Assessing the Relative Accuracy of Neural Network Models in Predicting Recidivism

Patricia L. McCall, North Carolina State University
William J. Smith, North Carolina State University
Denise L. Bissler, North Carolina State University

Neural network models have been proposed to provide superior predictive estimates of recidivism and infractions compared to conventional statistical techniques. The relative merits of neural network techniques is currently being debated among social scientists. Predictions of recidivism can have implications for assigning supervision in the community and for assessing the level of security needed among inmates in correctional facilities. The current project was funded by the North Carolina Governor's Crime Commission in order to assess the relative accuracy of neural network models compared to logistic regression models in predicting recidivism among three types of offenders: young offenders released to the community, young inmates between the age of 16 and 24, and juvenile delinquents. Data were analyzed to determine whether neural networks provides superior accuracy over logistic regression models. We were able to determine the extent to which neural networks provide more accurate prediction over conventional methods for predicting recidivism. preliminary results show that neural networks may be more accurate in predicting recidivism among true positives, but less accurate for true negatives.

(Return to Program Resources)

Updated 05/20/2006