Long-Term Body Weight Trajectories and Mortality in Older Adults: Hierarchical Clustering of Sparse Functional Data

Anna Zajacova, University of Wyoming
Huong Nguyen, Ohio State University
Snehalata Huzurbazar, University of Wyoming

To resolve the obesity paradox, researchers have increasingly focused on analyzing long-term weight changes and their effect on mortality. Analysts used fully parametric (regression) or semi-parametric (latent class) models, which required difficult-to-justify decisions that sometimes yielded conflicting findings. We propose a cutting-edge nonparametric approach --functional data analysis for sparse longitudinal data, specifically hierarchical clustering of functions estimated via the PACE algorithm-- to estimate classes of BMI functions and identify mortality risks in each. Data are from the Health and Retirement Study (N=9,893). We found three BMI trajectory clusters for each gender: normal stable, overweight gaining, and overweight losing. The mortality of the first two groups was similar while individuals in the overweight losing cluster experienced significantly higher risk of dying. The study highlights the potential of functional data analysis for BMI trajectories, as well as many other developmental and age-dependent processes relevant to population health.