[ADMB Users] Running models at different resolution

[H. Skaug]

Hi All,

I conjecture that 90% of all ADMB models would benefit from the inclusion
of some spatial aspect. However, if made spatial using the random
effects (RE) stuff, the same models would run so slow that they were useless.

If you know what the spatial correlation is, you can incorporate spatial effects
in non-RE models. (For instance, people are using random walks to
account for temporal trends in non-RE models). It is only for estimation of
the spatial correlation and some related variance parameters that you
need the RE stuff. So, if you scale down your model considerably (in a
way that will be case specific)
you may be able to estimate spatial correlation using a RE model.
This estimated spatial correlation can then be plugged into
the full scale non-RE model.

This two stage approach is slightly inconvenient in that you have to maintain
two different programs. You may be able to achieve this from within a
single .tpl file, however.

It would be interesting to hear if people think this is a useful approach.

Hans