[R-sig-ME] underdispersion glmer/MCMCglmm
Simone Messina
S|mone@Me@@|n@ @end|ng |rom u@ntwerpen@be
Tue Apr 20 10:36:27 CEST 2021
Dear all,
I am running binomial models based on infection rates, but I have some doubts which I hope you can kindly help me to solve.
Briefly, I want to test for the effect of forest type (2 levels) on birds� infection rate (proportion data). My dataset is composed by 2 rows per each species (one row per each type of forest). Sample sizes per species in each forest type are mostly below 10, except for one species over 20.
Question 1: Is it possible that my model is under-dispersed (see below)?
To deal with under-dispersion shall I get the model estimates as in Quasi-distribution?
glm1<-glmer(infection ~ forest + (1|species), data=dataset, family=binomial, weights = SampleSize)
> summary(glm1)
Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
Family: binomial ( logit )
Formula: infection ~ forest + (1 | species)
Data: dataset
Weights: Size
AIC BIC logLik deviance df.resid
96 100 -45 90 25
Scaled residuals:
Min 1Q Median 3Q Max
-1.2925 -0.4949 0.1130 0.6178 1.1473
Random effects:
Groups Name Variance Std.Dev.
species (Intercept) 1.763 1.328
Number of obs: 28, groups: species, 14
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.2969 0.4515 -0.657 0.511
forestUNL -0.5426 0.3875 -1.400 0.161
Correlation of Fixed Effects:
(Intr)
forestUNL -0.414
Then I test for under- / over-dispersion, and I detect under-dispersion:
> overdisp_fun(glm1)
chisq ratio rdf p
14.729 0.589 25.000 0.947
Hence, I calculated model estimates as in Quasi-distribution.
> printCoefmat(quasi_table(glm1),digits=3)
Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.297 0.347 -0.86 0.392
forestUNL -0.543 0.297 -1.82 0.068 .
Question 2: in case I decide to run the model with MCMCglmm (to include phylogeny), shall I consider under- / over-dispersion? How? Is there a way to get estimates as in Quasi-distribution in MCMCglmm models?
The only info I could get so far is that MCMCglmm automatically accounts for over-dispersion in count data with poisson distribution, but I don�t understand if this is valid also in case of proportion data with binomial family.
Thank you very much for your attention and explanations.
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models
mailing list