n = 616, 15%) and robbery (n = 391, 9%). All other offence groups had shares of less than 3%.

The patients were treated in the 14 Bavarian forensic clinics, all of which pursue a multimodal approach. This generally includes group and individual therapy, sports, occupational and creative therapy and, if necessary, drug treatment. The wards are usually conceptualised as a therapeutic community.

Instrument and potential predictors

The data were collected by means of a questionnaire developed by the IFQM, which has been used for follow-up surveys in Bavarian forensic clinics since 2010 (described in detail by Probst and colleagues (40)). The aim of the developmental process was to create an exhaustive list of variables covering not only basic patient data but information on various levels and aspects of treatment success. T₀ (discharge) variables include historical and background information on sociodemographics, diagnoses and offences (we further refer to these variables as the historical set), as well as information on setting, the course of inpatient treatment and type of discharge (treatment set). With t₁, data on legal probation and information on living situation, health status, employment status, financial background and renewed treatment in and outside the forensic system are added, if applicable (follow-up set). In total, the pseudonymised data set comprises over 450 (sub-)variables. For the present study, we selected 182 variables as potential predictors of treatment outcomes: 81 belonged to the historical set, 44 to the treatment set and 57 to the follow-up set (see Supplement 1). The other (sub-)variables were omitted as they cover information not relevant to the present research questions (mainly administrative data or background information on the treatment trajectory and the process of data collection, e.g. the nomination of the ward, year of admission, the professional training of the interviewed expert).

Outcome variables

Three dichotomous treatment outcomes were investigated. The release mode after FAT served as the first outcome and was excerpted from the patients’ routine documentation. Either the patient completed therapy and was regularly released on parole (i.e. positive outcome) or treatment was terminated prematurely resulting mostly in a transfer to prison (i.e. negative outcome). In contrast to non-forensic treatment, the supervising court decides on the mode of release, but both patient and clinic can apply for premature termination.

Substance use was defined as the second outcome with abstinence during the follow-up period as positive and consumption of any prohibited substance as negative. This information was based on self-reporting by the patients and aftercare staff. In cases where the sources contradicted each other, a conservative approach was taken and substance use was assumed. Hereby, the supervising court decided which substance was prohibited in each individual case. Before release, the court issues instructions to the patients, which they must comply with during the forensic aftercare period. These include which substances may not be consumed, depending on the offence and the individual forensic assessment. The positive group therefore includes patients that consumed forensically irrelevant substances, patients undergoing substitution therapy without co-use of prohibited substances, and patients with stable abstinence after a single relapse.

Finally, illegal behaviour served as the third outcome. Any form of delinquent behaviour during the follow-up period was considered negative. This is regardless of whether it became a matter of formal record or even led to an indictment or conviction. Again, this information was based on self-reporting by the patients and aftercare staff.

Data analysis

Statistical analysis was performed using IBM SPSS Statistics version 28. In a first, univariate step, we used Pearson χ² tests for categorical and t-tests for continuous variables in order to identify relationships between the variables of the historical, treatment and follow-up sets and the three outcome variables. As effect sizes, Cramer’s V and Cohen’s d were calculated.

Prior to step 2, variables were classified as predictor candidates (PC) if they showed a significant relationship to the respective outcome variable at step 1, and if effect sizes were relevant (i.e. d > .2 or V >.1). If the percentage of missing data did exceed 10%, or if the PC was logically related to the outcome criterion it was not considered in step 2.

In this second, multivariate step, we calculated for every outcome variable various logistic regression models with the previously identified specific PC. In the first model (model A), only historical variables were considered, in the second, historical and treatment variables (model AB) and in the third, variables of all sets (model ABC). As follow-up variables were raised after the first outcome (i.e. the release mode) was determined, these sorts of variables cannot serve as predictors of the first outcome and model ABC therefore was omitted. In order to get efficient models, we applied a forward stepwise regression procedure (only including predictors if they undercut the α = .05-level; if an included variable exceeded the α = .1-level it was excluded).

Taken together, eight logistic regression analyses were calculated and the resulting models were compared to each other on the parameters Nagelkerke’s R² (as an estimator of variance explanation, see (41)) and AUC as well as the Gini index (following a ROC-analysis for each model). According to the guidelines of Rice and Harris (42), AUC values between .56 and .64 are considered a low effect (Cohen’s d .20), values between .64 and .71 are a medium effect (Cohen’s d .50), and values of .71 or above a high effect (Cohen’s d .80). The models’ goodness of fit was estimated by the Hosmer & Lemeshow test (43). As classification markers, we report specificity as well as sensitivity and – due to the imbalance of group members – balanced accuracy (BAC, i.e. [sensitivity + specificity]/2, see (44)).

Results

Descriptives

With regard to the release mode as first outcome, 2,824 patients (67%) were released on parole, while FAT was terminated prematurely for 1,369 patients (33%; n = 8 missing data). Both groups differ significantly in treatment duration (t = 28.1, p < .001): patients released on parole had been treated for an average of 25.9 months (SD = 9.2), patients with premature termination for an average of 16.1 months (SD = 11.2). Among the regularly released patients, 93.9% were instructed by the supervising courts to attend forensic aftercare.

Out of the group of patients with regular release and a valid dataset at t₁, 1,790 (75%) did not consume prohibited psychoactive substances during follow-up (or had a single relapse without further consequences) versus 590 who did so (25%; n = 11 missing data with respect to substance use as the second outcome). There were 1,924 patients (81%) who did not show any delinquent behaviour during the follow-up period versus 462 who did so (19%; n = 5 missing data with respect to re-delinquency as the third outcome).

