📊 statistics

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

This paper provides a hands-on tutorial guiding applied researchers through the step-by-step implementation of complex Dynamic Latent Class Structural Equation Models (DLCSEM) in JAGS, using a clinical psychology example to demonstrate how to integrate Hidden Markov Switching Models with Dynamic Structural Equation Models for analyzing intensive longitudinal data.

Original authors: Roberto Faleh, Sofia Morelli, Vivato Andriamiarana, Zachary J. Roman, Christoph Flückiger, Holger Brandt

Published 2026-03-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Roberto Faleh, Sofia Morelli, Vivato Andriamiarana, Zachary J. Roman, Christoph Flückiger, Holger Brandt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how a person's mood changes over the course of a therapy session. If you just ask them once a week, you get a blurry, low-resolution photo. But what if you could take a high-definition video, frame by frame, capturing every tiny shift in their feelings? That is what Intensive Longitudinal Data (ILD) is: a super-detailed, moment-to-moment record of a person's life.

This paper is a "how-to" guide for researchers who want to analyze this kind of video data. It teaches them how to build a very sophisticated mathematical tool called DLCSEM (Dynamic Latent Class Structural Equation Modeling).

Here is the paper broken down into simple concepts, using everyday analogies.

1. The Problem: Why Old Tools Fail

Imagine you are watching a movie of a patient going through therapy.

Old tools treat the movie like a stack of still photos. They might say, "On average, the patient felt better at the end." But they miss the story. They don't see the sudden panic attack on day 3, or the moment the patient finally "gets it" on day 7.
The Challenge: People aren't robots. They don't change in a straight line. Sometimes they are in a "bad mood" state, and suddenly, they switch to a "good mood" state. Sometimes, the way they answer questions changes completely (e.g., at first, they think "trust" and "tasks" are separate things, but later, they blend them into one big feeling of "alliance").

2. The Solution: Building a "Smart Movie Player"

The authors built a guide to construct a model that acts like a smart movie player. Instead of just playing the video, this player can:

Zoom in on specific moments (Time Series).
Compare different people's movies side-by-side (Multilevel).
Detect hidden scenes where the plot suddenly changes (Latent Classes).

They built this player in four steps, like building a house:

Step 1: The Foundation (CFA)

Before building the house, you check the bricks. The researchers first checked if the questions (like "I feel nervous") actually measured what they were supposed to measure. They confirmed that the "bricks" (the questions) were solid and consistent at the start and the end of the therapy.

Step 2: The Frame (Time Series)

Next, they looked at how one moment leads to the next.

Analogy: Think of a domino effect. If you knock over a domino today, it makes the next one fall tomorrow.
The Model: They checked if a patient's anxiety today predicts their anxiety tomorrow. They found that yes, it does! But they also noticed that for some people, the "domino effect" is strong, and for others, it's weak.

Step 3: The Walls (Multilevel Modeling)

Now, they added the walls to compare different people.

Analogy: Imagine a classroom. Every student has a different starting height (baseline anxiety) and grows at a different speed (slope).
The Model: The tool accounts for the fact that Person A might start very anxious but improve slowly, while Person B starts calm but has a sudden spike. It separates the "person" from the "time."

Step 4: The Secret Switch (The "Aha!" Moment)

This is the most exciting part. The researchers added a Hidden Switch (Hidden Markov Switching).

Analogy: Imagine a video game character. For the first half of the level, they are in "Stealth Mode" (moving slowly, hiding). Suddenly, at a specific moment, they switch to "Combat Mode" (moving fast, attacking). The game doesn't tell you when they switched; you have to infer it from their behavior.
The Discovery: The model found that patients didn't just slowly get better. They stayed in a "High Anxiety" state for a while, and then—snap—they switched to a "Low Anxiety" state. This switch happened around the 5th session for most people, which matched perfectly with when the therapy started teaching a specific skill (cognitive restructuring).

3. The Second Big Discovery: The "Fusion" Effect

The researchers also looked at how patients described their relationship with their therapist (the "Working Alliance").

Early on: Patients saw the relationship as three separate things: "Do we like the tasks?" "Do we agree on goals?" "Do we like each other?" (Three separate boxes).
Later on: The model detected a shift. The patients stopped seeing them as separate boxes. They "fused" into one big feeling: "I trust this person."
The Analogy: Imagine a band playing. At first, you hear the drums, the guitar, and the vocals as separate instruments. Later, the music blends so perfectly that you just hear "The Song." The model detected this "fusion" happening over time.

4. Why This Matters (The Takeaway)

This paper is a manual for researchers to stop guessing and start seeing the real story.

For Doctors: It helps identify the exact moment a patient "breaks through" and gets better, so they can replicate that moment in future treatments.
For Science: It proves that people aren't just "getting better" in a straight line. They have hidden states, sudden jumps, and changing ways of thinking.

Summary in One Sentence

This paper teaches researchers how to use a super-smart mathematical camera to film a patient's therapy, not just to see if they got better, but to spot the exact moment they switched from "struggling" to "healing," and to understand how their very way of thinking about their therapist changed along the way.

1. Problem Statement

Intensive Longitudinal Data (ILD), often collected via Ecological Momentary Assessment (EMA), captures high-frequency observations of individuals over time. While traditional methods like Dynamic Structural Equation Modeling (DSEM) effectively model intra-individual change and inter-individual heterogeneity, they often assume that the underlying structural relationships (e.g., factor structures) and dynamic processes remain constant over time.

