Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: RE: generalized Poisson and BePress Selected Works web site

From   [email protected]
To   [email protected]
Subject   st: RE: generalized Poisson and BePress Selected Works web site
Date   Tue, 13 Sep 2011 03:37:07 -0400 (EDT)

I have two notices to share with you. The first is regarding a new command on SSC, the second is
information on a technology which may be of interest to many of you.


I have had a number of requests for me to write, or locate, a Stata command for generalized Poisson regression. James Hardin wrote such a command for our book, Generalized Linear Models and Extensions, 2nd edition (Stata Press, 2007). However, it did not estimate models that were under-dispersed, and there was no help file. I created a generalized Poisson command several years ago, but did not have it published on the SSC with the other commands I had written. However, due to recent requests I decided to have it posted to SSC. Kit kindly had it posted, and I wish to advise you of its availability. I checked it over, assuring that it would accurately estimate both under and over dispersed models. It provides the full compliment of maximum likelihood options, allows for survey models, bootstrapping, a variety of standard errors, constrained estimates, offset and exposure, irr, and so forth. A help file is on the SSC site for the command as well. To install, type on the Stata command line:

                                  . ssc install gnpoisson

The value of generalized Poisson regression is its ability to model Poisson underdispersion, which is not possible for the traditional negative binomial model [-nbreg y xvars- or -glm y xvars, fam(nb ml)-]. One can use a hurdle model for underdispersed count data, as well as a generalized negative binomial (not Stata's generalized negative binomial, which I call a heterogeneous negative binomial in keeping with LIMDEP vocabulary). Generalized NB has a 3rd parameter to estimate, aside from the mean and heterogeneity or ancillary parameter. Generalized Poisson has an ancillary parameter like the negative binomial. If the value of the ancillary parameter is 0, or near 0, the model is Poisson, just like the traditional negative binomial model. Moreover, parameter estimates of the two models are typically quite close, unless the model is under-dispersed. Again, the value of the model is its ability to handle under-dispersion. It also does a better job than negative binomial when there are excessive 0's in the data, but in such a case it is preferable to use a zero-inflated Poisson or zero-inflated NB model anyhow. Note that like nbreg, convergence can be difficult when the ancillary parameter is 0, or very close to 0. Computers can differ in providing convergence and appropriate estimates, which are Poisson. When the model is Poisson, the estimates and SEs of gnpoisson, nbreg, and poisson are the same.

You can read about modeling with generalized Poisson regression in my book:

Hilbe, Joseph M (2011), Negative Binomial Regression, 2nd edition, Cambridge University Press
Hardin, JW and JM Hilbe (2007), Generalized Linear Models and Extensions, 2nd ed, Stata Press
          (3rd edition out before end of year)
Cameron, AC and PK Trivedi (1998), Regression Analysis of Count Data, Cambrige University Press Winkelmann, Rainer (2008), Econometric Analysis of Count Data, 5th ed, Springer

and in a number of journal articles and book chapters.


Some StataListers may not know about the Bepress "Selected Works" site. It is a web site which you may custom design and on which you can upload documents of various formats for others to download. For instance, I have put some of my old unpublished papers as well as some journal articles.

But most importantly from my point of view, I have placed data sets and functions/commands that are used in two of my books, "Logistic Regression Models" (2008, Chapman & Hall/CRC) and "Negative Binomial Regression, 2nd edition" (see above). I also posted an electronic ebook called "Negative Binomial Regression Extensions" that provides additional code used in the book, as well as new code that I have developed since the book was published, and older code that may be relevant to estimating various count models. I also posted the "Errata and Comments" page for both books on the site.

The advantage of this setup is that I can make updates anytime I want, and automatically notify those who have signed up that additions or amendments have been made to the site. Those signing up can cancel at any time. This way anyone who has obtained a copy of either of these books can get enhancements, new code, comments, or whatever related to the book. I also like being able to immediately post errata
the hour I find it, or am notified of it.

Many of you may find this type of Site to be of use to you in letting others have access to documents, data, and code that you wish to share. You can set this up for your classes as well - partitioning a site into separate classes where documents, etc, related only to each respective class are available to students for download. I have my site partitioned into 6 sections as I recall.

Bepress is short for Berkeley Press, a well known site for a number of electronic journals. Some are new well respected journals in their respective subject areas; eg Journal of Quantitative Analysis in Sports. The founder, a prof at UC Berkeley, wanted to design a site like this to help academics have a site to share their writings. But others in the research industry have found it useful too. After it is set up, a google search of your name will display the site, which can be accessed by a click. Others can find your site in this way, but you can also specifically give the address to them.
Or you can display the address on your university or business web site.

Technial support is available by email and phone if you have problems or want to do things with the site for which it was not designed. For instance, the site was not originally designed for the downlad of data sets contained in a large zip file. I inquired about it though, and the developers showed me how to use (trick) the software to allow this capability. I have found the assistance to be immediate and very helpful.

You can check out my site to see how it is organized by accessing


or access the main web page to learn more about it at


The above site describes the capabilities provided and gives several other example "Selected Works" sites by others in the academic and research industry (eg RAND) world. If you do get a site, you receive another address to edit the site, and to receive a report of how many times a document has been downloaded. Each month
you get a complete useage report emailed to you by BePress.

I found this site to be very useful in providing support to those who have obtained copies of my books. For the most part, I have placed only papers and articles that I think are related to the two books for which I first obtained the site. I put some others on the site as well for which I had been receiving quite few requests -- this way instead of me having to email (attachment) an article to someone who is requesting it, they can just download it on their own. You can design the site to meet your own needs, and add to or take from as you desire. There is a cost associated for institutional subscribers, but none for individuals. I believe that many in the Stata community can find having such a site to useful -- I highly recommend it. I've had mine for well over a year, with no problems.

Joseph Hilbe

Hilbe, Joseph M (2011), Negative Binomial Regression, 2nd edition, Cambridge University Press
Hilbe, Joseph M (2009), Logistic Regression Models, Chapman & Hall/CRC

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index