Correlations > 1
Message/Author
 Jordan Silberman posted on Friday, July 06, 2012 - 2:25 pm
Hello,

I'm attempting to run a cohort sequential model that has 4 outcome scores as well as linear, quadratic, and cubic growth factors.

Within age cohorts, I'm getting correlations that are > 1 and < -1. I suspect that it's these "out-of-bounds" correlations that are causing this error message:

WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IN GROUP 19 IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES. CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE C.

Also, all within-cohort correlations between intercept and other variables are listed as "999"; perhaps this is also causing the error message. The "999's" may be related to the fact that I had to constrain the intercept's variance to 0 in order to identify the model.

So, 3 questions:
1. What could be causing correlations that exceed 1?
2. Are there other things besides correlations > 1 that may be causing the error message?
3. What can I do to prevent this error message?

Thank you,
Jordan
 Linda K. Muthen posted on Friday, July 06, 2012 - 6:32 pm
 Jordan Silberman posted on Wednesday, July 11, 2012 - 8:13 pm
Dear Dr. Muthen,

Thanks so much for your quick response--much appreciated.

We're running a cohort sequential model with 43 age cohorts, each contributing 4 years of data to the curve. We hypothesize a cubic trajectory that occurs over several decades. Because each age cohort was followed for just 4 years, we don't expect to observe a cubic trajectory for any single cohort. This is somewhat analogous to predicting that the Earth's surface will appear round from space, but will appear flat when perceiving small areas of land on the ground.

So, to assess the model for each group separately, we ran linear growth models only, even though the combined cohort sequential model includes a cubic growth factor. Do you agree that this strategy is OK?

The cohort-specific linear models also show some correlations that exceed 1, and some that are "999." What can cause the correlations in the cohort-specific models to exceed 1? What can cause the program to be unable to estimate a value for some correlations (and to therefore output "999")?

Any advice you can offer, or any information you can direct us to, would be greatly appreciated.

Thanks,
Jordan
 Linda K. Muthen posted on Thursday, July 12, 2012 - 10:01 am