Tweeting the Debates: Sentiment in Twitter Over the Three Presidential Debates

Written by Lisa Singh, Stuart Soroka, Michael Traugott and Frank Newport as a part of the Gallup Working Group.

As part of an ongoing collaboration between Gallup, Georgetown University and the University of Michigan, researchers have been gathering campaign-related tweets – both throughout the campaign, and in particular on debate nights.

This week marked the last of three presidential debates scheduled during the 2016 election season. The debates received very large, even historic, viewership. They also featured subjects and a temperament that most commentators regard as fundamentally different from past election debates. Gallup data following the first debate suggested that a majority of respondents believed that Clinton won. Even so, a sizable minority of viewers believed that Trump did well; and there have been continuing, and divided, discussions of both leaders’ performances in the second the third debates.

Tracking the Tone on Twitter

Twitter data offer a unique, real-time perspective of viewers’ reactions to debate content. The research team’s focus for the current analysis is on the tone, or sentiment, of election- and debate-related tweets during each election debate. Tone is captured automatically using both a domain-general lexicon and an emoji sentiment lexicon. The researchers then used the hashtags #neverhillary/#VoteTrump and #nevertrump/#VoteHillary to determine words systematically related to pro-Trump and pro-Hillary discussion; finally, this combined set of words and emojis are used to determine the tone of all tweets.

The researchers were able to assign a positive or negative tone to nearly 80% of roughly 4 million tweets throughout the evenings of the three debates, including nearly 2.5 million tweets in the combined three 90-minute debates themselves. These were all tweets that contained one of the debate related hashtags that Twitter was promoting, e.g. #debates. The data was collected using Twitter’s stream API. Combining tweets over five-minute intervals and then subtracting the proportion of pro-Trump tweets from the proportion of pro-Clinton tweets produces a relatively simple measure of ‘Candidate Advantage’.


Tracking ‘Candidate Advantage’ on Twitter for each debate. The measure captures candidate advantages in both directions. Because the researchers subtract Trump tone from Clinton tone, when the measure moves below zero, that indicates that Trump-related tweets reflect a tone that is better than Clinton-related tweets. When the measure moves above zero, that indicates that Clinton tweets are more positive than Trump tweets.

The figure presented here shows smoothed trends in Candidate Advantage across each of the three debates. The line is blue when Candidate Advantage leans towards Clinton, and red when Candidate Advantage swings towards Trump. Results suggest that most of the time during the debates, tweets about Clinton reflected a more positive tone than tweets about Trump. Indeed, only in the last half hour of the second debate is there any period when Trump tweets were systematically more positive than Clinton tweets.

Sentiment Trends Over Time

There are some interesting commonalities across debates. First, tone on Twitter always trends towards Trump towards the end of debates; in his appearances, he was a strong finisher. In the first and third debates, Candidate Advantage moves towards Trump over the last half hour; in the second debate, Trump simply dominates during this final period. There also appears to always be a turning point at about the 50-minute mark. This was a peak for Clinton in the 2nd debate, and a trough for her in the first and third debates.

Note that these trends are easily linked to actual debate content. Consider the second debate, the one in which there is the greatest variation in candidate advantage over time. Around 9:10-9:15, Trump tone is improving. This is the point at which he is discussing ‘locker room talk,’ and suggesting that Bill Clinton is the worst abuser of women.  9:45-9:50 is Clinton’s best point in the debate – the point at which the gap between the tone of Clinton and Trump tweets are at their greatest. Clinton arguing that Trump should release his taxes, and that he should change his position on Russia. Clinton continues to dominate for a little longer, until about 10:10.  The end of the debate marks Trump’s best moment – the point at which the candidates are discussing things they respect about each other.

What about the third debate? The tone of tweets leans in favor of Clinton for the entire debate. Trump is strongest when he is talking about his “phenomenal company,” and comparing his success with what he argues was Clinton’s failure with ISIS. The balance shifts back in favor of Clinton over the next ten minutes, however, as the topic changes to the candidates’ fitness to lead the country, and charges of sexual harassment against Trump.

There won’t be any more presidential debates this election season. The trends shown here do more than confirm Clinton’s dominance on the debate stage, however – they also point towards arguments that may be most effective in the final weeks of the campaign. Trump is at his weakest and Clinton at her strongest when the focus is women, minorities, and taxes; the reverse is true when the focus is on business success or ISIS. These are arguments we might expect to see dominate in the closing weeks of the campaign.

Related Reading: Tracking the Tone of the Campaign by Stuart Soroka and Political Persuasion by Alex Piazza at the University of Michigan Office of Research.

The Georgetown team that supports this project also includes Chris Kirov, and Yanan Zhu.

Posted in sentiment, Twitter Tagged with: , ,