Bayesian Graded Response Models for Eating-Disorder Risk Estimation Using Screening Data

To facilitate eating disorder (ED) diagnosis, individuals usually undergo a brief screening before being referred for a full-scale diagnostic assessment. ED screening tools typically use self-report Likert-type questions to assess ED symptoms. Individuals are classified into low-risk and high-risk groups by comparing their aggregated scores with a Receiver-Operating-Characteristic-determined cut-off point. However, binary classification by cut-off points could result in low positive predictive value (i.e., true positives) in low-prevalence populations, neglects differences in item-level discriminability, and requires a large sample size to determine the cut-off point. In this study, we built a Bayesian Graded Response Model (Bayesian-GRM) to generate ED risk predictions from screening responses. The Bayesian-GRM was implemented in a dataset of 1397 college students screened using two screeners: the SCOFF and the Brief Assessment of Stress and Eating (BASE). First, the Bayesian-GRM, coupled with a logistic regression component to account for gender differences, was fit to a training dataset with both responses to screening questions and ED diagnoses. The fitted model was then used to predict ED risk based on responses to screening questions only. We showed that, by accounting for population-level prevalence and gender, the Bayesian-GRM approach provided accurate individual ED risk estimates that were true to data. By accounting for item-level discriminability, the Bayesian-GRM was able to distinguish individuals with or without an ED with a better accuracy. Using a bootstrap study, we also showed that the Bayesian-GRM could provide relatively robust and stable ED risk estimates in small samples (for example, a sample with 90 women and 90 men). We concluded that the Bayesian GRM method is highly feasible for analyzing ED screening data and identifying individuals with a high risk of ED, and its application would enable lower resources and costs for future studies seeking to evaluate the performance of ED screening tools.