A Comparison of Logistic Regression, Classification and Regression Tree, and Neural Networks Models in Predicting Violent Re-Offending

Previous studies that have compared logistic regression (LR), classification and regression tree (CART), and neural networks (NNs) models for their predictive validity have shown inconsistent results in demonstrating superiority of any one model. The three models were tested in a prospective sample of 1225 UK male prisoners followed up for a mean of 3.31 years after release. Items in a widely-used risk assessment instrument (the Historical, Clinical, Risk Management-20, or HCR-20) were used as predictors and violent reconvictions as outcome. Multi-validation procedure was used to reduce sampling error in reporting the predictive accuracy. The low base rate was controlled by using different measures in the three models to minimize prediction error and achieve a more balanced classification. Overall accuracy of the three models varied between 0.59 and 0.67, with an overall AUC range of 0.65-0.72. Although the performance of NNs was slightly better than that of LR and CART models, it did not demonstrate a significant improvement.