With social media platforms serving as soapboxes for politicians, campaigns and voters alike, pollsters may be able to better predict election results by tracking the buzz around each candidate on social media, according to a new study published by a team of researchers from Penn State, Stanford University and the University of Florida.
The findings, published Nov. 21 in Social Media+Society, suggest that the social media chatter surrounding each candidate may be a useful indicator of public sentiment. By analyzing the total daily number of mentions of candidates — referred to as “bandwagon” — in posts on X, formerly named Twitter, researchers said they could have predicted the outcomes earlier in the campaign season in close senate races of 2016 and 2018.
“We tracked the bandwagon surrounding each candidate by counting the total number of times their name was mentioned on Twitter, before it was renamed X, and found that for a majority of the races that we examined, both the volume and the sentiment of mentions were able to predict election results early on in the campaign season,” said S. Shyam Sundar, Evan Pugh University Professor and the James P. Jimirro Professor of Media Effects at Penn State and co-author of the study. “While X has changed recently and mentions on it may not continue to be as accurate for making predictions in the future, we can look at other social media platforms to track posts, using similar methods for upcoming elections.”
According to Sundar, traditional polling often relies on cold-calling registered voters, and it is increasingly harder to get people to participate and obtain a good sample.
“There is an urgent need to gauge public sentiment that does not rely solely on phones, as polls are becoming more unreliable in close races,” Sundar said. “This work adds more data and can help us to predict close races.”
On social media platforms such as X, popularity metrics are automatically generated to users through likes and shares of the posts. The researchers assembled their dataset from the Twitter Premium API, which provides access to all non-deleted public tweets since 2006, enabling data collection based on specific search keywords.
To analyze tweets, the research team developed a modeling tool to identify the point in time at which bandwagon support for competing candidates begins to diverge. The researchers then chose one competitive U.S. Senate race in 2014 for testing their methodology before implementing it for eight close races in 2016 and 2018. They collected daily tweet totals during the election cycle starting on Sept. 1, using hashtags specific to each election. They analyzed over 800,000 tweets for both sentiment and volume bandwagons. Volume here refers to the size of the bandwagon, or the number of posts, surrounding a candidate. Sentiment bandwagon refers to the content of the posts.
“Sentiment bandwagons performed slightly better than volume bandwagons in predicting election outcomes. By combining these two metrics, we could predict seven out of the eight elections in our sample,” said lead author Jinping Wang, assistant professor of journalism and communications at the University of Florida who earned her doctorate at Penn State.
In other words, the quality — not simply quantity — of posts about a candidate could better indicate their election outcome, according to the researchers.
“While negative mentions may be better than no mentions at all in terms of creating buzz, our data showed that the nature of mentions does make a difference, and the more positive mentions, the better,” said Sundar, who also serves as co-director of the Media Effects Research Laboratory. “It isn’t just about the number of tweets. Our words make a difference, and calling out misinformation associated with a candidate doesn’t perpetuate it.”
The researchers compared their results to polling data from sources such as Marist College, YouGov and Emerson College taken during the same timeframes.
“We were surprised to find that we could detect these bandwagon trends back to September in seven out the of the eight election cycles we looked at, and that our analysis was consistent with election results,” Wang said.
The team also examined bot activity, or posts generated by automated programs that mimic human behavior on social media platforms. The researchers found that tweets generated by bots only accounted for 3% to 5% of posts, compared to about 20% in a presidential election.
“The bandwagon-based predictions were highly accurate and could be formed much earlier than polls,” saidNilàm Ram, professor of communication and psychology at Stanford University and part of the research team. “Although we could identify the eventual winner earlier, we’re not suggesting replacing polls with social media analysis, but simply offering another value indicator. In competitive political contests like these, additional data sources, such as social media metrics, might improve the quality of election predictions.”
Wang noted that social media is likely to play a bigger role among younger users.
“Most people get news and information on their phones, increasingly from social media as the importance of traditional media gradually disappears. The bandwagon predictions may be better for some populations,” she said.
Bandwagon predictions are also better for tracking the ebb and flow of support for a candidate, the researchers said.
“The reason why social media can be used to predict outcomes and sentiments early in an election cycle is because they provide daily snapshots unlike polls which tend to be less regular,” Sundar said.
While the researchers focused on political tweets in the run-up to U.S. senatorial elections, they said the demonstrated bandwagon effects and analytical approaches used in this project could be applied to several other areas for predicting and understanding larger social phenomena, including the spread of misinformation and trends in health-related phenomena such as vaccine uptake.
A grant from Penn State’s Social Science Research Institute supported this research.