Angry facial expressions bias gender categorization in children and adults: behavioral and computational evidence

Abstract

Angry faces are perceived as more masculine by adults. However, the developmental course and underlying mechanism (bottom-up stimulus driven or top-down belief driven) associated with the angry-male bias remain unclear. Here we report that anger biases face gender categorization toward "male" responding in children as young as 5-6 years. The bias is observed for both own- and other-race faces, and is remarkably unchanged across development (into adulthood) as revealed by signal detection analyses (Experiments 1-2). The developmental course of the angry-male bias, along with its extension to other-race faces, combine to suggest that it is not rooted in extensive experience, e.g., observing males engaging in aggressive acts during the school years. Based on several computational simulations of gender categorization (Experiment 3), we further conclude that (1) the angry-male bias results, at least partially, from a strategy of attending to facial features or their second-order relations when categorizing face gender, and (2) any single choice of computational representation (e.g., Principal Component Analysis) is insufficient to assess resemblances between face categories, as different representations of the very same faces suggest different bases for the angry-male bias. Our findings are thus consistent with stimulus-and stereotyped-belief driven accounts of the angry-male bias. Taken together, the evidence suggests considerable stability in the interaction between some facial dimensions in social categorization that is present prior to the onset of formal schooling.

Keywords: children; emotion; face; gender; representation; stereotype.

Figure 1

Example stimuli used in Experiments 1–3 (A) and in the control study (B). The identity of the faces used in Experiments 1–3 and in the control study were identical, but in the control study all faces were in neutral expression while faces in Experiments 1–3 had either angry, smiling or neutral expressions. Sixteen of the 120 faces from Experiments 1–3 had no neutral pose in the database.

Figure 2

Reaction times for gender categorization in Experiments 1 (adults) and 2 (children). Only reaction times from correct trials are included. Each star represents a significant difference between angry and smiling faces (paired Student t-tests, p < 0.05, uncorrected). Top: Caucasian (A) and Chinese (B) female faces. Bottom: Caucasian (C) and Chinese (D) male faces.

Figure 3

Sensitivity and male bias for gender categorization in Experiments 1 (adults) and 2 (children). Female faces were used as “signal” class. Each star represents a significant difference between angry and smiling faces (paired Student t-tests, p < 0.05, uncorrected). Top: Sensitivity for Caucasian (A) and Chinese (B) faces. Bottom: Male bias for Caucasian (C) and Chinese (D) faces.

Figure 4

Computational models. (A) Overall model specification. Each model had an unsupervised learning step (either PCA, ICA) followed by a supervised learning step (logistic regression or SVM). (B) Training, cross validation and test workflow. Stimuli were partitioned into a training set and a test set. Variables used in further analysis were the Leave-One-Out Cross-validation (LOOCV) accuracy, the test accuracy, and the log-odds at training. Human ratings were obtained in the control study (Supplementary Material).