[R] what does "rlm" do if it fails to converge within iteration limits?

Martin Maechler maechler at stat.math.ethz.ch
Tue Mar 13 18:25:08 CET 2012

>>>>> Michael  <comtech.usa at gmail.com>
>>>>>     on Mon, 12 Mar 2012 13:19:19 -0500 writes:

    > The problem is: by default shouldn't it use "Huber's"?
    > And it should be convex problem no?

    > so when I do rlm(y~x) which is a single-beta fitting problem,

    > shouldn't it always converge?

“In theory, theory and practice are the same. In practice, they are not.”
― Albert Einstein
  	 [according to http://www.goodreads.com/quotes/show/66864 ]

Theory says that convergence happens in an infinite number of
iterations, but then theory also says convergence means that the
coefficients don't change any more ;-)

[... etc.]

    > --------------------

    > Psi functions are supplied for the Huber, Hampel and Tukey bisquare
    > proposals as psi.huber, psi.hampel and psi.bisquare. Huber's corresponds to
    > a convex optimization problem and gives a unique solution (up to
    > collinearity). The other two will have multiple local minima, and a good
    > starting point is desirable.

which also mentions "up to collinearity" (theory).
"Practice" would add "near-collinearity" and many other
practical border line issues that can happen. 

As maintainer of the  robustbase  package,
I'm slightly intrigued by your example.
1) Can you provide it reproducible  dput(data)  "cut & paste"
   into your e-mail if small;  available as mydata.rda after
   save(., file="mydata.rda")  for download

2) What's the result of using  lmrob()  {package 'robustbase'}
   instead of  rlm()  {package 'MASS'} ?

Best regards,
Martin Maechler, ETH Zurich

    > On Fri, Mar 9, 2012 at 1:21 PM, Berend Hasselman <bhh at xs4all.nl> wrote:

    >> On 09-03-2012, at 20:00, Michael wrote:
    >> > Hi all,
    >> >
    >> > In using "rlm" I've got a bunch of warnings... "failed to converge in 20
    >> > steps", etc.
    >> >
    >> > My question is:
    >> >
    >> > what are the results then after the failure?
    >> >
    >> They haven't converged. So inaccurate. Maybe your model is badly
    >> formulated or ill conditioned.
    >> > Will "rlm" automatically downgrade back to "lm" upon failure?
    >> >
    >> Help says nothing about that so most likely no.
    >> Why don't you try and raise maxit? Use maxit=40 in the call of rlm. And
    >> see what happens.
    >> Berend

    > [[alternative HTML version deleted]]

    > ______________________________________________
    > R-help at r-project.org mailing list
    > 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.

More information about the R-help mailing list