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

Area of Science:

• Psychiatry and Mental Health
• Statistical Modeling
• Public Health

Background:

• Eating disorder (ED) diagnosis relies on screening tools, often using Likert-type questions.
• Current methods use cut-off scores for classification, which can lack precision in low-prevalence groups and ignore item-level differences.
• Traditional screening requires large sample sizes to establish reliable cut-off points.

Purpose of the Study:

• To develop and validate a Bayesian Graded Response Model (Bayesian-GRM) for enhanced ED risk prediction.
• To assess the model's accuracy in estimating individual ED risk, accounting for prevalence and gender.
• To evaluate the Bayesian-GRM's performance in distinguishing individuals with and without EDs, and its robustness in smaller sample sizes.

Main Methods:

• Implemented a Bayesian-GRM incorporating logistic regression for gender differences.
• Trained the model on a dataset of 1397 college students using SCOFF and BASE screeners with known ED diagnoses.
• Validated the model's predictive accuracy using screening responses only and assessed stability via bootstrap studies.

Main Results:

• The Bayesian-GRM provided accurate, data-driven individual ED risk estimates, considering population prevalence and gender.
• The model demonstrated superior accuracy in distinguishing individuals with and without EDs by accounting for item-level discriminability.
• Robust and stable risk estimates were achieved even with small sample sizes (e.g., 90 men, 90 women).

Conclusions:

• The Bayesian GRM is a feasible and effective method for analyzing ED screening data.
• This approach enhances the identification of individuals at high risk for eating disorders.
• Application of the Bayesian-GRM can reduce resources and costs in evaluating ED screening tool performance.