[R-sig-ME] GLMM for data with many zeros: poisson, lognormal-poisson or zero-inflated?
Alicia Valdés
aliciavaldes1501 at gmail.com
Thu May 12 12:35:38 CEST 2011
Dear Ben,
Thanks for your response. As I am not very used to these kind of lists, I posted
this response to my own message at first, so I will try to answer properly to
your message now (I hope this works).
Looking at what you said, the variance in my random factor "site" is nearly zero
in most of the cases (the higher value is 0.1). Do you think that then I can say
that differences in site are not important? Because I was thinking of just
removing "site" from the model and then fit a GLM instead of GLMM (as I will
have only fixed effects).
What do you mean when you say that "if the means are low enough then
zero-inflation may not be necessary"? Do you refer to the means of the "y"
values without zeros? And which value should I take as "low enough"? I believe
that the means of the response variable are around 2 with the zero values
included, and around 3-4 without the zeros. Is it low enough to not include
zero-inflation? Anyway, if I remove the random factor, I will try to fir a
zero-inflated GLM with zeroinfl in pscl package, and compare it with a poisson
or negative binoimial model. Should I have any problem in comparing AIC between
these different kinds of models?
I am still a bit confused about overdispersion. You said that in order to
evaluate if there was overdispersion in my model, I should look at the residual
deviance or sum of squared Pearson residuals, but these values are not given by
lmer. I only got a "deviance" value. How can I calculate these values, and which
threshold should I consider in order to say that I have (or not) overdispersion
in my data?
The "revisions" in my data are just temporal repetitions of the sampling. I
mean, I did different samplings of the same experimental units in different
dates, and I call each of these samplings a "revision". Anyway, I think that
fitting a model for each revision should be fine, because I can compare the
significant factors between them, as I don't have any specific hypothesis about
how the effects should change temporally.
Again, thank you very much for your time and suggestions!
Alicia Valdés Rapado
PhD Student
Dpto. Biología Organismos y Sistemas - Unidad de Ecología
Universidad de Oviedo
España - Spain
e-mail: valdesalicia.uo at uniovi.es
More information about the R-sig-mixed-models
mailing list