[R-lang] Singular Convergence in lmer on a medium-sized dataset

[John K Pate]

Hello all,

I'm currently trying to fit a linear mixed effects model (using lmer
from lme4) to a smallish to medium-sized dataset (about 10,000 items).
As I have seven continuous predictors and three random effects (two of
which are explicitly nested, i.e. talkerID and sentenceID), I'm using a
conservative stepwise forward selection procedure for my final model (I
can share the code if that would be relevant). This procedure allows for
random slopes and up to three-way interactions among fixed effects.

The final model fits fine, with no singular convergences. However,
during the model evaluation procedure, I get three singular convergence
warnings. My data contain no NAs, and every level of each random effect
has at least four data points. Should I be concerned about these
singular convergences, or do they not matter as they occur only on the
way to producing the final model? What steps should I take to prevent
these warnings?

I have tried running lme4a, but it seems to have a memory leak of some
sort. I run out of 4GB RAM before the model finishes examining just the
main effects.

Thank you!

John

==

John K Pate
Student, PhD Informatics
Informatics Forum 3.35
The University of Edinburgh
http://homepages.inf.ed.ac.uk/s0930006/


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.