Susan Scott posted on Thursday, July 20, 2006 - 11:38 am
I am a new Mplus user, and I am trying to run match up output from Mplus with output from models I ran with Lisrel (matrix language). I am having some difficulties.
1. My model in Lisrel included a categorical variable as well as a request for a Satorra-Bentler correction to produce robust standard errors and chi-sq test. I do not seem to be able to do this in Mplus. If I set all vars as continuous, I can obtain robust standard errors. Most (though not all) estimates are similar to those from my Lisrel output. However, the chi-square is very different from the chi-square I obtained in Lisrel. I am wondering if would be because of the one categorical variable (which loads, along with a continuous var, onto a dependent LV), or if there are other differences between the models of which I am unaware. 2. Also, even with MRM estimation, there are some fit statistics I cannot find. For example, I would like a confidence interval for RMSEA. I saw this question posted on the discussion board, but the response that it could be found in Tech12. When I request Tech12, I get an error message that this is only for Mixture Models. 3. When I switch to MLR or MLF estimation so that I can classify the one var as categorical, I seem to lose all fit statistics. Do I need to add something to the output line? 4. Also, I notice from Tech1 that Mplus is estimating 43 parameters, including 9 means for the 9 dependent vars. Lisrel indicates it is estimating 35 (which seems to be 43-9 +1 variance in Phi for the xLV (which is not really a LV, since it has a single indicator and the variance is set to 0). However, from the fit statistics, I see that both models have 20df. I am guessing this has something to do with the way measured variables that arenít part of a LV are coded in Lisrel, i.e. they appear as single-indicator LVs rather than measured variables. Even if my guess is correct, I would like to understand better.
I would prefer not to repeat the work I have already done in Lisrel, but rather pick up where I left off. Can you help me understand the differences I am obtaining between my Lisrel and Mplus output.
1. The Satorra-Bentler correction was developed for continuous variables not for categorical variables. The default estimator in Mplus when one or more dependent variables is categorical is WLSMV. I don't know what LISREL is doing when it does Satorra-Bentler when one dependent variable is categorical. 2. As far as we know, the confidence interval for RMSEA for MLM has not been developed. 3. When you use maximum likelihood with categorical dependent variables, fit statistics are not available because sample statistics are not sufficient statistics for model estimation. 4. To answer this, you would need to send your Mplus input, data, and output and the LISREL output and your license number to firstname.lastname@example.org.
Jon Elhai posted on Friday, October 22, 2010 - 9:26 am
In using the MLM estimator today using Mplus 6, I noticed that confidence intervals were given for RMSEA. Is that a new feature in Mplus? Do you happen to have any citations for how these CIs were developed for RMSEA?