Predicting Violence, Protecting Victims: On the Potential for Improving Pretrial Risk Assessment in Domestic Violence Cases

• In Event: Method Meets Mission: Empirical Applications Across Crime, Justice, and Community Response

Abstract

This paper evaluates how pretrial risk assessment tools can be optimized for domestic violence (DV) cases to better inform detention decisions. Using comprehensive data from New Jersey's post-2017 bail reform system, we first establish the causal impact of pretrial detention on reducing violent recidivism in DV cases. Through instrumental variables analysis leveraging variation in judicial assignment, we find substantial reductions in violent re-arrest in the one-year following initial arrest. Given these meaningful effects, we then analyze whether current risk assessment tools may systematically underperform when applied to DV defendants. We demonstrate that standard tools' reliance on conviction history creates predictive disparities due to significantly lower conviction rates in DV cases compared to non-DV cases. Our empirical analysis shows that specialized risk models that incorporate DV-specific factors and alternative data sources, including protection orders and arrest histories, can markedly improve both accuracy and fairness in risk prediction. However, our policy counterfactuals reveal that these potential improvements may yield limited practical benefits due to constraints in decision-making processes. These findings highlight the importance of considering both algorithmic design and institutional implementation when developing risk assessment systems for specialized criminal domains like domestic violence.

Authors

• Zubin Jelveh, University of Maryland, College Park

• Ashna Arora, Arnold Ventures

• Erick Andrade, University of Chicago