Introduction
People rapidly sort strangers into groups based on visible social traits (Dovidio et al. Reference Dovidio, Kawakami, Johnson, Johnson and Howard1997; Fiske and Neuberg Reference Fiske and Neuberg1990; Willis and Todorov Reference Willis and Todorov2006), including faces (Hugenberg and Sczesny Reference Hugenberg and Sczesny2006). Such inferences can prime in- and out-group processing (Bin Meshar, Stolier, and Freeman Reference Bin Meshar, Stolier and Freeman2022) through the use of stereotypical facial cues to infer partisanship (Rule and Ambady Reference Rule and Ambady2010) even in the absence of explicit political cues, such as campaign pins or hats. While stereotypical facial cues can shape partisan categorization, research finds that people interpret ambiguous emotions, like surprise, more negatively when the face belongs to an out-party member (Basyouni et al. Reference Basyouni, Harp, Haas and Neta2022). This suggests that clearly displayed emotions might further cue partisan categorization.
Most categorization research utilizes designs asking participants to make judgments under relatively high-information conditions. For example, conjoint experiments provide multiple attributes to isolate specific features (e.g., Norman and Green Reference Norman and Green2025), but everyday political impressions are often formed quickly with minimal information. We mimic low-information environments using three survey experiments to demonstrate that visible emotions cue political categorization. Our results demonstrate that negative emotions (fear and anger) are consistently associated with out-party faces, whereas positive emotions (happiness) are assigned to co-partisans.
Method and results
We fielded three experiments across two surveys to study categorization behavior in Canada. The first survey was fielded between June 19 and July 1, 2023 (N = 1,199) to a sample provided by Abacus Data. The second survey was fielded between August 28 and September 2, 2024 (N = 730) to a sample provided by Léger. Both surveys were administered with Qualtrics and included quotas for age, region, and gender.
Experiment 1 was fielded in both surveys,Footnote 1 twice per survey. Respondents were asked to “drag and drop” eight randomly assigned photos of faces from the Chicago Face Database (CFD) (Ma, Correll, and Wittenbrink Reference Ma, Correll and Wittenbrink2015)Footnote 2 into boxes corresponding to the Canadian political party they believed “the person would most likely support” (Figure 1). Facial stimuli were randomly assigned across participants, allowing us to isolate the causal effect of facial features on partisan categorization. Our outcome of interest is whether a face is assigned to the participant’s own party (co-partisan) or any other party (out-partisan). Respondents were not required to assign the faces to a party, reducing demand characteristics.
Task example (Experiment 1).
Experiment 2, fielded in the second survey, used the same randomly assigned CFD facial stimuli as Experiment 1. Participants were shown four faces and asked to select those they believed supported a given party leader, completing this selection separately for three major Canadian party leaders. The outcome of interest remains whether a face is assigned to an out-party leader. The distinction between political parties and party leaders adds nuance to our understanding of political categorization, as partisanship is often considered more flexible in Canada compared to other countries (Gidengil Reference Gidengil2022).
Figure 2 presents the Average Treatment Effects for Experiments 1 and 2, estimated using OLS regressions with robust standard errors clustered at the respondent level with 95% confidence intervals for both outcomes: (1) political parties and (2) the names of party leaders. The estimate highlights the impact of a given variable on the likelihood of being assigned to the out-party across the entire sample (top panels).
Average treatment effect of face characteristics on political categorization of citizens (Experiments 1 and 2).
Our results demonstrate that emotions are a significant driver of political categorization.Footnote 3 Participants are more likely to categorize faces with negative emotions, compared to neutral faces, into the out-group. When shown an angry face, participants are 9 and 8 percentage points (p < 0.001) more likely to categorize the face as an out-partisan. Similarly, fearful faces are 8 points (p < 0.001) more likely to be classified as out-partisans. When shown a happy face, respondents are less likely to categorize the face as out-partisan (5 points, p < 0.001), but this result is not statistically significant in the 2024 survey sample (p < 0.4). The pooled results for political parties are broadly consistent (right panel of Figure 2). Respondents are more likely to categorize angry (8 points, p < 0.001) and fearful (4 points, p = 0.01) faces as supporters of the out-party leader. Conversely, they are less likely to categorize happy faces as supporting the out-party leader (4 points, p = 0.003).Footnote 4
In an additional experiment, respondents were randomly assigned to complete the categorization task or a control condition before playing a partisan trust game. Neither task exposure nor the tendency to categorize negative faces as out-partisans predicted subsequent transfers to out-party members, suggesting this categorization operates independently of out-partisan hostility (see Supplementary Information Section F).
Conclusion
People often form impressions rapidly, based on minimal cues like facial appearance, emotions, or demographic features. Our ‘drag and drop’ method uniquely captures how people may categorize strangers in contexts that closely mirror real-world interactions. Two interconnected mechanisms might explain the sorting identified in these studies. The first, rooted in social identity theory, suggests individuals categorize those displaying negative emotions as out-partisans to maintain a flattering view of their own group (Tajfel and Turner Reference Tajfel, Turner, Austin and Worchel1979; Tajfel Reference Tajfel1981). Second, respondents may misperceive out-party members as more negative, reinforcing false polarization beliefs (Druckman et al. Reference Druckman, Klar, Krupnikov, Levendusky and Ryan2022; Levendusky and Malhotra Reference Levendusky and Malhotra2016).
Both mechanisms assume authentic categorization, yet our design cannot fully adjudicate between sincere assessment and alternatives such as task-induced priming or expressive partisan cheerleading (Bullock et al. Reference Bullock, Gerber, Hill and Huber2015). Our additional experiment offers some evidence against non-automatic interpretations, as categorizing negative faces as out-partisans does not predict subsequent out-partisan discrimination. These tendencies nonetheless carry real consequences (Flynn, Nyhan, and Reifler Reference Flynn, Nyhan and Reifler2017; Webster, Connors, and Sinclair Reference Webster, Connors and Sinclair2022), and future research should distinguish between interpretations by priming accuracy, adding non-partisan categorization conditions, or asking respondents to describe faces prior to any partisan framing.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/XPS.2026.10041.
Data availability
The data and code necessary to replicate all the analyses in this research are available at the Journal of Experimental Political Science Dataverse within the Harvard Dataverse Network (see Bergeron, Lee-Whiting, and Goel Reference Bergeron, Lee-Whiting and Goel2026).
Competing interests
The authors declare no conflict of interest.
Ethics statement
This project was approved by the University of Toronto Institutional Review Board (protocol #44379).
Open access
