Reconceptualizing Conflict Emergence: Evidence from Machine Learning

Abstract

Despite extensive research on the causes of civil war, many explanations remain contested. Critical theoretical omissions and methodological shortcomings inhibit our understanding of conflict emergence. Civil war is treated as a dichotomous phenomenon (peace/war), which overlooks the fact that armed conflict typically arises from preceding low-intensity collective action, such as protests and rebel mobilization. This dissertation makes two main contributions: (1) Integrating structural and procedural theories to develop a reconceptualization of conflict emergence, and (2) leveraging tools from machine learning to capture the complexity of civil war.

Civil war is a multistaged process, comprising initial collective action (stage i) and the transition to civil conflict (stage ii). Armed conflict entails significant risks and logistical challenges, making a direct link between structural risk factors and civil war seem implausible. Instead, less extreme types of collective action, such as protest, terrorism, and low-intensity armed violence, constitute a critical precursor to civil conflict. Collective action mediates the relationship between structural risk factors and civil conflict. In particular, there are two pathways to civil war: The protest and the clandestine pathway. Protests might gradually transition to armed violence, or the non-state group immediately decides to pursue violence, yet needs to mobilize first. The continued interactions between the non-state group and the government might spark an escalation process, where both sides consecutively apply heavier tactics, ultimately leading to more substantial levels of armed conflict.

Null-hypothesis significance testing encounters key challenges when studying intricate phenomena, potentially producing models that fall short in accurately capturing the underlying data-generating process. Conversely, machine learning allows constructing an empirical model with high complexity, capable of mirroring the non-linear and conditional associations in civil conflict data, while ensuring to discard sample noise. Post-hoc interpretability tools are leveraged to reveal how each variable affects the prediction, which assists in deriving theoretical conclusions. Machine learning is likewise relevant from a practical perspective, given that conflict early warning can inform policymakers where and when to intervene.

This thesis contains four papers, each addressing a unique aspect of the topic. Paper i explores the interactions between grievances and opportunity to uncover stereotypical pathways to civil war onset. Paper ii demonstrates that structural risk factors predict collective action more broadly, rather than civil war exclusively, and investigates similarities and differences in the causes of collective action. Paper iii identifies dangerous protest patterns that precede high fatalities in civil conflict, and argues that protest is a dynamic phenomenon, whose inherent mechanisms go beyond what static protest events can account for. Paper iv introduces a new conflict prediction model-the 'Onset catcher'-derived from a reconceptualization of conflict emergence and therefore tailored towards anticipating conflict onsets.