Skip to content

robust model selection criteria

3 messages · Carsten.Colombier@efv.admin.ch, Bert Gunter, Brian Ripley

Carsten.Colombier@efv.admin.ch
# 
Dear R-help-team,

do you know if there is a package for R available that contains a function,
which calculates a robust model selection criterium like robust AIC and has
a robust selection function like "step" for lm-objects, for an  rlm-object.
Unfortunately, functions like "step" or "stepAIC" cannot be applied to
rlm-objects. Moreover, these functions do not use  robust AIC.

Thanks for your help!

With best regards,
Carsten Colombier

Dr. Carsten Colombier
Economist
Group of Economic Advisers
Swiss Federal Finance Administration
Bundesgasse 3
CH-3003 Bern

phone +41 31 322 63 32
fax +41 31 323 08 33
email: carsten.colombier at efv.admin.ch
www.efv.admin.ch
Bert Gunter
# 
-- Bert Gunter
Genentech Non-Clinical Statistics
South San Francisco, CA
 
"The business of the statistician is to catalyze the scientific learning
process."  - George E. P. Box
??? How could this be meaningful? The robust "likelihood" need not increase
as more parameters are added because of the robust reweighting (points would
be downweighted differently in the different models). How do you account for
the number of "parameters" in a robust model given that it is in essence
nonlinear?

(This comment subject to correction/expansion by wiser heads than me)

-- Bert Gunter
Genentech Non-Clinical Statistics
South San Francisco, CA
 
"The business of the statistician is to catalyze the scientific learning
process."  - George E. P. Box
Brian Ripley
#
On Fri, 29 Apr 2005, Berton Gunter wrote:

            

      
      
      
    

        
More fundamentally, `AIC' is about maximum-likelihood fitting of true 
models.  Now rlm does usually correspond to ML fitting of a non-normal 
linear model, so it would be possible to compute a likelihood and hence 
AIC.  The point however is that the model is assumed to be false.  There 
are AIC-like criteria for that situation, but they are essentially 
impossible to compute accurately as they depend on fine details of the 
unknown true error distribution (and still assume a linear model).