priors with MCMCglmm binomial analysis
Hi Massimo, I'm not sure, but I was under the impression that binomial analysis in MCMCglmm should be done with the family ="categorical". I don't know if this will make a huge difference, but hopefully it helps! Good Luck, Eryn Eryn McFarlane MSc Candidate Department of Integrative Biology University of Guelph Guelph, ON, Canada N1G 2W1
On 2011-12-07, at 11:14 AM, m.fenati at libero.it wrote:
dear R users,
I don't know how to set a MCMCglmm for binomial analysis. I'm trying to fit
several model but bad results were always observed if compared with the glm
results.
My model is:
tc<-matrix(c(124,184,33,18,86,9), ncol=2)
tco<-data.frame("specie"=c("bov","ov","cap"),"pos"= tc[,2],"tot"=tc[,1])
prior<-list(R=list(V=1,nu=0.002))
m.1<-MCMCglmm(cbind(pos,(tot-pos))~specie,prior=prior,data=tco,nitt=900000,
thin=100,burnin=300000,family="multinomial2",verbose=FALSE)
Large coefficient values and intervals occurred after fitting the model:
summary(m.1)
Iterations = 300001:899901
Thinning interval = 100
Sample size = 6000
DIC: 401.758
R-structure: ~units
post.mean l-95% CI u-95% CI eff.samp
units 594395401 0.0002638 1.418e+09 5106
Location effects: cbind(pos, (tot - pos)) ~ specie
post.mean l-95% CI u-95% CI eff.samp pMCMC
(Intercept) 59.24 -20274.73 22702.05 6000 0.697
speciecap -62.28 -33022.96 28154.63 6000 0.815
specieov -142.82 -28811.83 30635.06 6000 0.717
When I fit a glm model, different and more reasonable results are observed:
m.1<-glm(cbind(pos,(tot-pos))~specie, data=tco, family="binomial")
summary(m.1)
.........
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.7731 0.2549 -6.955 3.52e-12 ***
speciecap 0.7922 0.4667 1.698 0.0896 .
specieov 1.6424 0.2947 5.574 2.49e-08 ***
Where is the error? a mistake in the model specification, priors setting, ...
I dont'know!
I tried with other priors but the results never agree with the glm results.
Could someone help me?
Thanks in advance!
Regards
Massimo
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models