Risk Assessment and Risk Management of Violent Reoffending Among Prisoners (PDFDrive)

Version 4.
3, 26/05/2015; To appear in Expert Systems with Applications, 2015,
Risk assessment and risk management of violent

reoffending among prisoners
Anthony Costa Constantinou*,1, Mark Freestone2, William Marsh3, Norman Fenton4, Jeremy Coid5.
1. Risk and Information Management Research Group, School of Electronic Engineering and Computer
Science, Queen Mary, University of London, London, UK, E1 4NS. E-mail: anthony@constantinou.info
2. Violence Prevention Research Unit, Centre for Psychiatry, Wolfson Institute of Preventive Medicine, Barts
and The London School of Medicine and Dentistry, Queen Mary, University of London, London, UK,
EC1A 7BE. E-mail: m.c.freestone@qmul.ac.uk
Science, Queen Mary, University of London, London, UK, E1 4NS. E-mail: d.w.r.marsh@qmul.ac.uk
Science, Queen Mary, University of London, London, UK, E1 4NS. E-mail: n.fenton@qmul.ac.uk
5. Violence Prevention Research Unit, Centre for Psychiatry, Wolfson Institute of Preventive Medicine, Barts
and The London School of Medicine and Dentistry, Queen Mary, University of London, London, UK,
EC1A 7BE. E-mail: j.w.coid@qmul.ac.uk
THIS IS A PRE-PUBLICATION DRAFT OF THE FOLLOWING CITATION:
Constantinou, A. C., Freestone, M., Marsh W., Fenton, N., & Coid, J.W. (2015). Risk assessment and risk
management of violent reoffending among prisoners. Expert Systems with Applications, 42(21): 7511-7529.
DOI: 10.1016/j.eswa.2015.05.025
CONTACT: Dr. Anthony Constantinou, anthony@constantinou.info, a.constantinou@qmul.ac.uk
ABSTRACT
Forensic medical practitioners and scientists have for several years sought improved decision support for
determining and managing care and release of prisoners with mental health problems. Some of these prisoners
can pose a serious threat of violence to society after release. It is, therefore, critical that the risk of violent
reoffending is accurately measured and, more importantly, well managed with causal interventions to reduce
this risk after release. The well-established predictors in this area of research are typically based on regression
models or even some rule-based methods with no statistical composition, and these have proven to be unsuitable
for simulating causal interventions for risk management. In collaboration with the medical practitioners of the
Violence Prevention Research Unit (VPRU), Queen Mary University of London, we have developed a Bayesian
network (BN) model for this purpose, which we call DSVM-P (Decision Support for Violence Management -
Prisoners). The BN model captures the causal relationships between risk factors, interventions and violence and
demonstrates significantly higher accuracy (cross-validated AUC score of 0.78) compared to well-established
predictors (AUC scores ranging from 0.665 to 0.717) within this area of research, with respect to whether a
prisoner is determined suitable for release. Even more important, however, the BN model also allows for
specific risk factors to be targeted for causal intervention for risk management of future re-offending. Hence,
unlike the previous predictors, this makes the model useful in terms of answering complex clinical questions
that are based on unobserved evidence. Clinicians and probation officers who work in these areas would benefit
from a system that takes account of these complex risk management considerations, since these decision support
features are not available in the previous generation of models used by forensic psychiatrists.
Keywords: Bayesian networks, belief networks, causal intervention, forensic medicine, mental health, released
prisoners, violent offence
*
Corresponding author.
1
Version 4.3, 26/05/2015; To appear in Expert Systems with Applications, 2015,
1 INTRODUCTION
Violence is a major global public health and social concern. While violence can generally be
described as an extreme form of aggression, the many different types of violence in
conjunction with the limited understanding of their links with certain mental states make
violent behaviour difficult to assess and predict. Previous research in criminology, forensic
psychology and psychiatry has discovered both weak and strong associations between
violence and various other demographic, environmental and individual factors; often referred
to as ‘risk factors’. Some of the factors that predict violence most strongly are ‘static’ or
unchangeable measures of past behaviour, such as personality disorder, previous convictions
for violence or violence at a young age (Monahan, 1984); other factors such as criminal
networks, substance use/misuse, or serious mental illness, may be amenable to treatment or
resolve over time and are therefore considered ‘dynamic’ (Hanson & Harris, 2000). Yet some
factors, such as active symptoms of mental illness, or intoxication, are subject to minute-to-
minute or hour-to-hour fluctuations and may be considered as ‘acute’ factors, that influence
violent outcome but remain relatively unpredictable (McNeil et al., 2003).
Accuracy in risk assessment plays a major role in identifying the small group of
individuals thought to pose a very high risk of harm to society and in monitoring their level
of risk during and after treatment (Douglas et al., 2005). Accurate prediction for violence,
even from the same data, can be heavily influenced by the analytical method (Elbogen &
Johnson, 2009; van Dorn et al., 2012), suggesting that the true underlying causes of violence
are yet to be fully understood.
Prediction of violence by individuals in psychiatric and criminal justice services has
evolved from simple unstructured estimation of risk based on clinical knowledge and
intuition, through an ‘actuarial’ approach based on static predictors of violence, to structured
professional judgement (SPJ), in which a list of static risk factors is considered alongside
dynamic factors as well as idiosyncratic factors specific to the individual to provide a guided
formulation of an individual’s risk of violence. There are many SPJ tools following this
template available to the clinician, including the HCR20 (Webster et al., 1997; Douglas et al.,
2013) or Violence Risk Scale (VRS; Wong & Gordon, 2003). Although intended as guides to
clinical practice, accurate validation of these risk assessment tools requires summation of the
values assigned to each item and the use of the resulting numerical scale to create a
‘predictive’ model of future violence (e.g. Doyle et al, 2014). However, any large scale
analysis of these predictive models finds that, on aggregate, neither SPJ measures nor
actuarial lists of static factors perform above a ‘threshold’ AUC (Area Under Curve) value of
0.70 (Fazel, Singh & Grann, 2012; Yang, Wong & Coid, 2010) or correctly classifying only
60% of cases (Troquete et al, 2014), suggesting that the evidence base for such predictive
models is not compelling. Additional research has also raised concerns that involvement in
these studies by original authors of the risk assessment tools may have led to inflated
estimates of accuracy (Singh, Grann & Fazel, 2010); and that, with some offender
populations, predictive efficacy is no better than chance (Coid, Ullrich & Kallis, 2013).
Further, while previous research may aid clinical decision-makers, who are
responsible for future detention or release of prisoners, in formulating possible specific risk
scenarios, none of the previous studies take explicit account of the underlying causal factors
of violence, and the dependencies between these and any interventions. Instead, they mostly
rely on the association between variables of interest, and checklists with no statistical
composition. As a result, the previously used modelling techniques are inadequate when it
comes to risk management, whereby repeated and frequently updated assessment of an
individual must take into consideration the effectiveness of causal interventions, thereby
going beyond a classification and regression framework, and into causal analysis for
2
simulating potential interventions. Clinicians and probation officers who work in these areas
would benefit from a decision support system that takes account of these complex risk
management considerations, and this can be achieved, as we show in this paper, by the use of
causal Bayesian networks (BNs).
BNs, sometimes also called belief networks or causal probabilistic networks, can be
applied to model complex problems, where variables and knowledge from different sources
need to be integrated within a single causal framework (Pearl, 1988; Heckerman et al., 1995;
Jensen, 1996). The use of BNs for risk assessment and risk management of violent behaviour
has not previously been studied in this area of research, yet it bears similarities with other
areas of critical risk assessment and decision making where properly developed BNs have
provided transformative improvements (Fenton & Neil, 2012). For instance, BNs have been
employed for analysis and knowledge representation with success in diverse domains such as
computational biology and bioinformatics (Friedman et al., 2000; Hohenner et al., 2005;
Jiang et al., 2011), gaming (Lee & Park, 2010), computer science and artificial intelligence
(de Campos et al., 2004; Pourret et al., 2008; Fenton & Neil, 2012), medicine (Heckerman et
al., 1992; Diez et al., 1997; Nikovski 2000), and law (Fenton & Neil, 2011; Fenton et al.,
2013; Taroni et al.,2014). Especially relevant recent use of BNs include management of
project maintenance delays based on expert judgments (de Melo & Sanchez, 2008), risk
analysis in large projects to extend their understanding of project risks within the Korean
shipbuilding industry (Lee et al., 2009), systematic development of causal interventional
systems for prognostic decision support (Yet et al., 2011), qualitative examination and
evaluation of service offered by the loan departments of Greek Banks (Tarantola et al., 2012),
safety control decision support in dynamic complex project environments (Zhang et al.,
2013), football match prediction (Constantinou et al., 2012; 2013) and inference of referee
bias (Constantinou et al., 2014), detection of problems in software development project
processes (Perkusich, 2015), and jointly monitoring internal and external performance of a
Master’s programme of an Italian University in a holistic approach (Di Pietro et al., 2015).
Despite the significant benefits demonstrated, BNs are still under-exploited in clinical
assessment. Experts may be challenged to express their knowledge in probabilistic form, and
for complex problem domains elicitation of expert knowledge may require an extensive
iterative process to ensure that the experts a) agree on the structure of the model and the
variables to be considered for inference; and b) are comfortable with the nodes, states, and
conditional dependences before they make any statements of probability.
In this paper, we present a BN model, which we call DSVM-P, for risk assessment
and risk management of violent reoffending for released prisoners. The paper contributes to
forensic psychiatry research with a novel causal probabilistic model that challenges the well-
established regression and rule-based predictors (which currently represent the state-of-the-
art in violence prevention) with higher predictive accuracy, superior decision support, and
superior risk management via the simulation of causal interventions. The paper also
contributes to expert systems research by showing how an expert-constructed BN model that
learns from complex questionnaire and interviewing data (that was never intended for causal
analysis) is still capable of outperforming the relevant state-of-the-art predictors, in terms of
whether a mentally ill prisoner is determined suitable for release, by assessing the risk of
violence over a specified time period after release.
The paper is organised as follows: Section 2 describes the data and methodology
behind the development of DSVM-P; Section 3 describes the model; Section 4 discusses the
results; Section 5 discusses model benefits and limitations relative to the current state-of-the-
art; Section 6 provides our concluding remarks and direction for future research.
3
2 DATA AND METHODOLOGY
Extensive statistical analysis of cohort data, primarily focusing on classification, has

