Hi,
I'm trying to use Robust non-linear regression to fit dose response curves.
Maybe I didnt look good enough, but I dind't find robust methods for NON
linear regression implemented in R. A method that looked good to me but is
unfortunately not (yet) implemented in R is described in
http://www.graphpad.com/articles/RobustNonlinearRegression_files/frame.htm
<http://www.graphpad.com/articles/RobustNonlinearRegression_files/frame.htm>
in short: instead of using the premise that the residuals are gaussian they
propose a Lorentzian distribution,
in stead of minimizing the squared residus SUM (Y-Yhat)^2, the objective
function is now
SUM log(1+(Y-Yhat)^2/ RobustSD)
where RobustSD is the 68th percentile of the absolute value of the residues
my question is: is there a smart and elegant way to change to objective
function from squared Distance to log(1+D^2/Rsd^2) ?
or alternatively to write this as a weighted non-linear regression where the
weights are recalculated during the iterations
in nlme it is possible to specify weights, possibly that is the way to do
it, but I didn't manage to get it working
the weights should then be something like:
SUM (log(1+(resid(.)/quantile(all_residuals,0.68))^2)) / SUM (resid(.))
the test data I use :
x<-seq(-5,-2,length=50)
x<-rep(x,4)
y<-SSfpl(x,0,100,-3.5,1)
y<-y+rnorm(length(y),sd=5)
y[sample(1:length(y),floor(length(y)/50))]<-200 # add 2% outliers at 200
thanks a lot
Hans Vermeiren
[[alternative HTML version deleted]]
______________________________________________
[email protected] mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
