Cluster analysis of longitudinal profiles for compositional count data
The Annals of Applied Statistics · 2025 · Volume 19 · Issue 2
Abstract
To classify biological roles of different species in an ecological system, modern studies collect longitudinal and compositional counts of DNA sequences of taxonomically diagnostic genetic markers to measure the abundance of species over time. The major challenges of conducting this analysis are twofold: how to accommodate the complex dependence in this data type and how to model the longitudinal trajectories of the species’ abundances. In this paper we propose a novel method named COMPARING to cluster longitudinal profiles for compositional count data to address these challenges. In COMPARING, generalized estimating equation is used to account for both the compositional and longitudinal dependence structures, nonparametric B-spline approximation is used to model the longitudinal curves, and a pairwise-distance penalization is used to identify subgroups with similar longitudinal patterns. We establish the convergence rate of the estimated curves and conclude that the true subgroups can be correctly identified with a high probability. We also conduct simulation studies to show the advantage of COMPARING over its competitors in clustering longitudinal trajectories from compositional count data. Finally, we apply COMPARING to study the co-existence of blood-borne parasites in African buffalo and demonstrate how the method successfully detects biologically meaningful subgroups of parasites for competition-colonization trade-off.
Keywords
B-spline, competition-colonization trade-off, generalized estimating equation, linearized alternating direction method of multipliers, working correlation structure
Enable JavaScript for citation tools and additional paper interactions.