previously been carried out by the research team, leading to the development of a conceptual
staged assessment and management model for individual patients and released prisoners
(Coid et al., 2009; Ullrich & Coid, 2011). While this statistical analysis has identified useful
predictors for violent behaviour, it has also shown that none have sufficient predictive
accuracy for a purely statistical approach to be effective for decision support.
The data used is the Prisoner Cohort Study (PCS) dataset (Coid et al., 2009) which
consists of interview and assessment data on 1717 prisoners serving sentences of at least 2
years for sexual or violent offences (Coid et al., 2007). Interviews were performed over two
phases; phase 1 interviews took place during prison sentence approximately 2 years before
release, and phase 2 interviewing approximately 2 years after release. However, only 1004 of
these cases were interviewed at phase 2, of whom 13 cases could not be matched to the
criminal records of the Police National Computer (PNC), and a further 38 were lost to follow-
up. Therefore, 953 individual cases were considered for parameter learning; 778 males and
175 females.
2.1. Expert-driven causal structure
BNs provide us with the flexibility to construct the causal structure of the model purely by
expert judgment. DSVM-P was built by combining data and knowledge. The development of
DSVM-P was supported by two clinically active experts in forensic psychiatry (JC) and
forensic psychology (MF), each with at least 8 years’ experience in forensic mental health
research, having published widely on: criminal justice outcomes (Fox & Freestone, 2008;
Coid et al., 2011; Coid et al., 2013), psychopathy and personality disorder (Coid et al., 2012;
Freestone et al., 2013), and mental illness (Coid et al., 2013). Overall, the model development
process first determined the structure and then the parameters of the model. The structure was
mainly based on expert knowledge while the parameters were learnt from data. We consider
these two stages in turn.
The primary steps for model development were: a) expert driven identification of
model variables which were considered to be important for estimating the risk of violence,
and b) expert constructed causal model structure based on the variables identified at step (a).
The model structure was divided into a number of key model components which we explain
in detail later in Section 3.
Much of the current research on BN construction assumes that sufficient data may be
available to make the experts’ input redundant. However, while 953 cases may seem like a lot
of data, it is actually insufficient for constructing a causal structure as complex as that
presented in Figure A.1 (for reasons explained in (Fenton, 2015)). Also, relying purely on
data-driven solutions in such problems can miss explanatory or intervention information. To
understand this, consider Figure 1 which presents three different models for head injury, and
demonstrates how an expert constructed BN model that incorporates both expert knowledge
and data (Figure 1c), can be more sensible than both a causal model learnt purely from data
(Figure 1b; Sakellaropoulos & Nikiforidis, 1999) and a standard data-driven statistical
regression predictor (Figure 1a).
The motivation behind Figure 1 is to demonstrate that purely data-driven models are
bound to fail in generating a sensible causal structure when important factors (e.g. in this
example “seriousness of injury”, and “treatment”) are absent from the dataset. When an
observation is provided into the causal network, the model informs us about factors that are
directly or indirectly causally linked to the observed event. This highlights the importance of
4
the causal structure for inference, and perhaps the requirement in spending effort with the
domain experts in ensuring that causality between factors in the BN flows sensibly.
a) Standard statistical regression model b) 'Causal' model learnt purely from c) Sensible causal model with missing/
learnt from data: Outcome= f(inputs) data (Sakellaropoulos & Nikiforidis, unobserved variables
1999)
Figure 1. The problem with data-driven models (Fenton, 2014).
2.2. Parameter learning
The model is parameterised using data from the PCS. The first step is to link relevant
questionnaire data to model variables, with the help of the experts, and a BN variable is
linked to one or more relevant questionnaire answers. For example, in the case of the variable
Financial difficulties, the sources of information for learning were answers provided to
questions "Are you behind paying bills?", "Have you recently had any services cut off?", and
"What is your average weekly income". We assume p(Financial difficulties=Yes) if evidence
of financial difficulties are observed for at least one of those responses.
The next step is to learn the model parameters. To deal with missing data we use the
Expectation Maximisation (EM) algorithm (Lauritzen, 1995). The experts were then asked to
review the model (by playing with the model in AgenaRisk), in terms of inferred outcomes at
different parts of the model, and suggest further revisions where necessary. In particular, after
model reviewing, revisions were normally suggested (or had to be performed) in cases where:
 disagreements between experts initially existed about the inclusion or not of

one or more variables in the model;
 disagreements between experts initially existed about the link (or the
direction of a link) between model variables;
 disagreements between experts initially existed about the formulation of one
or more model variables from questionnaire data;
 data indicated no effect between causally defined model variables;
 causal model links that were initially creating an endless loop for a set of
variables (cycles are not allowed in BNs);
 further analysis revealed very strong correlation between non-linked model
variables.
5
Since many of the steps were expert driven, disagreements between experts about
both model the structure and the variable identification were encountered due to the high
complexity of the domain. Extensive iterative process for expert knowledge elicitation
ensured eventual agreement between experts on both the structure of the model and the data
variables considered for inference.
3 THE MODEL
DSVM-P was built using the AgenaRisk BN tool (AgenaRisk, 2012). As well as the standard
discrete variables, AgenaRisk also supports continuous state variables which are
approximated using dynamic discretisation (Neil et al., 2010); we make use of this unique
feature for a number of variables as described later in this section.
The model is constructed on the basis of six generic factors: Criminal attitude,
Personality disorder, Socioeconomic factors, Mental illness, Substance misuse, and
Treatment responsivity. There are model components corresponding to each of the six factors.
A seventh component called Violence and other static risk factors links dynamic and static
risk factors for assessing violence. Figure 2 demonstrates a simplified model component
topology of the overall BN, and the complete BN model is presented in Figure A.1. Table B.1
provides detailed description of all the model variables. Note that, although at this schematic
level (Figure 2) there is a cycle, no cycles exist in the full model.
Figure 2. Simplified model component topology of the overall BN. Dual-directed links between components
indicate multiple dependencies between variables of one component to another.
We provide a brief description for each of the six model components and demonstrate
their direct interactions with child/parent nodes from other components in the subsections that
follow. In addition, we also provide a detailed description on the design of the seventh
component, which is responsible for linking all of the parts of the model for future violence
estimation. There are four categories of nodes/variables:
 Oval nodes with solid border representing observable variables;

 Oval nodes with dashed border representing definitional or any other latent variables;
 Square nodes with solid border representing interventions (i.e. treatments or
therapies);
6
 Square nodes with dashed border representing latent unobservable variables as a

result of an intervention (i.e. post-treatment effect).
3.1. Model component: Criminal attitude
Involvement in crime and a criminal lifestyle has been long known to be associated with
violence (Andrews & Bonta, 1994), either through the instrumental use of force by criminals
to obtain goals (e.g. in robbery) or through a tendency for criminal activities that may not be
violent in themselves (e.g. sale of illegal drugs) to be associated with a more violent lifestyle
due to operating outside the scope of the law (White, 1997). Involvement in criminal
activities is hypothesised to be positively influenced by the presence of criminal activity in
familial or peer groups, which may in turn lead causally to the development of attitudes
supportive of crime in an individual (Patterson et al., 1989).
Figure 3. Criminal attitude component and its direct interactions with child/parent nodes from other
components.
In the development of this component (Figure 3) we reasoned that markers of criminal

attitude (expression of criminal attitude; criminal network; criminal family background)
would be more causally related to violence if they were accompanied temporally by acute
risk factors such as anger, victimisation, the presence of violent thoughts, or gang
membership. Similarly but conversely, social withdrawal (e.g. due to symptoms of mental
illness as demonstrated later in Figure 6) may work protectively in this regard as it would
remove offenders from a context in which they may act violently as part of a criminal group.
3.2. Model component: Personality disorder
Personality disorders are chronic mental disorders which are characterised by a pervasive
pattern of disturbed thought and behaviour persisting from early adulthood (APA, 2013),
7
some of which have links with thought and behavioural patterns associated with violence. For
example, borderline personality disorder (BPD) is characterised by disturbed identity,
impulsive behaviour and self-harm; antisocial personality disorder (AsPD) is characterised
by high levels of anger and aggression, deception failure to obey social norms – for instance,
through criminality – and a lack of remorse. Arguably another form of personality
disfunction, psychopathy is not currently a medical diagnosis, but is an accepted condition
within forensic services measured by a 20-item checklist called the Psychopathy Checklist-
Revised (Hare, 2003) and comprising two separate but correlated factors each consisting of
two ‘facets’: Factor One is characterised by the absence of empathy and remorse (‘affective’
facet) together with interpersonally manipulative traits (‘interpersonal’ facet); and Factor
Two comprises mostly behavioural dysfunction relating partly to impulsivity (‘impulsive’
facet, or Facet Three) or the tendency to act without thinking; and criminality (‘antisocial’
facet). Some traits indicative of antisocial personality disorder - particularly impulsivity - are
shared by those comprising Factor Two of psychopathy (Coid & Ullrich, 2010). Where the
presence of Factor Two traits have been found to correlate directly with criminal violence
(Skeem & Mulvey, 2001), Factor One traits predict violence only weakly (Skeem et al.,
2002), but has a strong negative influence on treatment outcome (Olver et al., 2013).
Figure 4. Personality disorder component and its direct interactions with child/parent nodes from other
components.
When constructing this component (Figure 4), we considered personality disoders and
psychopathy to be static, lifetime constructs (in the manner suggested by (Douglas et al.,
2013) with potential antecedents in childhood abuse or neglect (Johnson et al., 1999) that
8
increase vulnerability to impulsive and aggressive behaviour – which in turn increase risk of
violence – and can interfere with treatment response. Anger within the component is
modelled as if it was a trait in personality disorder (what is known as ‘trait’ anger;
Spielberger & Sydeman, 1994); however in the dataset used for validation we only had
access to information about ‘state’ anger, which details the individual’s feelings of anger at
the time of interview. Using ‘state’ as a proxy for ‘trait’ anger may lead to some inaccuracies
as the individual may have been angry at the time of interview for legitimate reasons (length
of the interview; victimisation in prison, etc) unrelated to personality.
3.3. Model component: Socioeconomic factors
Low or unstable socioeconomic status has been shown to be associated with violent crime,
but only causally in the case of acute stress (i.e., hour-to-hour fluctuations in status such as
being made homeless) or in the context of a general ‘stain theory’, by which violence can be
explained as the product of multiple overlapping stressors upon an individual (Agnew, 1992).
In this model component (Figure 5) our intention was to model social stresses upon an
individual that might lead to violence in an attempt to cope – e.g. through robbery or
displaced aggression against family or friends – and to see how an individual’s social
resources – education, intelligence, social network – might counteract the effects of the stress.
Mental disorders such as anxiety or depression, which may also negatively influence an
individual’s ability to cope, were linked in from component 3.4 (below).
Figure 5. Socioeconomic factors component and its direct interactions with child/parent nodes from other
components.
9
In terms of individual resources, higher – or more stable - socioeconomic status,

