Harvard University researcher Ben Green offers a succinct reason why his recent work addresses gun violence as a public health epidemic rather than as a matter strictly for law enforcement personnel
"Certainly there is a role for police, though what we've seen is that you can't just arrest your way out of the problem of gun violence," Green, a doctoral student at Harvard, said. "And, notably, what we studied is not gun crimes—who was shooting—but rather, who was getting shot. Given that we were modeling victimization of gun violence, that's much more of a role for public health.
"We're trying to prevent people from getting shot in the first place."
The approach Green and his colleagues (fellow doctoral student Thibaut Horel and Yale University sociology professor Andrew Papachristos) took in their latest research, published in JAMA Internal Medicine, used machine learning—a self-exciting network that infers the likelihood that close social "neighbors" of a gunshot victim might also become victims—to model the social networks surrounding gun violence in Chicago (a city so plagued by violence that President Donald J. Trump has described it as the site of "horrible 'carnage'"). By employing a probabilistic contagion model common in tracking some types of epidemics, the three hoped to uncover whether public health officials and their partner agencies might someday be able to stage timely, targeted interventions in a more efficient way than approaches such as simply increasing policing in "hot" neighborhoods.
As the foundational data for their hypothesis, the researchers used police data on 138,000 people who had been co-offenders, arrested together for the same crime, over an eight-year period, and the records of 16,400 shootings recorded by police during the same period, excluding suicides, accidents, and police-involved shootings. The study expanded on a previous study by Papachristos, who had discovered that more than 70% of gun violence victims could be accounted for through networks that accounted for less than 5% of Chicago's population.
"Because everybody in the network we're studying has been arrested before, it's tempting to say, 'they're criminals, why should we be worried about them, the cops should just arrest them,'" Green said. "It may be true that these same people may have been arrested before, but they are the very same people who are most likely to be victims of trauma."
The results reinforced the idea that modeling the spread of gun violence on social contagion (the spread of beliefs, attitudes, and behaviors through social interaction) was a better predictor of who would become a victim of gun violence than modeling demographics such as age, gender, and residence. Also, a combined contagion-demographic model outperformed both individual approaches. One of the key lessons, in addition to the relative accuracy of the social contagion concept itself, was that the spread of gun violence was more akin to the contagion of behavior-related diseases such as sexually transmitted infections than airborne outbreaks such as influenza.
The differentiation is important, Green said, because the airborne pathogen philosophy is what has led to the "more cops on the beat" approach in neighborhoods he contends are already over-policed. The blood-borne contagion theory allows a public health approach, in which individuals who are calculated to be at highest risk of being shot at a specific period of time—such as the period directly following the time when a first-degree acquaintance is shot—can be targeted with interventions via social agencies or public health or health system staff in a much more targeted, dynamic way.
Charles Branas, chair of the department of epidemiology at the Mailman School of Public Health at Columbia University in New York City, noted in an editorial comment accompanying the study that Green, Horel, and Papachristos have advanced the contagion theory by modeling actual data, a first in the field.
Green said the research could be just the first step in finding and modeling complementary data sets. "It's certainly possible, once you understand the particular dynamic structures of these different networks, to layer them on top of each other and try to understand the interplay between them as opposed to just how the network itself works," Green said. "That's an area of future work we see as promising."
For example, Green said, two people may appear to present the same risk through the co-offender model, but if another mitigating model, such as those who are also in a protective church network, could be laid atop the co-offenders' network, the results could help explain why one person in one network gets shot, and another does not.
Another network analysis that might someday complement the "on-the-ground" work being done by Green, Horel, and Papachristos, is underway under the direction of Desmond Patton, a professor of social work at Columbia University who is developing natural language processing algorithms to analyze the online entries of those who may be at risk of violence.
Patton said he and Papachristos attended graduate school at the University of Chicago together and that "there are clearly some key overlaps. The key difference is he is looking at social networks on the ground, and I am looking at social networks via social media."
Patton's approach leverages some of his previous work, in which he discovered that young people living in violent neighborhoods taunt each other, make threats, and brag about violence on social media platforms in ways that may lead to firearm violence, a behavior he calls "Internet banging" or "cyberbanging."
The NLP development project employed two local residents who annotated the language and slang used by those who may become gun violence victims; that data was turned over to social work students to look for patterns of expressions—of aggression, or grief and loss, for example—and labeled. The labeled data sets were then turned over to Columbia data scientists who created computational systems to automatically detect and predict the themes Patton's team has found emerging from the data.
"We have created a computational system that can accurately predict aggression and loss in a small Chicago-based Twitter data set ," Patton said, "and we are working now to scale up that data, so we are collaborating with various organizations in the city that have been doing their own social media collection. We are planning to apply our computational system to a much larger data set to see if our predictive models work on that scale."
The recent work by Green and his colleagues and Patton illustrate how data sources not traditionally considered public health data are now being manipulated to further public health research, and perhaps enable interventions.
Analyzing user-generated big data to try to predict health outcomes goes back to the debut of Google Flu Trends in 2008, in which user queries that employed words such as "flu" were analyzed to try to detect where outbreaks may be occurring prior to the release of vetted data from public health officials. The uneven performance of the platform in matching confirmed flu cases led many epidemiologists to hesitate to place much stock in such data. However, social data that has met academically rigorous standards may present epidemiologists with useful insights.
"I think the epidemiologists are opening up to the possibility of more types of data," Branas said. "They would never close out an opportunity to look at big data. The application of this to gun violence really sets down the idea that the outcome here is a diseased outcome, and we can model it using the same approaches we have used for decades for things people traditionally consider disease-based entities."
Gregory Goth is an Oakville, CT-based writer who specializes in science and technology.
No entries found