Newly Released Ballot Data Finds Ticket Splitting Among Republican, Democratic Voters

In the 2020 presidential election, 1.9% of Republican voters in battleground states who supported Republican candidates for both Congress and state legislative seats split their ticket with a vote for Joe Biden, according to a new analysis funded in part by Yale’s Institution for Social and Policy Studies.

Among Democrats who were solidly supportive of lower-office Democrats in the same states, 1% split their ticket for Donald Trump.

“People often focus on anti-Trump Republicans, but this finding reminds us that there are Democrats who didn’t vote for Biden,” said ISPS faculty fellow Shiro Kuriwaki, who co-led the project with researchers at Yale; the Massachusetts Institute of Technology; the University of California, Los Angeles; Columbia University; and Harvard University. “At the same time, I had personally anticipated that more than 2% of solid Republican voters would have voted for Biden.”

The project, in press at a peer-reviewed journal, featured a unique dataset comprised of digital records of actual ballots cast by 42.7 million voters in 20 states voting for more than 2,200 candidates.

Voting districts regularly report vote totals for candidates, but rarely do researchers gain access to individual ballots. At the same time, more people are expressing concerns about voter fraud and election security, and academic evaluations can play a role in assessing those concerns.

The rise of computer-assisted voting machines opened a new avenue for investigators. The machines can produce a record of a scanned paper ballot, allowing for easier analysis of voter behavior, including how an individual, anonymous voter in a specific geographic location chose among candidates for various offices.

“For political scientists, it’s the only data set that measures in a precise way how people actually voted,” said Kuriwaki, who researches American elections. “Surveys are often good enough measures, but they are subject to questions of memory and honesty. They are not a complete census of the population. And they are inherently limited by the questions that a surveyor chooses to ask at the time.”

Following unfounded accusations of voting irregularities in the 2020 election, officials have fielded a flood of requests for such electronic files, known as cast vote records.

Kuriwaki and his collaborators, many affiliated with the MIT Election Data Science Lab, accessed these publicly available, unstandardized cast vote records and converted them into a standardized multi-state database. They then performed a detailed comparison between these totals and the certified election results. In validating the results across 23,000 voting precincts, vote tallies in the cast vote records matched the official, publicly available vote records in over 20,000 precincts, and nearly 22,000 matched within three votes.  

“The custom data processing and validation tools we use are innovative in the field, and they let us efficiently clean and release data on scales not previously possible”, said Mason Reece, a Ph.D. student at MIT and co-lead of the article with Kuriwaki. “We hope that both the data we’ve published and our custom tools enable new analyses and inspire future research.”

The researchers expected discrepancies because counties may choose to hand-count a batch of ballots or process them differently and exclude them from the electronic record. For example, some cast vote records did not include votes cast by mail or by provisional ballot.

“Cast vote records are not sufficient or necessary to prove an election was valid,” said Jeffery Lewis, a professor of political science at UCLA who collaborates with Kuriwaki.  “These records should be considered byproducts or an electronic trail of the ballot tabulator. That said, I would say it’s still quite reassuring that in these counties you could get datasets of millions of vote choices and reproduce the vote count exactly.”

In a paper from April, Kuriwaki, Lewis, and Michael Morse of the University of Pennsylvania, found that the release of such ballot records protect the anonymity of 99.8% of voters. To reveal any voter’s identity through the release of cast vote records would require a small precinct with either a completely unanimous voting result or the collusion of voters in the district intent on revealing someone else’s vote.

Co-authors on the ballot-splitting study included Lewis; Taran Samarth of Yale; Samuel Baltz, Joseph Loffredo, Kevin Acevedo Jetter, Zachary Djanogly Garai, Kate Murray, and Charles Stewart III of MIT; Aleksandra Conevska, Can Mutlu, and James Snyder Jr. of Harvard; and Shigeo Hirano of Columbia.

Kuriwaki said he would like to see more counties regularly release cast vote records. ISPS has funded a new project in which he is building an online dashboard to better disseminate voting data for the public and other researchers.

“For election administrators and the public in general, cast vote records provide additional transparency for how we conduct elections,” he said. “They are another important piece of evidence that votes were counted and recorded accurately.”