Multilevel Mediation Analysis (2-1-1 and 1-1-1-Mediation models)

[Amelie Nickel]

Dear Statalistusers,

i want to perform a 2-1-1 and a 1-1-1 multileve mediationanalysis. One antecedent Variable is messured at Level 1 (X_ij) and the other one at level 2 (X_j). The mediatorvariable (M_ij) and the dependent Variable (Y_ij) at level 1. I want to analyse if the effect from X_ij and/ or X_j on Y_ij is mediated by M_ij. I am using Stata 12.


I read the following articles, which suggest different methods:

• Krull & MacKinnon (2001): Multilevel Modeling of Individual and Group Level Mediated Effects: https://pdfs.semanticscholar.org/fe1...756.1531144934
  
  I found the following stata command that refers to this article and is provided from UCLA: Statistical Consulting: https://stats.idre.ucla.edu/stata/fa...data-method-1/
  

  ml_mediation

• Bauer, Preacher & Karen (2006): Conceptualizing and Testing Random Indirect Effects and Moderated Mediation in Multilevel Models: New Procedures and Recommendations: https://pdfs.semanticscholar.org/8a5...cfc2c053c9.pdf
  
  For this article UCLA uploaded a stata-syntax as well: perform-mediation-with-multilevel-data-method-2/

  This approach combines the dependent variable and the mediator into a single stacked response variable and runs one mixed model with indicator variables for the DV and mediator to obtain all of the values needed for the analysis.

There are more articles that discuss the topic of multilevel mediation but they seem very complex to me (at least for me they are very complex).
Is there anyone how has had to deal whith this problem as well? Can anyone recommend a stata command or a approach for doing a multilevel-mediation?

I am conscious that this isn't a very specific question but i would be very greatful for any advice.
Thanks in advance and kind regards,
Amelie

[Roman Mostazir]

The UCLA site you referred has this red flag! "NOTE: We are not fully confident that the methods on this page are valid for testing for mediated effects in multilevel models. Proceed at your own risk.".

I think you will be better off if you study Stata's 'sem/gsem' suits for structural equation modeling (type help sem). The pdf manual has examples of mediation and you need to fit them in multi-level setting. Sem cannot fit multi-level data where gsem can, but it is always good to start from -sem- as it will gradually build up your understanding of structural equation modeling domain.

[Amelie Nickel]

Thanks Roman for your answer,

so you think there is no adequate way to combine a multilevel analysis with

xtmixed

and a test of mediation?

Kind regards,
Amelie

[Amedeo Pugliese]

Dear Amelie,
this is a tough question. In our study (not published yet - therefore handle with care) this is what we did:
'The STATA command recommended was gsem (gsem (MV <- IV + Controls) (DV <- MV + IV + Controls M1[org_id]) together with the nlcom and bootstrap routines to test for indirect effects. Identical results are obtained when estimating the models as separate stand-alone using the mixed command (e.g: mixed DV MV + Controls || cluster_id: IDENTIFIER, reml). The latest routine also provides a general model fit (Wald-Chi), that we report in the table'.

Note that DV is level 1; MV is Level 2 and IV is Level 2, Controls: both at L1 and L2.


The references are Lehmann-Willenbrock Meinecke, Rowold & Kauffeld, 2015; Koopmann et al., 2016 for a similar approach.

If you need more context I could share some of the write up of the estimation strategy.

Hope it helps,
amedeo

[Amedeo Pugliese]

Dear Amelie
I was lucky enough to attend Robert Vandenberg ' course on multi-level analysis and he pointed me to a Chapter (14) in a book her co-authored with Charles Lance in which they argue that 2-2-1 or 2-1-1 models are hard if not impossible to estimate. The chapter is called The not so direct cross level direct effect. I think it can change our views on this multi-level mediation and moderation. But please read carefully as it may not apply to you theoretical relationships of interest.

[Amelie Nickel]

Dear Amedeo,

thanks a lot for your hint. I think it is indeed changing the view on multilevel modelling in general.

If i understood it correctly the problem is that is not possible to test the effect of a level 2 variable on a level 1 outcome because the level 1 DV is the mean of the dependent variable between each group of level 2 (especially problamtic when there is a high within-group variation) and therefore the cross-level-effect is the effect of the level 2 variable on the group mean of the DV.

In case of testing multilevel mediation (2-1-1) Zhang, Zyphur & Preacher (2009) suggested to test mediation by differentaiting a within- and a group mediationeffect. They do so by centering the mediator within the context and reintroductin of the subtracted means (CWC(M)). According to this the "true" mediation effect should be calculated by using the between mediation coefficent bbetween.


I wonder if it would be a solution to apply the CWC(M) centering-method to the outcome variable? This would mean that in step two of a multilevel mediation when the mediator serves as the outcome and the effect of X on M is being tested, only the between group mediator is used as the DV and therefore you could calculate the "true "path a_between?

Thanks again and kind regards,

Amelie