|UFDC Home||myUFDC Home | Help|
This item has the following downloads:
1 BUILDING A BETTER MODEL OF VIOLENCE PREDICTION: CLINICAL PREDICTORS OF AGGRESSION AND VIOLENCE AMONG RESIDENTS OF A MAXIMUM SECURITY STATE FORENSIC PSYCHIATRIC FACILITY By DOMINIQUE D. DELALOT A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010
2 2010 Dominique D. Delalot
3 To those who might ben e fit from this work
4 ACKNOWLEDGMENTS I would like to extend my sincere thanks to the many people who have played an integral part in the development and execution of this work: First, I would like to acknowledge my dissertation chair and supervisor Eileen Fennell for her encouragement and support throughout my traini ng. Her advice and direction have been greatly appreciated. I would like to thank my committee members, Duane Dede, Christiana Leonard, and Russell Bauer, for their integral contributions in the delivery of this work. I wish to thank my colleagues at the North Florida Evaluation and Treatment Center, particularly members of the Department of Psychology for their guidance and support throughout thi s project. Specific mention needs to be given to members of the Information Services department for their invaluable assistance during collection of these data. I am also deeply indebted to my friends and classmates, without whom the years would have been much less kind. Finally, and most of all, I wo uld like to thank my mom for her boundless faith, patience, and love, and for always encouraging me to explore.
5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ...............................................................................................................4 L IST OF TABLES ...........................................................................................................................8 LIST OF FIGURES .........................................................................................................................9 ABSTRACT ...................................................................................................................................10 CHAPTER 1 INTRODUCTION ..................................................................................................................12 2 REVIEW OF THE LITERATURE ........................................................................................15 Historical Overview of Violence Prediction ...........................................................................15 The Development of Violence Prediction .......................................................................15 Assessing Curr ent Dangerousness ..................................................................................16 The Problem of Defining Violence .................................................................................18 Research on Prediction General Issues ........................................................................19 Older Studies of Violence Risk Releva nt Variables ....................................................21 Newer Studies of Violence Risk Relevant Variables ...................................................22 Risk Factors For Violence ......................................................................................................25 History of Violence .........................................................................................................25 Socio D emographic Factors ............................................................................................26 Environmental Factors ....................................................................................................28 Substance Abuse .............................................................................................................29 Mental Illness ..................................................................................................................30 Psychosis .........................................................................................................................31 Mood Disorders ...............................................................................................................35 Personality Factors ..........................................................................................................36 Cognitive Impairment .....................................................................................................38 Clinical Judgment in Risk Assessment ...................................................................................40 Actuarial Versus Clinical Methods of Violence Prediction ............................................43 Risk Assessment Instruments ..........................................................................................46 3 MATERIALS AND METHODS ...........................................................................................51 Methodology ...........................................................................................................................51 Study Setting ...................................................................................................................51 Data Collection ................................................................................................................52 Measures .................................................................................................................................53 The Overt Aggres sion Scale (OAS) ....................................................................................53 Historical/Clinical/Risk Management 20 Item Scale (HCR 20) ........................................59
6 Assessment of Psychiatric Symptoms Using the Brief Psychiatric Rating Scale (BPRS) ................................................................................................................................60 Analytic Strategy ....................................................................................................................63 Treatment of Outliers ..............................................................................................................66 4 RESULTS ...............................................................................................................................70 Univariate Analyses ................................................................................................................72 Multivariate Analyse s .............................................................................................................74 Aim 1 Predicting Aggression vs. Non A ggression .......................................................75 Normality ....................................................................................................................75 Model A .......................................................................................................................76 Model B ......................................................................................................................79 Test Model (Model T) .................................................................................................80 Graphical Evaluation of Predictive Models ................................................................84 Aim 2 Predicting Aggression Severity .........................................................................87 Normality and Statistical Considerations ....................................................................87 Model A .......................................................................................................................89 Model B ......................................................................................................................90 Model T .......................................................................................................................91 Aim 3 Predicting Aggression Category ........................................................................91 Normality ....................................................................................................................92 Model A .......................................................................................................................92 Model B ......................................................................................................................93 Model T .......................................................................................................................94 Receiver Operating Characteristics (ROC) Analysis of Violence Subtype ................95 5 DISCUSSION .......................................................................................................................110 Aim 1 Predicting Aggress io n vs. Non A ggression ......................................................110 Aim 2 Predicting Aggression Severity .........................................................................114 Aim 3 Predicting Aggression Category........................................................................115 Additional Findings ........................................................................................................116 Methodological Considerations ................................................................................117 Sample Size ..............................................................................................................119 Unrecorded Aggression ............................................................................................120 Stre ngths of The Study Design ..................................................................................122 6 SUMMARY AND CONCLUSIONS ...................................................................................125 APPENDIX A RATING SCALES ...............................................................................................................128 Modified Overt Aggression Scale Coding Scheme ..............................................................128 HCR 20 Risk Assessment Instrument Items and Scoring Guidelines ..................................129 Brief Psychiatric Rating Scale Instrument Items and Scoring Guidelines ..........................130
7 B SAMPLE CHARACTERISTICS .........................................................................................131 Descriptive Characteristics of Categorical Variables ..........................................................131 Descriptive Characteristics of Continuous Variables ...........................................................132 Correlations Between HCR and BPRS Instruments and Aggression Measures ..................133 Correlations Between Demogra phic Variables and Aggression Measures .........................134 C ANALYSIS OF UNIVARIATE OUTLIERS ......................................................................135 Cases With Extreme Scores on Primary Measures ..............................................................135 LIST OF REFERENCES .............................................................................................................136 BIOGRAPHICAL SKETCH .......................................................................................................153
8 LIST OF TABLES Table page 31 Summary of univariate analyses conducted. ......................................................................68 32 Summary of multivariate analyses conducted. ..................................................................69 41 Success rate s and strength of overall models in classifying cases .....................................97 42 Correlations, odds ratios, and strength of individual predictors in each model in correctly classifying violent cases ......................................................................................97 43 Su ccess rates and strength of overall models in classifying cases with outliers excluded .............................................................................................................................98 44 Correlations, odds ratios, and strength of individual predictors in each model in correctly classifying aggressive cases with o utliers excluded ...........................................98 45 Areas Under the Receiver Operating Characteristic Curve (AUCs) for each model in classifying Aggressive and non aggressive cases ..............................................................99 46 Areas Under the Receiver Op erating Characteristic Curve (AUCs) for each model in classifying aggressive and non aggressive cases with outlier excluded ............................99 47 Model B descriptive characteristics .................................................................................100 48 Summary of mu ltiple regression analysis for variables predicting aggression severity ..101 49 Correlations between predictors and aggression subtypes ...............................................102 410 Success rates and strength of overall models in distinguishing between verbally and physically aggressive cases ..............................................................................................102 411 Correlations, odds ratios, and strength of individual predictors in each model to distinguish between verbally and physically aggressive cases ........................................103 412 Success rates and strength of overall models in distinguishing between verbally and physically aggressive cases outliers excluded from the analysis ..................................103 413 Correlations, odds ra tios, and strength of individual predictors in each model to distinguish between verbally and physically aggressive cases outliers excluded from the analysis ..............................................................................................................104
9 LIST OF FIGURES Figure page 31 Receiver Operating Characteristics (ROC) Curves of Varying Predictive Value .............67 41 Classplot of correct and incorrect predictions of aggression under Model T ..................105 42 Rec eiver Operating Characteristics (ROC) curve for Model A (HCR total score only) and classification of aggressive cases.. ............................................................................105 43 Receiver Operating Characteristics (ROC) curve: Model A (HCR total score only) and classificat ion of non-aggressive cases. ..................................................................................106 44 Receiver Operating Characteristics (ROC) curve: Model B (HCR plus BPRS) and classification of aggressive Cases. .......................................................................................106 45 Receiver Operating Characterist ics (ROC) curve: Model B (HCR plus BPRS) and classification of non aggressive Cases .............................................................................107 46 Receiver Operating Characteristics (ROC) curve: Model T (age, history, activation, emotional withdrawal) and classification of aggressive cases. ................................................107 47 Receiver Operating Characteristics (ROC) curve: Model T (age, history, activation, emotional withdrawal) and classification of non-aggressive cases. .........................................108 48 Receiver Operating Characteristics (ROC) curve: Models A, B, and T, and classification of non-aggressive cases ..........................................................................................................108 49 Receiver Operating Characteristics (ROC) curve: Model T (BPRS tension i tem) and classification of physically aggressive cases. ........................................................................109 410 Receiver Operating Characteristics (ROC) curve: Model T (BPRS tension item) and classification of physically aggressive cases With Outliers Excluded. ....................................109
10 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy BUILDING A BETTER MODEL OF VIOLENCE PREDICTION: CLINICAL PREDI CTORS OF AGGRESSION AND VIOLENCE AMONG RESIDENTS OF A MAXIMUM SECURITY STATE FORENSIC PSYCHIATRIC FACILITY By Dominique D. Delalot August 2010 Chair: Eileen Fennell Major: Psychology Three models were compared in order to evaluate their predic tive utility in distinguishing aggressive from non aggressive cases, characterizing aggression severity, and classifying verbally aggressive versus physically aggressive subjects during their first 5 weeks of treatment in a sample of 102 mentally ill criminal offenders residing in a maximum security forensic treatment facility. A modified version of the Overt Aggression Scale (OAS) was used to record instances of aggressive behaviors retrospectively within its 4 subtypes: verbal aggression, physical aggression against objects, physical aggression against self, and physical aggression against others. A combined HCR 20 Historical and Clinical subscale score was used as a measure of violence risk. While weekly Brief Psychiatric Rating Scale (BPRS) ratings serve d as a measure of clinical symptom severity during the evaluation period. Model A represented a baseline function based on the use of the combined HCR 20 Historical and Clinical subscale scores. Model B represented a function based on the additive benefits of BPRS items on this first model. And finally, a test model (Model T) was developed based on conceptual formulations and associations observed in preliminary analyses. Overall correct classification of aggression was aided by selective inclusion of BPRS symptom variable scores, primarily based on this
11 measures ability to more accurately classify non aggressive rather than aggressive cases. Aggression severity was primarily associated with young age and symptoms of agitation, and to a lesser degree patien t history. A measure of subjective tension successfully distinguished between verbally and physically aggressive cases, but classification rates were poor across all models in predicting membership in these aggression subtypes. The benefits of correct clas sification of non aggressive criminal offenders are recognized in light of current policy trends in forensic treatment settings.
12 CHAPTER 1 INTRODUCTION Overview Violence risk is a priority issue in inpatient forensic psychiatric settings. The nature of t reatment and rehabilitation of mentally disordered patients who are also criminal offenders requires constant monitoring of their level of dangerousness and predictions of future violence risk by mental health professionals. At the same time, it is commonl y held that consistently accurate predictions of future instances of violence cannot be made by mental health professionals because the likelihood that a person will behave aggressively is a function of a variety of dispositional and situational factors in cluding history, temperament, and provocation, which cannot all be known in advance. However, it is possible to consider the available historical data, socio demographic variables, mental status features, and the anticipated placement and other situationa l factors to establish an estimate of relative risk. In clinical forensic settings, decisions based on evaluation of these variables are often considered advisory in nature. For instance, the ultimate decision to recommend release to a less restrictive se tting is based upon the entirety of information available to the treatment team. Assessment of future risk of violence in forensic populations relies on the identification of reliable predictors in forecasting future adjustment, but many of these, such as clinical symptom variables, can vary over time. It is also important to note that risk factors increase the probability of violence but they are not necessarily causative; they are often descriptors, not predictors. What follows is a study examining indiv idual patient characteristics including socio demographic variables, and historical and clinical data in order to identify risk factors for predicting violence and aggression in an inpatient psychiatric forensic setting. Recognition and selection of these predictive factors was guided by a thorough review of the literature and a
13 determination of their particular relevance to individuals in these settings. This review also helped in identifying and disentangling the relationships between relevant variables t hat were found. To this end, this investigation began with a review of issues germane to violence prediction, including the following: relevance of risk assessment in various clinical settings; factors that have been found to be associated with aggression and violence in a variety of contexts and within various populations including inpatient and forensic psychiatric settings; the debate over and strengths and weakness of clinical and actuarial approaches to violence prediction; and an appraisal of common r isk assessment instruments. Finally, a method for identifying factors related to aggression and violence committed by current forensic psychiatric inpatients is put forward. Violence and aggression engender numerous emotions among mental health professiona ls for a number of reasons, ranging from fears of personal safety to the legal and ethical implications of ones patients injuring other individuals. Additionally, public concern about violence by the mentally ill has accompanied the dissemination of repor ts of such violence through the popular media and has generated increased examination of the legal and clinical processes of violence management. Mental health professionals are increasingly asked to assess the risk of violence in many psychiatric populati ons. Clinicians attempting to avoid these issues by steering clear of treatment and assessment situations involving violent clients are likely to find this an unrealistic strategy, as violent patients can be found in any treatment setting, including private practice, and various medical and psychiatric inpatient units. A more sensible approach is offered by the growing body of literature on the causes of violence among psychiatric patients, and the management of these patients, intended to inform practitio ners and equip them with the tools necessary to deal with violence in various treatment
14 and assessment situations. Increased interest has generated a growth in the number of relevant publications on these topics as well. This literature identifies the comp lex interface between individual characteristics and environmental influences in the assessment of violence risk and behavior. Although this paper focuses less on placing violence among psychiatric patients within the larger perspective of societal violenc e, knowledge of the interaction between the causes of violence and these larger contextual elements is useful in educating practitioners and informing the planning for and treatment of individual patients. The management of aggressive and violent patients involves a myriad of legal issues. These stem primarily from balancing the individual rights and liberties of the patient with the need to protect others from being harmed, and from the patient harming him or herself. These concerns arise in such areas as informed consent, forced medication adherence through emergency treatment order (ETOs), involuntary psychiatric hospitalization, and restraint and seclusion. These rights and responsibilities are informed by guiding ethical principles as well as specific c ase law, but vary by treatment context and jurisdictions (APA, 2002; Florida Statutes, 2008).
15 CHAPTER 2 REVIEW OF THE LITERA TURE Historical Overview of Violence Prediction The Development of Violence Prediction The evaluation of a patients violence potent ial by mental health professionals is a controversial one. Some maintain that the prediction of violent behavior by experts, who are presumed to possess greater expertise in such matters, is no more accurate than that of intelligent laypersons and that in some cases the accuracy of predictions of dangerousness by laypersons has been found to be greater than that of trained clinicians (Menzies, Webster, McMain, Staley, & Scaglione, 1994) But Otto (2 000; 2008) maintains that such conclusions are based on faulty interpretation of the research data. Another interpretation of the data suggests that criticisms of the accuracy of psychologists and psychiatrists in particular are based not on their worse t han chance performance in evaluating the risk of future violence (since it has been shown that this is not the case either), but that they are more likely to be correct when predicting that a client will not be violent, than when predicting that a client w ill be violent. According to Otto (2008) the probl em has instead been seen as a problem of over prediction, and that expert evaluators have essentially been basing their judgments on information in ways similar to laypersons. Criticisms of these earlier studies point out that they typically employed follo w up periods that were very short, which artificially inflated accuracy of prediction. Also, violence was measured solely based on arrests and convictions, which sorely underestimates incidents of violence (which typically go underreported as they seldom c all upon police involvement). It is also suggested that individuals released into the community were generally older. All of these variables have one common result: low base rates of subsequent violent behavior.
16 According to Rice, Harris, and Quinsey (2002) base rates of violence are significantly greater when violence is operationally defined to include self reported acts and violence reported by collaterals (other involved parties and significant others), when follow up periods are longer than those traditionally examined, and when individuals are younger when released. In general, the field of violence risk assessmen t has made notable progress since these early studies, and a number of factors have been found to be associated with violence risk. Assessing Current Dangerousness Before assessing future violence potential clinicians must determine an individuals present state of dangerousness and, therefore, particular attention be must be paid to their current psychological and emotional state. Menzies, Webster, and Sepejak (1985) identified various affective states thought to be associated with an increased risk of violence, and found that anger and lack of empathy were highly predictive of violent behavior. Similarly, Monahan et al. (2001) and Monahan et al. (2005) found that individuals followed in the influential MacArthur Study of Mental Disorder and Violence who rated higher on scores of anger were twice as likely as those with lower scores to act violently. This violence risk assess ment study was funded by the National Institute of Mental Health and employed public access data from the MacArthur Foundation to develop violence risk assessment software, and has been validated on numerous independent samples of patients. Findings by Ber g, Bell, and Tupin (2000) and Bell (2000) also suggest that mental health professionals should be watchful for precursors of imminent violence, such as changes in an individuals symptoms and other physical signs. They identify several such signals of impending violence, including erratic eye movements, clenched jaw or hands, unusual vocalizations (chanting, grunting), flared nostrils, non compliance with directives from psychiatric staff or family, and violation of a clinicians personal space.
17 Although these behavioral signals do not constitu te out and out dangers in and of themselves, in each case they can lead to actual threats that need to be taken seriously by clinicians and appropriately documented and explicated. Assessment of the degree of danger of a threat is aided by a thorough evalu ation of how an individual might carry out the threat, their understanding of the possible consequences of their actions, and an assessment of an individuals perceptions of alternatives to these behaviors. Findings from the MacArthur Foundation studies (Monahan and Steadman 1994; Bonnie and Monahan 1997; Silver, Mulvey, & Monahan, 1999) indicate that considera tion of these alternative behaviors and perceived coping strategies by a potential offender is fundamental. For instance, a psychiatric patient may want to avenge himself against an attacker for a misperceived slight or transgression against them but may instead choose to turn and walk away when confronted by the possibility of incarceration should they physically attack the individual. Brizer and Crowner (1989) and Convit, Jaeger, Lin, Meisner, and Volavka (1988) suggest that additional information which should be gathered in violence risk assessm ent includes an individuals fantasies about violence, whether they keep a list of possible victims and even, to some degree, an appraisal of the intended targets. All of this information ultimately serves to inform on the relative likelihood of future vio lent behavior towards others. In murdersuicides the taking of a perpetrators own life has been conceptualized as an extension of their violent impulses turned inward when, for instance, they feel that they can no longer live without their victim. Researc h suggests that evaluations of violence threat requires an assessment of suicide risk. Brizer and Crowner (1989) and Convit et al. (1988) found that a history of violent suicide attempt greatly increased the probability of future violence towards others. This relationship is even more compelling w hen a homicidal threat has been made.
18 Asnis, Kaplan, Hundorfean, and Saeed (1997) found that 91% of psychiatric outpatient attempted murderers had also attempted suicide sometime in their lives. They also found that 86% of psychiatric patients reporting homicidal ideation also reported concurrent suicidal ideation. Research ex amining variables thought to be contributory to violence has led to a distinction in the literature between static and dynamic risk factors As the name implies, static factors are thought to be those factors that are unchangeable and therefore not amena ble to intervention, such as an individuals personal history and demographic group membership. By contrast, dynamic factors are those typically targeted for change. These can include psychiatric symptoms, substance use habits, access to weapons, and devia nt peer associations. This distinction allows for a systematic strategy for organizing and weighing risk factors, and ultimately can serve to inform clinicians in implementing prevention strategies for those at risk for future violence, as well as treatmen t and management of those who have already offended. The Problem of Defining Violence While patient pathology remains a central theme of violence prediction, the causes of aggression and violence are multifactorial. This is likely one of the reasons for t he diversity of definitions, models, and terminology associated with studies of aggression and violence in psychiatric populations. For instance, Swanson, Bland, and Newman (1994) define violence as a physical encounter with someone, other than a spouse, wh ich involves the use of a weapon. While Chen, Hwu, and Williams (2005) treat aggression as incorporating both verbal and physical assaults. Ryan and Maguire (2006) and Maguire and Ryan (2007) suggest that the diversity and inconsistencies in definitions, and therefore in the techniques employed to measure these constructs, represent major obstacles to the meaningful measurement and evaluation of
19 aggression and violence in psychiatric populations. In the present study, for instance, aggression and violence are used synonymously. Research on Prediction General Issues Clinicians are frequently called upon to evaluate whether an in dividual who has acted violently in the past is likely to re offend. Such evaluations are particularly common among psychiatric patients who become involved in criminal or other legal processes where admission, discharge, sentencing, probation, and any pla nning for eventual placement depends on an appraisal of the likelihood for future violent behavior. Beyond the context specific interest of clinicians as to whether an individual who has acted violently in the past is likely to exhibit future antisocial be havior, there is also a concern for public safety which drives public policy, defines the context in which medico legal decisions are made, and ultimately helps identify those individual characteristics on which clinical judgments about future violence are made. Violence prediction is predicated on the assumption that individuals who have offended in the past possess particular attributes that are associated with a greater likelihood to offend again in the future. As Martin (2003) indicates, if environmental factors and measurement error accounted for all of the variance in violent behavior, and none was accounted for by individual characteristics, then clinical judgment would co ntribute considerably less in decisions about disposition and placement of psychiatric patients. The literature, however, is replete with research on the many predictors of violent behavior among individuals who have previously offended, and supports the n otion that underlying individual characteristics are examples of such predictors. It should be noted that research examining predictors of aggression has varied greatly in terms of the populations and contexts in which they have been studied. While the fo cus of the present study is identifying predictors of aggression among adult male psychiatric residents of a state run maximum security forensic treatment facility, others have focused on juvenile offenders
20 (Farrington, Loeber, & Kalb, 2001) non criminal non institutionalized psychiatric patients (DeMatteo, Heilbrun, & Marczyk., 2006) women (Putkonen, Komulainen, Virkkunen, Eronen, & Lnnqvist, 2003) and numerous other groups and settings. It should also be noted that the majority of studies have been conducted in North America and Europe. While some have argued that the variations in demographic characteristics from study to study necessarily limits the confidence with which statements regarding the likelihood of future violence can be made, the literature overwhelmingly suggests that the personality characteristics predictive of future violence are shared across psychiatric and offender populations (Quinsey, Jones, Book, & Barr, 2006) Another source of potential variability in study results arises from the ways in which future risk and violent behavior (particularly re offending and recidivism) are defined. For instance, investigations of recidivism among parolees often include behaviors considered to be violatio ns which would not be considered antisocial or criminal were not the individuals on parole, such as failing to inform a parole officer of a change of residence or failing to meet other terms of conditional release. This definitional problem is also encountered in studies of forensic patients who may be counted as recidivist when re hospitalized or admitted to civil treatment facilities for non criminal behaviors. As Quinsey et al. (2006) aptly point out, the majority of crimes (such as theft, prostitution, and drug possession) are not related to violence or safety issues. Most research into violent offending among forensic psychiatric patients has involved follow up studies of individuals discharged from various institu tional settings in order to identify those characteristics which are related to later violence. Measuring these characteristics, and identifying a statistically significant relationship between distinguishing attributes and future acting out is typically a ccomplished through examination of hospital and police records, and
21 other available official documentation. While such studies focus on the risk of violent behavior among those who have already offended, these investigations are ultimately driven by an int erest in identifying those predictors of violent behavior among those who have never been apprehended, thereby identifying the precursors of violent behavior. The present study confines itself to the evaluation and prediction of violence defined by a wide spectrum of behaviors, among psychiatrically hospitalized residents of a maximum security forensic treatment facility, in relation to demographic, historical, clinical, and hospitalization characteristics. Older Studies of Violence Risk Relevant Variable s Despite the prevalent need for violence risk evaluations, the ability of mental health professionals to accurately assess the risk of aggression is inadequate. Studies on risk assessment have generally failed to provide solutions that lend themselves to the precise and practical prediction of violence risk (Amore et al., 2008) Swans on, Holzer, Ganju, and Jono (1990) and McNiel, Binder, and Greenfield (1988) long ago demonstrated that individual socio demographic variables such as male gender, low socio economic status (SES), young age, and even single marital status are positively associated with increased violence risk in community samples. Their data suggested that these variables are more predictive of violence risk than clinical variables such as psychiatric diagnosis and symptom profile. Results of other studies, by contrast, provide strong evidence that clinical variables are more predictive than historic and demographic factors in predicting risk, including personality disorder, alcohol abuse, seizure disorder, organic psychosis, schizophrenia, and mania (Taylor, 1985; Link & Stueve, 1994; Junginger & McGuire, 2004) Amore et al. (2008) argue that the emphasis on specific ps ychiatric diagnoses in violence prediction has been over emphasized in the literature and that it is not the specific diagnosis, per se, that is a risk factor but rather symptom characteristics and profiles of certain disorders, such as hallucinations,
22 del usions, and thought disorder. They utilized the Brief Psychiatric Rating Scale (BPRS) to evaluate 24 symptom constructs (each rated along a 7 point scale of severity ranging from not present to extremely severe) that best describes a patient's present condition, in order to identify the best predictors of violence in an acute care psychiatric population; their results were mixed. Studies by Monahan and Steadman (1983) echoed a plethora of research which showed that a history of violence is the most sign ificant predictor of future violence. Another risk factor which has been identified as predictive of future violence is a history of incarceration (McNiel, Binder, & Greenfield, 1988) But there is less consensus regarding the factors that distinguish violence risk in populations in which everyone presumptively shares such histories. Therefore, what makes someone more likely to aggress, and what form is this subsequent aggression likely to take? The present study seeks to address these questions, and has as its main objectives to 1) identify the de mographic and clinical factors associated with aggressive acts among those admitted to a maximum security forensic treatment facility in the 5 weeks following admission; 2) examine how these relate to the intensity/severity of aggressive acts; and 3) ident ify and distinguish the types of aggression and their association to these variables. It is believed that a model will immerge that will predict membership of residents into groups of violent offenders by the types of acts they are likely to commit (and to clarify whether differences represent distinct groups or continuums of violence). Ultimately, it is hoped that the predictive utility of clinical variables, rather than historical variables alone, will be demonstrated in assessment of violence risk. Newer Studies of Violence Risk Relevant Variables Monahan (1997) aptly points out that the investigation of violence risk in actual clinical practice has suffered from problems of replication and generalizability. This is largely due to the
23 multitudes of variables proposed as predictive of future violence. In fact, Menzies and Webster (1995) and Menzies et al. (1994) point out that with rare exception, practically every variable investigated has been found, to some degree, to have an association with future violence. Duxbury (2002) and Scott and Resnick (2006) and Resnick and Scott (1997) argue that this does little to aid clinicians in accurately predicting which patients will display aggressive behavior. And according to Webster, Douglas Eaves, and Hart (1997) and Secker et al. (2004) much of the literature has failed to demonstrate a firm linear relationship between these identified risk factors and actual episodes of violence. Therefore, they stress that attentiveness to the interaction amongst these factors is critically important if they are to be unraveled. According to Watts, Leese, Thomas, Atakan, and Wykes (2003) those violenc e risk factors which have been identified in epidemiological studies of are of little added benefit to the clinical judgments of mental health professionals because they have shown that the risk factors for violence among those with mental illness are the same as for those in the non psychiatric population (including gender, age, employment status, and criminal history). Watts et al. (2003) also stress the importance of considering the base rate (prevalence) of violence in any population und er consideration when evaluating their likelihood of future violence and the accuracy of risk assessment approaches. This often problematic issue can be addressed using Receiver Operating Characteristics (ROC) (Mossman, 1994; Rice & Harris, 1995) which provide indexes of effectiveness derived from measures of sensitivity and specificity and are analyzed similarly to effect sizes (Cohen, 1988; Cohen, 1992; Hasselblad & Hedges, 1995) While studies of violence risk proliferate, there is a relative paucity of r esearch into risk assessment in inpatient and residential settings, particularly during the more acute period of care. And there is an even greater scarcity of work on the clinical prediction of assaultive behavior in
24 forensic settings. The work of Hoptman Yates, Wack, and Convit (1999) stands as an exception. Their work compared patient characteristics generally associated with violence in forensic psychiatric hospitals with actual assaultive behavior in order to evaluate the acc uracy of clinical decisions based on these patient variables. They found that clinicians in their sample showed a high rate of accuracy of 71%, with a diagnostic sensitivity of 54% and a diagnostic specificity of 79%. Owen, Tarantello, Jones, and Tennant, (1998) described severa l surveys of inpatient violence, and several studies by McNiel and Binder (1994; 1995) and McNiel, Sandberg, and Binder (1998) in which mental health professionals identified those individual attribu tes they felt were relevant to violence prediction, and also explored the accuracy of their clinical judgments based on these characteristics. Another significant problem intrinsic to studies of violence risk prediction is their lack of amenability to strictly experimental designs. Such studies need to take place in real world clinical settings, in which mental health professionals and other treatment staff are obligated to act when an individual is either acting aggressively or has been deemed to be at sig nificant risk of committing violence. In such settings, the prediction of risk is affected by the act of responding to that risk and of intervening in an aggressive act. In particular, Davison (1997) notes that in such settings a variety of remedies typically follow to prevent the anticipated outcome, and that this has the effect of increasing false positive rates. Therefore, while the present study identified instances of specific aggressive behaviors for each participant there was also a careful accounting of targeted clinical interventions given in response to these behaviors over the course of the study period. In this way the value of the predictive variables could be assessed both with and without the influe nce of clinical interventions.
