Positive definite vs. positive semidefinite required for phylogenetic regression - error message in brms

Interfaces brms
1

Hi!

Running R v. 3.6.0, rstan v. 2.19.3, brms v. 2.13.0 on Ubuntu 18.04.2 LTS.

I’m trying to run a relatively simple phylogenetic regression for some COVID work (I’m afraid it involves confidential patient data, so I can’t share it). Normally I’d just code this up in a full Stan model, but seeing as it was simple, I thought I’d give brms a go.

I’m getting an error when I try to run it (in the test case of the minimal model with just the intercept and the phylogenetic random effect):

Error: Within-group covariance matrices must be positive definite.

Now phylogenetic covariance matricies are only guarenteed to be positive semi-definite (and this one is positive semi-definite not positive definite), I was wondering, is the positive definite condition required for some computational reason I’m missing, or is this just a case of the test being applied being overly strict?

Thanks!

2

brms has to enfore positive definitness to reliably get the choselky factor of the covariance matrix. You may be able to move your matrix from positive semi-definite to positive definite by adding a very small value to the diagonal.

3 Likes
3

Thanks for getting back so fast. Adding 0.00000002 to the diagonal did it.

1 Like