Structural multilevel models for longitudinal mediation analysis: a definition variable approach

Mediation analysis is used to assess the direct effect of an exposure on an outcome, and the indirect effect transmitted by a third intermediate variable. Longitudinal data are the most suited to address mediation, since they allow mediational effects to manifest over time. There exist several approaches to deal with longitudinal mediation analysis, and one of the most widely spread, especially in social and behavioural sciences, consists of using multilevel models. However, when applied to mediational settings, these models present some limitations that can be overcome moving to a structural perspective. In this paper we propose a new formalisation of multilevel models within a structural framework combining the reticular action model notation and the definition variable approach. We reconsider two multilevel mediation designs very frequent in longitudinal settings from this structural perspective, discuss the advantages and limitations of such an approach and provide an empirical example.