Longitudinal analysis of terror incidents: a comparison of count models

Date

2025-8

Author

Cebeci, Zehra

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

431
views

0
downloads

Cite This

This study aims to analyze longitudinal count data comprising the annual number of terrorist incidents across countries. Due to its structure, the dataset exhibits character istics such as overdispersion and excess zeros, which limit the applicability of classic Poisson models. The study employs both marginal models (Generalized Estimating Equations; GEE), which examine population-averaged effects, and random effects models (Generalized Linear Mixed Models; GLMM), which account for country specific heterogeneity. Furthermore, to better capture the structural properties of the data, models such as Negative Binomial (NB), Zero-Inflated Poisson (ZIP), Zero Inflated Negative Binomial (ZINB), and hurdle models are utilized. These models aim to provide more robust estimates by addressing the challenges posed by excess zeros and high variability. By associating terrorism incidents with socioeconomic variables such as population, GDP, unemployment, urban population ratio, and governance indicators, this study compares model performances to determine the most accurate predictive model. The findings show that governance, population, and time trends consistently predict terrorism levels, while the effects of GDP, unemployment, and urbanization vary. GLMMs outperformed GEE due to their ability to capture unobserved heterogeneity. Zero-inflated and hurdle models effectively handled excess zeros and overdispersion. Among these, Poisson and Zero-inflated Poisson GLMMs demonstrated the best predictive accuracy according to RMSE and MAE values, while Negative Binomial models provided superior model fit based on AIC and BIC values. Hurdle models further refined the determinants of attack occurrence and frequency.

Subject Keywords

Zero-inflated models, Longitudinal count data, Marginal models, Random effect models, Hurdle models

URI

https://hdl.handle.net/11511/115669

Collections

Graduate School of Natural and Applied Sciences, Thesis