Model is nearly unidentifiable with glmer
Dear David,
On Wed, 26-02-2014, at 01:26, David.Duffy at qimr.edu.au wrote:
On Tue, 25 Feb 2014, Ramon Diaz-Uriarte wrote:
http://ligarto.org/rdiaz/nearly-unidentifiable.tar.gz a small (230 KB) file with: - data - R code to show the problem - two runs of that code on different machines
I don't seem to get those warnings when I run your "f02" ("very large
eigenvalue") problem here - get the same parameter estimates (x86_64,
lme4_1.0-6, Matrix_1.1-0). The estimated RE variance is very small.
I think the main difference can be the version of lme4: I am using 1.1-4 (and Matrix v. 1.1-2) and I think between v.1.0 and 1.1 of lme4 there is an increase in the reporting of warnings (I think I've got this impression from messages in the list and from the comments at the github repository ---e.g., https://github.com/lme4/lme4/issues/120).
Your cut-down "singularity" example gives a variance for dataSetID with a starting value for theta, eg glmer(nS ~ Method + (1|dataSetID), family=poisson, data = d2, start=0.01) - again v. small.
If I do that (again, using lme4 v.1.1-4 and of course with the check.conv.singular set to "warning") I still get the "singular fit" warning (though not if I use bglmer, from package blme). Best, R.
| David Duffy (MBBS PhD) | email: David.Duffy at qimrberghofer.edu.au ph: INT+61+7+3362-0217 fax: -0101 | Genetic Epidemiology, QIMR Berghofer Institute of Medical Research | 300 Herston Rd, Brisbane, Queensland 4006, Australia GPG 4D0B994A
Ramon Diaz-Uriarte
Department of Biochemistry, Lab B-25
Facultad de Medicina
Universidad Aut?noma de Madrid
Arzobispo Morcillo, 4
28029 Madrid
Spain
Phone: +34-91-497-2412
Email: rdiaz02 at gmail.com
ramon.diaz at iib.uam.es
http://ligarto.org/rdiaz