[Rd] Parametric links for glm?
roger koenker
roger at ysidro.econ.uiuc.edu
Sun Jul 30 17:29:37 CEST 2006
At useR 2006 I mentioned that it would be nice to have a way to
specify binomial links
that involved free parameters and described some experience with a
Gosset link involving
a free degrees of freedom parameter, and a Tukey-lambda link with two
free parameters.
My implementation of this involved some rather kludgey modifications
of binomial,
make.link and glm that (essentially) added a "..." argument to
binomial permitting the
parameters to be passed down to the fitting routines. So, in the
Gosset case, for example,
had the following passage in make.link, involving a free degrees of
freedom parameter nu:
}, Gosset = {
linkfun <- function(mu) qt(mu,nu)
linkinv <- function(eta) {
thresh <- -qqt(.Machine$double.eps,nu)
eta <- pmin(thresh, pmax(eta, -thresh))
pt(eta,nu)
}
mu.eta <- function(eta) pmax(dt(eta,nu), .Machine$double.eps)
valideta <- function(eta) TRUE
},
Just prior to the useR meeting, there were some changes in R-devel to
binomial that allow one to pass an object of class "link-glm" -- if I
understand properly, this allows one to more easily
roll one's own links, but it leaves the question of passing
parameters on the table. That is,
presumably, I could pass a structure, like that above of class "link-
glm" and thereby avoid
the Cerberus of officially sanctioned links, but I still don't see
how to pass my parameters.
Can someone enlighten me? Or, would it still be necessary to add
"..." to the binomial
function?
url: www.econ.uiuc.edu/~roger Roger Koenker
email rkoenker at uiuc.edu Department of Economics
vox: 217-333-4558 University of Illinois
fax: 217-244-6678 Champaign, IL 61820
More information about the R-devel
mailing list