[R-sig-Geo] geoR likfit: confusing AIC results?

[Timothy W. Hilton]

Dear P.J.,

Thanks very much for such a helpful (and fast!) response.  I need to
increase my understanding of the optimization issues surrounding the
distances and orders of magnitude of the parameters.  If you are able to
spare another few moments, your advice raised two questions for me.

> 1. your area is of the size 6000x6000
> This is related with distances and order of magnitude of the
> parameters, in particular to phi.
> Since you total variances are around 3 the maximization may need
> some tunung for dealing with different oreders or magnitude
> 
>    1.a a simple suggestion would be replace 6000 by 6 and see whether it
>        change the estimates

I reduced the size of my area from 6000x6000 to 6x6, and it did indeed
change the estimates -- now the pure nugget AIC is less than or equal to
the exponential AIC in six of ten cases.  In two of the four cases where
the exponential AIC was lower, the AIC values did not differ until the
10th or 11th decimal place.  I expect this is more in the range of
floating point error, making the AIC values practically equal.

One question related to this: the difference between the exponential
loglik and the pure nugget loglik is greater than 1e-14 in only two of
the ten cases.  Because the exponential model has three estimated
parameters (sigmasq, phi, tausq) to the pure nugget's one (tausq),
shouldn't the pure nugget AIC be lower in all of those cases where the
loglik values are essentially equal?

>    1.b an alternative would be to pass values of fnscale to the
> optimizer (optim) through the appropriated ... mechanism

I will look into this also.  Thanks!

> 2. Notice that summary() on a likfit object (or elements of the
> list) already give you the fit and likelihood values for the pure
> nugget (independence) model such that you do not need to run the
> maximizer over it

Ah, very helpful.  Is this in the "nospatial" element of the list
reported by likfit?  I wasn't sure what that represented, exactly.

Again, thank you for taking the time to respond to my question, and
thank your for providing geoR to the R community.

Cheers,
Tim

--

Timothy W. Hilton
PhD Candidate, Department of Meteorology
The Pennsylvania State University
503 Walker Building, University Park, PA   16802
hilton at meteo.psu.edu