395 McNeil Building
3718 Locust Walk
Like most criminal justice risk assessments, risk assessments for intimate partner violence (IPV) typically use very broad definitions of the forecasting target. Often the forecasting target is simply the presence or absence of any actions that qualify under existing statutes. A loud argument can suffice. At the other extreme can be a lethal assault. Consequently, the usual search for risk factors can be compromised by very heterogeneous outcomes. An important risk factor for an argument may be an unimportant risk factor for an assault causing injuries. A few studies have narrowed their focus to very serious forms of intimate partner violence in which the victim is injured or even killed. Such outcomes make the research extremely important. But to be effective, the research must overcome very low base rates making identification of risk factors immensely difficult. This talk addresses and tries to circumvent the problems caused by low base rates for IPV in which the victim is injured.Using a unique dataset, I focus on the attributes of very high-risk IPV perpetrators and the circumstances associated with their IPV incidents reported to the police. Very high risk is characterized as having a high probability of committing a repeat IPV assault in which the victim is injured. Rather than rely solely on an analysis of IPV incidents, I apply sequentially to data from a large metropolitan police department three algorithms: stochastic gradient boosting, an algorithm inspired by natural selection, and agglomerative clustering. The first is used to define a fitness function, the second is used to construct a population of very high-risk IPV offenders, and the third is used to help visualize the results. The constructed population does not have a problem with low base rates and instructive results are obtained. Several risk factors are universally shared by all high-risk IPV offenders that would otherwise be obscured by the low base rate. Some aggravate risk, and some mitigate risk.