Dynamic Structural Equation Models

• Dynamic Structural Equation Models are statistical frameworks that blend time series dynamics, multilevel structures, and latent variables to capture evolving and complex systems.
• They employ state space representations and Kalman filter techniques to enable efficient Bayesian inference and scalable gradient-based posterior sampling.
• Applications span intensive longitudinal, spatiotemporal, and social network data, supporting causal analysis and dynamic topology tracking through advanced optimization algorithms.

Dynamic Structural Equation Models (DSEMs) generalize traditional structural equation modeling by integrating time series dynamics, multilevel hierarchical effects, and latent variable structures. They offer a unifying statistical framework for modeling both the evolution of unobserved processes over time and structured variation between units ( e.g., individuals or spatial locations). DSEMs have become central for the analysis of intensive longitudinal, spatiotemporal, and complex social network data, as well as for causal inference in time-dependent systems.

1. Mathematical Formulations and Model Classes

A general DSEM partitions observed series $\mathbf y_{it}$ (dimension $U$) for individual $i=1,\dots,N$ and time $t=1,\dots,T$ into additive latent components: within-person dynamics, between-person effects, and between-timepoint effects: $$\mathbf y_{it} = \underbrace{\mathbf y_{1,it}}_{\text{within}} + \underbrace{\mathbf y_{2,i}}_{\text{between-person}} + \underbrace{\mathbf y_{3,t}}_{\text{between-time}}$$ Between- and within-level models are specified as multivariate normal SEMs with hierarchical factor structure and time-varying latent states: $$\begin{aligned} \mathbf y_{3,t} &= \boldsymbol\nu_3 + \Lambda_3\,\boldsymbol\eta_{3,t} + K_3\,X_{3,t} + \boldsymbol\epsilon_{3,t},\quad \boldsymbol\epsilon_{3,t}\sim N(\mathbf0,\Sigma_3),\ \boldsymbol\eta_{3,t} &= \boldsymbol\alpha_3 + B_3\,\boldsymbol\eta_{3,t} + \Gamma_3\,X_{3,t} + \boldsymbol\xi_{3,t},\quad \boldsymbol\xi_{3,t}\sim N(\mathbf0,\Psi_3), \end{aligned}$$ with analogous formulations for between-person terms. The within-person latent process is a vector-autoregressive SEM of lag $L$: $$\begin{aligned} \mathbf y_{1,it} &= \nu_{1,it} + \Lambda_{1,it}(L)\,\boldsymbol\eta_{1,it} + R_{it}(L)\,\mathbf y_{1,it} + K_{1,it}\,X_{1,it} + \boldsymbol\epsilon_{1,it},\ \boldsymbol\eta_{1,it} &= \alpha_{1,it} + B_{1,it}(L)\,\boldsymbol\eta_{1,it} + Q_{it}(L)\,\mathbf y_{1,it} + \Gamma_{1,it}\,X_{1,it} + \boldsymbol\xi_{1,it}, \end{aligned}$$ with Gaussian process and innovation errors.

DSEMs subsume several cases:

• Latent variable vector autoregressive (VAR) models, with measurement structure for multiple indicators.
• Multilevel (hierarchical) extensions, including random coefficients on dynamic parameters.
• Cross-classified and spatial dynamic factor models, via structured loadings and VARX components (Valentini et al., 2013).

2. State Space Representation and Efficient Bayesian Inference

Key computational advances arise from expressing the (within-level) time series SEM as a linear Gaussian state space model: $$\begin{aligned} \tilde{\boldsymbol\eta}_{i,t+1} &= T_{it}\,\tilde{\boldsymbol\eta}_{it} + c_{it} + w_{it},\quad w_{it}\sim N(\mathbf{0},W_{it})\ \mathbf y_{1,it} &= Z_{it}\,\tilde{\boldsymbol\eta}_{it} + d_{it} + v_{it},\quad v_{it}\sim N(\mathbf{0},H_{it}), \end{aligned}$$ where $\tilde{\boldsymbol\eta}_{it}$ augments current and $U$0-previous latent states and observations. This reduction is nontrivial and leverages block-matrix manipulations to recover a standard linear state-space form, critical for scalable Bayesian inference (Sørensen, 4 Mar 2026).

Analytical integration of latent trajectories via the Kalman filter leads to a collapsed marginal likelihood: $U$1 Gradient-based posterior sampling on the collapsed surface (e.g., via Hamiltonian Monte Carlo) is feasible, as the required derivatives can be propagated through the Kalman recursions using automatic differentiation.

3. Online and High-Dimensional Dynamic Topology Tracking

DSEMs are applied to dynamic network topology inference by modeling a time-varying adjacency matrix $U$2 within a SEM: $U$3 where sparsity is imposed on $U$4 (typically with $U$5 penalty), and adaptation to nonstationarity is achieved via exponential forgetting factors. Efficient online proximal algorithms (ISTA, FISTA, or SGD) accommodate large problem dimensions and enable pseudo–real-time updating (Zaman et al., 2020, Baingana et al., 2013). Theoretical guarantees, such as dynamic regret bounds, quantify tracking error in terms of the pathwise drift of the ground truth topology.