Univariate relationships

Release mode

As the result of univariate analyses, 23 variables of the historical set and 22 variables of the treatment set met the criteria of statistical significance and a relevant effect size (as an overview, see Supplement 2.1 containing all significant and relevant variables). Among the historical variables related to release mode were, e.g. main diagnosis, main offence, level of schooling, vocational training, marital status, length of concurrent prison sentence and indicators of criminal history. Factors identified from the treatment set included age at discharge, clinic location, medication, incidents during treatment (i.e. misuse of privileges, abscondence, substance use, violence against fellow patients), clinical assessments of treatment progress, and treatment duration.

(Prohibited) substance use

Thirteen variables of the historical set, seven of the treatment set and 28 of the follow-up set turned out to be significantly associated to substance use to a relevant degree (as an overview, see Supplement 2.2 containing all significant and relevant variables). Besides historical variables of criminal, clinical and educational history, the main offence and duration of concurrent prison sentence were associated with substance use as well. Among treatment variables were clinic location, incidents during treatment (i.e. misuse of privileges, abscondence, substance use) and clinical assessments of treatment progress. The resulting list of follow-up variables included several indicators of resocialisation (e.g. occupational, living and financial situation), re-delinquency, contact with forensic aftercare and other medical treatment providers.

Re-delinquency

The criteria of statistical significance and relevant effect sizes were met by 15 variables of the historical set, eight of the treatment set and 28 of the follow-up set (as an overview, see Supplement 2.3 containing all significant and relevant variables). Historical variables of educational and criminal background, the current legal case and age at admission were related to re-delinquency. Among treatment variables were clinic location, incidents during treatment (i.e. misuse of privileges, abscondence, substance use, violence against fellow patients), clinical assessments of treatment progress and age at discharge. The resulting list of follow-up variables included several indicators of resocialisation (e.g. occupational, living and financial situation), substance use, contact with forensic aftercare and other medical treatment providers.

Predictive models on release mode

In order to establish exhaustive statistical models to predict regular release on various predictor levels, we first inserted 18 out of the 23 previously identified historical PC (five variables were omitted for exceeding 10% missing data) and then these 18 historical together with seven out of 22 treatment PC in forward stepwise binary logistic regression procedures (nine treatment variables were omitted for exceeding 10% missing data, ‘administrative district’ was logically linked to ‘forensic hospital’ and therefore omitted, ‘improvement of addiction disorder’ as indicated by the treatment providers was a logical precursor of the outcome criteria and – due to the legal peculiarities – ‘duration of treatment’ as well as ‘age at discharge’ were also logically linked to the first outcome, i.e. regular release, so were omitted). Table 1 gives an overview of the resulting models (model A and model AB) after having terminated the regression procedure.

Table 1.

Comparison of model characteristics, forward stepwise logistic regression models with outcome release mode.

	Model Aa			Model ABb
Valid cases (% missings)	3,482 (17%)			3,415 (19%)
No. of steps (variables)	14 (12)			12 (12)
No. of variables	Set Ac	Set Bd	Set Ce	Set Ac	Set Bd	Set Ce
Provided	18	–	–	18	7	–
Included	12	–	–	7	5	–
Nagelkerke R²	.255			.460
Omnibus model test	Chi²	df	p	Chi²	df	p
	706.2	39	<.001	1368.9	42	<.001
Goodness of fit, Hosmer & Lemeshow test	Chi²	df	p	Chi²	df	p
	13.6	8	.094	9.7	8	.289
ROC-analysis	AUC	p	Gini-Ind.	AUC	p	Gini-Ind.
	.77	<.001	.53	.86	<.001	.72
Correctly classified cases	spec.	sens.	BAC	spec.	sens.	BAC
	88.3%	44.1%	66.2%	89.5%	60.7%	75.1%

Notes. ^aOnly historical predictor candidates were provided for this model (set A); ^bhistorical and treatment-related predictor candidates were provided for this model (set A and set B); ^cset-A comprised historical and background variables; ^dset-B comprised treatment variables; ^eset-C comprised follow-up variables; abbreviations: AUC = area under curve; BAC = balanced accuracy; df = degrees of freedom; Gini-Ind. = Gini-index; ROC = receiver operating characteristic; sens. = sensitivity; spec. = specificity.

While both models include 12 variables, they differ largely in their degree of variance explanation (model A: 26%, model AB: 46%, see Table 1). Thus, by additionally taking into account variables on treatment course and setting, the percentage of explained variance can be increased by 20%. Both models show high effect sizes, as each AUC exceeds .71 (42).

Table 2 shows the odds ratios (OR) of those variables included in at least one of the resulting models. The following initially provided historical PC were not included in model A or AB: working situation, longest employment, inpatient treatment, prior treatment in forensic psychiatry and age at first criminal act. Among the provided treatment PC, misuse of privileges and violence against fellow patients were not included in model AB.

Table 2.

Variables and coefficients in resulting forward stepwise logistic regression models with outcome release mode.

