How negativity and policy content drive the spread of political messages

ABSTRACT

Research demonstrates that negative messages spread more online, both at the elite and mass levels. We know comparatively less about the role that policy content plays, and whether that might be responsible for the effect of negativity. We examine over 1.4 million congressional tweets to test the effect of message tone and topic on the number of retweets that messages receive. We find that elite messages are retweeted more when they contain political attacks. However, even while controlling for tone, we find that messages are shared more when they discuss some “culture war” issues or when they discuss Donald Trump.

KEYWORDS:

• Social media
• political communication
• policy
• negativity
• congress
• culture war

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19331681.2024.2352475.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Notes

1. A handful of member accounts have since been deleted, but most have been captured by ProPublica’s Politwoops archive. We supplement the existing tweets from Twitter with those from Politwoops where possible.

2. We omit quote tweets to better isolate messages that members generate on their own, rather than responding directly to another tweet where the original author would have varying levels of followers. It is possible that quote tweets would exhibit distinct patterns, though that is beyond the scope of the current study.

3. For our dependent variable, the variance is larger than the mean, suggesting that over dispersion is present. A likelihood ratio test of alpha = 0 tests whether a negative binomial distribution is more appropriate than a Poisson distribution. For this model, the Chi-square statistic is 3.9e + 8, which is statistically significant at the 0.000 level, demonstrating that the dependent variable is overdispersed and that a negative binomial regression is appropriate.

4. The majority of tweets in our sample do not contain attacks, which parallels patterns found in previous studies (e.g., see Pedro-Carañana et al., 2020).

5. The results of this robustness check are available upon request.

6. In analyses not reported, we replicate our models but include an additional control that captures the age of the member of Congress. This variable is not statistically significant in our models explaining the number of retweets, and our other variables are unchanged both in sign and significance when this variable is included.

7. Negative binomial regression assumes that the logarithm of the number of retweets is linear in relation to the explanatory variables. Using the number of followers directly as an explanatory variable effectively assumes that the number of retweets is exponential in the number of followers. It is, therefore, more appropriate to use the logarithm of the number of followers as our explanatory variable.

8. In analyses not reported, we have run models that interact our Attack variable with each of the policy classifications in separate models. For the environment, healthcare, infrastructure, and the military, attack messages about each of these topics are retweeted significantly more, while non-attack messages are retweeted less.

9. It is possible that some of these effects are driven more by questions of “identity” than the “culture war.” In analyses not reported, we break our women’s issues keywords into those more likely to discuss abortion (e.g., pro-choice, pro-life, Roe v. Wade) and non-abortion keywords. Abortion topics are central to the culture war, while others are more related to female identity, discrimination, etc. In this supplemental model, non-abortion topics are retweeted significantly more, while abortion messages are retweeted less. Unpacking this further would be a worthwhile avenue for future research.

Additional information

Notes on contributors

Jeffrey A. Fine

Dr. Jeffrey A. Fine, PhD in Political Science, is Professor of Political Science at Clemson University (Clemson, SC). His research largely focuses on social media use by political elites, representation in Congress, and public policy.

D. Hudson Smith

Dr. D. Hudson Smith, PhD in Physics, is Director of Applied Machine Learning at Clemson University (Clemson, SC). His research focuses on novel and impactful applications of Artificial Intelligence tools, with an emphasis on Machine Learning applications spanning diverse fields such as politics, economics, agriculture, healthcare, and education.

Cierra Oliveira

Cierra Oliveira holds a BA in computer science from Clemson University (Clemson, SC). She is a research scientist at The Bail Project, where she conducts quantitative research and data analysis related to cash bail and pre-trial detention. Her research interests lie at the intersection of data, policy, and justice.

Nicholas Deas

Nicholas Deas, is a doctoral student in Computer Science at Columbia University (New York, NY). His research focuses on the intersection of Natural Language Processing and the social sciences. In particular, he focuses on the evaluation of societal biases in language technologies and their application to psychology, political science, and sociology.

Spencer Shellnutt

Spencer Shellnutt, has a degree in political science from Clemson University (Clemson, SC). Her research examines the nature of political communication on social media, particular those from members of the U.S. Congress.

Riley Stotzky

Riley Stotzky, has degrees in political science and economics from Clemson University (Clemson, SC). Her research interests focus on economic development and the role of social media in political life.

Rachel Clyburn

Rachel Clyburn, JD, is a graduate of Clemson University (Clemson, SC) and the William & Mary Law School (Williamsburg, VA). Her research explores the factors that affect why social media messages spread online.