[R-sig-ME] singularity - differences in model output

Ben Bolker bbo|ker @end|ng |rom gm@||@com
Thu Jun 13 06:19:01 CEST 2019

  This new warning could be due either to changes in reporting or to
slight changes in the optimization defaults. The NEWS file (run
news(package="lme4") from inside R, or see
<https://github.com/lme4/lme4/blob/master/inst/NEWS.Rd>) tells you
what's changed, and when, in the package.

  In the CRAN version (1.1-21) the default tolerances for the
singularity check are *either* 1e-5 (isSingular) or 1e-4 (in the
built-in check); both change to 1e-4 in the development version (see
). In principle singular fits only occur when the estimated variance
is *exactly* zero, which is possible in lme4 (because it uses a
box-constrained optimizer with a bound at zero), but checking with a
tolerance of 1e-4 should also identify models that are very close to
singular (and might "really" be singular if there were no numeric
'fuzz' from finite-precision computation).

  The GLMM FAQ has more discussion of singular fits:

   Ben Bolker

On Thu, Jun 13, 2019 at 3:17 AM Alessandra Bielli
<bielli.alessandra using gmail.com> wrote:
> Hi everyone,
> I recently re-ran a model that I wrote about 5 months ago.
> The model is a GLMM of the type :
> m1 <- glmer(response ~ fixed  + random effect, data = d, family =
> "binomial")
> At that time, everything ran smoothly, with no error/warning messages.
> However I have just run it again and I get a message saying the model is
> singular.
> 1) The first question is, why didn't it happen then? Is it simply because
> the package was updated?
> I am asking because I need to justify my changes with the reviewers of the
> journal where I'd like to publish my study.
> 2) I used the function isSIngular() to check my model, using the default
> tolerance value. How do I decide the tolerance for detecting singularity?
> Thank you so much for your help
> Alessandra
>         [[alternative HTML version deleted]]
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

More information about the R-sig-mixed-models mailing list