When veteran crime forecaster Richard Berk was contacted by the National Research Council (NRC) (the research arm of the National Academy of Sciences) for a project funded by the National Science Foundation (NSF), he was asked, “What can be done about mass violence?” His first thought: How do you predict a crime that is statistically rare, but increasingly dire? If you can’t predict it, how would you intervene to prevent it?
“If I were to say that for all the schools in Philadelphia there won't be a mass shooting in the next 12 months, I would be correct 99.99 percent of the time.” says Berk, Professor of Criminology and Statistics and Chair of the Department of Criminology. "My forecasts would be very accurate, but virtually useless. Still, the more I thought about the problem, it became an interesting and worthwhile challenge—finding a needle in the haystack.”
NRC’s pitch was rooted in a request for Berk to present any forthcoming research at a conference which would be attended by academics and various federal agency representatives charged with ensuring public safety. As an international leader in the use of machine learning, Berk uses very large datasets with hundreds of thousands of observations and many hundreds of variables to forecast crime.
The first step in this new research undertaking was to acquire the necessary data to formulate a test bed. Berk notes: “You can shoot 20 people, but if nobody dies, it's not officially mass violence and data on mass violence is very spotty in any case.” Another approach was required.
Berk was already actively involved in crafting strategies to learn more about Intimate Partner Violence (IPV). With Berk’s assistance, Susan B. Sorenson, Professor of Social Policy and Professor of Health and Societies, had designed a special offense form for Philadelphia police to complete when responding to calls involving IPV and other forms of domestic violence. The data gathered from thousands of these reports turned out to be a starting point for thinking about perpetrators of violence. While IPV reported to the police is very common, it is relatively rare for perpetrators to re-offend with same victim such that the victim is physically injured. Although the number of such events is large, they are a small proportion of all IPV incidents. One can treat repeat IPV incidents with injuries as statistically rare events.
The police offense forms provided a very rich set of predictors. "We can tell whether there were children in the house at the time of the incident, and we know the nature of the relationship between the perpetrator and the victim. For example, if couples are breaking up, that’s a situation known to put victims at greater risk," says Berk, who worked with Sorenson on the research paper for the NRC. "We know whether the perpetrator has been stalking, whether he broke in, whether there's property damage, whether he had a firearm. Many of these predictors are known to be associated with IPV, but many had not been studied before.”
Berk and Sorenson’s paper, “An Algorithmic Approach to Forecasting Rare Violent Events: An Illustration Based in IPV Perpetration,” applies three algorithms to the police data to identify not just perpetrators who are predicted to reoffend, but ones who are very likely to reoffend in a manner that leads to victim injuries. Because such individuals represent a very small fraction of all IPV perpetrators, they may represent a useful test case for forecasting mass violence.
"In order to construct to a predictive algorithm in such circumstances, you need a large data set,” Berk says. “You can then construct a sizeable, synthetic population of the offenders, which allows you to study simulated rare events as if they were real. If the attributes of these hypothetical offenders subsequently turn out to forecast well for real data, you are on your way to forecasting rare events of IPV. Ideally the same strategy would work for mass violence.”
The NRC conference was a who’s who of departmental agencies. “The FBI was there, the CIA, the Secret Service, Homeland Security—everybody was there,” says Berk. “Mostly it was people scratching their head because no one has ever tried this before in criminal justice, and it took a while for people to sort of digest what we accomplished.”
The question is, how do you employ these procedures? Berk says that the issues characterizing mass violence are complex and very challenging.
"We don't know much in part because a lot of the perpetrators don't survive the incident,” Berk says. “And too often we know little about their personal history. Take the mental health perspective, for instance. In order to use mental health as a predictor, you'd need to have somebody who is already in touch with providers of mental health services, so that you'd have a professional evaluation of whether the person is dangerous. But most mass murders are not in the mental health system and commonly not on anyone’s radar. Moreover, the vast majority of individuals diagnosed with mental health problems pose absolutely no risk to public safety. They would be false positives. These are not circumstances in which background checks are likely to be effective.”
While modern forecasting methods are sometimes useful in flagging certain high-risk individuals, old-fashioned police work is also essential.
“Before people commit mass shootings, there's a period of preparation. They usually have to buy or acquire semi-automatic weapons, lots of ammunition, high capacity magazines and often a bullet-resistant vest. In the recent Virginia Beach incident, a silencer was employed. Shooters also commonly broadcast their motives and intentions on social media,” Berk says. “And sometimes, friends, acquaintances, co-workers, and other notice these preparations. So we need to get better at spotting these immediate warning signs. You can bet law enforcements agencies are all over this, but such information is often not easily obtained.”
In addition to his research on mass violence, Berk also recently authored, "Almost Politically Acceptable Criminal Justice Risk Assessment," which examines accuracy and fairness in criminal forecasting and how best to respond to ideological criticisms that may be largely fact-free. His and Sorenson's paper on mass violence is set to appear in the journal Criminology and Public Policy later this year.