[Rd] rstandard.glm() in base/R/lm.influence.R

[Roger Bivand]

I contacted John Fox about this first, because parts of the file are 
attributed to him. He says that he didn't write rstandard.glm(), and 
suggests asking r-devel.

As it stands, rstandard.glm() has summary(model)$dispersion outside the
sqrt(), while in rstandard.lm(), the sd is already sqrt()ed. This seems to
follow stdres()  in VR/MASS/R/stdres.R.

Of course for the c("poisson", "binomial")  families,
summary(model)$dispersion is 1, so sqrt doesn't make a difference, but
working with an epidemiologist, Marilia Carvalho, we wanted to map the
standardised residuals of a quasipoisson null model of cases offset by
log(pop.at.risk), and rstandard.glm() seemed to compress the values more
than we'd have expected. Her colleague: "Valeska (L.  Andreozzi) says that
she has already noticed that, and as McCullagh & Nelder (2nd ed, pg 397)
says, the dispersion is inside the sqrt". (That's the way I read 12.5 on 
p. 397 too). I can't see that summary(model)$dispersion has already been 
subject to sqrt() in sum(object$weights * object$residuals^2)/df.r in 
summary.glm().

Should:

    res / (summary(model)$dispersion * sqrt(1 - infl$hat))

be:

    res / (sqrt(summary(model)$dispersion * (1 - infl$hat)))

on line 117 in base/R/lm.influence.R?

Best wishes,

Roger

-- 
Roger Bivand
Econonic Geography Section, Department of Economics, Norwegian School of 
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen, 
Norway, voice: +47-55959355, fax: +47-55959393; Roger.Bivand at nhh.no