Participants

Participants (N = 110; 54 male, 53 female, 3 non-binary) were recruited from the undergraduate psychology research pool at North Dakota State University (NDSU), located in the midwestern United States. Data for eight participants could not be analyzed due to incomplete visual tasks or survey questionnaires. A further ten participants were excluded due to technical difficulties during the task. Finally, five more individuals were excluded due to responding too quickly (responses faster than 250ms for more than 5% of task trials). No other inclusion or exclusion criteria were applied. The final sample for this study consisted of 87 individuals (37 male, 47 female, 3 non-binary). An attrition analysis was conducted to confirm that excluded participants did not differ on questionnaire variables. This analysis did not reveal a significant difference between the groups on symptoms of ED (p = .755), depression (p = .441), anxiety (p = .264), alexithymia (p = .976), and emotion regulation (p = .187). Participants were primarily Caucasian (84%), with 5% of East Asian descent, 3% of African descent, 2% of Hispanic descent, and 2% of Native American descent. They reported a mean age of 19.04 years (SD = 1.99).

MaterialsQuestionnaires

All questionnaires have been scored so that higher values imply more severe symptoms.

Eating disorder diagnostic scale

The Eating Disorder Diagnostic Scale [40, 41] is a brief 22-item self-report questionnaire measuring symptoms of three key eating disorders (AN, bulimia nervosa, and binge-eating disorder) based on diagnostic criteria from the 5th edition of the Diagnostic and Statistical Manual). The scale provides a continuous symptom composite score, which represents an individual’s overall level of ED pathology and was utilized in all analyses. A score of greater than 16.5 points implies clinically significant symptoms [24]; 49 participants (56%) scored above this threshold. In the current sample, the Cronbach’s alpha value for the EDDS symptom composite score was 0.83, 95% CI: [0.77, 0.87].

Beck depression inventory

The Beck Depression Inventory, Second Edition (BDI-II; Beck et al., [5]) is a 21-item self-report instrument intended to assess the severity of depression symptoms. The scale was chosen due to its prevalence in the field, and the consistently high reliability and validity scores across studies. A total score of 14–19 implies mild, 20–28 moderate, and 29–63 severe depression; 13 participants (15%) scored as mild, 16 participants (18%) scored as moderate, and 8 participants (9%) scored as severe. In the current sample, the Cronbach’s alpha value of BDI-II was 0.93, 95% CI: [0.91, 0.95].

State-trait anxiety inventory

The State-Trait Anxiety Inventory (STAI; Spielberger et al., [39]) is a 40-item self-report scale assessing current and typical levels of anxiety. The scale was chosen due to its prevalence in the field, as well as the option of evaluating both trait and state anxiety. In the current study, only the trait component of the scale was used in order to maintain consistency with the other disorder scales. A score of greater than 40 points implies clinically significant symptoms; 11 participants (13%) scored above this threshold. In the current sample, the Cronbach’s alpha value of STAI was 0.93, 95% CI: [0.91, 0.95].

Toronto alexithymia scale

The Toronto Alexithymia Scale-20 [2] is a 20-item scale assessing the severity of alexithymia (i.e., difficulty identifying and describing one’s own emotions). The TAS was chosen due to being the most validated and prevalent measure of alexithymia in the literature, as well as offering measures of separate alexithymia constructs (difficulty identifying feelings, difficulty describing feelings, and externally oriented thinking). The current study used the summed score across the three components in all analyses. A score between 51 and 60 implies possible alexithymia, with scores of above 60 implying definite alexithymia; 27 participants (31%) scored as “possible”, and 20 participants (23%) scored as “definite”. It should be noted that Bagby has since preferred to use the TAS as a continuous dimensional construct [3], which was done in this study. In the current sample, the Cronbach’s alpha value of TAS was 0.77, 95% CI: [0.70, 0.84].

Difficulties in emotion regulation - short form

The Difficulties in Emotion Regulation - Short Form (DERS-sf; Kaufman et al., [22]) is a shortened version of the Difficulties in Emotion Regulation Scale [18], which involves half the items of the original (from 36 to 18), while retaining over 80% of the variance relative to the full measure. The DERS-sf was chosen to reduce participant burden without substantially sacrificing reliability. The DERS-sf consists of six subscales, which were combined into a general index of emotion regulation difficulties in the current study. As this is considered a continuous dimensional measure, no threshold cut-offs were provided. In the current sample, the Cronbach’s alpha value of DERS-sf was 0.91, 95% CI: [0.88, 0.93].

Demographic information

Demographic information including age, gender, ethnicity, and education was collected. To assess gender, participants responded to the question “What gender do you identify as?”. Response options included male, female, non-binary, and other. Participants also reported whether they had ever received a previous diagnosis of an ED, anxiety disorder, or major depressive disorder.

