Linear Mixed Models for Assessing Longitudinal Mediation

Most mediation models utilize cross-sectional data despite the fact that mediation often consists of processes that unfold over time. Maxwell and Cole (2007) demonstrated that cross-sectional approaches to mediation typically generate biased estimates of longitudinal parameters even under the ideal conditions. Kenney et al. (2003) discuss multilevel models in which there is lower level mediation and the mediational links vary randomly across upper level units. Bauer et al. (2006) proposed procedures for evaluating direct, indirect, and total effects in multilevel models when all relevant variables are measured at the lower level and all effects are random. These articles discussed the multilevel model as a “hierarchical” structure (e.g., siblings nested in pedigree). Although one may assume their findings would generalize to the longitudinal multilevel model (i.e., repeated measures nested in individuals), there are many intricacies with covariance structures of a longitudinal model that do not occur in nested data. In the multilevel longitudinal analysis framework, a time-varying covariate would be considered a lower level mediator. Thus, for many time-varying mediators, the Kenney et al. (2003) and Bauer et al. (2006) approaches may be appropriate; however, multivariate time-varying mediators may need to be conceptualized as time-varying latent mediators. Curran and Bauer (2007) demonstrate how to build path diagrams for multilevel models; however, only the slopes and intercepts are conceptualized as latent variables. We propose linear mixed (multilevel) mediation analyses that will incorporate both time-invariant (upper level) and time-varying (lower level) covariates as putative mediating variables through individual level models of longitudinal outcomes.