Comparing Model-Based, Feature-Based, and Deep Learning Approaches to Time-Series Clustering for Intensive Longitudinal Data: Monte Carlo Studies and an Empirical Demonstration

Abstract

Intensive longitudinal data allow researchers to ask whether individuals form subgroupswith different within-person dynamics. This thesis compared five parameter-based, feature-based, and convolutional neural network representations for clustering in two Monte Carlostudies, robustness analyses and extensions, and an empirical demonstration. In order-homogeneous autoregressive populations, direct coefficient estimates produced the strongestsubgroup recovery. In mixed-order populations, deep learning representations were morecompetitive for shorter series, whereas direct estimation gained ground as series lengthened. InStudy 1, series length accounted for 50.7% of the variation in Model-Based recovery and samplesize for 0.1%. In extensions where scale or volatility defined membership, representationssensitive to those properties performed best. Cluster-number selection was a separate problem:re-expressing the same fitted autoregressive information in redundant, strongly dependentcoordinates markedly changed the selected count. In an empirical demonstration adapting theprocedure of Li et al. (2025) to affect data, the Bayesian information criterion with a Gaussianmixture selected two clusters for three representations and one for three others, whereas the gapstatistic with k-means selected one for every representation. Overall, recovery depended onalignment between the representation and the subgroup-defining signal, within-seriesinformation, and the geometry presented to the clustering algorithm. Membership recovery andsubgroup enumeration should therefore be evaluated separately, as properties of the wholerepresentation-and-algorithm pipeline.