A final, data-driven look at Trump v. Clinton on social media
By Jennifer Stromer-Galley
After the surprise win by Donald J. Trump, several postmortems have contemplated how Trump elevated Twitter to drive attention to his candidacy. A scholarly forum in the journal Political Communication highlights how the news media’s attention to Trump’s tweets helped propel him to victory during the primaries.
But, how did his use of Twitter and Facebook compare to Hillary Clinton’s?
I offer here a different kind of post-mortem, based on my work on the Illuminating 2016 project. Over the course of the presidential campaign, our team collected all of the Facebook and Twitter messages by the 17 Republican and five Democratic candidates. We also analyzed the comments left on their Facebook posts.
Surprisingly, Clinton’s campaign was overall more active on Twitter and on Facebook than Trump’s, generating 19 percent more messages (11,475 messages by Clinton to 9,390 by Trump). On Facebook, Clinton generated 500 more messages than Trump. While Trump’s tweets seemed to garner more news coverage, Clinton’s campaign was actually substantially more active on social media, generating 25 messages a day on average to Trump’s 20.
Yet, Trump’s social media following was larger than Clinton’s. In November 2015, Clinton had 1.7 million followers on Facebook. By Election Day that had grown to 8.4 million, a 394 percent increase. Trump had 4.2 million Followers on Facebook in November 2015. By Election Day, that number jumped to 12.35 million, a 194 percent increase. So, while Clinton saw a greater increase, Trump still had nearly 4 million more followers.
We also examined share of voice on Twitter — the volume of activity on each campaign’s main Twitter hashtag. For Clinton’s campaign, that key hashtag was #Imwithher. For Trump, it had been #makeamericagreatagain, shortened over time to #MAGA (See Figure 1). We also track the amount of engagement with the candidate’s twitter handle. By these two measures, during the last four months of the campaign, there was more social media buzz about Trump than about Clinton (See Figure 2).
Figure 1: Share of Voice On Trump’s and Clinton’s Twitter Hashtags
Figure 2: Volume of Tweets that Mention the Candidates’ Twitter Handles Over the Last Four Months of the Campaign
We also collected and categorized the public comments on the two campaigns’ Facebook walls. We looked at the amount of talk on the candidate’s wall in support of and attacking the candidate. Overall, we found that on Donald Trump’s wall, there were more posts in support of him than opposing him. By contrast, almost half of the comments on Clinton’s wall attacked her.
All of this suggests that while Clinton’s campaign was overall more active on its social media accounts, it did not receive the same amount of attention and support on social media as compared with Donald Trump.
As we collected messages, our team also put them into several categories. The categories we developed were generated in part by looking at prior research on political campaign television ads. These categories focus primarily on the persuasive messages by the campaign: advocacy, which articulate support for the candidate; attack, which challenge or criticize an opponent; image, messages about the candidates character or personality; and, issue, messages about the candidate’s policy positions or values.
We then added to those categories messages that are somewhat unique to social media. The categories include: calls-to-action, which ask people to do something; informative, which provide basic information about the campaign without urging action.
We categorized these messages using human-supervised machine learning. Our team initially categorized a sample of candidate messages, and then used a machine-learning tool called Scikit Learn. It looks at the structural features and vocabularies, looking for commonalities in messages of the same category, and inductively generates an algorithm based on those commonalities. We also added additional features to the algorithm, such as political party, to further improve the accuracy of the algorithm. After rounds of further improvements on the algorithm by providing corrections to incorrectly categorized messages, our final algorithm performs as well or better than the humans who originally categorized candidate messages. Overall, our categorization is over 75 percent accurate on all of the categories that are discussed in the remainder of this blog.
The Surfacing Stage
Early in the election season, during the surfacing stage (August 1–December 31, 2015), there were noteworthy differences in the messaging styles of Trump and Clinton in terms of the volume of attack and advocacy messages.
First, Trump attacked his opponents in the Republican Party for their policy positions and on their character and personality twice as much as Clinton (Figure 3). He also advocates for himself at a higher rate than Clinton (Figure 4).
The volume of attack at this stage by Trump is noteworthy. Generally in the surfacing stage, presidential candidates stay focused on advocating for themselves. Part of the work in this stage is to introduce themselves to the American people. Most candidates have low name recognition, which means that their first job is just to get people to know their name. This was one of the challenges for example for Marco Rubio, the junior Senator from Florida.
Trump, on the other hand, did not have to generate name recognition. He was already a household name across America as a prominent businessman with a visible chain of hotels, apartments, and corporate buildings emblazoned with his name and host of The Apprentice. Thus he was able to start attacking his opponents, using his star power and the visibility that came with it to try and bring down some of his key opponents, such as Jeb Bush — the party establishment’s presumptive nominee — from the start.
Figure 3: Trump and Clinton on the Attack in the Surfacing Stage
Figure 4: Trump and Clinton Advocating for their character and ability to lead
Trump also used social media heavily to promote his image — or, in marketing terms, his brand. During the surfacing stage, Trump’s messages are overwhelmingly image focused — with 84 percent of his persuasive messages focused on image over policy. Bush, by contrast, was more policy-focused, with only 43 percent of his persuasive messages focused on his image. For Rubio it was 50 percent, and for Cruz 60 percent.
