Skip to content

[RsR] Non-linear robust method

6 messages · A@Teyteiboym m@iii@g oii ise@@c@uk, Stromberg, Arnold, Bruno L. Giordano +3 more

Original
A@Teyteiboym m@iii@g oii ise@@c@uk
# 
Hello,

I am based at STICERD at the London School of Economics and have been
using R for about a year. At the moment I am working on programming an R
algorithm for fitting a Pareto distribution robustly (this is a part of
package to supplement a research paper on income distribution). It isn't
going too well. I am aware that there are routines in R to fit GLM using
robust methods, but I am not sure whether any work has been done in the
direction of non-linear parametric distributions (such as gamma or
Pareto). 

I hope you hear from any of you soon and sorry for the trouble, 

Alex

PS Some of you may be aware that there is some excellent work by A
Marazzi, who programmed many such routines in FORTRAN about 15 years
ago, which run on older R interfaces.


Please access the attached hyperlink for an important electronic communications disclaimer: http://www.lse.ac.uk/collections/secretariat/legal/disclaimer.htm
Stromberg, Arnold
# 
I did work on robust nonlinear methods in the early nineties. To the best of my knowledge, none of my algorithms have been redone (and hopefully upgraded) in R. I may put a Ph.D. student on it, but that'll take a few years. In anyone knows of robust nonlinear code in R, let us know.

Thanks,
Arny


Arnold J. Stromberg, Ph.D.
Professor and Chair, Department of Statistics
817 Patterson Office Tower
Lexington, KY 40506-0027
ph: 859-257-8859
fax: 859-323-1973

-----Original Message-----
From: r-sig-robust-bounces at r-project.org [mailto:r-sig-robust-bounces at r-project.org] On Behalf Of A.Teytelboym at lse.ac.uk
Sent: Friday, August 24, 2007 7:45 AM
To: r-sig-robust at r-project.org
Subject: [RsR] Non-linear robust method

Hello,

I am based at STICERD at the London School of Economics and have been
using R for about a year. At the moment I am working on programming an R
algorithm for fitting a Pareto distribution robustly (this is a part of
package to supplement a research paper on income distribution). It isn't
going too well. I am aware that there are routines in R to fit GLM using
robust methods, but I am not sure whether any work has been done in the
direction of non-linear parametric distributions (such as gamma or
Pareto).

I hope you hear from any of you soon and sorry for the trouble,

Alex

PS Some of you may be aware that there is some excellent work by A
Marazzi, who programmed many such routines in FORTRAN about 15 years
ago, which run on older R interfaces.


Please access the attached hyperlink for an important electr...{{dropped}}
Bruno L. Giordano
# 
Hello,

as a side note, the Matlab function nlinfit (Statistics toolbox) for 
nonlinear fitting has a robust option.

It shouldn't be incredibly hard to translate it in R code.

However, I have to say that the routine itself does not perform 
incredibly well in case of outliers.

	Bruno


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Bruno L. Giordano, Ph.D.
Music Perception and Cognition Laboratory
CIRMMT http://www.cirmmt.mcgill.ca/
Schulich School of Music, McGill University
555 Sherbrooke Street West
Montr?al, QC H3A 1E3
Canada
Office: +1 514 398 4535 ext. 00900
http://www.music.mcgill.ca/~bruno/
Stromberg, Arnold wrote:
Martin Maechler
# 
BLG> Hello,
    BLG> as a side note, the Matlab function nlinfit (Statistics toolbox) for 
    BLG> nonlinear fitting has a robust option.

    BLG> It shouldn't be incredibly hard to translate it in R code.

    BLG> However, I have to say that the routine itself does not perform 
    BLG> incredibly well in case of outliers.

But the  robustbase package *has* had robust nonlinear
regression, almost since its beginning,
 nlrob() !

Probably because Alex used a somewhat misleading subject line in
his posting, I think you (Arnold and Bruno) have both been answering
the wrong question.

If I understand correctly, Alex was rather looking for help on
doing robust modelling for a specified non-*normal* distribution 
``for the good data'', whereas most available robust functions
assume that the "good data" is normally distributed and then
there's a fraction of "arbitrarily distributed" data points
(sometimes called "outliers" ...).

IIUC, Alex wants the "good data" to be Pareto ...
and he mentioned Marazzi's code and papers which did this for
the Gamma (and 'Weibull', BTW).

Martin Maechler,
ETH ZUrich
Eva Cantoni
# 
Quoting Martin Maechler <maechler at stat.math.ethz.ch>:
It is maybe worth looking at what kind of code has been used in:

Victoria-Feser, M.-P. and E. Ronchetti (1994). "Robust methods for  
personal income distribution models". The Canadian Journal of  
Statistics 22, 247--258.

Certainly not R, but maybe Splus (I have no idea, and I haven't checked).

Best,
Eva Cantoni
Ruckstuhl Andreas (rkst)
# 
Some additional remarks to Martin's and Eva's answer:
See Alfio Marazzi's page on Software: 
	 http://www.iumsp.ch/Unites/us/Alfio/msp_programmes.htm

See also the Maria-pia Victora-Feser's page
	http://www.hec.unige.ch/www/index.php?&pid=189

and there the papers 
	Dupuis, D. and M.-P. Victoria-Feser (2007). "A Robust Prediction Error Criterion for Pareto Modeling of Upper Tails". The Canadian Journal of Statistics. To appear. (proof's version)

	Maria-Pia Victoria-Feser, Elvezio Ronchetti (1997). "Robust Estimation for Grouped Data"
	  Journal of the American Statistical Association, Vol. 92, No. 437 (Mar., 1997), pp. 333-340


All the best
Andreas Ruckstuhl