Hi, I'm trying to estimate a multivariate ACE model. I started by checking a correlated factors solution (umxACEv), but there are some correlations that just can't be real (e.g., 2.02).
| rA1 | rA2 | rA3 | rA4 | rC1 | rC2 | rC3 | rC4 | rE1 | rE2 | rE3 | rE4 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| x | 1.00 | 1.00 | ||||||||||
| y | -0.40 | 1 | 0.17 | 1 | ||||||||
| z | 0.03 | 0.38 | 1 | -0.03 | -0.03 | 1 | ||||||
| w | 0.15 | 2.02 | -0.21 | 1 | 0.06 | -0.41 | -0.07 | 1 |
Standardized variance-based models may yield negative variances...
Warning message:
In sqrt(I * C) : NaNs produced
So I checked what are the computed correlations in a cholesky model (umxACE), and there are different correlations, in this case they are all in bounds, but still a correlation of 0.98 between two completely different variables, that were measured in two different ages, seem to be really weird. Furthermore, there are a lot of estimates that are drastically different both in magnitude (C correlations) and in direction, between the two models.
| rA1 | rA2 | rA3 | rA4 | rC1 | rC2 | rC3 | rC4 | rE1 | rE2 | rE3 | rE4 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| x | 1.00 | 1.00 | 1 | |||||||||
| y | -0.39 | 1 | -0.11 | 1 | 0.16 | 1 | ||||||
| z | -0.07 | 0.27 | 1 | -0.97 | -0.15 | 1 | . | -0.02 | 1 | |||
| w | -0.20 | 0.98 | 0.25 | 1 | -1.00 | 0.17 | 0.95 | 1 | 0.08 | -0.33 | -0.14 | 1 |
I'd appreciate any thoughts of why such values can occur.
Thanks!
Printer-friendly version
please share the MZ and DZ correlation tables.
I'm guessing this is some kind of oddity from small-n data
umxSummarizeTwinData()will generate the needed tables to say more.This is the correlation table from umxSummarizeTwinData().
However, w comes from a different wave than the x and y. So when checking the overlap for w and y for example (the most out of bound correlation), there are only 15-17 families overlapping in MZ twins, so I'm guessing this is the explanation for the out of bounds correlations..
Thanks!
Based on the above, my guess is that the out-of-bounds correlations are a result of the A, C, and E matrices not being positive-definite under the parameterization (direct-symmetric?) being used. Since those matrices are not proper covariance matrices, their standardized forms won't be proper correlation matrices. There was a recent thread about that specifically.