25 Risk Factors for Violence History of Violence Traditional wisdom holds that the best single predictor of future violence is a past history of violence (Klassen and O'Connor 1988) Studies based on the MacArthur Foundation data ( Monahan, 2001) found that all measures of previous violence were strongly associated with future violence (these included data from such sources as hospital and arrest records, and self report). Scott and Resnick (2006) suggest that obtaining information on the most violent act an individual has ever co mmitted can be helpful in predicting the likelihood and severity of subsequent violent behavior. Certainly, as with any comprehensive clinical interview, obtaining a detailed violence history is critical to the evaluation of risk. Such an evaluation should include information on the type of violent act, why the violent act occurred (from the perspective of all relevant parties), who was involved, the degree of injury, and the contribution of other relevant variables (such as intoxicating substances, frequen cy of the type of violent act, etc.). Evaluation of an individuals history of violence is aided by official court and criminal records, which typically contain a plethora of information not only on illegal behavior but useful demographic data. One such pi ece of useful information is an offenders age at first arrest, which in the case of serious offending has been found to be a variable which correlates highly with recidivist criminal offending (Borum, Swartz, & Swanson, 1996) According to Wolfgang (1987) following four arrests the probability of a fifth is 80%. Military and work records also contain a wealth of relevant information useful to evaluating an individuals violence potential. Catalano, Dooley, Novaco, Wilson, and Hough (1993) suggest that clinicians conducting risk assessments for individuals who have served in the military, should pay particularly close attention to any history of fights, unexcused absences, disciplinary actions, type of discharge, and reas ons for reassignments. While work history
26 evaluations should also include a review of the incidence of job changes and any history of terminations from employment. Catalano et al. (1993) also reported that individuals who are frequently fired from their job are at increased risk for future violence. In fact, they found a 6 fold increase in the likelihood of future violence among those laid off from work. Act s of violence fall along a continuum of severity often based on the lethality of the instrument of violence. This bears not only on the outcome of the violent act but upon the determination of future risk potential and re offending. Studies by Zimring (1993) reveal that an assault with a gun was 5 times more likely to result in a fatality than an assault with a knife. Those who have used weapons in the past pose a serious risk of future violence, and the severity of this risk has been found to vary by the l ethality of the weapon used. Therefore, a thorough risk evaluation must include an appraisal of an individuals access to weapons. According to a 2001 Department of Justice report it is estimated that approximately 40% of American homes contain a firearm a nd 20% of these maintain the gun in an unlocked and loaded condition (Webster, Vernick, Ludwig, & Lester, 1997) Therefore, asking an indi vidual whether they have access to a weapon is critical. Quanbeck et al. (2004) also point out that tracking the movement of a weapon, for instance by asking an individual whether they recently moved it from a locked case in the attic to the nightstand by their bed, can provide useful information as well. Reports of such behavior could clue a clinician into an exacerbation of paranoid symptoms (for instance, of someone suffering from a psychotic disorder or other mental illness) that would make them more likely to aggress against someone they misperceive as a threat. Socio D emographic Factors The assessment of risk of dangerousness and fu ture violence by mental health clinicians must also be based on a review of demographic risk factors that have been found to increase the probability of future violence. Some of these include younger age (Monahan 1981; Swanson &
27 Holzer, 1990) male gender (Tardiff & Sweillam, 1980; Tardiff, 1996) low economic status (Swanson et al., 1997) lower intelligence (Quinsey & McGuire, 1986; Borum et al., 1996) criminal history (Monahan, 1981; Convit et al., 1988) childhood abuse (Widome, 1989) childhood environment (Convit et al., 1988) and substance abuse (Steadman et al., 1998) It should be noted that while race has been explored as a violence risk factor it has been associated with errors in violence prediction (Hoptman et al., 1999) Violent behavior is associated with membership in a younger age group. Using data from the Epidemiologic Catchment Area study Swanson et al. (1997) found a rate of violent behavior of 7.34% among those between 18 and 29, 3.59% for those between 30 and 44, 1.22% for those between 45 and 64, and a rate of less than 1% for those 65 or older. According to the Federal Bureau of Investigation (2004) males commit violent acts at a rate 10 times greater than females. However, according to Lidz, Mulvey, and Gardner (1993) and Newhill, Mulvey, and Lidz (1995) among individuals with mental illness men and women do not differ significantly in their base rates of violent behavior. They found that, in fact, among the mentally ill base rates of violent behavior in some cases are slight ly higher for women than men. The seminal MacArthur Foundations Violent Risk Assessment Study (Monahan, 2001) followed male and female psychiatric inpatients between 18 and 40 years of age who were subsequently released into the community. These individuals, all of whom continued to suffer from mental disorders, were monitored for 1 year for acts of violence towards others. Results indicate that men were only slightly more likely than women to be violent. In fac t, it was found that women were more likely to aggress towards members of their family and those in their immediate home environment. At the same time, violence perpetuated by men was more likely to lead to arrest and medical intervention.
28 Krakowski and C zobor (2004) examined the relationship between gender and violence in mentally ill individuals within a psychiatric institutional setting. Among their 155 male and 67 female psychiatric inpatients they found little difference in the rates of physical assault between men and women in this setting. However, women had a higher rate of physical assaults within the first 10 days of the study, while assaults by men we re more likely to result in injury. Quanbeck et al. (2004) and Maden, Scott, Burnett, R., Lewis, and Skapinakis (2004) argue that risk evaluation should also inclu de consideration of an individuals economic status. Borum, Swartz, and Swanson (1996) found a rate of violence 3 time greater amongst individuals defined as belonging to lower income brackets. By contrast, another study by Silver, Mulvey and Monahan (1999) su ggests that neighborhood poverty was more predictive of violent behavior than an individuals economic status. Violence risk has been found to increase with lower intelligence, cognitive disabilities, and mental retardation (Quinsey & McGuire, 1986; Borum et al., 1996) In a study by Hodgins (1992) examining the relationship between intellectual functioning and violence risk, men with intellectual deficits were 5 times more likely to commit violent offense, while similar women were 25 times more likely. Many studies have also identified thos e with less education as having higher rates of violent behavior (Link, Andrews, & Cullen, 1992; Bor um et al., 1996) Environmental Factors Among the mentally ill certain environmental variables have been found to consistently contribute to future aggression across treatment settings. HaggardGrann, Hallqvist, Lngstrm, and Mller (2006) found that i nterpersonal friction among patients, among staff, and between patients and staff in hospital treatment wards contributes to the occurrence of violence in these settings. They also found that being denied treatment can have similar negative consequences. K err and Roth (1986; 1987) linked aggressive behavior in psychiatric units to patie nt perceptions
29 of slow recovery from illness. Lanza (1983) and Kindy, Peterson, and Parkhurst (2005) identified a relationship between inadequate hospital staffing and the incidence of violence. Finnema, Dassen, and Halfens (1994) found that lack of freedom and privacy can contribute to aggression in a psychiatric ward. Institutional disorganization (Duxbury, 2002) and limited options for therapeutic activities (Finnema et al., 1994; Kindy et al., 2005) have also been cited as contributing to institutional violence. Perry, Pescosolido, Martin, McLeod, and Jensen (2007) and Martin (2003) aptly point out the relevance of these and other environmental variables not only on violence prediction but on the overall clinical management of psychiatric patients by reminding us that treatme nt environments are typically completely different in every respect to what patients are accustomed to outside of these settings. Substance Abuse A number of studies have reported that the majority of violent offenders are under the influence of alcohol or drugs at the time of their violent crimes and that in general there is a very strong association between violent behavior and substance abuse (Murdoch, Pihl, & Ross, 1990; Tardiff, 1999; Monahan, 2001) Roth (1994) found that i n half of all violent crimes, the offense was preceded by alcohol consumption by either the perpetrator, the victim, or both. Goldstein, Bellucci, Spunt, and Miller (1991) examined the relationship between violenc e and stimulants, including powder and crack cocaine, amphetamines, and PCP (angel dust). They reported that these drugs appear to be of particular concern to violence risk as they typically result in disinhibition, instability, grandiosity, and paranoia which are all consistent with psychotic spectrum symptoms. They also found that the violence associated with stimulant ingestion varies by gender, with men being more likely to engage in violent crimes while under the influence and women more likely to b e the victims of violence when similarly impaired.
30 Within a psychiatric population the potentiating effects of substance abuse on mental illness are of particular concern. Data from the MacArthur Foundation Studies illustrate that co occurring diagnoses o f mental illness and substance abuse strongly predicts future violence. In their study examining the additive effects of mental illness and substance abuse on violent behavior Steadman et al. (1998) compared rates of violence of community based psychiatri c patients and non patients and found that the presence of substance abuse increased the rate of violence among non patients 3 times, while it increased the rate of violence 5 fold in their outpatient psychiatric sample. Mental Illness One of the stigmas o f mental illness is the depiction of the mentally ill as dangerous and unpredictable. At the very least these perceptions have contributed to the proliferation of studies examining whether individuals with psychiatric disorders are indeed more violent than those in the general population. The sum of the research has painted a mixed picture (Swanson et al., 1990; Link et al., 1992; Torrey, 1994; Steadman et al., 1998) Monahan (1997) examined rates of violence among psychiatric patients committed to civil setting s who were subsequently discharged to the community. While he found a very weak relationship between mental illness and violence, the majority of psychiatric patients were not violent. He also noted that the rate of violence among the sample of mentally ill patients was greater only when their psychiatric symptoms flared up, such as during acute, transient exacerbation of psychotic symptoms. He also found that a diagnosis of schizophrenia, despite the common perception of an increased association with dange rous acting out and criminality, was actually associated with lower rates of violence than diagnoses of bipolar disorder or even major depression. As previously mentioned, drug abuse is strongly associated with violence risk. According to Monahan, Steadman et al. (2005) it is also a much stronger predictor of violence than mental illness. In their examination of
31 violence risk factors among the mentally ill Steadman et al. (1998) found no significant difference in the rates of violence between non substance abusing patients discharged from psychiatric facilities and non substance abusing individual s in the general population who were not suffering from mental illness. In other words, the presence or absence of mental illness was not the determining factor when considering the prevalence of violence. Psychosis The assessment of violence potential req uires the careful review of specific psychiatric symptoms. Humphreys, Johnstone, MacMillan, and Taylor (1992) argued that the presence of a diagnosis of psychosis should be of parti cular relevance when assessing violence risk potential. They examined 253 psychiatric patients undergoing their first psychotic episode and found that 52 had demonstrated behaviors prior to hospitalization that were threatening towards the lives of others. Of these, 36 exhibited dangerous behaviors towards others up to a year prior to hospitalization, while the remaining 16 displayed such behaviors for more than a year prior to their first schizophrenic episode. Krakowski, Volavka, and Brizer (1986) argue that because they have increased access to weapons people with symptoms of paranoia present as an increased risk of dangerousness. Scott and Resnick (2006) suggest that violence among paranoid psychotics is typically deliberate and premeditated and compatible wi th their delusional beliefs. As such, an explicit target is typically identified and the violence directed towards this misperceived tormenter (often a member of the immediate family or friend of the aggressor). While the research is far from unequivocal, Wesley (1993) examined the contribution of specific delusions in violence prediction and found that persecutory delusions in particular (those associated with a perceived threat and being controlled by others) were strongly associated with an increased ris k of violent behavior. Link et al. (1992) and Li nk and Stueve (1995) identified
32 additional types of delusional beliefs that tend to be associ ated with violence among mentally ill patients, including various beliefs regarding the permeability of the mind including thought insertion ( the delusion that thoughts are being inserted into one's mind by someone else), thought interference, that others wish to do them harm (persecutory delusions), and that they are being followed (persecutory and referential delusions). Link and Stueve (1995) found that these delusional beliefs in particular are associated with increased rates of violence among persons w ith psychotic spectrum disorders. Similarly, Swanson and Holzer (1990) Swanson and Holzer (1991) and Swanson et al. (1997) used data from the Epidemiologic Catchment Area surveys to identify risk o f violence among individuals reporting delusions related to perceived external control of their thoughts and external threat. They reported a rate of assault twice as high among psychotic patients experiencing these specific delusions compared to psychotic patients with other symptoms. Similarly, a recent study by Stompe, Ortwein Swoboda, and Schanda (2004) found that delusions related to perceived threat or thought control were not associ ated with an increased risk of violence, or the severity of violence, when comparing male criminal offenders with a diagnosis of schizophrenia who were found Not Guilty by Reason of Insanity (NGI) to non offending schizophrenics. They also reported no diff erence in the rates of specific psychotic symptoms between each group. Several recent, large scale, and prominent studies seem to contradict these findings however. Studies based on data from the MacArthur Study of Mental Disorder and Violence (Appelbaum, Robbins, & Monahan,. 2000; Monahan et al. 2001) reported that the presence of delusions among discharged psychiatric patients, regardless of the particular type of delusions, was not predictive of rates of violent behavior, nor was it associated with higher rates of violence in particular. Howeve r, Monahan and his colleagues draw a distinction between delusional and
33 non delusional suspiciousness, where the former is characterized by more disorganized and bizarre content, and the latter is based on misperception of hostile intent due to misinterpre tation of the behavior of others. They found that non delusional suspiciousness was indeed related to later violence. The literature presents a mixed picture of the importance of various clinical features in determining the dangerousness of mentally ill individuals. In addition to the form and content of various delusional subtypes, the affective latency of these underlying thought disorders has been examined. For instance, Appelbaum, Robbins, and Roth (1999) examined how certain delusional beliefs make psychiatric patients feel, and whether these self reported feelings are later associated with an increased risk of aggression Results indicate that when patients experience delusions that make them unhappy, angry, scared, or anxious, they are more likely to engage in violent behavior. It should be noted, however, that the evaluation of violence risk potential should include que stions regarding an individuals tendency to act on their delusions even when these acts are not violent. Research by Monahan (2001) also reveals that an individuals tendency to act on their del usional beliefs (non violently) regardless of the content of that delusion is also associated with a greater tendency towards violence. An evaluation of the form and content of hallucinatory behaviors would also seem to be instrumental to thorough violence risk assessment. However, research into whether the presence of hallucinations increases an individuals likelihood to commit violent acts is also mixed. The presence of hallucinations in and of itself does not appear to be related to an increased rate of violence (Zisook, Byrd, Kuck, & Jeste 1995) But just as with delusions the presence of hallucinations has been associated with greater violent acting out when they generate negative emotions in schizophrenic patients (fear, sadness, anger, anxiety). According to Cheung,
34 Schweit zer, Crowley, and Tuckwell (1997) this relationship is even more pronounced when the patient has not developed appropriate strategies for coping with their hallucinatio ns and for managing these negative emotions. Rudnick (1999) conducted a review of 7 controlled studies on the association between auditory command hallucinations and violence; none succeeded in demonstrating a positive relationship, and one managed to demonstrate an inverse relationship. In a similar study Scott and Resnick (2006) found a positive relationship between command hallucinat ions and violence when the form of the hallucination was of a person familiar to the patient and the commanded act was of a less serious nature. In stark contrast, a study by McNiel, Eisner, and Binder (2000) found that patients were twice as likely to be violent when experiencing command hallucinations to harm others. Their study examined 103 psychiatric patients within a civil facility; 33% of which rep orted command hallucinations over the previous year, while 22% indicated complying with these commands. Psychiatric symptoms rarely occur in isolation, even within a specific diagnosis. Command hallucinations are often accompanied by a delusional thought structure whose content is consistent with the hallucination (Junginger, 1990) While data from the MacArthur Foundation St udies failed to reveal a relationship between non violent command hallucinations and subsequent violence, or between general hallucinations and violence, other data support a relationship between command hallucinations to commit violence and subsequent vio lent acts. People with schizophrenia commonly lack insight about their illness or the consequences of their behavior. According to Buckley et al. (2004) when compared to a non violent control group, violent schizophrenics tended to display a particular lack of awareness regardin g their mental state and the legal ramifications of their behaviors. Therefore, they argue, clinicians
35 should evaluate not only the positive and negative symptoms of psychotic disorder but an individuals insight into their illness and legal situation. Wa lsh, Buchanan, and Fahy (2002) report that while the majority of individuals with schizophrenia do not engage in violent behavior, there is growing evidence of an association between the diagnosis of schizophrenia and criminal acts. Wallace, Mullen, and Burgess (2004) retrospectively reviewed 2861 patients with schizophrenia who had been followed over a period of 25 years, and found that patients with a diagnosis of schizophrenia had a greater number of criminal convict ions (both violent and non violent) compared to a matched comparison group. However, they also indicated that these results could be accounted for by concurrent treatment interventions, substance abuse, or active symptoms. Mood Disorders There is evidence of an association between depression and violent behavior as well. Numerous studies indicate that depression is the most frequent diagnosis in murder suicides (Coid, 1983; Marzuk, Tardiff, & Hirsch, 1992; Scott, Harrington, House, and Ferrier, 1996) and there is ample evidence suggesting that individuals may commit violent acts on others when anguished, hopeless, and despaired. An early investigation of the association between depressive sy mptoms and specific crimes by Resnick (1969) identif ied a pattern of murder suicide in which depressed mothers kill their children and then take their own life. Today this has generally been reframed as postpartum depression, where hormonal changes in a woman's body after childbirth are thought to trigger s ymptoms of depression. In extreme cases, this explanation has been used as a defense in cases involving women claiming that postpartum depression moved them to behave erratically or even kill their children. Rosenbaum (1990) found that murder suicide cases in couples are often associated with depressive symptoms related to feelings of jealousy and loss.
36 He argues that depression in these cases is the driving force for the externalized violence, in that the homici dal act is simply an extension of the suicide. Other mood conditions have been found to be associated with increased violence. Krakowski, Volavka, and Brizer (1986) found that while serious violence is rare among patients with s ymptoms of mania, they show a greater rate of assaultive and threatening behaviors. Nonetheless, they also found that individuals with manic symptoms engage in fewer criminal acts of all kinds compared to individuals diagnosed with schizophrenia. According to Tardiff and Sweillam (1980) individuals with mania are more likely to be violent or act out aggressively when restrained or when their behavior is limited in some way. More recent studies on the association between mania and violence, however, suggest a substantial connection between manic symptoms in psychiatric patients and criminal behavior. Quanbeck et al. (2004) examined 66 criminally incarcerated inmates diagnosed with bipolar disorder, and found that 77% displayed manic symptoms and 59% were psychotic at the time of arrest. They also foun d that these manic inmates were later discharged from inpatient community treatment at a much higher rate, and that they were seldom involved in follow up outpatient treatment after discharge, effectively increasing their likelihood of re offending. Person ality Factors In addition to its association with a number of psychiatric disorders, violent behavior has been found to be more prevalent among individuals with certain types of personality disorders, as defined by the DSM IV (1994). Tardiff (1999) and Tardiff and Sweillam (1980) found that borderline personality disorder is associate d with a greater risk of violence, while Meloy and Gacono (1992) identified sadistic traits as more likely to lead to greater violence. Overwhelmingly, however, antisocial personality disorder has been found to be the most common type associated with violence (Monahan, 2001) Violence committed by those with
37 antisocial personality disorder has been found to be characterized by a number of features, including a motivation of revenge and co occurring alcohol abuse (Scott, Harrington et al., 1996) and that the acts are devoid of emotionality or any pretense of moral justification (Williamson, Hare, & Wong, 1987) Heilb run (1990) found that the combination of low IQ and anti social personality disorder significantly increases an individuals risk of future violence. Joyal, Putkonen, Paavola, and Tiihonen (2004) emphasize the importance of assessing antisocial personality traits among the mentally ill in order to adequately ascertain violence risk potential. They found that individuals diagnosed with both schizophrenia and antisocial personality disorder are actually less likely to engage in violent acts that are provoked by their psychotic symptoms, compared with schizophrenic patients without antisocial personality disorder. They indicate that the presence of violence in individuals with these comorbid diagnoses is more likely to be associated with other factors, including excessive alcoho l use, a prior squabble with the victim before the violent act, or a victim who is not a family member. Differentiating the individual contributions of subclinical features on violence risk is a constant challenge. While certain personality disorders have been found to be associated with increased dangerousness, many individual characteristics constitute these diagnoses. The presence of some of these traits alone may not meet criteria for a diagnosis of personality disorder but may be extremely predictive o f future violence. For instance, Borum et al. (1999) identified impulsivity ostentation, a sense of entitlement, low frustration tolerance, intolerance of criticism, and even reckless driving to be personality traits associated with increased violence. Reid, Balis, and Sutton (1997) reported similar findings, adding that such individuals also tend to display little insight into their behavior and frequently externalize blame for their problems.
38 These personality traits tend to be associated with sudden, sporadic violence rather than premeditated or well planned violence. While obtaining a history of repetitive antisocial behavior is well recognized as instrumental in any violence risk evaluation, the concept of psychopathy has received a great deal of attention in the literature as well. First described systematically by Dr. Hervey Cleckley in 1941 (Cleckley, 1982) psychopathy includes a specific set of personality traits and behaviors. Psychopaths are said to be superficially charming, often appear remarkably normal to observers, and tend to mak e a good first impression. But they are self centered, impulsive, dishonest, they lack empathy, and they frequently engage in irresponsible behavior just for fun or to satisfy their own needs. They are driven primarily by self gratification. They are mostl y devoid of guilt, empathy, or love. When involved in interpersonal or romantic relationships they are described as unfeeling, cruel, and heartless. Psychopaths frequently make excuses for their reckless and outrageous behaviors, externalize blame for thei r misdeeds, and seldom learn from their mistakes or benefit from negative feedback. Hare, Hart, and Harpur (1991) developed the Psychopathy Checklist Revised (PCL R) in order to tap into these attrib utes and offer a validated measure of psychopathy. Salekin, Rogers, and Sewell (1996; 1997) and Salekin, Rogers, Ustad, and Sewell (1998) found that psychopathy is a strong predictor of violence and criminal behavior, and therefore emphasize the importance of assessing psy chopathy in any evaluation of dangerousness. Cognitive Impairment There is ample evidence in the literature to indicate that brain injury can lead to aggressive behavior. Data from the National Institutes of Health (1998; 2001) show that following brain injury individuals may display verbal and physical aggression when no such tendencies existed prior to injury. Studies have also identified some of th e typical features of aggression following
39 brain injury; these tend to cluster around the themes of impulse control deficits and absence of remorse. Specifically, such individuals display violent behavior with no purposeful objective or intent, they tend t o exhibit reactive aggression that is often set off by trivial stimuli, these eruptions occur suddenly and without any steady buildup, and there is a notable absence of reflection or planning prior to the violent act or concern and remorse preceding the ev ent. A further complication to assessing risk in such individuals is the episodic nature of their violence, which is characterized by frequent and long periods of relative tranquility. Kalunian, Binder, and McNiel (1990) examined the relationship between violence and organ ic psychosis in senile geriatric patients, and found this population, more than individuals with any other psychiatric diagnosis, to engage in more aggressive acts. Winstanley and Whittington (2004) investigated incidents of aggression in a general hospital setting, and found that individuals were experiencing some kind of cognitive deficit in 64% of these e pisodes. They went on to suggest that these cognitive impairments likely resulted in their inability to inhibit there behaviors, or significantly affected their perception of the intentions of staff and peers so that they felt compelled to aggress against them. Studies by Quinsey and Maguire (1986) Borum et al. (1996), and Emerson (1998) have each confirmed a negative correlation between general level of cognitive functioning and violence risk. And more recently, Van Winkel et al. (2006) reported significant changes in IQ scores in a longitudinal study of schizophrenic patients, and patterns of deterioration related to these IQ scores. Perry, Pescosolido, Mar tin, McLeod, and Jensen (2007) suggest that their findings support a claim of impaired cognitive functioning related to the more acute phases of severe psychiatric illness, which in tur n may explain the greater frequency of aggression in those
40 who, due to mental illness, may not be able to fully appreciate the consequences of their behaviors. Clinical Judgment in Risk Assessment Determination of an individuals risk of future violence is frequently a critical component of a mental health assessment. Evaluation of an individuals level of dangerousness is particularly common in a number of situations, including emergency psychiatric evaluations, involuntary commitments, seclusion and restr aint, release and discharge decisions, probation and parole, and death penalty evaluations. Risk and dangerousness are not psychiatric diagnoses, but rather clinical judgments based on legal (and frequently political) policy, of an individuals procliv ity for violent behavior. And there are no diagnostic techniques, interviews, or psychometric instruments which predict future violence with a great degree of accuracy in all populations. Relatively infrequent events are particularly hard to predict. Murder, for instance, is more difficult to predict than more common offenses, such as domestic violence, due to its low base rate of occurrence. The accuracy of such predictions, however, varies greatly based on a number of variables, not the least of which is the period of time for which the behavior is being predicted. Accuracy is low for long term clinical prediction of violence, because of low base rates of various assaultive behaviors, but also due to often vague measures of violence which can produce a hi gh rate of false positives (Monahan, 1981) Apperson, Mulvey, and Lidz (1993) found that the accuracy of clinical violence prediction can be influenced by a host of factors including t he prediction techniques and outcome measures, the composition of the study sample, as well as the duration of the follow up period. Interestingly, they also found that 80% of violent episodes in their sample took place over 72 hours post admission, which speaks to the importance of longer follow up periods, particularly
41 since many studies have restricted their prediction time frame to the first few days after psychiatric admission (McNiel, Binder, & Greenfield, 1988) By virtue of the limited period of time they cover these studies tend to confine violence risk assessment to patients displaying more acute psychotic symptomotology. But Hoptman et al. (1999) question the relevance of such studies to the assessment of patients in forensic patients, where residents are frequen tly transferred from other supervised settings in which their more severe psychotic symptoms are likely to have been remediated through some kind of treatment. Early studies on the accuracy of clinical judgment in predicting future instances of violence su ggest accuracy rates by trained professionals (psychologists, psychiatrists, social workers, and other forensic experts) no greater than those of laypersons (Swanson, 2008) In an early review by Monahan (1981) it was found that psychiatrists and psychologists evaluating institutionalized mentally ill patients, who had committed violence in the past, were accurate in no more than 1 of every 3 predict ions of violent behavior over the 2year period during which they were followed. A striking limitation of such findings, however, is the lengthy follow up periods on which they are based. More recent studies examining clinical judgment in predicting futur e violence indicate that the accuracy of such predictions improves dramatically when limited to shorter periods of time (McNiel et al., 1988; McNiel and Binder, 1991; Apperson et al., 1993) McNiel and Binder (1991; 1995) examining risk assessment with psychiatric patients in civil hospitals, also found that clinical predictions of violence were more accurate than chance when the period of prediction was restricted to 1 week. In one such study, in a sam ple of 226 patients, they reported a sensitivity of 67% and specificity of 69% of their risk assessments. Scott and Resnick (2006) identify 5 components of dangerousness that may aid clinicians conducting violence risk assessments. The first is the magnitude of the potential harm, which
42 includes the like ly target of violence, whether the behavior is likely to involve harm to property or to person, and whether this harm is physical or psychological. All of these elements are said to be critical to understanding the degree of anticipated harm. The second of these main components is the likelihood of this harm, that is, the probability of the violent behavior and therefore the seriousness of an individuals intent to harm. Evaluation of an individuals past violent behavior, of course, can be helpful in predi cting whether violent acts are likely to be carried out in the future. The third and related component is an examination of the frequency of the threatened behavior. Beyond simply being a measure of past behavior, however, it takes into account the number of times a particular act has occurred over a particular span of time, with greater frequency aggressive behaviors suggesting a greater risk of those behaviors occurring again in the future. The fourth component speaks to the immediacy of the threat, for i nstance, whether an individual is threatening to act on violent thoughts within a few hours versus within a few days or months. The final component Scott and Resnick identify includes a complex of situational factors, such as association with deviant peer group, involvement with illicit drugs or alcohol, access to weapons, and a host of socio economic variables, from limited financial resources to family association with criminality. In predicting future violence Scott and Resnick also propose dividing aggressive behavior into affective violence and predatory violence. Where the former involves violent behavior in response to a perceived threat, the latter is thought to be more dangerous in that it often occurs in the absence of identifiable antecedents. H owever, in affective aggression, where an individual identifies a target of violence, the hostile behavior can be a reaction to a perceived threat based on an internal sense of anxiety or fear rather than from the external environment. Both types of aggres sion can lead to deliberate, purposeful, and goal directed violence. If there is one
43 distinguishing feature of the types of aggression identified by Scott and Resnick, it is the level of remorse typically experienced and displayed by their perpetrators; pr edatory aggression is seldom accompanied by remorse for the violent act. Actuarial Versus Clinical Methods of Violence Prediction While violence risk assessment is commonplace in forensic psychiatric settings, the focus on the development of assessment ins truments and decision making models in violence prediction has generally proceeded along two lines of research, one focusing on clinical models and the other on actuarial models. Traditional reliance on individual clinical impression as the sole source of decision making has come increasingly under fire. Grove and Meehl (1996) and Grove (2005) echo the sentiments of many critics when they described the clinical method as informal and subjective and impressionistic. (Grove & Meehl, 1996, pp 293294). These same critics have hailed the actuarial method as more formal and precise. The last two decades have seen the growth of these actuarial violence risk assessment instruments based on studies identifying predictor variables shown to be related to future violence. Once evaluated on these risk variables (for instance, during an evaluation of a psychiatrically hospitalized forensic patient, in consideration of placement in a less restrictive setting) an individual is later followed up in order to eva luate what combination of variables were best predictive of later violence. Another approach to the evaluation of violence risk comes from the development of structured clinical instrument s. These tools were developed out of previous research and clinical experience rather than derived through follow up of samples of individuals. They guide clinicians decision making in evaluating an individuals violence risk potential. These seemingly disparate techniques have led to what Monahan (1981) long ago called the Clinical Versus Actuarial Debate.
