How computers are identifying violence hotspots

Computers have long been used to identify potential areas of conflict, but a new model being used in Liberia means that violence can now be predicted on a grassroots level

A statistical model that can accurately predict where violence will occur in Liberia has been developed. The advance raises the possibility that developing countries could use similar models to target peacekeeping measures at predicted hotspots, potentially preventing violence escalating into regional clashes or wars.

Statistical models that point to where violence may occur already exist, but they typically use ‘big data’ sources – such as databases of news reports or social media – to predict where regional or national violence might take place at short notice.

These models break down where insufficient data is available, for example at the level of villages with small populations or in countries where news reports are unreliable and few people use social media.

Political scientist Robert Blair of Yale University in the USA, one of the new model’s developers, says he was particularly keen to see if it was possible to predict violence at the level of small towns or villages.

“It’s small-scale incidents that you want to be able to predict in order to stop them escalating into these much larger problems”

In Liberia, the courts and police are “very weak – inaccessible and corrupt – so citizens don’t have institutions that can adjudicate when fist fights or murders happen”, he says.

That means minor clashes, especially those involving ethnic factions, can quickly spread to whole regions. “It’s precisely these small-scale incidents that you want to be able to predict in order to stop them escalating into these much larger problems which can have longer-term national implications,” says Blair.

A refined version of the model Blair and his colleagues have developed – described in a working paper published online in November 2014 on the Social Science Research Network – could suggest ways of deploying Liberia’s underfunded and understaffed police force in the most at-risk areas.

Blair began working on the model in 2009 with fellow Yale researcher Alexandra Hartman and Christopher Blattman of Harvard University. At the time, the team was working on a Liberian government-funded evaluation of a programme to resolve village disputes.

As part of this, the researchers surveyed 242 villages and towns across Liberia’s violent northern border in 2008 and 2010. They gathered a wide range of information on potential risk factors from community representatives, from the number of people who accept interreligious marriage to the number who believe other tribes are ‘dirty’.

Alongside that, they asked local leaders in 2010 and 2012 about any instances of violence in their village or town that year.

Blair and the team then plugged the 2008 survey results into their statistical models. This showed which variables tallied most closely with the violence that happened in 2010. Then they repeated the exercise with the survey results from 2010 and used it to predict violence that occurred in 2012, which they measured separately.

They found that one model correctly predicted 88% of violent incidents that occurred using just five variables.

Like what you’re reading? Positive News depends on your support to publish quality inspiring content. Please donate to help us continue pioneering a more constructive news media.

The research hints that it could be easier than previously thought to collect specific data to predict violence in other nations. That is “because we’re not collecting data across 87 different variables, we’re really looking at these four or so things,” says Michael Kleinman, director of investments at NGO Humanity International, which part-funded the research.

But he notes that fresh research would be needed to confirm whether a similarly small selection of risk factors could predict violence in other nations or in special circumstances such as the current Ebola crisis.

The 88% correct prediction rate comes at the cost of many false positives – instances where violence is predicted, but never occurs – says Blair. Yet he does not necessarily see this as a problem. What it means, he says, is that “the Liberian government could say: ‘Hey, we have the resources to send cops to 10 villages. Can you tell us which are the 10 villages with the highest risk of violence?’ Well, yes, the model can do it.”

First published by SciDev