The prediction of violent reoffending on release from prison: derivation and external validation of a scalable tool

Background
Over 30 million persons are released from prison worldwide every year, who comprise of a high risk group for perpetrating interpersonal violence. Currently there is considerable inconsistency and inefficiency in identifying those who would benefit from interventions to reduce risk.
Methods
We developed predictive models for violent reoffending on a total cohort of all 47 326 prisoners at release in Sweden 2001-2009, with 11 263 individuals who violently reoffended. First, a derivation model was developed to determine strength of pre-specified routinely collected criminal history, socio-demographic and clinical risk factors, and testing them in an external validation. We measured discrimination and calibration for prediction of violent reoffending at 1 and 2 years using specified risk cut-offs. TRIPOD guidelines were followed.
Findings
A 14 item model was developed from pre-specified routinely collected criminal history, socio-demographic and clinical risk factors, and tested in an external validation. The model showed good measures of discrimination (c-index 0.74) and calibration. For risk of violent reoffending at 1 year, sensitivity was 76% and specificity was 61%. Positive and negative predictive values were 21% and 95%, respectively. At 2 years, sensitivity was 67% and specificity was 70%. Positive and negative predictive values were 37% and 89%, respectively. Of those with predicted risk of violent reoffending of more than 50%, 88% had drug and alcohol use disorders. The model was used to generate a simple web-based risk calculator (OxRec).
Interpretation
We have developed a prediction score in a Swedish prison population that can assist in decision making on release identifying those who are at low risk of future violent offending and higher risk prisoners who may benefit from drug and alcohol treatment. Further evaluation in other populations and countries is needed.