Weights of Evidence Model and Onset of Civil War: Favourability Analysis and Evaluation
MetadataVis full innførsel
- Institutt for geografi 
This thesis explores the use of the weights of evidence (WOFE) model for studying civil war onset in a spatial context. Previous study of civil war has extensively studied a range of geographic factors, and how these affect the outbreak of internal armed conflicts. This thesis models these geographic factors within a geographic information system (GIS), together with geographic data on the location of civil war onsets from 1960 to 2005. Hence, the analysis may be applied on a truly local level, compared to studies using country level data. Through the use of the WOFE model, each geographic factor is classified, and each class’ association with the onset point is calculated and weighted. Hence, each class’ spatial association with the outbreak of civil war is derived as weights. The calculated weights are then combined to form a posterior probability map, indicating where future onset of civil war is most likely to occur. The WOFE model is applied to both territorial and governmental outbreaks of civil war to assess differences and similarities. As the posterior probability is based on a cross sectional analysis applying onsets from 1960 to 2005, I apply a more temporally restricted analysis using onsets which have taken place after the cold-war. Accordingly, I compare the weight tables between the full analysis and the post cold-war analysis, in addition to investigating how each cells probability has changed between the two analyses posterior probability maps. In addition to a temporal restricted control analysis, I apply three model evaluation techniques, to assess the predictive power of the model. By splitting the onset sample in a training point layer and a control point layer I am able to find how well the model predicts future outbreaks of civil war. In general, the results from using the WOFE model shows that governmental onsets are associated with high population, prosperous areas, areas close to the capital and areas near natural resources. On the contrary, territorial onsets are associated with relatively low population, disadvantages regions, periphery of the country and resource scarce areas. Posterior probability maps were calculated using 80 % of the onset events, and the predicted values were compared to the location of the other 20 % of onset events. Through the use of efficiency of prediction curve and receiver operating characteristic curve, the WOFE model was assessed to accurately predict future civil war onsets. The findings of this thesis through the use of the WOFE model support previous research findings using more traditional regression analysis methods.