On Wed, 14 May 2008, Mike Ryckman wrote:
Hello,
Thank you for your replies. I think that information largely focuses on
robust standard errors... maybe I wasn't clear.
No, you were not and, as pointed out previously (and described in the
posting guide), a reproducible example would clearly help.
By default Stata parameterizes the dispersion. From the stata help files
it is calculated as 1 + alpha*exp(xb + offset).
If I interpret this correctly, this is would be a vector of length n, not
a scaler dispersion.
As I understand it, this process should happen outside of a decision to use
robust errors. I think the normal glm.nb function in MASS uses some kind of
constant in estimating the error process in a manner more analogous to how
they would be estimated in any normal glm model.
There are different parametrizations of the negative binomial model but
afaik both have dispersion = 1 (in the GLM sense) but have a variance
function that can account for over-dispersion (compared to the Poisson
variance function).
However, if you want to use a different dispersion, you can plug it into
the summary method
summary(fm, dispersion = ...)
Z
I think the problem is that stata is using a different method to calculate
the standard errors - and I'm not sure what that method is or how to do
it in R.
The gamlss package produces the closest results - but even those are still
not
exactly the same.
Mike
-----Original Message-----
From: Ted Harding [mailto:[EMAIL PROTECTED]
Sent: Wednesday, May 14, 2008 2:38 AM
To: r-help@r-project.org
Cc: Mike Ryckman
Subject: Re: [R] Negative Binomial Model
On 14-May-08 09:11:31, Achim Zeileis wrote:
On Wed, 14 May 2008, Mike Ryckman wrote:
Hello,
I am trying to run a negative binomial regression model in R and
can't get the standard errors to match the output I get from the
Stata nbreg command. I've tried a few different options but
haven't had much luck. The closest I've found is:
gamlss(formula, family = NBI, sigma.formula = ~ 1,data=dataframe)
...But this is still a little off most of the time and pretty far off
at
other
times (compared with the Stata output). The glm.nb from the MASS
package
produces the correct coefficients, but different (usually very
different)
standard errors.
Could anybody explain this and point me in the right direction? I'd
really
appreciate it.
Well, you don't really give us enough information to know (a
reproducible
R example and the desired standard errors from Stata would have been
helpful). My guess is that Stata uses some sort of "robust" standard
errors, i.e., sandwich standard errors. Try something like:
library("MASS")
library("sandwich")
library("lmtest")
fm <- glm.nb(...)
coeftest(fm, vcov = sandwich)
See also the following thread for some more discussion for count data
regression in R and Stata:
https://stat.ethz.ch/pipermail/r-help/2008-May/161640.html
Better links for following this thread are:
Start:
https://stat.ethz.ch/pipermail/r-help/2008-May/161591.html
Then click on "Next message:" for the second. Unfortunately,
at that point the thread broke (Paul next responded to the list
as a reply to an off-list message I sent him). So to continue,
next take Achim's URL above:
https://stat.ethz.ch/pipermail/r-help/2008-May/161640.html
and thereafter continue to click on "Next message:" until the
thread runs out.
That was a very helpful thread, for me!
Best wishes,
Ted.
--------------------------------------------------------------------
E-Mail: (Ted Harding) <[EMAIL PROTECTED]>
Fax-to-email: +44 (0)870 094 0861
Date: 14-May-08 Time: 10:38:24
------------------------------ XFMail ------------------------------
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
