Modeling Conflict Duration:

Insights from Ensemble Learning


I investigate why some civil wars last longer than others. Challenging the dominant rationalist paradigm that champions credible commitment problems as the primary culprit, I focus on making accurate temporal predictions based on structural determinants that leverage the maximum variation in data. More specifically, I posit a theory of strategic decision-making under constraints, in a world where absolute and relative capabilities—both materially and politically—determine the outcome.

The timeline of my empirical strategy is as follows. First, I inserted absolute and relative capability variables into existing large-n studies on armed conflict duration and replicated them (2015). Then, I created an ensemble learner consisting of about a dozen machine learning algorithms that maximize predictive accuracy (2016). Based on these results, I selected an in-depth case study to gain qualitative insights, which led to two months in Sierra Leone for fieldwork (2017). Finally, I employed Bayesian geostatistics on the Sierra Leone case, testing the hypotheses generated by the case study on about 9k geo-coded instances of territorial control and attack patterns.

Although ensemble learning is a machine learning concept—a committee of weak learners aggregating into a strong learner—I utilize it in two ways in my dissertation. First, I use a diverse set of algorithms, which have their own strengths and weaknesses, to maximize predictive accuracy. But the whole enterprise can be seen as ensemble learning as well—by combining traditional large-n research, machine learning, case study, and GIS analysis, I aim to make a strong empirical case for conceptualizing conflict duration in a novel way.

The dissertation is set to be submitted by June 2018.