On Tue, 29 Jul 2003, Volker Bahn wrote:
> Dear list members,
>
> I posted the question below a few weeks ago on this list and didn't receive
> any answers (but an email from one other person showing interest in an
> answer). I was wondering whether this was because of summertime or my poor
> phrasing of the question or did I really land somewhere outside the area of
> expertise of this enlightened and helpful group of people? If the latter was
> true, could you maybe suggest colleagues who might be able to help?
>
> Thank you.
>
> Volker Bahn
>
> Old question:
>
> I'm running spatial linear models (slm) of the CAR family in Splus 6.1 on a
> 7.2 RedHat Linux machine (but the same effect described below is true for my
> Splus 6.1 Windows Version running on a XP machine).
> I want to compare the slm model to a similar lm model by AIC and use the
> log-likelihood to calculate the AIC. However, the loglik given for the slm
> model by Splus does not seem to be correct to me. Two observations lead me
> to this conclusion:
>
> 1) If I calculate loglik myself based on the residual standard error (RSE)
> with a formula from Burnham and Anderson (2002) I get a value completely
> different from Splus:
>
> Loglik = (-n / 2) * LN(RSE^2 * df / n) - (n / 2) * LN(2*PI()) - (n / 2)
>
> With
>
> RSE = residual standard error (in output)
>
> df = degrees of freedom on the RSE
>
> n = sample size
That's not the right formula for a spatial linear model: where in B&A
(2002) are such models discussed (I could not find them)?
Log-likelihoods and hence AICs are only defined up to an additive constant
(which depends on the dominating measure used). Are you sure you are
using comparable ones?
[...]
--
Brian D. Ripley, ripley@stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595
|