ID No.	Variables in the equation (index category, if applicable)	Coefficients after stepwise inclusion of significant predictor candidates
		Model Aa			Model ABb
	Category	OR / p	95% CI of OR		OR / p	95% CI of OR
Set A: historical and background variables
5	Marital status (single)	p < .001			p < .001
	Married	.46***	.34	.64	.56***	.39	.80
	Divorced/separated	.85	.69	1.05	.70**	.55	.89
	Widowed	.61	.22	1.65	.79	.25	2.44
	Stable partnership	.28***	.23	.34	.23***	.18	.29
23	Level of schooling (none)	p < .001			p < .001
	Special school	.63*	.41	.97	.75	.45	1.23
	Secondary school (without degree)	.48***	.35	.65	.47***	.33	.67
	Secondary school (with degree)	.46***	.33	.63	.45***	.31	.65
	Qualified secondary school	.42***	.29	.62	.38***	.25	.59
	College	.45*	.22	.91	.50	.24	1.03
	Technical college	.42*	.19	.92	.32*	.12	.82
	Other	.53*	.29	.96	.45*	.22	.92
24a	Level of vocational training (none / in training)	p < .01			n.s.
	Voc. training dropout	.74**	.60	.92
	Graduated voc. training	.65***	.52	.81
	Academic degree	1.52	.62	3.72
	Multiple degrees	.83	.46	1.51
	Other	.77	.38	1.53	n.s.
34	Age at first verdict	.98**	.97	.99
35	Legal file records	n.s.			1.02*	1.01	1.04
36	Total duration of previous imprisonment	1.004***	1.002	1.006	1.003**	1.001	1.006
39	Violent physical assault as predelinquency	1.40***	1.17	1.67	1.40***	1.15	1.71
42	Robbery and extortionate robbery as predelinquency	1.32*	1.06	1.65	n.s.
43	Property crime as predelinquency	1.31**	1.09	1.56	n.s.
46	Other offence as predelinquency	1.34***	1.14	1.57	n.s.
48a	Main diagnosis (multiple drug use)	p < .001			n.s.
	Other (no addiction disorder)	1.29	.51	3.27
	Alcohol	.67***	.52	.85
	Opioids	1.29*	1.02	1.64
	Cannabinoids	.66*	.47	.93
	Sedative	.49	.04	5.44
	Cocaine	.71	.35	1.43
	Other stimulants	.71**	.54	.92
64a	Main offence (drug offence)	p < .001			p < .001
	Homicide	1.06	.46	2,46	1.36	.52	3.54
	Violent physical assault	1.80***	1.40	2,32	1.62***	1.25	2.11
	Sexual offence against adults	.79	.34	1,86	1.12	.41	3.01
	Sexual offence against minors	–			–
	Arson	2.22	.99	4.96	1.45	.57	3.68
	Robbery	1.79***	1.35	2.37	1.57**	1.15	2.16
	Property offence	2.41***	1.89	3.06	2.47***	1.89	3.24
	Traffic offence	1.77*	1.09	2.90	1.45	.84	2.50
	Other	1.50	.94	2.41	1.67	.98	2.85
75	Duration of concurrent prison sentence	.986***	.982	.990	.985***	.980	.990
Set B: treatment and setting variables
1	Forensic hospital (Regensburg)				p < .001
	Hospital 1				.29***	.19	.45
	Hospital 2				.81	.44	1.48
	Hospital 3				.71	.50	1.02
	Hospital 4				.47***	.34	.66
	Hospital 5				.80	.27	2.35
	Hospital 6				.96	.66	1.40
	Hospital 7				1.07	.46	2.48
	Hospital 8				1.14	.74	1.75
	Hospital 9				.05***	.01	.18
	Hospital 10				7.05***	4.68	10.61
	Hospital 11				1.62**	1.18	2.21
	Hospital 12				.25**	.11	.60
	Hospital 13				1.81**	1.19	2.75
	Hospital 14				1.23	.70	2.18
4a	Abscondence				11.27***	7.20	17.64
7	Re-delinquency during treatment				2.94***	1.91	4.53
8a	Use of (prohibited) psychoactive substances during treatment				4.51***	3.67	5.54
24	Any medication at discharge				1.58***	1.29	1.94
Constant		2.03**			.40***

Notes: ^aOnly historical predictor candidates were provided for this model; ^bhistorical and treatment predictor candidates were provided for this model; * p < .05, ** p < .01, *** p < .001; significant OR are highlighted in bold; abbreviations: CI = confidence interval; OR = odds ratio.

Predictive models on (prohibited) substance use

Eleven out of 13 previously identified historical PC (two variables were omitted for exceeding 10% missing data) were provided for inclusion in the first forward stepwise binary logistic regression procedure in order to predict substance use after release as outcome criteria. In the second analysis, these 11 historical variables were inserted together with five out seven treatment PC (the variable ‘administrative district’ was omitted because of the logical linkage to the variable ‘forensic hospital’, ‘improvement of comorbid disorder’ exceeded the cut-off of 10% missing data). In the last procedure, 11 out of 28 follow-up PC were additionally provided (16 variables were omitted for exceeding 10% missing data, ‘living situation’ was logically linked to ‘living in complementary institution’ and therefore omitted). Table 3 gives an overview over the resulting models (model A, model AB and model ABC) after having terminated the regression procedure.

Table 3.

Comparison of model characteristics, forward stepwise logistic regression models with outcome substance use after treatment.

