Phil posted on Tuesday, February 21, 2006 - 3:53 pm
Under "TESTS of MODEL FIT" in the Mplus results, there are two model fits: (1) Chi-Square Test of Model Fit; (2) Chi-Square Test of Model Fit for the Baseline Model. Which one should I use? Is good model with small Chi-Square statistic? How to use this Chi-Square statistic to evaluate a model (e.g., Chi-Square/df < 5)?
If it is not given for your analysis, it has most likely not been developed for that situation. If you want a more specific answer, send your input, data, output, and license number to email@example.com.
Joseph posted on Tuesday, February 28, 2006 - 11:24 am
What is the criterion for the WRMR in Mplus? What does "StdYX" mean in model results? Thanks.
We recommend a value less than or equal to one. We do not recommend this for growth models with many time points or for multiple group analysis.
See pages 503-504 of the Version 4 Mplus User's Guide for a description of StdYX.
hans beeman posted on Thursday, November 30, 2006 - 5:04 pm
Hello Drs. Muthén,
using the MLR estimator, MPlus does not give RMSEA confidence intervals. I read in the MPlus forum that a technique to compute confidence intervals for RMSEAs while using MLR has not yet been developed. Hence, I was surprised to read a paper recently (Zimprich, Allemand, & Hornung, 2006 [European Journal of Psychological Assessment]) that reported 95% confidence intervals for RMSEA, having used the MPlus MLR estimator.
It appears the authors have a posteriori implemented the T2* statistic into the formula given in MacCallum, Browne, and Sugawara (1996) [Psychological Methods].
My question is: As I need CIs for my RMSEAs (using MLR) too - is this procedure possible?
We have looked at the paper and think this looks good. We will add it in a future update.
J.-T. Kuhn posted on Wednesday, June 20, 2007 - 5:52 am
I have a question relating to the computation of the RMSEA in multiple group SEM/CFA in MPlus (Type=meanstructure). I have conducted a multiple groups CFA, constraining loadings and error variances across groups. How is the RMSEA calculated in MPlus for this?
The usual formula is
RMSEA = SQRT[d/df],
with d = chi^2 - df / (N - 1).
I have the following parameters:
Overall fit N = 1328 df = 267 chi^2 = 518.178
Groupwise fit Group 1 N = 305 df = 69 chi^2 = 149.139
Group2 N = 354 df = 66 chi^2 = 121.517
Group3 N = 358 df = 66 chi^2 = 136.735
Group4 N = 311 df = 66 chi^2 = 110.787
Using the parameters from the overall model, I get a RMSEA of 0.053 in MPlus, but using the formula above, I get 0.027. What is the reason for this? How is the RMSEA computed in this case?
Or is - in the multiple group setting - the RMSEA calculated groupwise and then averaged?
I'd greatly appreciate your help! I need to calculate RMSEA confidence intervals for MLM-estimates (rescaled), that's why I ask (RMSEA CIs are not given for MLM estimates).
There is no confidence interval available for the MLM estimator. I would look at the point estimate and use the suggested cutoff along with the other fit statistics.
M Hamd posted on Saturday, April 03, 2010 - 5:28 pm
Heck (2001) say on p.109, "To test the complete (i.e., two-group) multilevel model with unbalanced group sizes, Muthen's robust quasi-likelihood estimator (MLM) should be used because it provides the correct chi-square coefficient and standard errors.With unbalanced group sizes, the RMSEA is incorrect, so only the chi-square can be used to estimate the overall model fit."
Is this correct that when I have unbalanced design, I can only interpret ch-sq. But for large sample sizes ch-sq test will seldom pass? I will appreciate any comments on this.
Ref: Heck, R. H. 2001. Multilevel modeling with SEM. In G. A. Marcoulides & R. E. Schumacker (Eds.), New developments and techniques in structural equation modeling: 89-127. Mahwah, NJ: Lawrence Erlbaum Associates, Inc.
Perhaps Heck was referring to the "MUML" estimator which is only ML with balanced data. In the current Mplus the regular ML estimator is the default and recommended estimator. Here, chi-2, RMSEA, and SEs are all correct also with unbalanced designs.
You are either using an older version of Mplus or the CI and probability have not yet been developed for your situation.
Jenny L. posted on Tuesday, June 04, 2013 - 9:10 pm
Thank you for your advice. I'm using version 7 so it should not be the concern. May I ask what might be some reasons why "CI and probability have not yet been developed"? Is it possibly because the lower bound is 0 and the upper bound is also very close to 0 and thus the CI was not displayed? (I had a chi square(df=2)=.01, CFI=1.000, TLI=1.029). Thank you for your help!
No one has taken the time to research the proper CI and probability for all situations.
Li posted on Thursday, December 19, 2013 - 6:55 am
Drs. Muthen, I did a factor analysis with categorical outcomes using Mplus 7, and the reported confidence interval is 90%. I was wondering why Mplus does not report a 95% CI because that's more conventional. I ask this question because a reviewer of my article asked why I reported 90% CI instead of 95%. I sense there is probably a reason behind the decision made by Mplus. Thanks for your explanation. Li