←Back

653 – Longitudinal and Survival Analysis in Observational Studies

Clustering Longitudinal Unbalanced Data: An Application to the Early Childhood Growth Pattern

at Nemours/Alfred I. duPont Hospital for Children/University of Delaware

Investigating the presence of distinct trajectories, forming groups of individuals with similar trajectories, and identifying individual and group-level factors contributive to distinct trends are areas of growing interest in the analysis of longitudinal data. Methods based on dissimilarity in shapes across trajectories are mainly used for clustering unbalanced longitudinal data. Model-based approximation of curves improves precision for sparse and irregular spaced data. This paper used the empirical best linear unbiased prediction (BLUP) of random coefficients in the piecewise linear mixed effects model for dissimilarity measures, and then used heuristic as well as model-based algorithms for clustering BLUP and their functionally transformed scores. The paper also used a model-based evaluation of resulting clusters by adding one cluster at a time and its interaction with time variables to the above piecewise linear mixed effects model to select an optimum cluster solution. An application of this clustering technique to a dataset with a population growth curve of cubic polynomial during ages 1 and 60 months for 3365 children identified a cluster solution with 6 distinct patterns of growth trajectories as the best solution.