1

I have a data set which contains cell counts across the 250 brain regions in 12 animals. I want to construct a correlation matrix and a graph theoretical model using the grouped lasso correction. However it says my matrix is not positive definite.

I have tried to use the caret package to select correlations above .xx in order to have a positive definite matrix. However, the highest correlation allowed between variables without throwing the error is .5. This reduces the number of columns from 250 to 8, which I can't use. Can anyone help with suggestions on how to move forward? Could this be do to the fact that I have many more variables than observations?

Antonio.Aubry
  • 41
  • 2
  • [This](https://stats.stackexchange.com/questions/30465/what-does-a-non-positive-definite-covariance-matrix-tell-me-about-my-data) might help! – kedarps Jun 05 '19 at 15:11
  • I've seen this thread before. My worry is if I do it one by one, it won't be different from using the caret package, ie: I'll only have 10 variables left. Any other ideas? – Antonio.Aubry Jun 05 '19 at 15:16
  • The thread at https://stats.stackexchange.com/questions/16327/testing-for-linear-dependence-among-the-columns-of-a-matrix/74328#74328 provides some solutions. It's unclear, though, what "it" is and why it is objecting to a non-pd matrix. The solution might just be to use different software, but you would need to provide more information before we could know. – whuber Jun 05 '19 at 15:20
  • Sorry, I'm using the qgraph package to construct the graphical analysis. The function "cor_auto" which is essentially a wrapper around the LavCor function. I could simply chose not to force a PD, but this is not suggested for the EBICglasso model I want to use. Perhaps its time to chose a different model? – Antonio.Aubry Jun 05 '19 at 15:42

0 Answers0