Hidden Social Biases in Emotion Recognition Emerge From Reaction Time Variability

For decades, psychologists studying how quickly people recognize emotions on faces have relied on a single, deceptively simple number: the average reaction time. Averaging is the workhorse of experimental psychology, but it flattens the rich structure hidden inside response-time data, discarding information about how consistent people are and how often they produce unusually slow responses. A new study published in Attention, Perception, & Psychophysics by Jason Tipples of the University of Sunderland argues that this averaging habit has been masking something important. By decomposing reaction times into their distributional components, the research shows that the well-known tendency to categorize happy faces faster than fearful ones is only one thread in a far more intricate pattern, one in which social categories such as facial sex quietly shape not just how fast we respond, but how consistent and how error-prone our emotional judgments are.

The phenomenon at the center of the study is the happy-face advantage, the robust finding that people are quicker to classify a smiling face as happy than to classify a fearful face as afraid. The effect has been replicated across laboratories and stimulus sets, and it has become a standard benchmark for theories of emotion perception. Yet the advantage is typically established by comparing mean reaction times across conditions, a practice that treats every response as equally informative and assumes the distribution of responses shifts as a whole. That assumption, Tipples argues, is precisely where the trouble begins. Reaction-time distributions in speeded categorization tasks are famously skewed: most responses cluster near a fast modal value, but a long tail of slow responses stretches out toward several seconds. A mean blends these two regimes together, and any experimental manipulation that affects only one regime can be diluted, exaggerated, or rendered invisible by the averaging process.

To get around this limitation, the study turned to ex-Gaussian modelling, a statistical technique that describes each participant’s reaction-time distribution as the sum of a Gaussian component and an exponential component. The model yields three parameters, each with a distinct psychological interpretation. The mu parameter reflects the mean of the Gaussian component, capturing the speed of typical, fast responses. The sigma parameter reflects the standard deviation of that Gaussian component, indexing the variability or consistency of responses in the fast body of the distribution. The tau parameter captures the exponential tail, quantifying the frequency or magnitude of unusually slow responses. Rather than asking whether happy faces are categorized faster than fearful faces overall, the model asks a sharper question: which component of the distribution carries the effect? Does happiness speed up typical responses, tighten their consistency, or eliminate occasional slow stumbles, or does fear do the opposite?

The results revealed that the answer is all of the above, and that different components tell different stories. The mu parameter replicated the classic happy-face advantage: typical responses to happy expressions were faster than typical responses to fearful expressions, consistent with the conventional mean-based picture. But the sigma parameter told a complementary tale. Responses to happy faces were not merely faster; they were also more consistent, showing less variability in the fast body of the distribution. Fearful expressions, by contrast, produced more scattered responding. And the tau parameter showed that fearful expressions selectively increased the number or duration of unusually slow responses, those occasional trials on which categorization seems to stall. In other words, the happy-face advantage is not a single uniform shift in processing speed. It is a composite of faster typical responses, tighter response consistency, and fewer catastrophic slowdowns, each of which can be influenced independently by the stimulus and by the social context in which it appears.

The most provocative part of the study concerns social-category cues. A long line of research has shown that emotion perception is not socially neutral. Faces do not arrive in the mind as pure expressions; they arrive as men or women, as members of one race or another, and these category memberships interact with emotional expectations. Earlier work by Hugenberg and colleagues demonstrated that the happy-face advantage for White faces can shrink or reverse for Black faces, and studies by Becker, Bijlstra, Craig, Lipp, and Martin have shown that stereotypes linking particular groups to particular emotions, such as the association of anger with men and happiness with women, modulate how quickly expressions are recognized. Tipples’ own prior work had already suggested that reanalysing existing data with ex-Gaussian models could reveal interactive effects of face race and face sex that mean-based analyses had missed. The new study extends that logic systematically, showing that social-category influences are distributed across the mu, sigma, and tau components rather than confined to a single speed measure.

This distributional view has consequences for how the field interprets social-category effects. If a stereotype-congruent combination, say a happy female face or an angry male face, speeds up typical responses but a stereotype-incongruent combination produces a burst of slow responses, the mean will register only a blended difference, and the underlying mechanisms will remain opaque. Two experimental conditions could produce identical mean reaction times while differing dramatically in consistency and tail behavior, implying entirely different cognitive processes. One might reflect a smooth shift in the speed of evidence accumulation, the other an intermittent conflict or hesitation that arises only on some trials. Conventional analyses cannot distinguish these possibilities; distributional analyses can. The study therefore functions as both a substantive finding and a methodological warning: the happy-face advantage literature, and by extension the broader literature on social influences on perception, may contain effects that are real but mischaracterized, or effects that have failed to replicate because the wrong statistic was used to detect them.

The technical machinery behind the analysis is worth appreciating. Ex-Gaussian models were fitted to the full reaction-time distributions, and the parameters were then analyzed with generalized additive models for location, scale and shape, a flexible regression framework developed by Rigby and Stasinopoulos that allows each distributional parameter to be modeled as a function of experimental predictors. This approach, which Tipples has applied in earlier papers on facial expression decision times and on diffusion-model accounts of reacting to angry and happy faces, treats variability not as noise to be minimized but as signal to be explained. The study used the NimStim set of facial expressions, a widely used stimulus battery validated with untrained observers, and the data and analysis code were made openly available on the Open Science Framework, allowing other researchers to verify the distributional decomposition for themselves.

Why should anyone outside the laboratory care about sigma and tau? Because the stakes of emotion perception are social. The speed and consistency with which we read expressions shapes first impressions, threat detection, hiring decisions, and clinical judgments. If social categories systematically alter not just the average speed of emotion recognition but the reliability of that recognition, then the biases embedded in everyday perception are more textured than a simple speed advantage suggests. A face that produces slower, more variable categorization in some observers is a face that is harder to read on the moments that matter, the ambiguous glance, the fleeting expression. Understanding that variability is patterned by social categories opens a route to interventions that target consistency and hesitation, not merely speed.

The study also carries a message for experimental psychology at large. Reaction-time data are collected in virtually every subfield of the discipline, from attention and memory to decision-making and social cognition, and the mean remains the default summary. The present findings join a growing chorus arguing that the mean is not enough. When an experimental manipulation changes the shape of a distribution rather than its location, mean-based analyses will misestimate, or entirely miss, the effect. The happy-face advantage, one of the most familiar effects in emotion research, turns out to be a case in point: its social-category modulation lives partly in the variance and partly in the tail, places the mean cannot see. As distributional and computational modelling becomes more accessible, the hidden structure of response variability is likely to yield further surprises, and studies like this one suggest that some of psychology’s most familiar effects are still waiting to be fully understood.

Subject of Research: Ex-Gaussian analysis of how social-category cues shape reaction-time distributions in facial-expression categorization

Article Title: Response variability reveals hidden social-category effects in facial-expression categorization

Article References: Tipples, J. (2026). Response variability reveals hidden social-category effects in facial-expression categorization. Attention, Perception, & Psychophysics, 88(7), Article 198. https://doi.org/10.3758/s13414-026-03348-y

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03348-y

Keywords: happy-face advantage, emotion categorization, ex-Gaussian modelling, reaction time distributions, response variability, social categorization, face perception, facial sex, social perception, Attention Perception & Psychophysics, Jason Tipples, cognitive psychology