[R] algorithmic method quantile regression

[Cade, Brian]

>From ?rq.fit.pfn I see:

Details:

     Preprocessing algorithm to reduce the effective sample size for QR
     problems with (plausibly) iid samples.  The preprocessing relies
     on subsampling of the original data, so situations in which the
     observations are not plausibly iid, are likely to cause problems.
     The tolerance eps may be relaxed somewhat.

And from 1 in the Quantreg FAQ that Roger indicated there is:

>From ?rq.fit.fn:

   eps: tolerance parameter for convergence.  In cases of multiple
          optimal solutions there may be some discrepancy between
          solutions produced by method '"fn"' and method '"br"'.  This
          is due to the fact that '"fn"' tends to converge to a point
          near the centroid of the solution set, while '"br"' stops at
          a vertex of the set.



So it sounds like the pfn version of quantile regression estimates might
differ because it is intended for independent identically distributed data
(think homogeneous), it involves subsampling of the data set, and
convergence criteria are slightly different for fn and pfn than for the br
(standard Barrodale and Roberts simplex linear program) algorithm.  All
conditions that could lead to different estimates.  My recommendation for
the sample sizes you are considering is to stick with the Barrodale and
Roberts algorithm as it is the best understood, most reliable procedure.

Brian


Brian S. Cade, PhD

U. S. Geological Survey
Fort Collins Science Center
2150 Centre Ave., Bldg. C
Fort Collins, CO  80526-8818

email:  cadeb at usgs.gov <brian_cade at usgs.gov>
tel:  970 226-9326


On Wed, Oct 14, 2015 at 3:03 PM, T.Riedle <tr206 at kent.ac.uk> wrote:

> The fn and br methods return the same results but the results provided by
> pfn differ. I do not find an explanation for this observation in the papers
> on quantile regression. Therefore my question.
>
> -----Original Message-----
> From: Roger Koenker [mailto:rkoenker at illinois.edu]
> Sent: 14 October 2015 22:33
> To: T.Riedle
> Cc: r-help at r-project.org
> Subject: Re: [R] algorithmic method quantile regression
>
> Did you read item 1 in the quantreg FAQ()?
>
>
> url:    www.econ.uiuc.edu/~roger            Roger Koenker
> email    rkoenker at uiuc.edu            Department of Economics
> vox:     217-333-4558                University of Illinois
> fax:       217-244-6678                Urbana, IL 61801
>
> > On Oct 14, 2015, at 2:56 PM, T.Riedle <tr206 at kent.ac.uk> wrote:
> >
> > Greetings R Community,
> > I am trying to run a quantile regression using the quantreg package. My
> regression includes 7 independent variables with approx. 800 daily
> observations each. Thus, I think that the Barrodale and Roberts algorithm
> should do the trick. However, the Frisch-Newton after preprocessing returns
> different results and more significant coefficients than the br method.
> Which algorithmic method should I use now? Do the results mean that the
> Frisch-Newton after preprocessing dominates the br method?
> >
> >       [[alternative HTML version deleted]]
> >
> > ______________________________________________
> > R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> > https://stat.ethz.ch/mailman/listinfo/r-help
> > PLEASE do read the posting guide
> > http://www.R-project.org/posting-guide.html
> > and provide commented, minimal, self-contained, reproducible code.
>
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

	[[alternative HTML version deleted]]