including both high intelligence and higher levels of educational attainment, may act
protectively in terms of preventing an individual’s involvement in crime (de Vogel et al.,
2011). Evidence suggests that stable intimate relationships and appropriate, supervised living
circumstances for prisoners and patients nearing discharge are important factors in preventing
violence (Ullrich & Coid, 2011; de Vries Robbé et al., 2013).
3.4. Model component: Mental illness
Mental illness in this component refers to a specific set of mental disorders – mood disorders
or psychoses – that may differentially affect risk of violence. Mental illness and violence
have long been stereotypically linked in Western culture through archaic representations of
the ‘mental patient’ but the reality is that they have been said to have an ‘intricate link’ which
may be explained by other risk factors such as substance misuse (Elbogen & Johnson, 2009)
or may depend upon specific markers for mental illness such as childhood abuse or neglect
(van Dorn et al., 2012). In either case, effective treatment for mental illness is widely
understood to be critical in preventing violence in individuals with such a condition.
In constructing this component (Figure 6), our approach was to build nodes relating to
individual symptoms or traits of mental illness, rather than diagnostic categories. Diagnostic
categories can be difficult to ascertain to all but the best-trained of clinicians; and even then
reliability of diagnosis between clinicians can be very poor (McGorry, 2013). Further, recent
research has demonstrated that specific symptoms, rather than the clusters of symptoms
represented by diagnoses, may have links to violence, particularly when mediated by
affective states such as anger. Examples of this include: a subset of delusional beliefs being
causally linked to violence (Coid et al., 2013; Keers et al., 2013); or command hallucinations
directing the patient to harm others (McNiel et al., 2000).
Figure 6. Mental illness component and its direct interactions with child/parent nodes from other components.
10
3.5. Model component: Substance misuse
The substance misuse component (Figure 7) assesses the risk level for violent re-offending
based on the misuse of a number of drugs and/or hazardous alcohol consumption. Substance
abuse is a clinically identified psychiatric disorder characterised by distress caused to an
individual due to the use of a psychoactive drug (APA, 2013), including alcohol, that may
also manifest in extreme cases as substance dependence where it leads to increased need for
the drug. The relationship between substance abuse and violence is complex: it may be
causative in the sense that some stimulants directly increase aggressive or violent behaviour
through their psychopharmacological action (e.g. Davis, 1996); that substance abuse or
dependency stimulates acquisitive violence to fund addiction (‘economic compulsive
violence’; (Goldstein, 1985)) or it may be that use of illegal substances implies involvement
in social systems where violence is more likely (‘systemic violence’; Boles & Miotto, 2003).
Whatever the case, substance misuse has been found to increase risk of violence by up to four
times in most populations, particularly in individuals suffering from existing mental illness
(Steadman et al., 1998; Elbogen & Johnson, 2009).
Figure 7. Substance misuse component and its direct interactions with child/parent nodes from other
components.
11
3.6. Model component: Treatment responsivity
Poor adherence or response to treatment in individuals with severe mental illness, are known
risk factors for violence (Witt et al., 2013). Equally, the effect of successful treatment on
either substance misuse or symptoms of mental illness may be to nullify the relationship of
these disorders to violence by removing the underlying cause (addiction compulsion;
command hallucinations, etc).
The Treatment responsivity component is represented by two factors: 1) the
responsiveness to any given treatment, and 2) the risk of refusing or failing to attend any
given therapy. We have already demonstrated in the previous subsections how treatment
responsivity is individually linked to the components of mental illness, personality disorder,
substance misuse, and criminal attitude. Figure 8 demonstrates these links collectively.
Figure 8. Treatment responsivity component and its direct interactions with child/parent nodes from other
components.
3.7. Model component: Violence and other static risk factors
In the previous six subsections we have demonstrated the six model components
corresponding to each of the six dynamic factors. Four danger level variables and one
protective level variable are associated with these components. The four danger levels and the
protective level variables are binary defined with states Low and High, indicating relative low
and high risks for violent re-offend based on key-variables within those components.
Specifically, the danger indication will be High for the combination of observed key-
12
variables for which the highest rate of violence is observed, and vice versa. For any other
combination of observations, for which the rate of violence is between the minimum (Low)
and maximum (High), the probabilities for Low and High will adjust relatively (i.e.
Low=High=50% when the rate of violence is equal to the average rate).
Figure 9. Violence and other static risk factors component and its direct interactions with child/parent nodes
from other components.
The component presented in Figure 9 can be described in five steps. In brief, from
steps 1 to 3 the Violent convictions rate is inferred hierarchically and respectively for each
step, based on a) the danger levels, b) the protective level, and c) the number of days the
released prisoner has already spent out of prison (with or without evidence of violent re-
offence). Further, at step 4 the revised Violent convictions rate (step 3) is considered for
predicting the expected number of violent reconvictions over a specified period of time in the
future, before this information is revised at step 5 on the basis of the five static risk factors.
Each of the five steps is described in detail as enumerated below:
1. We assume that the Violent reconviction rate (step 1) follows a distribution

which is estimated on the basis of the four danger levels; where hyperparameter is the
number of violent convictions observed over the observation period and hyperparameter
is the observation period (in days) minus .
While the beta distribution assumes the combinations of Low and High for the danger
levels, we have instead provided the combinations of ¬High and High as demonstrated in
Appendix C. This was done to ensure that sufficient data points are generated for a
13
reasonably well informed prior for p(Violent reconviction rate (step 1)); if we were to
follow the proper set of combinations no prior information would had been available for
many of those combinations due to an insufficient number of instances in the dataset.
Further, the high complexity behind the definition of each danger level made conditional
probabilities between danger levels highly uncertain and not feasible for expert
probability elicitation.
2. A revised beta distributed Violent reconviction rate (step 2) is generated based on the
social protective level.
3. A revised beta distributed Violent reconviction rate (step 3) is generated based on Time
since initial release (assessed in number of days), and Violent reconvictions since initial
release. We assume that the three variables follow a Beta-binomial approach such that the
Beta distribution Violent reconviction rate (step 3) serves as conjugate distribution of the
distribution Violent reconvictions since initial release, formulating a
compound distribution such that the parameter is randomly drawn from the Beta
distribution. The variable Time since initial release serves as the input (in days) for the
Binomial distribution. Consequently, the process assumes constant probability † for
violence over each trial (day).
So, for example, if we are monitoring an individual over a period of two years and we
observe no evidence of violent re-offence, then our belief for that individual becoming
violent in the future diminishes (in comparison to what it was immediately after release).
Figure 10 demonstrates the reduction in the risk of violent re-offence over a period of
2,000 days with no evidence of violent re-conviction (and the prediction given assumes
further 2,000 days in the future; i.e. p(Time at risk=2000)). The reduction effect is subject
to exponential decay. For example, after 1,000 days out of prison, without evidence of
violence, the reduction is approximately 10 absolute percentile points (i.e. down to ~18%
from ~28%), whereas after further 1,000 days the risk is further reduced by 4.5 absolute
percentile points (i.e. down to ~13.5% from ~18%).
Figure 10. Risk reduction for p(Violence=Yes) over the specified number of days out of prison with no
evidence of violent reconviction. This assessment assumes p(Time at risk= 2,000).
Alternatively, Figure 11 demonstrates how the risk of violent re-offence would have
increased had we observed violent reconvictions for that individual and over the same
†
Time-series analysis was not possible with our dataset, and no other relevant published research study has
attempted to answer this question.
14
period (and with the same assumption for Time at risk). In this case, the increase in the
risk of violence follows a logarithmic growth. For example, when we observe 2 violent
reconvictions (after 2,000 days spent in the community) the risk of violence over the next
2,000 days follows an increase of a massive 56 absolute percentile points (as opposed to
observing 0 violent reconvictions), whereas in the case 4 violent reconvictions the risk of
violence is increased by an additional 19.5 absolute percentile points (which is still a
significant increase, but considerably lower than the increase in the first scenario).
Figure 11. Risk increase for p(Violence=Yes) over the specified number of violent reconvictions observed
over 2,000 days out of prison. This assessment assumes p(Time at risk=2,000).
4. A prediction for Violent convictions is generated on the basis of a repeated Beta-binomial

