Collaborative Research:Modeling Issues with Time-Series--Cross-Section Data

点击翻译按钮获取中文摘要

Time-Series-cross-section (TSCS) models have become popular in political science, since they allow students of comparative politics (broadly defined) to use powerful statistical methods that have been the province of students of American politics (typically studying voting behavior via large surveys). While there are many applications, the prototypical application is the study of political economy, and in particular the impact of political arrangements on economic performance in advanced industrial societies. TSCS data has also become of interest in International Relations (IR). Many quantitative IR researchers use a "dyad-year" design, where pairs of nations are observed annually for long periods of time (ranging from 40 to over 100 years). The dependent variable of interest in these studies is often the binary indicator of whether a dyad was in conflict in a given year. Binary dependent variables cause special problems.Most studies either ignored TSCS issues or treated those issues as a nuisance, using a Feasible Generalized Least Squares (FGLS) estimation method to treat those nuisances. These FGLS procedures either have poor statistical properties (in finite samples) or seem dangerous on other grounds. One reason that TSCS data is of interest is that the richness of the data allows researchers to do many things; but many of those things should not be done. By now most political science scholarly publications appear to use the methodology developed by the researchers in earlier research where they recommended the use of Ordinary Least Squares estimates coupled with panel correct standard errors, with dynamics modeled via a lagged dependent variable. However, this still treats TSCS issues as problems of estimation and not specification. The researchers will develop the next generation TSCS methodology that focuses on specification. They are interested in studying the best way to model unit heterogeneity directly as well as to account for spatial correlation that is typically seen in these data. In addition, for binary TSCS data they explore how best to model temporal dynamics.The research entails developing appropriate estimators and testing them on both actual and simulated data. In addition, the investigators develop practical advice and software for the applied researcher using TSCS data that implements this methodology. In addition to developing the next generation TSCS methodology they gather and catalog a set of reference datasets for use in evaluating TSCS models. By developing a reference set of datasets that will be made freely available, the researchers will make it easier to evaluate future proposed estimators for TSCS data in political science and therefore improve the field of political methodology.