Unexpected Partisan Unity Among Congressional Leaders and Legislators Using New Latent Variable Estimation Techniques and Frameworks

Citation

Ebanks, Daniel C. (2024) Unexpected Partisan Unity Among Congressional Leaders and Legislators Using New Latent Variable Estimation Techniques and Frameworks. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/ja3j-rm49. https://resolver.caltech.edu/CaltechTHESIS:09192023-153757164

Abstract

This dissertation is intended as a collection of essays which explore innovations in the development and estimation of latent variable models. These methods have many applications, including Natural Language Processing and latent correlation structures, which this dissertation explores. In addition to the statistical challenge of innovating on this class of model, latent variable methods are computationally demanding, requiring research insights related to how to render such methods feasible, both in terms of memory constraints and in terms of achieving rates of convergence in realistic time frames. The overall substantive angle of this dissertation is related to political representation, in particular to U.S. Congress. This substantive focus allows me to study the quality of our democratic institutions and their responsiveness to and interactions with the public. This dissertations harnesses novel, large datasets, which demand the innovative methods developed throughout this dissertation to answer pressing questions related to these issues of political representation. The dissertation focuses on three main data sources: social media data from members of the U.S. House on Twitter, public speech data derived from Congressional Record from 1877-2016 and elections data from the U.S. House from 1956 to 2022. All of these data relate how politicians relate to their constituents: by communicating with them through social media or in public speeches in the first and second cases; and further by trying to earn their votes in the third case. Thus, this dissertation aims to answer questions relating to the use of innovative statistical methods to recover latent features of the data. It explores these questions through the lens of their applications to questions of American legislature. A key finding across domains is the relative unity and stability between legislative leaders and members of their respective parties. In fact, this stability is apparent both in contemporaneous studies of social media, electoral representation in the post-war era, and over historical speeches on the floor of the U.S. House.

Methodologically, this dissertation argues for new frameworks for thinking about large data in political science contexts. It emphasizes the importance of descriptive statistical approaches that consider the full distribution of the data generation process, including higher-order moments beyond the mean. In Chapter 2, it shows how calibrating statistical models for accurate generative descriptions can significant implications for how researchers interpret their statistical results and can accurately uncover important quantities of interest. In Chapter 3, this dissertation proposes new ways to think about external validity when using unsupervised methods for textual analysis. Finally, all three chapters contend with approaches to latent features in the data: topical structure in chapters 1 and 3, and contemporaneous correlations in Chapter 2. All three chapters employ these latent variable methods while proposing solutions to contend with the estimation and computational obstacles imposed by using such methods on large-scale data. In doing so, these papers find unexpected stability and unity among congressional leaders and legislators, with important implications for legislative representation in the United States.