Answers are grounded in this paper and cite their sources.
Background Latent variable models (LVMs) are fundamental in health research for capturing unobservable phenomena. However, their broader application is hindered by compartmentalised disciplines, inconsistent terminology, and a lack of methodological standardisation. This complexity often leads to methodological inertia, where researchers default to familiar rather than optimal models. This study aims to map LVMs using a computationally reproducible and expert-auditable approach to design a structured orientation tool. This framework helps researchers navigate existing models and narrow down candidate families based on data characteristics and hypotheses. Methods We applied a hybrid computational approach to map LVMs, beginning with a structured PubMed API search (from inception to 10 December 2025). First, LVM mentions were extracted from titles and abstracts using large language model (LLM) extraction based on mathematical criteria. Mentions were standardised through string processing, then grouped using natural language processing and clustering (TF-IDF and DBSCAN) to identify conceptually identical models and select representative exemplars. Subsequently, LLM-assisted data extraction from a targeted corpus, followed by rigorous expert validation, populated a characterisation grid detailing each exemplar’s theoretical, computational, and software attributes. Finally, selected exemplars were classified into a three-level hierarchical taxonomy based on their directional architectures and methodological objectives. Results A total of 8,495 unique papers were included, yielding 9,555 extracted LVM mentions. After standardisation and filtering, 643 unique model mentions were clustered, yielding 52 representative exemplar models. These exemplars were organised into the proposed exploratory taxonomy, which distinguishes three super-families. These comprise the reflective group (the vast majority, including classical latent class and growth curve models), the formative group (e.g., probabilistic principal component analysis), and the hybrid group (encompassing generative AI and bidirectional architectures). These super-families were further subdivided into nine meta-families relevant to health research. Conclusions This study provides an exploratory taxonomy of LVMs in health research, offering a structured orientation tool. By integrating a hybrid computational approach with expert validation, this framework assists model selection across epidemiological, biostatistical, and data science applications. It provides a practical foundation to help researchers navigate disciplinary silos and identify modelling strategies suited to their specific data characteristics.
No takes yet. Share an insight, caveat, or question.
Marchi et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: