PREDICTING VIOLENT OFFENDING AMONG MENTALLY ILL IN A MULTINATIONAL SETTING – NAIVE BAYESIAN FUSION AND MODEL MERGING WITH P-COURSE
Author(s)
Erkki JO Soini, Student(HE), RN, Researcher1, Tarja Rissanen, MSc, Project worker1, Jari Tiihonen, MD, PhD, Professor2, Sheilagh Hodgins, Phd, Professor, Head3, Markku Eronen, MD, PhD, Medical Director4, Olli-Pekka Ryynänen, Phd, MD, Professor, docent51University of Kuopio, Canthia, Kuopio, Finland; 2 University of Kuopio, Niuvanniemi Hospital, Kuopio, Finland; 3 Institute of Psychiatry, London, United Kingdom; 4 Vanha Vaasa Hospital, Vaasa, Finland; 5 University of Kuopio, Kuopio, Finland
Violent crimes perpetrated by the mentally ill are a global problem causing significant financial and human burden. Severe mental disorders are associated with an increased risk of violent offending. However, predicting violent offending in the mentally ill population has been difficult. OBJECTIVES: To explore protective and risk factors for a later severe violent offense (i.e. forensic admission) among the mentally ill (schizophrenics) and to do thorough investigation for appropriate models. METHODS: Between 1998-2000, 308 forensic and general psychiatry patients were recruited to case-control study in Canada, Finland, Germany, and Sweden. An innovative greedy naive Bayesian (NB) algorithm P-Course was utilized to search generalized factors from national datasets into a merger model (naive Bayesian fusion). The evidence was assessed with posterior odds (PO) and the predictions with accuracy, diagnostic odds ratio (DOR), area under ROC curve (AUC), Gini coefficient and logarithmic loss (log score). The models were tested using leave-one-out cross-validations and substitution procedures in which data was divided into separate teaching and test sets. RESULTS: The most influential predictive risk factors for violent offence include violent behavior prior to the index hospitalization (PO 11.8; 95% credibility interval 5.9-30.3), biological father's conviction (9.4; 4.7-20.0), no use of psychotropic medications before the age of 18 (7.7; 4.3-16.5), and biological father's substance abuse (2.6; 1.7-4.1). The merger model indicates very good discriminative power (DOR from 30.25; AUC from 0.82; Gini from 0.63; log score below 0.65) as well as robustness and accuracy (Canadian 82.5%; Swedish 87.1%) for the test sets. CONCLUSIONS: The characteristics of biological father and no use of psychopharmaceuticals before the age of 18 among mentally ill were related to committing a later severe violent offense. The exploration and predictions were carried out with P-Course and the multinational merger NB model indicated high discriminatory power and robustness compared to previous studies.
Conference/Value in Health Info
2006-10, ISPOR Europe 2006, Copenhagen, Denmark
Value in Health, Vol. 9, No.6 (November/December 2006)
Code
PMH23
Topic
Health Policy & Regulatory, Health Service Delivery & Process of Care, Medical Technologies, Patient-Centered Research, Real World Data & Information Systems, Study Approaches
Topic Subcategory
Health & Insurance Records Systems, Hospital and Clinical Practices, Medical Devices, Patient Behavior and Incentives, Prescribing Behavior, Pricing Policy & Schemes, Registries, Treatment Patterns and Guidelines
Disease
Mental Health
Explore Related HEOR by Topic