In a heated, polarized U.S. election season, campaigns, pollsters, and political scientists dedicate significant resources to efforts to predict voter behavior. Conventional wisdom holds that demographic factors reliably signal partisanship; this means sorting potential voters by race, age, gender, and economic class in order to guess how they will vote.

However, recent work by SPA Assistant Professor Seo-young Silvia Kim (coauthored with Jan Zilinsky of the TUM School of Social Sciences and Technology in Munich) suggests that this “social sorting” does little to predict partisanship and the outcome of elections. In “Division Does Not Imply Predictability: Demographics Continue to Reveal Little About Voting and Partisanship,” published this August in the journal Political Behavior, Kim used models trained on demographic labels from a half-century of public opinion surveys (1952-2020) to determine how well these factors predict partisanship and voting decisions.

“Social sorting is not about ideology,” said Kim. “It's this idea that the link between individual-level characteristics and partisanship is tightening. I think a lot of people simply assume that your demographics are your identities.”

Kim suggests that the media has clung to these relationships to create content on political races (i.e., ‘Black voters prefer Democratic candidates’ and ‘non-college educated white men vote for Trump’), and campaigns seize the opportunity to easily target voter outreach, creating persistent stereotypes regarding voter behavior.

“A lot of campaigns focus on trying to advance their hold of particular demographic groups, because this [e.g., voter age, race, or origin] is readily-available information,” she continued. “Also, every exit poll tries to cover the demographic breakdown of partisanship, [which] shapes this idea that demographics determine your partisanship.”

Kim’s findings, however, refuted this trope. While random guessing alone would predict with 50% accuracy, this carefully trained model could only accurately predict a meager 63.9% of two-party vote choices and 63.4% of partisan IDs, short of the 80-90% accuracy range needed to say with certainty that demographics influence voter behavior.

“Normally, such models work fairly well, but this one performed pretty poorly,” she explained. “You can't guess [voter behavior] very accurately if the only data you have is demographics, especially since we tend to look at only one particular characteristic. We should separate our judgment about individual partisan affiliations from group-level statistics that partisan campaigns and the media care about. I think it's a dangerous idea that you can scan a person and then figure out whom they're going to vote for.”

The study also tested how well partisanship predicted vote choice: this more reliable factor became a better predictor over time, reflecting increased polarization. This polarization, Kim said, prevents civil conversations about issues and erodes trust in the political process, leading, eventually, to the extreme behavior and threats of violence displayed in the most recent election.

“I think it's the political scientist's role to try to determine what exactly is going on,” she added. “Why are some of the functions that we relied on, and this idea that we could have a decent conversation––why is that going away? And what should we do about it?”

Kim pointed out that the public opinion surveys allowed respondents to self-report demographic factors like race and gender, which is complicated by how people identify themselves, and how they associate these identities with their public participation. She hopes to tease out these relationships in future studies.

“One explanation of why our research has provided contradicting results is that identities are not necessarily equal to labels,” she said. “How much emphasis do voters place on their social identities? I think that has a lot of nuances.”