[R-sig-ME] starting value setting in glmer

[Ben Bolker]

  [It's admittedly a judgement call, but it is in general preferable,
  and especially if you think your problem *might* be a general one,
  to send first to the r-sig-mixed-models at r-project.org mailing list,
  where (a) the questions and answers will be visible to everyone, and
  archived publicly for future reference, and (b) someone other than
  the lme4 maintainers might be able to answer.  Forwarding ...]

On 14-05-19 11:39 AM, Yuanbo Song wrote:
> Hi Ben,
> 
> I am currently using lme4.0 for my research and I really appreciated
> your and other author's contribution for this package. Unfortunately, I
> encoutered a problem about starting value setting when I used glmer
> function. I have five fixed effects and seven random effects, so I set
> my starting value as "start=list(fixef=start.fixef, theta=start.theta)",
> where start.fixef is a vector of 6 elements (including intercept and
> five fixed effects) and start.theta is a vector of 7 elements indicating
> the 7 random effects. However, I got a warning message saying:
> 
>  In sort(names(start)) == sort(names(FL)):
>  longer object length is not a multiple of shorter object length.
> 
> I do not quite understand why I got this warning message. Did I
> misunderstand the syntax of "start = "? If so, what should be done in my
> situation.
> 
> Thanks,
> Bob

This looks like a bug in lme4.0; at first glance it is just a
false-positive test, but on second glance (looking through the
internals of lme4.0) I'm not at all sure that I would trust
the "start" argument of lme4.0 at all for GLMMs without significant
testing on your side to ensure that it's giving sensible answers.
On a cursory glance, it looks like a lot of the code was copied
from the LMM functionality and may never have been tested properly
for GLMMs.  (It's possible that the code is cleverer than I think
and is actually working just fine ...)

  However, it looks in this case as though you're using the
syntax from lme4 >= 1.0.0 (not lme4.0 syntax): if you read the
help page for ?glmer in lme4.0 you'll see it's quite a bit different.

  More broadly, I would wonder why you're using the old version
of the package -- there *are* reasons to use it (that's why we
made it available in the first place), but I would strongly encourage
anyone doing a new analysis to use the most recent version(s) -- and
if they don't work for you for some reason, please let the maintainers
know so that we can try to resolve those issues!

  cheers
    Ben Bolker