Variances for count outcomes at withi... PreviousNext
Mplus Discussion > Multilevel Data/Complex Sample >
Message/Author
 Robert A. Ackerman posted on Wednesday, August 12, 2015 - 8:16 am
I’m trying to specify a multilevel model for count data in Mplus that I have run with MLwiN.

Because Mplus does not permit estimation of variances for count variables at the within level, I specified single-indicator latent variables and correlated the latent factors instead. Below is my model:

MODEL: %WITHIN%
FOCAL BY fPositw@1;
PEERS BY pPositw@1;
FOCAL WITH PEERS;

%BETWEEN%
fPositw;
pPositw;
fPositw WITH pPositw;

However, I get somewhat different results from MLwiN (especially for the fixed effects and the within-level variances).

Fixed Effects:
fPositw: (MLwiN: 0.590; Mplus: -0.904)
pPositw: (MLwiN: 0.625; Mplus: -0.677)
Random Effects:
%BETWEEN%
Focal variance: (MLwiN: 0.265; Mplus: 0.263)
Peer variance: (MLwiN: 0.186; Mplus: 0.209)
Focal WITH Peer covariance: (MLwiN: 0.217; Mplus: 0.231)
%WITHIN%
Focal-to-Peer variance: (MLwiN: 4.184; Mplus: 2.972)
Peer-to-Focal variance: (MLwiN: 3.801; Mplus: 2.596)
F-to-P WITH P-to-F Covariance: (MLwiN: 3.497; Mplus: 2.777)

Any idea on how I can change my model to obtain results that align more closely with MLwiN?
 Bengt O. Muthen posted on Wednesday, August 12, 2015 - 4:52 pm
It sounds like you are using a random intercept Poisson model in Mplus. Are you sure you are using the same model in MLwiN and not for example negbin? Make sure the programs show the same number of parameters and ompare loglikelihood value and if Mplus has a less high value you can sharpen the convergence criterion (mconv).
 Robert A. Ackerman posted on Thursday, August 13, 2015 - 7:58 am
Dear Bengt,

Thank you so much for your response. I checked MLwiN and I am indeed specifying a random intercept Poisson with the same number of parameters (although I am using first-order marginal quasi-likelihood for estimation). Whereas the log-likelihood value for Mplus is -5611.421, MLwiN provides -2*log(lh) = 18554.5. I tried sharpening the convergence criterion (MCONVERGENCE = .00001), but still got the same results.

I should also note that when I treat the variables as continuous, I get the same exact results across programs.

Should I try lowering the convergence criterion even more? If so, what would you recommend?

Thank you for all of your help!
 Bengt O. Muthen posted on Friday, August 14, 2015 - 6:31 am
For count outcomes the quasi-likelihood estimator gives different results than the ML approach of Mplus. If you use ML in MLwiN you should get the same results.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: