[R] stepAIC {MASS}

[Frank E Harrell Jr]

Nathan Leon Pace, MD, MStat wrote:
> In a generalized linear model with k covariates, there are 2(kth power) - 1
> possible models (excluding interactions).
> 
> Awhile ago a posting to R-help suggested Model Selection and Multimodel
> Inference, 2nd ed, by Burnham and Anderson as a good source for
> understanding model selection. They recommend (page 71) computing AIC
> differences over all candidate models in the set of possible models.
> 
> After looking at the help page for stepAIC and checking MASS, I am not
> certain which AICs are estimated by stepAIC.
> 
> Comments will be appreciated.
> 
> Nathan

Nathan,

If we were analyzing a randomized experiment with 10 treatment groups 
would we report an unadjusted P-value from a t-test comparing treatments 
1,3,4,6 with treatments 2,5,7,8,9,10 where a stepwise process  used the 
response variable to group the treatments?  Would we even try to 
collapse the treatments into groups (i.e., drop some of the 9 dummy 
variables)?

Cheers,
Frank

-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University