44 The semin al work by the MacArthur Foundation (Monahan, 2001) and data on which the largest studies to date examining risk factors for violence in institutionalized patients are based, have tended to conclude that actuar ial methods were greatly superior to clinical methods of prediction, and have demonstrated the largest effect sizes on predictive accuracy across clinical samples. A specific criticism of structured clinical evaluations is the subjectivity of their methods in that they are thought to guide therapeutic decision making but that the final determination is arrived at, for all intents and purposes, intuitively and impressionistically (Loewenstein, Weber, Hsee, & Welch, 2001) These data suggest that the debate seems to be hardly a debate at all, and that unstructured clinical violence risk assessments were long ago shown to be inferior. Others, however, do not feel that the debate is dead, or at least not beyond resurrecting (Buchanan & Leese, 2001; Litwack, 2001) They argue that th e implied superiority of actuarial methods over clinical assessment has not been established, and that such judgments are based on faulty interpretation of study data. For instance, Rice, Harris, and Quinsey (2002) and other critics suggest that the accuracy of actuarial methods has been greatly overstated in light of the low base rate of violence, which produces an i nordinate rate of false positive errors. That is, many non violent individuals are falsely classified as high risk and may be falsely sanctioned or detained in order to accurately capture the few individuals who represent an actual threat of violence. But even the base rate of violent offending in psychiatric populations is hotly debated, as studies have found rates from 25% (Monahan et al., 2001) 30% (Harris, Rice, & Quinsey, 1993) to even 50% (Rice, Harris, & Quinsey, 2002) which greatly affects the confidence with which predictions can be made. While extreme positions on this debate and over reliance on solitary methodologies or philosophies should be suspect, there is growing consensus in the literature that risk assessment
45 decisions that are based exclusively on clinical impression are inadequate. Studies have demonstrated that actuarial techniques can lead to high levels of predictive accuracy in va rious psychiatric populations (Buchanan & Leese, 2001) At the same time, strict adherence to actuarial prediction methods has also been criticized. While the use of actuarial models has gained in popularity, several studies have raised specific criticisms regarding their use. Douglas, and Ogloff (2003) suggest that they offer only vague estimates of risk, and that they lack generalizability beyond specific populations. Silva, Weinstock, and Leong (2003) suggest that such instruments do little to guide violence prevention and management of risk, and that results obtained are generally insensitive to change. The distinction between actuari al and clinical instruments is not always so clear. Litwack (2001) conduct ed a review in which he compared actuarial and clinical violence risk instruments, and pointed out that many actuarial methods include elements which require clinical judgment. In any case, the literature appears to bear out the dangers of using any of the se measures in isolation. Use of actuarial instruments represents an effort to standardize violence risk assessment but is not meant to replace all other methodologies. Borum (1996) indicates that these measures are simply meant to guide clinical judgment so that it is consistent with sound and prevalent research. Glancy and Chaimowitz (2005) remind us that while the debate continues, the preeminent goal of risk assessment in clinical practice is the reduction of harm and the management of threat. As such, all techniques and tools which inform clinical decision making should be conside red. Structured professional judgment has evolved as another research based technique for the assessment of violence risk. This model utilizes a professional guideline approach to decision making (Hart, 1998; Otto, 2000; Hart, 2001; Hart, 2001; Douglas & Kropp, 2002; Douglas &
46 Ogloff, 2003; Douglas et al., 2003) and has led to the development of numerou s schemes for guiding clinical judgment. As is common to other methods based on the professional judgment technique of violence risk assessment, mental health clinicians review all clinically relevant and available data in order to identify whether particu lar risk factors are present. These risk factors are described, defined, and operationalized based on the research literature examining their relationship with violence, and a final judgment of violence risk is then made based on these factors. At the same time, the benefits of actuarial prediction cannot be discounted. Their pitfalls are many, and include limited generalizability, the potential exclusion of low base rate but crucial information, the use of relatively static variables, and their arguably li mited contribution towards informing treatment, prevention, and management of violence. As elucidated above, however, actuarial prediction increases reliability and predictive accuracy, and promotes accountability and transparency in risk evaluation. Three of the most popular risk assessment schemes, which vary in their clinical vs. actuarial orientation, are the Historical/Clinical/Risk Management 20item (HCR 20) scale (Webster, Douglas, Eaves, & Hart 1997) Psychopathy Checklist Revis ed (PCL R) (Hare et al., 1991) and the Violence Risk Appraisal Guide (VRAG) (Harris & Rice, 1991; Harris, Rice, & Cromer, 2002) These instruments attempt to provide more structured and systematic approaches to the prediction of violence and the determination of dangerousness, as well as offering a clearer path to decision making and a view of the factors considered in making those determinations, while allowing for some sc rutiny of the rationale behind the decisions that are made. Risk Assessment Instruments Clinicians who conduct violence risk assessments are increasingly relying on standardized instruments to guide their evaluations and aid in their predictions. These t ools not only direct information gathering but can tie clinical judgment to existing research on the variables most
47 relevant to the assessment of future violence. The PCL R is among the most validated and widely used risk assessment instruments. As its nam e implies, it is designed to measure an individuals psychopathic characteristics, such as callousness, remorselessness, lying, and failure to accept responsibility for behavior. Higher scores on this measure have been found to be associated with higher ra tes of recidivism in various clinical populations (Hare, Clark, Grann, and Thornton, 2000) The PCL R uses specific scoring criteria to rate each of 20 items along a 3 point scale (from 0 to 2) derived from a semi structured interview and available historical information. An individuals total score can range from 0 to 40, according to how much their responses are representative of the psychopathic prototype, with scores over 30 generally being reflective of significant psychopathic tendencies (Vitacco & Rogers, 2001) While it was originally developed as a measure for this specific personality disorder, it has become increasingly used as a risk assessment tool beca use it has demonstrated predictive accuracy in many populations and for all types of violent offenses (Gray et al., 2004) Hart et al. (1991) demonstrated that the PCL R has a stable factor structure which capt ures affective and interpersonal traits as well as the behavioral elements of psychopathy. And Cooke (1998) has demonstrated the favorable psychometric properties of the PCL R. Hare also developed a 12item screening version of the PCL R called the PCL SV. Hill, Rogers, and Bickford (1996) demonstrated that this brief adaptation has good predictive validity when used to estimate violence within institutional settings. Monahan et al. (2000) have also demonstr ated its predictive validity in community violence. However, Freedman (2001) and Salekin, Rogers, and Sewell (1996) suggest that the PCL R has the potential to over predict violence because it may fail to measure factors which mediate and moderate the likelihood of futu re violence, such as protective positive attributes that mitigate against violence including resiliency and social competence.
48 Also, Dolan and Doyle (2000) point out a potential confound in that some of the PCL R and PCL VR items may be bound to the outcome variables of interest, like violence itself, in such checklist items as past criminal activity. They reco mmend various methods for statistically controlling for this potential problem, including removing these items from analysis. Nonetheless, there is ample evidence for the efficacy of this instrument for predicting violence in mentally ill criminal offender s (Gray et al., 2007) The Historical/Clinical/Risk Management 20 item (HCR 20) scale is a risk assessment instrument developed by Webster et al. (1997) which combi nes clinical judgment of risk with an evaluation of relevant historical factors. The HCR 20 is named for its 20 risk factors which comprise the 3 domains of historical factors, clinical factors, and risk management factors. One of the clinical items is psy chopathy, and is determined through use of the PCL R. The HCR 20 was developed to be used with a wide variety of clinical and offender populations, and its effectiveness in predicting violent behavior has been demonstrated in numerous studies (Gray, Hill, McGleish, & Timmons, 200 3) In separate retrospective analyses Strand, Belfrage, Fransson, and Levander (1998) and Douglas et al. (1999) have demonstrated t he predictive accuracy of the HCR 20 in various forensic and civil psychiatric populations. While Grann, Belfrage, and Tengstrm (2000) and Belfrage, Fransson, and Strand (2000; 2004) have obtained simila r results in prospective analyses. According to Gray and colleagues and Douglas and colleagues, the HCR 20 has also been found to be a better predictor of community violence than the PCL SV. They also noted that in regression analyses the HCR 20 added significant predictive accuracy beyond the PCL SV alone, and that the growing body of data appears to suggest that the HCR 20 may be a superior instrument for the prediction of violent recidivism in psychiatric populations than the PCL R or PCL SV.
49 The Viole nce Risk Appraisal Guide (VRAG) (Harris et al, 1991) differs signific antly from the risk assessment instruments described above in that it represents an attempt at a statistically derived and algorithmic approach to evaluation of violence risk. It is an empirically derived actuarial instrument based on the authors review o f files from 1965 to 1980, of adult male psychiatric patients treated in a maximum security treatment facility, to predict violent recidivism. Specifically, they examined recidivism among more than 800 serious offenders, approximately half of whom were pat ients in a forensic psychiatric facility while the other half consisted of non psychiatric convicted criminals. The VRAG uses 12 personal characteristics of an offender to give the probability, in terms of a percentage, that a criminal offender will commit a new violent offense within a specified period of time (typically with access to the community), and provides a comparison of an offenders risk potential to other offenders. It should be noted that the 12 items on the VRAG are scored using a weighing procedure developed by the authors based on a calibration sample of 618 men who committed serious, violent and/or sexual offenses and that the heaviest weighting is the PCL R score (Dolan & Doyle, 2000) In addition to psychopathy, items are related to general antisocial tendencies including childhood behavior problems, a history of non violent offending, and other personality disorders. The VRAG is one of the most extensively researched actuarial risk assessment instruments, and its utility and predictive validity have been widely demonstrated with a variety of populations, client cultures, and treatment settings (Lindsay et al., 2008) including individuals with intellectual deficits (Quinsey, Book, & Skilling, 2004) Using Rece iver Operator Characteristics (ROC) curves, Rice and Harris (1995) found that the VRAG predicted violent re offending with Area Under the Curve (AUCs) of 0.75, 0.74, and 0.74 for 3.5, 6, and 10 years, respe ctively. However, they reported less impressive results in their examination of the VRAGs predictive accuracy with
50 sexual recidivism in a population of pedophilic sex offenders. Webster et al. (1994) criticize the VRAG for its reliance on static factors and recommends that it be supplemented by a clinical checklist which may better capture the contribution of dynamic factors. Numerous studies are beginning to emerge that directly compare the predictive accuracy of both actuarial and structured clinical risk assessment instruments on the same sample of offenders. Remarkably, many of these studies have reported few significant differences between these instruments in their ability to predict future violence (Barbaree, Seto, Langton, & Peacock, 2001; Kroner & Mills, 2001; Sjostedt & Langstrom, 2001; Tengstrm, et al., 2001; Glover, Nicholson, Hemmati, Bernfeld, & Quinsey, 2002) The prediction of future violence is far from an exact science. But as clinicians are increasingly called upon to assess patients level of dangerousness and their potential for future aggressive behavior, they must increasingly rely on various techniques guided by research findings on which to ground their decisions. Mental health clinicians typically do not make unequivocal statements regarding an individuals future behaviors, but rather make probabilisti c statements about the likelihood of future behavior. As such, their predictions will not always be correct. The use of valid and appropriate risk assessment tools in conjunction with a detailed history and clinical evaluation, however, greatly increases t he accuracy of assessment of future dangerousness.
51 CHAPTER 3 MATERIALS AND METHOD S Methodology Study Setting The present study is an attempt to characterize the utility of diagnostic variables as predictors of violence in a population of residential ment ally ill criminal offenders during the acute care phase of their treatment. This study examined residents of the North Florida Evaluation and Treatment Center, admitted over a 3 year period (between July 2006 and July 2009). The center is a 216bed maximum security evaluation and treatment facility where individuals are committed by various Florida courts as either Incompetent To Proceed to trial (ITP) or Not Guilty by reason of Insanity (NGI) by virtue of major mental disorder, and on a multitude and range of criminal offenses. The center averages approximately 20 admissions per month. Residents are admitted from a number of jurisdictions compromising the criminal courts system of the State of Florida, from county jails, state psychiatric hospitals, and community placements. The stated mission of the facility is to evaluate, treat and discharge men with mental disorders, committed by criminal court, in a safe and secure environment and in a manner which ensures proper safety, security and respect for human r ights. A team of psychiatrists, nurses, psychologists, counselors and paraprofessionals develops a service plan for the evaluation and treatment of each resident with the goal of a successful return to the court or community. Final diagnostic formulations are based on comprehensive review of data obtained at this secure forensic mental health treatment facility. All residents are directly observed by treatment staff at maximum intervals of 30 minutes, 24hours a day, 7days a week. Residents of this facil ity reside in single rooms but interact with peers and staff, including treatment and security staff, and other personnel throughout the center.
52 A multidisciplinary treatment team evaluates each resident at least once per week during their first month of a dmission and then each month thereafter. All staff members maintain chart notes and share observations about each residents behavior and their abilities and interactions with other members of the facility. This information is then used in the evaluation a nd treatment of residents with the goal of restoring them to competency and providing for their eventual placement needs. Data Collection All residents of the forensic psychiatric facility receive a comprehensive clinical evaluation consistent with treatme nt standards. This evaluation includes psychiatric and psychological interviews and testing by licensed clinicians, information obtained by daily observations from staff, records from previous mental health providers, and review of court and other legal documents. Therefore, the majority of assessment procedures included in the present study were part of the usual treatment approaches and practices of the facility, with the exception of the coding of chart notes retrospectively by a single evaluator and sub sequent data analysis. All medical, psychological, and legal information, along with all relevant demographic characteristics were entered into a secure database in accordance with established privacy statutes. Recording of aggressive and noncompliant beha viors are incorporated into the treatment facilitys daily operations and standard operating procedures. While all staff members are involved in reporting of these incidents, documentation is typically carried out by each residents counselor, psychiatrist nurse, case worker, and unit staff members, each of whom has the closest and most frequent contact with each resident. During the study period several hundred ratings of aggression and disruptive behavior were retrospectively coded from these chart notes in order to characterize aggressive behavior in the sample.
53 Clinical, socio demographic, and legal information was obtained through direct interview with each resident and extensive review of medical and other records. Psychiatric symptoms and diagnoses w ere formulated from discrete clinical judgments based on criteria from the Diagnostic and Statistical Manual, Fourth Edition, Text Revision ( DSM IV TR, APA 2000) When available, historical and clinical information was also obtained from collateral sources, including family members and other appropriate informants. Incidents of verbal and physical aggressive behavior while at the treatment facility were assessed and documented by ward and other treatment staff and recorded in each residents chart. These incidents were then retrospectively coded using a modified version of the Overt Aggression Scale (OAS) in order to determine the frequency and nature of violence perpetuated by each resident in the sample over the first 5 weeks of admission to the treatment facility. And the presence of historical and clinical variables were examined in order to identify factors associated with these aggressive behaviors. Interventions by treatment staff, also contained in ward records, were also included in the analyses in order to determine the predictive utility of these variables in light of treatments each participant may have received in direct response to their aggressive be haviors. Measures The three main measures used in this study were a modified version of the Overt Aggression Scale (OAS), the Historical/Clinical/Risk Management 20 item (HCR 20) Scale, and the Brief Psychiatric Rating Scale (BPRS). Each was used to operat ionalize aggression severity and type, violence risk, and clinical symptom severity, respectively. The Overt Aggression Scale (OAS) A modified version of the OAS by Yudofsky, Silver, Jackson, Endicott, and Williams (1986) was used to describe the frequency and severity of aggressive behavior of each
54 participant. The original instrument was designed to be used by treatment professionals to record incidents of aggression and violence as they occur, in a variety of settings (including workplace, medical, and psychiatric). It was developed to assess not just the prevalence of aggressio n but the nature of that violence in psychiatric populations The OAS is a standardized behavior checklist for the rating of episodes of four types of aggression as they occur: verbal aggression, physical aggression against others, physical aggression aga inst self, and physical aggression against objects. The OAS appraises each of these types of aggressive acts along a range of severity by adding weightings to violent incidents in each of these categories. While ward chart notes documented all types of act ing out observed by staff as it occurred, the study analyses required the retrospective coding of these behaviors for each participant along the dimensions of the OAS checklist, adapted to include each 1 week period of the first 5 weeks after each partici pants arrival. A variation of the OAS, called the Modified Overt Aggression Scale or MOAS (Sorgi, Ratey, Knoedler, Markert, and Reichman, 1991) was developed during the course of a later study to assess the effects of Nadolol (a beta b locker) on aggressive behavior when the original OAS was found to lack usefulness because it required evaluators to record isolated incidents of aggressive behaviors every time an individual displayed that behavior. This met with significant practical limitations. It is not difficult to imagine the challenge this would pose to treatment staff in typical psychiatric settings, where a patient is likely to display any number of behaviors many times over the course of a day. It would simply be impractical for t hem to stop what they are doing and separately record each behavior every time it occurs. The MOAS represented an attempt to correct this limitation by evaluating longitudinal trends. Unfortunately, while it included the same 16 behavior items as the OAS i t eliminated the section on interventions. For
55 the purposes of the current study, the contribution of therapeutic interventions, particular the direct responses by staff to instances of aggressive behavior were instrumental to the analyses. Therefore, the original OAS was chosen as the method of recording aggressive behavior but was employed as a retrospective scale to capture each 1 week period of behavior, much as the MOAS was designed to do. The OAS scale has been found to demonstrate good internal, inte rrater, and test retest reliability (Kay, Wolkenfeld, & Murrill, 1988) It has also demonstrated utility in capturing more subtle forms of aggression that might otherwise go unrecorded. Reports of inappropriate behavior in inpatient settings frequently underestimate the prevalence of aggressive behavior (Larkin, Murtagh, & Jones, 1988) and incidents of verbal a buse, destruction of property and attempted assaults are often ignored. According to Kay et al. the OAS allows for balanced representation of overt aggression, though it has acknowledged limitations as a result of its subjectivity. Knoedler (1989) reported on the usefulness of the MOAS in discriminating among different patterns of aggressive behavior. According to his data on the validity and reliability of the MOAS subscales, the types of behaviors which distinguish each subscale are appropriate in terms of the developmental course, as it were, of the progression of violence it attempts to capture. He also demonstrated that different raters select similar subscales to characterize observed aggressive behaviors. Smith (1988) argued that these types of scaling procedures represent necessary requirements if one is to conduct clinical studies involving the use of ordinal statistics for comparing MOAS subscale levels. Smith (1988) also suggested that the four behavioral subscales of the MOAS shoul d be scored separately, because each subscale describes a different type of aggressive behavior. He
56 argued that listing the scores for each subscale separately allows clinicians to identify changes, for instance, from a pattern of physical aggression to on e of verbal aggression or vice versa, while an overall score remains unchanged. This change over time, however, was not a focus of the present study. A criticism of both the OAS and MOAS that has been levied in the past, and which is important to recognize is that both treat data as if they were measurable on an interval scale. Smith (1988) argued that aggression does not lend itself to such measurement, and if misused this instrument can lead to significant misinterpretation of results. He provided a useful example in illustrating this dang er by noting that one black eye [does not] equal three swearwords plus a mouthful of spittle. (p. 1081). At the same time, the overall OAS score provides a metric for aggressive severity that lends itself well to analyses supporting the use of continuous variables. Therefore, while the analysis of aggression severity in this study will proceed along these lines, in order to address the prior criticism, each type of aggressive behavior defined by the OAS will also be treated as discrete categories in an an alysis characterizing aggressive subtypes in the population. When properly used the OAS has been a successful adjunct to a variety of assessment and intervention studies. It represents an attempt at a concise, standardized clinical scale for the measuremen t of aggressive behavior that can ultimately enhance behavior and mental status reporting. The strength in the OAS, as it was utilized in the present study, was in its utility to assess trends in aggressive behavior over time and with repeated measurement. By setting limits on how the scale is to be applied, and acknowledging that the derived numerical scores have specific limitations in how they may be interpreted and applied, the OAS proved to be a useful scale for evaluating aggressive behavior in this s tudy setting.
57 As used in the present study, the scale consists of four categories of aggression (Appendix A 1), verbal aggression, self directed aggression, aggression against others, and aggression against objects. Each of these categories contains 4 st atements that describe aggressive behaviors of increasing severity. When a patient was noted to have engaged in a violent behavior, all of those statements which are applicable to that behavior were checked off by the evaluator and assigned a weighted scor e whose range varies by the scale: verbal aggression is scored on a scale from 1 to 4; aggression against objects on a scale from 2 to 8; physical aggression against self is scored on a scale from 3 to 12; and physical aggression is scored on a scale from 4 to 16. Additionally, an intervention rating is given in order to assess the severity of a specific behavior by capturing the force of a direct response by staff to that behavior. For instance, the use of physical restraints or intramuscular medication t o address a residents misbehavior garners a higher aggression severity score than a behavior requiring no intervention or only a verbal reprimand. Additionally, because the highly structured and controlled setting could reasonably be thought to limit the full expression of an individuals aggression potential if, for instance, they were put into isolation or restraints in response to misbehavior thereby limiting their ability to re offend (and leading to a dramatic underestimation of their level of violenc e potential) use of this intervention score was not only substantiated in the present study but offered an approach to conducting evaluations of this type in extremely structured environments. The frequency and severity of these four types of aggression we re thereby documented, and a global measure of aggression severity was derived from the sum of the weighed scores (with verbal aggression being assigned the lowest weight and physical aggression the highest) using the most serious behaviors in each categor y (with a range of possible scores of 0 to 16),
58 and a score representing the sum of all interventions conducted to address the behavior. Thus, the overall total score reflects the number of incidents over time, their severity, the type of aggression commit ted, and the response to the aggression, and is a psychometrically validated method developed by the OAS authors. The OAS has also demonstrated adequate interrater and test retest reliability coefficients, typically in the range of 0.78 and 0.87, respectively (Amore et al., 2008) and as high as 0 .9 for each (Krakowski et al., 1986; Krakowski et al., 2004) As mentioned, the OAS was used to rate all incidents of violence retrospectively in the current study, based on careful monitoring of residents and all their beh aviors occurring in the treatment setting over the evaluation period. These included behaviors within their rooms, in common areas, during both group and individual activities, and on the grounds of the facility. This was aided by video surveillance and cy cles of routine monitoring by personnel in 15 and 30minute intervals. Staff reported incidents of non compliance and violence, along with impressions of each resident, including their mood, demeanor, subjective distress, and verbalizations during convers ation were also typically included in ward notes. When overt aggression occurred, it necessitated detailed information and documentation from treatment staff of the event and staff response. The retrospective nature of behavior reporting in the present st udy successfully captured a variety of aggressive behaviors performed by patients daily, allowing estimates of the frequency and intensity of each type of aggressive behavior for each week of the 5 week study period. While this method may not represent an exact accounting of every aggressive act during this period, the benefits of a concise and standardized measure of aggression appears to outweigh the limitations of this strategy.
59 In addition to identifying those variables most strongly associated with vi olence, the effects of various interventions were also taken into account, as each residents medication regimen and status (e.g. restriction, seclusion, and restraint) had a potential effect on the measurement of their overall overt aggression scores. For the purposes of this study, aggression was operationally defined as any single occurrence of a behavior listed on the OAS instrument. Historical/Clinical/Risk Management 20 item (HCR 20) Scale For each subject, demographic and psychiatric information obta ined through clinical interview and extensive review of patient records were used to code the 10 Historical (static) and 5 Clinical (dynamic) correlates of violence risk identified by the HCR 20 and supported by research and established practice guidelines elucidated above (Appendix Figure B). Historical subscale items include the following: Previous Violence; Young Age at First Violence; Relationship Instability; Employment Problems; Substance Use Problems; Major Mental Illness; Psychopathy; Early Maladjus tment; Personality Disorder; and Prior Supervision Failure. Clinical subscale items include the following: Lack of Insight; Negative Attitudes; Active Symptoms of Major Mental Illness; Impulsivity; and Unresponsiveness to Treatment. Each item is coded on a 3point scale with values of 0 (available information does not indicate the presence of the item), 1 (available information suggests the possible presence of the item), and 2 (available information indicates the presence of the item). Because use of the P sychopathy Checklist (PCL) is required in order to complete the psychopathy item of the HCR 20, subject data were also coded in order to capture the 12 items encompassing the screening version of this instrument (the PCL SV). These items include an evaluat ion of emotional and interpersonal characteristics (including grandiosity, superficiality, lack of empathy, deceitfulness, and lack of remorse or personal responsibility) and behavioral attributes (lack of behavioral control and impulsivity, childhood and adult antisocial behavior, and irresponsibility and lack of goals).
