Clustering longitudinal data: comparison of model-based and distance-based approaches using simulated and real-world data in psychiatric research.

Abstract

Background: Clustering longitudinal symptom trajectories is increasingly used to characterise clinical heterogeneity in psychiatric disorders. However, the relative performance of distance-based and model-based approaches remains insufficiently studied in psychiatric settings, where limited follow-up waves, measurement variability, and heterogeneous patient responses may affect clustering reliability. We aimed to systematically compare distance-based and model-based (Latent Class Mixed Models, LCMMs) methods, evaluating their performance in recovering the number of groups, classification accuracy, and clinical interpretability.

Methods: We analysed longitudinal symptom data from 237 first-episode psychosis patients (PEPs cohort). We also performed simulations (500 replicates per scenario) with known group structures (3 or 4), varying group-size distributions (balanced/unbalanced) and levels of group separation (well-separated, partially overlapping, strongly overlapping). Recovery of the true or optimal number of groups was assessed using multiple validity indices. Classification accuracy was measured by weighted Cohen's [Formula: see text]. Clinical interpretability was evaluated in terms of within-group homogeneity, stability across replications, and minimum group size.

Results: In the PEPs cohort, distance-based methods consistently identified two stable and clinically interpretable trajectory groups, whereas LCMMs frequently suggested more complex structures, often including very small classes. Concordance between methodological families was low, indicating that patient stratification was strongly method-dependent: distance-based methods grouped patients primarily by overall symptom severity, whereas LCMMs distinguished groups according to trajectory slope. In simulations, both approaches performed well when groups were clearly separated, but performance declined with increasing overlap and group-size imbalance. Distance-based methods generally produced more stable and homogeneous partitions, whereas LCMMs showed greater variability across replications.

Conclusions: In psychiatric longitudinal studies with few assessment waves and moderate sample sizes, distance-based clustering methods may provide a robust strategy for trajectory-based partitioning. Model-based approaches such as LCMMs may offer advantages when complex nonlinear trajectories are expected, but their stability may be limited under noisy or weakly separated conditions. Combining multiple evaluation criteria and considering clinical interpretability are essential for reliable subgroup identification.