[R-sig-ME] Zero-inflated binomial (ZIB) models in glmmADMB:, warnings and errors

[Highland Statistics Ltd]

> ------------------------------
>
> Message: 2
> Date: Wed, 23 Oct 2013 15:48:42 +0100 (BST)
> From: Stefan Ferger <stefan.ferger at yahoo.de>
> To: "r-sig-mixed-models at r-project.org"
> 	<r-sig-mixed-models at r-project.org>
> Subject: [R-sig-ME] Zero-inflated binomial (ZIB) models in glmmADMB:
> 	warnings and errors
> Message-ID:
> 	<1382539722.1360.YahooMailNeo at web171701.mail.ir2.yahoo.com>
> Content-Type: text/plain; charset="utf-8"
>
>


> 1) None of the models shown above produces any warnings or errors when I run them with glmer (and thus ignore the zero-inflation).
> 2) When I sum all pecks on plot level and leave out cluster completely, I get much less zeros (but also much less replicates). When again modeled in glmer without controlling for zero-inflation the results are qualitatively the same as in 1), but the overdispersion is becoming much stronger (ratio ~ 3.5). Correcting for this by incorporating an observation-level random effect, however, leads to strong underdispersion (ratio ~ 0.3) and differing results.
>
> 3) I nevertheless think that the model structure should be determined by the design and therefore I would like to go for the version that appreciates the clusters, which means that I have to deal with these zeros.
>
> Does anybody know what I could do to fit these models in glmmADMBsuccessfully? Do I actually need to account for this zero-inflation, the consensus seems to be that one should do so if there are "over-proportionally" many zeros - but when is this threshold passed?
>
> As fitting the models successfully in glmmADMB might be tricky, does anybody know how the model structure should look like in one of the other packages, i.e. MCMCglmm (I found the model specification very tricky here) or R-INLA?
>
>
> Thank you for your help,
> Stefan



I would certainly include the plot effect, and cluster effect within plot. .....it imposes a correlation structure on your data.
Not sure whether I understand the color thing though.

If your binomial GLMM is overdispersed then you should try to figure out why this is. Common causes are:

1. Outliers
2. Non-linear patterns
3. Wrong link function
4. Wrong distribution
5. Too many zeros
6. Dependency structure that has not been included...or included in the wrong way.

For 2...plot your residuals vs each continuous covariate. If 5 is the cause, then yes...a zero inflated binomial is an option.
For 4 you may want to consider the beta-binomial distribution, see for example the book from Ben Bolker, or our 'Beginner's Guide to GLM & GLMM'; it
contains code to fit a beta-binomial GLMM in JAGS. A zero inflated version of a beta-binomial GLMM/GAMM requires similar code.


As to your question how many zeros means zero inflated models.....it all depends. I have data sets with 70% of zeros...and a
Poisson/NB GLM still do the job...and I have data sets where 25% of zeros already means ZIP. Same holds for ZIB.


I would strongly suggest to do this in JAGS.

Have fun

Alain








-- 
Dr. Alain F. Zuur
First author of:

1. Analysing Ecological Data (2007)
2. Mixed effects models and extensions in ecology with R (2009)
3. A Beginner's Guide to R (2009)
4. Zero Inflated Models and GLMM with R (2012)
5. A Beginner's Guide to GAM (2012)
6. A Beginner's Guide to GLM and GLMM (2013)

Highland Statistics Ltd.
6 Laverock road
UK - AB41 6FN Newburgh
Tel: 0044 1358 788177
Email: highstat at highstat.com
URL: www.highstat.com
URL: www.brodgar.com