60 The HCR 20 normally includes 5 Risk Management Items that capture elements related to release and treatment planning, and required services and support. Risk Management items were not com pleted as a matter of standard practice at this facility, as residents are considered to have just arrived at the time their first comprehensive psychological evaluation is conducted, and during which much of this studys data were obtained. As Gray et al. (2003) point out in their study of the efficacy of the HCR 20 in predicting self harm in mentally disordered offenders, omission of the Risk Management items is justified in cases where the projection of conditions relevant to release and discharge planni ng on which these scores are based are either not possible or not relevant to the imminent evaluation of risk. Therefore, the HCR 20 measure in actuality captured only the15 Clinical and Historical items, and will instead be referred to simply as the HCR for the purposes of the present study. Assessment of Psychiatric Symptoms Using the Brief Psychiatric Rating Scale (BPRS) In order to evaluate a broader range of psychological symptom variables than the HCR Clinical items are able to capture, each subjec t was assessed using the Brief Psychiatric Rating Scale (BPRS). Ratings were completed weekly, with scores averaged across the 5 week evaluation period. The BPRS was originally developed by Overall and Brown (1962) as a quick way to measure psychiatric symptom change over time, in various clinical p opulations and with individuals with a wide range of diagnoses. It has since become one of the most commonly used semi structured instruments for the assessment of psychiatric symptoms in both clinical and research settings (Jacobs, Ryba, & Zapf, 2008) The original 16item BPRS was expa nded to 18 items with the addition of the two symptom items of excitement and disorientation (Overall & Klett, 1972) and later to 24 items through the addition of six more symptom items, including bizarre behavior, elevated mood, motor hyperactivity, distractibility, self neglect, and suicidality (Lukoff, Nuechterlein, & Ventura, 1986) These additions were made in order to increase the
61 instruments sensitivity to a broader range of psychotic and affective symptoms (Ventura, Nuechterlein, Shaner, & Liberman, 2000) The BPRS is particularly useful in helping to judge the effectiveness of treatment in patients who have moderate to severe psychoses (Fayek, Kingsbury, & Simpson, 2002) Trained clinicians rate each of these symptom constructs along a 7 point Likerttype scale of severity ranging from 1 indicating not present/no impairment to 7 ind icating extremely severe in order to best describe a patients present state. These judgments are based on both observation and verbal reports by the patient, and depend on the symptom domains being assessed. The BPRS allows for the extraction of symptom clusters and dimensions, which vary somewhat from study to study, but typically include positive and negative symptoms of psychosis, mania, and depression and anxiety. The symptoms comprising each cluster of the 24item BPRS are as follows: a) positive sy mptoms encompass bizarre behavior, unusual thought content, disorganization, hallucination, and suspiciousness; b) negative symptoms include blunted affect, motor retardation, emotional withdrawal, and self neglect; c) mania comprises motor hyperactivity, elevated mood, excitement, distractibility, hostility, and grandiosity; and d) depression/anxiety encompass depression, anxiety, suicidality, and guilt (Lukoff et al., 1986) Each BPRS provides well elucidated anchor points and probe questions for each item which, when accompanied by appro priate clinician training, have been shown to increase the reliability of BPRS scores (Ventura, Green, Shaner, & Liberman, 1993) Numerous studies have demonstrated the sensitivity, effectiveness, and good interr ater reliability of the BPRS as a measure of psychiatric symptoms over time (Hedlund & Vieweg, 1980; Hafkenscheid, 1993; Ventura et al., 1993; Roncone et al., 1999; Lachar et al., 2001) These studies have also examined the BPRS across psychiatric settings, and with clinicians who varied in their level of
62 training on the instrument. Concordance rates across all BPRS items were found to be adequate when bas ed on patient verbal report and behavior, demonstrating acceptable internal validity (Greenwood & Burt, 2000) And good concurrent validity has been demonstrated, with high correlations between global and subscale scores between the BPRS and the Brief Symptom Inventory. BPRS symptom data have been factor analyzed in a number of studies, in order to clarify facets of mental illness gathered from various psychiatric populations. But the vast majority of these analyses have been conducted using the 16 and 18item BPRS, and generally have limited their investigations to a subset of patients, such as chronically ill schizophrenics (Malla & Norman, 1983; Overall & Beller, 1984; Long, Harring, Brekke, Test, & Greenberg, 2007) There exist o nly a few published factor analyses of the 24item BPRS. Van der Does, Dingemans, Linszen, Nugter, and Scholte (1993) reported a fourfactor model, similar to the one described above, using a population of recent onset schizophrenia patients. Their principal components analysis of the BPRS scores yielded a four dimensional structure (positive symptoms, negative symptoms, disorganization, and depression) and a categorical analysis resulted in predominantly positive, negative, and disorganized clusters (though more than half of their sample reportedly could not be classified to any of t hese groups). Dingemans, Linszen, Lenior, and Smeets (1995) demonstrated a five factor solution using a more diverse population of psychiatric patients. They added the BPRS items of Elevated Mood, Distractibility, and Motor Hyperactivity, to the 18item version in order to form a disorganization factor. Van der Does and colleagues (1993) had used t his same structure but termed the factor mania. Using a sample of mentally ill homeless subjects Burger, Calsyn, Morse, Klinkenberg, and Trusty (1997) also reported a five factor solution for the 24 i tem BPRS. They termed these factors positive symptoms, negative
63 symptoms, anxious depressive symptoms, hostile suspiciousness, and activity. Overall, it would seem that the research supports these types of factor structures but that the specific composit ion of factors for each version of the BPRS varies across samples. Therefore, the present study utilized the 18 item BPRS instrument with 5 factor structure that has been well validated for use with populations of psychiatrically hospitalized forensic pat ients (Appendix Figure A 3). These 18 clinical variables can be viewed as clusters of closely related symptoms, with each factor indexing a single underlying latent construct whose individual items correlate more highly with each other than with symptom items in other clusters. These factors are Affect Symptoms (which includes somatic concern, anxiety, guilt, and depressed mood), Positive Symptoms (consisting of conceptual disorganization, grandiosity, hallucinatory behavior, and unusual thought content), Negative Symptoms (consisting of emotional withdrawal, motor retardation, and blunted affect), Resistance Symptoms (including hostility, suspiciousness, and uncooperativeness), and Activation (with the symptom variables of tension, mannerisms and posturing excitement, and disorientation). Correlations between BPRS symptom clusters as well as BPRS items were examined to determine which psychiatric symptoms and symptom clusters were related to the types and frequencies of aggressive behaviors observed, and t o individual historical and demographic variables. Subsequent analyses were conducted to examine the multi determined relationships between symptoms, aggression, and patient history variables. Analytic Strategy Descriptive statistics and frequency distribu tions of the averaged aggression rating were examined to characterize the nature and prevalence of aggression in the study sample during the evaluation period. Initial analyses using Spearmans correlations were used to examine the relationship among conti nuous variables, while Chi square tests were used to test the difference
64 between categorical variables. Principle analyses consisted of using logistic and multiple regressions to compare the utility of the clinical, historical, and demographic characterist ics in building predictive models of aggression and violent behavior. Subsequent analyses utilized Receiver Operating Characteristics (ROC) curves, areas under the curve (AUCs), and odds ratios in order to address traditional criticisms of measures of pred ictive accuracy when they are highly dependent on the base rate of the criterion variable, such as correlation and indexes obtained from 2x 2 contingency tables (false positives and false negatives). Because violence is a relatively low base rate event, ev en in this population, ROC provided a more efficient method of estimating the predictive relevance of individual patient variables. According to Rice and Harris (1995) studies of risk assessment typically demonstrate a base rate of violence below 50%. At t hese levels, they argue, the base rate of the criterion variable significantly weakens the size of correlations. Additionally, ROC curves can be applied to data consisting of continuous predictor variables (such as symptom severity) and a dichotomous depen dent measure (such as membership in a subtype of violent offending). Areas under the curve (AUCs) were used as a measure of accuracy for identified predictive factors. While odds ratios were used to evaluate the effect of individual patient characteristics on violence group membership (as defined by the OAS). Douglas et al. (1999) recommend a cutoff of 2.5 or 3.00 as the lower demarcation of a strong association. The AUC value (also known as the concordance index or c statistic) provided by the ROC curve i s a combined measure of sensitivity and specificity that serves as an indicator of the overall performance of the logistic regression functions. In other words, this statistic is interpreted as the percentage of all possible pairs of cases in which the mod el assigned a greater probability to a correct case assignment than to an incorrect case assignment. It is the average
65 value of sensitivity for all possible values of specificity, with values ranging from 0 (no discriminating power) to 1 (perfect accuracy) Chance level performance is indicated by an AUC of 0.5, represented by a diagonal line on the graph. Therefore, while the numerical value provided by the AUC is informative ROC plots can also be used to define detection cut off points. A general guide to the accuracy of a function is as follows: 0.90 1, excellent; 0.800.90, good; 0.70.080, fair; 0.600.70, poor; 0.500.60, fail. ROC curves can have different levels of diagnostic value (Figure 3 1, left). Each of these ROC curves have different AUCs. Li ne A, which is never observed in real world cases, represents a perfect test, where the area under the ROC curve=1. The diagonal line represents chance level and has an AUC of .5. Curves B and C, which lie between these two extremes, represent curves with actual predictive value. Line B, with the wider arch that hugs the X axis has the greater AUC and, therefore, would have a greater ability to distinguish between violent and non violent cases in the sample. Additionally, ROC curves with equal AUCs can have different levels of diagnostic value. (Figure 3 1, right). Both of these curves have the same areas under the ROC curve, but they are not equivalent. In the high sensitivity range (high false positive range), test B is superior. While in the low sensitiv ity range (false positive rate range) test A is superior. In order to perform the full ROC curve analysis the residuals obtained in the logistic regression analyses were entered as independents, with the dichotomous violence dependent serving as the classi fication variable. Residual analysis proceeded utilizing an analog of Cook's distance D which is a measure of the influence of each case in the regression function taking into account the impact of individual cases as well as the magnitude of their stan dardized
66 residuals. Given the concern regarding the effect of extreme scores in the sample, this statistic also provides a measure of the effect that deleting cases has on residuals for all other cases. It should be noted that the determination of adequate sample size presented as a complex problem throughout the analyses. Sample size calculations were based on the number of covariates and the smallest proportion of negative cases in the sample, in accordance with criteria suggested by Peduzzi, Concato, Kem per, Holford, and Feinstein (1996). Additionally, regarding parameters for the standard error of area, analyses were conducted assuming a nonparametric distribution. (The smooth or fitted ROC curves represent parametric estimates, while the empirical (no n parametric) ROC curve are rutted due to the discrete points on which they are based). Tables 3 1 and 32 display summaries of each statistical strategy. Treatment of Outliers Extreme scores were expected within a population of mentally ill criminal offen ders on measures of mental illness (symptom severity) and maladaptive behaviors (aggression). Because a conceptual priority throughout the design and implementation of this study was its real world applicability it was felt that elevated scores, in particu lar, represented naturally occurring phenomena and statistically valuable information. When outliers were identified, an examination of these cases was conducted to see whether individual scores on relevant measures differed significantly so that these de viations could be adequately characterized. Capturing the influence of all scores in the data set was consistent with the conceptual and design priorities of the author. However, the potential impact of outliers on a sample of this size could not be taken too lightly. Therefore, each analysis was conducted both with and without the identified outliers excluded from the data set.
67 Figure 31. Receiver Operating Characteristics (ROC) Curves of Varying Predictive Value 1Specificity 1Specificty Note : Adapted from Park, Mo Goo, and Jo (2004)
68 Table 3 1. Summary of principal univariate analyses conducted. Variables Purpose Statistic Examples Categorical Sample characteristics Frequencies, Percentiles Primary diagnosis Group differences Pears ons Multidimensional Chi square Ethnicity Abuse history Head trauma Religious affiliation Admission number Charge severity Presence of violence Type of violence (verbal/physical) Veteran status Treating psychiatrist Marital Status Continuous Sample characteristics Descriptives Age Correlations Spearmans Rho Education Mini Mental Status Exam score Beck Hopeless Scale score HCR scores (Total, Clinical, and Historical subscales) OAS scores (Aggressi on Severity) BPRS scores (Affect, Positive, Negative, Resistance, and Activation clusters, and Total score) Note. BPRS = Brief Psychiatric Rating Scale, HCR= Historical/Clinical/Risk Management violence risk assessment scheme; OAS = Overt Aggressio n Scale.
69 Table 3 2. Summary of principal multivariate analyses conducted. Purpose Statistic Variables in each equation Aggression vs. non aggression Logistic Regression Model A (HCR Total) Receiver Operating Characteristics (ROC) Model B (HCR Total + BPRS Total) Area Under the Curve (AUC) Model T (HCR Historical + BPRS Activation + BPRS Withdrawal + Age) Aggression severity Multiple regression Model A (HCR Total) Model B (HCR Total + BPRS Affect + BPRS Positive + BPRS Resistance + BPRS Activa tion) Model T (HCR Historical + BPRS Activation + BPRS Withdrawal + Age) Aggression category Logistic Regression Model A (HCR Total) (verbal/physical) Receiver Operating Characteristics (ROC) Model B (HCR Total + BPRS Total) Area Under the Cu rve (AUC) Model T (BPRS Subjective Tension) Note. BPRS = Brief Psychiatric Rating Scale, HCR= Historical/Clinical/Risk Management violence risk assessment scheme; OAS = Overt Aggression Scale.
70 CHAPTER 4 RESULTS The final sample consisted of 102 mentally d isordered male criminal offenders residing in a maximum security forensic treatment facility. These were sequentially assigned cases to this single author in the Department of Psychology of the North Florida Evaluation and Treatment center (NFETC). The sam ple initially consisted of 105 participants but 3 were excluded due to incomplete records (this unavailability of pertinent records also precluded an assessment of any differences between the excluded cases and the final study sample). Sample characteristi cs are presented in Appendix B. Of these cases, 92 had been adjudicated by their committing court and admitted to the NFETC as ITP, while 10 had been adjudicated as NGI. Treatment modalities for residents at the facility did not differ between type of disp osition. Multi dimensional Pearsons Chi square confirmed that there was no relationship between disposition (NGI vs. ITP) and subject classification as aggressive or non aggressive, 2 = (1, N =102)=1.448, p=0.315, and no relationship between disposition a nd verbal aggression, Pearsons 2 (1, N =102)=0.222, p = 0.638, or physical aggression, 2 =(1, N =102)=1.478, p=0.224. The ethnic/racial composition of the study sample was as follows: 67 African Americans (65.7%), 30 Caucasians (29.4%), 4 Hispanic (3.9%), and 1 Asian (1%). The mean age of the sample was 36.75 ( SD =12.44) with a range of 19 to 77 years. The mean level of education was 12.58 ( SD =11.82). For 74 of these residents (72.5%) this was their first admission to this facility, while 24 (23.5%) of the residents in this sample were on their second admission, and 4 (3.9%) were on their third. In each case, the evaluation period consisted of the first 5 weeks of the most recent admission for each subject.
71 The distribution of primary intake diagnoses for pa rticipants, corresponding to DSM IV classifications (2000), was as follows: 83 (81.4%) Schizophrenia and Other Psychotic Disorders, 9 (8.8%) Mood Disorders, 1 (1%) Anxiety Disorders, 1 (1%) Adjustment Disorder, 7 (6.9%) Substance Related Disorders, and 1 ( 1%) Delirium, Dementia, and Amnestic and Other Cognitive Disorders. It should be noted that many of residents carried overlapping psychiatric diagnoses, particularly comorbid substance abuse/dependence as well as a host of medical problems. Nonetheless, t hese primary diagnoses correspond to and are consistent with established procedures favoring parsimony in computerized record keeping. Primary intake diagnosis was made by one of 5 attending psychiatrists who also completed the weekly BPRS ratings. Multi dimensional Pearsons Chi square confirmed that there was no relationship between rater and symptom severity as assessed by the BPRS factor scores. The severity of participants criminal offenses was assessed by noting the most severe charge in each case a ccording to the Florida Criminal Code (2009) which is consistent with established protocol in the reporting of crime statistics (FBI Unified Crime Reports, 2004). The distribution of charge severity in this sample was as follows: 7 (6.9%) 1st degree feloni es, 40 (39.2%) 2nd degree felonies, 44 (43.1%) 3rd degree felonies, and 11 (10.8%) capital offenses. Appendix tables B 1 and B 2 present the descriptive characteristics of the sample, including the distribution of scores across relevant measures (i.e. HCR BPRS, and OAS scores) and other variables, including total, subscale, and factor scores in each case. The mean HCR Total score was 18.7 ( SD =4.99, range: 730). The mean HCR Historical items subscale score was 12.73 ( SD =3.66, range : 3 20). And the mean HC R Clinical items subscale score was 6.02 ( SD =2.2, range: 110). The tables of sample characteristics also include the proportion of aggressive and non aggressive cases based on the authors operational definitions (n=59 and 43,
72 respectively), the frequency of specific behaviors in each category of aggression captured by the OAS, and other relevant features of the sample. Univariate Analyses Bivariate correlations were conducted in order to identify salient associations between variables. It was hypothesized that these relationships were likely to be influenced by some combination of several factors. While non parametric correlations were initially proposed in order to sidestep all parametric assumptions about the nature of the underlying data distribution, a number of methods were employed to examine the univariate normality of the data. First, the author employed the visual method for detecting outliers from scatterplot and boxplots of the relationship between the interval level variables to identify cases that had an unusual value for a single variable. This was followed up by conversion of each variable score to standard scores. Due to the sample size univariate outliers were identified as cases with standard scores of 3.0 or beyond. Using these technique s, a few outliers were identified in both the dependent and independent measures, including 3 subjects who consistently produced exceedingly high scores on most measures. Before deciding whether to retain or omit outliers from the analyses an examination w as conducted of the raw scores that made these cases outliers (Appendix C 1). In each case scores were consistent and in the correct direction of the anticipated relationship. Because nonparametric tests are based upon rankings of magnitudes the contribut ion of extreme values in terms of the interval measures was greatly reduced compared to their effect in parametric statistical tests. Therefore, no transformations were undertaken and it was decided that these scores would be retained. Multivariate normali ty was addressed separately in subsequent analyses.
73 Spearmans rho was utilized as the non normal distribution of some data (particularly from demographic and lower order variables rather than principle dependent measures) required the use of non parametri c tests. As indicated previously, numerous demographic variables have been found to be associated with violent behavior. The majority of these relationships, however, were not replicated in this sample and very few were found to be associated with the pre sence of aggression, or associated with its severity or type. As expected, an association was found between the BPRS total score (the measure of symptom severity) and aggression frequency ( r =.281, p<.001) and aggression severity ( r =.298, p<.001). A small a ssociation was also found between aggression severity and Positive Symptoms ( r =.251, p=.001), Resistance ( r =.248, p=.001), and Activation ( r =.307, p=<001), and between aggression frequency and these same symptom clusters ( r =.261, p=.001; r =.237, p<.001; an d r =.314, p<.001, respectively). Surprisingly, no association was discovered between Negative Symptoms and aggression. Likewise, there was no relationship found between Beck Hopelessness Scale score (BHS; a measure of pessimism about the future and a usefu l diagnostic measure of suicidality in clinical populations) and aggression. Therefore, the anticipated protective benefits of depression in individuals with mental illness, as it relates to aggressive behavior, was not observed in this sample. In both the aggressive and non aggressive samples Pearsons multidimensional Chi square tests failed to reveal significant relationships between aggression severity, and between aggression type (verbal vs. physical) and any of the following categorical variables: primary diagnosis, treating psychiatrist, disposition (ITP vs. NGI), charge severity, admission number, ethnicity, veteran status, religious affiliation, marital status, trauma/abuse history, head trauma
74 history, or suspicion of malingering. However, the r elationship between charge severity and aggression category approached significance in the aggressive group, Pearsons 2 (3, N =102)=7.526, p=.057. While an association between admission number and aggression severity trailed a little further behind when t he whole sample was considered, Pearsons 2 (74, N =102)=92.933, p=.068. Due to the presence of multiple predictor variables comparisons of the individual contributions of each could not be accomplished by contrasting the simple correlation coefficients. Subsequent statistical techniques were chosen to capitalize on both the nominal and interval nature of the measures in exploring the linear relationships between predictor and criterion variables. Risk, symptom, and aggression severities (as assessed by HC R, BPRS, and OAS, respectively) were each measured on a continuous scale by using their overall/total scores, as was aggression frequency (also provided by the OAS instrument). Multivariate Analyses Correlations amongst predictor variables was expected giv en that both the BPRS and the HCR Clinical subscale, in particular, examine psychological functioning and assess aspects of an individuals presenting symptoms. This initially presented as a concern in what inferences could eventually be drawn regarding th e relative contribution of each predictor variable to the success a given model. However, correlations typically revealed weak associations between HCR, BPRS, and OAS scores. As many of the variables were correlated, however weakly, it was impossible to a ttribute the variance in the dependent variable explicitly to any one independent variable. On this basis, logistic regression was used to test several models based on a combination of HCR subscale scores, BPRS individual and item factor scores, and other salient personal variables, on the dichotomous dependent variable of identification of subjects as aggressive vs. non -
75 aggressive, with any codable behavior lending to a classification as aggressive. Multiple regression techniques were then used to test t he predictive utility of these models in assessing an individuals level of aggression severity during the evaluation period, with scores on the OAS serving as the criterion. Finally, frequency of aggressive acts was used to identify subjects as primarily verbal aggressors or physical aggressors, and the utility of the models in making these predictions was evaluated. It was thought that, taken together, these could characterize aggressive behavior in the sample and provide useful estimates of a participant 's likelihood of violent behavior and level of aggression severity. While variables could not be manipulated directly as in the case of a controlled experiment, these techniques allowed a test of several models in order to clarify precisely which of the o bserved sets of variables (scores on a number of predictors) are influencing the results and which of these gives rise to the best prediction of violence (the criterion variable). It should be noted that multiple regression typically requires a large numbe r of observations in order to have sufficient power to obtain meaningful results. The number of cases should far surpass the number of predictor variables included in the regression. According to Tabachnick and Fidell (2001) a minimum of five times as many participants as predictor variables is necessary, while ten times is considered acceptable. The current investigation utilizes a subject sample 11 times the number of predictors under investigation. Aim 1 P redicting Aggression vs. Non A ggression Normali ty Non parametric alternatives were employed for all analyses. However, an examination of residuals was conducted to assess the normality of errors and homoscedasticity assumptions. An assessment of QQ plots indicated that linearity assumptions were not vi olated. And inspection of the correlation matrix revealed most correlations were in the range of 0.2 0.3, with only one
76 reaching 0.40, indicating that no problems of multicollineraity between constructs existed. An analysis of the standardized residuals id entified 1 case beyond the 2 SD multivariate cutoff commonly recommended (case# 6635, SD=2.08552; standardized residuals greater than 2.58 are outliers at the .01 level, while those greater than 1.96 are outliers at .05 level). Because outliers are cases t hat have an unusual combination of values for a number of variables, and can significantly affect regression results, serious consideration was undertaken to evaluate whether this case should be excluded from the analysis since, as previously mentioned, ca pturing extreme scores was consistent with the conceptual priorities of this study. Visual inspection of this individual case revealed relationships between variable scores that were in the anticipated and logical directions. And because this case essentia lly fell on the cut off point (which could as easily have been set at 3 SD given the conceptual framework of this study), it was felt that this case should be included in the analysis. However, due to the relatively small sample size the potential effect of outliers is amplified considerably. For this reason, rather than removing this case from the analysis it was modeled separately. Model A In order to test a real world model representing standard practice in violence risk assessment, an analysis was ru n with HCR Total score as the sole regressor (hereafter referred to as Model A). Examination of the case processing summary revealed that no cell was empty or particularly small, and that all cases were used in the analysis. The classification table prese nted the initial success rate for this analysis before the regressors (predictors) were included. This intercept only (or null) hypothesis provided a baseline for comparing the improvement of subsequent models in correctly classifying cases. It indicate d that 100% of the observed aggressive subjects would be correctly categorized prior to inclusion of these predictors, while none of the non aggressive subjects would be correctly predicted. It also estimated an overall
77 correct classification rate of 57.8% With the predictor included the number of cases correctly classified rose marginally to 59.8%. Correct assignment of aggressive subjects was 79.7%, while correct assignment of non aggressive subjects was 32.6%. Omnibus test results provided a measure of the ability of all predictors in the model to jointly predict the dependent variable. The Enter method was employed in which all model terms were entered in one step. The finding of a significance likelihood ratio 2 (1, N =102)=7.849, p<0.1) indicated th at the predictor in Model A was significantly related to the response variable (classification as aggressive versus non aggressive), indicating adequate fit of the data to the model. The 2 log likelihood value (132.041) indicated a good fit when compared with the null model, which also indicated that the predictor was significantly related to the response variable. Assessment of the Hosmer and Lesmeshaw test values provided an additional measure of goodness of fit between the observed and predicted number of cases in the aggressive vs. non aggressive conditions. This test is considered more robust than the traditional 2 test, particularly in cases such as those in this study when continuous covariates are in the model and sample size is relatively small. T he high p value, 2 (8, N= 102)=6.71, p=.568, is another indication that the model adequately fit the data. As HCR Total score was the only regressor, it provided the only coefficient (along with the constant) for the linear logit equation. The coefficien t in the equation for violence, which was found to be significant in Model A is indicated as 0.110, Walds 2 (1, N = 102)=6.277, p<0.05, while the Exp(B) multiplier is 1.116. The coefficient for HCR Total score means that, on average, a one point increase in HCR score adds 0.1 to the log odds in favor of a participant being correctly identified as aggressive. That is, the odds in favor of a subject being aggressive are multiplied by e 0.1=1.116. This small effect was confirmed by the Nagelkerke R2 value in the
78 model summary which indicated that the predictor was accounting for only as much as 8.7% of the variation in classification. Therefore, the effect of the independent variable on the odds ratio in the final model was negligible. The analysis was re run with the identified outlier excluded. The case processing summaries revealed that the excluded case was one that was classified as non aggressive, and that the null hypothesis now demonstrated a 58.4% overall accuracy of case classification (up slightly f rom 57.8% in the original null model). With the outlier removed from the regression analysis correct assignment of violent subjects remained unchanged at 79.7%, while correct assignment of non violent subjects increased modestly from 32.6% to 33.3%, with a proportional increase in overall accuracy from 59.8% to 60.4%. The overall model was still significant, 2 (1, N = 102)=7.849, p<0.001, indicating that the predictor in the model was still significantly related to the response variable of violence classifi cation. Other model fit indices were comparable to the previous analysis, including 2 log likelihood of 129.291, and Hosmer and Lesmeshaw 2 (8, N = 102)=7.889, p=0.444, all indicating adequate fit of the data. The effect on the revised sample was still sm all, with the predictor in the model now accounting for up to 10.1% of the variation in classification (a modest increase from 8.7%). This indicates that the effect of the independent variable on the odds ratio in the final model was negligible, and that r emoval of the outlier from the analysis added little to the strength or predictive utility of the HCR Total score. The coefficients in the equation for violence changed negligibly, though the HCR Total score, which was approaching the .001 level of signifi cance now achieved it, Walds 2 (1, N = 102)=7.090, p<0.01. The coefficient in the equation for aggression increased from 0.110 to 0.120, while the Exp(B) multiplier increased from 1.116 to 1.128. These values, therefore,
79 remained essentially unchanged, suggesting that the identified outlier should be retained in subsequent analyses in order to preserve the integrity of the data. Model B The effects of including the BPRS Total score in this basic model were then evaluated, as the additive benefit of this clinical symptom measure in predicting violence above and beyond the HCR alone is a basic question clinicians utilizing this measure would pose. Model B, therefore, includes HCR and BPRS as regressors in the 2predictor model. The classification table for the null hypothesis indicated the same initial overall classification rate of 57.8% prior to inclusion of the regressors in the model. With inclusion of the Model B regressors this rose to 62.7%, with a correct classification rate of 78% for aggressive ca ses (down from 79.7% in the HCR only model) and 41.9% correct classification of non aggressive cases (up from 32.6%). The omnibus test for model coefficients indicated that the model as a whole was significant 2 (2, N = 102)=9.161, p=0.01, and indicators o f model fit were comparable to those in Model A (with 2 log likelihood=129.72; and Hosmer and Lesmeshaw Test 2 (8, N = 102)=6.149, p=.631). The combined variable model accounted for 11.5% of the variation in classification (up slightly from 8.7% in Model A). Additionally, examination of the regression coefficients indicated that the BPRS variable contributed negligibly to the model, as HCR Total was the only significant variable, B =0.100, 2 (1, N = 102)=5.054, p<0.05. The overall correct assignment of case s increased from 58.4% to 62.4% in the outlier excluded condition, with correct assignment of violent cases down from 79.7% to 77.6%, and correct classification of non aggressive subjects up from 33.3% to 41.9%. With the outlier excluded the overall model continued to be significant 2 (2, N =102)=9.003, p=.010, and model fit indices are still adequate ( 2 log likelihood=128.776, and Hosmer and Lesmeshaw Test 2 (8,
80 N =102) = 6.273, p=.617). The amount of explained variation in classification remained at 11.5 %, with or without the outlying case in the analysis. And once again, the HCR Total score continued to be the only statistically significant variable in the model, Walds 2 (1, N = 102)=4.957, p<0.05, B =0.099. Exclusion of the outlier had no appreciable ef fect on the model or the observed contribution of its variables. Test Model (Model T) Model T is a 4 predictor model consisting of variables chosen based on both theoretical considerations as well as the correlations observed in the univariate analyses. Th is model includes HCR Historical score (Historical), BPRS Activation cluster score (Activation), BPRS Emotional Withdrawal item score (Withdrawal), and participant age. Results of the regression analysis with these added variables indicated that the overall percentage of correctly assigned cases to aggression group increased from Model A, from 59.8% to 66.7%. The percentage of correctly classified violent cases was unchanged at 79.7%, but the correct classification of non violent cases rose from 32.6% to 48.8%. These represent minor improvements in the initial model, and speak to the questionable utility of the added variables in adding to the accurate prediction of aggressive subjects, though there appears to be a marked improvement in the classificat ion of non aggressive cases. The measure of the ability of all predictors in the model to predict the dependent variable was again given by the omnibus test, which indicated a significant likelihood ratio 2 (4, N = 102)=18.403, p=.001, and a good model fi t. The observed 2 log likelihood value (120.478) also indicated a good fit when compared with the null model, meaning that at least one of the predictors in the Test Model was significantly related to the response variable. This statistic is necessary in contrasting models with dissimilar numbers of regressors as it evaluates the
81 significance of the change in 2 that results from the removal of additional regressors from the full model. The Hosmer and Lesmeshaw test table confirmed the model fit between th e observed and predicted number of cases in the aggressive vs. non aggressive conditions, 2 (8, N =102) = 2.237, p=.973, indicating once again that the model adequately fit the data. The coefficients for the linear logit equation indicated that the interce pt (constant) and the coefficients of each regressor produced the following logit equation representing the contribution of each of the predictors: Logit= 0.111(HCR H) + 0.175(BPRS Activation) 0.065(Age) + 0.532 However, the p values from Walds Tests fo r all but the age and BPRS Activation variable were non significant, indicating that History and Withdrawal had a negligible effect. The BPRS Activation variable appeared to exercise the most influence on whether or not someone was classified as aggressive or non aggressive, (Activation B =0.175, Walds 2 (1, N = 102)=.730, p<.05) followed by participant age (Age B = 0.056, Walds 2 (1, N =102)=8.086, p=.01). The effect of the independent variables on the odds ratio is given by Exp(B), with the odds ratio det ermined by dividing 1 by the Exp(B) value. This indicates that participants of greater age were 0.945 times more likely to be aggressive (therefore less likely to be aggressive). Also, participants with higher Activation cluster scores (i.e. who experience symptoms related to general agitation) were 1.191 times more likely to be aggressive. This small effect is confirmed by the Nagelkerke R2 value that measures the strength of these of associations and indicates that the predictors in the test model are acc ounting for up to 22% of the variation in classification; a relatively small amount, but a considerable increase from the HCR only model of 8.7% The interpretation of the coefficients can be awkward. In this case, for instance, for a unit increases in age the log odds of being aggressive versus non aggressive increases by B = 0.056.
