Do Primaries Work? Bayesian Causal Models of Partisan Ideology and Congressional Nominations

Dissertation by Michael G. DeCrescenzo, Political Science, UW–Madison. Defended Oct 16, 2020.

The version of the document submitted for archiving was generated at a prior commit. Click here for that PDF.

Abstract

Conventional wisdom surrounding primary elections in the United States holds that primary competition is a standout reason why classic predictions from formal models of elections---candidates should take ideological positions near the district's median voter---fail to manifest in the real world. Because partisan candidates must win their party's nomination before advancing to the general election, and primary voters tend to hold non-median beliefs on issues of public policy, candidates may feel stronger incentives to take strong partisan stances on issues to appease their primary electorates than they feel to take moderate stances to appease the general electorate.

This dissertation tests this theory by employing a novel measure of district-party ideology, the ideological preferences of Democratic and Republican partisans in each district, estimated with a Bayesian group-level item–response model. Under the assumptions of structural causal models, I find that primary candidates position themselves to match the ideological preferences of their primary electorates, even controlling for the level of general election competition in their district. Primary voters, meanwhile, prefer primary candidates that broadly represent the ideological "core" of their party, but I find little evidence that primary voters discriminate much between relative ideologues or relative moderates except in extreme cases. This project also establishes and exemplifies a Bayesian framework for estimating causal effects, which is undeveloped in political science but is valuable for estimating causal effects using structural models and data estimated from measurement models.

Aims (short)

1. Create novel estimates for the policy preferences of partisan groups within Congressional districts, using customized Bayesian IRT modeling approach.

2. Apply novel estimates of local partisan preferences to test key theoretical claims about representation in primary elections: does the extremism/moderation of local preferences (X) meaningfully affect the extremism/moderation of primary candidates for Congress (Y1) and the extremism/moderation of the candidate eventually nominated to run in the general election (Y2)?

3. Explore a Bayesian framework for causal inference in political science: formal notation, theoretical clarity on the meaning and application of priors, and practical guidance for pragmatic causal inference with Bayesian value-added. Apply Bayesian estimation to structural causal models with causal diagrams/graphs.