s-news
[Top] [All Lists]

Re: ancova model checking

To: Christoph Scherber <Christoph.Scherber@uni-jena.de>
Subject: Re: ancova model checking
From: Frank E Harrell Jr <feh3k@spamcop.net>
Date: Mon, 23 Feb 2004 08:23:06 -0500
Cc: s-news@lists.biostat.wustl.edu
In-reply-to: <4039F5EB.1000607@uni-jena.de>
Organization: Vanderbilt University
References: <4039F5EB.1000607@uni-jena.de>
On Mon, 23 Feb 2004 13:45:31 +0100
Christoph Scherber <Christoph.Scherber@uni-jena.de> wrote:

> Dear all,
> 
> I have several partially correlated explanatory variables that I want to
> 
> analyse using analysis of covariance. 
> 
> The problem now is that order in model specification matters. I would 
> therefore like to have an automated routine which permutates the 
> positions of the explanatory variables, and then compares all resulting 
> model versions using AIC.
> 
> What I´d like to do is something like:
> 
> explanatories_c(explanatory.1,explanatory.2,explanatory.3....explanator
> y.n) for (i in 1:n) explanatory[i]_sample((explanatories[i])
> model[i]_aov(response~explanatory[i])
> AIC(model[i+1],model[i])
> 
> this whole procedure would then be replicated until the full nr of 
> explanatory variables (say, 10) is tested in all possible positions 
> (e.g. 10!)
> 
> I would really appreciate any suggestions on this.
> 
> Best regards
> Christoph.

There are faster systematic ways to go about this, but any of these
procedures are likely to destroy the meaning of P-values, confidence
limits, and result in biased regression models with overly optimistic R^2
and betas.  If you are doing ANCOVA in order to test for a treatment
effect you are really in trouble.  Pre-specify the model.

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

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