82 For the sake of clarification coefficients are exponentiated and interpreted as odds ratios, given by the value of Exp(B). In this case, for each year a participants age increas es their odds of being violent decreases by a factor of 0.945 (values less than 1 indicate that an increase in the value of the predictor variable is associated with a decrease in the odds of the event). The 95% confidence interval values indicate that the magnitude of this decrease is likely to be in the range of 0.910 to 0.983. And for each unit increase in BPRS Activation score, the log odds of being aggressive versus non aggressive increases by B =0.175, corresponding to each 1 point increase on the Acti vation score increasing the odds of being aggressive by a proportional amount (by a factor of Exp(B)= 1.191 ( CI.95=1.035,1.369). This 4 predictor regression analysis was re run with the previously identified outlier excluded from the analysis. Classificatio n accuracy was essentially unchanged across each group. The rate of correctly identified violent subjects remained unchanged at 79.7%. Classification accuracy of non violent subjects increased slightly from 48.8 % to 52.4%. And there was a small increase i n overall correct classification from 66.7% to 68.3%. Re examination of goodness of fit measures including likelihood ratio 2 (4, N =101)=21.532, p<.001), 2 log likelihood (115.609), and the Hosmer and Lesmeshaw Test 2 (1, N = 101)=3.691, p=.884, continue d to indicate adequate fit of the data to the model, and that at least one of the predictors was significantly related to the response variable. It should be noted that the Omnibus Tests of Model Coefficients was significant at a greater statistical level. With the outlier removed from the analysis b oth the Age and Activation items remained statistically significant (Activation, B =0.193, Wald 2 (1, N =101)=6.725, p=.01; Age, B = 0.064, Wald 2 (1, N =101)=9.600, p<.01). Removal of the identified outlier also allowed the HCR Historical item to achieve statistical significance ( B = 0.131, Walds 2 (1, N = 101) = 4.138, p<.05).
83 Only Withdrawal failed to influence the model. History now exercised a comparable degree of influence classifying cases as Activation ( B =. 131, with associated odds ratio Exp(B)= 1.140, CI.95=1.005,1.293). Participant age exercised the least amount of influence while still remaining significant, and the predictors now accounted for up to 25.8% of the variation in classification, an increase f rom 22.2% in the function that included the outlying case. Taken together, elimination of the outlier from the regression analysis had a significant impact in that the relative contribution of the History variable achieved statistical significant. However, its relative contribution was quite small, and the relationship between variables and their relative combined effects in the model predicting aggression classification changed marginally. Additionally, the overall predictive accuracy with the outlier remo ved increased very little, from 66.7% to 68.3%, with an equally small increase in the explained variation in classification from 22% to 25.8%. Results of these analyses suggest that young age and increased Activation cluster symptoms had a significant effe ct in the prediction of violence in the Test Model, while the Historical subscale score of the HCR (which characterizes static variables related to participants past experiences) contributes as well, but only when extreme cases are excluded from the analy sis. Otherwise it contributes negligibly to the correct classification of individuals as aggressive or non aggressive. Despite findings that mood symptoms (and depression in particular) may serve as a protective factor in violence risk, Emotional Withdrawa l (a BPRS item that captures this clinical entity) appears to have negligible predictive value. Table 4 1 displays the success rates of each model, while Table 4 2 displays the correlations, odds ratios, and influence of individual predictors in each model in correctly classifying non aggressive and
84 aggressive cases. Tables 4 3 and 44 display these same values derived from analyses with the outlier excluded from consideration. Graphical Evaluation of Predictive Models Model T, which includes the variables of age, Activation, Withdrawal, and History appears to be render better overall rates of accurate classification of cases than those that are based solely on the HCR Total score (Model A, and the traditional measure of violence risk), and Model B (the addi tion of BPRS Total score to the HCR measure) which speaks to their relative predictive utility. However, additional diagnostic measures should be applied before such assertions can be made. The histogram of predicted probabilities or classplot (Figure 4 1) provides an alternative way of assessing correct and incorrect predictions under the Test Model. The X axis indicates the predicted probability from 0.0 to 1.0 of the dependent classification (aggressive vs. non aggressive) being classified 1. Zero corre sponds to predicting that a subject will be non aggressive, while 1 indicates that he will be aggressive. The line of N's and Y's under the X axis indicate that a prediction of 0 to 0.5 corresponds to the case being classified as non aggressive, and 0.6 to 1 corresponds to a case being classified as aggressive.. The Y axis indicates the number of cases classified (frequency). Inside the plot are columns of observed Ns and Y's corresponding to noaggression and yes aggression classifications. A look at the plot of observed groups and predicted probabilities of the Test Model using the whole sample reveals a principally normal distribution rather than the desired U shaped distribution, which indicates that the predictions were not well differentiated. Instead the classplot suggests that many predictions were close to the cut point of 0.5, which would in turn suggest poor model fit. There also appear to be a number of errors; Y's to the left are false predictions of subjects as non violent, while the Ns to th e right are false predictions of subjects as aggressive. However, the plot also reveals that Model T classified difficult cases fairly well
85 (those at p = 0.5), which likely explains the adequate numerical goodness of fit indicators reported above. A final method for evaluating the logistic regression model makes use of Receiver Operating Characteristics (ROC) curve analysis in which the model's predicted values are used to differentiate between positive and negative cases. The power of these values to accom plish this differentiation in cases is quantified by the Area under the ROC curve (AUC). The logistic regression analysis of Model A (the real world model) identified a significant function using HCR Total score as the sole regressor, Walds 2 (1, N = 1 02) = 6.277, p<.05. It produced an overall correct classification rate of 59.8%, with a correct assignment of violent subjects of 79.7%, and a correct assignment of non violent cases of 32.6%. The ROC curves based on this model reveal the true diagnostic utility of the HCR total score when used alone, with a fair AUC value of 0.649 ( p=.011, CI.95=.543755) for aggressive cases (Figure 42), but poor prediction of non aggressive cases, with an AUC=0.351 ( p=.011, CI.95=.245457) (Figure 4 3). An additional Model A ROC curve was created with the outlier excluded from the dataset. This produces very similar results in both the aggressive classification AUC (.660, p<.01, CI.95=.554.765; rather than .649) and non aggressive classification (.340, p<.01, CI.95=.235.446; down from .351). Additionally, the form of the ROC curves are essentially identical and do not reveal any substantial differences in predictive reliability from the analysis utilizing all cases. With the addition of BPRS Total score Model B was al so found to be significant overall 2 (2, N =101)=9.161, p=.01, producing a slightly improved overall correct classification rate of 62.7%. But this improvement was based on an increase in the successful classification of non aggressive cases (to 41.9%) ra ther than better identification of aggressive subjects (which
86 actually fell 1 point to 78%). This is reflected in the ROC curve for Model B, which describes a very different function from the one produced by Model A. It describes an observed correct class ification of aggressive cases far below chance (AUC=.266, p<.001, CI.95=.169 .364) (Figure 44) but superior correct classification of non aggressive cases (AUC=.734, p<.001, CI.95= .636 .831) (Figure 45), with similar values observed with the outlier rem oved; AUC=.717, p<.001, CI.95=.616.817, for non aggressive cases and .283, p<.001, CI.95=.183.384, for aggressive cases. Additionally, with the outlier removed, the forms of the ROC curves are essentially identical and do not reveal any substantial diffe rences in predictive reliability from the analysis utilizing all cases. The numerical and graphical output for the ROC analyses (Table 4 5, Figures 46 and 47) indicates the true diagnostic value of the proposed model (Model T). Once again, the addition o f the BPRS items in the model appears to have created a function that describes a very different relationship from the one observed in the HCR only Model A, and one which is similar to Model B, which not coincidentally also utilized a BPRS variable (Total score). Violence group classification drops sharply in Model T compared to the HCR only model when predictions are made based on the proposed logistic regression (AUC=.312, p=.001, CI.95=.210 .141) (Figure 46). However, prediction of non aggressive cases based on the Test Model is vastly superior (AUC=.688, p=.001, CI.95=.586.790). (Figure 47). Examination of the combined plot (Figure 4 8) suggests that an essentially inverse but proportional relationship is obtained, with both the Test model and Model B evidently predicting opposite events from the HCR only model, and doing so fairly well, and showing proportional deficiencies in the prediction of membership in the other group.
87 With the outlier removed these values did not change appreciably. As before, the nature and magnitude of the observed relationships were essentially unchanged, with a similar and equally poor AUC for classification of aggressive cases in the Model T (.312, p=.001, CI.95=.210.414, down from .313) and fair AUC for classification of non aggressive cases (.687, p=.001, CI.95=.585.790, down from .688). The form of the curve also did not reveal substantial differences in the previously observed relationships. Aim 2 Predicting Aggression Severity Normality and statistical considerations The use of multiple regression requires adherence to similar restrictive assumptions of logistic regression, including multivariate normality of the data, the assumption of homogeneity of variance covariance matrices, and the avoidance of multicolineari ty. Given a larger sample size these procedures are sufficiently robust to cope with some skewness. But the problem of outliers is particularly serious given the study sample, and required probing of the data for violations of these underlying assumptions. In order to evaluate the presence of outliers for subsequent multivariate analyses the author evaluated Mahalanobis D2, a multidimensional version of the z score that measures the distance of a case from the multidimensional mean (centroid) of a distribu tion given the multidimensional variance (covariance) of the distribution. A case is considered a multivariate outlier if the probability associated with its D2 is 0.001 or less. It should be noted that D2 follows a 2 distribution with degrees of freedom equal to the number of variables included in the calculation. Mahalanobis D2 requires that the variables be metric, therefore this statistic is applicable to the interval level variables used to explore aggression severity in this study, which is measured as a continuous variable. In the dataset, no cases had Mahalanobis D with a probability less than or equal to 0.001, and therefore no cases were excluded from these analyses.
88 Though there was an adequate theoretical basis to expect some variables to exer t a greater influence over the Model T (for example, ratings of hostility more than motor retardation in predicting aggression within BPRS symptom clusters), the vast majority did not lend themselves easily to combinations reflecting firm theoretical c onsiderations or grounding in previous research findings. Therefore, the Enter (simultaneous) method was used to evaluate the relative contribution of each predictor variable in the multiple regression analyses. This also served to address concerns regar ding sample size. The square of the measure of correlation ( R2) specifies the proportion of the variance in the criterion variable which is accounted for the proposed model, thereby providing a measure of how good a prediction one can make of the criterion variable by knowing the predictor variables. It should be noted, however, that an adjusted R2 value is used that takes into account the number of variables in each of the models as well as the number of subjects/observations on which each model is based. This provides the most constructive measure of a models success. Standardized Beta weights ( ) measured in units of standard deviation, are calculated to provide a measure of how strongly the predictors impact the criterion variable, with higher beta values indicating greater influence of the predictor variable on the criterion variable. Beta re gression coefficients were computed in order to assess these associations and compare the strength of each predictor and criterion variables relationship. The first multiple regression analysis was conducted in order to once again evaluate the real worl d model (Model A) in which HCR Total score is the sole regressor. The next analysis evaluated the inclusion of BPRS factor scores to this function (Model B). Finally, the Test Model T variables of participant age, History, Activation, and Withdrawal were entered into a separate regression in order to evaluate their utility in predicting aggression severity.
89 Model A Examination of the model output revealed that Model A was significant, F (1, 2)=11.308, p=.001, but that it accounted for only 9.3% of the varia nce in aggression severity scores; not a particularly good model (Adjusted R2=0.093). This corresponds to a small effect size ( f2) of 0.1025. As Model A had only one regressor, ANOVA results confirmed the same overall model significance. As the HCR total s core was the sole regressor in this analysis, it is necessarily the sole influence ( =0.319) that along with the absolute t value ( t = 3.363, p=0.001) suggests that this variable was having a moderate impact on the aggression severity score. As Model A rep resents an evaluation of the traditional violence risk assessment instrument, the observed relationship follows the expected course: as the total score increases on this measure of violence risk there is an observed increase in the severity of aggression o bserved in the sample. Specifically, the Model A value indicates that a change of one point in this predictor variable will result in a positive change of 0.319 standard deviations in aggression severity score. Model B The resilience of the multiple regression procedure allowed for the exploration of t he contribution of a greater number of predictors than in the previous version of Model B and capture the utility of BPRS factor scores without the need to subsume them under the BPRS Total score. A priori sample size calculations for multiple regression indicated that with a desired statistical power level of 80%, an alpha level of .05, and a small anticipated effect size, the number of predictors (not including the regression constant) could be safely increased to 5 (the calculations in fact revealed that a sample size of 97 would have sufficed). Univariate analyses identified 4 of the 5 BPRS factor scores as correlated, though weakly, with aggression severity. These included the following: Affect ( r =0.212, p<.05), Positive Symptoms ( r =0.244, p<.05), Resis tance ( r =0.251, p<.05), and Activation ( r =0.336, p<.001). No Negative Symptom
90 cluster items (which are comprised of Emotional Withdrawal, Motor Retardation, and Blunted Affect) were found to be associated with aggression severity. It is noteworthy, however that Emotional Withdrawal was found to be associated with the presence of violence in the sample but not its severity, which supports its previous inclusion in Model T. These BPRS factor scores were then entered into the Model B multiple regression analy sis with the HCR Total score, in order to compare this combined model to the HCR only Model A. Table 4 7 indicates the descriptive characteristics of the study sample based on these new variables. The overall model was found to be statistically significan t, F (5,96)=3.908, p<.01) and now explained 12.6% of the variance in aggression severity. This represents a very modest improvement over Model A (at 9.3%). The Adjusted R2 of 0.126 corresponds to a moderate increase in effect size ( f2= 0.144165). But furthe r examination of the coefficients revealed that only History was significant ( t =2.853, p=.005). The BPRS factor scores appeared to contribute negligibly to the Model B function. Model T Examination of model output confirmed the overall significance of the Model T F (4,97)=11.308, p=.001) and indicated that the model now accounted for a similar percentage of the variation in aggression severity scores at 12.3%, down from 12.6% in Model B but still improved from Model A. This also corresponded to an increase i n effect size ( f2=0.140251). Overall, however, this was still not a great model. Examination of the coefficients further revealed that only Activation ( t =3.227, p<.001) and History ( t =2.40, p=.05) were significant variables, with the former exercising a sl ightly greater influence on the model (Activation =0.303; History =0.225). Age and Withdrawal appear to have had a negligible impact on the model. This is not surprising as neither was found to be associated with aggression severity in the earlier univar iate analysis, but both were retained in order to preserve the consistency of the
91 model. It should be noted that colinearity diagnostics revealed no significant correlations between predictor variables (History Tolerance value= 0.987, VIF =1.013; Activation Tolerance value=0.985, VIF =1.015). Post hoc statistical power calculations given the observed alpha level, the number of predictors, the observed R2, and the sample size indicate an observed power of 0.9377. Therefore, though the effect size of this model falls within the moderate range, the odds of observing an effect when it occurs are quite good. Table 48 displays a summary of each regression analysis with the relative contribution of each predictor in modeling aggression severity. Aim 3 Predicting A ggression Category The final analysis attempted to characterize the nature of violence in the study sample. Once subjects were identified as either aggressive or non aggressive (based on the same criteria employed in the initial logistic regression) and v ariables were evaluated for the degree of their association with aggression severity, final analyses were conducted to characterize aggression subtypes and test the predictive utility of Model T compared to Models A and B in successfully classifying these cases. The 4 violence groupings articulated in the modified OAS aggression instrument (verbal aggression, physical aggression against objects, physical aggression against self and physical aggression against other persons) were re worked in order to char acterize subjects as primarily verbal aggressors (Verbal) and physical aggressors (Physical). This transposition was easily affected as the aggressive behaviors of all subjects but one were characterized primarily by one of these 2 types based on a simple frequency of behaviors. The remaining case achieved higher scores within the self harm grouping of the OAS instrument than in the other 3. As he was the only subject to do so, and because self harm was felt to represent a unique type of behavior fairly dis tinct from those guiding the authors research hypotheses, this
92 case was coded according to the category in which this subject attained his second highest score, that of verbal aggression. Normality Because these analyses were conducted on a subset of the total sample, an assessment of normality was again conducted, with an emphasis on evaluating the normality of errors and homoscedasticity assumptions. An analysis of the standardized residuals identified 3 cases beyond the recommended 2 SD multivariate cut off (case# 705724, SD= 2.09379; case# 706549, SD=2.06410; and case# 706843, SD= 2.02036). These values fall between the .05 and .01 significance level cutoffs (of 2.58 and 1.96, respectively). For the reasons enumerated earlier, it was decided that these cases would be modeled separately in the subsequent analyses. Model A In order to evaluate the utility of Model A in differentiating the Verbal group from the Physical group HCR Total score was again entered into a logistic regression as the sole predicto r using all cases previously identified as aggressive (N=59). The classification table for the null presents the initial success rate for this analysis before the regressor is included, and indicates that 100% of the observed Verbal subjects would be co rrectly classified prior to inclusion of the predictor, while none of the Physical cases would be correctly classified. This represents a chance level prediction of the hit rate based on using the largest category to classify all cases. The null model, the refore, estimates an overall correct classification rate of 79.7%. With the HCR predictor included the classification rates are identical. The overall model does not achieve significance, 2 (1, N = 59)= 0.052, p=.820. While the Hosmer and Lesmeshaw model f it index appears adequate, 2 (8, N = 59)=3.690, p=.884, results of the omnibus test indicates that the HCR Total predictor does a poor job classifying
93 subjects. Model A captures a maximum of only 1% of the variation in classification of aggression type. W ith the identified outliers excluded from this analysis the function still fails to achieve statistical significance 2 (1, N = 59)=1.580, p=.209. Taken together, The HCR fails to differentiate between verbal and physical aggressors in the sample. Model B D ue to the limitations imposed by the sample size on the number of permissible predictors in the logistic regression, evaluation of the additive benefits of Model B variables to the HCR only model were once again based on the BPRS total score rather than fa ctor scores, giving us the 2 predictor model originally evaluated in the analysis of aggressive/non aggressive classification. However, while univariate analyses indicated that HCR Total was associated with violence category in the total sample ( r =.218, p< .01, N =102), there was no correlation found in the aggressive only sample, and no association indicated between BPRS Total (or factor scores for that matter) and aggression subtype. It is therefore not surprising that addition of the BPRS symptom variables did not improve classification of aggression subtype, as the overall model again failed to achieve significance, 2 (2, N = 59)=4.531, p=.104, and the HCR term was also not found to be significantly related to the criterion, Walds 2 (1, N = 59)=.441, p=.5 06. However, the BPRS variable came very close to achieving significance in this model, with Walds 2 (1, N = 59)=3.751, p=.053. For this reason it was thought that perhaps the model was being weighted down, so to speak, by inclusion of the HCR term. In order to test this notion and evaluate the full benefit of the BPRS the model was subsequently tested with the BPRS Total score as the sole regression term. Once again, however, the BPRS Total score variable failed to achieve significance in the model, Wald s 2 (1, N = 59)=3.456, p=.063. Taken together BPRS
94 aggregate scores do not prove to be useful in successfully distinguishing verbal from physical aggressors in the full sample of violent subjects. With the previously identified outliers removed from the analysis, the model subsequently demonstrated an ability to distinguish between verbal and physical cases. The overall model was significant 2 (2, N = 59)=10.839, p<.01, and demonstrated good model fit ( 2 Log likelihood=38.536, and Hosmer and Lemeshow Tes t 2 (7, N = 59)=5.297, p=.624. But most notably, the BPRS regressor achieved significance, Walds 2 (1, N = 59)=5.967, p=..05). The model now explained an even greater proportion of the variation in classification; up to 30.0%, which is the most of any mod el thus far. Examination of the coefficients also revealed that the impact of the BPRS variable had increased ( B = 0.125; Exp(B) =.883), but the HCR variable remained non significant ( B =.174, Walds 2 (1, N = 59)=3.410, p=.065. The odds ratio, given by Exp( B) =.910 indicates that for each unit increase in BPRS score the odds of being classified as physically aggressive decreased by a factor of .910 Model T Finally, Model T was evaluated for its ability to successfully distinguish between aggression subtypes. The original 4 predictor Model T successfully contributed to increasing classification rates of aggressive and non aggressive cases. This model included participant age, History, Activation, and Withdrawal as regressors. However, univariate analyses faile d to demonstrate an association between any of these variables and aggression category, and identified only one variable as correlated with aggression group: the BPRS item of Subjective Tension. Therefore, the revised Model T included this variable as the sole predictor in the logistic regression.
95 The overall correct assignment of cases was unchanged, but the revised Model T was found to be significant 2 (1, N =59)=5.585, p=.05, with adequate model fit indicated by the 2 log likelihood (54.013) and Hosmer and Lesmeshaw Test 2 (1, N =59)=8.133, p=.321. Between 9.0% and 14.2% of the variation in classification is explained by this model. The influence of Tension was significant, though barely, 2 (1, N =59)=4.215, p=.040 and its impact on the model given by it s coefficient of 0.691, while the odds ratio is given by the Exp(B) multiplier of 0.501, This indicates that subjects with higher scores on the Tension measure are about half as likely to be classified as Physical. This also indicates that for each 1 point increase on the BPRS Tension item score, the odds of a subject being classified as Physical decreased by half, CI.95=0.259,0.969. Evaluating this effect with the outliers excluded produced omnibus tests showing that the overall model was still significa nt, 2 (1, N =56)=6.416, p=.05). Model fit indices were again adequate, including 2 log likelihood (42.960) and Hosmer and Lesmeshaw Test, 2 (7, N =56)= 7.172, p=.411. But the influence of the Tension regressor, while still significant, Walds 2 (1, N =56) =4.339, p=.05, decreased very slightly to Exp(B )=0.410. Altogether, the direction and nature of the relationships were maintained. Table 4 9 displays the correlations between predictors in each model and aggression subtypes. Table 4 10 summarizes the succe ss rates and predictive efficacy of each model in distinguishing between Verbal and Physical cases. Receiver Operating Characteristics (ROC) Analysis of Violence Subtype Examination of the ROC curve based on the Model T regression for the differentiation o f verbally aggressive from physically aggressive cases (Figure 4 9) indicates an AUC value of .696, p<.05, CI.95=.551,.841. With the outlying cases excluded from the analysis, this value jumps to .738, p<.05, CI.95=.592,.883. Additionally, the form of the ROC curves changes, so that
96 in the low sensitivity range (low false positive rate range) the model with the outliers excluded is superior. Table 411 indicates the relative influence of variables within models in correctly classifying cases Table 4 12 and 413 provide these same results when outlying cases were excluded from the analyses.
97 Table 4 1. Success rates and strength of overall models in classifying cases (N=102) Correct Classification % Omnibus Test Model Overall Non Aggressive Aggressive Chi square df p % Max Explained Variance A 79.7 32.6 59.8 6.841 1 p<.010 8.7 B 78.0 41.9 62.7 9.161 2 p=.010 11.5 T 79.7 48.8 66.7 18.403 4 p=.001 22.2 Note: Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence ri sk assessment scheme total score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale total score (BPRS). Model T : 4 predictor model comp rised of historical subscale score of the Historical/Clinical/Risk Management violence risk assessment scheme (History), participant age, activation cluster symptoms of the Brief Psychiatric Rating Scale (Activation); and Emotional Withdrawal item score of the Brief Psychiatric Rating Scale (Withdrawal). Table 4 2. Correlations, odds ratios, and strength of individual predictors in each model in correctly classifying aggressive cases (N=102) Spearman's Wald's Odds Ratio Model Variables r p Chi square df p Exp(B) 95% CI A HCR 0.255 .01 6.277 1 .010 1.116 1.024 1.217 B HCR 0.255 .010 5.054 1 .030 1.106 1.013 1.207 BPRS ns ns 2.212 1 .030 1.029 0.991 1.069 T History 0.196 .050 3.186 1 ns 1.117 0.989 1.264 Age 0.259 .010 8.086 1 .000 0.945 0.91 0.098 Activation 0.240 .020 5.987 1 .010 1.191 1.035 1.369 Withdrawal 0.327 .000 0.73 1 ns 0.852 0.591 1.200 Note: BPRS = Brief Psychiatric Rating Scale; AUC = area under the receiver operating characteristic cu rve; HCR=Historical/Clinical/Risk Management violence risk assessment scheme total score; History= historical subscale score of the Historical/Clinical/Risk Management violence risk assessment scheme; Age = participant age; Activation=activation cluster sy mptoms of the Brief Psychiatric Rating Scale; Withdrawal = Emotional Withdrawal item of the Brief Psychiatric Rating Scale.
98 Table 4 3. Success rates and strength of overall models in classifying cases with outliers excluded (N=101) Correct Classificatio n % Omnibus Test Model Overall Non Aggressive Aggressive Chi square df p % Max Explained Variance A 79.7 33.3 60.4 7.849 1 .01 10.1 B 62.4 41.9 77.6 9.000 2 .01 11.5 T 79.7 52.4 68.3 21.532 4 .000 25.8 Note: Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale tota l score (BPRS). Model T : 4 predictor model comprised of historical subscale score of the Historical/Clinical/Risk Management violence risk assessment scheme (History), participant age, activation cluster symptoms of the Brief Psychiatric Rating Scale (Act ivation); and Emotional Withdrawal item score of the Brief Psychiatric Rating Scale (Withdrawal). Table 4 4. Correlations, odds ratios, and strength of individual predictors in each model in correctly classifying aggressive cases with outliers exclude d (N=101) Spearman's Wald's Odds Ratio Model Variables r p Chi square df p Exp(B) 95% CI A HCR 0.254 .01 7.090 1 .008 1.128 1.032 1.232 B HCR 0.254 .01 4.957 1 .026 1.140 1.012 1.205 BPRS 0.188 ns 2.191 1 ns 1.029 0.991 1.069 T History 0.19 ns 4.138 1 .042 1.140 1.005 1.293 Age 0.273 .006 9.600 1 .002 0.938 0.900 0.977 Activation 0.223 .025 6.725 1 .01 1.213 1.048 1.404 Withdrawal 0.012 ns 1.386 1 ns 0.795 0.542 1.165 Note: BPRS = Brief Psychiatric Rating Scale; AUC = area under the receiver operating characteristic curve; HCR=Historical/Clinical/Risk Management violence risk assessment scheme total score; History= historical subscale score of the Historical/Clinical/Risk Management violence risk assessmen t scheme; Age = participant age; Activation=activation cluster symptoms of the Brief Psychiatric Rating Scale; Withdrawal = Emotional Withdrawal item of the Brief Psychiatric Rating Scale.
99 Table 4 5. Areas Under the Receiver Operating Characteristic Curve (AUCs) for each model in classifying aggressive and non aggressive cases (N=102) Area Under the Curve (AUC) Model Non Aggressive Aggressive p A 0.351 0.649 .011 B 0.734 0.266 .000 T 0.688 0.312 .001 Note. Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale total score (BPRS). Model T : 4 predictor model comprised of historical subscale score of the Historical/Clinical/Risk Management violence risk assessment scheme (History), participant age, activation cluster symptoms of the Brief Psychiatric Rating Scale (Acti vation); and Emotional Withdrawal item score of the Brief Psychiatric Rating Scale (Withdrawal). Table 4 6. Areas Under the Receiver Operating Characteristic Curve (AUCs) for each model in classifying aggressive and non aggressive cases with outlie r excluded (N=101) Area Under the Curve (AUC) Model Non Aggressive Aggressive p A .353 .647 .012 B .717 .283 .000 T .684 .316 .002 Note. Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme to tal score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale total score (BPRS). Model T : 4 predictor model comprised of historical sub scale score of the Historical/Clinical/Risk Management violence risk assessment scheme (History), participant age, activation cluster symptoms of the Brief Psychiatric Rating Scale (Activation); and Emotional Withdrawal item score of the Brief Psychiatric Rating Scale (Withdrawal).
100 Table 4 7. Model B descriptive characteristics Variable Mean Std. Deviation N OAS 15.450 29.673 102 HCR 18.700 4.997 102 Affect 3.458 2.366 102 Positive 7.100 4.437 102 Resistance 4.179 3.287 102 Activation 4.240 3.735 102 Note: OAS=OAS total score; HCR=HCR total score; Affect=BPRS affect cluster symptoms; Positive=BPRS positive symptoms cluster score; Resistance= BPRS Resistance cluster symptoms; Activation=BPRS activation cluster symptoms.
