Bayesian measurement variance PreviousNext
Mplus Discussion > Categorical Data Modeling >
 Jen Rose posted on Monday, September 23, 2013 - 7:36 am

I want to assess measurement invariance across 24 groups for 6 binary indicators of a single factor. I'd like to use Bayesian SEM to do this, but not all of my indicators are measured in every group. Is it possible to do Bayesian SEM with this kind of missing data?
 Bengt O. Muthen posted on Monday, September 23, 2013 - 9:26 am
Bayes multiple-group is done by mixture knownclass. I think that will work fine with missing on some variables for everyone in the group because a covariance matrix is not computed for each group.
 Jen Rose posted on Monday, September 23, 2013 - 2:10 pm
Thanks. Can this be done in Mplus 6.1?
 Linda K. Muthen posted on Monday, September 23, 2013 - 2:22 pm
No, this was added in Version 7.
 Baoyue Li posted on Thursday, April 03, 2014 - 6:13 am
Dear Dr.Muthen,

I am trying to run a MIMIC model on 32 ordinal indicators (1-4) using BSEM method. We have 6 latent factors and a few binary variables used as the explanatory variables for the 6 factors (i.e. the indirect effects). All direct effects of the binary variables and the cross-loadings are estimated, being given informative priors. The sample size is around 800. Here are some questions we have now:

Q1. Is it better to report all standardized coefficients with the stdY option in OUTPUT statement? we saw some raw loading estimates exceeded 2.

Q2. For the choice of the informative prior variances of the cross-loadings and direct effects, can we just simply follow your examples in the BSEM technical report that uses variance 0.01 and 0.04 respectively? And further run separate sensitivity analysis with several other choices of the prior variances for the cross-loadings and direct effects? Basically, we do not have much prior information about this two kinds of parameters.

Thanks very much for your time in advance,
 Bengt O. Muthen posted on Thursday, April 03, 2014 - 8:06 am
I assume you are declaring your ordinal indicators as categorical.

Q1. I would report both unstand'd and stand'd estimates.

Q2. Those priors are probably a good starting point for a sensitivity analysis.
 Baoyue Li posted on Thursday, April 03, 2014 - 8:46 am
Thanks for your prompt replies. Yes, we declared the indicators as categorical.

Further on Q2, how do we choose the different prior variances for the sensitivity analysis? Is it good practice to first find the largest value that keeps the model identified, and then try a few values smaller than it? and do this for the cross-loadings and direct effects separately?

Many thanks!
 Bengt O. Muthen posted on Thursday, April 03, 2014 - 3:53 pm
You should relate the prior variance to the magnitude of the parameter. For instance, if you regard a direct effect of magnitude + - x to be the limit of what is ignorable, then say 2 SD of the prior should not go beyond that. See my Psych Methods paper for this kind of reasoning with respect to the loadings.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message