R-square in Mplus
Message/Author
 Jon Matthew posted on Friday, January 12, 2018 - 1:21 pm
Hi there,

I am running a multilevel mediation model (2-2-1), whereby I am running between coefficients to estimate the variance in the outcome variable (measured at level 1).

The outcome variable measured in level 1 is a latent construct. These are measured through observed variables. When I run the Mplus multilevel mediation analysis, I get a R-Square values for each of the observed variables of that specific latent construct.

Is it possible to get a R-square value for the whole latent construct rather than the variance explained by individual observed items in Mplus?
 Bengt O. Muthen posted on Friday, January 12, 2018 - 5:05 pm
The output gives you R2 also for the latent DVs.
 Jon Matthew posted on Saturday, January 13, 2018 - 2:22 am
Thanks for the response. It does gives the R2 Square values for the Between (latent) DVs and Within (for each Observed item) for the DV latent construct.
 Jack Peltz posted on Friday, June 01, 2018 - 5:59 am
Dear Dr. Muthen, I am currently running analyses using Preacher et al.'s (2010) syntax for 2-1-1 mediation in a multi-level model. What is the best way to estimate how much variance (what %) is being explained in the outcome as well as in the mediator? (STANDARDIZED cannot be used in this instance, so r2 are not immediately accessible.)
 Bengt O. Muthen posted on Friday, June 01, 2018 - 5:37 pm
Why not use standardized - do you have a random slope?
 Jack Peltz posted on Friday, June 01, 2018 - 8:01 pm
We z-scored the variables in Level-1 units, and we are using a random slope model. Can we simply subtract the within and between residual variance from 1 to get the variance accounted for (from the outcome)? Alternatively, for the mediator, can we get the variance accounted for by subtracting the between-level residual variance from 1 as well?
 Bengt O. Muthen posted on Monday, June 04, 2018 - 1:17 pm
That's not the correct way to get R-square with random slope models. Instead, use Estimator = Bayes where you get an R-square that is within-level averaged over clusters. This is described in the paper

Schuurman (2016) How to Compare Cross-Lagged Associations in a Multilevel Autoregressive Model. Psych Methods.
 Jack Peltz posted on Thursday, June 07, 2018 - 8:02 am
Thank you, Dr. Muthen. I may have been mistaken as we are using a fixed slope model. (Sorry about that.) Would the Estimator=Bayes method still apply or would you suggest a different way for determining how much variance is accounted for in our mediator and outcome in the 2-1-1 model?
 Bengt O. Muthen posted on Thursday, June 07, 2018 - 5:50 pm
In that case you can use any of our estimators and just request Standardized in the Output command.