INLA_spatio-temporal models

[a_r]

Dear all, 

I've been running a spatio temporal model in R-INLA. the time scope is 46 years, and it's yearly data. 

The models run pretty smoothly for the Knorr-Held formulation described in "spatio and spatio-temporal bayesian models with R-INLA", with the following code:

control <- list(

  predictor = list(compute = TRUE),

  results = list(return.marginals.random = TRUE, return.marginals.predictor=TRUE),

  compute = list(hyperpar=TRUE, return.marginals=TRUE, dic=TRUE, mlik = TRUE, cpo = TRUE, 

                 po = TRUE, waic=TRUE, graph=TRUE, gdensity=TRUE, openmp.strategy="huge"), 

  group = list(model="rw2"))

#non-parametric dynamic trend

#knorr-held formulation

formula2 <- numberTorn ~1 + f(ID, model = "bym", graph=tornb.inla) + f(Year, model="crw2")+

  f(Year2, model="iid")

modelsptp2 <- inla(formula = formula2, family = "poisson", 

                   quantiles = c(.05, .5, .95), 

                   data = TornInla,

                   control.compute = control$compute,

                   control.predictor = control$predictor)

summary(modelsptp2)

But I also would like to understand the space-time interactions, so I started to dig into it, and this is what happens:

#SP-TIME INTERACTION

#TYPE I

TornInla$area.year <- seq(1,length(countyTorn))

formTypeI <- numberTorn~ + f(ID, model="bym", graph=tornb.inla)+

  f(Year, model="crw2") + f(Year2, model="iid")+

  f(area.year, model="iid")

mod.intI <- inla(formTypeI,family="poisson",data=TornInla,

                 control.predictor=list(compute=TRUE),

                 control.compute=list(dic=TRUE,cpo=TRUE))

--------------- type I works fine

#Type II

TornInla$area.int <- TornInla$ID

TornInla$year.int <- TornInla$Year

formTypeII <-numberTorn ~ + f(ID, model="bym", graph=tornb.inla)+

  f(Year, model= "rw2") + f(Year2, model="iid") + 

  f(area.int, model="iid", group = year.int, 

    control.group=list(model="rw2"))

mod.intII<-inla(formTypeII,family="poisson",data=TornInla, 

                control.predictor=control$predictor, 

                control.compute=list(dic=TRUE,cpo=TRUE))

----------------------throws an error:

Warning in inla(formTypeII, family = "poisson", data = TornInla, control.predictor = control$predictor,  :

  f(area.int, ...) :  There is no indices where group[]=1, this is *usually* a misspesification

Warning in inla(formTypeII, family = "poisson", data = TornInla, control.predictor = control$predictor,  :

  f(area.int, ...): Number of unused groups >= the number of groups used: 1969 >= 47, this is *usually* a misspesification

Error in inla.inlaprogram.has.crashed() : 

  The inla-program exited with an error. Unless you interupted it yourself, please rerun with verbose=TRUE and check the output carefully.

  If this does not help, please contact the developers at <he...@r-inla.org>.

In addition: Warning message:

running command '"C:/Users/rp/Documents/R/win-library/3.3/INLA/bin/windows/64bit/inla.exe"  -b -s -v  "C:/Users/rp/AppData/Local/Temp/RtmpY3Jd2d/file1df8390cd9/Model.ini"' had status 1 

> 

#Type III

formTypeIII<- numberTorn~+ f(ID, model="bym", graph= tornb.inla)+

  f(Year, model="crw2") + 

  f(Year2, model="iid") +

  f(year.int, model="iid", group =area.int, 

    control.group=list(model="besag", graph = tornb.inla))

mod.intIII <- inla (formTypeIII,family="poisson",data=TornInla, 

                    control.predictor=control$predictor, 

                    control.compute=control$compute)

###Type IV 

formTypeIV<- numberTorn ~+ f(ID, model="bym", graph=tornb.inla)+

  f(Year, model="crw2") + f(Year2, model="iid") + 

  f(area.int, control.group= list(model="rw2"))

mod.intIV <- inla (formTypeIV,family="poisson",data=TornInla, 

                   control.predictor=control$predictor, 

                   control.compute=control$compute)

Type III and IV just take too long (more than 12 hours so far). I even set in control.compute   openmp.strategy="huge", and I am running the scripts in MRO, which is supposelly faster than R, but still it isn't fast enough. Is this normal, or is there something wrong with the code? And why does it crash when I try to run the type II?

Any advice is welcome!

Thanks in advance, 

Angela

[Haakon Bakka]

Partial answer:

To speed up spacetime, use 

           control.inla=list(int.strategy='eb')

(No integration of hyper-param)

and select good staring values, to input in

control.mode=list(theta=c(   <starting values in internal scale>   ), 

                             restart=TRUE), 

See the SPDE tutorial for more space-time code.

Kind regards,

Haakon

--
You received this message because you are subscribed to the Google Groups "R-inla discussion group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to r-inla-discussion-group+unsub...@googlegroups.com.
To post to this group, send email to r-inla-discussion-group@googlegroups.com.
Visit this group at https://groups.google.com/group/r-inla-discussion-group.
For more options, visit https://groups.google.com/d/optout.

[Finn Lindgren]

For case 2, focus on the warning messages; the error is very likely an effect, and what the warnings are about is the cause.

Finn

Finn

To unsubscribe from this group and stop receiving emails from it, send an email to r-inla-discussion...@googlegroups.com.
To post to this group, send email to r-inla-disc...@googlegroups.com.

[Paolo Girardi]

I think you have to scale the temporal variable starting from one.

To post to this group, send email to r-inla-disc...@googlegroups.com.
Visit this group at https://groups.google.com/group/r-inla-discussion-group.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "R-inla discussion group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to r-inla-discussion-group+unsub...@googlegroups.com.

[Finn Lindgren]

Indeed, good catch Paolo!

All indexed models start on 1, so a model with years 1950,1951,..,1960, for example, needs to have indexes 1,2,...,11. Otherwise the model will have 1960 timesteps instead of just 11.

Finn

Finn

To unsubscribe from this group and stop receiving emails from it, send an email to r-inla-discussion...@googlegroups.com.

[Håvard Rue]

On Tue, 2017-03-28 at 14:24 +0100, Finn Lindgren wrote:
> Indeed, good catch Paolo!

Indeed!!!!

--
Håvard Rue
Department of Mathematical Sciences
Norwegian University of Science and Technology
N-7491 Trondheim, Norway
Voice: +47-7359-3533 URL : http://www.math.ntnu.no/~hrue
Mobile: +47-9260-0021 Email: havar...@math.ntnu.no

R-INLA: www.r-inla.org

[Angela Rodrigues Afonso]

Great! 

Thank you all.