How Covid-Related Cues & Policies Affect Beliefs, Preferences, and Evaluations

• In Event: Elite Cues & Motivated Reasoning

Abstract

An open question in the field of American politics is the degree to which voters punish politicians for deviating from the voter’s preferred policies, or whether voters adjust their policy preferences to match those of their co-partisan elites (Lenz, 2012; Mason, 2018). Convincingly answering this question has remained an elusive goal for political science research due to data availability. To avoid the problem of observational equivalence, we need observations of individuals' beliefs both before and after either political figures have changed their position, or facts on the ground have changed. We obtain richly detailed panel data from multiple sources allowing us to observe the joint beliefs of the public, politicians, experts, and the media together over time in the context of the Covid-19 pandemic between January and December of 2020. We test whether the public reacts to policy pronouncements by politicians by updating their views of the seriousness of the pandemic, or by updating their views of the politicians themselves (i.e., did people who believed the pandemic was a hoax downgrade their view of Trump when he declared a state of emergency).

To test our hypotheses, we measure public opinion and elite cues with rich text data from social media, media coverage, and policy pronouncements. For public opinion, we collected tweets posted by a random sample of 1.4 million Twitter users observed continuously throughout 2020. To measure elite pronouncements, we collected Covid-related tweets by national and local media, congresspersons, President Trump, and governors. We also scraped the press releases of governors, as well as written newspaper articles and TV news transcripts by major media outlets throughout the year.

We created text-based measures of public opinion and elite pronouncements about the pandemic using a supervised machine learning approach. A team of research assistants labeled the various types of documents for: the seriousness of the pandemic, reference to information or misinformation, expression of policy preferences, and evaluation of government performance in handling the pandemic. This included labelling over 42,000 tweets with 78 unique labels, covering dimensions of the discourse ranging from the health risks of the disease to evaluations of public policy and officials to the sharing of (mis)information. We then trained transformer models, state-of-the-art deep learning models for text classification, to machine-code the unlabeled data. Our classifier performs well, with all labels achieving AUC scores in excess of 0.8.

Our preliminary results suggest that early in the pandemic, the public was more willing to update their views of the pandemic based on pronouncements by politicians. However, later in the pandemic, when voters had formed opinions on the issue, they were more likely to update their attitudes toward politicians when politicians endorsed or enacted policies they opposed. These findings suggest that the debate over who leads and follows in a representative democracy is not static, but rather evolves as both the public and politicians update their positions on a new issue.

Author

• James Hodgdon Bisbee, NYU