Which p value to use?
Message/Author
 Qi posted on Thursday, October 09, 2008 - 9:20 am
I am running an SEM model, and Mplus 5 now gives me p values for each coefficient. However, I wonder which p value I should use as the test of whether a particular path coefficient is different from 0. I found in my output, the p values for unstandardized coefficent test vs. STDYX coeffcients are different, p=.029 for unstandardized vs. p=.085 for STDYX coefficent. In Mplus 4, I used the z value in output to get p, and that z was based on (unstd coefficent/s.e. of unstd coefficent); if following this logic, shall I use p=.029?

 Qi posted on Thursday, October 09, 2008 - 9:21 am
I am also wondering what the test under the STDYX section is testing?
 Linda K. Muthen posted on Thursday, October 09, 2008 - 9:46 am
In both cases, raw and standardized, the test is that the coefficient is different from zero. The test is the ratio of the parameter estimate to its standard error. You should use the p-value for the estimate that you report.
 Qi posted on Thursday, October 09, 2008 - 10:06 am
Linda,

Thanks for your answer! I am little puzzled though, usually we would report both the raw and std coefficents, but if their p values differ so much, how can we explain this to the reviewers? How did this difference happen? Does it mean that the values we used in Mplus4 era were not accurate for the STD coefficients test?

Qi
 Linda K. Muthen posted on Thursday, October 09, 2008 - 10:27 am
 David Merolla posted on Thursday, October 23, 2008 - 7:34 am
I am also confused by this. In other programs that compute a standard error for standardized coefficients they are standardized in the same way as the coefficients (i.e., multiplied by the quotient of s.d x/s.d y) and thus the t-values for the unstandardized and standardized coefficients are the same. How does m-plus compute the standardized standard errors to yield drastically different t-values?
Or
Are drastically different p-values between the coefficients a sign of other problems in the model?

Thanks
 Linda K. Muthen posted on Thursday, October 23, 2008 - 9:57 am
Raw and standardized coefficients have different standard errors. Using the raw standard errors for standardized coefficients is not correct. Go to Technical Appendices on the website where there is a document that describes how Mplus computes standard errors for standardized coefficients.

If you have drastically different z-values, it could be caused by a problem with the model such as variances close to zero. Small differences are to be expected. If you are not sure, send your output and license number to support@statmodel.com and we will take a look at it.
 MAH posted on Thursday, June 25, 2009 - 12:09 pm
Quick question. I have a standardized standard error of 0 surrounding a standardized path coefficient of 0. I am trying to obtain CI's for this standardized path coefficient. Are the CI's 0? Or is there some problem with the model? I did not get an error message.
 Bengt O. Muthen posted on Thursday, June 25, 2009 - 5:28 pm
An estimated SE of zero gives a CI of length zero. So you should ask yourself why SE=0.