INSTITUTE FOR POLICY RESEARCH

New Tools for Studying Bias and Belief

IPR researchers share methods for discovering hidden patterns in our social attitudes

In the Social Cognition and Intergroup Processes (SCIP) lab, IPR graduate research assistant Jonathan Doriscar reviews a project with IPR psychologist Sylvia Perry.

People are messy. We say one thing and do another. We have principles but make exceptions. We change.

For Jonathan Doriscar, an IPR graduate research assistant pursuing a PhD in social psychology and a master’s in statistical methods and data science, that complexity is frustrating—but mostly fascinating. And it’s why he believes researchers need a diverse toolkit to understand human behavior.

His new paper published in Social Cognition, whose co-authors include IPR psychologist Sylvia Perry and Kellogg social psychologist and IPR associate Tessa Charlesworth, offers a guide for researchers who want to use computational methods to make sense of a messy and fascinating world. Those insights could inform efforts to reduce prejudice, improve education, address political polarization, and more.

Investigating the Intersections

Doriscar studies social cognition, which he describes as “the process through which we interpret and move through the social world.” Humans are constantly navigating multiple intersecting situations, stimuli, and relationships.

“People don’t just hold one belief at a time,” Doriscar said. “Social beliefs interact. They reinforce each other. Sometimes they conflict with each other. So the meaningful unit to study is not necessarily one attitude, but the intersection of several attitudes.”

Traditional statistical approaches typically begin with a specific outcome in mind. In social cognition, these approaches often include classic statistical models, such as regression, that test whether one variable predicts another. Researchers might ask, for example, whether political ideology predicts racial bias.

If researchers use machine learning for prediction questions like these, they tend to use what’s called supervised machine learning. “Supervision” in this context refers not to the researcher, but to the presence of a pre-specified outcome that the algorithm learns to predict.

Unsupervised machine learning (UML) works differently. Instead of starting with a target, researchers let the data speak more freely. They analyze large datasets without a predefined outcome to see what patterns, groupings, or relationships naturally emerge.

Revealing Hidden Patterns

“Social science datasets are larger and more complex than ever before. Our tools need to be able to map out and explore that new complexity, even if the map surprises us,” said Charlesworth, the senior author on the paper. UML offers the tools to uncover insights that might otherwise remain hidden.

In their paper, Doriscar and his colleagues walk researchers through four specific UML methods and when to use them, and link to R code, data files, and video tutorials. Two methods, K-means and DBSCAN, identify clusters within the data. Another pair, principal component analysis and market basket analysis, uncover structure in complex datasets by highlighting what tends to vary together and what tends to show up together.

The authors demonstrate the power of these methods with examples from large-scale projects looking at racial attitudes. Rather than simply predicting bias from ideology, they apply UML methods to identify nuanced psychological profiles.

For example, using K-means clustering, they identified five groups of respondents, each defined by a particular mix of political ideology and racial attitudes. Some clusters fit familiar narratives. Others did not.

One expected cluster leaned liberal and reported warm feelings toward Black people, but showed implicit bias favoring White people in their Implicit Association Test responses. That combination aligns with what prior research has described as “aversive racism,” when someone believes themself to be egalitarian but nonetheless harbors bias.

More surprising was another cluster that leaned slightly liberal but reported relatively cold explicit attitudes toward Black people—attitudes the team expected to be concentrated mostly among conservatives.

Such findings capture more nuance than broad generalizations like “liberals tend to feel warmer toward Black people.” They reveal variation within groups, including thorny contradictions.

That nuance has practical implications. If researchers or policymakers are designing interventions to reduce prejudice, different psychological profiles may require different approaches.

“Perhaps we need to target these people differently, given that there is heterogeneity,” Doriscar said.

Complementary Approaches

Doriscar stresses that unsupervised and supervised machine learning are “two approaches that don’t have to be siloed.” UML does not replace theory-driven research or traditional statistical inference. In fact, combining the two can strengthen research.

One practical entry point for researchers? When a study doesn’t work as expected.

If a theory-driven intervention fails to show the predicted effect, researchers might be tempted to scrap the project. Instead, Doriscar suggests they probe deeper and ask what they might have missed. UML may reveal subgroups the intervention did work for, pointing to hidden variability that standard analyses miss.

Still, UML presents challenges. Diving into a dataset without a clear target can feel overwhelming. And not every pattern uncovered will fit neatly into existing theories.

“You may find patterns that are not guided by an underlying theory,” Doriscar said. “You may not be able to explain what the underlying mechanism is.”

That tension, he argues, is productive. His goal is to help researchers “weld” theory-driven and data-driven approaches and use them simultaneously “to extract higher-quality insight” that can be applied in schools and other real-world settings.

As scientists confront increasingly complex questions—from racism to political conflict to misinformation—tools like UML that embrace, rather than simplify, that complexity will become more important.

Tessa Charlesworth is an assistant professor of management and organizations and an IPR associate. Jonathan Doriscar is a PhD Candidate in social psychology, a master’s student in data science and statistics, and an IPR graduate research assistant. Sylvia Perry is an associate professor of psychology and an IPR fellow.

Photo credit: Rob Hart

Published: March 10, 2026.