STAT 592

Lab 4

Markov Random Fields

Today's lab will deal with another built in geoR data set, SIC, which deals with Swiss precipitation. While R is installed on the virtual mahine (see below), geoR is not, so you need to use the R command install.package("geoR"). We will focus on the subset of 367 sites called sic.367. Have a look at it to start with. You may want to look at various transformations that make the data look close to Gaussian.

In order to fit an MRF we will first create a triangulation of the points. The software we will be using is called spatio_temporal_GMRF and is available on the Linux virtual machine stat592a-linux.stat.washington.edu .

The complete help file for the program is available here, or by the command spatio_temporal_GMRF -H. There is a short version by the command spatio_temporal_GMRF -h .

In order to fit a GMRF model, we will focus on the Laplace approximation method (INLA) developed by Håvard Rue and implemented for non-equispaced data by Johan Lindström. First make a directory on the Linux machine (assumed to be called lab/swiss in what follows). The needed files, which you want to copy to the directory, are here, or in this compressed tar file.

The Readme. file tells you how to go about running the program.
The read.data.r file has a possible way of displaying aspects of the output.

The task is to compare the MRF predictions at (some of) the reserved sites, sic.100, to an appropriate kriging at the same sites.


Here is the code that we are using (in compressed tar format). It requires Håvard Rue's GMRFLIB software library.