Conflict Victimization Model in Southern Thailand: An Event-Level Analysis Using Multinomial Logistic Regression

บทคัดย่อ

This study investigates the determinants of conflict-related victimization severity in Thailand’s southernmost provinces using event-level data from 9,310 southern unrest incidents recorded between 2004 and 2020. Victimization outcomes were classified into three ordered severity categories—safe, injured, and dead—and analyzed using Multinomial Logistic Regression (MLR) to capture differentiated risk mechanisms. The dataset, compiled by the Deep South Coordination Center (DSCC), integrates police, military, and administrative records to provide comprehensive coverage of subnational conflict events. Feature selection was conducted using chi-squared screening followed by backward refinement, yielding nine key predictors encompassing temporal, spatial, contextual, and tactical dimensions. Model estimation was performed under both baseline and imbalance-adjusted weighted specifications, with robustness assessed through cross-validation and sensitivity analyses. Results indicate that non-residential locations, particularly public places, roads, and agricultural areas, substantially increase the likelihood of injury and fatal outcomes. At the same time, rural settings are associated with lower severity risk, and attacks targeting military personnel exhibit markedly elevated injury and fatality odds relative to civilian targets. Shooting attacks emerge as the most lethal modality, with an estimated 15.8-fold increase in fatal risk compared to bombings. Predictive evaluation demonstrates strong discrimination for safe outcomes (AUC = 0.94), with good to moderate performance for injury (AUC = 0.80) and fatal events (AUC = 0.75). The findings highlight the dominant roles of spatial context and attack modality in shaping victimization severity and demonstrate the value of interpretable statistical modeling for informing targeted security planning and conflict-prevention strategies.