Model-based clustering of high-dimensional longitudinal data via regularization.

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Longitudinal data analysis presents challenges in high dimensions.
• Clustering and variable selection are often performed sequentially, limiting efficiency.
• Understanding physical activity trajectories in adolescent girls requires advanced statistical methods.

Purpose of the Study:

• To develop a novel model-based clustering method for high-dimensional longitudinal data.
• To simultaneously identify subject groupings and relevant predictors within each group.
• To analyze the Trial of Activity in Adolescent Girls (TAAG) dataset.

Main Methods:

• A doubly penalized likelihood approach within linear mixed-effects models (LMMs) for sparsity.
• Simultaneous clustering and variable selection using a coordinate descent algorithm within an Expectation-Maximization (EM) framework.
• Gaussian mixture distribution assumption for subject grouping and Bayesian Information Criterion (BIC) for model selection.

Main Results:

• The proposed method demonstrates satisfactory performance in numerical studies.
• It effectively accommodates complex data with multilevel and/or longitudinal effects.
• The method allows for increasing dimensions of fixed and random effects with sample size.

Conclusions:

• The new method offers a unified approach for clustering and variable selection in high-dimensional longitudinal data.
• It provides a robust framework for identifying distinct patterns and influential factors in complex datasets.
• The approach is applicable to various fields requiring analysis of time-course data with subgroup identification.