The problem seems to be that the algorithm for coming up with a
starting guess for the phi (dispersion) parameter is getting a negative
number. It's not all that easy to figure this out ... The data set is a
little bit nasty (lots of points stacked on the equivalent of (0,0)),
but not pathological.
I'm cc'ing the maintainer of the package --
adding the lines
if (sum(phi_y)<0) {
stop("bad estimated start value for phi: consider setting start
values manually\n",
"(see ?betareg.control)\n",
"Estimated starting values for
mean:\n",paste(beta,collapse=","))
}
at line 162 of betareg.R in the current CRAN version provides
a more informative error message in this case.
See below for solutions.
=============
results <- read.csv2("betareg_tmp.csv")
results$drugcat <- cut(results$drug,c(0,0.005,0.06,0.17))
table(results$drug)
table(results$alcoh)
table(results$cond)
## shows that fairly large fractions of the data are
## in the lowest category:
## 165/209=0.001 'drug'
## 54/209=0.001 'alcoh'
## 38/209=0.001 'cond'
## so this will be a fairly challenging problem in any case
library(ggplot2)
ggplot(results,aes(x=alcoh,y=cond))+stat_sum(aes(size=..n..),alpha=0.7)+
facet_wrap(~drugcat)+theme_bw()
library(betareg)
## set phi link to logarithmic
## basic problem (digging through betareg.fit etc.) is
## that initial estimate of phi, based on
## linear model of logit(cond) ~ alcoh + cond, is NEGATIVE ...
## doesn't seem to be any way to override this starting value
## brute force
try(gyl<-betareg(cond ~ alcoh + drug, data=results,
link.phi="log"))
## pick through, debugging ... find starting values used
svec <- c(-1.6299469,0.8048446,1.7071124,0)
gyl<-betareg(cond ~ alcoh + drug, data=results,
link.phi="log",
control=betareg.control(start=svec))
## would work fine with more generic starting values
svec2 <- c(qlogis(mean(results$cond)),0,0,0)
gyl2<-betareg(cond ~ alcoh + drug, data=results,
link.phi="log",
control=betareg.control(start=svec2))
## before I got that to work, I also tried this (which
## will be slower and less efficient but is a useful
## alternative
library(bbmle)
## define a variant parameterization of the beta distribution with
## m=a/(a+b), phi=(a+b)
dbeta2 <- function(x,m,phi,log=FALSE) {
a <- m*phi
b <- phi*(1-m)
dbeta(x,shape1=a,shape2=b,log=log)
}
m1 <- mle2(cond~dbeta2(m=plogis(mu),phi=exp(logphi)),
parameters=list(mu~alcoh+drug),
data=results,
start=list(mu=qlogis(mean(results$cond)),logphi=0))
summary(m1)
p1 <- profile(m1)
plot(p1,show.points=TRUE)
confint(p1)
confint(m1,method="quad") ## not much difference
coef(m1)
coef(gyl)
On 11-03-13 12:59 PM, Vlatka Matkovic Puljic wrote:
> Sorry, here is my data (attached).
>
> 2011/3/12 Ben Bolker <bbol...@gmail.com <mailto:bbol...@gmail.com>>
>
> Vlatka Matkovic Puljic <v.matkovic.puljic <at> gmail.com
> <http://gmail.com>> writes:
>
> >
> > That was also my first thought.
> > But I guess it has something to do with W and phihat
> > (which I'm struggling to check
>
> Again, it would help to post a reproducible example ...
> hard to debug/diagnose by remote control. If you can't
> possibly post the data to the list, or put them on a web
> site somewhere, or randomize them a bit so you're not
> giving anything away, or find a simulated example that
> shows the same problem, you could as a last resort send them
> to me.
>
> Ben Bolker
>
> ______________________________________________
> R-help@r-project.org <mailto: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.
>
>
>
>
> --
> **************************
> Vlatka Matkovic Puljic
> +32/ 485/ 453340
>
______________________________________________
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.