101 Table 4 8. Summary of multiple regression analysis for variables predicting aggression severity (N = 102) Variable Model A Model B Model T B SEB B SEB B SEB HCR 1.893 0.563 0.319*** 1.631 0.572 0.275* BPRS Affect 0.113 1.207 0.009 n/s Positive 0.449 0.912 0.067 n/s Resistance 0.218 1.256 0.024 n/s Activation 1.842 1.136 0.2 32 n/s 2.407 0.746 0.303** Age 0.216 0.226 0.091 n/s History 1.827 0.761 0.225* Withdrawal 0.173 2.247 0.007 n/s R2 0.102 0.169 0.158 Adjusted R 2 0.093 0.126 0.123 F Chan ge 11.308*** 3.908** 4.543** Effect Size f 2 0.1026 0.112 0.140 Note. *p < .05 **p < .01 ***p<.001 Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HC R). Model B : 5 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), Brief Psychiatric Rating Scale total score (BPRS), BPRS Affect Cluster score (Affect), BPRS Positive Symptoms Cluster sc ore (Positive Sx), BPRS Resistance Cluster score (Resistance), and BPRS Activation Cluster score (Activation). Model T : 4 predictor model comprised of historical subscale score of the Historical/Clinical/Risk Management violence risk assessment sche me (History), participant age, activation cluster symptoms of the Brief Psychiatric Rating Scale (Activation); and Emotional Withdrawal item score of the Brief Psychiatric Rating Scale (Withdrawal)
102 Table 4 9. Correlations between predictors and aggression subtypes (N=59) Spearman's Correlations Model Variables Verbal Physical r p r p A HCR 0.212 0.033 0.102 ns B HCR 0.212 0.033 0.102 ns BPRS 0.262 0.008 0.135 ns T Tension 0.275 0.035 0.275 0.035 Note: HCR=HCR total score; BPRS=BPR S total score; Tension=BPRS Subjective tension item score. Table 4 10. Success rates and strength of overall models in distinguishing between verbally and physically aggressive cases (N=59) Model Correct Classification % Omnibus Test Overall Verbal Physical Chi square df p % Max Explained Variance A 79.7 100 0 1.58 1 ns 1 B 79.7 100 0 4.531 2 ns 11.6 T 79.7 100 0 5.585 1 0.018 14.2 Note. Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment sc heme total score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale total score (BPRS). Model T : 1 predictor model comprised of Tension item score of the Brief Psychiatric Rating Scale (Tension).
103 Table 4 11. Correlations, odds ratios, and strength of individual predictors in each model to distinguish between verbally and physically aggressive cases (N=59) Model Wald's Odd Ratio Variables Chi square df p Exp(B) 95% CI A HCR 0.052 1 ns 1.014 0.897 1.147 B HCR 0.441 1 ns 1.045 0.918 1.190 BPRS 3.751 1 ns* 0.938 0.878 1.001 T Tension 4.215 1 0.04 0.501 0.259 0.969 Note. Model A : 1 predictor model compris ed of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale t otal score (BPRS). Model T : 1 predictor model comprised of Tension item score of the Brief Psychiatric Rating Scale (Tension). Table 4 12. Success rates and strength of overall models in distinguishing between verbally and physically aggressive cases outliers excluded from the analysis (N=56) Model Correct Classification % Omnibus Test Overall Verbal Physical Chi square df p % Max Explained Variance A 83.9 100 0 1.58 1 ns 4.7 B 83.9 97.9 11.1 10.839 2 0.004 30 T 83.9 100 0 6.416 1 0.011 18.5 Note. Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR). Model B : 2 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale total score (BPRS). Model T : 1 predictor model comprised of Tension item score of the Brief Psychiatric Rating Scale (Tension).
104 Table 4 13. Correlations, odds ratios, and strength of individual predict ors in each model to distinguish between verbally and physically aggressive cases outliers excluded from the analysis (N=56) Model Wald's Odd Ratio Variables Chi square df p Exp(B) 95% CI A HCR 1.466 1 ns 1.098 0.944 1.279 B HCR 3.41 1 ns 1.19 0.989 1.431 BPRS 5.967 1 0.015 0.883 0.799 0.976 T Tension 4.339 1 0.037 0.41 0.177 0.949 Note. Model A : 1 predictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR). Model B : 2 p redictor model comprised of Historical/Clinical/Risk Management violence risk assessment scheme total score (HCR), and Brief Psychiatric Rating Scale total score (BPRS). Model T : 1 predictor model comprised of Tension item score of the Brief Psychiatric Rating Scale (Tension).
105 Figure 41. Classplot of correct and incorrect predictions of aggression under Model T Observed Groups and Predicted Probabilities 8 + + | | | | F | | R 6 + Y + E | Y | Q | Y Y | U | Y Y | E 4 + Y Y Y Y Y + N | Y Y Y Y Y | C | Y Y Y N YY Y YYY Y Y | Y | Y Y Y N YY Y YYY Y Y | 2 + NN Y N Y N N NYY YY YNNYY Y Y Y Y YY + | NN Y N Y N N NYY YY YNNYY Y Y Y Y YY | | N Y NN N NNNNYNNNNY NNY NN YN NYNNYN NY NNN NYNYYNNNNYY YYN Y YYYNY Y YN Y Y Y | | N Y NN N NNNNYNNNNY NNY NN YN NYNNYN NY NNN NYNYYNNNNYY YYN Y YYYNY Y YN Y Y Y | Predicted --------+ --------+ --------+ --------+ --------+ --------+ --------+ -------+ --------+ ---------Prob: 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 Group: NNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYY Note: Predicted probability is of membership for Yes Aggressive group. The Cut Value=.50 Symbols: N=No Aggression; Y=Yes Aggression. Each Symbol Represents .5 Cases. Figure 42. Receiver Operating Characteristics (ROC) curve for Model A (HCR total score o nly) and classification of aggressive cases Note: Sensitivity= true positive rate. 1 Specificity=false positive rate
106 Figure 43. Receiver Operating Characteristics (ROC) curve: Model A (HCR total score only) and classification of non aggressive cases Note: Sensitivity= true positive rate. 1 Specificity=false positive rate Figure 44. Receiver Operating Characteristics (ROC) curve: Model B (HCR plus BPRS) and classification of aggressive cases Note: Sensitivity= true positive rate. 1 Specif icity=false positive rate
107 Figure 45. Receiver Operating Characteristics (ROC) curve: Model B (HCR plus BPRS) and classification of non aggressive cases Note: Sensitivity= true positive rate. 1 Specificity=false positive rate Figure 46. Receiver Operating Characteristics (ROC) curve: Model T (age, history, activation, emotional withdrawal) and classification of aggressive cases Note: Sensitivity= true positive rate. 1 Specificity=false positive rate.
108 Figure 47. Receiver Operating Characteri stics (ROC) curve: Model T (age, history, activation, emotional withdrawal) and classification of non aggressive cases. Note: Sensitivity= true positive rate. 1 Specificity=false positive rate Figure 48. Receiver Operating Characteristics (ROC) cu rve: Models A, B, and T, and classification of non aggressive cases. Note: Sensitivity= true positive rate. 1 Specificity=false positive rate
109 Figure 49. Receiver Operating Characteristics (ROC) curve: Model T (BPRS tension item) and classification of physically aggressive cases (N=59) Note: Sensitivity= true positive rate. 1-Specificity=false positive rate. Figure 410. Receiver Operating Characteristics (ROC) curve: Model T (BPRS tension item) and classification of physically aggressive cas es With Outliers Excluded (N=56) Note: Sensitivity= true positive rate. 1 Specificity=false positive rate.
110 CHAPTER 5 DISCUSSION Aim 1 Predicting Aggression vs. Non A ggression A logit model was used to predict classification of subjects as aggressive or non aggressive. Age and Activation symptoms were shown to be statistically significant predictors of aggression, but aggression was significantly related to Historical subscale score only in the analysis that was conducted with the outlier excluded. F or every one unit increase in Activation score the odds of classification as aggressive (versus non aggressive) increased by a factor of 1.178, while for each 1 year increase in participant age the odds decreased by 0.94. The experimental function provided a maximum correct classification of cases of 25.8%, based largely on its greater ability to identify non aggressive subjects rather than aggressive subjects. Though this represents a fairly small proportion of cases, this is more than a two fold increase in the least successful model based on the traditional measure of violence risk, the HCR Total score. Direct comparison of the odds ratios obtained in the present study with those found in others using similar predictors and comparable forms of the OAS ins trument to measure aggression, suggest that the sample is not unique. Watts, Leese, Thomas, Atakan, and Wykes (2003) reported odds ratios of 1.120 and 1.800 for their modified OAS total score and BPRS symptoms cluster score (hostile suspiciousness), respec tively, for actual assault in a sample of patients in a locked psychiatric unit. They also reported odds ratios of 1.08, 0.60, and 1.99 on their modified OAS measure, anxiety depression BPRS symptom cluster score, and withdrawal retardation BPRS symptom cl uster score, respectively, and aggressive behavior. And Amore et al. (2008) reported odds ratios for BPRS subscales and symptoms clusters and a number of risk factors for physically aggressive behavior during psychiatric hospitalization within the same
111 ran ge: hostility suspiciousness (1.14), anxiety depression (0.95), thought disturbance (1.07), activation (1.14), and anergia (1.04). By comparison, the best fitting model for correctly classifying aggressive cases achieved odds ratios of 1.19 and 1.12 for B PRS Activation cluster symptom score and HCR Historical subscale score, respectively (and 1.21 and 1.14 when the effects of outliers were excluded from the analysis). According to Fleiss, Williams, and Dubro (1986) odds ratios above 2.50 should be conside red clinically important. But studies of violence prediction do not typically report such values. Additionally, the present study serves to address the paucity of reliable research addressing models of aggression in a population of hospitalized mentally il l criminal offenders. It should be noted that the traditional HCR variables were consistently strong predictors across each model for correctly classifying aggressive cases, which further legitimizes the continued use of the HCR 20 as one of the standard m easures of violence risk assessment, particularly when used in isolation. Initially, the ROC curves based on Model T seemed paradoxical. For the purpose of comparison, a subsequent logistic regression analysis was conducted with the 5 BPRS factor score as regressors, an analysis of dubious statistical validity given the small sample and number of predictors in the regression. Not surprisingly the function was not found to be significant and therefore no discernable contribution of individual variables coul d be identified. However, obtained residuals were used to construct ROC curves evaluating the predictive reliability of BPRS factor scores when used alone. These aggregated clinical symptom measures appeared to provide a fair method of predicting violent c ases (AUC = 0.668, p<.01), and was comparable to the HCR when used alone. And it proved to be similarly unreliable in discerning non violent cases (AUC=0.332, p<.01) These explorations of the data suggest that these clinical factor
112 clusters could prove t o be comparable to the HCR total score in correctly assigning violent cases, and that further investigation utilizing larger samples is indicated. Another subsequent post hoc exploration of the data revealed a comparable AUC value to the best model (Model A) when the BPRS Activation cluster (the most salient variable in the regression analysis) is considered alone (AUC=.645, p<.01). But this is not the case when the age variable was considered in isolation (AUC=.349, p<.01). It is striking that the age only predictive value is comparable to the combined function. Therefore, age appears to reduce the predictive utility of the model in predicting aggressive cases, but not in predicting non aggressive cases. In their evaluation of the predictive validity of sev eral risk assessment measures with male offenders with intellectual disabilities, Lindsay et al. (2008) demonstrated the utility of HCR subscale scores. They reported AUCs of .600 and .720 for HCR Historical and Total scores, respectively. A study by Banks et al. (2004) based on actuarial models of prediction achieved an AUC of .738 in predicting violent incidents in a mainstream offender sample. While Harris and Rice (2003) reported predictive accuracies as high as AUC=.900. Taken together the use of struc tured combinations of risk variables has been found to produce superior predictive models of aggression. In the present study the HCR only model produced a total sample AUC value of .649, which falls within the low range of values reported in the literatur e for the predictive abilities of this measure. This is likely due to the truncated form employed in this study, in which the 5 Risk Management items were not scored, thereby reducing the ability of the HCR measure to do what it reportedly does best: disti nguish violent from non violent cases. By comparison, the models that included BPRS symptom and demographic variables produced AUC values that fell far below even chance levels. Ironically, they produced superior classification of nonaggressive
113 cases. Mod el B, which examined the simple additive effect of the BPRS on the HCR model, produced and AUC of .734 ( p<.001) while the more selective Test Model produced an AUC of .688 ( p=.001) for the prediction of non aggressive cases. By comparison, the identificati on of these cases by the HCR fell far below chance levels (AUC=.351, p=.011). The prediction of non aggressive cases, it should be noted, has clinical utility and is not well addressed by traditional measures. This appears to support the authors initial hypothesis that supplementing commonly used risk assessment measures with those assessing clinical symptom variables has practical benefit in providing a complete predictive algorithm for violent and non violent behavior. But clearly these variables tell a different story depending on how they are combined. It is tempting to examine AUC values for all of the variables and draw causative conclusions based on these regarding their relationship with the yes/no prediction of aggression. The examination of ROC c urves is useful in refining the statistical inferences one makes based on the logistic regression models, but direct comparisons of multi variable effects should be made with extreme caution. However, it would seem that the correct classification of violen ce potential is aided by the consideration of symptom variables and their severity, particularly those related to general agitation. Also, throughout this analysis the effect of age appears in the anticipated direction, with younger age being related to aggression. A few variables appear to suffice towards increasing the accuracy of the prediction of aggressive behavior, supporting a less is more strategy to violence prediction. However, diagnostic tests of the strength and predictive benefit of the Test Mo del suggest that its true contribution is in its ability to help in the prediction of non violent rather than violent cases. As such, it demonstrates utility as a potential adjunct to more traditional risk assessment measures. The use of one such establish ed measure, the HCR, is further substantiated in the present study.
114 Aim 2 Predicting Aggression Severity Taken together, using the enter method, a significant model emerged with Activation (a BPRS factor cluster that includes the diagnostic items of subjective tension, mannerisms and posturing, excitement, and disorientation) and History (the HCR Historical subscale that captures 10 static variables related to a persons background) found to be predictive of aggression severity. Other variables in the mod el were not significant predictors. It is noteworthy that Activation was also found to be the sole statistically significant clinical predictor in the logit model used to predict classification of subjects as aggressive or non aggressive. Also, Model T was found to be superior to models based solely on the traditional risk assessment device (HCR) and to a model based on the simple additive influence of the HCR with a measure of aggregated symptom variables (BPRS Total score). Thus the selective addition of clinical symptom variables, as captured by the BPRS overall score (and by the BPRS factor scores to a lesser degree) is substantiated in the sample and provides an algorithm for predicting aggression intensity from this combination of risk assessment items and symptoms of mental illness. This algorithm is also superior to those offered by models in which these instruments are used in isolation. The research on violence prediction has typically concerned itself with yes/no classification of violence risk pot ential, and rarely on characterizing aggression intensity should violence occur. One exception includes a study by Nicholls, Brinka, Greaves, Lussier, and Verdun Jones, (2008) who used the OAS to explore differences between male and female forensic psychia tric inpatients in order to identify the mental health, psycho social, and criminogenic contexts in which aggression occurs. Many studies exploring the severity of aggression are primarily concerned with identifying gender differences in their samples. Ste inert, Wiebe, and Gebhardt (1999) conducted a similar evaluation of aggressive behavior against self
115 and others among newly admitted schizophrenic patients, employing a stepwise multiple regression to identify predictors of the number and duration of re ho spitalization. The present study appears to address a marked scarcity of research regarding the relationship between aggression severity, risk assessment, and clinical variables not only in forensic patients, but psychiatric patients in general. Aim 3 P redicting Aggression Category Identification of verbal aggressive and physical aggressive subtypes represented a simple extension of analyses distinguishing between aggressive and non aggressive groups, with the dichotomy in this case characterizing the na ture of aggression rather than its presence. Taken together, logistic regression analysis performed with aggression category as the criterion variable and a score on a measure of subjective tension as the predictor significantly distinguished between verba l aggression and physical aggression subtypes. While univariate analyses were expected to reveal significant intercorrelations for each type of aggressive behavior and ratings of risk and symptoms severity, no other demographic, clinical, or historical var iable was found to even be associated with aggression grouping. The degree of this models predictive utility in correctly classifying aggression type is at least comparable to measures used to initially distinguish aggressive and non aggressive cases and is, in fact, superior when extreme cases are excluded from the analysis. An older study by Kay, Wolkenfeld, and Murrill (1988) used a modified version of the OAS to characterize the nature and prevalence of aggression among psychiatric patients, and found that this instrument successfully discriminated between cases of autoaggression, physical assault, and general aggression, and could serve to clarify the interaction of demographic, clinical, affective and behavioral characteristics. Gray et al. (2003) co nducted prospective evaluation of the utility of various measures in predicting aggression subtypes. They explored the
116 benefits of the HCR 20, Psychopathy Checklist (PCL), Beck Hopelessness Scale (BHS), BPRS, and age at first psychiatric admission to distinguish between verbal aggression, violence to property, and physical violence. They found that the HCR Clinical and Historical subscales were good predictors of these behaviors, though the combined benefit of the HCR 20 total score was superior to the use of either subscale alone. They reported correlations between .53 and .56 between subscale scores and these behaviors, AUCs between .79 and .83, and odds ratios between 2.55 and 8.85. By comparison, bivariate correlations between predictors in each of the m odels and aggression subtypes were very small in the present sample, and ranged from .102 to .275. Similarly, while Gray and colleagues identified the BPRS as a good predictor of verbal and physical aggression, with correlations between .58, AUCs between .81.84, and odds ratios between 2.314.00, only the BPRS Tension variable even achieved statistical significance in the current sample. In comparison, the models utilized in the present study are less effective in classifying aggression subtypes. It shou ld be noted, however, that studies employing retrospective analyses, as in the present study, have tended not to find all HCR 20 scales to be such good predictors of all forms of aggression (Douglass et al., 1999). This may be due to their dependence on ch art notes rather than direct observation for obtaining relevant information regarding target behaviors. Additional Findings The present study failed to replicate meaningful associations between violence type or severity and age, education, history of head trauma (with loss of consciousness), history of abuse (physical, emotional, sexual, or combat related), or primary psychiatric diagnosis, which have been reported elsewhere. Moreover, the dimensions of psychopathology proposed in the present study appear t o be mostly non contributory. Nonetheless, associations were found between at
117 least one symptom cluster that characterizes participants subjective impression of tension, excitement, and disorientation, and the frequency and severity with which they aggres s. Examining the relationship between both independent measures and other demographic variables replicated some interesting findings, such as the well known relationship between youth and increased impulsivity. Another finding of interest in the data is o f the association between head injury and affective symptoms, particularly somatization. Studies have supported a link between a history of head injury and violent behavior, even many years following injury (Kim, 2002). While the present study failed to re plicate this result, head injury was found to be associated with the BPRS clinical variables of Somatic Concern and Subjective Anxiety ( r =.243, p<.05 and r =.205, p<.05, respectively). It is possible that the experience of multiple head traumas increases an individuals subjective concern for their own medical wellbeing. Alternatively, these associations may be related to organic trauma related personality changes. This study also confirmed the absence of significant redundancy between the HCR and BPRS clini cal items. This may explain why, according to the results, they characterize opposite behavioral states equally well, with the BPRS essentially predicting nonviolence, and the HCR doing what it does best, correctly assigning violent cases. To some degree, they can be interpreted as complimentary measures of self control and adaptation of subjects to their environment, which are clinically relevant in their own rights and could provide useful information to clinicians when used in tandem. Methodological Con siderations The choice of L ogistic coefficients was based on both conceptual and statistical considerations. The identification of significant bivariate correlations guided much of the preliminary consideration of relevant variables. Regression functions c an be found to be non significant, as in the case of Model T, when their corresponding correlations are found to be
118 significant. This is because logistic coefficients represent only partial coefficients controlling for other variables in the model, while c orrelation coefficients are uncontrolled. Also logistic coefficients represent both linear and nonlinear relationships, while correlations reflect only linear relationships. And finally, significant parameter estimates indicated that there was a relationsh ip between the predictors and criterion variables, though these relationships may not have applied overall. However, in order to make overall statements about the significance of the independent variables both the parameter estimates and correlations need to be statistically significant. For this reason the identification of salient variables proceeded from not only conceptual and methodological priorities but from simple univariate analyses. A major concern of the author was the operational definition of aggression as an all or none construct where, for the purposes of group membership, a single event characterized an individual as aggressive. This rarely happened however, as aggressive behaviors were seldom isolated and typically aligned under dominant subtypes that lent themselves easily to one grouping. Other studies have employed median cutoff scores to characterize violent behavior (Gray et al., 2003), rather than the single event classification scheme employed here. But the limited sample and relat ively low base rate of aggressive behavior in the sample precluded use of this method. Additionally, as the intent of the study was not to differentiate between severe aggression and average or mild aggression, the use of the single event scheme was subs tantiated. No doubt, however, it provided very different information regarding the sample than other predictive studies employing the alternate technique. Gray and colleagues used a median split cut off point to differentiate cases scoring higher and lower on measures of their target behavior, and then employed a Mann Whiteny U comparison of means for those scoring above and below this median value for each of their predictors. Due to limited sample size, the
119 present study approached the research question from the opposite end by dividing subjects into aggressive and non aggressive cases and then comparing these groups on their scores of the predictor. This technique has been widely used (Belfrage et al., 2000) as it lends itself to the type of retrospectiv e analysis conducted in the present study. An additional methodological concern was that while the overall scores on each measure were treated at ordinal or interval level (and provided the means to operationalize target behaviors of interest) the items on which they are based were sometimes categorical. However, according to Wagner, Beiden, and Metz (2001) ROC curves are not affected by the use of a Likert type and other scales that use discrete multi point systems of the type employed in both of the risk and clinical measures (HCR and BPRS). According to them, ROC curves built on discrete rating and continuous scales are not typically different, particularly when resulting operating points are well distributed. A number of individual scale items (9 from the BPRS and 2 from the HCR) were found to be correlated with the dependent measure to some degree. But statistical limitations precluded evaluation of the predictive contribution of these underlying clinical substrates. Therefore they were aggregated into their superordinate categories for computational reasons. Sample Size A major limitation to achieving multivariate significance is sample size. Reliability of estimates decline heavily when there are few cases for each observed combination of independent variables. In extreme cases, if there are too few cases compared to the number of variables, it becomes increasingly difficult to converge on a solution. Peduzzi et al. (1996) recommend that the smaller of the classes of the dependent variable have at leas t 10 events per parameter in the model. While the study analyses proceeded in accordance with these criteria,
120 these essentially represent a minimum requirement. Salient relationships may have been missed given the size and nature of the sample. Due to the stringent statistical limitations imposed by the methodologies employed, it was necessary to combine some categories of both the independent and dependent variables not only to make the analysis simpler, but because the number of cases in some categories w ere too small to conduct certain tests. As the number of categories increased more parameters needed to be estimated, and it became increasingly difficult to obtain significant results. For this reason, for instance, aggression type was dichotomized. But this resulted in a fundamental alteration of the research question. Similarly, it would have been preferable to model aggressive cases separately, once identified, in the assessment of aggression severity. But this would have reduced the sample size to whe re only 1 predictor could have been employed in the multiple regression analysis. Additional univariate analyses could have been used to further explore the relationships between aggression severity and the BPRS factor scores, but the intent of this study was to build models of prediction based on multiple variables, which required multivariate techniques, but also larger sample sizes.. Unrecorded Aggression An examination of what constituted aggressive behavior begs the question of what overt behaviors did not meet this definition while still representing deviant or unwanted behaviors in this population. For instance, it was common to find case notes in subjects charts referring to disregarding rules, non compliance, resisting instructions, being demand ing, requiring multiple prompts by staff, manipulativeness, and refusing participation in therapeutic activities. These appeared to constitute a passive aggressive subtype that, although not captured by the OAS instrument, could offer therapeutically use ful information and help characterize yet another model of aggressive behavior. In their study of the changes in aggressive behavior throughout
121 the course of dementia, Keene et al. (1999) referred to aggressive resistance as, Resisting anyone's attempt to help or being uncooperative in any situation (p 543). Though their definition captures resistance in an entirely different psychiatric population (and includes resistance by patients during intimate care), this concept seems applicable to this sample as well, as it describes a form of deliberate though non violent behavior that straddles the line between defiance and active aggression. Resistance and uncooperativeness in forensic psychiatric patients has been widely explored in the context of malingering research. Merten, Friedel, and Stevens (2006) explored the prevalence and influence of uncooperativeness in producing invalid forensic neuropsychological test profiles due to negative response bias in patients. But there is a dearth of research on this phenomenon as it relates to the behavioral management of inpatient criminal offenders. The literature has more often addressed this issue in terms of treatment non compliance among community dwelling offenders and practices of outpatient forensic mental he alth treatment (Lamb, Weinberger, & Gross, 1999). This would appear to represent an opportunity for future research. Similarly, review of patient charts often revealed references to a host of sexually inappropriate behaviors including flirtatious behavior towards female staff, making sexual comments, masturbating in public, and preoccupation with inappropriate sexual themes. While not coded as aggressive according to the norms of the OAS instrument, sexual aggression has been linked to a host of behavioral disturbances in psychiatric patients. The issue of psychiatric patients who sexualize their encounters with treatment staff (particularly female nurses) is specifically addressed by Massoc (2003) and Higgins, Barker, and Begley (2009). They describe unwant ed sexual attention and behaviors from patients as a form of sexual aggression and
122 boundary violation similar to physical violence. Here as well, research has rarely addressed these behaviors within the context of an inpatient psychiatric treatment facility for criminal offenders. This phenomenon also warrants further consideration. The reality of the coding scheme necessitated some interpretation of intent. For example, the measure did not differentiate between the instigators and victims of aggression, f or instance, separating those who might be defending themselves against unprovoked attacks by becoming violent themselves. Also, a behavior such as making a mess is a codable form of aggression against objects according to the OAS instrument. In one ca se a participant was noted to have clogged his toilet and flooded his room. But whether this constituted aggression or accident was subject to some interpretation, as was the evaluation of treatment notes with vague references to residents as argumentative and inappropriate without mention of specific behaviors. Strengths of the Study Design Despite the low base rate of severe violence in this setting, because the OAS captures a range of types and severity of aggressive behaviors many residents were fo und to engage in some of these acts. Common aggressive behaviors included making threats of violence, throwing or kicking objects, head banging, hitting, pushing, and kicking others. While many studies have examined aggression and violence within in patien t psychiatric and criminal justice settings, and have noted contributing factors as diverse as patient perception of staff behaviors (interpersonal conflict, limitsetting by staff) and psychotic symptoms and tension (Nolan, Shope, Citrome, and Volavka, 2009), it is reasonable to question whether the highly structured and controlled nature of this environment, as a maximum security forensics unit, may ultimately have limited the full expression of aggression in this population. The modified OAS coding schem e, which captured not only the severity of behavior but the intensity of responses (interventions) to these behaviors, was an attempt to account for this possible
123 limitation and preserve the potential for results to have practical application in the envi ronment from which these data were derived. The effects of various interventions were expected play a role in modifying the predictive value of identified variables and factors. But while the coding scheme incorporated a unique violence rating system that captured the force of direct clinical intervention in this population, a subsequent examination of the bivariate correlations reveals that this addition did little to affect the strength of the associations that were found, either negatively or positively. The correlation coefficients observed were almost identical in each relationship found when this weighted measure was removed. The retrospective nature of this study, one in which aggressive behaviors were evaluated for the first 5 weeks following a resid ents admission to the treatment facility, utilizing various archival sources, may be considered a weakness of this study by some. But the design benefited largely from well defined and standardized instruments in the scoring and evaluation of aggressive b ehaviors in this population. The traditional reliance on subjects autobiographical verbal report was strengthened through review of various collateral and historical sources, and extensive review of comprehensive chart notes from each residents treatment team. Previous studies have suffered from imprecise definitions of aggression and violence, use of non standardized scales in the evaluation of these behaviors, and poorly defined assessment periods. Some studies have favored prospective designs, arguing that only these can fully capture aggression as it occurs in a naturalistic fashion. But these methods fail to address the practical needs of those conducting research and working within in patient mental health and residential environments. Chief among th ese is the impracticality of study designs requiring 24 hour 1 to 1 observation of all participants. Such studies not only lack workability but have limited generalizability.
124 This study attempted to identify dynamic variables that are associated with aggr essive and violent behavior in a residential forensic psychiatric setting. A focus on standardized clinical variables, and psychiatric symptom profiles in particular, represented a departure from early investigations of the prediction of violence based on clinical judgment alone, which rarely produced better than chance results. It was also an attempt to supplement later investigations that relied on general historical, demographic, and static risk factors for violence. As previously enumerated, many studi es have revealed at least some association between violence and a number of these variables (such as age, mental health status, and drug abuse history) while equal numbers of investigations have failed to reveal significant relationships. The majority of t hese relationships were not supported in the present study sample.
