Predicting War: Challenges and statistical modeling approaches in analysing the occurrence of military conflicts
The increase in available data about military conflicts, e.g. collected by the correlates of war project (Palmer et al. , 2015) or the Uppsala Conflict Data Program (Themné́r & Wallensteen, 2014), provides the political science researcher with ample opportunities to study the onset of wars and other armed conflicts. This has led to a significant increase in application papers using quantitative analysis techniques to model armed conflicts as one major aspect of international relations. Having now both the data and the analysis software at hand generates a plethora of studies and publications. From a methodological point of view the logistic regression has become kind of a panacea for these analyses and rests as the cornerstone in this field. Since the major aim in political science rests on the empirical evaluation of theories and hence on the explanatory aspects of statistical modelling, the preference for the logistic regression approach seems a logical consequence.
In this study, we look at a number of modern classification algorithms, such as CART, AdaBoost, neural nets, and support vector machines, as alternatives for the logistic regression model. Addressing the explanatory power of the resulting models as well as the predictive accuracy of the classifiers, a comparative evaluation is given. While the overall results appear to be fairly stable across algorithms, there are subtle differences in the resulting models that require proficient choices by the data analyst.
For further information about CoWell and the Brown Bag Research Seminars please contact Prof. Dr. Guido Möllering.