Filed under: Spatial Models
Poisson GLMM on large spatial grid
It has for a long time been possible to fit GLMMs (Generalized Linear Mixed Models) in ADMB-RE. A typical example is correlated count data with Poisson distribution. However, when the observation are located on a spatial grid the number of latent variables (random effects in the ADMB-RE terminology) grows quadratically in the number of grid points in each geographical direction. The large number of random effects causes a computational challenge.
###Example:
tpl-file and data files needed for running the example. gmrf10.dat contains a 10x10 grid, and should run fast. gmrf100.dat contains a 100x100 grid, and will require >8Gb of memory. In addition you should use the command line
./gmrf -est -ilmn 5 -ind gmrf100.dat -shess -ndi 10000000 -noinit -gbs 100000000
Note: Current research on the ADMB source code has shown that relatively small modifications will make it feasible to fit a 200x200 grid on a 1.5 Gb machine. These improvements will find their way into the official ADMB source soon.
###Files: