[RsR] robustbase: precision of nlrob with plinear algorithm.

[Jerry Lewis]

Thanks for your response.  Would a better workaround be to use the plinear estimates as starting values for the full Gauss-Newton?

Jerry

-----Original Message-----
From: Andreas Ruckstuhl [mailto:rkst using zhaw.ch] 
Sent: Monday, February 01, 2016 3:20 PM
To: Jerry Lewis
Cc: Martin Maechler; r-sig-robust
Subject: Re: robustbase: precision of nlrob with plinear algorithm.

Dear Jerry

As in the help file to nlrob is noted the conditionallinearity approach fails for robust fitting methods. Unfortunately not anymore by an error message.

If you checkthe output more carefully, you will note that the result of nlrob(... , algorithm="plinear" ) is not reliable:
DNrfit1 <- nlrob( density ~ Asym/(1 + exp(( xmid - log(conc) )/scal ) ),
                  data=DNase1, start=list(Asym=3,xmid=0,scal=1) )
DNrfit2 <- nlrob( density ~   1 /(1 + exp(( xmid - log(conc) )/scal ) ),
                  data=DNase1, start=list(xmid=0,scal=1), algorithm="plinear" )

summary(DNrfit1)
## Residuals:
##        Min         1Q     Median         3Q Max
## -0.0322811 -0.0130976 -0.0008932  0.0095784 0.0404174

summary(DNrfit2)
## Residuals:
##      Min       1Q   Median       3Q      Max
## 0.003927 0.107372 0.280145 0.641955 1.002490

Here all residuals are positive!

I have not (jet) understood what goes wrong. Hence,we should disable this argument (for the moment).

Thanks for your hint.

All the best
Andreas



Am 28.01.2016 um 06:37 schrieb Jerry Lewis:
>
> In nls, plinear simplifies the search for the LS estimate by 
> simplifying the optimization through dimension reduction.  Output 
> (including standard errors are completely equivalent, provided that 
> that the higher dimensional ordinary Gauss-Newton optimization is able 
> to adequately find the least squares estimates.
>
> The first example for nlrob suggests that this is only partially true 
> for nlrob.  To the displayed precision almost everything (robustness 
> weights, robust residual SE and parameter estimates) are identical 
> between plinear and ordinary Gauss-Newton, but the standard errors of 
> parameter estimates from plinear are 2.769 times larger for each 
> parameter.  Is this a bug in the code, or is there a theoretical 
> reason why this should be so?
>
> Thanks,
>
> Jerry W. Lewis, PhD
>
> Principal Biostatistician
>
> Biogen
>
> 225 Binney St
>
> Cambridge, MA 02142
>
> library(robustbase)
>
> DNase1 <- DNase[ DNase$Run == 1, ]
>
> with(DNase1,plot(conc,density,log="x"))
>
> # plinear and Gauss-Newton get the same robustness weights, robust 
> residual SE, and parameter estimates,
>
> # why then are the standard errors different?
>
> # Given that, why are the plinear standard errors larger when when the 
> comment in the example
>
> summary( nlrob( density ~ Asym/(1 + exp(( xmid - log(conc) )/scal ) ),
>
>          data=DNase1, start=list(Asym=3,xmid=0,scal=1) ) )
>
> summary( nlrob( density ~   1 /(1 + exp(( xmid - log(conc) )/scal ) ),
>
>          data=DNase1, start=list(xmid=0,scal=1), algorithm="plinear" ) 
> )
>

--
----------------------------------------------------------------------

Prof. Dr. Andreas Ruckstuhl
Schwerpunktleiter Statistische Datenanalyse IDP Institut für Datenanalyse und Prozessdesign ZHAW Zürcher Hochschule für Angewandte Wissenschaften
Rosenstrasse 3                    Tel.  : +41 (0)58 934 78 12
Postfach                          Fax   : +41 (0)58 935 78 12
CH-8401 Winterthur                e-Mail:Andreas.Ruckstuhl using zhaw.ch
                                   WWW   :http://www.idp.zhaw.ch