Predicting local violence

Abstract

Riots, murders, lynchings, and other forms of local violence are costly to security forces and society at large. Identifying risk factors and forecasting where local violence is most likely to occur should help allocate scarce peacekeeping and policing resources. Most forecasting exercises of this kind rely on structural or event data, but these have many limitations in the poorest and most war-torn states, where the need for prediction is arguably most urgent. We adopt an alternative approach, applying machine learning techniques to original panel survey data from Liberia to predict collective, interpersonal, and extrajudicial violence two years into the future. We first train our models to predict 2010 local violence using 2008 risk factors, then generate forecasts for 2012 before collecting new data. Our models achieve out-of-sample AUCs ranging from 0.65 to 0.74, depending on our specification of the dependent variable. The models also draw our attention to risk factors different from those typically emphasized in studies aimed at causal inference alone. For example, we find that while ethnic heterogeneity and polarization are reliable predictors of local violence, adverse economic shocks are not. Surprisingly, we also find that the risk of local violence is higher rather than lower in communities where minority and majority ethnic groups share power. These counter-intuitive results illustrate the usefulness of prediction for generating new stylized facts for future research to explain. Ours is one of just two attempts to forecast local violence using survey data, and we conclude by discussing how our approach can be replicated and extended as similar datasets proliferate.

The Promise and Pitfalls of Conflict Prediction: Evidence from Colombia and Indonesia

“…We are particularly effective at identifying "hot spots" with high concentrations of violence, defined as five or more incidents in a single year. Indeed, our ensemble model, which leverages the best new methods, performs better than previous sub-national attempts (Blair et al, 2017;Colaresi et al, 2016;Weidmann and Ward, 2010;Witmer et al, 2017).…”

“…While LASSO penalizes the sum of the absolute value of the regression coefficients, ridge regression penalizes the sum of squares. We follow best practices as in Blair et al (2017) and use a weighted average of the two penalties, where the weight for the LASSO penalty is α = 0.95 and the weight on the ridge penalty is 1 − α.…”

The Promise and Pitfalls of Conflict Prediction: Evidence from Colombia and Indonesia

“…We are particularly effective at identifying "hot spots" with high concentrations of violence, defined as five or more incidents in a single year. Indeed, our ensemble model, which combines the best new methods, performs better than previous subnational attempts (Blair et al, 2017;Colaresi et al, 2016;Weidmann and Ward, 2010;Witmer et al, 2017). 2 We view these results as especially important given that such local hot spots can pose a serious risk of regional or national escalation, and some of these locations will not be known to policymakers, especially in large, diverse countries.…”

On the Risk Assessment of Terrorist Attacks Coupled with Multi-Source Factors

“…The previous related studies primarily involved research from three aspects, as shown in Table 1. A terrorist attack prediction project led by Blair et al used a neural network to successfully predict the conflict in Liberia in 2010 with the data in 2008; the accuracy was between 0.65 and 0.74 [19]. Dong used the 2010-2016 forecast of terrorist attacks in India as an example to empirically examine the effectiveness of machine learning based on back propagation (BP) neural networks in real-life terrorist attacks.…”