Modeling Challenges in Intensive Longitudinal Health Data: Methods for Non-Random Missingness and Early Prediction

Google map

Abstract: This talk is based on her dissertation proposal.

Intensive longitudinal data (ILD) allow detailed study of health behaviors in real-world settings, capturing dynamic, within-person processes and enabling early, personalized interventions. However, ILD pose unique analytical challenges, including non-random missingness, high dimensionality, temporal irregularity, and nonlinear behavioral dynamics.

This dissertation tackles two key challenges. First, it develops a multilevel joint modeling approach to handle non-random missing data, accounting for within- and between-person heterogeneity and temporal clustering. Second, it evaluates methods for long-term health forecasting. In this talk, I focus on the forecasting component, describing how early-stage data can be used to predict long-term outcomes in the presence of contextual shifts. I compare conventional statistical approaches with deep learning models, including data augmentation strategies, to improve predictive accuracy despite modeling challenges in ILD. Simulation studies and empirical analyses are used to evaluate model performance.