process, such that the Beta distribution Violent reconviction rate (step 3) serves as
conjugate distribution of the distribution Violent convictions,
formulating a compound distribution such that the parameter is randomly drawn from
the Beta distribution. The variable Time at risk serves as the input for the Binomial
distribution. Specifically, when we provide information for Time at risk DSVM-P
generates Binomial distributed prediction for the number of violent convictions expected
over the specified (time at risk) period.
5. The variable Violence indicates a binary prediction for future violent reconviction, and
which is translated from Violent reconvictions such that 0 violent reconvictions indicate
p(Violence=No) and 1≥ violent reconvictions indicate p(Violence=Yes). The prediction
for violence (and consequently violent reconvictions) is then revised based on the five
static risk factors of Age, Gender, PCLR Score, Prior violent convictions and Prior
acquisitive crime convictions. All the of above static risk factors serve as strong
predictors for violence but none of which serves as an underlying cause of violence.
4 RESULTS AND DISCUSSION
In this section we assess the performance of DSVM-P and comment on the results.
Specifically, Section 4.1 assesses the predictive accuracy of DSVM-P, Section 4.2 analyses
interventions, and Section 4.3 analyses the danger levels.
4.1. Assessment of predictive accuracy
While there are several scoring functions available to assess the predictive accuracy of a
probabilistic model of violent reoffending, the area under the curve (AUC) of a receiver
operating characteristic (ROC) is the standard method in this domain for binary predictive
15
distributions. Hence, we use the AUC of ROC to compare the predictive performance of
DSVM-P against other well-established probabilistic predictors in this area.
Some advantages, such as independence of both base rate and selection ratio, over
other measures are appreciated in this field (Hanley & McNeil, 1982a, 1982b; Rice & Harris,
1995), and in (Rice & Harris, 2005) the authors outlined why the AUC is the preferred
measure of predictive or diagnostic accuracy in forensic psychology or psychiatry. However,
the AUC has also been subject to criticism. Singh (2013) explains why AUCs do not capture
how well a risk assessment model’s predictions of risk agree with actual observed risk,
indicating that the AUCs provide an incomplete portrayal of predictive validity. While there
is a long debate in the literature (Lobo et al., 2007) on how to interpret AUCs, still more than
half of violence risk assessment validation studies report only the AUC (Singh, 2013) since
there is no other agreed measure for violence accuracy in this domain.
Typically, the AUCs are either reported based on the whole development sample or
based on a cross-validated sample. An AUC score of 0.5 indicates forecasting capability no
better than chance, whereas a score of 1 (or 0) corresponds to a perfect predictive model.
Evidently, AUCs reported on the whole development sample are likely to be
optimistic, especially when the model is optimised for the sample upon which they were
developed in which case running the danger of overfitting the model. DSVM-P generates an
AUC score of 0.79 (95% CI: 0.7552-0.8215. Performing a 10-fold cross-validation the AUC
score only drops to 0.78 (95% CI: 0.7449-0.8149). This suggests no danger for model
overfitting and that the predictive accuracy of DSVM-P is expected to be very good for other
similar data samples.
Table 1 shows how the cross-validated performance of the DSVM-P compares against
the three well established predictors within this area of research when employed with the
same dataset. All three predictors are used in clinical practice in the UK and internationally
and have been previously validated through the use of AUC statistics calculated against the
‘sum’ of the items as described above. These predictors are:
a) Violence Risk Appraisal Guide (VRAG, Quinsey 2006): is a model developed in

Canada for predicting reoffending by mentally ill offenders on the basis of 12
variables linked to violence risk – such as maladjustment at school or a history of
alcohol use – that correlated best with reoffending as determined by multiple
regression analysis (Harris et al., 1993). It has the advantages of brevity and
significant use in, and validation with, clinical justice populations, but does not
consider dynamic, or changeable items, associated with violence risk, and which can
be targeted for intervention.
b) Health-Clinical-Risk 20 (HCR-20; Webster et al, 1997): is a 20-item checklist of

static and dynamic risk factors associated with violence in psychiatric patients, such
as acute symptoms of mental illness or stable relationships after discharge (Webster et
al., 2005). The HCR-20 is an SPJ measure used primarily by clinicians seeking to
assess readiness for discharge amongst patients whose mental disorder is linked to
their offending: it is relatively quick to complete but requires training and extensive
clinical experience.
c) Psychopathy Checklist-Revised (PCL-R; Hare, 2003): is a checklist of 20 variables

measuring psychopathy, a disorder characterised by a callous lack of empathy and
remorse, shown to be strongly associated with offending behaviour in prisoner
populations, although some meta-analyses have drawn its performance as a risk
16
assessment measure into question (Singh, Grann & Fazel, 2011). Like the HCR20, it
requires extensive training and experience to use accurately.
Table 1. Comparison of AUC scores between the DSVM-P and the three well established predictors for violence
risk assessment, when employed with the same dataset.
Model AUC
DSVM-P 0.78
VRAG 0.7171
HCR-20v2 0.665
PCL-R 0.6648
DSVM-P provides a significant increase in predictive accuracy for violent recidivism, over
the three predictors discussed above. Figure 12 presents the partial AUCs for DSVM-P (left
graph) generated at 100-90% specificity and sensitivity, and superimposed ROC curves for
DSVM-P (95% CI), VRAG, HCR-20, and PCL-R predictors; indicating the significance
levels between DSVM-P and the other three predictors, as well as trade-off between
sensitivity and specificity.
Figure 12. Partial AUCs for DSVM-P (left graph), and superimposed ROC curves (right graph) for DSVM-P(95%
CI), PCL-R, VRAG, and HCR-20 predictors.
4.2. Analysis of the interventions
Table 2 below demonstrates the expected reduction in the risk of violence for each
intervention introduced in the model. Over each iteration, the what-if analysis (or sensitivity
analysis) assumes p(Violence=Yes) for five years forward (i.e. p(Time at risk=1,825)),
observable active symptoms for the intervention under analysis, and observable inactive
symptoms for the remaining three interventions (with all of the other model factors
unknown).
Assuming no intervention (i.e. no treatment/therapy), the results show that psychotic
symptoms generate a considerable higher risk for violence (i.e. 42.85%) over hazardous
drinking, drugs and anger. When intervention is advised, the results suggest that there is not
much difference between partial and full responsiveness to treatment over all four
interventions, and show that psychiatric treatment can be very effective with 42.88% relative
17
reduction in the risk of violent re-offence, followed by alcohol treatment with a relative risk
reduction of 24.43%, but drug treatment and anger management less effective. However, as
stated in Section 3.2, results relating to anger management should be interpreted with caution
due to the temporal unreliability of the ‘state’ model used to measure anger levels in sample
participants.
The same experiment is repeated, but this time the assumption is that the symptoms
associated with the remaining three interventions (over each iteration) are also unknown
(instead of inactive). As expected, the results (Table 3) demonstrate a decrease in intervention
effectiveness for all cases, but the relative impact between interventions remains similar to
that presented in Table 2. Repeating the experiment for a third time, but with all symptoms
associated with each of the interventions being active over all iterations, the results
demonstrated that none of the treatments was capable of individually providing any
meaningful reduction in the risk of violent re-offence; implying that the active symptoms
associated with the remaining three interventions (over each iteration) were strong enough to
maintain the risk for future re-offending at the same high risk level.
Table 2. Sensitivity analysis for p(Violence=Yes) assuming 5 years forward, with sensitivity variables each of the
four interventions assessed individually and relative to the specified treatment responsiveness. The analysis
assumes observable active symptoms for the intervention under analysis, and observable inactive symptoms
for the remaining three interventions, over each iteration.
Intervention Treatment="No" Responsiveness Responsiveness Sensitivity Reduction rate

for: (i.e. no to treatment to treatment to for
intervention) ="Partly" ="Yes" Violence p(Violence=Yes)
Alcohol 0.1392 0.0945 0.0789 0.0633 24.43%
Drugs 0.1028 0.0871 0.0822 0.0209 9.49%
Anger 0.1444 0.1388 0.1382 0.0062 4.29%
Psychiatric 0.4285 0.2655 0.2448 0.1838 42.88%
Table 3. Sensitivity analysis for p(Violence=Yes) assuming 5 years forward, with sensitivity variables each of the
four interventions assessed individually and relative to the specified treatment responsiveness. The analysis
assumes observable active symptoms only for the intervention under analysis, over each iteration.
Intervention Treatment="No" Responsiveness Responsiveness Sensitivity Reduction rate

for: (i.e. no to treatment to treatment to for
intervention) ="Partly" ="Yes" Violence p(Violence=Yes)
Alcohol 0.4183 0.3656 0.3483 0.0700 16.73%
Drugs 0.3610 0.3445 0.3395 0.0215 5.96%
Anger 0.3448 0.3374 0.3369 0.0079 2.29%
Psychiatric 0.4406 0.3505 0.3388 0.1018 23.10%
4.3. Analysis of the danger levels
Table 4 demonstrates the impact for each of the danger levels, when are individually and
collectively observed, for p(Violence=Yes), again assuming five years forward. The results
clearly demonstrate that the risk for future re-offending is extremely low when all of the four
danger levels indicate Low danger. When only one of the danger levels is observed as being
High, the substance misuse appears to be most dangerous with 32.44% probability for future
re-offence, whereas aggression the least dangerous with 17.38% probability. Combining two
High danger levels, the combination of aggression and mental illness appears to be
significantly more dangerous than residual combinations (with 63.55% probability for future
re-offence), whereas the combination of aggression and attitude (with 30.49% probability for
18
future re-offence) appears to be the least dangerous. Combining three High danger levels the
risk for future re-offence is increased under all scenarios; but for the combination of
aggression, mental illness and substance misuse the risk drops considerably. This result needs
further exploration, but could be due to the cluster of symptoms representing a disturbed but
non-criminal group of individuals whose aggression was associated with mental illness and
substance use but who mostly lacked motive or capacity for violence.
Table 4. Danger level analysis for p(Violence=Yes) over five years forward. A √ indicates High observable
danger level, and ¬High (or Low) otherwise (Appendix C provides more details on these combinations).
Aggression Attitude Mental Substance p(Violence=Yes)

