Comparing conventional and machine-learning approaches to risk assessment in domestic abuse cases

We compare predictions from a conventional protocol-based approach to risk assessment with those based on a machine-learning approach. We first show that the conventional predictions are less accurate than, and have similar rates of negative prediction error as, a simple Bayes classifier that makes use only of the base failure rate. A random forest based on the underlying risk assessment questionnaire does better under the assumption that negative prediction errors are more costly than positive prediction errors. A random forest based on two-year criminal histories does better still. Indeed, adding the protocol-based features to the criminal histories adds almost nothing to the predictive adequacy of the model. We suggest using the predictions based on criminal histories to prioritize incoming calls for service, and devising a more sensitive instrument to distinguish true from false positives that result from this initial screening.