Eye movements to natural images as a function of sex and personality

Abstract

Women and men are different. As humans are highly visual animals, these differences should be reflected in the pattern of eye movements they make when interacting with the world. We examined fixation distributions of 52 women and men while viewing 80 natural images and found systematic differences in their spatial and temporal characteristics. The most striking of these was that women looked away and usually below many objects of interest, particularly when rating images in terms of their potency. We also found reliable differences correlated with the images' semantic content, the observers' personality, and how the images were semantically evaluated. Information theoretic techniques showed that many of these differences increased with viewing time. These effects were not small: the fixations to a single action or romance film image allow the classification of the sex of an observer with 64% accuracy. While men and women may live in the same environment, what they see in this environment is reliably different. Our findings have important implications for both past and future eye movement research while confirming the significant role individual differences play in visual attention.

Figure 1. Images that produced the most distinct eye movements largely depicted social scenes.

Significant differences (blue = men; women = red; dark = p<.05; light = p<.01) displayed for the top fifteen images that produced the most discriminating eye movements and the image that produced the least (bottom right). These images (displayed in full color during the experiment) largely depicted social scenes.

Figure 2. Women consistently fixated lower than men while there fixation distributions were more spread out than those of men.

Panel A illustrates how the mean Y component of female fixations were lower than their male counterparts, especially during the potency block. This effect was replicated using a different, more accurate eye tracker and different participants. Panel B shows entropy calculations of the fixation maps show how, as expected, entropy increased with fixation number. Men's fixation distributions contained higher information than women's indicating women were employing more exploratory and diverse visual strategies, especially around the seventh fixation. Error bars are the standard error of the mean.

Figure 3. Sex classification accuracies spanned from 40% to almost 80% while fixations from different image categories produced significantly different levels of performance.

Panel A displays the distribution of accuracies. Panel B shows which image categories produced the most discriminable fixations. Women, in particular produced more predictable fixations when viewing images that typically contained people. Error bars are the standard error of the mean.

Figure 4. Particularly while viewing images depicting people, women looked marginally below salient features.

Violin plots illustrate how the difference in the distribution of Y-component fixations when fixating faces is likely to be behaviorally significant. While the male distributions tend to center on the eyes of the faces, the distribution of female fixations are shifted down to the nose or even the mouth.

Figure 5. Personality predicts the accuracy of fixation-based sex classification.

Standardized beta values of a logistic regression model trained with personality data to predict sex classification accuracy. Positive beta values represent traits that are likely to be seen in correctly classified individuals while negative betas indicate traits prevalent in misclassified participants. After Bonferroni correction, extraversion (EX), premeditation (PR), perseverance (PE) and conscientiousness (CO) were still significant for both men and women. Openness to experience (OP) was also left significant for women and urgency (UR) for men. Emotional stability (EM), agreeableness (AG) and sensation-seeking (SE) were not significant for either men or women. Error bars represent the standard deviation of the 200 bootstrap estimates.