[R-sig-ME] Fwd: about graphical checking on glmm and contradiction between parsimony and AIC values

[Fernando Schmidt]

Now for all group.

Dear Glenda,

Recently, I ran a model selection using the dredge function in MuMin
package, which ranks the models according with AIC, AICc. I used this for
count data using mixed effect models.

This analyses follows the Information-theoretic approach (Burnham &
Anderson 2002), which don’t have a significance value “p” associated.
However, based on “delta” between the models and model “weight” you could
determine which model is the best approximated model for your data.

I am starting to use this approach and I am not so familiar with all the
terms and concepts. Maybe the information-theoretic approach could be
useful for your aims.

I attached some a paper that used it for count data. I hope this could help
you.

All the best,

Fernando


2011/12/13 glenda mendieta <glendamendieta at gmail.com>

> Dear list members,
> I am running GLMMs with count data, Laplace approximation, poisson family,
> using glmer {lme4}, the last version of R and R studio in windows platform.
> When fitting my final models, I run an anova to look at the "significance"
> of a term inclusion (random effect term), of course, this does not apply
> for random effects, but at least it gives me differences in DFs and whether
> the models are significantly different or not.
> It basically tells me that the least parsimonious model is the one with
> the lowest AIC value. As you can see below the difference in AIC values is
> pretty big.
>
> > anova(g,gwi)
> Data: db.e
> Models:
> gwi: abundance ~ census * avail.surface + (1 | tree) + (1 | spp)
> g: abundance ~ census * avail.surface + (1 | tree) + (1 | spp) +   (1 |
> spp:tree)
>        Df     AIC     BIC   logLik Chisq Chi Df Pr(>Chisq)
> gwi  12 22342.0 22442.9 -11159.0
> g      13  9702.5  9811.8  -4838.3 12641      1 < 2.2e-16 ***
>
> Then I run a qqplot on residuals vs. fitted values for each model and I
> can see that the more parsimonious model (*gwi*) is the one with the better
> fit (the points are pretty well alined across a straight line); whereas in
> the other model (*g*) the points are in a curved line.
> Would this be because random crossed effects should not be included as an
> interaction term (like in the last model [g]) (Johnson & Omland, 2004)? or
> I am overfiting?
>
> My intuition tells me I should go for the most parsimonious model, since
> the graphical checking works. But, I wonder if there any advise you can
> give me to improve the fit of this model?. There is still a lot of variance
> unexplained there, the model with the random effect interaction term
> "spp:tree" has a variance of 59.87 sd. 7.73, in comparison to the model
> with only tree and spp (2.45 sd. 1.55 & 3.5 sd. 1.90).
>
> Greetings and thanks in advance for your time,
>
> Glenda Mendieta-Leiva
> PhD candidate
> University of Oldenburg
>
> PS.  Johnson, J. B. and Omland, K. S. 2004. Model selection in ecology and
> evolution. Trends Ecol. Evol. 19: 101-108.
>
> ______________________________**_________________
> R-sig-mixed-models at r-project.**org <R-sig-mixed-models at r-project.org>mailing list
> https://stat.ethz.ch/mailman/**listinfo/r-sig-mixed-models<https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
>



-- 
Fernando Augusto Schmidt
Lab. de Ecologia de Comunidades
PPG - Entomologia. Universidade Federal de Viçosa
Viçosa, MG - Brazil 36570-000
www.labecol.ufv.br
+55 3182388810




-- 
Fernando Augusto Schmidt
Lab. de Ecologia de Comunidades
PPG - Entomologia. Universidade Federal de Viçosa
Viçosa, MG - Brazil 36570-000
www.labecol.ufv.br
+55 3182388810
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Barlow_et_al._2010_Dredge.pdf
Type: application/pdf
Size: 359533 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20111213/15e3446f/attachment-0004.pdf>