A partial ordering constraint or a priori structure can be seamlessly incorporated in high-dimensional SVARs. Polyhedral and smooth surrogate constraints on acyclicity are enforced alongside $U$6 sparsity via DC–ADMM, with convergence to stationary points guaranteed under mild regularity (Lin et al., 2023).

4. Dynamic Latent Class Structural Equation Models

DSEMs can be further generalized to Dynamic Latent Class SEMs (DLCSEM), in which the parameter regime switches among a finite set of latent states governed by a hidden Markov process. For instance, with latent state $U$7,

$U$8

with state-specific parameters for measurement and (AR) dynamic structure. Bayesian implementation proceeds via hierarchical models (e.g., random effects on intercepts and AR terms) plus Markov transitions for class membership. Applied workflows are specified for intensive longitudinal data incorporating abrupt or regime-dependent latent changes (Faleh et al., 18 Aug 2025).

5. Spatio-Temporal and Multivariate Dynamic SEMs

Spatial Dynamic Structural Equation Models (SD-SEM) extend DSEM methodology to lattice (regional) data by introducing spatially structured factor loadings modeled as Gaussian Markov random fields: $U$9 where columns of $i=1,\dots,N$0 encode smooth spatial patterns, and the latent factors $i=1,\dots,N$1 follow VAR or cointegrated error correction dynamics. Proper Bayesian inference is conducted via Gibbs sampling, Metropolis–Hastings, and stochastic search variable selection for cointegration structure (Valentini et al., 2013).

This approach enables the identification of "similarity regions" where multiple units (e.g., US states) share common temporal features, informative for out-of-sample forecasting and for cluster-based policy analysis.

6. Dynamic Structural Causal Models and Continuous-Time Extensions

Recent work formalizes Dynamic Structural Equation Models as Dynamic Structural Causal Models (DSCMs), broadening the scope from discrete-time SEMs to models whose endogenous variables are entire time trajectories (either deterministic or stochastic, possibly cyclic, and with latent confounders) (Rubenstein et al., 2016, Boeken et al., 2024). These frameworks:

• Formalize how ordinary or stochastic differential equations (ODEs/SDEs) induce DSCMs whose solution maps (Itô-maps) satisfy global Markov properties expressed in terms of $i=1,\dots,N$2-separation in mixed graphs.
• Enable reasoning about time-dependent interventions, including both static and dynamic "do" operations.
• Support time-splitting and subsampling, yielding discrete-time SEMs compatible with constraint-based causal discovery.
• Provide identification tools for causal effects under time-dependent (dynamic) interventions, with a full extension of do-calculus.

The DSCM formalism subsumes classical SVARs, discrete-time dynamic SEMs, and continuous-time causal analysis.

7. Computational Benchmarks and Practical Scalability

Recent advances in Kalman filter–marginalized MCMC offer dramatic gains in efficiency relative to traditional Gibbs or joint-parameter-sampling approaches for Bayesian estimation of DSEMs. The per-iteration sampling dimension is reduced from $i=1,\dots,N$3 to $i=1,\dots,N$4, with runtime improvement by factors ranging from $i=1,\dots,N$5 to $i=1,\dots,N$6 depending on model structure (number of indicators, lag, and presence of cross-classification), as documented in simulation studies using the NUTS-Kalman algorithm (Sørensen, 4 Mar 2026). Bulk-ESS and tail-ESS metrics, as well as convergence diagnostics ($i=1,\dots,N$7), indicate superior efficiency and mixing for gradient-based methods leveraging analytic Kalman marginalization.

Method	Computational Complexity	Relative Bulk-ESS Efficiency	Notes
Gibbs (Full latent)	$i=1,\dots,N$8	Baseline (1x)	Scales poorly in $i=1,\dots,N$9
NUTS-Joint	$t=1,\dots,T$0	~0.6–1x	Joint latent and parameter sampling
NUTS-Kalman	$t=1,\dots,T$1	$t=1,\dots,T$2–$t=1,\dots,T$3	Latent states analytically marginalized

• Efficient Bayesian Estimation of Dynamic Structural Equation Models via State Space Marginalization (Sørensen, 4 Mar 2026)
• Dynamic Regret Analysis for Online Tracking of Time-varying Structural Equation Model Topologies (Zaman et al., 2020)
• Principled Identification of Structural Dynamic Models (Francis et al., 18 Dec 2025)
• Dynamic Structural Equation Models for Social Network Topology Inference (Baingana et al., 2013)
• Structural Discovery with Partial Ordering Information for Time-Dependent Data with Convergence Guarantees (Lin et al., 2023)
• From Deterministic ODEs to Dynamic Structural Causal Models (Rubenstein et al., 2016)
• Modeling US housing prices by spatial dynamic structural equation models (Valentini et al., 2013)
• Dynamic Latent Class Structural Equation Modeling: A Hands-On Tutorial for Modeling Intensive Longitudinal Data (Faleh et al., 18 Aug 2025)
• Dynamic Structural Causal Models (Boeken et al., 2024)