Terrorist Attack Attribution with Machine Learning based Multiple Imputation
32 Pages Posted: 11 Aug 2020
Date Written: July 11, 2020
Empirical analysis of terrorism is mired by the inherent clandestine nature of the subject material. The central difficulty of the analysis is the fact that the majority of terrorist attacks are not claimed. According to the Global Terrorism Database, over half of all terrorist attacks recorded since 1997 went unclaimed. Because of this, researchers tend to either use all attacks in a specific area or to focus on claimed attacks only. Both of these approaches limit inference and become problematic when evaluating the behavior of specific terrorist organizations. We present an alternative approach: applying machine learning-based multiple imputation models to attribute unclaimed attacks. This imputation allows researchers to obtain a more complete view of a terrorist organization. Machine learning provides predictive accuracy and multiple imputation improves the robustness of those predictions. Using a combination of attack characteristics and strategic factors, we predict the probability an unclaimed attack was committed by a specific group. A case study on the terrorist group Tehrik-i-Taliban Pakistan finds that even under conservative classification thresholds, the group likely committed at least 39% more attacks than it claimed. We also find that there are substantial differences between claimed and unclaimed attacks in terms of inflicted casualties and target selection. This is a methodology that can be applied throughout terrorism research, thereby allowing for more in-depth group-level analysis.
Keywords: Terrorism, Multiple Imputation, Random Forest
JEL Classification: C80, F52
Suggested Citation: Suggested Citation