predictions

[mbr...@cs.cas.cz]

I am puzzled by the inla behavior when computing predictions via adding NA observations in a stack for prediction.

I do a spatial (spde) model (with two spatial fields of different ranges) and want to compute predictions on a fine and large grid. So I make one stack for data and other for predictions. Then I find that the grid is too large for one run, so I split the grid into two parts, make two stacks for them and then call inla twice. First with the data stack + part1 prediction stack, then with the same data stack and part2 prediction stack. 

So, I do use the same data and two different set of positions (plus covariates) in the two runs, resulting in objects v.cast1 and v.cast2. There I would expect maybe some small discrepancies for the estimated parameters between v.cast1 and v.cast2.

To my surpise, the estimated parameters are quite different (see the attached file "two parts.txt"), leading to a clear cut in the picture that merges the summary.linear.predictor predictions and their variability from v.cast1 and v.cast2 ("yhat.png" and "sd.png").

Although, the particular model I use might be somewhat badly conditioned and might need some prior finetuning at least, I do not understand why adding two different no-data (zero additional information) pieces changes the estimates from the same data so profoundly. Is this something specific to the two-spatial-components models or is it more general? 

In any case, what would be a general and safe way to do the predictions from the same data in several steps when memory does not allow them to be done in one piece?

All the Best

 Marek 

    

[INLA help]

This can be due to numerical issues. Can you retry with a newest version and adding

library(INLA)

inla.setOption(inla.mode=“experimental”)

And let us know…

thx

H

--
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/c357266d-5d6d-4035-9ef8-bab5cec59b67n%40googlegroups.com.
<sd.png><two parts.txt><yhat.png>