Message/Author 


Hello everybody, I'd like to know what type of estimation should I use if I have a model with sampling weights? 


It's not as estimation choice. You would use the WEIGHT option of the VARIABLE command so that estimation would take into account sampling weights. 


Hi Linda, Thanks... I did that you told me but I got the following message: *** ERROR in Analysis command Sampling weights is not available for estimator ML. Can you help me? Thanks again 


Sampling weights are available for MLR, MLM, and MLMV. I would suggest MLR. 


Thanks a lot... Can you suggest some reading about these methods: MLR, MLM and MLMV? Thanks again 


The user's guide has brief descriptions. Look under ESTIMATOR in the index. Also Technical Appendix 4 which can be accessed from the website. 


Hi, I receive the following message, when I try to use my sampling weight: "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED", while there is no problem in estimating the same model without weights (I use MLR as estimator for the weighted estimation). How can I solve this problem? Thank you very much. 


Please send the input, data, output, and license number to support@statmodel.com. 


Hello, I applied sampling weights for a latent variable SEM model estimated with WLSMV estimation. (Because I have many dichotomous indicators, WLSMV was recommended by a professor.) I ran one of my models and it had 139 degrees of freedom. When I applied sampling weights, without otherwise changing the model, the model had 91 degrees of freedom. It it typical to lose degrees of freedom simply by applying sampling weights? Why? Thank you, Lisa 


It seems you are using a version of Mplus prior to Version 6. The degrees of freedom for WLSMV were not computed in the regular way. You should use WLS or WLSM to see the expected degrees of freedom for your model. Prior to Version 6, the chisquare and degrees of freedom were adjusted to obtain a correct pvalue and only this pvalue was interpretable. 


Dr. Muthen, Because I have many dichotomous indicators, I have been using WLSMV estimation. You mentioned that with versions of Mplus prior to Version 6, the ChiSquare values and degrees of freedom in the output are not correct, and that one can get the appropriate degrees of freedom from a model estimated with WLS. Should one get the ChiSquare fit statistic from a WLSestimated model as well? (I think the fit would be different if I used WLS instead of WLSMV, though, because I am working with several dichotomous indicators, and my matrices do break assumptions of multivariate normality.) Also, are the ChiSquare statistics for the DIFFTEST command (used for ChiSquare fit comparisons when one is using WLSMV) accurate with versions of Mplus prior to Version 6? I ran DIFFTEST several times in my work, but I am using Version 5.2, and I wonder if the results from the DIFFTESTs that I am reporting are accurate. Thank you. 


Prior to Version 6, chisquare and degrees of freedom were adjusted to obtain a correct pvalue. This is what should be reported. You should not report a chisquare from a different estimator. DIFFTEST is fine in Version 5.2. 


Thanks, Dr. Muthen. I am confused about the prior statement to "use WLS or WLSM to see the expected degrees of freedom for your model" if I am using a version prior to Version 6. It seems that you are saying that the ChiSquare and degrees of freedom are fine for WLSMVestimated models with Version 5.2. Isn't this inconsistent with the first statement? Thanks again for clarifying this for me. 


In Version 5.2, you should report only the pvalue for the chisquare when WLSMV is used for the reasons stated above. If you want the degrees of freedom calculated in the regular way, you will find them if you use WLS or WLSM or you can calculate them yourself. 


OK, I wonder if I can report the degrees of freedom from my output (using WLSMV in Version 5.2) and simply state that these are adjusted degrees of freedom? 


This would need to be your decision. 


OK, thanks for your help! 

Cecily Na posted on Friday, June 01, 2012  11:09 am



Hello, Linda I think SEM with continuous outcomes assumes multivariate normality. How can I check multivariate normality if I use sampling weights? Thanks 


Multiv nonnormality is not a problem when you use the MLR estimator because it is nonnormality robust. 

Back to top 