	Model Aa			Model ABb			Model ABCc
Valid cases (% missings)	2,165 (9%)			2,142 (10%)			1,643 (31%)
No. of steps (variables)	5 (5)			9 (9)			10 (10)
No. of variables	Set Ad	Set Be	Set Cf	Set Ad	Set Be	Set Cf	Set Ad	Set Be	Set Cf
Provided	11	–	–	11	5	–	11	5	11
Included	5	–	–	5	4	–	1	2	7
Nagelkerke R²	.087			.185			.373
Omnibus model test	Chi²	df	p	Chi²	df	p	Chi²	df	p
	130.1	10	<.001	285.2	28	<.001	464.85	31	<.001
Goodness of fit, Hosmer & Lemeshow test	Chi²	df	p	Chi²	df	p	Chi²	df	p
	7.5	8	.483	9.5	8	.304	15.9	8	.044
ROC-Analysis	AUC	p	Gini-Ind.	AUC	p	Gini-Ind.	AUC	p	Gini-Ind.
	.66	<.001	.32	.73	<.001	.46	.84	<.001	.70
Correctly classified cases	spec.	sens.	BAC	spec.	sens.	BAC	spec.	sens.	BAC
	99.4%	1.9%	50.7%	94.8%	20.2%	57.5%	94.2%	42.6%	68.4%

Notes. ^aOnly historical predictor candidates were provided for this model; ^bhistorical and treatment-related predictor candidates were provided for this model; ^chistorical, treatment-related and follow-up predictor candidates were provided for this model; ^dset-A comprised historical and background variables; ^eset-B comprised treatment variables; ^fset-C comprised follow-up variables; abbreviations: AUC = area under curve; BAC = balanced accuracy; df = degrees of freedom; Gini-Ind. = Gini-index; ROC = receiver operating characteristic; sens. = sensitivity; spec. = specificity.

The higher the level of predictors, the more variables were included and the higher the proportion of variance explained. Only when follow-up variables are included in addition to historical and treatment variables (as in model ABC) does variance explanation exceed one-third. With AUC = .66, model A reveals a medium effect size (42) and models AB and ABC high effect sizes (as both AUC > .71).

Table 4 shows OR of those variables included in at least one of the resulting models. In none of the models were the following initially provided historical PC included: level of vocational training, age at first criminal act, legal file records, predelinquency and main offence. Among the treatment PC, number of misuses of privileges was not included in either model AB or ABC. In the latter model, occupational situation, secured financial situation, source of income and marital status as provided follow-up PC were not included.

Table 4.

Variables and coefficients in resulting forward stepwise logistic regression models with outcome substance use after treatment.

ID No.	Variables in the equation (index category, if applicable)	Coefficients after stepwise inclusion of significant predictor candidates
		Model Aa			Model ABb			Model ABCc
	Category	OR / p	95% CI of OR		OR / p	95% CI of OR		OR / p	95% CI of OR
Set A: historical and background variables
25	Working situation (unemployed)	p < .001			p = .034			n.s.
	In training	.61	.32	1.14	.73	.37	1.44
	Retired	.54	.15	1.88	.67	.18	2.42
	Regular employment	.70**	.56	.88	.83	.65	1.05
	Freelance work	.55*	.33	.92	.62	.36	1.07
	Mini-job	.27**	.11	.69	.33*	.13	.85
	Other	.35**	.17	.76	.41*	.19	.90
27	Longest employment	n.s.			.997*	.994	1.000	n.s.
28a	Age at first manifestation of disorder	.97***	.95	.99	n.s.			n.s.
34	Age at first verdict	.97***	.95	.99	.97**	.96	.99	.97**	.95	.99
36	Total duration of previous imprisonment	1.004***	1.002	1.007	1.003*	1.001	1.006	n.s.
75	Duration of concurrent prison sentence	.991***	.986	.996	.991**	.987	.997	n.s.

Set B: treatment and setting variables
1	Forensic hospital (Regensburg)				p < .001			p = .002
	Hospital 1				.97	.64	1.49	40**	.21	.74
	Hospital 2				.49*	.24	.98	.46	.18	1.18
	Hospital 3				1.20	.82	1.77	.63	.36	1.09
	Hospital 4				.64*	.45	.91	.57*	.35	.94
	Hospital 5				1.38	.26	7.32	3.95	.57	27.34
	Hospital 6				.53**	.32	.86	.55*	.30	.99
	Hospital 7				1.23	.32	4.73	2.72	.30	24.81
	Hospital 8				.43**	.24	.77	.35**	.18	.70
	Hospital 9				.71	.42	1.19	.74	.36	1.51
	Hospital 10				.67	.32	1.40	.61	.25	1.47
	Hospital 11				.71	.48	1.05	1.08	.62	1.88
	Hospital 12				.39*	.18	.88	.44	.16	1.21
	Hospital 13				.38**	.18	.76	.23*	.07	.78
	Hospital 14				.51	.25	1.06	.21*	.06	.73
4a	Abscondence				2.55*	1.22	5.79	n.s.
8a	Use of (prohibited) psychoactive substances during treatment				2.54***	2.02	3.20	1.74***	1.30	2.34
16	Improvement of addiction disorder (no)				p < .001			n.s.
	Moderate				.31*	.13	.75
	Very				.22***	.09	.54
Set C: follow-up variables
1	Interviews conducted with both patients and professionals							.57**	.38	.87
8	Living in complementary institution (no, single/family)							p = .004
	Yes (residential or assisted living)							.54*	.30	.98
	Homeless							5.30	.49	57.72
	Other (prison, forensic hospital, foreign country, etc.)							2.42**	1.24	4.74
23	Any inpatient treatment							1.75**	1.20	2.55
36	Frequency of contact with forensic aftercare (weekly)							p < .001
	Never							–
	Sporadic							.78	.44	1.39
	Every second week							.38***	.23	.61
	Monthly							.20***	.12	.34
	Other (regular) frequency							.18***	.07	.44
39a	Re-delinqueny according to professionals (no)							p = .006
	Yes							.98	.35	2.73
	Unclear (assumed re-delinquency)							4.17*	1.16	15.00
55	Re-delinquency combined and conservative							2.85*	1.08	7.52
57	Revocation of parole (no)							p < .001
	Yes							6.54**	1.69	25.31
	Probable							4.48**	1.81	11.13
Constant		1.80*			3.23*			1.73

Notes. ^aOnly historical predictor candidates were provided for this model; ^bhistorical and treatment-related predictor candidates were provided for this model; ^chistorical, treatment-related and follow-up predictor candidates were provided for this model; * p < .05, ** p < .01, *** p < .001; significant OR are highlighted in bold; abbreviations: CI = confidence interval; OR = odds ratio.

