>>>I have tested for overdispersion and this is not a problem.
I'm always a bit wary of tests of overdispersion, mainly because I don't
have a clear idea of what the power of this test is. In other words,
how much overdispersion would there have to be before the test was
significant.
>>>Second, in theory, there is an unlimited number of groups that could
be formed in any given year, and thus there appears to be no
heterogeneity of risk, if I understand that concept correctly.
Well there must be some heterogeneity, otherwise there would be no point
in to modelling the data.
When I'm modelling disease incidence in a cohort of people, I'm assuming
that everyone in the cohort is at risk of the disease, but I'm not
assuming that they are at the same level of risk (no heterogeneity).
I'm assuming that there is heterogeneity in risk and I'm looking for
factors that explain this - factors associated with an increase or
decrease in risk.
>>>As for Stas's comments, I wholeheartedly agree that the reviewer in
question is not much of a reviewer.
I publish in medical journals and it is still the case that many medical
journals do not employ statistical reviewers. Consequently you can get
reviewers comments on the statistical analysis that are idiotic or
frankly bizarre. Often my colleagues (medical doctors) will want me to
simply do what the reviewer wants in order to get the paper published
thus turning a good paper into a bad paper (I wonder how often this
happens), whereas my response is to say "No, the reviewer is an idiot
and it is our obligation to point this out to them". Heated discussions
usually follow.
Without knowing exactly what your data looks like, it's difficult to
give you any more advice, but checking the residuals along the lines
that David suggested should enable you to check for autocorrelation.
______________________________________________
Kieran McCaul MPH PhD
WA Centre for Health & Ageing (M573)
University of Western Australia
Level 6, Ainslie House
48 Murray St
Perth 6000
Phone: (08) 9224-2140
Phone: -61-8-9224-2140
email: [email protected]
http://myprofile.cos.com/mccaul
_______________________________________________
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Antonio Silva
Sent: Thursday, 7 August 2008 5:01 AM
To: [email protected]
Subject: st: Autocorrelation in Poisson regression
I am very impressed with the quality of responses I have gotten, thank
you so much.
In response, I have some comments and a few more questions. First, for
Kieran, let me clarify: My dependent variable is "# of groups founded."
What this means is that for each year of the study (there are 40 years)
there is a number, which represents the number of new organizations that
come into existence in that year. So, for example, in 1965, 3 new groups
were formed, in 1966, 2 new groups were formed, and in 1967 0 new groups
were formed. So I have a number for each year in the period under study.
Does that make sense? I have tested for overdispersion and this is not
a problem.
Second, in theory, there is an unlimited number of groups that could be
formed in any given year, and thus there appears to be no heterogeneity
of risk, if I understand that concept correctly. Of course, Kieran may
feel that this particular count variable is simply not appropriate for
use in Poisson regression, and I am curious to hear your thoughts on
this.
As for Stas's comments, I wholeheartedly agree that the reviewer in
question is not much of a reviewer. In fact, he/she even included in
his/her review that he/she "was not sure if autocorrelation is even a
problem in Poisson regression," but that I should discuss it anyway. I
have looked everywhere, and all the books and articles I read on Poisson
basically imply (but do not explicitly state in a way that is quotable)
what Kieran said-they say that if a process truly is Poisson,
autocorrelation is not a problem. I think that Stas' suggestion (that I
include some language about there not being a standard test for
autocorrelation) is a very good one, and may well work. Though I have to
be honest, I am not sure what he means by discretization. Could you
indulge me a little more and tell me what you mean by this?
Finally, David, can you give me an idea of how I can generate the
deviance residuals after using the Poisson command in Stata? I thought
this option was available for other methods but not Poisson. And what
should I look for in the correlogram?
I am sorry to be asking such basic questions, but this is the first time
I have ever used Poisson regression. And to be honest, the reason I am
using it is because some other reviewer told me to because my dependent
variable was a count variable it was the best way to go.
Thanks again. This has been very helpful and useful to me.
_________________________________________________________________
Your PC, mobile phone, and online services work together like never
before.
http://clk.atdmt.com/MRT/go/108587394/direct/01/
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
