[Rd] weights in glm (PR#8720)

Ben Bolker bolker at ufl.edu
Wed Mar 29 01:51:58 CEST 2006


Peter Dalgaard <p.dalgaard <at> biostat.ku.dk> writes:

> 
> I'm not even sure we want to fix this up. I recall some nasty issues
> with DF that have no proper solution that way - an observation with a
> tiny weight represents an observation with a large variance and
> contributes 1DF to the residual, with weight zero it is not supposed
> to contribute at all, so there's a discontinuity for weights
> approaching zero.
> 

  with respect: is it worth adding a note to the documentation
and/or a warning to the code?  I understand that it
becomes unwieldy to warn/protect against all
"stupid"/unforeseen/suboptimal uses of the software, but I also 
understand how the original poster could have decided that setting
a weight to zero was a plausible way to ignore cases ... there is 
already a warning that "observations with zero weight not used for
calculating dispersion" -- could one modify this
slightly to warn people against using zero weights?  (Are zero
weights always a bad idea?)

   Ben Bolker



More information about the R-devel mailing list