[R-sig-ME] calculation of AIC

[Ben Bolker]

  Well, my two cents is (are?) that the AICs should definitely be
presented via round(AIC,0) or round(AIC,1) and not with signif()
as they are.  Deviance/AIC differences are really only important
on an absolute scale (agreeing with Murray Jorgensen).  Thus
I would consider the use of signif() rather than round() in
print.glm() a "misfeature" ...  (Does anyone know of a context
where *relative* differences in AIC or deviance are important?)

  Ben Bolker

Jeremiah Rounds wrote:
>> Date: Sun, 1 Feb 2009 11:41:44 -0800> From: adik at ilovebacon.org>
>> To: orzack at freshpond.org> CC: r-sig-mixed-models at r-project.org>
>> Subject: Re: [R-sig-ME] calculation of AIC> > > On Sun, 1 Feb 2009,
>> orzack wrote:> > > Speaking of this, does anybody know how to
>> change the default rounding for > > glm (and lmer) OR for an R
>> session in general (e.g., so that a regular call > > to glm would
>> generate AIC values with more digits)?> > > 1/7> [1] 0.1428571> >
>> options(digits=22)> > 1/7> [1] 0.1428571428571428> >
>> options(digits=23)> Error in options(digits = 23) :> invalid
>> 'digits' parameter, allowed 1...22> > options(digits=22)> > ...but
>> this of course won't help if there is explicit rounding programmed>
>> into glm/lmer. I also do not understand what would motivate this
>> code,> instead of a more straightforward round(aic,0).> > --Adam
> 
> First, glm apparently is not using rounding from "round".  glm is
> using signif or equivalent logic.  There is a difference. The
> difference is significant digits is a fairly precisely defined
> notion.  It is the number of digits you keep on the front of the
> power of 10 in "a X 10^b".  Round is much more cosmetic from what I
> can tell.
> 
> Second,  round(aic,0) is not more straightforward in the presence of
> very small AIC.  Here the distinct difference from round(a,0) and
> signif(a,4) is that you never know prior to viewing the aic how many
> digits round(a,0) will be keeping from the original unrounded AIC
> value.  With signif(a,4) you always know there will be a 4 digit
> number times 10 to some power.
> 
> Third, in my limited statistical experience (I am a master's student)
> AIC is not a method where those extra digits ever really matter.  I
> did a project on AIC/BIC model selection.  IMO in order to use AIC
> properly you have to consider the models in a nearby neighborhood to
> the best AIC as just as good as the model with the best AIC and
> consider sensitivity in your estimates.    There is no real context
> where you can properly say "the fifth significant digit of AIC has
> decided that model A is better than model B, and so I discard model
> B."
> 
> That is what I think, Jeremiah
> 
>>> _______________________________________________>
>>> R-sig-mixed-models at r-project.org mailing list>
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> _________________________________________________________________
> 
> 
> [[alternative HTML version deleted]]
> 
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


-- 
Ben Bolker
Associate professor, Biology Dep't, Univ. of Florida
bolker at ufl.edu / www.zoology.ufl.edu/bolker
GPG key: www.zoology.ufl.edu/bolker/benbolker-publickey.asc