How to Predict Which Soldier Will Commit Violent Acts

By
The President Visits The Army National Training Center
Photo: Charles Ommanney/Getty Images

Nipping violence in the bud before it happens is an area of major, obvious concern for behavioral researchers. To the extent violence can be predicted, researchers face the kindest odds in situations where they have access to a great deal of data about a given set of individuals — that’s what allows them to identify potential risk factors, tweak models that predict who’s most likely to commit violence, and so on. The military, it turns out, is a potentially fruitful area of research, simply because the armed services collect such a heaping quantity of data about their members. In a new study in Psychological Medicine, a team of 13 researchers used Army data to develop a machine-learning algorithm that they say can help predict which soldiers are most likely to commit violent crimes.

Machine-learning” sounds complicated, but the basic idea is relatively simple: The researchers fed a bunch of the Army’s administrative data from the almost million soldiers who served from 2004 to 2009 into a computer, and programmed the computer to develop a model of which variables and combinations of variables were correlated with a soldier being investigated by the Army for a “first accusation of a major physical violent crime, not occurring within a soldier’s family” — crimes committed against family members, researchers think, are influenced by a different set of factors. Once the model was developed, the researchers took it and applied it to a different slice of data — the 2011 to 2013 sample — to see how well it did.

Dr. Ronald Kessler, a health-policy researcher at Harvard Medical School and the paper’s corresponding author, said the team was happy with both the amount of data the Army records provided and with the accuracy of the predictive models that popped out. “We had an incredible amount of information about [the soldiers],” said Kessler, “and while you can’t predict these things with perfect accuracy, it’s pretty amazing how you can do a heck of a lot better than random.”

Specifically, the authors write, “disadvantaged social/socioeconomic status, early career stage, prior crime, and mental disorder treatment” were all predictive of violence. It’s important to note that we’re not talking about Minority Report–level accuracy here; overall, very few soldiers commit violent acts — and they don’t do so at a higher rate than the general population, Kessler was quick to point out. But there was a markedly higher risk of violence among those in the group deemed by the algorithm to be at the highest risk. According to Kessler, they were seven to ten times more likely than regular soldiers to commit a violent crime, which he compared to the increased risk of developing cancer for someone who smokes three packs of cigarettes a day versus a nonsmoker. Put differently, in a given year about 2 out of every thousand soldiers will commit a violent crime, versus 15 out of every thousand among those in the highest-risk group.

As the authors point out, while they did have access to a giant trove of data, there was some key stuff missing: The Army data didn’t include “significant predictors found in previous studies … [like] personality traits, social networks, [and] early life experiences.” But for the next wave of research conducted by Kessler and his colleagues, they will have access to more and better data — Kessler explained that since 2009 or so, the Army has begun building two new data sets with more in-depth information about soldiers’ “personalities, values, and dispositions,” as he put it. “So now we’re building a new file starting 2010 going forward where we’re going to have that [enhanced data] for everybody,” he said.

This will likely add a great deal of richness to the data, Kessler explained. He pointed out that, at the moment, the Army adopts a very egalitarian model in which all soldiers receive the same basic sorts of services and attention. But this might not be the most efficient approach, he argued — rather than provide suicide-prevention services to a massive group of soldiers who mostly aren’t at risk for suicide, for example, why not figure out who is at an elevated risk and direct more resources to them? He hopes the Army will eventually be able to apply this same logic to other-directed violence prevention as well. More sophisticated models will certainly help.