| The interdependence of international events is obvious to even casual observers of global politics. History is replete with examples of events repeating within states and/or being influenced by outcomes in other states. Despite this, much of the current literature in International Relations either mishandles or outright neglects this dependence, thereby threatening the credibility of our inferences. In large part, this stems from the difficulty of modeling such dependence when one's data are binary and rare, as they often are for many of the most widely-studied phenomena in IR (e.g., violent conflict, economic crises, etc...). For data of this type, commonly-used strategies to capture dependence are frequently ill-suited and, as such, new approaches are required. Therefore, this thesis aims to clarify the empirical challenges which arise from these data, detail the problems with existing approaches, and offer alternatives which should be preferred.;The focus is principally on two potential (and related) sources of bias which may arise within binary time-series cross-sectional (b-TSCS) data: true (inter-)dependence and unit heterogeneity. In the first, the outcomes, actions, and/or choices of some unit-times depend directly on those of other unit-times. To model both spatial and serial dependence in such data, a spatiotemporal-lag probit model estimated using maximum-simulated-likelihood using recursive-importance-sampling (MSL-by-RIS) is presented. This allows us to directly model the dependence of the lagged-latent outcomes, which is shown to have several advantages over models using the observed indicator (e.g., model consistency, effects estimation, predictive accuracy). The second main focus is on the threat of unobserved unit heterogeneity, that is, when time-invariant unit-characteristics influence the outcome, action, choice, but go unmodeled. While fixed-effects estimators are traditionally the solution to this issue, such models have received heavy criticism in political science applications with b-TSCS data. In light of these criticisms, a penalized-maximum-likelihood fixed effects (PML-FE) model is proposed which suffers from few of these drawbacks and permits the estimation of novel unit-specific substantive effects. In addition, original analyses into intrastate conflict and financial crises are offered to highlight the value of these approaches for testing existing, and motivating new, theories of international behavior. |
