At 08:54 AM 5/10/00 -0400, Ben Bolker wrote:>
>I''ve seen profile.nls; I was thinking about extending it to do
general
>likelihood surfaces, but it looks like all I need to is write a
"profile"
>method for generalized likelihoods (I don''t know whether I should
expend
>the effort to follow the format of ms() as defined in S(-PLUS?) or not),
>and "plot.profile" will work automatically.
>
Two small points:
ms() in S+ is not a statistical function per se, it is a minimization
function only. This is why there are so few methods for the objects
generated by it. If you could assume that the function being minimized is
the negative of a log-likelihood then it would make sense to have all the
statistical paraphernalia for it, but you can''t. profile is somewhere
on
the border of statistical and non-statistical. It makes sense for any
function, but without a statistical context to guide you it is hard to know
where it might be sensible to stop profiling. In the other places (nls,
glm, at the moment) your stopping point is tied to the drop in likelihood
as you move away from the MLE.
Having said that, I have often thought it would be very sensible to have a
dedicated function that just minimized negative log-likelihoods in a
reasonably general case. The extension I would like to see provided would
be a facility to express the log-likelihood in terms of "metameters",
each
of which might be a true parameter, or could separately be specified as a
linear function of covariates. (I''m not sure how feasible this would
be if
you do not have symbolic derivatives, but always start off optimistically!)
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !) To: r-help-request at
stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._