illness misuse
- - - - 0.0242
- - - √ 0.3244
- - √ - 0.2600
- √ - - 0.2468
√ - - - 0.1738
- - √ √ 0.3798
- √ - √ 0.4901
√ - - √ 0.5405
- √ √ - 0.4318
√ - √ - 0.6355
√ √ - - 0.3049
- √ √ √ 0.6625
√ - √ √ 0.2569
√ √ - √ 0.7374
√ √ √ - 0.6578
√ √ √ √ 0.7784
Figure 13. Sensitivity analysis for target node p(Violence=Yes) on the basis of the nine specified sensitivity
nodes, where A is age, PVC is prior violent convictions, PACC is prior acquisitive crime convictions, SMDL is
substance misuse danger level, AtDL is attitude danger level, MIDL is mental illness danger level, AgDL is
aggression danger level, and G is gender. The analysis assumes that all four treatments are instantiated to
"No".
19
Figure 13 presents a sensitivity analysis for target node p(Violence=Yes) based on the
nine specified sensitivity nodes. The analysis assumes that all treatments are instantiated to
"No". The tornado graph reveals three apparent clusters of impact on future violence, based
on this BN structure. In the highest impact cluster we observe the factors of age, prior violent
convictions and PCL-R; in the second highest impact cluster we observe prior acquisitive
crime convictions and all four danger levels; whereas gender appears to be the least
significant factor of the nine considered. The tornado graph also demonstrates which state
corresponds to what increase/decrease for p(Violence=Yes). For example, when it comes to
the variable Age the state which results into the highest probability for p(Violence=Yes) is
"18-19", whereas the state "60+" generates the lowest probability.
5 MODEL BENEFITS & LIMITATIONS
This section provides a review of the benefits and limitations of the BN approach (and the
DSVM-P model in particular), and how these compare against the well-established classical
regression and rule-based predictors and methods within the domain of forensic psychiatry,
which represent the current state-of-the-art.
The practical and methodological benefits can be summarised as follows:
1. Improved accuracy: DSVM-P demonstrates significant improvements in the

predictive accuracy (cross-validated AUC score of 0.78), compared to the current
predictors that have been employed with the same dataset (AUC scores ranging from
0.665 to 0.717), with respect to whether a prisoner is determined suitable for release.
2. Interventional analysis: The BN approach generally allows for specific risk factors to
be targeted for causal intervention for risk management. In the specific case of
DSVM-P this is done by examining whether the risk of future re-offending can be
managed to acceptable levels as a result of one or more interventions, and this makes
the model useful in terms of answering complex clinical questions that are based on
unobserved evidence. This allows for analysis that goes beyond the predictive
accuracy and into improving risk management and decision support.
3. Inverse inference: In contrast to the current predictors, inference in BN models can

be performed from effect to cause rather than just from cause to effect. This unique
capability provides radically improved decision-support, since it enables extensive
what-if analysis and provides the decision maker with the ability to examine
surprising evidence. Notably, the capability of explaining away evidence of violence
and other related risk factors can be used in DSVM-P by decision makers to
investigate the reasons as to why a particular individual has been violent, when the
model was indicating otherwise.
4. More informative predictions: While, like other predictors, DSVM-P generates

binary predictions for violence (i.e. Yes/No), it also provides multinomial predicted
distributions for future violence (i.e. expected number of violent convictions).
DSVMP also provides predictions for future violent re-offence over a specified time
forward (i.e. for as little as one day, up to many years forward).
5. Handles missing evidence: While current predictors only consider what information
is available for predictive analysis, and hence ignore factors for which information is
20
unavailable, DSVM-P allows flexibility with model inputs due to the BN framework.
Further, when assessing a prisoner, missing evidence (i.e. when the prisoner does not
respond to specific questions) are not ignored but rather inferred from evidence
provided to other factors within the model, and which are linked (directly or
indirectly) to important unobserved variables;
6. Structural integrity: If required (i.e. in future studies, or when DSVM-P is learnt with
different datasets), expert knowledge can be easily incorporated for factors that are
important for prediction but which the historical database fails to capture. This allows
the model to retain its structure for future relevant studies, regardless how limited the
dataset might be in terms of the number of variables.
The practical and methodological limitations can be summarised as follows:
1. Extensive effort required for development: Developing a causal expert-driven BN,

such as DSVM-P, not only requires collaboration with domain experts, but also an
extensive iterative development process. Although recent research in BN modelling
(Fenton & Neil, 2012) has introduced a range of techniques for reducing the burden of
expert elicited models, this up-front development effort remains the primary barrier to
more widespread adoption of BNs.
2. Necessary use of subjectivity: Relying on expert judgment implies inevitable

subjectivity and also possible bias. This is partly addressed by using multiple experts.
This, in turn, may lead to disagreements between experts; especially with regards to
the causal structure of the model in such a highly complex domain.
3. Complexity: Because a causal expert-driven BN model incorporates additional

variables beyond what is in any dataset, the models are typically more ‘complex’
mathematically (although not conceptually) than the regression and rule-based
predictors. In DSVM-P, the model complexity results in a number of practical
limitations. Specifically:
a) some model variables could have been modelled with a higher number of states,
and others with a higher number of parent nodes, but this option was not feasible
due to insufficient data size;
b) the combination of the danger levels is modelled sub-optimally (Appendix C) to
ensure that sufficient data points are generated for a reasonably well informed
prior, and this approach is expected to generate slightly overestimated violent
reconviction rates;
c) DSVM-P assumes that there is a constant (daily) risk rate of violent re-offending
that does not vary with time.
The limitations 3(a) and 3(b) can be overcome with a sufficiently larger dataset, whereas
limitation 3(c) can be overcome when relevant data becomes available to allow time-series
analysis for the risk rate of future violent re-offence. Having appreciated the impact the data
size has on such a large and complex BN model, a richer dataset also promises even higher
forecasting capability and hence, superior decision support.
21
6 CONCLUDING REMARKS AND FUTURE WORK
We have presented a novel Bayesian network model, which we call DSVM-P, for risk
assessment and risk management of future violent re-offending for released prisoners who
suffer from mental illness. Specifically, in terms of risk assessment, the model can be used to
assess the risk of violence for a given individual, and over a specified time period assuming
release. In terms of risk management, the model can be used to examine whether the risk of
violence for the given individual can be managed to acceptable levels after release.
The need for such a system was evident by forensic medical practitioners and
scientists who work in this area of research and who, over a period of several years, remained
unimpressed by the decision support offered by the classical statistical, regression and rule-
based systems that still dominate this area of research (Coid et al., 2014). As a result, forensic
medical practitioners have identified the need to examine new ways of modelling that include
the representation of causal relationships. Hence, it was felt that causal BN models could
improve on the state-of-the-art. To our knowledge, this is the first BN system developed for
violence prevention management in forensic psychiatry. As a result, the implications of this
paper, even though it is simply a BN application of practical use, expand to both areas of
research (forensic medical sciences and expert and intelligent systems).
In terms of implications in forensic psychiatry and violence prevention research, the
resulting BN model presented in this paper provides an important step forward for decision
support and risk management. Clinicians and probation officers who work in these areas
would benefit from such a decision support system that handles the underlying complexity
and that is able to properly quantify uncertainty to improve risk management and decision
making by simulating the effect of potential interventions (e.g. treatments and therapies) for
prisoners who are about to be released.
In terms of implications in expert and intelligent systems research, we have shown
how an expertly constructed BN model, with parameter learning performed based on
complex questionnaire and interviewing patient data with missing values (data that was not
really suitable for causal analysis) is still capable of significantly outperforming the state-of-
the-art predictors within this area of research with respect to whether a mentally ill prisoner is
determined suitable for release. Specifically, the BN model demonstrates a cross-validated
AUC score of 0.78, and this compares well against well-established predictors such as the
VRAG, HCR20v2 and PCL-R, which demonstrate AUC scores ranging from 0.665 to 0.717
when employed with the same dataset (details in sections 4 and 5). The implications are
extended to the interventional modelling case in the sense that the BN demonstrates how
actions are supported by the model, with respect to determining whether a prisoner's risk of
violence can be managed to acceptable levels after release on the basis of some causal
intervention, such as treatment, therapy and/or medication. The outcome of this paper is in
general agreement with many other studies that demonstrate decision support benefits, in
various other domains, using probabilistic graphical models (de Melo & Sanchez, 2008;
Pourret et al., 2008; Lee et al., 2009; Lee & Park, 2010; Fenton & Neil, 2011; 2012;
Tarantola et al., 2012; Constantinou et al., 2012; 2013; Zhang et al., 2013; Di Pietro et al.,
2015; Perkusich, 2015).
In terms of AI and decision support research, the problem addressed in this paper is
typical of many critical decision-making scenarios (especially in medicine, forensics, and
transport safety assessment); specifically decision-makers are seeking improved methods for
prediction and risk assessment, but have either little relevant data, or have to rely on poorly
structured data. In such scenarios pure data-driven machine learning methods will not
produce models that provide the necessary accuracy and insights. However, in combination
with expert judgment, causal BNs provide the potential to do better. The method described in
22
this paper contributes to a new research framework for building BN models in such
situations. Extensive further research in this area is being carried out in the BAYES-
KNOWLEDGE project (Fenton, 2014).
Other planned research extensions will determine the usefulness of DSVM-P through
expert validation by carrying out pilot studies with clinicians and a qualitative assessment on
a graphical user interface which is planned for future development. The capability of BNs as
decision support tools will also be evaluated in individuals with serious mental health
problems who are about to be discharged from Medium Secure Services.
ACKNOWLEDGEMENTS
The authors were supported primarily by a Program Grant for Applied Research, program
RP-PG-0407-10500, from The National Institute for Health Research UK (NIHR) and also
EU grant ERC-2013-AdG339182-BAYES_KNOWLEDGE.
23
APPENDIX A
Figure A.1. The complete Bayesian network model.
24
APPENDIX B: Description of model variables
Table B.1. Description of the model variables.
Variable Node name Model Node type Node category Node states
No. component
1 Victimisation Labelled Observable No/Yes
2 Gang member Labelled Observable No/Yes
3 Criminal network Labelled Observable No/Yes
4 Criminal family Labelled Observable No/Yes
background Criminal
5 Criminal attitude attitude Labelled Observable No/Yes
6 Violent thoughts Labelled Observable No/Yes
7 Compliance with Labelled Observable No/Partial/Yes
supervision
8 Negative attitude Labelled Observable No/Partial/Yes
9 Attitude danger level Labelled Latent Low/High
10 Aggression danger Criminal Labelled Latent Low/High
level attitude/
Personality
disorder
11 ASPD Labelled Observable No/Yes
12 BPD Labelled Observable No/Yes
13 Abuse or neglect as a Labelled Observable No/Yes
child
14 Anger Labelled Observable No/Yes
15 Impulsivity Personality Labelled Observable No/Partial/Yes
16 PCLR factor 1 disorder Observable 0-16
17 PCLR factor 2 Observable 0-18
18 PCLR facet 3 Observable 0-10
19 Anger management Labelled Observable No/Yes
intervention
20 Anger management Labelled Latent No/Yes
given failure
21 Anger management Labelled Latent No/Yes
post-treatment
22 Intelligence Labelled Observable Extremely Low/
Borderline/
Low Average/Average/
High Average/Superior
23 Living circumstances Labelled Observable Homeless/Bail Hostel or
Shelter/Living alone/
Living with partner/
Socioecono Living with family or
mic factors friends/Other
24 Education Labelled Observable No/GCSE or O’Level/
A’Level+/Other
25 Stress Labelled Observable No/Yes
26 Financial difficulties Labelled Observable No/Yes
27 Employment or Labelled Observable No/Yes
training
28 Problematic life Labelled Observable No/Yes
events
25
29 Social withdraw Labelled Observable No/Yes