However, psychological processes are often non-linear and may undergo categorical shifts (regime changes) where an individual moves between distinct latent states (e.g., from a "high anxiety" state to a "low anxiety" state, or from a multidimensional to a unidimensional factor structure). Existing methods like Latent Transition Analysis (LTA) often assume homogeneous transition probabilities across individuals, while standard DSEM cannot easily capture abrupt structural changes in the measurement model itself. There is a need for a flexible framework that combines time-series dynamics, multilevel modeling, and latent class switching to model these complex, state-dependent processes in a confirmatory manner.

2. Methodology

The paper proposes a step-by-step tutorial for implementing Dynamic Latent Class Structural Equation Models (DLCSEM) using the Bayesian software JAGS (integrated with R). The methodology is built incrementally:

Foundational Steps:
1. Confirmatory Factor Analysis (CFA): Establishes measurement invariance and factor dimensionality at specific time points.
2. Time Series Modeling: Introduces Autoregressive (AR) processes (e.g., AR(1)) to model temporal dependencies.
3. Multilevel Modeling (MLM): Incorporates random intercepts and slopes to account for inter-individual differences (between-person) and intra-individual fluctuations (within-person).
4. DSEM: Combines CFA and MLM to model latent variables over time with random effects.
The Core Innovation: DLCSEM:
The authors integrate Hidden Markov Switching Models (HMSM) into the DSEM framework.
- Latent States: Individuals occupy one of $S$ discrete latent states at any given time $t$ .
- State-Specific Dynamics: Parameters (e.g., intercepts, autoregressive coefficients, factor loadings) can vary depending on the current state.
- Transition Mechanism: A Markov process governs the probability of switching from state $s'$ to $s$ . This transition probability can be modeled as a function of time-varying covariates (e.g., previous anxiety levels) using a logit link.
- Bayesian Implementation: The model is estimated using Markov Chain Monte Carlo (MCMC) via Gibbs sampling. The authors utilize weakly informative priors and address identification issues (e.g., label switching) through constrained priors (e.g., censoring intercept differences to ensure $State_1 > State_2$ ).

3. Key Contributions

Practical Tutorial: Provides a comprehensive, hands-on guide for applied researchers to implement complex DLCSEM in JAGS, bridging the gap between theoretical methodology and practical application.
Confirmatory Framework: Argues for and demonstrates a confirmatory approach to latent class modeling. Unlike exploratory latent class analysis, the authors specify states based on theoretical hypotheses (e.g., "improved" vs. "non-improved" states) to ensure interpretability and model stability.
Two Distinct Applications:
1. State-Dependent Dynamics (Illustration 5): Models a shift in the level of a latent variable (anxiety) while maintaining the same factor structure. This captures "sudden gains" in therapy.
2. State-Dependent Structure (Illustration 6): Models a shift in the factor structure itself. It demonstrates a transition from a multidimensional Working Alliance Inventory (WAI) structure (three factors) to a unidimensional structure (global factor), capturing the "fusion" of subscales over time.
Code and Reproducibility: All code is provided in a public GitHub repository, facilitating replication and adaptation.

4. Results

The methodology was applied to data from the IMPLEMENT randomized controlled trial involving 57 patients with generalized anxiety disorder over 15 sessions.

Illustration 1 (CFA): Confirmed configural measurement invariance for the Beck Anxiety Inventory (BAI) but suggested a structural shift in the WAI (increasing correlation among subscales over time).
Illustration 4 (DSEM): Confirmed significant inter-individual differences in baseline anxiety and strong temporal dependencies (high autocorrelation).
Illustration 5 (Sudden Gains DLCSEM):
- Identified two states: a "non-improved" state (higher anxiety intercept) and an "improved" state (lower anxiety intercept).
- Found that patients transitioned from the non-improved to the improved state around session 5, aligning with the timing of cognitive restructuring in the therapy protocol.
- The transition probability was significantly influenced by previous anxiety levels (lower anxiety increased the likelihood of switching to the improved state).
Illustration 6 (Fusion DLCSEM):
- Modeled a transition from a 3-factor WAI structure (State 1) to a 1-factor structure (State 2).
- Approximately 66% of patients switched states, with the average switch occurring at session 5.2.
- The transition was driven by the pattern of responses (correlations between items) rather than absolute score levels.
- The model successfully captured the "fusion" of alliance subscales into a global factor over the course of therapy.

5. Significance and Implications

Personalized Intervention: DLCSEM allows for the real-time classification of individual states, enabling clinicians to detect critical turning points (e.g., sudden improvement or relapse) and tailor interventions dynamically.
Theoretical Insight: The framework moves beyond static trait assumptions, allowing researchers to model how the very structure of psychological constructs evolves during interventions (e.g., how a therapeutic alliance consolidates from distinct components into a unified bond).
Methodological Rigor: The paper highlights critical challenges in estimation, such as convergence issues, sample size requirements (suggesting $N \ge 50$ and $T \ge 25$ ), and the necessity of strong theoretical priors to prevent label switching and non-identifiability.
Causal Caution: The authors emphasize that while DLCSEM captures dynamic associations and conditional dependencies, it does not inherently establish causality without experimental manipulation of the transition mechanism.

In summary, this paper provides a robust methodological roadmap for analyzing complex, non-stationary psychological processes, demonstrating that combining latent class switching with dynamic structural equation modeling offers superior insights into the mechanisms of change in clinical interventions compared to traditional static or purely continuous dynamic models.