[R] stepAIC

[Prof Brian Ripley]

On Tue, 12 Sep 2006, stephenc wrote:

> Hi
> 
>  
> 
> I hope this isn't off topics, but I have always found when I stepAIC() some
> glm I get an improvement in accuracy and kappa, but I have just done a case
> where I got a marginal deterioration.  Is this possible, or should I be
> going through my figures carefully to see if I have messed up?

It is certainly possible.

Optimizing AIC is only aimed at getting an improvement in one prediction 
measure, on average.

-- 
Brian D. Ripley,                  ripley at 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