Mediation analysis is a common statistical method in psychology and education research for learning how a treatment affects an outcome via intermediate variables known as mediators. Conventional methods for multiple mediators are predicated on linear models with strict (and often implicit) assumptions. Alternative causal inferential-based methods allow for non-linear models, but assume correctly knowing the mediators’ causal order, and that they share no hidden confounders. These restrictions impede their applications in many realistic settings, and when improperly used can lead to severely incorrect results even with just two mediators. Building on novel definitions of causal direct and indirect effects, I proposed estimators that do not have these restrictions. As an FWO postdoc, I will develop methods for two common and important settings in psychology and education research with more complex data structures: longitudinal mediators and multilevel data. Extensions for these settings are difficult, because they require carefully (i) adjusting for complex but unknown confounding patterns among the mediators, and (ii) conceptualizing multipart causal effects that are simultaneously transmitted within and between individuals. I will define causal effects, and propose estimators that rely on minimal assumptions for valid inference. The estimation procedures will be available via free software to address substantive questions in psychology and education research.