Predictive models on re-delinquency

In order to predict re-delinquency as outcome criteria, 12 out of 15 previously identified historical PC (three variables were omitted for exceeding 10% missing data) were provided for inclusion in the first forward stepwise binary logistic regression procedure. In the second analysis, these 12 historical variables were inserted together with six out of eight treatment predictor candidates (two variables exceeded the cut-off of 10% missing data). In the last analysis, 12 out of 28 follow-up PC were additionally provided (15 variables were omitted for exceeding 10% missing data, ‘living situation’ was logically linked to ‘living in complementary institution’ and therefore omitted). Table 5 gives an overview over the resulting models (model A, model AB and model ABC) after having terminated the regression procedure.

Table 5.

Comparison of model characteristics, forward stepwise logistic regression models with outcome re-delinquency.

	Model Aa			Model ABb			Model ABCc
Valid cases (% missings)	2,176 (9%)			2,146 (10%)			1,619 (32%)
No. of steps (variables)	8 (6)			8 (8)			11 (11)
No. of variables	Set Ad	Set Be	Set Cf	Set Ad	Set Be	Set Cf	Set Ad	Set Be	Set Cf
Provided	12	–	–	12	6	–	12	6	12
Included	6	–	–	5	3	–	3	1	7
Nagelkerke R²	.093			.140			.397
Omnibus model test	Chi²	df	P	Chi²	df	p	Chi²	df	p
	130.6	11	<.001	197.3	28	<.001	434.20	25	<.001
Goodness of fit, Hosmer & Lemeshow test	Chi²	df	P	Chi²	df	p	Chi²	df	p
	4.3	8	.825	6.6	8	.582	5.3	8	.720
ROC-Analysis	AUC	p	Gini-Ind.	AUC	p	Gini-Ind.	AUC	p	Gini-Ind.
	.67	<.001	.35	.71	<.001	.43	.87	<.001	.73
Correctly classified cases	spec.	sens.	BAC	spec.	sens.	BAC	spec.	sens.	BAC
	99.7%	1.7%	50.7%	98.8%	6.0%	52.4	96.1%	33.5%	64.8%

Notes. ^aOnly historical predictor candidates were provided for this model; ^bhistorical and treatment-related predictor candidates were provided for this model; ^chistorical, treatment-related and follow-up predictor candidates were provided for this model; ^dset-A comprised historical and background variables; ^eset-B comprised treatment variables; ^fset-C comprised follow-up variables; abbreviations: AUC = area under curve; BAC = balanced accuracy; df = degrees of freedom; Gini-Ind. = Gini-index; ROC = receiver operating characteristic; sens. = sensitivity; spec. = specificity.

Again, the higher the level of predictors, the more variables were included and the higher the proportion of variance explained. Only when follow-up variables were included in addition to historical and treatment course variables (as in model ABC) did variance explanation increase to almost 40%. Model A reveals a medium effect size, whereas models AB and ABC reveal high effect sizes (42).

Table 6 shows OR of those variables included in at least one of the resulting models. In none of the models were the following initially provided historical PC included: level of schooling, level of vocational training, age at first criminal act, age at first verdict and main offence. Among the treatment PC, forensic hospital, number of violent assaults against fellow patients, and use of (prohibited) psychoactive substances during treatment were not included in either model AB or ABC. In the latter model, interview conducted with, occupational situation, living in complementary institution, secured financial situation, and marital status as follow-up PC were not included.

Table 6.

Variables and coefficients in resulting forward stepwise logistic regression models with outcome re-delinquency.

