Impact of temporal order selection on clustering intensive longitudinal data based on vector autoregressive models

Area of Science:

• Longitudinal Data Analysis
• Statistical Modeling
• Psychometrics

Background:

• Model-based clustering of multivariate intensive longitudinal data relies on consistent temporal order across individuals.
• Psychological and behavioral processes often exhibit between-individual differences in temporal order.
• Existing methods for setting temporal order include using the highest order (HO) or lowest order (LO) for all processes, with limited study on their impact.

Purpose of the Study:

• To examine the performance of HO and LO methods in vector autoregressive (VAR)-based clustering.
• To compare Gaussian mixture models (GMM) and k-means algorithms within a two-step VAR-based clustering procedure.
• To identify optimal approaches for clustering individuals based on their longitudinal data dynamics.

Main Methods:

• A simulation study was conducted to evaluate the performance of HO and LO methods.
• The study utilized Gaussian mixture models (GMM) and k-means clustering algorithms.
• A two-step vector autoregressive (VAR)-based clustering procedure was implemented across various data conditions.

Main Results:

• The lowest order (LO) approach demonstrated superior performance in cluster identification compared to the highest order (HO) approach.
• The highest order (HO) approach was more favorable for estimating cluster-specific dynamics.
• Gaussian mixture models (GMM) generally outperformed k-means clustering.
• The combination of the LO approach and GMM yielded the best outcomes for cluster identification.

Conclusions:

• The choice of temporal order (LO vs. HO) significantly impacts VAR-based clustering results.
• Gaussian mixture models (GMM) combined with the LO approach are recommended for robust cluster identification in longitudinal data.
• Findings offer practical recommendations for empirical applications of model-based clustering techniques.