Efficient Bayesian Estimation of Dynamic Structural Equation Models via State Space Marginalization

Dynamic structural equation models (DSEMs) combine time-series modeling of within-person processes with hierarchical modeling of between-person differences and differences between timepoints, and have become very popular for the analysis of intensive longitudinal data in the social sciences. An important computational bottleneck has, however, still not been resolved: whenever the underlying process is assumed to be latent and measured by one or more indicators per timepoint, currently published algorithms rely on inefficient brute-force Markov chain Monte Carlo sampling which scales poorly as the number of timepoints and participants increases and results in highly correlated samples. The main result of this paper shows that the within-level part of any DSEM can be reformulated as a linear Gaussian state space model. Consequently, the latent states can be analytically marginalized using a Kalman filter, allowing for highly efficient estimation via Hamiltonian Monte Carlo. This makes estimation of DSEMs computationally tractable for much larger datasets – both in terms of timepoints and participants – than what has been previously possible. We demonstrate the proposed algorithm in several simulation experiments, showing it can be orders of magnitude more efficient than standard Metropolis-within-Gibbs approaches.

💡 Research Summary

Dynamic Structural Equation Models (DSEMs) have become a cornerstone for analyzing intensive longitudinal data because they simultaneously capture within‑person time‑series dynamics and between‑person hierarchical variation. However, the prevailing Bayesian estimation strategy—Metropolis‑within‑Gibbs as implemented in Mplus—requires sampling a latent state for every participant at every time point. This leads to a parameter space that grows as O(N·T·V₁) (where N is the number of participants, T the number of occasions, and V₁ the dimension of the within‑level latent factor), causing severe computational bottlenecks and highly autocorrelated draws.

The paper’s central contribution is a theorem showing that the within‑level part of any DSEM can be expressed exactly as a linear Gaussian state‑space model (LG‑SSM). By augmenting the state vector to include the current and L‑1 past latent factors together with the corresponding observed indicators, the authors construct a transition matrix Tᵢₜ and an observation matrix Zᵢₜ that preserve the Markov property. Consequently, the Kalman filter can be applied to compute the exact marginal likelihood of the data and its gradient with respect to all static parameters.

Marginalizing the latent states collapses the sampling problem from O(N·T·V₁ + N·V₂ + T·V₃ + |Θ|) to O(N·V₂ + T·V₃ + |Θ|), where V₂ and V₃ are the dimensions of the between‑person and between‑timepoint latent variables, and Θ denotes all remaining model parameters (loadings, autoregressive coefficients, variances, etc.). This reduction eliminates the need to draw the high‑dimensional latent trajectories at each iteration.

To exploit the reduced dimension while still updating all parameters jointly, the authors embed the Kalman‑filter‑based marginal likelihood into a Hamiltonian Monte Carlo (HMC) framework, specifically the No‑U‑Turn Sampler (NUTS). NUTS uses the exact gradient of the marginalized log‑posterior to propose distant, energy‑preserving moves, achieving high acceptance rates even in very high‑dimensional spaces. Because the Kalman filter is run only once per iteration, the computational overhead of each HMC step is modest, dominated by matrix‑vector operations on the static parameters.

A series of simulation experiments explores a wide range of DSEM specifications: varying latent factor dimensions, lag orders, numbers of indicators, missing‑data patterns, and dataset sizes (N = 500–2000, T = 50–200). Across all scenarios, the proposed algorithm outperforms the traditional Metropolis‑within‑Gibbs sampler by factors ranging from 10× to over 300× in effective sample size per second (ESS/s). The gains are especially pronounced when V₁ is large, confirming that the state‑space marginalization eliminates the dominant source of computational cost. Posterior means, credible intervals, and convergence diagnostics are comparable or superior to the baseline, indicating that efficiency is not achieved at the expense of statistical accuracy.

The authors acknowledge limitations: the current formulation assumes linear dynamics and Gaussian noise. Extending the approach to nonlinear state equations or heavy‑tailed measurement errors would require approximate filters such as the extended Kalman filter or particle filters, which would re‑introduce additional computational burdens. Moreover, the choice of priors for the static parameters can affect HMC performance, and guidance on robust prior specification is needed. Finally, the implementation is built on R and Stan; further speedups could be realized through parallelization, GPU acceleration, or a dedicated C++ library.

In summary, by recasting the within‑level DSEM as a linear Gaussian state‑space model, analytically marginalizing the latent trajectories with a Kalman filter, and employing NUTS‑based HMC for the remaining parameters, the paper delivers a dramatically more scalable Bayesian estimation procedure. This methodological advance opens the door to applying full Bayesian DSEMs to large‑scale intensive longitudinal studies that were previously infeasible, and it sets a clear agenda for future work on non‑linear extensions and high‑performance software.