Tactics and Tallies: A Study of the 2016 U.S. Presidential Campaign Using Twitter 'Likes'

We propose a framework to measure, evaluate, and rank campaign effectiveness in the ongoing 2016 U.S. presidential election. Using Twitter data collected from Sept. 2015 to Jan. 2016, we first uncover the tweeting tactics of the candidates and second, using negative binomial regression and exploiting the variations in ‘likes,’ we evaluate the effectiveness of these tactics. Thirdly, we rank the candidates’ campaign tactics by calculating the conditional expectation of their generated ‘likes.’ We show that while Ted Cruz and Marco Rubio put much weight on President Obama, this tactic is not being well received by their supporters. We demonstrate that Hillary Clinton’s tactic of linking herself to President Obama resonates well with her supporters but the same is not true for Bernie Sanders. In addition, we show that Donald Trump is a major topic for all the other candidates and that the women issue is equally emphasized in Sanders’ campaign as in Clinton’s. Finally, we suggest two ways that politicians can use the feedback mechanism in social media to improve their campaign: (1) use feedback from social media to improve campaign tactics within social media; (2) prototype policies and test the public response from the social media. Introduction Twitter is playing an important role in connecting the presidential candidates with voters (Sanders 2016). Between September 18, 2015 and January 23, 2016, Hillary Clinton posted 1316 tweets, Bernie Sanders 1698 tweets, Donald Trump 2533 tweets, Ted Cruz 1309 tweets, and Marco Rubio 908 tweets.1 These tweets constitute a valuable data source because they are explicitly political in nature, they are many, and, importantly, they carry feedback information from the voters in the form of ‘likes.’ In this paper, we solve two problems. We first study the tweeting tactics in the tweets: we analyze who are mentioned in these tweets and what issues are raised. We then use negative binomial regression to evaluate the effectiveness of these tactics by exploiting the variations in ‘likes,’ which we refer to as tallies. Our study focuses on the five Copyright c © 2017, Yu Wang ([email protected]), Xiyang Zhang, Jiebo Luo. All rights reserved. We do not count retweets, as retweets do not have as a feature the number of ‘likes.’ Figure 1: Selected tweets Donald Trump (R) posted on February 9th. leading candidates: Hillary Clinton (D), Bernie Sanders (D), Donald Trump (R), Ted Cruz (R), Marco Rubio (R).2 Figure 1 shall illustrate our points well. It shows three tweets that Donald Trump posted on February 9th, 2016, all of which are political in nature. The first tweet talks about drugs. The second tweet raises the issue of ObamaCare and points towards President Obama (D).3 The third tweet is about the ISIS. Trump supporters responded to these three tweets differently, assigning to the third tweet the most ‘likes’ and to the second tweet the fewest ‘likes.’ By connecting topics with responses, we are therefore able to infer the effectiveness of the tactics. Our approach has a distinct advantage over polls. IndividThis selection is based on both polling results and on candidates’ performance in the Iowa caucus. Throughout, we follow the convention that Republican candidates are marked with (R) and Democratic candidates are marked with (D). ar X iv :1 70 4. 02 04 2v 1 [ cs .S I] 6 A pr 2 01 7 uals surveyed in polls may not actually turn out to vote and even if they do vote they might change their mind and vote differently. By contrast, our data is built on strong revealed preference. Individuals voluntarily express their preferences, and thus the estimates we obtain are more reliable. We organize our paper as follows. Section 2 presents related work. Section 3 introduces the dataset US2016 and the methodology. Section 4 presents the tactics we uncover from the tweets. Section 5 evaluates these tactics. Section 6 ranks the campaign effectiveness of the five candidates. Section 7 concludes. Related Work Our work builds upon previous research in electoral studies in political science and behavioral studies based on social media. Political scientists have long studied the effects of campaigns and public debates. Many studies have found that campaign and news media messages can alter voters’ behavior (Riker 1986; Iyengar and Kinder 1987). According to Gabriel S. Lenz, public debates help inform some of the voters about the parties’ or candidates’ positions on the important issues (Lenz 2009). In our work, we assume that tweets posted by the presidential candidates reveal their policy positions in various dimensions and that supporters reveal their policy preference by deciding whether or not to ‘like’ the tweets. In a related strand, researchers have found that gender constitutes an important factor in voting behavior. One common observation is that women tend to vote for women, which political scientists refer to as gender affinity effect (King and Matland 2003; Dolan 2008). In our work, we test whether the women issue is emphasized in Hillary Clinton’s campaign and evaluate its effectiveness. There are also a large number of studies on using social media data to analyze and forecast election results. DiGrazia et al. (DiGrazia et al. 2013) find a statistically significant relationship between tweets and electoral outcomes. MacWilliams (MacWilliams 2015) suggests that a candidate’s number of ‘likes’ in Facebook can be used for measuring a campaign’s success in engaging the public. According to Williams and Gulati (Williams and Gulati 2008), the number of Facebook fans constitutes an indicator of candidate viability. According to O’Connor et al. (O’Connor et al. 2010), tweets with sentiment can potentially serve as votes and substitute traditional polling. Gayo-Avello, Metaxas and Mustafaraj (Gayo-Avello, Metaxas, and Mustafaraj 2011), on the other hand, report unsatisfactory performance of such predictions and advocate that scholarly research should be accompanied with a model explaining the predicative power of social media. Substantively, our paper is closely related to two existing studies of the 2016 U.S. presidential election. Wang, Li and Luo (Wang, Li, and Luo 2016b) study the growth pattern of Donald Trump’s followers. Wang, Li and Luo (Wang, Li, and Luo 2016a) use Twitter profile images to study and compare the demographics of the followers of Donald Trump and Hillary Clinton. The authors also study the effects of public debates on the number of Twitter followers. Our work uses both the number of candidate followers (as a control variable) and the number of ‘likes’ (as the dependent variable). Our contribution is to infer tactic effectiveness from these ‘likes.” There are also quite a few studies modeling individual behaviors on social media. Lee et al. (Lee et al. 2015) model the decision to retweet, using Twitter user features such as agreeableness, number of tweets posted, and daily tweeting patterns. Mahmud, Chen, and Nichols (Mahmud, Chen, and Nichols 2013) model individuals’ waiting time before replying to a tweet based on their previous replying patterns. Our study models the number of ‘likes’ that candidates’ tweets receive. Our innovation is to use tweet-specific features instead of individual-specific features, as is done in the abovecited literature. A closely related work is Wang, et al. (Wang et al. 2016), which uses Latent Dirichlet Allocation (LDA) to extract tweet topics and models supporter preferences. In this paper, we reply on domain knowledge and search for specific topics using predefined keywords. As a result, we will have definitive labels. Data and Methodology We use the dataset US2016, constructed by us with Twitter data. The dataset contains a tracking record of the number of followers for all the major candidates in the 2016 presidential race, including Hillary Clinton (D), Bernie Sanders (D), Donald Trump (R), Ted Cruz (R), and Marco Rubio (R). The dataset spans the entire period between September 18th, 2015 and January 23th, 2016, and covers four Democratic debates and four Republican debates. Dependent variable Our dataset US2016 contains all the tweets that the five candidates posted during the same period and the number of ‘likes’ that each tweet has received. In Table 1, we report the summary statistics of the dependent variable: ‘likes.’ Table 1: Summary statistics Variable Mean Std. Dev. Min. Max. N Hillary Clinton 1702.454 1682.262 12