Instead of requiring victims of to drive the interventions, public agencies in High Point has created several level of interventions that are based on criminal record and officer assessments.
The interventions vary from sending a form letter to a home visit by a detective. Our application could help High Point prioritize these interventions based on the history of police visits and arrests.
To help prioritize these interventions, we wanted to model the probability of another domestic violence related arrest given the history available to us. Because of some limitations of the data size and some missing information, we modeled something a bit different.
We modeled the probability of any arrest based on history, not just domestic violence arrests. And we made the simplifying assumption that address is a substitute for name. These assumptions should be relaxed before this system is used.
We are also concerned about the possibility that our model could reinforce any racial biases that might be present in police practices.
We focus attention to specific locations and specific individuals for interventions.
- All else equal, race is a significant factor in arrests
- All else equal, men are more likely to be arrested than women.
- All else equal, people are less likely to be arrested in Nov and Dec than other times of the year (even controlling for history and whatnot).
- The older you are, the less likely you are to get arrested
- If an offender is first arrested for domestic violence on a Wednesday or Friday, they
- Get more data
- Wire up the predictions to the frontend
- Work with HPPD to make it usable in their processes
- Understand the potential for reinforcing biases