Visual recognition tasksBubbles task

Participants sat 80 cm away from the computer screen. On each trial, the participant saw an emotional face image, partially obscured by the bubbles mask (see Figure S1 for a depiction of these images). Participants were required to make a choice between six emotion categories (anger, disgust, fear, happiness, sadness, surprise) within 2500ms. We employed the QUEST staircasing procedure [46] in order to maintain task difficulty at 75%. This was achieved by increasing the number of bubbles after each correct response, and decreasing the number after each incorrect response in a manner which provided enough information to achieve 75% accuracy. A participant saw 150 trials for each image of an emotion category, for a total of 900 trials, with breaks evenly spaced in-between. A single image for each category was used, with the specific images chosen based on emotion ratings from previous work within the lab. Visual information use was computed as the average Spearman ranked correlation between a Bubbles visualization image of a participant and the ideal observer image for that category. For details of generating bubbles visualization images for participants and the ideal observer, see Supplementary Materials.

Facial emotion recognition task

Participants sat 80 cm away from the computer screen. Participants were required to make a choice between six emotion categories (anger, disgust, fear, happiness, sadness, surprise), without a time limit. On each trial, the participant saw a single face depicting an emotion from one of the six categories. There were three identities per category, for a total of 18 unique images. A participant saw 10 repetitions of each identity, with 30 repetitions per emotion category, for a total of 180 trials. This task was always completed after the Bubbles tasks in order to avoid familiarizing participants with the stimuli that were used in the Bubbles task. Facial emotion categorization accuracy was computed as the average categorization accuracy on this task for each participant.

All questionnaires and behavioral tasks, stimulus sets, and other supplementary information is included on the OSF page for this study: https://doi.org/10.17605/OSF.IO/FH9VD.

Procedure

The authors assert that all procedures contributing to this work comply with ethical standards on human experimentation and the Declaration of Helsinki. All research procedures were approved by the NDSU Institutional Review Board, with the study being granted exempt status. Participants were recruited online via NDSU’s Psychology study registration system. The study description outlined that participants would answer questionnaires about disorder symptoms and perform multiple facial expression recognition tasks. All participants were compensated with course credit. All participants provided informed consent at the beginning of the study. All study tasks and questionnaires took place in a computer testing room, with dividers between individual computers. After consent, participants performed the Bubbles task, followed by the facial emotion recognition task. Self-report questionnaires and demographics were administered at the end of the study with their order randomized, but with the demographic questions always occurring last.

Data analytic approach

Multiple linear regression was used in order to examine the relationship between visual task performance and ED symptom severity. This analysis was conducted in structured steps with both models being nested with the same N. In Model 1, only variables representing overall task performance averaged across emotion conditions were entered into the model. In Model 2, variables representing symptoms of comorbid mental disorders were added. An additional regression analysis was conducted with the six category-specific variables replacing the average task performance. We did not add comorbid disorder symptoms to category-specific models as power for these analyses generally fell below 60%. All predictors, including questionnaires, were mean-centered. EDDS (the outcome variable) was not centered. RStudio (RStudio Team [37]), was used for all analyses.

Power analysis calculations assuming a linear regression analysis framework indicated that a total sample size of 55 participants would provide a power of 0.90 to detect a medium effect (i.e., f2 = 0.15, assuming an alpha of 0.05 and power of 0.80), when examining the contribution of a single predictor (task performance) while controlling for the four covariates (i.e., the comorbid disorder questionnaires). Cohen (1992)’s original effect size guidelines indicate that f2 of 0.15 is medium and an f2 of 0.02 is small effect. A post-hoc sensitivity analysis revealed that the smallest detectable statistically significant effect size is f2 = 0.09 with 80% power, and f2 = 0.12 with 90% power. Thus, the study was powered to detect small-to-medium effects.

Due to the lack of a significant difference in study variables by gender (visual information use, categorization accuracy, and EDDS; all ps > 0.05), and to preserve statistical power for the main analytic variables of task performance, the main analyses are reported for the full sample. See Tables S1-2 for bivariate correlations of all study variables separately for men and women. See Table S3 for all regression study results with gender as a covariate.

We performed two robustness checks. The first evaluated regression assumptions and multivariate outliers. We computed values for leverage, Cook’s distance, and studentized residuals in order to test for multivariate outliers. Participants were excluded if they exceeded the cut-off on at least two of these metrics, and the models were re-ran in a follow-up analysis. Additionally, linearity was examined with loess plots. Homoscedasticity was assessed with a fitted value vs. standardized residual plot, as well as the non-constant variance score test and the Breusch-Pagan test. Normality in the residuals was assessed with a residual qq-plot and the Shapiro-Wilk test. Beyond the potential multivariate outliers described above, no further problems with the regression assumptions were identified during this process.

The second robustness check examined whether our results were sensitive to the assessment of ED symptoms. Instead of using the EDDS composite symptom score,

we followed diagnosis coding, where a 0 means no ED pathology (44%) and a 1 means some ED pathology (56%). We used the score cut-off outlined in Krabbenborg et al., [24] where a score of at least 16.5 points implies clinically-significant ED symptoms. We then performed a logistic regression to compare these categorical groups. This analysis did not reveal a significant effect of EDDS diagnosis (see Table S4).