R help - Jul 2003 - robust regression

Hi,

trying to do a robudt regression of a two-way linear model, I keep 
getting the following error:

 > lqs(obs ~ y + s -1,method="lms",
contrasts=list(s=("contr.sum")))
Error: lqs failed: all the samples were singular

Robust regression with M-estimators works (also regular least square 
fits, of course):
rlm.formula(formula = obs ~ y + s - 1, method = "M", contrasts =
list(s
= ("contr.sum")))

I tried an exact sampling (psamp="exact"), but I keep getting syntax 
errors. Any idea how I can make the first one work?

Thanks,

joerg

-- 
----------------------------------------------------------
J?rg Schaber
Instituto Cavanilles de Biodiversidad y Biologia Evolutiva
Universidad de Valencia               Tel.: ++34 96 354 3666
A.C. 22085                            Fax.: ++34 96 354 3670
46071 Valencia, Espa?a                email : jos at uv.es

On Wed, 30 Jul 2003, Joerg Schaber wrote:
> Hi,
> 
> trying to do a robudt regression of a two-way linear model, I keep 
> getting the following error:
> 
>  > lqs(obs ~ y + s -1,method="lms",
contrasts=list(s=("contr.sum")))
> Error: lqs failed: all the samples were singular
> 
> Robust regression with M-estimators works (also regular least square 
> fits, of course):
> rlm.formula(formula = obs ~ y + s - 1, method = "M", contrasts =
list(s
> = ("contr.sum")))
> 
> I tried an exact sampling (psamp="exact"), but I keep getting
syntax
> errors. Any idea how I can make the first one work?
You may well not be able to.  lms relies on being able to find a good fit 
to (just over) 50% of the data.  Sounds like you can't fit to 50%.
Try increasing `quantile'.

Generally resistant methods should be employed with caution for designed 
experiments (with factor covariates), and your specification of contrasts 
suggest y is a factor.  So this may not make statistical sense either.


-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

Hi,

the package is called "quantreg" and yes, it solved my problem. LAD 
regression was actually the method I was looking for, thanks a lot.
However, some of my problems are rather large and even if I use the 
method 'fn' and 'pfn' recommended for large problems in
'rq' I get an
error:

 > res <- rq(o ~ y + s 
-1,tau=0.5,method="fn",contrasts=list(s=("contr.sum")))
Error: cannot allocate vector of size 1950212 Kb

Is there are way to avoid that? The least square procedures are able to 
handle these large problems.

best wishes,

joerg

Patrick Burns wrote:
> Two possible alternatives are:
>
> 1) Use least absolute deviation regression, which you can
> get from Roger Koenker's package that I think is called
> "quantile".
>
> 2) The LAD regression is essentially equivalent to median
> polish (see, for instance, "Understanding Robust and Exploratory
> Data Analysis" by Hoaglin, Mosteller and Tukey).  To gain some
> more efficiency for nearly Gaussian data, you could replace the
> median by a location estimator that is more efficient, such as a
> trimmed mean.
>
> Good luck,
>
> Patrick Burns
>
> Burns Statistics
> patrick at burns-stat.com
> +44 (0)20 8525 0696
> http://www.burns-stat.com
> (home of S Poetry and "A Guide for the Unwilling S User")
>
> Joerg Schaber wrote:
>
>> Hi,
>>
>> trying to do a robust regression of a two-way linear model, I keep 
>> getting the following error:
>>
>> > lqs(obs ~ y + s -1,method="lms",
contrasts=list(s=("contr.sum")))
>> Error: lqs failed: all the samples were singular
>>
>> Robust regression with M-estimators works (also regular least square 
>> fits, of course):
>> rlm.formula(formula = obs ~ y + s - 1, method = "M",
contrasts =
>> list(s = ("contr.sum")))
>>
>> I tried an exact sampling (psamp="exact"), but I keep getting
syntax
>> errors. Any idea how I can make the first one work?
>>
>> Thanks,
>>
>> joerg
>>
>
>
>
-- 
----------------------------------------------------------
J?rg Schaber
Instituto Cavanilles de Biodiversidad y Biologia Evolutiva
Universidad de Valencia               Tel.: ++34 96 354 3666
A.C. 22085                            Fax.: ++34 96 354 3670
46071 Valencia, Espa?a                email : jos at uv.es