Good point. If they're everywhere,
they're not outliers. The rhetoric
was way out of control there. Let's
try again. I want to make a distinction
like this.
1. "well-behaved" data + a "few" outliers
(n ~ 1) => it is sensible to use robust
regression as a check on standard.
2. long-tailed data => work on a different
scale.
Nick
[email protected]
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]]On Behalf Of Finne H�kon
> Sent: 30 January 2004 14:15
> To: '[email protected]'
> Subject: st: RE: RE: qreg versus rreg
>
>
> In the context of choosing an appropriate transformation or
> link function:
> in what dimension(s) can you have outliers "everywhere"?
> (many clustered
> ones, many dispersed ones, in both ends of a distribution, in repeated
> observations, on multiple variables ...)
>
> -- H�kon
> [email protected]
>
> -----Original Message-----
> From: Nick Cox [mailto:[email protected]]
> Sent: 30. januar 2004 14:58
> To: [email protected]
> Subject: st: RE: qreg versus rreg
>
>
> This raises the old classical trope,
> beaten almost to death by the late Sir Isaiah Berlin
> in many of his essays on intellectual history,
> that the fox knows many things, but the hedgehog
> knows one big thing.
>
> When attacked, the hedgehog has just one means
> of defence, although it is usually effective. -qreg-
> is a hedgehog. The fox has many different tricks. -rreg-
> is a fox. Its mixed strategy is an attempt to be smart
> in different ways.
>
> My experience loosely matches Richard's, certainly
> in terms of wanting to think that -qreg- is as good
> because of the much greater ease in explaining it.
> At the same time, if you have outliers everywhere,
> you are possibly working on inappropriate scales
> and should wonder about reaching for a transformation
> or, in some frameworks, a different link function.
>
> Nick
> [email protected]
>
> > -----Original Message-----
> > From: [email protected]
> > [mailto:[email protected]]On Behalf Of Richard
> > Williams
> > Sent: 30 January 2004 13:20
> > To: [email protected]
> > Subject: st: qreg versus rreg
> >
> >
> > This came up several months ago on the list but I am still
> > confused: As a
> > means for dealing with outliers, what are the relative merits
> > of -rreg- and
> > -qreg-? When should one be preferred over the other?
> >
> > As I understand it, -rreg- goes through this complicated
> > weighting scheme,
> > which causes outliers to be weighted less heavily. -qreg-
> > (by default)
> > does median regression, and the median is less affected by
> > outliers than
> > the mean is.
> >
> > In terms of giving a quick 30 second intuitive explanation, I like
> > -qreg-. On the other hand, in the few examples I've tried
> > myself or seen
> > elsewhere, the results from -rreg- seemed more plausible. On
> > the other
> > other hand, in those examples -rreg- basically just dropped
> > the extreme
> > outliers, and I could do that myself without a fancy program.
> >
> > This is one of the problems with using Stata in a stats
> > class. When I only
> > used SPSS, these issues never came up, because as far as I
> > can tell SPSS
> > can't do anything like this!
>
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
