Date:         Sat, 27 Dec 2008 08:44:08 -0500
Reply-To:     Peter Flom <peterflomconsulting@mindspring.com>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject:      Re: Comparison between Observed and Predicted Probability using
              Poisson Regression
Comments: To: Sassy <AugustinaO@GMAIL.COM>
Content-Type: text/plain; charset=UTF-8

Sassy <AugustinaO@GMAIL.COM> wrote
>
>I am doing a analysis were I want to compared the observed number of
>hospital visits patients had to the predicted number of hospital
>visits they are supposed to have based on the medication they are
>taking. we are looking at visits within one year only.  I think this
>analysis requires a poisson regression using proc genmod, which I have
>not used before.  I also want to control for other factors that might
>affect the frequency of visits.  Does anyone have any idea on how to
>use poisson regression for this analysis or have experience using
>poisson regression?

You can indeed do this in SAS, and Poisson regression is probably a good start, although I wouldn't be surprised if you wound up needing a negative binomial model to deal with overdispersion.  Poisson regression assumes that the conditional variance is equal to the conditional mean, and, very often, the variance grows faster than the mean does.  You might wind up needing a zero-inflated negative binomial, since (in most populations, anyway) there will be a huge number of people with no hospital visits.

All of these have been discussed on SAS-L.  Poisson regression and negative binomial regression can be done in GENMOD, but the zero inflated models require NLMIXED.  Dale McLerran has written code to fit these models.

HTH

Peter

Peter L. Flom, PhD
Statistical Consultant
www DOT peterflom DOT com