A staged Bayesian framework for longitudinal prediction of adolescent depression using add health data

Area of Science:

• Computational psychiatry
• Longitudinal data analysis
• Bayesian statistical modeling

Background:

• Accurate prediction of adolescent depression requires methods accounting for feedback between mental health and substance use.
• Existing methods may lack reproducibility and interpretability for longitudinal behavioral data.
• This study focuses on predictive modeling rather than causal inference.

Purpose of the Study:

• To develop and validate a staged Bayesian framework for predicting adolescent depression using longitudinal data.
• To incorporate temporal feedback between mental health and substance use in predictive models.
• To improve the accuracy (discrimination and calibration) and interpretability of adolescent depression prediction.

Main Methods:

• Utilized data from the U.S. National Longitudinal Study of Adolescent to Adult Health (Add Health).
• Developed a three-stage Bayesian framework: logistic models, joint modeling of depression and substance use, and temporal feedback modeling.
• Assessed model adequacy using posterior predictive checks and internal validation; reported discrimination (AUC) and calibration (Brier score).

Main Results:

• Lagged depression and lagged substance use were significant predictors of future depression.
• The staged Bayesian model demonstrated improved predictive performance across stages.
• Area Under the Curve (AUC) increased from 0.78 to 0.84, and Brier score decreased from 0.185 to 0.152 from Stage 1 to Stage 3.

Conclusions:

• The staged Bayesian pipeline offers a generalizable method for longitudinal prediction with temporal feedback.
• The framework enhances predictive discrimination and calibration while maintaining model interpretability.
• This approach provides a robust template for analyzing complex longitudinal behavioral data in adolescents.