Speaker
Description
Background
Longitudinal observational data frequently involve time-varying confounding, autoregressive dependence, and potential reciprocal feedback between processes. These features complicate the estimation of cross-lagged causal effects and challenge the assumptions underlying standard modelling approaches. Methodological evaluation requires transparent, reproducible simulation frameworks to verify that candidate methods can recover the underlying causal data structure.
Objective
To construct an empirically informed, reproducible simulation framework following the ADEMP (Aims, Data-generating mechanism, Estimands, Methods, Performance measures) structure, and to assess the ability of Structural Equation Modelling (SEM) and Bayesian dynamic multivariate panel modelling (dynamite) to recover predefined cross-lagged causal effects.
Methods
A data-generating mechanism was developed for N=1000 individuals measured at five timepoints (baseline and four follow-ups). It incorporated a time-varying confounder with deterministic drift and gamma-distributed increments, along with autoregressive processes for the exposure (X) and outcome (Y). Three scenarios were examined: (1) no cross-lagged effect, (2) a unidirectional effect from X→Y, and (3) bidirectional feedback. For each scenario, 500 Monte Carlo replications were generated. SEM and dynamite were applied to each dataset. Performance was evaluated using bias, root mean squared error (RMSE), and 95% interval coverage. All code, simulations, and analysis scripts were written for transparency and reproducibility. The computational time required was also recorded. .
Results
Dynamite consistently recovered the true cross-lagged parameters across all scenarios, with almost zero bias, low RMSE, and coverage near nominal levels. SEM displayed systematic negative bias and near-zero coverage across all scenarios, including when no true effect was present. Runtime benchmarking based on 2,000 Monte-Carlo replications showed that SEM executed all scenarios in approximately 1.3 hours. Dynamite required 6.2 hours for Scenario 1, 5.9 hours for Scenario 2, and 55.3 hours for Scenario 3, driven by the increased computational burden of posterior sampling in reciprocal-feedback structures. While the operational overhead for dynamite was substantially higher, its parameter recovery was materially superior.
Conclusions
Model performance depended critically on alignment between methodological assumptions and the data-generating mechanism. The framework demonstrated that Bayesian dynamic panel models offer robust estimation under time-varying confounding and temporal feedback, whereas SEM failed to recover the true causal effects. These results underscore the importance of rigorous simulation-based validation before applying causal methods to longitudinal observational data.
85717617289