Rich is concerned about a predicted Poisson rate above one (see below).
If the data consist of counts of successes out of a certain number of
attempts, I would use a binomial model, which in Stata can be fit using
blogit or the glm command with family binomial. In both cases you get to
specify the binomial denominator (the number of trials or attempts) and the
fitted counts will never exceed that number.
The Poisson distribution can be viewed as an approximation to the binomial
for rare events, typically large number of attempts with small probability
of success in each one. In that case the fitted count will rarely exceed the
number of attempts. But why use an approximation when Stata can do the right
thing?
Cheers,
German
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Richard Goldstein
Sent: Tuesday, October 11, 2005 2:18 PM
To: [email protected]
Subject: Re: st: RE: Re: predict after Poisson
Thank you for both explanations. I assume that your point about an IR above
1
still holds if exposure is a number of attempts (and the numerator is the
number of successes), correct? I ask this because that is my situation and
here an IR above 1 means a predicted value of successes that is greater than
the number of trials (the exposure).
Are there sensible constraints that I could use on the model to keep the
predicted values to no more than the exposure? Are there any alternative
models that I should look into?
Thanks,
Rich
[earlier bits of exchange deleted]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
