2 messages · Michael Lawrence, Brian Ripley
Hi all, I'm using nnet() for non-linear regression as in Ch8.10 of MASS. I understand that nnet() by default optimizes least squares. I'm looking to have it instead optimize such that the mean error is zero (so that it is unbiased). Any suggestions on how this might be achieved? Cheers, Mike
Mike Lawrence http://artsweb.uwaterloo.ca/~m4lawren "The road to wisdom? Well, it's plain and simple to express: Err and err and err again, but less and less and less." - Piet Hein
I'm using nnet() for non-linear regression as in Ch8.10 of MASS. I understand that nnet() by default optimizes least squares. I'm looking to have it instead optimize such that the mean error is zero (so that it is unbiased). Any suggestions on how this might be achieved?
What makes you think least-squares does not achieve that? At a guess you mean the average prediction error over the training data in a regression problem, and any non-linear regression with an intercept term achieves that. But 'it is unbiased' needs futher statements including what the model is that what 'it' is supposed to be estimating unbiasedly in that model. Now for a non-linear regression you should be using linout = TRUE (as in the reference you give), and then you do have a 'non-linear regression with an intercept term'.
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