Three-Level Longitudinal Mediation with Nested Units: How does an Upper-Level Predictor Influence a Lower-Level Outcome via an Upper-Level Mediator Over Time?

Abstract

Listen

Translate article

ABSTRACTIn the study, extending from the cross-lagged panel model (CLPM) and the 2-2-1 cross-sectional multilevel mediation model, we proposed a three-level longitudinal mediation model to evaluate the causal process among variables at different levels over time. Given the complexity of the proposed model, Bayesian estimation was used. A simulation study was conducted to examine the estimation accuracy of Bayesian estimation for the proposed model. Factors considered in the simulation study included average sample sizes of the lower-level units within each upper-level unit (or cluster size; ), numbers of upper-level units (or clusters; J), numbers of time points (T), fixed direct and indirect effect sizes, and variances and covariances of upper-level random effects. Results indicated that the fixed effect estimates for the indirect effect components (a and b), fixed direct effect (c), and the average mediation effect estimate were reasonably accurate when , J > 20, and T ⩾ 4. Comparing the proposed model against the CLPMs coupled with aggregation or disaggregation techniques, we found that the CLPMs yielded misleading fixed-effects estimates and low coverage rates when variation existed between upper-level units. The proposed model was illustrated using a study on early childhood education. We compared estimates and inferences from the proposed three-level longitudinal mediation model, the CLPMs with aggregation and disaggregation, and the traditional 2-2-1 mediation model. Limitations and extensions of the current study were discussed.