Dear Jarrod,
Thank you very much for your response. Unfortunately I still have issues
with convergence even when the residual variance is fixed. Part of the
problem I was having was the specification for visualizing the posterior
distributions of the phylogenetic portion of the analysis.
This may be a silly question, but what is the difference between:
plot(my2$VCV[,1])
plot(my2$VCV[,1]/(rowSums(my2$VCV)+pi^2/3)).
The two produce very obviously different graphs, and I was wondering
what exactly the second part of the latter plot is related to?
Even looking at the correct graph for the phylogenetic part of the
analysis, it seems I am having trouble with 'animal' converging, and
have massive issues with the variance. I have attached the posterior
distributions for both your example (PhyLogRegSIM.pdf) and my own data
(PhyLogRegDAT.pdf) to illustrate the problem. This occurs even when
running the models for a longer time.
I much appreciate your time and advice on this matter.
Thanks,
Joanna Baker
Below is the code for my model specification:
prior1=list(R=list(V=1, fix=1), G=list(G1=list(V=1, nu=1, alpha.mu=0,
alpha.V=1000)))
myAinv2<-inverseA(mytree)$Ainv
my2<-MCMCglmm(DV~IV, random=~animal, ginverse=list(animal=myAinv2),
family="categorical", prior=prior1, data=mydata,
nitt=10000000,thin=1000,burnin=100000)
plot(my2$Sol)
plot(my2$VCV[,1]/(rowSums(my2$VCV)+pi^2/3))
c2<-((16*sqrt(3))/(15*pi))^2
my2.int<-my2$Sol/(sqrt(1+c2))
my2.int
posterior.mode(my2.int)
summary(my2.int)
-----Original Message-----
From: Jarrod Hadfield [mailto:j.hadfield at ed.ac.uk]
Sent: 01 September 2012 20:06
To: Joanna L Baker
Cc: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] Phylogenetic Logistic Regression + MCMCglmm
Hi,
The most likely reason is that you have not fixed the non-identified
residual variance, although there are other possibilities. Here is an
example with a simulated tree and binary data:
tree<-rcoal(100)
x<-rnorm(100)
l<-rbv(tree, 1, nodes="TIPS")+x+rnorm(100) y<-rbinom(100, 1, plogis(l))
dat<-data.frame(y=y, x=x, species=tree$tip.label)
prior1=list(R=list(V=1, fix=1), G=list(G1=list(V=1, nu=1, alpha.mu=0,
alpha.V=1000)))
# residual variance fixed at 1.
Ainv<-inverseA(tree)$Ainv
m1<-MCMCglmm(y~x, random=~species, ginverse=list(species=Ainv),
family="categorical", prior=prior1, data=dat)
plot(m1$Sol)
# fixed effects should be zero and one
plot(m1$VCV[,1]/(rowSums(m1$VCV)+pi^2/3))
# phylogenetic ICC should be 1/(2+pi^2/3)=0.189. Note wdie credible
intervals - you need v.large phylogenies to get precise estimates with
binary data.
Cheers,
Jarrod
Quoting Joanna L Baker <J.L.Baker at 2008.hull.ac.uk> on Fri, 31 Aug 2012
11:39:08 +0100:
Hello all,
I am currently trying to use MCMCglmm to carry out a phylogenetic
logistic regression, with a single binary response and one or more
continuous or categorical predictors. However, I am having issues with
the phylogenetic component, 'animal', converging. This occurs when I
use different datasets, with high, moderate and low phylogenetic
signal.
I have tried improper (with the expected issues), proper, and
alpha-expanded priors but these do not seem to have much effect on the
results. I should note that I have no issues with convergence when my
response is a continuous variable.
Does anybody have a worked example with data and tree that I could use
to get started with phylogenetic logistic regression or could somebody
point me in the direction of any published work that has used the
package for such an analysis successfully? I'll much appreciate any
help with this.
Thanks again,
Joanna Baker
University of Hull
Cottingham Road
Hull
East Yorkshire
j.l.baker at 2008.hull.ac.uk
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
-------------- next part --------------
**************************************************
To view the terms under which this email is
distributed, please go to
http://www2.hull.ac.uk/legal/disclaimer.aspx
**************************************************