[cc'ed back to r-help]
I've started to take a look, and there's nothing immediately obvious
about the problem with the fit (the warnings and errors are about a
"non-positive-definite Hessian", which usually means an
overfitted/poorly identified model) -- still working on whether there's
a way to get more useful information.
As it turns out, glmmADMB's default behavior in this case is to just
stop with an error. I could re-write things to allow it to tell you
something (but it wouldn't be able to compute standard errors on the
parameters).
It's conceivable that you might get slightly different results
(different enough that it would work in one case and fail in another) on
different operating systems, different machines, etc. ... because the
failure is really a case of something that's numerically on the edge,
where negative values in a vector are illegal but the minimum might be
-1e-7 in one case and 1e-7 in another -- i.e., different but just by
numeric "fuzz".
My first response to this sort of problem would be to see if you can
simplify the model slightly -- e.g. are both random effects really
necessary, etc..
Several things do suggest themselves from looking at the data:
* your response variable ranges from 0 to 1, which doesn't make much
sense for a negative binomial response distribution. What does your
response variable represent? If it is overdispersed *binomial* data,
then (1) you need to use something like a beta-binomial distribution
instead (unfortunately not yet implemented in glmmADMB, but if you were
desperate I might give it a whack) (2) you need to specify the
denominators -- just the proportions won't do it. If you have just
proportions, then you might be looking for a beta distributed response
(although that has its own trickiness in dealing with response values of
exactly 0 or 1).
* your "Site" variable has only two levels. That makes dealing with
it as a random variable extremely questionable (from a *philosophical*
or experimental-design point of view it may be sensible to treat it as a
random variable, but not practical: see http://glmm.wikidot.com/faq for
discussion of this point). When I made it a fixed effect instead, I got
(apparently) sensible results.
* The plot of the data (see attached, I hope the attachment gets
through) makes it clear that you have some potential balance problems.
In section/segment combinations (C-D) x (3-5) you have only a few (often
only one) observation in the (beetle.ev=1) case. Mild lack of balance is
no problem in mixed models, but such severe lack of balance can be.
(Note this doesn't guarantee a problem, but it is one of the factors
that makes estimation less stable.)
On 11-10-16 05:41 PM, James McCarthy wrote:
> Hello Ben,
>
> Many thanks for offering to help with this. I have attached the data set
> “stain.csv”. I have also attached a script file with all of the code
> that should (hopefully) work for you.
>
> What I would like to get working is the function:
>
> nbin1<-glmmadmb(stainp~beetle.ev+Caged*Section/Segment+(1|Site)+(1|Log.code),data=dat1,family="nbinom")
>
> The issue is with the “beetle.ev” column (binary, presence/absence of
> insects). Below is some code that works, all that has changed is that
> “beetle.ev” is substituted with “Test”. This one works, and has some
> data in pretty much the same format as “beetle.ev” (as you can see in
> the .csv file).
>
> nbin2<-glmmadmb(stainp~Test+Caged*Section/Segment+(1|Site)+(1|Log.code),data=dat1,family="nbinom")
>
> For the life of me, I can’t figure out what’s wrong with this column.
>
> Thanks again for talking a look at it.
>
> Cheers,
> James
>
>
> James McCarthy
> MSc (Ecology) candidate, University of Canterbury
> c/o Scion (New Zealand Forest Research Institute)
> Forestry Rd, P.O. Box 29-237, Christchurch, New Zealand
> Ph: (03) 364 2987 Ext. 7820
> Cell: 027 268 5506
> Fax: (03) 364 2812
> Email: james.mccar...@scionresearch.com or
> james.mccar...@pg.canterbury.ac.nz
> www.scionresearch.com
>
> This email may be confidential and subject to legal privilege, it may
> not reflect the views of the University of Canterbury, and it is not
> guaranteed to be virus free. If you are not an intended recipient,
> please notify the sender immediately and erase all copies of the message
> and any attachments.
>
> Please refer to http://www.canterbury.ac.nz/emaildisclaimer for more
> information.
>
##
install.packages("glmmADMB",repos="http://glmmadmb.r-forge.r-project.org/repos",type="source";)
library(glmmADMB)
dat1<-read.csv("stain.csv")
## attach(dat1) ## dangerous
names(dat1)
dat1<-transform(dat1,
Site=factor(Site),
## BMB: fixed typo
Log.code=factor(Log.code),
Caged=factor(Caged),
beetle.ev=factor(beetle.ev),
Test=factor(Test))
## BMB: this could also be done via colClasses= argument to read.csv
nbin1<-glmmadmb(stainp~beetle.ev+Caged*Section/Segment+(1|Site)+(1|Log.code),data=dat1,family="nbinom",
verbose=TRUE)
nbin1B<-glmmadmb(stainp~beetle.ev+Caged*Section/Segment+Site+(1|Log.code),data=dat1,family="nbinom",
verbose=TRUE)
## install.packages("coefplot2",repos="http://r-forge.r-project.org";)
coefplot2(nbin1B)
library(ggplot2)
ggplot(dat1,aes(x=beetle.ev:Caged,y=stainp,colour=Site))+
stat_sum(aes(size=..n..),alpha=0.5)+
facet_grid(Segment~Section)+opts(panel.margin=unit(0,"lines"))
summary(nbin1)
nbin2<-glmmadmb(stainp~Test+Caged*Section/Segment+(1|Site)+(1|Log.code),data=dat1,family="nbinom")
summary(nbin2)
library(coefplot2)
coefplot2(nbin2)
VarCorr(nbin2)
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
