Message/Author 


I've got a problem with saving the SIGB and SAMPLE matrices. The inputfile is correct, as far as I can see and the SAVEDATA INFORMATION says that the 2 files are saved but they don't contain any data. I want to fit a 2levelmodel with ordinal data, equal thresholds, no specifications on the WITHINlevel and 1 factor on the BETWEENlevel (twolevel ordinal logit model with unconstrained covariance structure). Thank you for the your latest hints and comments, and thank you in advanced for those concerning this issue. Florian Fiedler. 

bmuthen posted on Monday, July 11, 2005  7:55 pm



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


I finally managed to get the SIGB and SAMPLE correlation matrices for continuous data using ESTIMATOR = MUML. At least the files containing the correlations really contain them. Now the next problem emerged: conducting an EFA works fine with the pooledwithin matrix, but it's not possible to do so on the between matrix. I get an error message saying: *** FATAL ERROR THE SAMPLE COVARIANCE MATRIX COULD NOT BE INVERTED. THIS CAN OCCUR IF A VARIABLE HAS NO VARIATION, OR IF TWO VARIABLES ARE PERFECTLY CORRELATED, OR IF THE NUMBER OF OBSERVATIONS IS NOT GREATER THAN THE NUMBER OF VARIABLES. CHECK YOUR DATA. THIS PROBLEM IS DUE TO: VARIABLE : S20F I had a look at the file already but I can't figure out what could be wrong. There are some values equal to one which are on the diagonal but the other values aren't equal to 1. The number of observations (number of cluster, right?) is 89 and I have 32 observed (pseudo) continuous variables in the model. 1) Any suggestion what go wrong here? 2) Why do these matrices only contain data when I use MUMLestimator? Thank you very much. Florian Fiedler. 

bmuthen posted on Tuesday, July 19, 2005  10:27 pm



1) You may want to read my 1994 Soc Meth & Res article related to this. The estimated SigB is not necessarily positive definite and this problem does not necessarily arise because of two variables being perfectly correlated. The S_B matrix is pos def, but it is the wrong matrix to analyze. These are the reasons why a separate analysis of the between level part of the model is difficult. Now, I think you only get this EFA message if you use ML, not if you use say ULS because ULS does not require a pos def sample matrix. 2) The Sig_B matrix is only estimated with MUML because it is based on the sample covariance matrices S_B and S_PW used by MUML (see the Mplus Tech Appendix); it is not the ML estimate  you get the ML estimate through a Basic run with ML. 


I'd like to continue at the point above. Is there a possibility to get the s_pw matrix from raw data consisting of categorical variables? What's the model I've got to specify? Or is the only solution to assume the data as being continuous? Regards, Florian Fiedler. 


Ok, I just found the section in the MPlus documentation about saving the sp_w with SAMPLE statement works only with continuous data and MUML. So there's no other possibility? 

bmuthen posted on Thursday, November 24, 2005  2:11 pm



S_PW is not available for categorical outcomes. So with categorical outcomes you have to go the ML 2level route. Stay tuned, however, for future developments. 

Anonymous posted on Sunday, December 11, 2005  2:45 pm



Hello Dr. Muthén, does "ML 2level route" refer to step 5 in your 1994paper without the exploratory steps 14? 


Yes. 

Marco posted on Saturday, January 14, 2006  12:18 pm



Will S_PW available for categorical outcomes with Mplus 4? I would like to analyse the withinstructure of likerttype data with considerable skewness, but my problem is that there are not enough clusters to model a MCA with unrestricted sigb. The idea is to analyse spw, where the sample size is high in relation to the number of withinparameters. Maybe you have another suggestion, thanks! 


No, SP_W for categorical outcomes will not be available in Version 4 of Mplus. As far as I know, there is no published way to do this so it would require research to develop. I think that in your situation it is more important to take into consideration that your variables are categorical given that you have floor or ceiling effects and ignore the clustering in the data. You should look at your intraclass correlations and get a rough idea of the design effect using the approximate formula: design effect = 1 + (average cluster size  1) * icc. If this value is less than 2, ignoring clustering is probably not a major problem. 

Fred Li posted on Monday, February 06, 2006  9:16 am



Dear Dr. Muthen: May I ask you to email your 1991 BW program for computing pooled withih and between matrices and ICC..? license number : SABA 04006063 Thanks in advance!! Fred Li 


We don't have this program to distribute. It is part of Mplus. You can obtain these matrices using the SAVEDATA command. You can contact Joop Hox at the University of Utrecht in the Netherlands. I believe he has it. 

Back to top 