Multicollinearity in Confirmatory Fac...
Message/Author
 marina saiz  posted on Monday, November 14, 2005 - 4:12 am
My main intention is to know if four variables form two constructs or just one.
The problem is that two of these four variables are highly correlated (r. 97), however, these two variables form part of a same construct and the other two form part of another construct. I want to know if these highly correlated variables can be employed in CFA despite the problem of high correlation. If so, can you give me any reference? I have always heard that in CFA, multicollinearity might be a problem.
I have tried Principal component analysis where multicollinearity is not a problem but I have a problem that one construct is dependent on the other.
Thanks a lot for your help
Marina
 bmuthen posted on Monday, November 14, 2005 - 5:58 am
Multicollinearity is a problem if you use the two constructs in the prediction of another variable. If for example they are used as dependent variables, the problem would not arise.
 Scott R. Colwell posted on Tuesday, March 14, 2006 - 6:41 am
I am running an multi-group model (cluster is 0 and 1) with TYPE=COMPLEX. I suspect that there is some degree of multicollinearity in group 1 that doesn't exist in group 0.

I have 2 questions that I'm hoping you can help me with.

1) Is there a way of incorpoating a "ridge regression" into the model to stablize the SE estimates, or is there a method of adjusting the SE's for multicollinearity? I am using the default MLR estimator, is there a better estimator when there are near multicollinearity cases?

2) Can I use the MODEL 1: function to add a second order factor behind the highly correlated factors in group 1 to try to control for this? If so, how do I specify a different factor structure in model 1 than in Model 0?

Thanks,
 Bengt O. Muthen posted on Tuesday, March 14, 2006 - 8:00 am
1. Ridging is only done for observed covariates and then it is automatic with no user input possible.

2. A second-order factor model can help by predicting from the second-order factor instead of the first-order factors. Each group can have a different model. You specify the full model in the general part and then you set second-order factor parameters to zero in the group-specific part for the group that doesn't have the second-order factor.
 Scott R. Colwell posted on Tuesday, March 14, 2006 - 8:37 am
Is it reasonable to say then that if one converts a latent varible into an observed one (exogenous variable) that the ridging would take place automatically?
 Bengt O. Muthen posted on Tuesday, March 14, 2006 - 9:17 am
Only if it is an exogeneous variable and only if the sample covariance matrix is singular so that ridging is necessary. The high correlation case does not qualify.
 Cheng posted on Tuesday, June 06, 2017 - 7:58 pm
a) Is there any method I can use in Mplus 7.4, to confirm whether my CFA model doesn't has multicollinearity problem? b)If yes, can we get the VIF values?

c) I got this message when run a CFA in Mplus "The latent variable covariance matrix (PSI) is not positive definite. This could indicate a negative variance/residual variance for a latent variable, or a correlation greater or equal to one between two latent variables, or a linear dependency among more than two latent variables."

Does this indicate that my model might have multicollinearity problem? If yes, can we check in TECH4 OUTPUT, which output section should I read?

Thank you very much
 Bengt O. Muthen posted on Wednesday, June 07, 2017 - 6:02 pm
A regular CFA without covariates doesn't have multicollinearity issues. The PSI problem is due to model mis-specification. Run an EFA to see where you might go wrong in terms of fixing cross-loadings to zero.