Multilevel mediation with count outcome
Message/Author
 Jonathan Mohr posted on Wednesday, May 22, 2019 - 2:52 pm
I want to test indirect effects in a 1-1-1 multilevel mediation model with a count outcome. The Mplus syntax I've put together runs, but I'm not sure it is correct.

I'm writing for guidance on
(1) if the syntax (below) seems OK,
(2) if Mplus can be used to generate confidence intervals for the indirect effect using bootstrapping or Monte Carlo methods. Here's the syntax (X = predictor, M = mediator, Y = outcome)

USEVARIABLES ARE X M Y ;
CLUSTER IS ID;
COUNT IS Y (nb);
MISSING = ALL (-99);

ANALYSIS: ESTIMATOR IS MLR; TYPE = TWOLEVEL RANDOM;

MODEL:
%WITHIN%
X*;
Y ON X* (cw)
M* (bw) ;
M ON X* (aw) ;

%BETWEEN%
B_X BY X@1;
X@0 B_X*; !
[X@0 B_X*];

B_M BY M@1;
M@0 B_M*;
[M@0 B_M*];

Y*;
[Y*];

Y ON B_X* (cb)
B_M* (bb);
B_M ON B_X* (ab);

MODEL CONSTRAINT:
NEW(indb indw totalw totalb);
indw=aw*bw;
indb=ab*bb;
totalw = indw + cw;
totalb = indb + cb;

Thanks,
Jon
 Bengt O. Muthen posted on Wednesday, May 22, 2019 - 4:09 pm
The a*b product formulas are not suitable for count outcomes. See our Short Course video and handout for Topic 11 on our website and also Chapter 8 of our RMA book.
 Noah Emery posted on Friday, August 30, 2019 - 11:27 am
I noticed that Mplus version 8 now allows standardized effects to be estimated for models with count outcomes (e.g., stdYX), but I cannot find any documentation regarding how these is handled computationally.

Are these estimates trustworthy?
 Bengt O. Muthen posted on Saturday, August 31, 2019 - 5:37 pm
For counts, standardization with respect to the X's is the only one that makes sense. You can obtain this by simply standardizing your continuous X's. Binary X's should not use standardized coefficients.