Separable space-time model + prediction via sampling: coding query

[Julianne Meisner]

Hello, 

I'm sure there's something I'm doing wrong but I can't figure it out. I'm fitting a separable space-time model in R-INLA, with the goal of doing prediction via sampling using inla.posterior.sample. The first issue has to do with the model fitting, which I am doing as follows:

knots=min(cattle_data@data$time):max(cattle_data@data$time)

mesh = inla.mesh.2d(loc.domain = loc.mesh, 

                    max.edge = c(0.3,0.6), 

                    offset = c(1, 1))

spde_RE<-inla.spde2.matern(mesh, alpha=2) 

A<-inla.spde.make.A(mesh, loc=coords_cattle, group=cattle_data@data$time, n.group=length(knots))

iset <- inla.spde.make.index('i', n.spde=spde_RE$n.spde, n.group=length(knots))

stk.dat.c<-inla.stack(

  data=list(resp=cattle_data@data$n.cattle), 

  A=list(A,1), 

  effects=list(c(iset, list(intercept=1)), list(cattle_data@data)), 

  tag='est')

pcprior<-list(theta = list(prior = "pc.prec", params = c(3, 0.05)))

f.spde <- resp ~ 0 + intercept + urban + f(i, model=spde_RE, hyper = pcprior, group=i.group, control.group=list(model="ar1")) 

res<-inla (f.spde, data=inla.stack.data(stk.dat.c), family="zeroinflatedpoisson1", offset=log(cattle_data@data$mem.cattle), control.predictor=list(A=inla.stack.A(stk.dat.c)), 

                      num.threads = inla.getOption("num.threads"), control.compute=list(config = TRUE), verbose=TRUE)

-------------------------

The dataset has 6,290 rows, the mesh has 1,428 vertices, and there are 17 knots in total. There is at least one row in the dataset (cattle_data) corresponding to each knot (i.e., each "time" value). My goal is to do prediction via sampling using inla.posterior.sample. 

The A matrix is 6290 x 24276 as expected, but res$summary.random$i has 1,428 rows (that is, equal to the number of mesh vertices). I believe this should have # rows = # mesh vertices * # knots, that is 24,276. Is that correct? Similarly, the corresponding entry in res$misc$configs$contents$length is 1,428. 

The second issue has to do with the mesh for prediction. My code for this is:

Aprd<-inla.spde.make.A(mesh, loc.pred, group=kIdx, n.group=length(knots))

Where "kIdx" is the group index (1-17), and there are 90,000 locations for prediction. This is a 1428 x 24276 matrix. 

My thought is if the corresponding draw from inla.posterior.sample is of length 24276, I can then do draw%*%Aprd for prediction. Is my Aprd code a matrix to project from the space-time mesh, to a space only mesh at the indicated knot, and then I need a second matrix to project to the prediction locations? 

Alternatively, if you can point me to an example of sampling via prediction on a separable space-time model, that would be helpful. As you may be able to tell, I am not a statistician!

Thanks,

Julianne

[Elias T. Krainski]

What is the output of

length(knots)

Elias

--
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...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/r-inla-discussion-group/a9ab763f-9b38-4b85-8490-754d67ebb7d3%40googlegroups.com.

[Julianne Meisner]

Any other assistance on this issue would be much appreciated-- I haven't made any headway in the past 4 days. 

Thanks,

Julianne

[Elias T. Krainski]

Dear Julianne,

It is hard to tell what is the problem without reproducible code. If you provide it, we can try and see what is going  on.

Best regards,

Elias

--
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...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/r-inla-discussion-group/69ea1e9b-83be-4f32-a198-bd775e2d468f%40googlegroups.com.

[Julianne Meisner]

Thanks, Elias, I was able to get back to this today and discovered it was an issue in how I was building my stack, which I have now resolved. 

I still have one remaining question, which I hope will be more straightforward to answer. When predicting by sampling (using inla.posterior.sample) using a space-time model, I should be building my projection matrix as Aprd<-inla.spde.make.A(mesh, loc.pred, group=kIdx), where kIdx is the index for the group (time) I want to do prediction for and loc.pred=# prediction locations, correct? The resulting matrix will have nrows=loc.pred and ncols=mesh$n.

Many thanks,

Julianne

To view this discussion on the web, visit https://groups.google.com/d/msgid/r-inla-discussion-group/68b0d1e1-de35-fc84-5614-8d7e45cc9da2%40gmail.com.

--

Julianne Meisner, BVM&S(Vet) MS

PhD student | Center for One Health Research

Department of Epidemiology | University of Washington

Travel photography

Relief veterinarian