The Paradox of Moderation in a Polarized Era

• In Event: Ideology in Times of Polarization

Abstract

One of the great challenges in the study of American politics is understanding the roots of polarization. In Congress, the rapid polarization is, at least in part, puzzling due to evidence that moderation is electorally useful (Brady, Canes-Wrone and Cogan 2002). Perhaps the rewards to moderation has disappeared or perhaps they are overwhelmed by other incentives.

In this paper we assess the effect of political position-taking on congressional election results in the 2020 election. The paper is based on two innovations. The first is the original collection of website data for over 1,400 candidates in the primary and general election. Coupled with publically available Twitter records, we are able to amass a very substantial documentation of the public presentation of congressional candidates.

The second innovation is the development of a statistical algorithm to generate ideology estimates based on the content of the web and social media of candidates. The foundation for our approach is the wordfish model from Slapin and Proksch (2008). The wordfish model uses count data from words in text to estimate ideology.

One challenge with wordfish is that it assume that words have a consistent ideological meaning across parties; that is, there is a single ideological weight for a given word or phrase that applies to all candidates.

In practice, terms often have very different meanings across parties: if a Democrat says “Green New Deal” a lot, they are likely liberal. If a Republican says “Green New Deal” a lot, they are likely a conservative, raising the specter of an unwanted policy. In current provisional results, we see a large number of such words (e.g., @seanhannity, wuhan, “obama biden” many curse words, harvard, idiots, "white privilege", libs, hitler, @rudygiuliani). We also see terms that moderates of both parties use more than extremists (e.g., "deliver real results", "bipartisan solutions", "dairy farmers", "bravely served", "bush family").

For all these terms, a direct implementation of wordfish would find the relationship between the word usage and ideological to be small or non-existent because a single linear model using ideology to predict word-count will poorly fit the actual shape, which is either “v-shaped” (for words used by extremists) or “mountain-shaped” (for words used by moderates).

Allowing words to have different meanings across parties, creates an identification problem however as, in essence, the word-to-ideology models are fit separately for each party. This identification problem resembles the problem of estimating preferences of political actors across institutions (Bailey 2013). The solution here is similar to the inter-institutional case: find “bridge” outcomes where the model coefficients are the same across contexts. In this case, we can use the congressional roll call voting by incumbents with the assumption that the mapping from ideology to roll call votes is the same across parties.

We generate ideal points for more than 1,200 congressional candidates based on 11,000 politically relevant words in the Twitter corpus and 5,800 politically relevant words in the website corpus.

We then analyze the effect of ideology on election outcomes. Preliminary results indicate that moderation continues to be electorally valuable, indicating that (frustratingly) congressional candidates are finding reasons to be polarized even when it hurts them electorally.

Our approach can also be extended to include measure of affinity with Donald Trump and policy interests such as covid, factors that we will also assess for possible electoral effects in both primary and general elections.

Author

• Michael A. Bailey, Georgetown University