[R-sig-ME] Model selection GLM vs. GLMMs

[Nelida Villaseñor]

Thanks Ben.

On: "I think you would find a bit of disagreement among experts about
the best procedure -- whether it would be to drop random effects until
you got a sensible non-singular fit, or to keep them in even though
they're singular"

Could you please suggest me some references supporting each of these procedures?

Cheers,
Nelida.


> Message: 1
> Date: Mon, 21 Apr 2014 00:15:04 +0000 (UTC)
> From: Ben Bolker <bbolker at gmail.com>
> To: r-sig-mixed-models at r-project.org
> Subject: Re: [R-sig-ME] Model selection GLM vs. GLMMs
> Message-ID: <loom.20140421T020951-335 at post.gmane.org>
> Content-Type: text/plain; charset=us-ascii
>
> Nelida Villasenor <nelida.villasenor at ...> writes:
>
>>
>> Dear all,
>
>> I'm performing model selection based on AICc on a set of GLMMs that
>> only vary in their fixed effects. The data comes from 12 transects
>> (5 measures along each transect), then each transect is modelled as
>> a random effect "+(1|transect)". As the response variable was
>> proportions (presences/n), I fitted the models using binomial family
>> and the total number of points (n) as weights.
>
>> Given that some models had boundary problems
>
>   meaning singular fits (estimated zero variances and/or +/- 1 correlations
> and/or values of estimated theta=0) ?
>
>> I ran the model
>> selection on a set of GLMs instead of GLMMs. The results were almost
>> identical in terms of the list of models with the highest support
>> (for 7 response variables where model selection was performed
>> independently).
>
>> I'm wondering which approach is correct? Or, as my results show, it
>> does not really matter, because the random effect does not change in
>> my GLMMs?
>
>   I think you would find a bit of disagreement among experts about the
> best procedure -- whether it would be to drop random effects until you
> got a sensible non-singular fit, or to keep them in even though
> they're singular.  Keep in mind that you should get the same estimates
> with a GLM or a GLMM if the variance estimates are all zero ... Since
> it doesn't sound as though it affects your einference/model selection
> on the fixed effects, I would say you could choose either approach
> (and explain clearly what you did).
>
>   Ben Bolker


-- 
Nélida R. Villaseñor
PhD Scholar
Fenner School of Environment and Society
The Australian National University