Mixing messages: How candidates vary in their use of Twitter

ABSTRACT

How do candidates campaign on social media? To date, political scientists have mostly considered the content or tone of candidates’ social media posts. In this paper, we combine these research areas and study how candidates’ campaigns vary the content and tone of their messages. We argue incumbents emphasize positive issue content and limit attacking their opponent. Competitive races encourage positive mobilization posts and aggressive attack messages. Using Twitter data from the final two months of the 2018 U.S. midterm elections, we find that candidates predictably emphasize different types of messages and, in the aggregate, campaigns’ social media feeds significantly vary.

KEYWORDS:

• Elections
• congress
• Twitter
• U.S. politics

Notes

1. See Gervais, Evans, and Russell’s (2020) analysis that largely supports these findings in the 2018 U.S. midterm elections.

2. Most commonly tone is measured as a dichotomous variable (e.g., whether a post was negative or an attack).

3. Researchers have categorized social media campaign posts in numerous categories, including: issue, mobilization, fundraising, attack, campaign activities, and interacting with other social media users (Jungherr, 2016)

4. Of course, Tweets can mention both issues and mobilization.

5. Fifteen percent is a lower bound as Evans et al. (2014) categorize posts into mutually exclusive categories. As subsequent research shows (e.g., Auter & Fine, 2016, 2018), social media posts cover multiple topics.

6. These percentages imply most other posts are on other content. In our data, we found most unclassified posts are best described as candidate advertising. These messages document the campaign by tracking the candidate’s daily activities, announcing events, or posting photos of the candidate with their supporters.

7. Non-quality, first time candidates likely adopt this strategy more readily than candidates holding lower office.

8. Even before Trump’s election, the 2016 presidential race featured negativity from both Clinton and Trump. It is possible candidates adapted their behavior based on the unusually vitriolic 2016 cycle (Evans et al., 2018).

9. Our data includes Democratic and Republican candidates in each race, as well as incumbent third-party candidates such as Bernie Sanders and Angus King. In states with top two and jungle primaries, we included all Democratic and Republican candidates. Our analysis excludes the 27 candidates who did not maintain Twitter accounts (all long-shot candidates). Additionally, 27 candidates, all who lost their races and were long-shots, deleted their Twitter accounts before we collected their tweets. The main concern is this missing data may affect our analysis of Hypothesis 3c, that long-shot candidates send more attack posts. However, we suspect this bitter, sore loser behavior is likely correlated with running negative, aggressive campaigns, not politeness. As such, it is likely this missingness leads us to underestimate our effects when evaluate Hypothesis 3c.

10. We collected challengers’ verified campaign accounts and officeholders’ campaign, official, and, if consistently used for political purposes, personal accounts. The House and Senate have guidelines regarding the differences between accounts, which mostly concern how legislators can spend official funds. As reelection-minded officeholders seek votes when communicating with the public in any venue, we include both account types. Other research often only uses campaign accounts. Our findings are robust to specifications that only include those accounts, with two exceptions. The Competition variable in Table 4, Model 3 is not statistically significant and as are our Incumbent and Long shot findings reported in Table 5. We could not verify the ages, a key control variable, for 15 candidates. As such, our effective sample size is 882.

11. The issue-based keyword list included 70 terms. The mobilization keyword list included 28 terms. Both lists are included in the Appendix, in Tables A and B.

12. A common tweet type not included in either category are posts that document a candidate’s campaign.

13. Less than 5% of tweets are classified as aggressive attacks.

14. The Cook Political Report rates races as Tossup, Lean Democrat or Republican, Likely Democrat or Republican, and Solid Democrat or Republican. In our measure, Tossup is the highest value (4), followed by Lean races (3), Likely races (2), and Solid races (1).

15. The baseline category is a governor’s race.

16. This alerts the other user you are mentioning them and the tweet is included when someone searches for the tagged screen name. This is one way a candidate can attack their opponent or other politicians. We discuss potential limits to our approach below.

17. We paired candidates running against one another and searched their Twitter feeds for instances in which they tagged their opponent.

18. Our results are robust to fractional logistic regression models.

19. Positive mobilization tweets increase 0.04% as a proportion of all tweets as competition increases one-level in our 4-point measure.

Additional information

Notes on contributors

Jeremy Gelman

Dr. Jeremy Gelman is assistant professor of political science at the University of Nevada, Reno. His research examines legislative partisanship and agenda-setting in the United States.

Steven Lloyd Wilson

Dr. Steven Lloyd Wilson is an assistant professor of politics at Brandeis University and co-PI of the Digital Society Project. His research focuses on comparative democratization, cyber-security, and the effect of the Internet on authoritarian regimes, particularly in the post-Soviet world.

Constanza Sanhueza Petrarca

Dr. Constanza Sanhueza Petrarca is a Research Fellow at the WZB Berlin Social Science Center. Her research examines democracy, representation and electoral behaviour.