Dynamic Latent Class Structural Equation Modeling: A Hands-On Tutorial for Modeling Intensive Longitudinal Data

Roberto Faleh ∗ ⁣ † {}^{*\,\dagger}, Sofia Morelli ∗, Vivato Andriamiarana, Zachary J. Roman, Christoph Flückiger, Holger Brandt · 2025

Read Dynamic Latent Class Structural Equation Modeling: A Hands-On Tutorial for Modeling Intensive Longitudinal Data by Roberto Faleh ∗ ⁣ † {}^{*\,\dagger}, Sofia Morelli ∗, Vivato Andriamiarana, Zachary J. Roman, Christoph Flückiger, Holger Brandt (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-12983.txt.

Introduction

Roberto Faleh and co-authors wrote a hands-on tutorial for a family of models used on intensive longitudinal data. Such data track people many times over days or weeks. Their 2025 guide builds up to dynamic latent class structural equation models, which let both the measures and the hidden states shift over time.

The tutorial starts with plain factor and time series models in the Bayesian program JAGS. It then adds multilevel structure, dynamic models, and hidden Markov switching, layer by layer. A therapy study on anxiety, with alliance scores between patients and therapists, runs through every step. Each block states the hypothesis, the model, the code path, and how to read the output. Applied researchers get a full ladder, not just the top rung.

Catalog record

CREATORRoberto Faleh ∗ ⁣ † {}^{*\,\dagger}, Sofia Morelli ∗, Vivato Andriamiarana, Zachary J. Roman, Christoph Flückiger, Holger Brandt
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-12983

