Multicollinearity bivariate autoregre...
Message/Author
 Jana Ebermann posted on Wednesday, October 19, 2011 - 4:52 am
Hello Dr. Muthen,
I modeled a bivariate autoregressive model. Looking at the standardized solution some estimates are bigger than 1. Normaly this could indicate multicollinearity. I can't find any information if this is as problematic as in normal regressions or if this is formed through the autoregressive process. Should I transform the data? Or isn't it problematic in those kind of models? Could you please help me?
Kind regards,
Jana
 Bengt O. Muthen posted on Wednesday, October 19, 2011 - 9:42 pm
You can request TECH4 and see what the estimated correlations are among your variables. I assume that you allow residual correlations for variables at the same time point.
 Jana Ebermann posted on Thursday, October 20, 2011 - 3:37 am
Dear Mr. Muthen,
as expected the estimated correlation among the variables with autoregressive relationship are very high (around 0.99). Your assumtion regarding the allowed residual correlations for variables at the same time point is right but i also tried to estimate the model with fixed residual correlations to 0 - without any changes. Should I try to find a solution without those problems? Or is it possible to report such a model with standardized estimates larger than 1? I found a text of joereskog about this problematic, but it only mentioned the same thing I already knew: higher values might suggest that there is a high degree of multicollinearity in the data. I didnt find any papers dealing with that problem within those autoregressive path models.
Kind regards,
Jana
 Bengt O. Muthen posted on Thursday, October 20, 2011 - 5:29 pm
Perhaps the two processes are just too highly correlated at each time point to be studied separately.
 Georgia Macnevin  posted on Sunday, November 18, 2012 - 6:25 pm
Dear Dr Muthen,

I'm currently trying to run a pathway analysis model with the following variables:

IV: X1 (experimentally manipulated between-subject variable with two levels)

Mediators: Y1, Y2, Y3, Y4 (all continuous variables measured after the manipulation of X1)

DV: Y5 (continuous variable measured after the mediators)

I set up the following model:

MODEL

Y1 ON X1
Y2 ON X1
Y3 ON X1
Y4 ON X1
Y5 ON Y1 Y2 Y3 Y4 X1
Y1 WITH Y2 Y3 Y4
Y2 WITH Y3 Y4
Y3 WITH Y4

MODEL INDIRECT

Y5 IND Y1 X1
Y5 IND Y2 X1
Y5 IND Y3 X1
Y5 IND Y4 X1
Y5 IND X1

I'm concerned that I might be having a
multicollinearity issue with my four mediating factors but I'm unsure how to test for multicollinearity in Mplus. I had a look through the manual and online but I have been able to find anything that helps me.

Any help would be greatly appreciated.

Thank you,
Georgia
 Linda K. Muthen posted on Monday, November 19, 2012 - 10:35 am
Mplus has no test of multicollinearity. You can look at correlations of factors in TECH4.
 Georgia Macnevin  posted on Monday, November 19, 2012 - 7:32 pm
Thank you for your speedy response Dr Muthen.

Sorry to bother you again but in the Tech4 output I have correlation that are ranging from .68- .80 between my mediators.

The mediator that is correlated at the .80 with another mediator is still producing a significant indirect effect in the model.

Do you think that I can report these results or do you think that they mediators are to highly correlated?

Thank you again

Georgia
 Linda K. Muthen posted on Tuesday, November 20, 2012 - 12:01 pm
These are pretty high. Perhaps you could make them into a factor and use the factor as a mediator.
 Georgia Macnevin  posted on Tuesday, November 20, 2012 - 5:48 pm
Thank you for all your help Dr Muthen. I'll give that a try.
 SocialPsychology posted on Monday, September 30, 2013 - 5:35 am
Hello Prof. Muthen,
I have got a question: I modeled a bivariate autoregressive model. Yet, looking at the standardized solution (stdyx)I found that some estimates are bigger than 1.
At first I thought it might be due to multicolinearity, so I did what you suggested above and had a look at the correlations using TECH4. However, the correlations between my latent variables vary between .12 and .38 so I don`t think that multicolinearity is relevant here. Do you have any idea why I got coefficients bigger than 1?

Thank you very much for your help!

Kind regards,
Michael
 Linda K. Muthen posted on Monday, September 30, 2013 - 6:54 am
See the FAQ on the website called Standardized Coefficients Greater Than One.
 SocialPsychology posted on Tuesday, October 01, 2013 - 3:54 am
Thanks a lot!