ID No	Variables in the equation (index category, if applicable)	Coefficients after stepwise inclusion of significant predictor candidates
		Model Aa			Model ABb			Model ABCc
	Category	OR / p	95% CI of OR		OR / p	95% CI of OR		OR / p	95% CI of OR
Set A: historical and background variables
25	Working situation (unemployed)	p = .035			n.s.			n.s.
	In training	.86	.46	1.61
	Retired	1.90	.65	5.54
	Regular employment	.79	.61	1.02
	Freelance work	.65	.37	1.16
	Mini-job	.47	.19	1.12
	Other	.35*	.15	.83
27	Longest employment	.996*	.993	< 1.000	.996*	.992	< 1.000	n.s.
28a	Age at first manifestation of disorder	n.s.			n.s.			1.04*	1.01	1.06
35	Legal file records	1.05***	1.02	1.07	1.04**	1.01	1.06	n.s.
36	Total duration of previous imprisonment	1.006***	1.003	1.009	1.005***	1.002	1.008	1.008***	1.004	1.013
75	Duration of concurrent prison sentence	.993**	.987	.998	.991***	.986	.996	n.s.
79a	Age at admission	.96***	.94	.97	.97***	.95	.98	.96**	.94	.99
Set B: treatment and setting variables
3a	Misuse of privileges (short leave etc.)				1.41*	1.05	1.88	1.57*	1.05	2.34
8a	Use of (prohibited) psychoactive substances during treatment				2.25***	1.77	2.85	n.s.
16	Improvement of addiction disorder (no)				p = .003			n.s.
	Moderate				.42*	.18	.97
	Very				.31**	.13	.71
Set C: follow-up variables
10	Source of income (wage)							p = .016
	Basic income scheme							1.37	.88	2.14
	Social welfare							1.09	.48	2.47
	Unemployment benefit							2.04*	1.14	3.66
	Retirement pension							.94	.27	3.27
	Other							2.21**	1.36	3.58
13a	Substance use acc. to professionals (none)							p < .001
	Permanent substance use							2.49	.80	7.82
	Abstinence after relapse							4.69***	2.85	7.74
	Intermediate substance use (no stable abstinence and neg. effects observable)							1.69	.57	5.05
	Substituted							1.20	.08	18.0
	Unclear (substance use probable)							2.77	.70	10.9
16	Abstinence (trichotomised and conservative) (yes)							p = .007
	Substance use (permanent or intermediate)							5.25**	1.88	14.69
	Substituted							3.18	.21	47.86
23	Any inpatient treatment							1.71**	1.15	2.56
30	Any outpatient treatment							1.95**	1.22	3.13
36	Frequency of contact with forensic aftercare (weekly)							p < .001
	Never							.89	.06	14.14
	Sporadic							2.59**	1.46	4.61
	Every second week							1.39	.86	2.24
	Monthly							.94	.57	1.56
	Other (regular) frequency							.11*	.01	.89
57	Revocation of parole (no)							p < .001
	Yes							7.70***	2.37	24.96
	Probable							3.56**	1.54	8.25
Constant		1.80*			3.23*			.03***

Notes. ^aOnly historical predictor candidates were provided for this model; ^bhistorical and treatment-related predictor candidates were provided for this model; ^chistorical, treatment-related and follow-up predictor candidates were provided for this model; * p < .05, ** p < .01, *** p < .001; significant OR are highlighted in bold; abbreviations: CI = confidence interval; OR = odds ratio.

Discussion

The present study sought to identify predictors of the success in forensic addiction treatment (FAT) and to analyse their relationship to each other and to three different outcome criteria. A negative outcome was defined as premature release, re-delinquency and/or use of prohibited substances after treatment. The predictive value of various types of variables was investigated by the application of stepwise logistic regression analyses.

Main findings on risk and protective factors

The results of the multivariate models indicate that the strongest protective factors for regular release are being married or living in a stable partnership, having any school degree and having been sentenced to a long concurrent prison term. Long duration of previous imprisonment, a history of violent offences, young age at first conviction and robbery, property offences and violent physical assaults as the main offences in the current court case, on the other hand, are the most relevant risk factors. During treatment, abscondences, re-offending, drug use and the need for medication also increase the risk of premature release.

Unemployment, young age at first manifestation of substance use disorder, young age at first conviction, long previous prison sentences and short concurrent prison sentences are the most important retrospective risk factors for substance use after release. Abscondences and substance use during treatment also increase the risk. As far as the period after treatment is concerned, the need for frequent contact with forensic aftercare, the revocation of probation and reoffending are closely linked to substance use.

For re-delinquency as the third outcome criterion, long previous prison sentences, a high number of legal file records, a young age at admission and short concurrent prison sentences are the most important retrospective risk factors, while a long previous period of employment serves as a protective factor. During treatment, substance use was identified as the strongest risk factor. In the post-treatment phase, abstinence is the most important protective factor. This result emphasises how closely substance use and offence risk are still intertwined even after regular release from FAT.

Comparison with previous findings

It is noticeable that the success rate in the present study is higher for all three outcome criteria than in a recently published systematic review (16). There, for example, the median proportion of premature termination was 47%. In the present study, it is only 33%. For substance use and re-delinquency after release, this is certainly related to the comparatively short time-at-risk; on the other hand, it could be an indication of selection effects attributable to voluntary participation in the present study. The difference in the criterion of substance use after FAT is certainly also influenced by our more liberal definition (only the use of prohibited psychoactive substances was considered negative and a single relapse did not count as negative outcome, as well).

Release mode

Compared to previous research on therapy completion as outcome criteria, the large number of PC identified here, which corresponds to the large number of significant predictors as a result of the multivariate analysis, is initially surprising. This is certainly related to the comprehensive set of variables and the large sample. It would go beyond the scope of this report to go into every single variable, so we will limit ourselves to a few that we consider to be particularly relevant. With 12 historical variables, one-fourth of the variance can be explained. This value even exceeds the previously reported shares (17,19,20) with its maximum of 21% variance explanation (18). Even if the predictor sets are not congruent, they are similar in terms of content. As with the present set of predictors, most sets reported in the literature include information on familiar/living situation prior to treatment, level of schooling, occupation (status or level of training), criminal history, main diagnosis (drug vs alcohol vs multiple SUD), main offences, further delinquency (predelinquency or secondary offences), criminal history (age at first verdict, number of file entries) and the duration of concurrent prison sentence.

In the present study, by including setting and process variables, the proportion of variance explained increases to just under half. This also significantly exceeds the values reported by Gericke and Kallert (21) who reported 26% (including incidents during treatment) and Querengässer et al. (18) reporting 32% (including the clinic factor). It is not surprising that variables related to particular incidents increase the risk of premature release by up to OR = 11 (for abscondence). After all, decisions about premature termination are made by courts on the basis of the course of treatment. It is noteworthy, however, that the clinic factor remains significant even in a multivariate analysis. The high range between OR = .25 and OR = 7.1 (compared to the index clinic) indicates that the differences in the proportion of irregular releases that the various clinics reveal cannot be explained by patient- or process-related factors. In other words, the chance of a regular release depends to a very large extent on the forensic hospital (or the responsible local supervising court as the deciding authority) in which the patient is treated, even when other influencing factors are controlled.