30 Ability to cope Labelled Observable Low/High
31 Domestic stability Labelled Observable Low/High
32 Social protective Labelled Latent Low/High
level
33 Symptoms of mental Labelled Observable No/Partial/Yes
illness
34 Depressive Labelled Observable No/Yes
symptoms
35 Anxiety Labelled Observable No/Yes
36 Thought insertion Labelled Observable No/Yes
37 Hallucinations Labelled Observable No/Yes
38 Strange experiences Labelled Observable No/Yes
39 Paranoid delusions Labelled Observable No/Yes
40 Psychiatric treatment Labelled Observable No/Yes
intervention
41 Depressive Labelled Latent No/Yes
symptoms post- Mental
treatment illness
42 Anxiety post- Labelled Latent No/Yes
treatment
43 Thought insertion Labelled Latent No/Yes
post-treatment
44 Hallucinations post- Labelled Latent No/Yes
treatment
45 Strange experiences Labelled Latent No/Yes
post-treatment
46 Paranoid delusions Labelled Latent No/Yes
post-treatment
47 Psychiatric treatment Labelled Latent No/Yes
given failure
48 Mental illness danger Labelled Latent Low/High
level
49 Cocaine before Labelled Observable No/Yes
prison sentence
50 Cannabis before Labelled Observable No/Yes
prison sentence
51 Ecstasy before prison Labelled Observable No/Yes
sentence
52 Cocaine during Labelled Observable No/Yes
prison sentence
53 Cannabis during Labelled Observable No/Yes
prison sentence
54 Ecstasy during prison Labelled Observable No/Yes
sentence
55 Cocaine after release Labelled Observable No/Yes
56 Cannabis after Labelled Observable No/Yes
release
57 Ecstasy after release Labelled Observable No/Yes
58 Hazardous drinking Labelled Observable No/Yes
after release
59 Cocaine dependence Labelled Observable No/Yes
60 Cannabis Labelled Observable No/Yes
dependence
Substance
61 Ecstasy dependence Labelled Observable No/Yes
26
62 Alcohol dependence misuse Labelled Observable No/Yes

63 Drug treatment Labelled Observable No/Yes
64 Alcohol treatment Labelled Observable No/Yes
intervention
65 Cocaine post- Labelled Latent No/Yes
treatment
66 Cannabis post- Labelled Latent No/Yes
treatment
67 Ecstasy post- Labelled Latent No/Yes
treatment
68 Hazardous drinking Labelled Latent No/Yes
post-treatment
69 Drug treatment given Labelled Latent No/Yes
failure
70 Alcohol treatment Labelled Latent No/Yes
given failure
71 Any drug Labelled Definitional No/Yes
dependence
72 Response given drug Labelled Latent No/Yes
dependence
73 Response given Labelled Latent No/Yes
alcohol dependence
74 Substance misuse Labelled Latent Low/High
danger level
75 Responsiveness to Labelled Observable No/Partly/Yes
treatment Treatment
76 Refuse or fail to responsivity Labelled Latent No/Yes
attend therapy
77 Prior acquisitive Labelled Observable 0-2/3-12/13+
crime convictions
78 Prior violent Labelled Observable 0/1/2-5/6+
convictions
79 PCLR total score Violence Labelled Observable 0-9/10-16/17-26/27+
80 Age and other Labelled Observable 18-19/20-21/22-25/
static risk 26-29/30-34/35-39/
factors 40-49/50-59/60+
81 Gender Labelled Observable Female/Male
82 Time at risk Observable 0-5000
83 Violent reconvictions Latent 0-1
rate (step 1)
84 Violent reconvictions Revised Latent 0-1
rate (step 2)
85 Violent reconvictions Revised Latent 0-1
rate (step 3)
86 Time since initial Observable 0-5000
release
87 Violent reconvictions Observable 0-inf
since initial release
88 Violent convictions Latent 0-inf
89 Violence Labelled Latent No/Yes
27
APPENDIX C: Combining danger level indications
Let us assume the combination where L is the input for the danger level of
substance misuse. Figure C.1 shows that the substance misuse danger level is low (i.e.
High ) when we observe N (i.e. no substance use) and high (i.e. High ) when we
observe AD (i.e. both alcohol and drug use), whereas for combinations of A and D the danger
level is uncertain. Hence, by providing the prior information of combination
for combination , the model considers all the combinations between N, A and D
(i.e. ¬High) iteratively, instead of simply N (i.e. Low), against the other three component
danger levels.
This sub-optimal approach was only introduced due to insufficient number of
instances in our dataset; it can be safely ignored for datasets with sufficiently larger number
of instances. It should also be noted that while the naive Bayesian classification could have
also been introduced to effectively deal with the insufficient sample size, it was considered
inappropriate (due to its naive independence assumptions) for this case, since we were only
interested in modelling the violence rate based on the combinations of those danger levels.
Figure C.1. Substance misuse danger level indications based on drug and alcohol use, where N indicates no
substance use, A indicates only alcohol use, D indicates only drug use, and AD indicates both alcohol and drug
use.
Table C.1. Danger level combinations provided for the