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dynamic Latent Class Structural Equation Modeling: A Hands-On Tutorial for Modeling Intensive Longitudinal Data Roberto Faleh ∗ ⁣ † {}^{*\,\dagger}, Sofia Morelli ∗, Vivato Andriamiarana, Zachary J. Roman, Christoph Flückiger, Holger Brandt Abstract In this tutorial, we provide a hands-on guideline on how to implement complex Dynamic Latent Class Structural Equation Models (DLCSEM) in the Bayesian software JAGS. We provide building blocks starting with simple Confirmatory Factor and Time Series analysis, and then extend these blocks to Multilevel Models and Dynamic Structural Equation Models (DSEM). Leading through the tutorial is an example from clinical psychology using data on a generalized anxiety treatment that includes scales on anxiety symptoms and the Working Alliance Inventory that measures alliance between therapists and patients. Within each block, we provide an overview, specific hypotheses we want to test, the resulting model and its implementation as well as an interpretation of the results. The aim of this tutorial is to provide a step-by-step guide for applied researchers that enables them use this flexible DLCSEM framework for their own analyses. † † footnotetext: ∗ These authors contributed equally to this work. † Corresponding author: roberto.faleh[at]uni-tuebingen.de Keywords: Dynamic Structural Equation Modeling; Dynamic Latent Class Structural Equation Modeling; Latent Class Analysis; Bayesian Methods; Intensive Longitudinal Data Translational Abstract In psychological treatment, people’s symptoms and emotional experiences often shift over time—and not everyone changes in the same way. Clinicians and researchers need tools to understand these patterns, especially when working with data collected repeatedly over days or weeks. This tutorial introduces a method called Dynamic Latent Class Structural Equation Modeling (DLCSEM) that helps identify different types of change in therapy outcomes. For example, some people may show steady improvement, while others follow more complex or delayed paths. Using real-world data from people undergoing treatment for anxiety, we show how this method can reveal important differences in how people respond to therapy. The goal is to give mental health professionals and applied researchers a tool to better understand individual progress, which can support more personalized treatment planning. The tutorial is designed to be practical and clear, with step-by-step guidance using free, open-source software—making it accessible for anyone working with rich, time-based psychological data. Introduction Intensive longitudinal data (ILD) have become an extremely important data source for researchers in clinical and health psychology (e.g., [ 34 , 50 ] ). ILD capture information from multiple subjects observed over numerous time intervals, ranging from thirty to several hundred distinct time points [ 6 ] . Often, ILD include information from several sources, such as self-reports, physiological, behavioral, environmental and geographic data (e.g., [ 27 ] ). The growing availability of mobile phones, smartwatches, and other personal wearable devices has made ILD collection more accessible and cost-effective [ 47 ] . The collection of ILD in intervention studies (Ecological Momentary Assessment; EMA) has several advantages compared to traditional intervention studies with few time points [ 31 ] . Self-reports are not as strongly affected by biases caused by retrospective recalls of events because data is obtained in real-time and natural settings, thus increasing their generalizability and ecological validity [ 52 ] . The richness of ILD enables researchers to capture both intra- and inter-individual processes and dynamic changes, enhancing the ability to predict future events and improve individual forecasts. This is particularly important for identifying potential relapse triggers or the conditions necessary for optimal treatment efficacy (e.g., [ 15 ] ). Recently, new Bayesian approaches of Dynamic Structural Equation Modeling (DSEM) and its extension to Dynamic Latent Class Structural Equation Modeling (DLCSEM) were developed that are targeted at analyzing ILD (e.g., [ 3 , 4 , 35 ] ). DSEM can be viewed as a combination of time-series models and multilevel SEM. The time-series model on the within-level can be set up for both observed and latent variables. It can include all types of auto-regressive (AR) and moving-average (MA) specifications. Each within-level coefficient can be specified as a random effect to account for person- and time-specific effects. Person-specific random effects provide information about individual differences. This information can produce more precise forecasts of individual trajectories and, for example, improve designing tailored interventions. Time-specific random effects describe dynamical (e.g., accelerated) changes over time [ 23 ] . The between model can include further information such as (observed or latent) baseline covariates. DLCSEM builds upon the foundation established by DSEM but introduces a key extension: a latent class model that captures categorical shifts over time, referred to as ’states.’ Unlike traditional Latent Class Analysis (LCA), where state membership remains static (as in growth curve mixture models), DLCSEM allows state membership to change dynamically over time. This flexibility allows individuals to follow different state-specific processes at each time point, offering a more detailed representation of temporal dynamics and deeper insights into the underlying dynamical process. The ability of DLCSEM to identify critical states, detect stable a priori differences, account for unobserved heterogeneity, and capture intra-individual changes extends the capabilities of traditional dynamic modeling. It enables real-time dynamic classification of individual states, providing a powerful tool for personalized evaluations and prompt interventions. Moreover, it offers aggregated results comparable to those obtained from DSEM, supporting the evaluation of treatments while also classifying individuals into distinct states based on their dynamic responses. In Kelava et al. 2022 , the framework was used to investigate student dropout from university. Their findings revealed that the latent state variable acted as a proxy for students’ intention to drop out. This variable was shown to predict actual dropout, which typically occurred, on average, six measurement occasions after the estimated intention to drop out. Additionally, they identified relevant time-varying covariates, such as negative affect, which significantly predicted this shift beyond baseline covariates like cognitive skills. Roman et al. 2024 used DLCSEM to identify inattentive response patterns in long questionnaires. In this application, one latent state represented the response pattern for attentive participants using a hypothesized factor structure; and the second state modeled inattentive response behavior using person-specific parameters that are independent of the underlying constructs that were supposed to be measured. Flückiger et al. 2022 applied the DLCSEM framework on the dimensional structure of Working alliance inventory WAI [ 32 ] . They showed that patients’ response patterns changed over the course of a psychotherapeutic intervention. While patients used a three-dimensional scale initially that represented the subfacets of the working alliance, the majority of the patients started to follow a single-dimensional scale (global factor) during later sessions. Scope and Outline In this tutorial, we will provide a step-by-step, didactic guide on implementing Bayesian latent variable models for ILD. We start from basic concepts to ensure a comprehensive understanding of the foundational tools and methods, and gradually build up to more complex frameworks. We will introduce simple Bayesian implementations of Confirmatory Factor Analysis (CFA), Time Series Models, and Multilevel Models (illustrations 1 through 3). Then, we will combine these approaches into a Dynamic Structural Equation Model (DSEM) in illustration 4. Finally, we will extend this approach to Dynamic Latent Class Structural Equation Modeling (DLCSEM) and provide to illustrations: a model that captures sudden gains in therapy progression (illustration 5) and a model that captures dimensional shifts in the underlying factor model (illustration 6). This tutorial is

Read the complete text