s-news
[Top] [All Lists]

Re: all possible combinations of predictors

To: Julia Linke <jlinke@ucalgary.ca>
Subject: Re: all possible combinations of predictors
From: Frank E Harrell Jr <fharrell@virginia.edu>
Date: Thu, 31 Oct 2002 06:58:03 -0500
Cc: s-news@lists.biostat.wustl.edu
In-reply-to: <3DC0A3DB.3554CF2@ucalgary.ca>
Organization: University of Virginia
References: <3DC0A3DB.3554CF2@ucalgary.ca>
On Wed, 30 Oct 2002 20:30:35 -0700
Julia Linke <jlinke@ucalgary.ca> wrote:

> Dear List,
> I have been looking into various different regression model
> selection strategies, and have read Quinn and Keogh's (2002) advise to
> select the best model based on all possible combinations of the multiple
> 
> predictors, rather through procedures such as backward, or
> forward stepwise selection.  I generally use the AIC as a criteria to
> select the best model, but now I am looking for a command (such as
> stepAIC) that integrates model selection based on lowest AIC using all
> possible predictor combinations.
> 
> Would you be able to suggest a S-Plus code that achieves
> running all possible predictor combinations (potentially including
> interactions) for a multiple regression model ?
> 
> Any insight would be greatly appreciated.
> Regards,
> Julia
> 
> 

I haven't read the reference you cited but I hope they added that all of these 
stepwise procedures can be disasters in terms of overfitting, non-preservation 
of confidence interval coverage and type I error, biased estimates of 
regression coefficients, inflated R^2, inflated adjusted R^2, etc., etc.  See 
more at http://www.pitt.edu/~wpilib/statfaq.html and 
http://hesweb1.med.virginia.edu/biostat/rms .
-- 
Frank E Harrell Jr              Prof. of Biostatistics & Statistics
Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences
U. Virginia School of Medicine  http://hesweb1.med.virginia.edu/biostat

<Prev in Thread] Current Thread [Next in Thread>