Predicting Domestic Abuse (Fairly) and Police Risk Assessment.

Area of Science:

• Criminology
• Data Science
• Public Health

Background:

• Effective domestic abuse victim risk assessment is vital for appropriate support.
• Current UK Domestic Abuse, Stalking, and Honour Based Violence (DASH) assessments fail to identify many vulnerable victims.
• There is a need for improved predictive models in domestic abuse cases.

Purpose of the Study:

• To develop and evaluate machine learning models for domestic abuse risk assessment.
• To compare the predictive performance of machine learning models against the DASH assessment.
• To identify key variables for predicting domestic abuse victim vulnerability.

Main Methods:

• Utilized data from 350,000 domestic abuse incidents from a large UK police force.
• Tested various machine learning algorithms, including logistic regression with elastic net.
• Incorporated police database information and census-area-level statistics into models.

Main Results:

• Machine learning models significantly improved predictive capacity for intimate partner violence (IPV) (AUC = .748) and non-IPV (AUC = .763) compared to DASH.
• Criminal history and domestic abuse history, especially time since the last incident, were the most influential variables.
• DASH assessment questions minimally contributed to the predictive performance of the developed models.

Conclusions:

• Machine learning offers a more accurate approach to domestic abuse risk assessment than the current DASH tool.
• Predictive models utilizing readily available police data can enhance victim identification and support.
• While fairness disparities exist across subgroups, model-based predictions offer overall accuracy improvements.