On the Democratic side, only 46 percent of Clinton’s messages were image focused, similar to Jeb Bush. Her competitors on the Democratic side produced the same amount of persuasive messages that were image focused. Only 40 percent of Bernie Sanders’ messages were image-focused, and Maryland Governor Marin O’Malley was also more issue than image focused, with only 46 percent of his messages about his character and abilities to lead.
It is important to look at Facebook and Twitter separately to see how the campaigns use these two applications differently. Twitter’s user base is much smaller than Facebook’s, with only about 23 percent of the American public that is online using the service. Twitter is where journalists, celebrities, and elites tend to reside. It is a public application, with little hidden behind privacy settings. As such, research has found that Twitter can be an important application for politicians and elites to try and set the news agenda. Facebook now has about the same reach as television, with 68 percent of internet users having a Facebook account. Facebook allows much longer posts, and campaign posts can have wide reach on the social media application given its large, heterogenous user base.
In comparing Twitter and Facebook during the surfacing period, we see that actually the campaigns used the two applications in distinctly different ways. On Twitter, Trump and Clinton both focused more of their messages on attacking opponents and advocating for their positions, but on Facebook, there was more emphasis placed on urging their supporters to action, such as encouraging them to buy a Trump Make America Great Again hat (see Figure 9) and generally advocating for their candidacies (See Figures 5–8).
Figure 5: Clinton’s Messages on Facebook During the Surfacing Stage
Figure 6: Clinton’s Messages on Twitter During the Surfacing Stage
Figure 7: Trump’s Messages on Facebook During the Surfacing Stage
Figure 8: Trump’s Messages on Twitter During the Surfacing Stage
Figure 9: Trump Urging Supporters to Buy a Hat on Facebook
During the primaries, the campaigns changed what they said on social media. Calls-to-action doubled on Trump’s social media, increasing from 6 percent in the surfacing stage to 13 percent in the primaries stage. Clinton’s campaign shifted little in terms of the volume of calls-to-action — but it included an increased number of calls for digital engagement (See Figure 10). Digital engagement messages urge people to take action on behalf of the campaign online, such as retweeting a message or watching an online video (See Figure 11).
Figure 10 Clinton vs. Trump on calls for digital engagement during the surfacing and primary periods
Figure 11 Example of Digital Engagement
On encouraging supporters to vote, there were distinct differences between Clinton and Trump. Although both candidates were involved in highly competitive primaries, Clinton was almost twice as likely as Trump to urge people on Twitter and Facebook to vote during the month of February (See Figure 12). In February it became clear that Vermont Senator Bernie Sanders would be a formidable opponent. This month was equally competitive for Trump, with a crowded field of 17 Republican candidates vying for strategic advantage in the early voting states, such as Iowa, New Hampshire, and South Carolina. Trump’s calls to vote remained fairly steady throughout the primaries, averaging a little under 40 a month on both Facebook and Twitter through May.
Figure 12: Clinton Vs. Trump Urging People to Vote
We also categorized calls to action to identify when campaigns urged supporters to engage in activities online and when they urged activities offline. Recall that Trump had large, sometimes raucous campaign rallies. Yet, while this was a prominent feature of his campaign, it was not widely touted on his social media accounts nor other types of offline engagement (See Figure 13). Clinton’s calls to action ranged from urging people to attend campaign events to seeking interns to help coordinate the campaign (See Figure 14).
Figure 13: Traditional Engagement
Figure 14: Clinton Seeks Interns In New York Via Twitter
Hillary Clinton and her campaign shifted from doing more advocating than attacking in the surfacing stage to attacking more during the primaries. Her volume of attacking surpasses Trump’s when you look at Facebook and Twitter combined (See Figure 15). The style of their attacks, though, is different. As I argued on The Conversation, Trump’s attacks tended to be more personal, calling his opponents names and attacking their character. Clinton’s attacks were more focused on policy and differences in perspectives. Even when she was critical of her opponent’s personality, she did not use pejoratives.
Figure 15: Trump and Clinton on the attack in the primaries on Social Media
But, as we saw in the surfacing stage, the campaigns used Facebook and Twitter differently. Looking at Facebook only, Clinton was less likely than Trump to attack her opponents, but going into June, Trump became substantially more likely to attack than Clinton (See Figure 16). Part of this shift can be explained by the fact that it became clear by June that Clinton would be the Democratic nominee. At this point, Trump began to attack Clinton aggressively, while Clinton also shifts to attacking Trump on Facebook, but not at the same rate.
Figure 16: Trump and Clinton on the Attack on Facebook
During the primary phase, Trump’s messages are no more substantive than during the surfacing stage in terms of articulating the policies that would drive his presidency. Of Trump’s messages that are persuasive messages (attack + advocacy on issue) 84 percent either attacked the image of his opponent or advocated for his own qualities and character to lead the country. By contrast, Clinton’s image talk is 53 percent of her persuasive messages, and she was slightly more likely to advocate for her image over attacking her opponents’ image.