This finding is reminiscent of a recently published study that examined lengths of stay in forensic psychiatry in Germany (45). It showed that these differed significantly between the facilities. Although structural differences in the characteristics of the clinics were also examined, it remained unclear to what this clinic effect was attributable. Against this background, further research into non-patient-related factors on the outcome of forensic treatment appears highly desirable.

(Prohibited) substance use

In the present study, the resulting set of five historical variables explain less than 9% of outcome variance on substance use after completed FAT. Compared to the release mode as first relevant outcome, this is a rather sobering finding. The resulting set of risk factors (unemployment, young age at first verdict and manifestation of SUD, short duration of concurrent prison disorder and long duration of previous imprisonment) widely correspond to previous findings (16,29,34,36,37).

The consideration of process and setting variables more than doubles variance explanation – without eliminating the historical predictors in the resulting model. Abscondences and substance use during treatment increase the risk of substance use after FAT; clinical assessment on the improvement of SUD is related to post-treatment substance use as well as the clinic variable. With respect to this outcome, the wide range of OR = .21 and OR = 4 should be interpreted with due caution as an indicator of systematic differential treatment effectiveness. Some clinics seem to be much more successful than others in their core mission, namely to treat SUD effectively.

Only if post-stationary factors are included can a variance explanation of over 30% be achieved. All but one of the historical variables are removed from the model and only the hospital factor and substance use during treatment remain from the treatment-related PC. Three out of seven resulting follow-up variables refer to the patient’s legal behaviour. This underlines the link between re-delinquency and substance use after FAT (29,34). Another two variables are linked to the patient’s behaviour (either towards the forensic aftercare or to the interviewer). Only the variable living in a complementary institution was a structural indicator of the risk to consume prohibited substances: it was lower in residential or assisted living. The correlation with the variable inpatient treatment is more likely to be a consequence of new consumption than a trigger. Taken together, there is very little overlap with those risk factors reported from non-forensic research (e.g. unemployment, comorbid personality or psychiatric disorder (38)).

Re-delinquency

With the variables at hand, less than 10% of the variance on re-delinquency can be explained by historical variables. The resulting list of risk factors comprises unemployment and short duration of employment prior to FAT, criminal history (high numbers of file entries and long duration of prison experience), short duration of concurrent prison sentence and young age at admission. This is in line with previous findings (14,16), even if these findings are very heterogeneous. In contrast to substance use as the second outcome, the integration of setting and process variables increases variance explanation by less than 5%. Misuse of privileges and substance use during treatment are related to re-delinquency. It is possible that antisocial behaviours are already evident in these two treatment variables, which reappear after the end of treatment in the form of renewed delinquency. Interestingly, the clinical assessment on the improvement of SUD is related to re-delinquency, as well. This again points to the link between consumption and delinquency after FAT. The exhaustive model with all three types of variables succeeds in explaining up to 40% of the variance on re-delinquency. Again, follow-up variables eliminate most of the other variables. While five of them concern substance use or treatment (inpatient, outpatient or forensic aftercare), only unemployment benefit or other (mostly unknown) sources of income can be identified as structural risk factors. The correlation with the variable revocation of parole is more likely to be a consequence of re-delinquency than a trigger.

Notably, in contrast to previous research (30,31), adverse childhood experiences or difficult family conditions played no role in any of the models predicting re-delinquency. The same can be stated for the clinic factor. In other words, differences in the structure and possibly in the interventions offered by the clinics are not reflected in different rates of re-delinquency.

Answers to the research questions

We further discuss the results based on the previously posed research questions. The question of whether distal outcome criteria can also be predicted on the basis of historical information cannot be affirmed. Variance explanations of under 10% and AUC near the border of a low effect suggest that the criteria of re-delinquency and substance use after FAT are hardly predictable by the historical and background information at hand. Instead, and leading to the second question, the inclusion of information on the treatment process and the follow-up period adds considerable predictive value.

Some differentiations must also be made here. Since the decision on premature termination of FAT (first outcome) is made by a supervising court based on clinicians’ recommendations, it is not surprising that taking the treatment process into account increases the proportion of variance explained to almost half. Compared to that, the results concerning re-delinquency and substance use after FAT are rather disappointing. Still, less than 20% (for substance use) or less than 15% (for re-delinquency) of the variance can be explained by historical and treatment process variables together. Only the inclusion of follow-up information allows a satisfactory prediction with an explained variance for each of around 40% and AUC greater than .8. This can be interpreted to mean that the time after treatment (and the work done by forensic aftercare (46)) is crucial and has a much greater influence on these criteria than the characteristics of the patients and the inpatient treatment itself. The latter underlines the conclusion of Walde and Lungwitz (28) who highlighted the importance of risk and protective factors that refer to the follow-up period.

