Bayesian Inference for Political Science Panel Data

Abstract

To answer substantive questions regarding individual change, it is necessary to collect information from each subject multiple times. For decades, political scientists have col- lected a great deal of panel data, relating to mass behavior, comparative politics, and international relations. Unfortunately, the most commonly used method of analyzing panel data – linear models with individual fixed effects – oftentimes masks important quantities that can be estimated using alternative strategies. In this paper, we re- view the literature on the general linear panel model, and discuss Bayesian estimation strategies using Markov chain Monte Carlo methods. The model is extremely flexible, allowing for multiple fixed and random effects, and can be estimated using standard Gibbs sampling. To illustrate the utility of the approach, we model party identification from the 1992-1996 American National Election Study panel. In addition, we provide easy-to-use software to estimate the models as part of the MCMCpack package for the R language.