Individual Submission Summary
Share...

Direct link:

Multivariate MRP

Sun, September 3, 10:00 to 11:30am PDT (10:00 to 11:30am PDT), LACC, 303A

Abstract

Multilevel Regression and Post-Stratification (MRP) has dramatically improved survey research by allowing scholars to produce high-quality estimates of public opinion at subnational levels of geography. However, MRP as currently practiced has a major limitation: It can extrapolate only one question at a time. This limits the ability of researchers to answer questions about the relationship between issues within subnational geographies (e.g., how partisanship and public opinion covary in individual states; the prevalence and type of split-ticket voting; etc.). For example, when attempting to measure nationalization using vote choice, even very large surveys have few respondents who report particular patterns of split-ticket voting (e.g., Republican for governor; Democrat for attorney general) at most subnational geographies. This data is unfortunately too limited to be used directly to draw reliable estimates.

We develop “multivariate MRP” to address this limitation by simultaneously modelling and post-stratifying multiple questions. It uses the predictive variables from traditional MRP to predict responses to multiple questions that have been combined into a single high-dimensional categorical variable. Designing an appropriate model is methodologically challenging as the type of regularization must be carefully tailored to allow an expressive model but prevent overfitting. Further, estimation using traditional methods is expensive (taking days or even weeks). Therefore, we also develop a new method for mean-field variational inference with multinomial hierarchical models.

After estimation, the researcher would report various quantities of interest (e.g., conditional distribution of one variable given others); we discuss different possible quantities and provide software for computing them. Substantively, we provide a new comprehensive measure of nationalized voting at different levels of geography (county, state, congressional district) in the United States using the CCES. We also show that our method can be used to estimate public opinion by party at subnational geographies, improving on existing methods that have difficulty propagating uncertainty.

Authors