Message/Author 

yao lu posted on Tuesday, March 13, 2012  1:57 pm



Hi Dr. Muthen, I have a structural regression model reads as: fcom on ccom; ftrust on ctrust; ftrust on fcom; fatt on catt; fatt on fcom; fatt on ftrust; ctrust on ccom; catt on ccom ctrust; rec pur on fatt; All the variables above are latent. ANALYSIS: Estimator = MLR; The SEM solutions indicated a very good model fit. The results show that the coefficients of ftrust on fcom, rec on fatt, and pur on fatt, were all very high, all exceeding .85. Although significant coefficient is desired, I hardly seen such high coefficients in other literature. I wonder should I be cautions that those high coefficients might indicate some technique errors I made in SEM? Model specification is theoretically sound. Thank you very much for your help in advance. 


You may have multicollinearity among you latent covariates. Ask for TECH4 in the OUTPUT command to see if this is the case. 

yao lu posted on Thursday, March 15, 2012  12:25 pm



Dr. Muthen, Thank you. Yes, multicollinearity was the case. There are no theoretical reasons for me to combine or delete any of highly correlated variables. Are there anyways to deal with this problem in Mplus? I tried centering all the observed indicators for all the latent variables in my model, however, the results were the same? My analysis is MLR. Thank you very much in advance. 


There is no really good way around multicollinearity. If the collinearity is among factors, you may want to investigate why they are so highly correlated, perhaps the fixed zero crossloadings aren't all exactly zero, in which case inflated factor correlations appear. 


Dear Professor Muthen, I'm running a longitudinal model with 4 measurement points and 4 latent constructs I tested my final model on the total sample (N=5000) with the results for the following paths (Standardized XY solution) Meff_G4 > Maff_G5 .09 (est.) .05 (S.E.) .079 (pvalue) Meff_G4 > Mcom_G5 .02 (est.) .04 (S.E.) .67 (pvalue) When I try to run a multigroup analysis for boys and girls (50%50%) I get the following (again XY standardized) results: GIRLS: Meff_G4 > Maff_G5 .00 (est.) .06 (S.E.) .997 (pvalue) Meff_G4 > Mcom_G5 .06 (est.) .05 (S.E.) .24 (pvalue) BOYS: Meff_G4 > Maff_G5 .74 (est.) .04 (S.E.) .000 (pvalue) Meff_G4 > Mcom_G5 .50 (est.) .03 (S.E.) .000 (pvalue) How is it possible that the standardized path coefficients for boys are so extremely high (even higher than path coefficients relating the same construct over time). The model terminated normally but I guess something 'computational' is going on but I can't figure out what it is. I do not expect the paths to be the same for boys and girls but this difference is just too large. Thank you! 


Please send the output and your license number to support@statmodel.com. 


Dear Linda and Bengt, I am currently trying to assess whether multicollinearity is a problem in my model. In a nutshell, I have 3 emotions that affect 2 types of motivation. This motivation, in turn, influences 2 types of wellbeing. The emotions and the 2 types of motivation are both observed variables; 1 of the wellbeing variables is observed and the other is a latent variable. The estimator is MLM. The model is: OtherFocusedWellBeing by mastery growth selfacceptance purpose; CommunionMotivation on gratitude compassion pride; AgencyMotivation on gratitude compassion pride; OtherFocusedWellBeing on AgencyMotivation CommunionMotivation; SelfFocusedWellBeing on AgencyMotivation CommunionMotivation; I have searched the mplus user guide and mplus discussions, but can’t seem to find a solution. What would be the best way to see if multicollinearity is a problem? Thank you, Elizabeth 


Multicollinearity is caused by high correlations so I would look at the correlations among the variables. I don't know of any cutoff for how high constitutes a problem. You might want to ask this on a general discussion forum like SEMNET. 


Dear dr Muthen, I am analysing fairly simple multilevel models, each having the following structure but with varying outcomes: %within% UO2 on predictor1 UO1; %between% UO2 on predictor2 UO1; Some of my results have very high s.e.. For example an estimate of 2.5 and s.e. of 306 on the between level. If I inspect tech4 I do encounter a high correlation between between UO2 and predictor2 (r > .80). 1. Would this indicate problematic multicollinearity? 2. If so, would it be appropriate to exclude UO1 from the between level? For one of the other outcomes using this model, I find a correlation of > .80 for predictor2 and UO2. Als the predictor is the variable of interest, I cannot simply delete this variable from the model. 3. Is there some other solution to take care of the high s.e.? Thank you so much! Kind regards, Aurelie 


How about following Table 5.11 of the Raudenbush & Bryk (2002) book and groupmean center UO1 on within and use its group mean version on between. Both can be obtained using Define. 

Back to top 