[R] pros and cons of "robust regression"? (i.e. rlm vs lm)

[Berton Gunter]

Spencer:

Your comment reinforces Andy's point, which is that purported outliers must
not be ignored but need to be clearly identified and examined. For reasons
that you well understand, robust regression methods are better for this in
the linear models context than standard least aquares. However, as I
understand it, the problem in the case that you describe is **not** that the
outliers weren't identified and examined, but that they were and were
dismissed as metrological anomalies. I would say this was an error of
scientific (mis)judgment, not data analystical methodology.

Anyway, this has taken us too far afield, so no more on-list comments from
me.

-- Bert 
 

> -----Original Message-----
> From: Spencer Graves [mailto:spencer.graves at pdf.com] 
> Sent: Thursday, April 06, 2006 10:30 AM
> To: Berton Gunter
> Cc: 'Liaw, Andy'; 'r user'; 'rhelp'
> Subject: Re: [R] pros and cons of "robust regression"? (i.e. 
> rlm vs lm)
> 
> 	  A great example of the hazards of automatic outlier 
> rejection is the 
> story of how the hole in the ozone layer in the southern 
> hemisphere was 
> discovered.  Outliers were dutifully entered into the data base but 
> discounted as probable metrology problems, which also plagued the 
> investigation.  As the percentage of outliers became excessive, 
> investigators untimately became convinced that many of the "outliers" 
> were not metrology problems but real physical problems.
> 
> 	  For a recent discussion of this, see Maureen Christie 
> (2004) "11. 
> Data Collection and the Ozone Hole:  Too much of a good thing?" 
> Proceedigns of the International Commission on History of Meteorology 
> 1.1, pp. 99-105 
> (www.meteohistory.org/2004proceedings1.1/pdfs/11christie.pdf).
> 
> 	  spencer graves
> p.s.  I understand that Australia now has one of the world's highest 
> rates of skin cancer, which has contributed to a major change 
> in outdoor 
> styles of dress there.