Does gaze to faces predict face expertise?
The Social Motivation theory of autism argues that varying levels of social motivation modulate experience with faces over the course of development, and ultimately impact children's face processing skills. A two-step multiple regression analysis was therefore used to discern whether visual attention to faces predicts face perception skill in the combined sample of ASD and TDC participants. Age was entered in Step 1, as preliminary analyses suggested that face processing skills are positively correlated with chronological age (Pearson's r
= 0.49, p
< 0.001) and prior research suggests that face expertise continues to develop throughout childhood and adolescence (Carey et al., 1980
; Thomas et al., 2007
). Proportion of total fixation duration to faces was entered into the model in Step 2. Consistent with our hypothesis, attention to faces accounted for a significant amount of variance in face processing skills above and beyond the effect of age, ΔF(1, 107)
= 5.64, p
= 0.02 (Table , Figure ).
Gaze predicts face processing skill—entire sample combined.
Partial regression plots. Gaze to faces predicting face skill (A) and social skill predicting face skill (B), after controlling for the effect of chronological age. (B) Additionally illustrates a group difference in social skill.
Next, we tested whether scores on the SCQ (a measure that evaluates autistic symptomatology, including social communication skills) predicted total fixation duration to faces and face perceptual skills. While the SCQ not a measure of social motivation per se, these analyses may serve as a springboard for future targeted research using a scale designed specifically to assess motivation. A regression entering SCQ total score as a predictor of attention to faces returned a null result. Next, to test the relationship to face perception skill, we conducted a regression with age entered in Step 1 and SCQ entered in Step 2. Results revealed that the SCQ score accounts for a significant amount of variance in face processing skills after accounting for the effect of age, ΔF(1, 107) = 23.92, p < 0.001 (Table , Figure ), with greater social impairment being associated with reduced face expertise.
Regression with SCQ score predicting face processing skill.
To determine whether total fixation duration to faces differed by stimulus type, salience level, and diagnostic group, a 2 (Type: face/object) × 2 (Salience: high/low) × 2 (Diagnosis: ASD/TDC) repeated measures ANOVA was conducted. This analysis revealed a main effect of Type, F(1, 108) = 61.63, p < 0.001, η2p = 0.36, a main effect of Salience, F(1, 108) = 131.07, p < 0.001, η2p = 0.57, and an interaction between Type and Salience, F(1, 108) = 44.17, p < 0.001, η2p = 0.30. Contrary to our hypothesis, however, there was no effect of Diagnosis, either as a main effect or as an interaction with Type [F(1, 108) = 0.13, p = 0.72, η2p = 0.001], Salience [F(1, 108) = 1.81, p = 0.18, η2p = 0.02], or Type x Salience [F(1, 108) = 0.27, p = 0.60, η2p = 0.003]. Post-hoc tests revealed that all participants looked significantly more at objects (63%) than at faces (37%), t(109) = −7.95, p < 0.001, and more at high salience stimuli (direct faces and high salience objects, 59%) than low salience stimuli (averted faces and low salience objects, 41%), t(109) = 11.58, p < 0.001. Diagnostic group differences were not significant: participants with ASD looked at faces 36% of the time compared to 38% of the time in the TDC group, t(108) = 0.37, p = 0.72, and at high salience stimuli (direct faces and high salience objects) 60% of the time compared to 50% in the TDC group, t(108) = 1.35, p = 0.18. Interestingly, gaze to direct and averted faces was tightly correlated across groups (direct: 20%, averted: 17%, r = 0.73, p < 0.001) but gaze to high versus low salience objects was not (high salience: 39%, low salience: 24%, r = 0.10, p = 0.32). This suggests that high salience objects were much more riveting than low salience objects, and that all faces were attended to similarly whether they faced the observer or were averted.
We began our analyses with very strong a priori hypotheses about gaze in ASD versus TDC participants, based on a significant body of research (Klin et al., 2002
; Nakano et al., 2010
; Rice et al., 2012
). Given that we purposefully calculated our eye tracking variables using Klin and colleagues' methods as a guide, the absence of diagnostic group differences was extremely surprising, and convinced us that the present data warranted a closer look. A number of strategies were used to probe the data and ensure that we did not miss a significant group difference in gaze. Our first follow-up analysis asked whether all children fixated on faces and objects equally quickly from the start of a trial or whether, perhaps, one group was slower to fixate on a certain stimulus type than the other. We hypothesized that the ASD group would fixate on objects more quickly than the TDC group, who would be faster to fixate on faces. As with Total Fixation Duration, however, there was no main effect of diagnosis, F(1, 108)
= 0.36, p
= 0.55, and no interaction between diagnosis and Type, F(1, 108)
= 0.14, p
= 0.71, or diagnosis and Salience, F(1, 108)
= 0.41, p
= 0.52, or diagnosis, Type, and Salience, F(1, 108)
= 0.18, p
= 0.67. Next we tested whether the ASD group might study faces and objects differently than the TDC group (e.g., by examining objects in greater detail than faces), which can be indexed by the number of times participants fixate within an AOI. Again, there was no interaction between diagnosis and Type, F(1, 108)
= 1.08, p
= 0.30, or diagnosis and Salience, F(1, 108)
= 2.36, p
= 0.13, or diagnosis, Type, and Salience, F(1, 108)
= 0.95, p
= 0.33. We then tested the hypothesis that children with ASD would visit object AOIs more frequently than face AOIs, and that this pattern would be reversed in the TDC group. Results revealed no interaction between diagnosis and Type, F(1, 108)
= 2.20, p
= 0.14, or diagnosis and Salience, F(1, 108)
= 1.55, p
= 0.22, or diagnosis, Type, and Salience, F(1, 108)
= 0.10, p
= 0.75. We further tested for differences in average visit duration. As with the other variables we explored, there was no interaction between diagnosis and Type, F(1, 108)
= 0.09, p
= 0.76, or diagnosis and Salience, F(1, 108)
= 1.32, p
= 0.25, or diagnosis, Type, and Salience, F(1, 108)
= 0.44, p
= 0.51. Finally, although our sample is matched on chronological age and GCA at the group level, we re-ran the original RMANOVA on total fixation duration, including age and IQ as covariates in the model in addition to diagnosis as a fixed factor. The interaction between diagnosis and Type was still not significant, F(1, 106)
= 0.33, p
= 0.57, nor was the interaction between diagnosis and Salience, F(1, 106)
= 1.59, p
= 0.21, or diagnosis, Type, and Salience, F(1, 106)
= 0.21, p
Long segments of gaze data may obscure meaningful eye movements that occur in the first few seconds of an experiment (Swingley et al., 1998
). For this reason, we decided to isolate and examine the first 3.5-s loop of gaze data in each trial. A repeated measures ANOVA on proportion of total fixation duration in the first 3.5 s of each trial revealed no interaction between diagnosis and stimulus Type, F(1, 108)
= 1.39, p
= 0.24, and no interaction between diagnosis and Salience, F(1, 108)
= 0.01, p
= 0.92, or diagnosis, Type, and Salience, F(1, 108)
= 1.77, p
After exhausting the possibilities, we determined that our original finding, while surprising given the broader literature, was undeniably accurate. As discussed below, we speculate that the object movies in our paradigm may have been too appealing to reveal group differences that other paradigms with more subtle manipulations were able to document.