distribution Violence reconvictions rate(step 1).
Combinations Combinations
assumed by provided to
the model the model
L,L,L,L L,L,L,L
H,L,L,L H,¬H, ¬H, ¬H
L,H,L,L ¬H,H, ¬H, ¬H
L,L,H,L ¬H, ¬H,H, ¬H
L,L,L,H ¬H, ¬H, ¬H,H
H,H,L,L H,H, ¬H, ¬H
H,L,H,L H, ¬H,H, ¬H
H,L,L,H H, ¬H, ¬H,H
L,H,H,L ¬H,H,H, ¬H
L,H,L,H ¬H,H, ¬H,H
L,L,H,H ¬H, ¬H,H,H
H,H,H,L H,H,H, ¬H
H,H,L,H H,H, ¬H,H
H,L,H,H H, ¬H,H,H
L,H,H,H ¬H,H,H,H
H,H,H,H H,H,H,H
28
REFERENCES
AgenaRisk. (2002). Bayesian Network and Simulation Software for Risk Analysis and Decision Support.
Retrieved February 5, 2015, from http://www.agenarisk.com
Agnew, R. (1992). Foundation for a General Strain Theory. Criminology 30(1), 47-87
American Psychiatric Association (APA). (2013). Diagnostic and Statistical Manual of Mental Disorders: 5th
edition (DSM-5). Arlington. VA: Amer Psychiatric Pub Incorporated.
Boles, S. M., & Miotto, K. (2003). Substance abuse and violence: A review of the literature. Aggression and
Violent Behavior, 8(2): 155-174.
Boylan, J. F., Kavanagh, B. P., & Armitage, J. (2011). Randomised controlled trials: important but overrated?
Journal of the Royal College of Physicians, 41: 126–31.
Cartwright, N., & Munro, E. (2010). The limitations of randomized controlled trials in predicting effectiveness.
Journal of Evaluation in Clinical Practice, 16: 260-266.
Coid, J.W., Yang, M., Ullrich, S., Zhang, T., Sizmur, S., Roberts, C., & Rogers, R. D. (2009). Gender
differences in structured risk assessment: Comparing the accuracy of five instruments. Journal of
Consulting and Clinical Psychology, 77(2): 337-348.
Coid, J. W., & Ullrich, S. (2010). Antisocial personality disorder is on a continuum with psychopathy.
Comprehensive Psychiatry, Vol. 51, Iss. 4: 426-433.
Coid, J. W., Yang, M., Ullrich, S., Zhang, T., Sizmur, S., Farrington, D., et al. (2011). Most items in structured
risk assessment instruments do not predict violence. The Journal of Forensic Psychiatry & Psychology,
22(1), 3-21.
Coid, J., Ullrich, S., & Kallis, C. (2013). Predicting future violence among individuals with psychopathy. British
Journal of Psychiatry, 203: 387–388. doi: 10.1192/bjp.bp.112.118471.
Coid, J. W., Ullrich, S., Kallis, C., Keers, R., Barker, D., Cowden, F., et al. (2013). The relationship between
delusions and violence: findings from the East London first episode psychosis study. JAMA Psychiatry,
70(5), 465-471.
Coid, J., Freestone, M. & Ullrich, S. (2012) Subtypes of psychopathy in the British household population:
findings from the national household survey of psychiatric morbidity. Social Psychiatry and
Psychiatric Epidemiology, 47(6): 879-891.
Coid, J. W., Ullrich, S., Kallis, C., Freestone, M., Gonzalez, R., Bui, L., Igoumenou, A., Constantinou, A. C.,
Fenton, N., Marsh, W., Yang, M., DeStavola, B., Hu, J., Shaw, J., Doyle, M., Archer-Power, L.,
Davoren, M., Osumili, B., McCrone, P., Barrett, K., Hindle, D., & Bebbington P. (2015). Improving
Risk Management in Mental Health Services – A Multi-Methods Approach. The National Institute for
Health Research (NIHR), UK.
Constantinou, A. C., Fenton, N. E. & Neil, M. (2012). pi-football: A Bayesian network model for forecasting
Association Football match outcomes. Knowledge-Based Systems, 36: 322, 339.
Constantinou, A. C., Fenton, N. E. & Neil, M. (2013). Profiting from an inefficient Association Football
gambling market: Prediction, Risk and Uncertainty using Bayesian networks. Knowledge-Based
Systems, 60-86.
Constantinou, A. C., Fenton, N. E., & Pollock, L. J. H. (2014). Bayesian networks for unbiased assessment of
referee bias in Association Football. Journal of Psychology of Sport and Exercise, Vol. 15, 5: 538-547.
de Campos, L. M., Fernández-Luna, J. M., & Huete, J. F. (2004). Bayesian networks and information retrieval:
an introduction to the special issue. Information Processing & Management, 40 (5): 727–733.
de Melo, A. C. V., & Sanchez, A. J. (2008) Software maintenance project delays prediction using Bayesian
Networks. Expert Systems with Applications, 34: 908-919.
de Vogel, V., de Vries Robbé, M., de Ruiter, C., & Bouman, Y.H.A. (2011). Assessing protective factors in
forensic psychiatric practice. Introducing the SAPROF. International Journal of Forensic Mental
Health, 10, 171-177.
de Vries Robbé, M., de Vogel, V., & de Spa, E. (2011). Protective factors for violence risk in forensic
psychiatric patients. A retrospective validation study of the SAPROF. International Journal of
Forensic Mental Health, 10, 178-186.
Davis. W.M. (1996) Psychopharmacologic violence associated with cocaine abuse: kindling of a limbic
dyscontrol syndrome? Progress in Neuro-Psychopharmacology & Biological Psychiatry, 20 (8) pp.
1273–1300
Di Pietro, L., Guglielmetti Mugion, R., Musella, F., Renzi, M. F., & Vicard, P. (2015). Reconciling internal and
external performance in a holistic approach: A Bayesian network model in higher education. Expert
Systems with Applications, 42: 2691-2702.
Díez,, F. J., Mira, J., Iturralde, E., & Zubillaga, S. (1997). DIAVAL, a Bayesian expert system for
echocardiography. Artificial Intelligence in Medicine, 10 (1): 59–73.
29
Douglas, K. S., Yeomans, M. & Boer, D. P. (2005). Comparative validity analysis of multiple measures of
violence risk in a sample of criminal offenders. Criminal Justice and Behaviour, 32: 479-510.
Douglas, K. S., Hart, S. D., Webster, C. D., & Belfrage, H. (2013). HCR-20 (Version 3): Assessing risk of
violence – User guide. Burnaby, Canada: Mental Health, Law, and Policy Institute, Simon Fraser
University.
Doyle, M., Archer-Power, L., Coid, J., Kallis, C., Ullrich, S. & Shaw, J. (2014) Predicting Post-Discharge
Community Violence in England and Wales Using the HCR-20v3. International Journal of Forensic
Mental Health, 13: 140-147. DOI: 10.1080/14999013.2014.906517
Elbogen, E. B., & Johnson, S. C. (2009). The intricate link between violence and mental disorder: results from
the National Epidemiologic Survey on Alcohol and Related Conditions. Archives of General
Psychiatry, 66(2): 152-161.
Fazel, S., Singh, J.P., & Grann, M. (2012) Use of risk assessment instruments to predict violence and antisocial
behaviour in 73 samples involving 24 827 people: systematic review and meta-analysis. BMJ, 345. doi:
http://dx.doi.org/10.1136/bmj.e4692
Fenton, N.E. & Neil, M. (2011). Avoiding Legal Fallacies in Practice Using Bayesian Networks. Australian
Journal of Legal Philosophy, 36: 114-151.
Fenton, N.E. & Neil, M. (2012). Risk Assessment and Decision Analysis with Bayesian Networks. CRC Press.
Fenton, N. (2014). Effective Bayesian Modelling with Knowledge Before Data. Retrieved April 29, 2015,
from BAYES_KNOWLEDGE: https://www.eecs.qmul.ac.uk/~norman/projects/B_Knowledge.html
Fenton, N. (2015). The problem with big data and machine learning. Retrieved April 29, 2015 from Probability
and Risk: http://probabilityandlaw.blogspot.co.uk/2015/03/the-problem-with-big-data-and-
machine.html
Fenton, N. E., Lagnado D., & Neil M. (2013). A General Structure for Legal Arguments Using Bayesian
Networks. Cognitive Science, 37: 61-102.
Fox, A. & Freestone, M (2008) An evaluation of the Good Lives and Discipline (GLAD) Programme at a UK
DSPD Unit. Prison Service Journal, 151
Freestone, M., Howard, R., Coid, J.W. & Ullrich, S. (2013) Adult antisocial syndrome with comorbid borderline
pathology: association with antisocial and violent outcomes. Personality and Mental Health, 7(1): 11-
21.
Friedman, N.F., Linial, M., Nachman, I., & Pe’er, D. (2000). Using Bayesian Networks to Analyze Expression
Data. Journal of Computational Biology, 7 (3/4): 601–620.
Goldstein, P.J. (1985). The Drugs/Violence Nexus: A tripartite conceptual framework. Journal of Drug Issues,
(39): 493-506.
Grann, M., Langstrom, N., Tengstrom, A., & Kuligren, G. (1999). Psychopathy (PCL-R) Predicts Violent
Recidivism Among Criminal Offenders with Personality Disorders in Sweden. Law and Human
Behavior, 23: 2, 205-218.
Hanley, J.A. & McNeil, B.J. (1982a). Maximum attainable discrimination and the utilization of radiologic
examinations. Journal of Chronic Disorders 35: 601–611.
Hanley, J.A. & McNeil B.J. (1982b). The meaning and use of the area under a receiver operating characteristic
(ROC) curve. Radiology, 143: 29–36.
Hanson, R. K., & Harris, A. J. (2000). Where Should We Intervene?: Dynamic Predictors of Sexual Offence
Recidivism. Criminal Justice and Behavior, 27(6): 6-34.
Harris, G.T., Rice, M.E., & Quinsey, V.L. (1993). Violent recidivism of mentally disordered offenders: the
development of a statistical prediction instrument. Criminal Justice and Behavior 20: 315–335.
Hare, R. D. (2003). Manual for the Hare Psychopathy Checklist-Revised: 2nd Edition. Toronto, Canada: Multi-
Health Systems.
Heckerman, D., Breese, J., & Nathwani, B. (1992). Toward normative expert systems I: the PATHFINDER
project. Methods of Information Medicine, 31: 90–105.
Heckerman D., Mamdani, A. & Wellman, M. (1995). Real-world applications of Bayesian networks.
Communications of the ACM, vol. 38, no. 3, pp. 25-26.
Hohenner, M., Wachsmuth, S., & Sagerer, G. (2005). Modelling expertise for structure elucidation in organic
chemistry using Bayesian networks, Knowledge-Based Systems, 18 (4–5): 207–215.
Jensen, F. V. (1996). An introduction to Bayesian networks. New York Springer.
Jiang, X., Neapolitan, R.E., Barmada, M.M., & Visweswaran, S. (2011). Learning Genetic Epistasis using
Bayesian Network Scoring Criteria. BMC Bioinformatics, 12: 89.
Johnson, J.G., Cohen, P., Brown, J., Smailes, E.M. & Bernstein, D.P. (1999) Childhood Maltreatment Increases
Risk for Personality Disorders During Early Adulthood. Archives of General Psychiatry 56 (7): 600-
606.
Lauritzen, S. L. (1995). The EM algorithm for graphical association models with missing data. Computational
Statistics & Data Analysis, 19: 191-201.
30
Lee, E., Park, Y., & Shin, J. G. (2009). Large engineering project risk management using a Bayesian belief
network. Expert Systems with Applications, 36: 5880-5887.
Lee, K. C., & Park, B. (2010). A general Bayesian network approach to analyzing online game item values and
its influence on consumer satisfaction and purchase intention. Communications in Computer and
Information Science, 114: 53-62.
Lobo, J. M., Jimenez-Valverde, A., & Real, R. (2007). AUC: a misleading measure of the performance of
predictive distribution models. Global Ecology and Biogeography, 1-7.
Kroner, C., Stadtland, C., Eidt, M. & Nedopil, N. (2007). The validity of the Violence Risk Appraisal Guide
(VRAG) in predicting criminal recidivism. Criminal Behaviour and Mental Health, 17: 89−100.
McGorry, P. D. (2013), The next stage for diagnosis: validity through utility. World Psychiatry, 12: 213–215.
McNeil D, Gregory A, Lam J, Binder, R.L. & Sullivan, G.R (2003) Utility of decision support tools for
assessing acute risk of violence. Journal of Consulting and Clinical Psychology 71:945—95
McNeil, D., Eisner, J.P. & Binder, R.L. (2000) The Relationship Between Command Hallucinations and
Violence. Psychiatric Services, 51(10): 1288-1292.
Mears, D. P. & Cochran, J. C. (2013). What is the effect of IQ on offending? Criminal Justice and Behavior,
doi: 10.1177/0093854813485736.
Monahan, J. (1984). The prediction of violent behaviour: Toward a second generation of theory and policy. The
American Journal of Psychiatry, 141, 10-15.
Neil, M., Marquez, D. & Fenton, N. E. (2010). Improved Reliability Modeling using Bayesian Networks and
Dynamic Discretization. Reliability Engineering & System Safety, 95(4), 412-425.
Nikovski, D. (2000). Constructing Bayesian networks for medical diagnosis from incomplete and partially
correct statistics. IEEE Transaction on Knowledge and Data Engineering, 12, 4: 509-516.
Olver, M., Lewis, K. & Wong, S.C. (2013) Risk reduction treatment of high-risk psychopathic offenders: the
relationship of psychopathy and treatment change to violent recidivism. Journal of Personality
Disorders
Pearl, J. (1988). Probabilistic reasoning in intelligent systems: networks of plausible inference. San Mateo,
California, Morgan Kaufmann Publishers.
Pearl, J. (2000). Causality: Models, Reasoning and Inference. Cambridge University Press.
Pearl, J. (2009). Causality: Models, Reasoning and Inference, 2nd Edition. CUP.
Perkusich, M., Soares, G., Almeida, H., & Perkusich, A. (2015). A procedure to detect problems of processes in
software development projects using Bayesian networks. Expert Systems with Applications, 42: 437-
450.
Pourret, O., Naim, P., & Marcot, B. (2008). Bayesian Networks: A Practical Guide to Applications. Chichester,
UK: Wiley.
Quinsey, V.L. (1998) Violent Offenders – Appraising and Managing Risk. Washington, DC: American
Psychological Association.
Rice, M. E. & Harris, G. T. (1995). Violent recidivism: assessing predictive validity. Journal of Consulting and
Clinical Psychology 63: 737–748.
Rice, M. E., & Harris, G. T. (2005). Comparing Effect Sizes in Follow-Up Studies: ROC Areas, Cohen’s d, and
r. Law and Human Behavior, 29: 5, 615-620
Sakellaropoulos, G. C., & Nikiforidis, G. C. (1999). Development of a Bayesian Network for the prognosis of
head injuries using graphical model selection techniques. Methods Inf Med, 38(1): 37-42.
Sansone, R. A., Songer, D. A., & Gaither, G. A. (2000). Medically self-harming behavior and its relationship to
borderline personality among psychiatric inpatients. The Journal of Nervous and Mental Disease, 188:
384-386.
Singh JP, Grann M, Fazel S. (2011) A comparative study of violence risk assessment tools: a systematic review
and metaregression analysis of 68 studies involving 25,980 participants. Clinical Psychology Reviews,
31(3):499-513. doi: 10.1016/j.cpr.2010.11.009
Singh, J. P. (2013). Predictive Validity Performance Indicators in Violence Risk Assessment: A Methodological
Primer. Behavioral Sciences and the Law, 31: 8-22.
Singh JP, Grann M, Fazel S (2013) Authorship Bias in Violence Risk Assessment? A Systematic Review and
Meta-Analysis. PLoS ONE 8(9): e72484. doi:10.1371/journal.pone.0072484
Skeem, J., Monahan, J., & Mulvey, E. (2002).Psychopathy, treatment involvement, and subsequent violence
among civil psychiatric patients. Law and Human Behavior, 26, 577-603
Skeem, J. & Mulvey, P. (2001) Psychopathy and community violence among civil psychiatric patients: Results
from the MacArthur Violence Risk Assessment Study. Journal of Consulting & Clinical Psychology,
69, 358-374
Smith, G. C., & Pell, J. P. (2003). Parachute use to prevent death and major trauma related to gravitational
challenge: systematic review of randomised control trials. British Medical Journal, 327: 1459-1461.
31
Spiegelhalter, D. J., Abrams, K. R., & Myles, J. P. (2004). Bayesian Approaches to Clinical Trials and Health-
care Evaluation. John Wiley and Sons.
Spielberger, C. D., & Sydeman, S. J. (1994). State-Trait Anxiety Inventory and State-Trait Anger Expression
Inventory. In M. E. Maruish (Ed.), The use of psychological testing for treatment planning and
outcome assessment. (pp. 292-321). Hillsdale, NJ: Erlbaum.
Stadtland, C., Kleindienst, N., Kröner, C., Eidt, M. & Nedopil, N. (2005). Psychopathic traits and risk of
criminal recidivism in offenders with and without mental disorders. International Journal of Forensic
Psychiatry 4: 89–97.
Stadtland, C. & Nedopil, N. (2004). Vergleichende Anwendung heutiger Prognoseinstrumente zur Vorhersage
krimineller Rückfälle bei psychiatrisch begutachteten Probanden. Monatsschrift für Kriminologie und
Strafrechtsreform 2: 77–85.
Steadman, H., Mulvey, E., Monahan, J., Robbins, P., Appelbaum, P., Grisso, T., Roth, L., & Silver, E. (1998).
Violence by people discharged from acute psychiatric inpatient facilities and by others in the same
neighborhoods. Archives of General Psychiatry, 55, 393-401.
Tarantola, C., Vicard, P., & Ntzoufras, I. (2012). Monitoring and improving Greek banking services using
Bayesian Networks: An analysis of mystery shopping data. Expert Systems with Applications, 39:
10103-10111.
Taroni, F., Aitken, C., Garbolino, P., & Biedermann, A. (2014). Bayesian Networks and Probabilistic Inference
in Forensic Science (2nd ed.). Chichester, UK: John Wiley & Sons.
Troquete, N.A., van den Brink, R.H., Beintema, H., Mulder, T., van Os, T.W., Schoevers, R.A., & Wiersma, D.
(2014) Predictive Validity of the Short-Term Assessment of Risk and Treatability for Violent
Behavior in Outpatient Forensic Psychiatric Patients. Psychological Assessment, Dec 15 2014. doi:
10.1037/a0038270
Ullrich, S., & Coid, J. W. (2011). Protective factors for violence among released prisoners - Effects over time
and interactions with static risk. Journal of Consulting Clinical Psychology, 79(3): 381-390.
Van Dorn, R., Volavka, J., & Johnson, N. (2012). Mental disorder and violence: is there a relationship beyond
substance use? Social Psychiatry and Psychiatry Epidemiology, 47(3): 487-503.
Webster, C.D., Douglas, K.S., Eaves, D.E. & Hart, S.D. (1997) HCR-20: Assessing Risk for Violence, Version 2.
Vancouver, BC: Simon Fraser University.
Witt, K., van Dorn, R. & Fazel, S. (2013) Risk Factors for Violence in Psychosis: Systematic Review and Meta-
Regression Analysis of 110 Studies. PLoS ONE, 8(2), doi:10.1371/journal.pone.0055942
Wong, S. & Gordon, A. (2003) Violence Risk Scale (VRS) Version 2. Saskatoon, CA: University of Saskatoon.
Yang, M., Wong, S.C., & Coid, J. (2010) The efficacy of violence prediction: a meta-analytic comparison of
nine risk assessment tools. Psychological Bulletin, 136 (5): 740-767. doi: 10.1037/a0020473.
Yet, B., Perkins, Z., Marsh, W., & Fenton, N. (2011). Towards a Method of Building Causal Bayesian Networks
for Prognostic Decision Support. Probabilistic Problem Solving in Biomedicine, Vol. 107.
Zhang, L., Wu, X., Ding, L., Skibniewski, M. J., & Yan, Y. (2013). Decision support analysis for safety control
in complex project environments based on Bayesian Networks. Expert Systems with Applications, 40:
4273-4282.
32

