Bayesian Multilevel Analysis of Binary Time-Series Cross-Sectional Data in Political Economy

Xun Pang
In this dissertation project, I propose a Bayesian generalized linear multilevel model with pth order autoregressive errors (GLMM-AR(p)) for modeling inter-temporal dependence, con-temporary correlation, and heterogeneity of unbalanced binary Time-Series Cross-Sectional data. The model includes two unnested sources of clustering in the unit-and time-dimensions for analyzing heterogeneities and contemporal correlation which are salient in the era of globalization. Group-level variations are
more » ... variations are further explained with unit-and time-specific characteristics. For handling dynamics in politics and political economy, I apply the autoregressive error specification to analyze serial correlation which may not be fully captured by the selected covariates. Two applications on civil war and sovereign default demonstrate how the proposed model controls for multiple potential confounders. It also improves reliability of statistical inferences and helps forecasts by more efficiently using the information in data. The first application focuses on the causal relationship between ethnic minority rule and civil war onset. The GLMM-AR(p) model helps study those background factors which affect the relationship under investigation. The second applied study considers how regime duration affects sovereign default conditional on regime type by putting the national policy-making regarding repaying external debt into the international iii context. To model the heterogeneous vulnerability or sensitivity of the developing countries to global shocks, I extend the GLMM-AR(p) model to analyze time-specific unit-varying effects. iv ACKNOWLEDGEMENTS Although only a few names appear on the cover of this dissertation, there are many people to whom I owe my gratitude, because it was they who made this dissertation project possible and turned my graduate experience into one that I will cherish forever. My deepest gratitude is to my dissertation chairs, Andrew D. Martin and Jeff Gill. I have been amazingly fortunate to have Andrew as my main advisor. As one of the leading scholars in the discipline, he gave me the most important and helpful advice and guidances in the whole process of this dissertation research. I also benefited a lot from his great mentorship in term of professional development, teaching skills, and even presentation style. His support and confidence on me helped me get through many tough times over the past five years. Jeff 's great passion for Bayesian statistics and Markov Chain Monte Carlo is always inspiring and encouraging me. I am grateful to him for carefully reading and commenting on countless revisions of this manuscript. I am also thankful to him for encouraging the use of correct tools for scientific research and consistent notation in my writings. My another advisor, Nathan Jensen, is always there to listen and give advice. He has contributed a lot to the production of this dissertation with his patience and insightful suggestions. I also want to say special thanks to Edward Greenberg who opened the door of Bayesian statistics to me. It was also he who encouraged me to design my first MCMC algorithm and programme it from scratch. A number of other Washington University faculty have been helpful to me along the way, including (but not limited to)
doi:10.7936/k7mc8x52 fatcat:lugsvblfcnbxdgntyeehnwswti