[R-sig-ME] pwrssUpdate Error with new version of lme4

[Ben Bolker]

Johannes Radinger <johannesradinger at ...> writes:

> Hi all,
> 
> I'd like to follow up an issue which had already been discussed some weeks
> ago
> about the pwrss error in the new version of lme4:
> https://github.com/lme4/lme4/issues/134
> Thanks to Ben Bolker who had worked to solve that problem and implemented
> solutions in the release branch in github.
> 
> I tried that version from github (lme4_1.0-5) but run again/still in
> problems (which I didn't have with the older lm4 version). Again, I can
> provide some data (see below) to reproduce following error:
> Error: (maxstephalfit) PIRLS step-halvings failed to reduce deviance in
> pwrssUpdate
> 
> The error occurs only when I include both predictors and all species:
> mod <- glmer(presabs~pred1+pred2+(1|species),family=binomial,data=mydf)
> 
> Models with single predictors work:
> mod1 <- glmer(presabs~pred1+(1|species),family=binomial,data=mydf)
> mod2 <- glmer(presabs~pred2+(1|species),family=binomial,data=mydf)
> 
> when I remove species "Rutiilus" the model also works:
> mod3 <-
> glmer(presabs~pred1+pred2+(1|species),family=binomial,
>   data=mydf[mydf$species!="Rutiilus",])
> 
> So there seems and issue with the combinations of that single species and
> the two predictors.
> However that was working in lme4_0.999999-2 (except for warnings: "glm.fit:
> fitted probabilities numerically 0 or 1 occurred ").
> 

 [snip snip snip ]

  In our defense, these are awfully messy data -- most of the predictor
values are concentrated very near zero, with a few values that are
many orders of magnitude larger ... and there seems to be an issue
of complete separation/large parameter values.  The proximal problem
was another underflow issue, which I have fixed in the development
branch on github.  The change also (mostly) fixes the other examples
reported at https://github.com/lme4/lme4/issues/138 , although they
still take a long time to run and end with a warning about the maximum
number of function evaluations being exceeded ...
  (This example no longer gives a warning with glmer: I'm not sure
whether it should or not)