| Scope and method of study. A backpropagation neural network and discriminant analysis were compared for their efficacy in the prediction of violent behavior. Forty-eight predictor variables including demographic data, criminal history, psychometric data, substance abuse history, and situational factors were collected from official records of male criminal offenders (N = 392) and used to predict the violent or nonviolent nature of the offense for which each subject was incarcerated. A 3-fold cross-validation technique was employed to estimate the true (population) hit rate. Subsets of the total set of predictor variables were selected by two methods: stepwise discriminant analysis and neural network weight matrix analysis. Each subset of predictor variables was also used to develop both neural network and discriminant analysis prediction models.; Findings and conclusions. Both neural network (NN) and discriminant analysis (DA) models showed statistically significant prediction accuracy of about 77% total hits on cross-validation, accounting for 38% of the variance in the criterion. As decision thresholds for classification were made increasingly stringent, however, the NN models held their accuracy better than the DA models. The highest levels of accuracy were achieved for both NN and DA models with a collection of 17 variables that included demographic data (age, income, race, unskilled labor), criminal history (probation and parole status, previous violent arrests), psychometric data (MMPI scales 1, 3, 8, 0; IQ), situational factors (being married, living with a mate, irregular work history, supporting a family), and substance abuse (benzodiazepines). The results suggest that neural networks offer a viable tool for the prediction of a low base rate human behavior such as violence. |
