Multilevel Modeling for the Study of Public Opinion and Voting

Columbia University, New York NY

Investigators

Abstract

This project will develop a general set of tools for understanding and checking the fit of multilevel models. The new tools include computations of average predictive effects for models with nonlinearity, interactions, and variance components, and generalization of simulation-based model checking for multilevel models. In parallel, multilevel models will be explored for public opinion and voting data. A central application area of this project is to use national poll data to estimate time trends in public opinion for different states, a problem that cannot be solved by existing approaches using state and national polls separately. A related area of work is to model dependence structures among individual voters; that is, voter-level models that can add up districts, states, and the country to predict realistic group-level opinion patterns. This has implications for voting power and also is related to studies of networks in probability theory and sociology. This project is anticipated to have broader impacts in two ways. First, the diagnostic methods for multilevel models will be relevant to a wide range of researchers in social science and survey sampling. Second, the modeling of public opinion and voting patterns will be relevant to studies of state-level opinion trends (an important topic in this modern era of geographically-polarized voting) and for understanding the quantitative relationships between opinion and voting.