Three-Level Longitudinal Mediation with Nested Units: How does an Upper-Level Predictor Influence a Lower-Level...

Area of Science:

• * Multilevel modeling
• * Longitudinal data analysis
• * Causal inference

Background:

• * Traditional models like cross-lagged panel models (CLPMs) and 2-2-1 mediation models have limitations in analyzing complex multilevel longitudinal data.
• * Existing methods may produce misleading estimates when significant variation exists between higher-level units.

Purpose of the Study:

• * To propose and evaluate a novel three-level longitudinal mediation model.
• * To assess the accuracy of Bayesian estimation for this complex model.
• * To compare the proposed model with existing methods, including CLPMs with aggregation/disaggregation techniques.

Main Methods:

• * Development of a three-level longitudinal mediation model.
• * Application of Bayesian estimation due to model complexity.
• * Conducting a simulation study to test estimation accuracy under various conditions (cluster size, number of clusters, time points, effect sizes, random effects variances/covariances).

Main Results:

• * Bayesian estimation for the three-level model demonstrated reasonable accuracy for fixed effects (direct and indirect components) when cluster size was adequate, J > 20 clusters, and T ⩾ 4 time points.
• * Cross-lagged panel models (CLPMs) with aggregation or disaggregation techniques produced misleading fixed-effect estimates and low coverage rates when inter-unit variation was present.
• * The proposed model was illustrated with an early childhood education study, comparing its inferences against CLPMs and the traditional 2-2-1 model.

Conclusions:

• * The proposed three-level longitudinal mediation model offers a more accurate approach to evaluating causal processes in multilevel longitudinal data compared to existing methods.
• * Bayesian estimation is a viable and accurate method for complex multilevel longitudinal mediation models.
• * The model provides reliable estimates and inferences, particularly when compared to CLPMs that can yield biased results in the presence of between-unit variation.