R devel - Apr 2009 - summary for negative binomial GLMs (PR#13640)

Full_Name: Robert Kushler
Version: 2.7.2
OS: Windows XP
Submission from: (NULL) (69.246.102.98)


I believe that the negative binomial family (from MASS) should be added to the
list for which dispersion is set to 1.

<kushler <at> oakland.edu> writes:
> 
> Full_Name: Robert Kushler
> Version: 2.7.2
> OS: Windows XP
> Submission from: (NULL) (69.246.102.98)
> 
> I believe that the negative binomial family (from MASS) should be 
> added to the
> list for which dispersion is set to 1.
  Could you please clarify? In what procedures, under what
circumstances?  Sounds like you mean l. 573 of glm.R:

if(object$family$family %in% c("poisson", "binomial"))  1

  The use case here is using negative.binomial with a fixed 
theta parameter, right?  Using glm.nb takes care of this problem
(it produces an object of class "negbin": MASS:::summary.negbin
shows that the dispersion gets set to 1 here).

   I guess there's a little bit of a jurisdictional
argument here, since the negative.binomial family is
in MASS, and summary.glm is in base R ... also, there's
a bit of a challenge in figuring out the test, because
object$family$family is not a fixed string for negative
binomial-family objects (e.g. "Negative Binomial(0.4)"
in the example below) -- I'm not sure of the cleanest
way to detect this case.

   I think I agree with you, but it would help to present
your case in more detail ...

  Ben Bolker

===================Example:

x <- rep(seq(0,23,by=1),50)
s <- rep(seq(1,2,length=50*24),1)

tmp2 <- data.frame(y=rnbinom(length(s),
                    mu=8*(sin(2*pi*x/24)+2),size = 0.4),x=factor(x),s=s)

library(MASS)
tmp.glm.nb2 <- glm.nb(y~factor(x)-1 +offset(log(s)),data = tmp2)
summary(tmp.glm.nb2)
## summary.negbin takes care of this case

tmp.glm.nb3 <- glm(y~factor(x)-1 +offset(log(s)),data = tmp2,
                   family=negative.binomial(theta=0.4))

summary(tmp.glm.nb3)