[R-sig-Geo] Binom.krige.bayes strange uncertainty outputs

[Nicola Batchelor]

Dear list,

I'm having a few problems with my uncertainty outputs from the command
binom.krige.bayes in geoRglm.  I am getting sensible looking predictive
medians, and 2.5% quantiles, but the upper quantile and also the uncertainty
outputs are very odd looking (I have attached images of the predictive
median, plus the data points and also the upper limit of the 95% CrI to
illustrate the problem).  

My commands are shown below, but as far as I can see I haven't done anything
that would give this problem.

Has anyone experienced anything like this before?  I've been trying to get
the upper quantile from my prediction simulations to compare the results,
but haven't yet managed (still quite new to R and so these things take me a
while to work out!)...is it possible that the predictive simulations could
be correct and so I could take the uncertainty values from there?  Perhaps
it is some sort of memory problem?

########################

nicola1<-read.csv("MBGdataset.csv", header=TRUE)
nicola1 <-
as.geodata(nicola1,coords.col=3:4,data.col=1,covar.col=5:27,units.m.col=2)
names(nicola1)
attach(nicola1)

########################

#CREATE PREDICTION GRID
predpoints2km<-pred_grid(c(440, 590), c(130, 270), by=2)
plot(predpoints2km)
points(nicola1, add=TRUE)

covars<-read.dbf("covars2km.dbf")
names(covars)

#######################

#Phi = 60
#Set a random seed
set.seed(548)

#Set the prior control
Pri60<-prior.glm.control(beta.prior="flat", sigmasq.prior="uniform",
phi.prior="fixed", phi=60, tausq.rel=1)
#Set the MCMC control
Mc<-mcmc.control(S.scale=0.0133, burn.in=1000000, thin=1000, n.iter=5000000)
#Set the model control
Mod<-model.glm.control(trend.d=trend.spatial(~MARKET+MINLST+HCDIST,
nicola1),trend.l=trend.spatial(~MARKET+MINLST+HCDIST,covars),
cov.model="matern",kappa=0.5, lambda=0)
#Set the output control
Out<-output.glm.control(sim.posterior=TRUE,sim.predict=TRUE,
keep.mcmc.sim=TRUE, quantile=TRUE, inference=TRUE)

#Run the model with inputs (prior control, MCMC control, trend
specification, model control, and output control)
test60<-binom.krige.bayes(nicola1, prior=Pri60, model=Mod,
locations=predpoints2km, mcmc.input=Mc, output=Out)

########################

Thanks,

Nicola

-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: error.doc
Type: application/msword
Size: 79872 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20090803/f8f03317/attachment.doc>