R-alpha: GLM (from R-core)

Peter Dalgaard BSA p.dalgaard@kubism.ku.dk
05 Sep 1997 16:48:24 +0200

As some of you may know, there has been formed a "core team" of
developers, who can dig directly into the R source at Auckland and fix
bugs etc., in order to speed up development. There is a mailing list
for that group. This is necessary, because some of the bugs discussed
are not publicly available (and hopefully will never be!), but one
potential risk of the structure is that we forget about R-devel. One
recent case was this:

I discovered that summary.glm was calculating "t" statistics (est./SE)
correctly in the case were the dispersion parameter of the GLM is
fixed, but referring it to a t distribution with the residual DF. This
is wrong, badly so in the case of tabular data.

This is just a bug, but then Thomas Lumley added:

"We should also allow the df to be specified as an argument to
summary.glm(). The gaussian case is the only one where there is a real
honest t-distribution and in other cases the user should be free to use
other values of the df that might improve the approximation.  This is
particularly true when the user specifies the dispersion, as then we have
no idea what df is appropriate."

That kind of topic is about "policy" as opposed to bugfixing and
codewriting, and - as Martin Maechler pointed out - there are people
on R-devel who have broad experience in GLMs and may want to
contribute to the discussion. 

Also other things with GLMs are debatable, e.g. whether it is always
correct to use the weighted RSS and not the deviance for the
dispersion estimate.

(The fixed-dispersion bug is fixed in the current sources, using the
simple logic that if the dispersion in not estimated from the RSS, we
use the normal distribution (or "infinite df" if you like))

   O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard@biostat.ku.dk)             FAX: (+45) 35327907
r-devel 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-devel-request@stat.math.ethz.ch