Risk Assessment and Risk Management of Violent Reoffending Among Prisoners (PDFDrive)

Uploaded by

Copyright:

Available Formats

Risk Assessment and Risk Management of Violent Reoffending Among Prisoners (PDFDrive)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Risk Assessment and Risk Management of Violent Reoffending Among Prisoners (PDFDrive)

Uploaded by

Copyright:

Available Formats

Version 4.

3, 26/05/2015; To appear in Expert Systems with Applications, 2015,

Risk assessment and risk management of violent

THIS IS A PRE-PUBLICATION DRAFT OF THE FOLLOWING CITATION:

2 DATA AND METHODOLOGY

Extensive statistical analysis of cohort data, primarily focusing on classification, has

2.1. Expert-driven causal structure

Figure 1. The problem with data-driven models (Fenton, 2014).

2.2. Parameter learning

 disagreements between experts initially existed about the inclusion or not of

 Oval nodes with solid border representing observable variables;

 Square nodes with dashed border representing latent unobservable variables as a

3.1. Model component: Criminal attitude

In the development of this component (Figure 3) we reasoned that markers of criminal

3.2. Model component: Personality disorder

3.3. Model component: Socioeconomic factors

In terms of individual resources, higher – or more stable - socioeconomic status,

3.4. Model component: Mental illness

3.5. Model component: Substance misuse

3.6. Model component: Treatment responsivity

3.7. Model component: Violence and other static risk factors

1. We assume that the Violent reconviction rate (step 1) follows a distribution

4. A prediction for Violent convictions is generated on the basis of a repeated Beta-binomial

4 RESULTS AND DISCUSSION

4.1. Assessment of predictive accuracy

a) Violence Risk Appraisal Guide (VRAG, Quinsey 2006): is a model developed in

b) Health-Clinical-Risk 20 (HCR-20; Webster et al, 1997): is a 20-item checklist of

c) Psychopathy Checklist-Revised (PCL-R; Hare, 2003): is a checklist of 20 variables

4.2. Analysis of the interventions

Intervention Treatment="No" Responsiveness Responsiveness Sensitivity Reduction rate

Intervention Treatment="No" Responsiveness Responsiveness Sensitivity Reduction rate

4.3. Analysis of the danger levels

Aggression Attitude Mental Substance p(Violence=Yes)

5 MODEL BENEFITS & LIMITATIONS

The practical and methodological benefits can be summarised as follows:

1. Improved accuracy: DSVM-P demonstrates significant improvements in the

3. Inverse inference: In contrast to the current predictors, inference in BN models can

4. More informative predictions: While, like other predictors, DSVM-P generates

The practical and methodological limitations can be summarised as follows:

1. Extensive effort required for development: Developing a causal expert-driven BN,

2. Necessary use of subjectivity: Relying on expert judgment implies inevitable

3. Complexity: Because a causal expert-driven BN model incorporates additional

6 CONCLUDING REMARKS AND FUTURE WORK

Figure A.1. The complete Bayesian network model.

APPENDIX B: Description of model variables

Table B.1. Description of the model variables.

29 Social withdraw Labelled Observable No/Yes

62 Alcohol dependence misuse Labelled Observable No/Yes

APPENDIX C: Combining danger level indications

Table C.1. Danger level combinations provided for the

You might also like