[R-sig-ME] help with model convergence

[Ben Bolker]

  [cc'ing r-sig-mixed-models: it's best to keep sending replies back to
the list so they can be archived and others can read them, or offer input]

On 10-11-01 01:28 PM, David Stainbrook wrote:
> Ben,
> 
> Thanks for your input. I read that article that you suggested and it
> appears they used SAS and Genstat to do their analysis. 

  Yes (although as I said at the time, I wouldn't actually trust the
methods that they used in Genstat for this problem.  I just think their
description of the problem is clear).

> Is it possible
> to use the Poisson-lognormal model in R or translate the model to this
> using R and lmer? Another professor mentioned that I may be able to get
> it to work using a negative binomial model in SAS or ADModel Builder.
> What do you suggest?

  Yes, you can use the Poisson-lognormal in recent versions of lme4,
simply by including an individual-level random variable.  You may get
warnings.
  You could indeed use a negative binomial model in SAS or AD Model
Builder (in ADMB you could also use the lognormal-Poisson model).

> Do you have any idea why Doug allowed the lmer function to fit
> quasipoisson if he doesn't feel that the results will be reliable? I
> would have trusted my results and wouldn't have had any idea that they
> might have been unreliable if he had not said that.

  I believe he implemented it a while ago and his opinions have now
changed. (I agree that it might be a good idea to disable this
functionality.)

> Also, do you have any idea how to increase both the default number of
> function evaluations and iterations with the control statement within
> the lmer model statement?
> The lme4 user's guide did not give an example for both.
> I tried this <control=list(maxIter=2000), control2=list(maxFN=3000)>,
> but I believe that only the default number of iterations was changed. I
> still got the error saying that it reached the function evaluation
> limit, so I don't think the second part actually worked. I stuck a "2"
> after control because otherwise I would get an error message saying that
> there were multiple control arguments used.

...,control=list(maxIter=2000,maxFN=3000),... should work. I would
expect "control2" would have exactly no effect.

  I went back and read your message
<https://stat.ethz.ch/pipermail/r-sig-mixed-models/2010q4/004653.html>.
 A few comments:
  * there was a bunch of HTML junk at the beginning of the file
(probably an accidental cut-and-paste error from Word) -- that may have
put people off from reading farther ...
  * I think your zip file may have been stripped -- the R mailing lists
don't allow every type of attachment, and zipped files can often contain
nasties.
  * you seem to be tackling a difficult problem.  I appreciate that
you're offering full details on your problem (full scripts and data),
but it's going to take someone else at least half an hour (and probably
quite a bit more) to get up to speed on what you're doing and what's not
working; unfortunately, that's more than most anyone has time for,
unless the problem happens to be something very close to their
interests. Unfortunately, you may well need to find local help for this
(your advisor? a friendly stats professor or graduate student?)
  * it's possible, depending on the complexity of your model, that
you're simply trying to fit too complicated a model.  You do have a lot
of data points, but some of your covariates may be strongly correlated.
Have you tried:
   - seeing if you can successfully fit a subset of the data points
(this could be faster, allowing you to debug quicker)?
   - seeing if you can successfully fit a subset of the covariates, or
which covariates or combinations of covariates are problematic?
   - seeing if you can successfully fit a non-mixed (GLM) model,
treating 'individual' as a fixed effect?
   - simulating data, possibly in a simplified form, to see if you can
get the right answer when you know what it is?
  * lme4 is quite finicky about convergence, on the philosophy that it's
better not to give an answer than to give a wrong one.

  R does have its advantages, but if you're up to working with SAS or AD
Model Builder I would recommend you also try those approaches -- see if
you run into the same problems.  But I would definitely try some of the
trouble-shooting strategies above, first.

  good luck,
    Ben Bolker


> Thanks again,
> 
> David

> 
> On Thu, Oct 28, 2010 04:49 PM, *Ben Bolker <bbolker at gmail.com>* wrote:
> 
>        My advice would be to use an individual-level random variable
>     (translating to a lognormal-Poisson model, which is qualitatively
>     similar to a negative binomial) -- see e.g. Elston et al 2001 for a
>     decent explanation, although you should not necessarily trust the
>     numeric methods they use ...
> 
>      [Elston, D. A., R. Moss, T. Boulinier, C. Arrowsmith, and X. Lambin.
>     2001. Analysis of Aggregation, a Worked Example: Numbers of Ticks on Red
>     Grouse Chicks. Parasitology 122, no. 05: 563-569.
>     doi:10.1017/S0031182001007740.
>     http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=82701.]
> 
> 
>       (Doug, thanks for the vote of confidence!)
> 
>       cheers
>         Ben
> 
> 
> 
> 
>