125 CHAPTER 6 SUMMARY AND CONCLUSI ONS Doctrine holds that past violent behavior is the best predictor of future violence. Indeed, research shows that many measures of prior violence, including self report, arrest records, and hospital records, are strongly related to future violence, and that this relationship holds in the case of institutional violence (Sirotich, 2008) A number of clinical variables, particularly variables rel ated to psychopathology, were expected to emerge as significant predictors of aggressive behavior in this population. It was also expected that some of these might mediate the predictive utility of purely historical factors. The aim of the present study wa s to determine violence risk factors and ascertain the level of predictive utility of these factors in this setting. It was thought that the addition of these variables could increase the accuracy of violence predictions when based on systematic review and a comprehensive inventory of symptom variables provided by the BPRS, and their selective application within models of prediction. It was hypothesized that this would aid in the identification of individuals likely to commit aggressive acts during their tr eatment, could provide an algorithm for expectancies regarding the intensity of aggression when it occurs, and help predict what form this violence might take. Iterations of each model were compared against a form of what has become the standard risk assessment instrument in the field, the HCR, in order to assess the benefits of including a wider array of symptom items than is provided by this brief measure. Aggressive acts were coded retrospectively during thorough review of ward charts, and according to s tandardized criteria of a modified version of the Overt Aggression Scale in order to capture a full array of both verbally and physically aggressive acts. Taken together, the addition of these symptom variables produced models with a small added benefit in their ability to correctly classify and characterize aggressive behavior. And for those who did aggress, these models did a meager job of predicting
126 the form this violence might take. However, in each case, these models proved better than those based on b usiness as usual approaches. Ultimately, violence risk assessment was augmented, however slightly, by a combination of clinical and non clinical predictors. At the same time, this study illustrated the importance of considering more than just static histor ical variables in the evaluation of risk. So while studies have frequently demonstrated (and is has become a veritable mantra in the social sciences) that past behavior is the best predictor of future behavior, despite its correlation with a number of meas ures this single variable (History) did not contribute significantly to any of the models evaluated, except in a single case when extreme scores were excluded from the analysis. Sometimes less is more. The addition of a more comprehensive assessment of sym ptoms of mental illness did not ultimately increase the utility of the more established and parsimonious risk assessment device in identifying aggressive cases. However, the author established the complimentary utility of a test model by demonstrating its predictive strength in correctly identifying exactly those types of cases the traditional method fails to uncover. And a subset of variables was successfully identified, including age, tension, and agitation that proved useful in predicting frequency and s everity of aggression in this population, though they typically added minimally to successfully predicting the form this aggressive behavior is likely to take. Even more precise approaches to the prediction of aggressive behavior may lead to the identific ation of additional subsets of violence and enable more effective and specific treatment outcomes. The classification of violent subjects continues to be a priority in many settings. There remain significant social and financial implications to misclassif ying non violent subjects as violent, in terms of restriction of freedoms, unnecessary psychopharmacological intervention, and stigmatization. The authors hypothesized model, which takes into account a wider array of
127 symptom variables associated with ment al illness, is demonstrated as superior in correctly classifying this group; that is, in identifying non aggressive cases. This has clinical utility in a population in which the modal behavior is violence. As a matter of public policy based increasingly on the application of least restrictive measures in managing mentally ill criminal offenders, the proposed model appears tailor made to the task of tackling modifiable risk factors of aggression, and serving as a compliment to the mission of identifying individuals who might benefit from interventions for persistently violent mentally ill men within the forensic mental health system of Florida.
128 APPENDIX A RATING SCALES A 1. Modified version of the Overt Aggression Scale. AGGRESSIVE BEHAVIOR (code most severe ) Subject # Category Behaviors Item Score Multiplier Total Severity Verbal Aggression Makes loud noises, shouts angrily 1 x1 Yells mild insults (E.g. "You're stupid!) 2 x1 Curses viciously, uses foul language in anger, makes moderate th reats to others or self 3 x1 Makes clear threats of violence toward others or self ("e.g. Im going to kill you"), or requests to help to control self 4 x1 Physical Aggression Against Objects Slams door, scatter clothing, makes a mess. 1 x2 Th rows objects down, kicks furniture without breaking it, marks the wall. 2 x2 Breaks objects, smashes windows 3 x2 Sets fires, throws objects dangerously 4 x2 Physical Aggression Against Self Picks or scratches skin, hits self, pulls hair (with no or minor injury only) 1 x3 Bangs head, hits fist into objects, throws self onto floor or into objects (hurts self without serious injury) 2 x3 Small cuts or bruise, minor burns. 3 x3 Mutilates self, makes deep cuts, bites that bleed, int ernal injury, fracture, loss of consciousness, loss of teeth. 4 x3 Physical Aggression Against Other People Makes threatening gesture, swings at people, grabs at clothes. 1 x4 Strikes, kicks, pushes, pulls hair (without injury to them) 2 x4 At tacks others, causing mild to moderate physical injury (bruises, sprain, welts). 3 x4 Attacks others, causing severe physical injury (broken bones, deep lacerations, internal injury) 4 x4 INTERVENTION (code all that apply) None 0 Talking to patient 1 Closer observation 2 Holding patient 3 Immediate medication given by mouth 4 Immediate medication given by injection 4 Isolation without seclusion (time -out) 3 Seclusion 5 Use of restrai nts 5 Injury requires immediate medical treatment for pateint 5 Injury requires immediate medical treatment for other person 5 Final Score
129 A 2. HCR risk assessment instrument (adapted from the HCR 20) HCR 20 Coding Sheet Su bject # __________ Date __________ Historical Items Code (0,1,2) H1 Previous Violence H2 Young Age at First Incident H3 Relationship Instability H4 Employment Problems H5 Substance Use Problems H6 Major Mental Illness H7 Psychopatholog y H8 Early Maladjustment H9 Personality Disorder H10 Prior Supervision Failure Historical Items Total /20 Clinical Items Code (0,1,2) C1 Lack of Insight C2 Negative Attitudes C3 Active Symptoms of Major Mental Illness C4 Impulsivi ty C5 Unresponsive to Treatment Clinical Items Total /10 Risk Management Items In___ Out___ (placement) Code (0,1,2) R1 Plans Lack Feasibility R2 Exposure to Destabilizers R3 Lack of Personal Support R4 Noncompliance with Remediation Attempts R5 Stress Risk Management Items Total /10 HCR 20 Total /40 Note: Scoring: 0=not present; 1=possible/less serious; 2=definite/serious; I=insufficient information.
130 A 3. Brief Psychiatric Rating Scale (BPRS) 18 Item Subject # Rating 0=Not assessed 4=Moderate 1=Not present 5=Moderately severe 2=Very mild 6=Severe 3=Mild 7=Extremely severe Factor Item Description Score Affect Somatic Concern Preoccupation with physical health, fear of physical illness, hypochondriasis. Anxiety (subjective) Worry, fear, over concern for present or future, uneasiness. Guilt Feelings Self blame, shame, remorse for past behavior. Depressed Mood Sorrow, sadness, despondency, pessimism. Factor Score Po sitive Conceptual Disorganization Thought processes confused, disconnected, disorganized, disrupted. Grandiosity Exaggerated self opinion, arrogance, conviction of unusual power or abilities. Hallucinatory Behavior Perceptions without normal exte rnal stimulus correspondence. Unusual Thought Content Unusual, odd, strange, bizarre thought content. Factor Score Negative Emotional Withdrawal Lack of spontaneous interaction, isolation deficiency in relating to others. Motor Retardati on Slowed, weakened movements or speech, reduced body tone. Blunted Affect Reduced emotional tone, reduction in formal intensity of feelings, flatness. Factor Score Resistance Hostility Animosity, contempt, belligerence, disdain for others. Suspiciousness Mistrust, belief others harbor malicious or discriminatory intent. Uncooperativeness Resistance, guardedness, rejection of authority. Activation Tension (subjective) Physical and motor manifestations of nervousness, over a ctivation. Mannerisms & Posturing Peculiar, bizarre, unnatural motor behavior (not including tic). Excitement Heightened emotional tone, agitation, increased reactivity. Disorientation Confusion or lack of proper association for person, plac e or time. Factor Score BPRS Total
131 APPENDIX B SAMPLE CHARACTERISTI CS B 1. Descriptive characteristics of categorical variables (N=102) DSM Primary Diagnosis N (Frequency) Percent Admission Number N (Frequency) Percent Schizo & Other Ps ychotic d/o 83 81.4 1 74 72.5 Mood d/o 9 8.8 2 24 23.5 Anxiety d/o 1 1 3 4 3.9 Adjustment d/o 1 1 Substance Related d/o 7 6.9 Charge Type N (Frequency) Percent Delirium, Dementia, Other Cognitive d/o 1 1 First Degree Felonies 7 6.9 Seco nd Degree Felonies 40 39.2 Ethnicity N (Frequency) Percent Third Degree Felonies 44 43.1 African American 67 65.7 Capital Offenses 11 10.8 Caucasian 30 29.4 Hispanic 4 3.9 Violent N (Frequency) Percent Asian 1 1 No 43 42.2 Yes 59 57.8 Ab use History N (Frequency) Percent None 43 42.2 Violence Type N (Frequency) Percent Physical 16 15.7 None 42 41.2 Sexual 3 2.9 Verbal 46 45.1 Combat 3 2.9 Objects 3 2.9 Multiple Trauma 18 17.6 Self 1 1 Unknown 19 18.6 Others 4 3.9 Ver bal/Self 1 1 Head Trauma with LOC N (Frequency) Percent Verbal/Others 5 4.9 None 56 54.9 Single Episode 25 24.5 Multiple Episodes 21 20.6 Veteran Status N (Frequency) Percent Religious Affiliation N (Frequency) Percent None 86 84.3 No Affiliation 23 22.5 Army 13 12.7 Protestant 1 1 Air Force 1 1 Catholic 12 11.8 Coast Guard 1 1 Jewish 2 2 Unknown 1 1 Muslim 5 4.9 Baptist 13 12.7 Psychiatrist N (Frequency) Percent Church of Christ 1 1 1 17 16.7 Pentecostal 4 3.9 2 1 4 13.7 Jehovah Witness 2 2 3 30 29.4 Generic Christian 18 17.6 4 12 11.8 Methodist 3 2.9 5 6 5.9 Other 1 1 6 18 17.6 Unknown 17 16.7 7 5 4.9
132 B 2 Descriptive characteristics of continuous variables. Mean Std. Error of Mean Median Mode Std. Dev iation Variance Minimum Maximum Sum Percentiles 25.0 50 75 90 AGE 36.75 1.23 36 21 12.44 154.8 19 77 3749 25.8 36 46 56 EDUCATION LEVEL 12.58 1.17 11 12 11.83 139.91 6 99 1283 9.0 11 12 14 MMSE 22.96 0.51 24 26 4.86 23.64 7 30 2089 20.0 24 26 29 BHS 4.910 0.52 3 1 4.82 23.19 0 19 422 1.0 3 9 12 HCR Scores Historical Subscale 12.73 0.36 13 10 3.66 13.37 3 20 1298 10.0 13 16 17 Clinical Subscale 6.02 0.22 6 7 2.18 4.7 5 1 10 614 4.0 6 8 9 Total Score 18.7 0.49 18 16 5.00 24.97 7 30 1907 15.0 18 22 26 Aggressive Incidents Verbal 2.20 0.38 1 0 3.86 14.93 0 23 224 0 1 3 7.7 Objects 0.41 0.12 0 0 1.25 1.55 0 8 4 2 0 0 0 1.0 Self 0.11 0.07 0 0 0.72 0.51 0 7 11 0 0 0 0 Others 0.38 0.10 0 0 1.02 1.05 0 7 39 0 0 0 1.0 Total 3.10 0.53 1 0 5.37 28.82 0 25 316 0 1 3 9.7 Aggression Severity 15.45 2.94 2 0 29.67 880.49 0 150 1576 0 2 15 64 BPRS Affect 3.46 0.23 3.38 0 2.37 5.6 0 9.25 353 1.58 3 5 6.6 Positive 7.1 0.44 6.65 6.5 4.44 19.69 0 18.50 724 3.69 7 10 14 Negative 2.93 0.25 2.63 0 2.53 6.39 0 9.25 299 0.69 3 5 7.2 Resistance 4.18 0.33 3.71 0 3.29 10.8 0 14.25 426 1.69 4 6 9.5 Activation 4.24 0.37 3.59 0 3.74 13.95 0 18.75 432 1.69 4 5 9.5 Overall 21.85 1.14 20.5 11 11.54 133.22 2.33 52.75 2228 13.2 21 28 39 Note: MMSE=Mini Mental Status Examina tion; BHS=Beck Hopelessness Scale.
133 B 3. Correlations between HCR and BPRS instruments and aggression measures. N=102 N=102 N=59 N=102 N=59 Measure Subscale/Factor Variable Violence Y/N OAS Severity OAS Severity Verbal vs. Physical Verbal vs. Physical r r r r r HCR 20 Total Score .255** .324** .273* .218** History Subscale .196* .251* .180* Previous Violence .358** Employment Problems .276* Clinical Subscale .284** .367*** .334** .227** Lack of Insight Negative Attitudes .265* Impulsivity .208* .302* BPRS Total .313*** .412** Affect Somatic Concern Anxiety (subjective) .222* .212* Positive Symptoms .244* .355** Conceptual Disorganization .299* Grandiosity .238* .217* .174* Hallucinatory Behavior .228* .332** Unusual Thought Content .283* Negative Symptoms .292* Emotional Withdrawal .327*** Resistance .251* .264* Hostility .288*** .365*** .311* .194* Uncooperativeness Activation .240* .336*** .372** Tension (subjective) .241* .294** .281* .275* .275* Mannerisms & Posturing Excitement .327*** .388*** .303* Disor ientation .266* Note: *p < .05 **p < .01 ***p<.001
134 B 4. Correlations between demographic variables and aggression measures N=102 N=102 N=59 N=102 N=59 Measure Subscale/Factor Variable Violence Y/N OAS Severity OAS Severity Verba l vs. Physical Verbal vs. Physical r r r r r Demographics Primary Dx 0.268* Psychiatrist Charge Severity Age 0.259** Admission # Veteran Status R eligion Marital Status Education MMSE Head Trauma Abuse BHS Malingering Ethnicity Disposition Note : *p < .05 **p < .01 ***p<.001
135 APPENDIX C ANALYSIS OF UNIVARIA TE OUTLIERS C 1. Cases with extreme scores on primary measures. Subject Variable Mean SD Subject's Score Z Score 7044 BPRS Resistance Factor 4.18 3.29 14.25 3.06 7099 BPRS Act ivation Factor 4.24 3.74 18.75 3.88 7044 BPRS Activation Factor 4.24 3.74 9 3.88 2042 OAS Total (Aggression Severity) 21.85 11.54 112 3.25 2689 OAS Total (Aggression Severity) 146 4.40 6781 OAS Total (Aggression Severity) 150 4.53 6781 Frequency of Aggression 3.10 5.37 21 3.33 7053 Frequency of Aggression 3.10 5.37 22 3.52 2689 Frequency of Aggression 3.10 5.37 24 3.89 2042 Frequency of Aggression 3.10 5.37 25 4.01
136 LIST OF REFERENCES Amor e, M., Menchetti, M., Tonti, C., Scarlatti, F., Lundgren, E., Esposito, W., & Berardi, D. (2008). Predictors of violent behavior among acute psychiatric patients: clinical study. Psychiatry and Clinical Neurosciences 62 (3): 24755. American Psychiatric A ssociation (2000). Diagnostic and Statistical Manual of Mental Disorders, 4th ed, Text Revision (DSM IV TR ). Washington, DC. Appelbaum, P.S., Robbins, P.C., & Monahan, J. (2000). Violence and delusions: data from the MacArthur Violence Risk Assessment Stu dy. American Journal of Psychiatry 157 (4): 56672. Appelbaum, P.S., Robbins, P.C., & Roth, L.H. (1999). Dimensional approach to delusions: comparison across types and diagnoses. American Journal of Psychiatry, 156(12): 1938 43. Apperson, L.J., Mulvey, E .P., & Lidz, C.W. (1993). Short term clinical prediction of assaultive behavior: artifacts of research methods. American Journal of Psychiatry, 150(9): 13749. Asnis, G. M., M. L. Kaplan, Hundorfean, & G., Saeed, W. (1997). Violence and homicidal behavior s in psychiatric disorders. Psychiatric Clinics of North America 20(2): 40525. Barbaree H.E., Seto, M.C., Langton, C.M., & Peacock, E.J. (2001). Evaluating the predictive accuracy of six risk assessment instruments for adult sex offenders. Criminal Just ice and Behavior 28 : 490521. Banks, S., Robbins, P., Silver, E., Vesselinov, R., Steadman, H., Monahan, J., Mulvey, E., Appelbaum, P., Grisso, T., & Roth, L. (2004). A multiple models approach to violence risk assessment among people with mental disorde r. Criminal Justice and Behavior 31, 324340 Belfrage, H., Fransson, G., & Strand, S. (2000). Prediction of violence using the HCR 20: A prospective study in two maximum security correctional institutions. Journal of Forensic Psychiatry 11 : 167175. Be lfrage, H., Fransson, G., Strand, S. (2004). Management of violent behavior in the correctional system using qualified risk assessments. Legal and Criminological Psychology 9 : 11 22. Bell, H. S. (2000). A potentially violent patient? American Family Phy sician, 61(7): 22378. Berg, A.Z., Bell, C.C., & Tupin, J. (2000). Clinician safety: assessing and managing the violent patient. New Directions in Mental. Health Services, ( 86): 9 29. Bonnie, R. J. and Monahan, J. (1997). Mental disorder, work disability and the law Chicago, University of Chicago Press.
137 Borum, R. (1996). Improving the clinical practice of violence risk assessment. Technology, guidelines, and training. American Psychologist, 51(9): 94556. Borum, R., Swartz, M., Riley, S., Swanson, J., Hiday, V.A., & Wagner, R. (1999). Consumer perceptions of involuntary outpatient commitment. Psychiatric Services, 50 (11): 1489 91. Borum, R., Swartz, M. & Swanson, J. (1996). Assessing and managing violence risk in clinical practice. Journal of Practical Psychiatry and Behavioral Health, 2(4): 205215. Brizer, D. A. and Crowner, M. (1989). Current approaches to the prediction of violence. Washington, DC, American Psychiatric Press. Buchanan, A. and Leese, M. (2001). Detention of people with dangerous se vere personality disorders: a systematic review. Lancet 358(9297): 19559. Buckley, P.F., Hrouda, D.R., Friedman, L., Noffsinger, S.G., Resnick, P.J., & Camlin Shingler, K. (2004). Insight and its relationship to violent behavior in patients with schizop hrenia. American Journal of Psychiatry, 161(9): 17124. Burger, G.K., Calsyn, R.J., Morse, G.A., Klinkenberg, W.D., & Trusty, M.L. (1997). Factor structure of the expanded Brief Psychiatric Rating Scale. Journal of Clinical Psychology 53(5): 4514. Cata lano, R., Dooley, D., Novaco, R.W., Wilson, G., & Hough, R. (1993). Using ECA survey data to examine the effect of job layoffs on violent behavior. Hospital & Community Psychiatry, 44(9): 8749. Chen, S.C., Hwu, H.G., & Williams, R.A. (2005). Psychiatric nurses' anxiety and cognition in managing psychiatric patients' aggression. Archives of Psychiatric Nursing, 19 (3): 1419. Cheung, P., I. Schweitzer, Crowley, K., & Tuckwell, V. (1997). Violence in schizophrenia: role of hallucinations and delusions. Schi zophrenia Research, 26(2 3): 18190. Cleckley, H. (1982). The Mask of Sanity, Revised Edition. St. Louis: Mosby Medical Library. Cohen, J. (1988). Statistical Power Analysis for the Behavioural Sciences. New York, Academic Press. Cohen, J. (1992). A Pow er Primer. Psychological Bulletin, 112 : 155159. Coid, J. (1983). The epidemiology of abnormal homicide and murder followed by suicide. Psychological Medicine, 13 (4): 85560. Convit, A., Jaeger, J., Lin, S.P., Meisner, M., & Volavka, J. (1988). Predicting assaultiveness in psychiatric inpatients: a pilot study. Hospital & Community Psychiatry, 39(4): 42934.
138 Cooke, D. J. (1998). Crosscultural aspects of psychopathy. New York, New York: Guilford Press. Czobor, P. and Volavka, J. (1996). Dimensions of th e Brief Psychiatric Rating Scale: an examination of stability during haloperidol treatment. Comprehensive Psychiatry, 37(3): 20515. Davison, J. (1997). Domestic violence: the nursing response. Professional Nurse 12(9): 6324. DeMatteo, D., K. Heilbrun, & Marczyk, G. (2006). An empirical investigation of psychopathy in a noninstitutionalized and noncriminal sample. Behavioral Sciences & the Law, 24(2): 13346. Dingemans, P. M., D. H. Linszen, Lenior, M.E., & Smeets, R.M. (1995). Component structure of t he expanded Brief Psychiatric Rating Scale (BPRS E). Psychopharmacology (Berl), 122(3): 2637. Dolan, M. and Doyle, M. (2000). Violence risk prediction. Clinical and actuarial measures and the role of the Psychopathy Checklist. British Journal of Psychiat ry, 177 : 30311. Douglas, K. and Kropp, P. (2002). A prevention based paradigm for violence risk assessment: clinical and research applications. Criminal Justice and Behavior, 29 : 617658. Douglas, K. and Ogloff, J. (2003). The Impact of Confidence on th e Accuracy of Structured Professional and Actuarial Violence Risk Judgments in a Sample of Forensic Psychiatric Patients. Law and Human Behavior 27(6): 573587. Douglas, K.S., Ogloff, J.R., & Hart, S.D. (2003). Evaluation of a model of violence risk asse ssment among forensic psychiatric patients. Psychiatric Services, 54(10): 13729. Douglas, K.S., Ogloff, J.R., Nicholls, T.L., & Grant, I. (1999). Assessing risk for violence among psychiatric patients: the HCR 20 violence risk assessment scheme and the P sychopathy Checklist: Screening Version. Journal of Consulting and Clinical Psychology, 67(6): 91730. Duxbury, J. (2002). An evaluation of staff and patient views of and strategies employed to manage inpatient aggression and violence on one mental health unit: a pluralistic design. Journal of Psychiatric and Mental Health Nursing, 9(3): 32537. Emerson, E. (1998) Working with people with challenging behaviour. In Clinical Psychology and People with Intellectual Disabilities (eds E. Emerson, C. Hatton, J. Bromley, et al), pp. 127153. Chichester: John Wiley & Sons.
139 Farrington, D.P., Loeber, R., & Kalb, L.M. (2001). Key research and policy issues. Child Delinquents: Development, Intervention, and Service Needs R. Loeber and D. P. Farrington. Thousand Oaks CA, Sage Publications, Inc. : 385394. Fayek, M., Kingsbury, S.J., & Simpson, G. (2002). Treatment Resistant Schizophrenia: Making the Determination. Psychiatric Times 19 (5). Federal Bureau of Investigation (2004). Uniform Crime Report Crime in the Uni ted States Washington, DC., U.S. Department of Justice : Table: Violent Crimes. Finnema, E.J., Dassen, T., & Halfens, R. (1994). Aggression in psychiatry: a qualitative study focusing on the characterization and perception of patient aggression by nurses working on psychiatric wards. Journal of Advanced Nursing, 19(6): 108895. Fleiss, J. L., J. B. Williams, & Dubro, A.F. (1986). The logistic regression analysis of psychiatric data. Journal of Psychiatric Research, 20 (3): 195209. Florida (1982). Criminal laws of Florida: Florida Criminal Code : Binghamton, N.Y. : Gould Publications. Freedman, D. (2001). False prediction of future dangerousness: error rates and psychopathy checklist revised. Journal of the American Academy of Psychiatry and the Law, 29(1): 8995. Glancy, G. D. and Chaimowitz, G. (2005). The clinical use of risk assessment. Canadian Journal of Psychiatry 50(1): 127. Glover, A. J., D. E. Nicholson, Hemmati, T., Bernfeld, G.A., & Quinsey, V.L. (2002). A comparison of predictor s of general and violent recidivism among high risk federal offenders. Criminal Justice and Behavior, 29(235249). Goldstein, P. J., Bellucci, P. A. Spunt, B.J., & Miller, T. (1991). Frequency of cocaine use and violence: a comparison between men and women. NIDA Research Monographs, 110 : 11338. Grann, M., Belfrage, H., Tengstrm, A. (2000). Actuarial assessment of risk for violence: predictive validity of the VRAG and the historical part of the HCR 20. Criminal Justice and Behavior, 27 : 97 114. Gray, N. S., Fitzgerald, S., Taylor, J., Macculloch, M.J., Snowden, R.J. (2007). Predicting future reconviction in offenders with intellectual disabilities: the predictive efficacy of VRAG, PCL SV, and the HCR 20. Psychological Assessment, 19(4): 4749.
140 Gray, N. S., Hill, C., McGleish, A., Timmons, D., MacCulloch, M.J., Snowden, R.J. (2003). Prediction of violence and self harm in mentally disordered offenders: a prospective study of the efficacy of HCR 20, PCL R, and psychiatric symptomatology. Journal of Consul ting and Clinical Psychology 71(3): 44351. Gray, N.S., Snowden, R.J., MacCulloch, S., Phillips, H., Taylor, J., MacCulloch, M.J. (2004). Relative efficacy of criminological, clinical, and personality measures of future risk of offending in mentally dis ordered offenders: a comparative study of HCR 20, PCL:SV, and OGRS. Journal of Consulting and Clinical Psychology 72(3): 52330. Greenwood, A. and Burt, G. (2000). Validity of the Brief Psychiatric Rating Scale within a forensic inpatient hospital. Exper t Evidence 8(1): 1530. Grove, W. M. (2005). Clinical Versus Statistical Prediction: The Contribution of Paul E. Meehl. Journal of Clinical Psychology, 61 : 12331243. Grove, W. M. and Meehl, P. E. (1996). Comparative efficiency of informal (subjective, impressionistic) and formal (mechanical, algorithmic) prediction procedures: The clinical statistical controversy. Psychology, Public Policy, and Law, 2 : 293323. Hafkenscheid, A. (1993). Reliability of a standardized and expanded Brief Psychiatric Rating Scale: a replication study. Acta Psychiatrica Scandinavica, 88(5): 30510. Haggard Grann, U., J. Hallqvist, Lngstrm, & N., Mller, J. (2006). Short term effects of psychiatric symptoms and interpersonal stressors on criminal violence a case crossover s tudy. Social Psychiatry and Psychiatric Epidemiology, 41(7): 53240. Hare, R. D., D. Clark, Grann, M., Thornton, D. (2000). Psychopathy and the predictive validity of the PCL R: an international perspective. Behavioral Sciences & the Law, 18(5): 623 45. Hare, R.D., Hart, S.D., & Harpur, T.J. (1991). Psychopathy and the DSM IV criteria for antisocial personality disorder. Journal of Abnormal Psychology, 100(3): 3918. Harris, G. T. and Rice, M. E. (2003). Actuarial assessment of risk among sex offenders. Annals of the New York Academy of Sciences, 989: 198210; discussion 23646. Harris, G., M. Rice, Quinsey, V.L. (1993). Violent recidivism of mentally disordered offenders: the development of a statistical prediction instrument. Criminal Justice and Behav ior, 20 : 315335. Harris, G. T., Rice, M. E., & Cormier, C.A. (1991). Psychopathy violent recidivism. Law and Human Behavior 15 : 625637.
141 Harris, G.T., Rice, M.E., & Cormier, C.A. (2002). Prospective replication of the Violence Risk Appraisal Guide in p redicting violent recidivism among forensic patients. Law and Human Behavior, 26(4): 37794. Hart, S. (1998). The role of psychopathy in assessing risk for violence: conceptual and methodological issues. Legal and Criminological Psychology, 3 : 121137. H art, S. D., Ed. (2001). Assessing and managing violence risk. HCR 20 Violence Risk Management Companion Guide Burnaby, BC, Simon Fraser University, Mental Health, Law, and Policy Institute. Hart, 2001 S.D. Complexity, uncertainty, and the reconceptualization of violence risk assessment Keynote address at the Annual Meeting of the European Association of Psychology and Law, Lisbon, Portugal (2001 June). Hartvig, P., Alfarnes, S., Ostberg, B., Skjnberg, M., & Moger, T.A. (2003). The prediction of violenc e in acute psychiatric units. International Journal of Forensic Mental Health 2 : 173180 Hasselblad, V. and Hedges, L. V. (1995). Meta analysis of screening and diagnostic tests. Psychological Bulletin, 117 (1): 16778. Hedlund, J. L. and Vieweg, B. W. (1980). The Brief Psychiatric Rating Scale (BPRS): A comprehensive review. Journal of Operational Psychiatry, 11 : 48 62. Heilbrun, A. B., Jr. (1990). The measurement of criminal dangerousness as a personality construct: further validation of a research inde x. Journal of Personality Assessment, 54(1 2): 1418. Higgins, A., Barker, P., & Begley, C.M. (2009). Clients with mental health problems who sexualize the nurse client encounter: the nursing discourse. Journal of Advanced Nursing, 65(3): 61624. Hill, C .D., Rogers, R., & Bickford, M.E. (1996). Predicting aggressive and socially disruptive behavior in a maximum security forensic psychiatric hospital. Journal of Forensic Science, 41 (1): 56 9. Hodgins, S. (1992). Mental disorder, intellectual deficiency, a nd crime. Evidence from a birth cohort. Archives of General Psychiatry, 49 (6): 47683. Hoptman, M.J., Yates, K.F., Patalinjug, M.B., Wack, R.C., & Convit, A. (1999). Clinical prediction of assaultive behavior among male psychiatric patients at a maximum s ecurity forensic facility. Psychiatric Services 50(11): 14616. Humphreys, M.S., Johnstone, E.C., MacMillan, J.F., & Taylor, P.J. (1992). Dangerous behaviour preceding first admissions for schizophrenia. British Journal of Psychiatry, 161 : 5015.