During this period, Trump also became more policy-focused on Facebook than on Twitter, but Clinton did not. On Facebook less than half (47 percent) of Clinton’s persuasive messages are about the issues, but on Twitter 58 percent are on the issues. By comparison, 30 percent of Trump’s persuasive messages on Facebook are on the issues, and that figure drops to 24 percent on Twitter. Thus, Clinton is more policy focused on Twitter than on Facebook, while it’s the opposite for Trump.
When the campaign turns to the general election phase, roughly July 1 until Election Day, Trump and Clinton were now the only remaining major party candidates and squarely focused on campaigning against each other.
In the last months of the campaign, generally the focus shifted to voter registration and then get-out-the vote efforts. Social media can be a useful starting place for helping give supporters events and activities to do to be part of the campaign and to help with the effort of winning the election. Although both campaigns, indeed, increased their calls-to-action in the last two months of the campaign, Clinton beat Trump in volume of such messages on Facebook and Twitter, producing a third more call-to-action type messages (See Figure 17). If we only look at Facebook, however, Trump’s campaign produced as many call-to-action type message as Clinton in October.
Figure 17: Trump vs. Clinton Calls to Action on Facebook
When it came to asking people to vote, the Clinton campaign produced more than twice as many messages asking for people to vote on election day on the two platforms (See Figure 18), but most of that was on Twitter. On Facebook, both campaigns urged people to vote at the same rate, but on Twitter, Clinton’s campaign produces three times more appeals for votes than does Trump.
Figure 18: Clinton Vs. Trump Calls to Vote
Although one of the running themes of the Trump candidacy was that he was a highly negative candidate on social media, in actuality, Clinton was substantially more negative. The volume of attack messages by Hillary Clinton increases dramatically in this final phase, reaching a peak of attack messages in October.
Indeed, if we look over the entire course of the campaign, from August of 2015 until Election Day in November of 2016, we see that Clinton became substantially more negative, and in the month leading up to the general election she produced more than double the attacks on Trump than Trump on Clinton (See Figure 19).
Figure 19: Trump vs. Clinton: Attack Messages Over the Entire Campaign
Coming into the general election, the two campaigns continued to devote more of their social media messages to talking about their image rather than the issues. Clinton’s campaign increased the volume of image-focused messages to 65 percent, with nearly twice as many of these being attacks on Trump’s image. By comparison, while 65 percent of Trump’s persuasive messages are also focused on image, half of those advocated for his character and ability to lead.
One other thing to note: Trump’s campaign in the early, pre-primary, actively engaged in conversation with his supporters on Twitter (See Figure 20). That engagement dropped precipitously when voting in the primaries began. There’s an uptick in May, but during the general election there were few interactions. As I report in my book, Presidential Campaigning in the Internet Age (2014), candidates and campaign staff generally do not engage in interaction online with the public. They don’t for a variety of reasons, including a lack of staff and resources to respond and engage with the massive volume of messages, and a feeling that such direct interaction does not pay off in terms of fundraising, engagement, or votes. Given limited resources, campaigns just don’t see much payoff in direct interaction. But, Trump, unlike any of his other opponents during the primaries or general election, seemed to have enjoyed responding and interacting with supporters. But, as his campaign grew and he took on more professional staff, that interaction fell off.
Figure 20: Clinton and Trump Engaging in Conversation on Twitter
Social Media and the 2016 Campaign
The pattern of communication on social media supports the perspective that Trump’s cult of personality, his image-focused campaign, and his selective attacks on his opponents were more compelling and effective than Clinton’s. Although Trump did not use his social media to engage his supporters to get involved in his campaign to the same extent that Clinton did, his novel and aggressive messaging seems to have played well to a segment of Internet users that shared and retweeted his messages at a greater rate. Clinton’s efforts to bring Trump down by attacking him more aggressively in the general election did not earn her the same kind of buzz and support on social media as Trump’s self-aggrandizing style.
Of course, Trump is president because of the Electoral College, Clinton actually earned more votes during the general election than Trump. Her social media was more focused on urging people to vote and on traditional and digital engagement efforts by her supporters. Although I cannot make a direct causal relationship, there is some suggestion here that candidates who put energy on social media to getting their supporters working for the campaign has payoff. The power of social media is in directly and repeatedly engaging supporters. Facebook, in particular is an important application for candidates to energize their base and mobilize them to pitch in and help their candidate win.
But it is striking how negative the campaign was from the very start. Trump went on the attack at the start of the campaign, which is unusual. Generally, candidates don’t begin to attack until closer to the vote. Generally, early in the campaign, they spent more of their time touting their policies and their ability to lead. But, this research focused on television advertisements. Perhaps in the social media age, negativity and personal attacks are more the norm; Trump capitalized and amplified what most of us on Facebook and especially Twitter routinely experience.
Jennifer Stromer-Galley is a fellow at the Tow Center for Digital Journalism and Professor in the School of Information Studies and Director for the Center for Computational and Data Sciences at Syracuse University.