Hi I have datasets from longitudinal school-based group-randomized trials and am interested in performing mediation analyses where I have “2-1-1” designs: X>M>Y where X is treatment group and varies only at the group level, M and Y are a mediator and outcome respectively that vary across students within schools. I want to confirm that one can now perform a parallel process LGC mediation analysis (as in Cheong and MacKinnon, 2003) using the model INDIRECT statement together with the TYPE=TWOLEVEL option and that this then somehow decomposes the variance into between and within cluster variance components when obtaining standard errors of parameter estimates rather than just applying a post-hoc DEFF adjustment to SEs (as would happen if TYPE=COMPLEX were specified, right?)? Thus, this analysis would perform a true multilevel (time within student within school) mediation latent growth curve analysis where what would be the 1st level in a repeated measures MLM is treated as multivariate observations on M and Y and the 2nd and 3rd levels are explicitly modeled as 2-level multilevel models would be in MLwiN. Not sure if I am stating this correctly and would love any corrections. Thanks very much.
That is the right way to look it. And it is doable in an Mplus Type = Twolevel analysis where on the Between level your treatment affects the between-level part of the student M and Y variables. The longitudinal aspect does not add a level but is handled in a wide format. You can form any indirect effects that you want using Model Constraint.
Thank you, this is very helpful. Based on your last sentence it sounds like Mplus can also be used to analyze models that are an extension of the MSEM cross-sectional models presented in Preacher et al. (2010), with the extension being that you can incorporate the longitudinal nature of the data using LGC within the MSEM framework. Do you know of any place I might find how one would actually parameterize these models? Thanks again.
Thank you! We will take a stab at it. One more follow-up question. Is it possible to test whether the difference between the between-group and within-group components of the indirect effect of a mediator are significantly different? Again, I haven't seen anything written on this in the literature but suspect it can be done in Mplus.