Repo Contents:
- Thesis Code for Reproduction
- Data
ABSTRACT: Early warning systems and other conflict prediction techniques aim to anticipate when and where the next conflict event may occur in a given region based on available data. Similar models may be applied in predicting natural disasters, economic collapse, or academic performance. Military forces, humanitarian actors, and international organizations like the United Nations are increasingly relying on conflict prediction and early warning systems to plan for conflict events, allocate resources, and ensure safety for personnel and civilians. However, despite the growing demand for robust early warning models, there is no consensus on the best approach for variable selection or model design. Scholars have used machine learning techniques such as random forest to compare multiple models, but other models like Bayesian techniques have yet to be explored at length. Moreover, the use of prior data, which may introduce historical bias to the model, is still a contested issue.
When implementing these models, several questions must be considered, most notably the implications of failure. Inaccurate predictions by an early warning model could result in loss of life, waste of resources, and unnecessary dispersion of personnel if the predicted dates or locations are incorrect. Even if researchers can produce models with acceptable prediction accuracy thresholds, there must be a discussion regarding the potential risks of errors and how to handle them.
This thesis aims to compare the accuracy of a Bayesian logistic regression model across multiple case studies of countries facing different types of conflicts including active warfare, ceasefire, and armed group conflict. Additionally, the ethical implications of applying such a model in a military or peacekeeping environment will be discussed. The primary research question is: Can the inclusion of prior conflict data in a Bayesian logistic regression model improve the performance of early warning models to a degree that outweighs the drawbacks of error that accompany conflict prediction?
LICENSE: Apache License 2.0