Dear all,
I have been using rlm() for robust regression. Could someone please
suggest an appropriate measure of goodness-of-fit [1]? All I've found
after trawling the web, literature databases, and previous r-help posts,
is the "robust R^2" on pp. 362-363 of the S-plus manual, which is
available at
http://web.mit.edu/afs/athena/software/splus_v7.0/www/statman1.pdf
(7.57 MB)
It is based on M-estimators as detailed in Chapter 11 of statman1.pdf. I
think the s-plus method lmRobMM() returns this "robust R-squared", but
I
would like to stay within R. Implementing the abovementioned formulas in R
was not straightforward because the definition and notation of the loss
functions (psi,rho,Tukey's bisquare) differ between statman1.pdf and
http://spider.stat.umn.edu/R/library/MASS/html/rlm.html
I would be most grateful for any references or tips.
[1] I'm picturing my data as possibly contaminated by some other
phenomenon than the one I'm interested in. What I think I want is a robust
estimator of the R-squared which would be obtained in the absence of such
contamination.
Sincerely,
Jon Olav Vik
