Multilevel Mediation 2 2 1 with observed data

[Amedeo Pugliese]

Hi Statalisters,

I am trying to fit what I thought would be a simple multilevel mediation model of the type 2-2-1 (Preacher et al, 2010; Krull & MacKinnon, 2001 and many others).

This is what the variables look like:
DV is a level 1
IV is a level 2
MV is a level 2
CNT1 is a level 1
CNT2 is a level 2
org_id is the level 2 clustering variable

For the sake of completeness, I have data on directors nested within 65 boards. Data are taken at the same point in time and are all observed data (e.g. no latent).

I have some questions that are bugging me as I do not have access to MPLUS that seems used largely in prior studies fitting 2-2-1 models (Super et al 2016; Lehman-Willenbrok et al 2015).

I guess the main question is: which command should I use in STATA 14.2? I have tried with the following ones but not sure which one is more appropriate in the estimation to avoid conflating results:

1. ml_mediation

code:
ml_mediation, dv(dv) iv(iv) mv(mv) l2id(org_id) cv(CNT1 CNT2)

When I try to estimate the model I do not succeed (e.g. Iteration 59: log likelihood = 8131.7637 (not concave)).
It seems to me that given the path a is a 2-2 (MV on IV) the model cannot converge.

Next I try to bootstrap as follows:
bootstrap indeff=r(ind_eff) direff=r(dir_eff) toteff=r(tot_eff), ///
reps(500) seed(1) cluster(org_id) idcluster(norg_id): ///
ml_mediation, dv(dv) iv(iv) mv(mv) l2id(org_id) cv(CNT1 CNT2)

estat boot, percentile bc

2. xtmixed

I get the same when I try to use xtmixed and specify the clustering at the org_id level. The problem occurs in the estimation of path a (MV --> IV) whereas I am able to estimate Eq2 and Eq3 using the xtmixed, rmle.

The code is as follows:
*Eq - path c (total effect) (DV --> IV)
mixed dv iv CNT1 CNT2 || org_id:, reml

*Eq - path a (MED --> IV)
mixed MV IV CNT1 CNT2 || org_id:, reml

** this one cannot be estimated. Only if I remove the clustering it will work. **

*Eq - path b (DV --> MED) & c' (DV --> IV)
mixed DV MV IV CNT1 CNT2 || org_id:, reml

3. GSEM

I have tried with the GSEM routine

gsem (MV <- IV CNT1 CNT2) (DV <- MV IV CNT1 CNT2 M1[org_id])

I can estimate this one but if I introduce clustering in the first bracket - nothing works.
Then I test the indirect/direct effects as follows:
nlcom _b[DV:MV]*_b[MV:IV]

nlcom _b[DV:IV] + _b[DV:MV]*_b[MV:IV]

Could you please help me understand which one of these is correct and what I am doing wrong?

The alternative I see is using a MLM with estimation of two models as separate (Chen et al 2007 J of App Psy is a great example but the execution seems old style).

Best regards and I am extremely thankful for your support.
Amedeo

[Roman Mostazir]

Try fitting a simple model first whether in ml_mediation or in gsem. Then gradually build your model. See which one is causing trouble. I have not used ml_mediation so can't comment much on this but one problem might be that you are trying to predict a level-2 mediator (MV) with a level-1 predictor(CNT1) (just a wild guess). I am not sure how a level-1 predictor can affect a level 2 mediator. Try removing that from MV equation.

The gsem model is not identified. You need to add error covariance path between the endogenous and the outcome variable i.e. cov(e.dv*e.mv). Again fit a simple random intercept model first than add other variables.

[Amedeo Pugliese]

Thank you Roman.
I can fit gsem model and have no issues even without adding an error covariance term. The issue is that I do not get model fit estimation. When I run mixed || org_id separately on the three models I get the same estimates and Wald as well.
Given that I do not use latent variables, but constructs are mean or average, using GSM is not strictly necessary.
A mistery why I cannot fit ml_mediation.

Amedeo

[Diana Hechavarria]

I am having similar issues as well.