142 Jacobs, M.S., Ryba, N.L., & Zapf, P.A. (2008). Competence related abilities and psychiatric symptoms: an analysis of the underlying structure and correlates of the MacCAT CA and the BPRS. Law & Human Behavior, 32(1): 6477. Joyal, C.C., Putkonen, A., Paavola, P., & Tiihonen, J. (2004). Characteristics and circumstances of homicidal acts committed by offenders with schizophrenia. Psychological Medicine 34(3): 43342. Junginger, J. (1990). Predicting compliance with command hallucinations. American Journal of Psy chiatry, 147(2): 2457. Junginger, J. and McGuire, L. (2004). Psychotic motivation and the paradox of current research on serious mental illness and rates of violence. Schizophrenia Bulletin, 30(1): 21 30. Kalunian, D.A., Binder, R.L., & McNiel, D.E. (1990). Violence by geriatric patients who need psychiatric hospitalization. Journal of Clinical Psychiatry, 51(8): 3403. Kay, S.R., Wolkenfeld, F., & Murrill, L.M. (1988). Profiles of aggression among psychiatric patients. I. Nature and prevalence. Journa l of Nervous & Mental Disorders, 176(9): 539 46. Kay, S.R., Wolkenfeld, F., & Murrill, L.M. (1988). Profiles of aggression among psychiatric patients. II. Covariates and predictors. Journal of Nervous & Mental Disorders 176(9): 54757. Keene, J., Hope, T., Fairburn, C.G., Jacoby, R., Gedling, K., & Ware, C.J. (1999). Natural history of aggressive behaviour in dementia. International Journal of Geriatric Psychiatry, 14(7): 5418. Kerr, C. A. and Roth, J. A. (1986). Populations, practices, and problems in forensic psychiatric facilities. ANNALS of the American Academy of Political and Social Science, ( 484): 127 43. Kerr, C., and Roth, J. (1987). Survey of facilities and programs for mentally disordered offenders DHHS Publication (ADM) 86 1493. Rockville, MD: U. S. Department of Health and Human Services. Kim, E. (2002). Agitation, aggression, and disinhibition syndromes after traumatic brain injury. NeuroRehabilitation, 17(4): 297310. Kindy, D., Petersen, S., & Parkhurst, D. (2005). Perilous work: nurs es' experiences in psychiatric units with high risks of assault. Archives of Psychiatric Nursing, 19 (4): 16975. Klassen, D. and O'Connor, W. A. (1988). Crime, inpatient admissions, and violence among male mental patients. International Journal of Law and Psychiatry, 11(3): 30512.
143 Knoedler, D. W. (1989). The Modified Overt Aggression Scale American Journal of Psychiatry, 146(8): 10812. Krakowski, M. and Czobor, P. (2004). Gender differences in violent behaviors: relationship to clinical symptoms and ps ychosocial factors. American Journal of Psychiatry 161 (3): 45965. Krakowski, M., Volavka, J., Brizer, D. (1986). Psychopathology and violence: a review of literature. Comprehensive Psychiatry, 27(2): 13148. Kroner, D. G. and Mills, J. F. (2001). The a ccuracy of five risk appraisal instruments in predicting institutional misconduct and new convictions. Criminal Justice and Behavior 28: 471489. Lachar, D., Randle, S.L., Harper, R.A., Scott Gurnell, K.C., Lewis, K.R., Santos, C.W., Saunders, A.E., Pear son, D.A., Loveland, K.A., & Morgan, S.T. (2001). The brief psychiatric rating scale for children (BPRS C): validity and reliability of an anchored version. Journal of the American Academy of Child and Adolescent Psychiatry, 40(3): 33340. Lamb, H.R., Wei nberger, L.E., & Gross, B.H. (1999). Community treatment of severely mentally ill offenders under the jurisdiction of the criminal justice system: a review. Psychiatric Services, 50 (7): 90713. Lanza, M. L. (1983). Origins of aggression. Journal of Psychosocial Nursing & Mental Health Services 21(6): 116. Larkin, E., Murtagh, S., & Jones, S. (1988). A preliminary study of violent incidents in a special hospital. British Journal of Psychiatry, 153 : 226231. Lidz, C.W., Mulvey, E.P., & Gardner, W. (1993) The accuracy of predictions of violence to others. Journal Of the American Medical Association, 269(8): 100711. Lindsay, W.R., Hogue, T.E., Taylor, J.L., Steptoe, L., Mooney, P., O'Brien, G., Johnston, S., & Smith, A.H. (2008). Risk assessment in offen ders with intellectual disability: a comparison across three levels of security. International Journal of Offender Therapy and Comparative Criminology, 52(1): 90111. Link, B., and Stueve, A. (1994). Psychotic symptoms and the violent/illegal behavior of mental patients compared to community controls. Chicago, University of Chicago Press. Link, B., Andrews, H., & Cullen, F. (1992). The Violent and Illegal Behavior of Mental Patients Reconsidered. American Sociological Review, 57( 3): 275292.
144 Link, B. G and Stueve, A. (1995). Evidence bearing on mental illness as a possible cause of violent behavior. Epidemiologic Reviews, 17 (1): 17281. Litwack, T. (2001). Actuarial versus clinical assessments of dangerousness. Psychology, Public Policy, and the Law, (7) : 409443. Loewenstein, G.F., Weber, E.U., Hsee, C.K., & Welch, N. (2001). Risk as feelings. Psychological Bulletin, 127 (2): 26786. Long, J.D., Harring, J.R., Brekke, J.S., Test, M.A., & Greenberg, J. (2007). Longitudinal construct validity of Brief Symptom Inventory subscales in schizophrenia. Psychological Assessment, 19(3): 298308. Lukoff, D., Nuechterlein, K.H. & Ventura, J. (1986). Manual for the Expanded Brief Psychiatric Rating Scale. Schizophrenia Bulletin, 12 : 594602. Maden, A., Scott, F., Burnett, R., Lewis, G.H., & Skapinakis, P. Offending in psychiatric patients after discharge from medium secure units: prospective national cohort study. BMJ, 328 (7455): 1534. Maguire, J. and Ryan, D. (2007). Aggression and violence in mental health ser vices: categorizing the experiences of Irish nurses. Journal of Psychiatric and Mental Health Nursin g, 14(2): 1207. Malla, A. K. and Norman, R. M. (1983). Mental hospital and general hospital psychiatric units: a comparison of services within the same ge ographic area. Psychological Medicine, 13(2): 4319. Martin, M. A. (2003). Predicting institutional behavior in youthful offenders: The role of individual and family factors in risk assessment. Psychology Denton, TX, University of North Texas : 139. Massoc, S. (2003). "[Manipulation and seductive behavior of patients towards nursing personnel]." Soins Psychiatrie ( 224): 18 22. Marzuk, P.M., Tardiff, K., & Hirsch, C.S. (1992). The epidemiology of murder suicide. Journal Of the American Medical Associatio n, 267(23): 317983. McNiel, D. E. and Binder, R. L. (1991). Clinical assessment of the risk of violence among psychiatric inpatients. American Journal of Psychiatry, 148(10): 131721. McNiel, D. E. and Binder, R. L. (1994). The relationship between ac ute psychiatric symptoms, diagnosis, and short term risk of violence. Hospital & Community Psychiatry, 45(2): 1337.
145 McNiel, D. E. and Binder, R. L. (1995). Correlates of accuracy in the assessment of psychiatric inpatients' risk of violence. American Jo urnal of Psychiatry, 152(6): 9016. McNiel, D.E., Binder, R.L., & Greenfield, T.K. (1988). "Predictors of violence in civilly committed acute psychiatric patients." American Journal of Psychiatry, 145(8): 96570. McNiel, D.E., Eisner, J.P., & Binder, R.L (2000). The relationship between command hallucinations and violence. Psychiatric Services, 51 (10): 128892. McNiel, D.E., Sandberg, D.A., & Binder, R.L. (1998). The relationship between confidence and accuracy in clinical assessment of psychiatric pati ents' potential for violence. Law & Human Behavior, 22(6): 65569. Meloy, J. R. and. Gacono, C. B (1992). A psychotic (sexual) psychopath: "I just had a violent thought ..." Journal of Personality Assessment, 58(3): 48093. Menzies, R. and C. D. Webster (1995). Construction and validation of risk assessments in a six year follow up of forensic patients: a tridimensional analysis. Journal of Consulting and Clinical Psychology, 63 (5): 76678. Menzies, R. J., Webster, C. D., McMain, S., Staley, S., & Scagli one, R. (1994). The Dimensions of Dangerousness Revisited: Assessing Forensic Predictions about Violence. Law and Human Behavior, 18(1): 1 28. Menzies, R.J., Webster, C.D., & Sepejak, D.S. (1985). The Dimensions of Dangerousness: Evaluating the Accuracy o f Psychometric Predictions of Violence among Forensic Patients. Law and Human Behavior, 9( 1): 49 70. Merten, T., Friedel, E., Stevens, A. (2006). [Insufficient cooperativeness in forensic neuropsychiatric assessment: prevalence estimates of negative resp onse bias]. Versicherungsmedizin 58 (1): 1921. Monahan, J. (1981). The clinical prediction of violent behavior. Rockville, MD, National Institute of Mental Health. Monahan, J. (1997). Clinical and Actuarial predictions of violence. St Paul, West Publish ing Company. Monahan, J. (2001). Rethinking risk assessment : the MacArthur study of mental disorder and violence Oxford ; New York, Oxford University Press. Monahan, J., Steadman, H. J., Silver, E., Appelbaum, P. S., Robbins, P. C., Mulvey, E. P., Roth L., Grisso, T., & Banks, S. (2001). Rethinking Risk Assessment: The Macarthur Study of Mental Disorder and Violence. New York, Oxford University Press.
146 Monahan, J. and Steadman, H. J. (1994). Violence and mental disorder : developments in risk assessmen t Chicago, University of Chicago Press. Monahan, J., Steadman, H.J., Appelbaum, P.S., Robbins, P.C., Mulvey, E.P., Silver, E., Roth, L.H., & Grisso, T. (2000). Developing a clinically useful actuarial tool for assessing violence risk. British Journal of Psychiatry, 176 : 3129. Monahan, J., Steadman, H.J., Robbins, P.C., Appelbaum, P., Banks, S., Grisso, T., Heilbrun, K., Mulvey, E.P., Roth, L., & Silver, E. (2005). An actuarial model of violence risk assessment for persons with mental disorders. Psychiat ric Services, 56(7): 810815. Mossman, D. (1994). "Assessing predictions of violence: being accurate about accuracy. Journal of Consulting & Clinical Psychology, 62(4): 78392. Murdoch, D., Pihl, R.O., Ross, D. (1990). Alcohol and crimes of violence: pre sent issues. International Journal of the Addictions, 25(9): 106581. National Institutes of Health (U.S.). Office of Communications. (1998). HEALTHWise newsletter consumer health information based on research from the National Institutes of Health Bethe sda, MD, National Institutes of Health. Newhill, C.E., Mulvey, E.P., & Lidz, C.W. (1995). Characteristics of violence in the community by female patients seen in a psychiatric emergency service. Psychiatric Services, 46 (8): 7859. Nicholls, T.L., Brink, J., Greaves, C., Lussier, P., & Verdun Jones, S. (2009). Forensic psychiatric inpatients and aggression: an exploration of incidence, prevalence, severity, and interventions by gender." International Journal of Law and Psychiatry, 32(1): 2330. Nolan, K.A ., Shope, C.B., Citrome, L., & Volavka, J. (2009). Staff and Patient Views of the Reasons for Aggressive Incidents: A Prospective, Incident Based Study. Psychiatric Quarterly 2009 May 2. [Epub ahead of print]. Otto, R. (June 27, 2008). An Introduction to Assessing & Managing Violence Risk Gainesville Criminal Court House, Florida Mental Health Institute. Otto, R. K. (2000). Assessing and managing violence risk in outpatient settings. Journal of Clinical Psychology, 56 (10): 123962. Overall, J. E. and B eller, S. A. (1984). The Brief Psychiatric Rating Scale (BPRS) in geropsychiatric research: I. Factor structure on an inpatient unit. Journal of Gerontology 39(2): 18793. Overall, J. E. and Brown, & W. L. (1962). A factor analysis of several measures of choice behavior from a probability learning situation. Journal of General Psychology 66 : 115 28.
147 Overall, J. E. and Klett, C. J. (1972). Applied Multivariate Analysis. New York, McGraw Hill. Owen, C., Tarantello, C., Jones, M., & Tennant, C. (1998). "Vi olence and aggression in psychiatric units. Psychiatric Services 49(11): 14527. Park, S.H., Goo, J.M., & Jo, C.H. (2004). Receiver operating characteristic (ROC) curve: practical review for radiologists. Korean Journal of Radiology 5(1): 11 8. Peduzzi P., Concato, J., Kemper, E., Holford, T.R., & Feinstein, A.R. (1996). A simulation study of the number of events per variable in logistic regression analysis. Journal of Clinical Epidemiology, Dec;49(12) :13739. Perry, B.L., Pescosolido, B.A., Martin, J.K., McLeod, J.D., & Jensen, P.S. (2007). Comparison of public attributions, attitudes, and stigma in regard to depression among children and adults. Psychiatric Services, 58(5): 6325. Putkonen, H., Komulainen, E.J., Virkkunen, M., Eronen, M., & Lnnqvi st, J. (2003). Risk of repeat offending among violent female offenders with psychotic and personality disorders. American Journal of Psychiatry, 160(5): 94751. Quanbeck, C.D., Stone, D.C., Scott, C.L., McDermott, B.E., Altshuler, L.L., & Frye, M.A. (2004). Clinical and legal correlates of inmates with bipolar disorder at time of criminal arrest. Journal of Clinical Psychiatry 65(2): 198203. Quinsey, V. L., A. Book, & Skilling, T. (2004). A follow up of deinstitutionalized men with intellectual disabilities and histories of antisocial behavior. Journal of Applied Research in Intellectual Abilities, 17 : 243254. Quinsey, V.L., Jones, G.B., Book, A.S., & Barr, K.N. (2006). The dynamic prediction of antisocial behavior among forensic psychiatric patients: a prospective field study. Journal of Interpersonal Violence, 21(12): 153965. Quinsey, V. L. and McGuire, A. (1986). Maximum security psychiatric patients: Actuarial and clinical prediction of dangerousness. Journal of Interpersonal Violence, 1 : 143171. Reid, W. H., G. U. Balis, & Sutton, B.J. (1997). The treatment of psychiatric disorders. Bristol, PA, Bruner/Mazel. Resnick, P. J. (1969). Child murder by parents: a psychiatric review of filicide. American Journal of Psychiatry, 126(3): 32534. Resnic k, P. J. and Scott, C. L. (1997). Legal issues in treating perpetrators and victims of violence. Psychiatric Clinics of North America, 20(2): 47387. Rice, M. E. and. Harris, G. T (1995). Psychopathy, schizophrenia, alcohol abuse, and violent recidivism. International Journal of Law and Psychiatry, 18(3): 33342.
148 Rice, M. E. and Harris, G. T (1995). Violent recidivism: assessing predictive validity. Journal of Consulting & Clinical Psychology 63(5): 73748. Rice, M. E., Harris, G. T, & Quinsey, V.L. (2002). The appraisal of violence risk. Current Opinion in Psychiatry 15(6): 589593. Roncone, R., Ventura, J., Impallomeni, M., Falloon, I.R., Morosini, P.L., Chiaravalle, E., & Casacchia, M. (1999). Reliability of an Italian standardized and expanded Brief Psychiatric Rating Scale (BPRS 4.0) in raters with high vs. low clinical experience. Acta Psychiatrica Scandinavica, 100(3): 229236. Rosenbaum, M. (1990). The role of depression in couples involved in murder suicide and homicide. American Journal of Psy chiatry, 147(8): 10369. Roth, J. A. (1994). Understanding and Preventing Violence. Public Welfare, 52(4): 3544. Rudnick, A. (1999). Relation between command hallucinations and dangerous behavior. Journal of the American Academy of Psychiatry and the Law, 27(2): 2537. Ryan, D. and Maguire, J. (2006). Aggression and violence a problem in Irish Accident and Emergency departments? Journal of Nursing Management, 14(2): 10615. Salekin, R. T., Rogers, R., & Sewell, K.W. (1996). A review and meta analysis of the Psychopathy Checklist Revised: Predictive validity of dangerousness. Clinical Psychology Science and Practice, 3: 203215. Salekin, R.T., Rogers, R., & Sewell, K.W. (1997). Construct validity of psychopathy in a female offender sample: a multitrait multimethod evaluation. Journal of Abnormal Psychology, 106(4): 57685. Salekin, R.T., Rogers, R., Ustad, K.L., & Sewell, K.W. (1998). Psychopathy and recidivism among female inmates. Law & Human Behavior, 22(1): 10928. Scott, C. L. and Resnick, P. J. (2006). Violence risk assessment in persons with mental illness. Aggression and Violent Behavior, 11 : 598611. Scott, J., Harrington, J., House, R., & Ferrier, I.N. (1996). A preliminary study of the relationship among personality, cognitive vulnerabilit y, symptom profile, and outcome in major depressive disorder. Journal of Nervous & Mental Disorders, 184 (8): 5035. Secker, J., Benson, A., Balfe, E., Lipsedge, M., Robinson, S., & Walker, J. (2004). Understanding the social context of violent and aggress ive incidents on an inpatient unit. Journal of Psychiatric and Mental Health Nursing, 11(2): 1728.
149 Services, U. S. Department of Health & Human Services (2001). Youth violence: A report of the Surgeon General Rockville, MD, U. S. Department of Health a nd Human Services, Centers for Disease Control and Prevention, National Center for Injury Prevention and Control, Substance Abuse and Mental Health services Administration, Center for Mental Health Services, and National Institutes of Health, National Inst itute of Mental Health. Silva, J., Weinstock, R., & Leong, G. (2003). Forensic psychiatric report writing. Principles and Practice of Forensic Psychiatry. R. Rosner. London, Arnold : 3136. Silver, E., Mulvey, E.P., & Monahan, J. (1999). Assessing violenc e risk among discharged psychiatric patients: toward an ecological approach. Law & Human Behavior, 23(2): 23755. Sirotich, F. (2008). Correlates of Crime and Violence among Persons with Mental Disorder: An Evidence Based Review. Brief Treatment and Crisi s Intervention, 8(2): 171194. Sjostedt, G. and Langstrom, N. (2001). Actuarial assessment of sex offender recidivism risk: a cross validation of the RRASOR and the Static 99 in Sweden. Law & Humand Behavior, 25(6): 62945. Smith, R. S. (1988). Comments on the Overt Aggression Scale (letter). American Journal of Psychiatry, 145(1610). Sorgi, P., Ratey, J., Knoedler, D.W., Markert, R.J., & Reichman, M. (1991). Rating aggression in the clinical setting. A retrospective adaptation of the Overt Aggression Sc ale: preliminary results. Journal of Neuropsychiatry and Clinical Neurosciences, 3(2): S526. Steadman, H.J., Mulvey, E.P., Monahan, J., Robbins, P.C., Appelbaum, P.S., Grisso, T., Roth, L.H., Silver, E. (1998). Violence by people discharged from acute ps ychiatric inpatient facilities and by others in the same neighborhoods. Archives of General Psychiatry, 55(5): 393401. Steinert, T., Wiebe, C., Gebhardt, R.P. (1999). Aggressive behavior against self and others among first admission patients with schizop hrenia. Psychiatric Services, 50(1): 85 90. Stompe, T., Ortwein Swoboda, G., Schanda, H. (2004). Schizophrenia, delusional symptoms, and violence: the threat/control override concept reexamined. Schizophrenia Bulletin 30(1): 3144. Strand, S., Belfrage, H., Fransson, G., & Levander, S. (1998). Clinical and risk management factors in risk prediction of mentally disordered offenders more important than actuarial data? A retrospective study of 40 mentally disordered offenders assessed with the HCR 20 violenc e risk assessment scheme. Legal and Criminological Psychology, 4 : 67 76.
150 Swanson, J., Estroff, S., Swartz, M., Borum, R., Lachicotte, W., Zimmer, C., & Wagner, R. (1997). Violence and severe mental disorder in clinical and community populations: the effe cts of psychotic symptoms, comorbidity, and lack of treatment. Psychiatry, 60(1): 1 22. Swanson, J. W. (2008). Preventing the unpredicted: managing violence risk in mental health care. Psychiatric Services, 59(2): 1913. Swanson, J. W. and Holzer, C.E. 3 rd (1991). Violence and ECA data. Hospital & Community Psychiatry, 42(9): 9545. Swanson, J.W., Holzer, C.E. 3rd, Ganju, V.K., & Jono, R.T. (1990). Violence and psychiatric disorder in the community: evidence from the Epidemiologic Catchment Area surveys. Hospital & Community Psychiatry 41(7): 76170. Swanson, M.C., Bland, R.C., & Newman, S.C. (1994). Epidemiology of psychiatric disorders in Edmonton. Antisocial personality disorders. Acta Psychiatrica Scandinavica Supplementum, 376: 63 70. Tabachnick, B.G. and Fidell, L.S. (2001). Using Multivariate Statistics Fourth Edition Needham Heights, MA: Allyn & Bacon Tardiff, K. (1996). Concise guide to assessment and management of violent patients Washington, DC, American Psychiatric Press. Tardiff, K. (1 999). Medical management of the violent patient : clinical assessment and therapy New York, M. Dekker. Tardiff, K. and Sweillam, A. (1980). Assault, suicide, and mental illness. Archives of General Psychiatry, 37(2): 1649. Tardiff, K. and Sweillam, A. (1980). Factors related to increased risk of assaultive behavior in suicidal patients. Acta Psychiatrica Scandinavica, 62(1): 63 8. Taylor, P. J. (1985). Motives for offending among violent and psychotic men. British Journal od Psychiatry, 147 : 4918. Te ngstrm, A., Hodgins, S., & Kullgren, G. (2001). Men with schizophrenia who behave violently: the usefulness of an early versus late start offender typology. Schizophrenia Bulletin, 27 : 205218. Torrey, E. F. (1994). Violent behavior by individuals with s erious mental illness. Hospital & Community Psychiatry, 45(7): 65362.
151 Van der Does, A.J., Dingemans, P.M., Linszen, D.H., Nugter, M.A., & Scholte, W.F. (1993). Symptom dimensions and cognitive and social functioning in recent onset schizophrenia. Psychol ogical Medicine, 23 (3): 74553. van Winkel, R., Myin Germeys, I., Delespaul, P., Peuskens, J., De Hert, M., & van Os, J. (2006). Premorbid IQ as a predictor for the course of IQ in first onset patients with schizophrenia: a 10year follow up study. Schizo phrenia Research, 88 (1 3): 4754. Ventura, J., Green, M.F., Shaner, A., and Liberman, R.P. (1993). Training and quality assurance on the BPRS: The Drift Busters. International Journal of Methods in Psychiatric Research. 3 : 221244. Ventura, J., Nuecht erlein, K.H., Subotnik, K.L., Gutkind, D., & Gilbert, E.A. (2000). Symptom dimensions in recent onset schizophrenia and mania: a principal components analysis of the 24item Brief Psychiatric Rating Scale. Psychiatry Research 97(2 3): 12935. Vitacco, M. J. and Rogers, R. (2001). Predictors of adolescent psychopathy: the role of impulsivity, hyperactivity, and sensation seeking. Journal of the American Academy of Psychiatry and the Law, 29(4): 37482. Wagner, R.F., Beiden, S.V., & Metz, C.E. (2001). Cont inuous versus categorical data for ROC analysis: some quantitative considerations. Academic Radiology, 8 :328334. Wallace, C., Mullen, P.E., & Burgess, P. (2004). Criminal offending in schizophrenia over a 25year period marked by deinstitutionalization and increasing prevalence of comorbid substance use disorders. American Journal of Psychiatry, 161 (4): 71627. Walsh, E., Buchanan, A., Fahy, T. (2002). Violence and schizophrenia: examining the evidence. British Journal of Psychiatry, 180 : 4905. Watts, D., Leese, M., Thomas, S., Atakan, Z., & Wykes, T. (2003) The prediction of violence In acute psychiatric units. International Journal of Forensic Mental Health, 2 (2003), pp. 173180. Webster, C. D., K. Douglas, Eaves, D., & Hart, S.D. (1997). HCR 20: A ssessing Risk for Violence, Version 2. Burnaby, BC, Canada,, Simon Fraser University, Mental Health, Law, and Policy Institute. Webster, C. D., G. T. Harris, Rice, M.E., Cormier, C., & Quinsey, V.L. (1994). Violence Prediction Scheme: Assessing Dangerousness in High Risk Men. Toronto, ON, University of Toronto. Webster, D.W., Vernick, J.S., Ludwig, J., & Lester, K.J. (1997). Flawed gun policy research could endanger public safety. American Journal of Public Health, 87(6): 91821. Widome, M. D. (1989). On the relevance of poor judgment. Pediatrics, 84(4): 7246.
152 Williamson, S., Hare, R. D.,& Wong, S. (1987). Violence: Criminal psychopaths and their victims. Canadian Journal of Behavioral Science, 19 : 454462. Winstanley, S. and Whittingto, R. (2004). Aggr ession towards health care staff in a UK general hospital: variation among professions and departments. Journal of Clinical Nursing, 13(1): 3 10. Wolfgang, M. E., Thornberry, T.P. & Figlio, R.M. (1987). From Boy to Man, from Delinquency to Crime. Chicago, University of Chicago Press. Yudofsky, S.C., Silver, J.M., Jackson, W., Endicott, J., & Williams, D. (1986). The Overt Aggression Scale for the objective rating of verbal and physical aggression. American Journal of Psychiatry, 143(1): 359. Zisook, S., Byrd, D., Kuck, J., & Jeste, D.V. (1995). Command hallucinations in outpatients with schizophrenia. Journal of Clinical Psychiatry, 56 (10): 4625. Zimring, F. E. (1993). Policy research on firearms and violence. Health Affairs (Millwood), 12(4): 10922.
153 BIOGRAPHICAL SKETCH Dominique Delalot was born in New York City New York, where he lived throughout his formative years. He graduated from the United Nations International School in 1987, completing the International Baccalaureate degree. He attended the aeronautics programs at Hawthorne College in Antrim New Hampshire from 1987 to 1988, and Bridgewater State College in Bridgewater Massachusetts from 1988 to 1991, earning certifications in flight training towards becoming a commercial pilo t. He subsequently completed a Bachelor of Science in p sychology at Bridgewater Sate College, and returned home to complete a Master of Arts in general psychology at New York University in 1994. Since then, he has worked as a study coordinator and resear ch assistant at numerous New York area hospitals. From 1994 to 1995, he conducted research at Mount Sinai Medical Centers Positron Emission Tomography (PET) Laboratory on the means to reduce and reverse memory loss associated with aging, and worked on dev eloping innovative approaches to learning and memory enhancement in normal older individuals. From 19972000, he worked at New York University Medical Center and the affiliated Veterans Affairs New York Harbor Healthcare System on studies examining treatm ent refractory schizophrenia, the efficacy of pharmacological interventions in the treatment of cocaine dependent outpatients, and the cognitive effects of antidepressant medications on healthy individuals. And, from 2001 to 2004, he worked at the Function al Brain Imaging Laboratory of North Shore University Hospital / Long Island Jewish Medical Center studying brain dysfunction underlying cognitive and affective abnormalities in Parkinsons disease. He arrived at the University of Florida in 2004, where he focused his efforts on gaining further specialized training in neuropsychology with a variety of medical populations, gaining additional experience in assessment and intervention in cases of severe psychopathology and
154 exposure to additional theoretical models of therapy, and working in a variety of inpatient and outpatient settings. The pursuit of a variety of experiences led to an opportunity to work with mentally ill criminal offenders at the North Florida Evaluation & Treatment Center in Gainesville, Fl orida. The present research on the static and dynamic risk variables predictive of violence in a population of psychiatrically hospitalized criminal offenders was based on this work. He complete d his Doctor of Philosophy in the Department of Clinical and Health Psychology (Neuropsycholo gy Track) in the summer of 2010, and continue d his clinical work with individuals suffering from various mental diseases, and research efforts examining the neurological underpinnings of psychological disorders.