[R] Negative Binomial Model

[Mike Ryckman]

Hello,

Thank you for your replies. I think that information largely focuses on
robust
standard errors... maybe I wasn't clear. By default Stata parameterizes the
dispersion. From the stata help files it is calculated as 1 + alpha*exp(xb +
offset).

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.

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:Ted.Harding at manchester.ac.uk] 
Sent: Wednesday, May 14, 2008 2:38 AM
To: r-help at 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) <Ted.Harding at manchester.ac.uk>
Fax-to-email: +44 (0)870 094 0861
Date: 14-May-08                                       Time: 10:38:24
------------------------------ XFMail ------------------------------