The next question was whether different prediction models that depend on the outcome criterion lead to comparable lists of predictors. The answer to this question is predominantly no. Across the three outcome criteria, a total of 17 variables were included in the respective historical models (model A). Of these, only the total duration of previous imprisonment and duration of concurrent prison sentence turn out to be predictive for all three outcomes. Another two variables were predictive for at least two outcomes: working situation for substance use and re-delinquency, and age at first verdict for release mode and substance use. The other 13 variables (76%) had a predictive value for only one of the three outcomes. There was somewhat greater agreement for the models in which historical, treatment and setting variables were included (models AB). Here, 10 of the 18 total variables included (56%) had a predictive value for only one of the outcomes. Three variables turned out to be predictive for every outcome; besides previous imprisonment and short concurrent prison sentence, substance use during treatment increased the risks as well. The other five variables showed predictive values for two outcomes: legal file records (release mode and re-delinquency), longest employment and clinical assessment on improvement of SUD (each substance use and re-delinquency), forensic hospital and abscondence (each release mode and substance use). In terms of the most comprehensive regression models that included all three types of variables (and were only calculated for the outcomes substance use and re-delinquency), only three out of 18 total predictors identified had a predictive value for both criteria: renewed inpatient treatment, contact with forensic aftercare and revocation of parole. Hence, the vast majority of predictors (83%) were related to only one of the outcome criteria.

The answer to the last research question probably reveals the highest practical relevance. Our findings indicate that the final predictor sets are not suitable for the creation of a prognostic tool for predicting treatment success in FAT prior to its onset. Before FAT begins, only the patient’s historical data is available for making any prognostic assessment. As we have seen, the historical data models for the three investigated outcome criteria result in heterogeneous lists of predictors and only show a reasonably satisfactory explanation of variance for the criterion release mode. Not only can the forensically at least equally relevant distal outcome criteria (substance use and re-delinquency) hardly be predicted by historical information only, but the identified sets of historical risk factors also show hardly any congruence.

Limitations and outlook

The present study pursues an explorative approach. Due to the large number of variables analysed, significant results are very likely for theoretical reasons alone. From a methodological point of view, the omission of a Bonferroni correction and the avoidance of preliminary analyses must be viewed critically. However, as the main objective (in addition to identifying potential predictors) was to determine how the predictors relate to each other and to the different outcomes, this limitation is somehow unavoidable. A confirmatory approach with adjustment for multiple testing would have limited the scope for achieving this objective too much. Although attempts were made to minimise the methodological problems, for example by selecting the predictors on the basis of the effect size and not significance alone, this all impacts reliability and generalisability of the present results and underlines our study’s exploratory nature and the need for further, hypothesis-driven research.

The present data set is routine quality assurance data. As this is not a full survey, but data transmission is voluntary, selection effects cannot be ruled out. Instead, they are actually quite probable. The fact that only data from just over half of all patients discharged during the study period could be included is certainly the greatest limitation of this study. It is inevitable that interviews can only be conducted with patients who agree to be interviewed. However, making the use of routine data that has already been collected and the anonymised questioning of staff members also dependent on patients’ consent effectively prevents the implementation of much more meaningful full surveys.

The collected variables were not selected against the background of our research questions. Despite the very broad range of variables, the investigation of further and possibly more complex variables would be desirable (see Tomlin et al. (16) who drew the same conclusion in their review). On the other hand, the present variables reflect well those characteristics that can be used by forensic experts for the required success prognosis. Unfortunately, some of the variables have not been routinely surveyed for the entire collection period. This also applies to some of the significant predictor candidates, so that these could not be considered further in the multivariate analyses because of missing data.

One of eight logistic regression analyses resulted in a significant Hosmer & Lemeshow test (ABC model for substance use, see Table 3) that challenges the goodness of fit assumption. As this test is considered to be very sensitive (43), it quickly falls below the significance level for large samples. In the present case, despite p = .04 (see Table 3), sufficient goodness of fit can nevertheless be assumed, as the difference between what was expected and observed was greater than 10 only in one out of 10 groups.

A special feature of the study is that two of the three outcome criteria are based on self-reports. There is a risk that substance use or delinquent behaviour was concealed by the patients interviewed. However, the additional questioning of experts and the conservative approach in the case of contradictory information increases the probability that such potential biases will be minimised. The advantage of information based on self-reports is that events that have never been recorded officially can also be included. Moreover, a validation study on the approach applied here showed that patients report previously unknown delinquent behaviour more frequently than known behaviour is concealed (47). Studies based on laboratory data on substance use and police register extracts on criminal offences would nevertheless be desirable in order to validate the present results.

Implications

In line with various previous findings, our results suggest that in the context of FAT, a change in the overarching understanding of therapeutic success would be appropriate. Success is not a static construct that can easily be predicted, but a dynamic process that is strongly influenced by alterable variables that take effect during and after inpatient treatment. This understanding of ‘success as a process’ means permanent monitoring and management of forensic patients’ (post-)treatment progress and needs and has significant implications for both forensic treatment and the legal system. Legislators and the courts should refrain from demanding a prognosis that is not sufficiently valid from an empirical point of view. Instead, the decision of whether FAT will be ordered should be made solely on the basis of the criminorelevance originating from the SUD, that is, according to which substance-related offences are to be expected and with what probability.

In turn, forensic treatment providers should shift their focus to long-term stability. Ongoing support for patients even after discharge and the provision of treatment services, for example, as part of expanded forensic aftercare, appear promising in order to at least mitigate the influence of post-inpatient risk variables.

Funding Statement

Data collection and analysis were sponsored by public institutions at local government level in the Federal State of Bavaria, Germany.

Supplemental material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/13218719.2025.2549827.

Ethical standards

Declaration of conflicts of interest

Jan Querengässer has declared no conflicts of interest.

Adelheid Bezzel has declared no conflicts of interest.

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the Ethics Committee at University of Regensburg (Vote No. 01/082; 2009-10-08) and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards.

Informed consent

Informed consent was obtained from all individual participants included in the study.