Date:         Thu, 15 May 1997 06:03:37 GMT
Reply-To:     Gerard Smits <gerards@DELTANET.COM>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         Gerard Smits <gerards@DELTANET.COM>
Organization: Delta Internet Services, Inc.
Subject:      Re: Test of Poisson Distribution

I recently tried genmod on a distribution I though might be Poisson.  It
was the number
of days in ICU (hospital).  Clearly, the data do not meet the independence
criterion, but
I thought I'd check it out and compare it to a rank sum test.  I used
resampling with a
monte carlo loop of 1000.  I found, after feeding a 1 day difference (this
was a test of power)
that the genmod was not as powerful as the rank sum.
 
More inportantly, I reran the simulation with no difference between the two
"identical"
resampled groups.  Using an alpha of .05, I got a Type 1 error rate of 5%
with the
rank sum, but a 35% error rate with the genmod poisson regression.  I later
generated
true poisson variates and got the error rate I expected.
 
So, in short, the poisson model does not appear to be robust with respect
to violation
of the distrbution assumptions.
 
 
 
Charlie Hofacker <chofack@cob.fsu.edu> wrote in article
<3378BD8D.7D7@cob.fsu.edu>...
> Dear SAS wizards and assorted hangers on,
>
> I have a set of count data which represents the incidences of
> certain behaviors across a large sample.  A frequency distribution
> reveals something that looks a lot like a Poisson with zero
> being the most common frequency, followed by 1 and so forth.  Also
> the mean and the variance are pretty similar.
>
> I would like to test the hypothesis that the incidence of this
> behavior is Poisson, preferably by maximum likelihood.  I originally
> thought I could use GENMOD to do this with an intercept only model,
> but it didn't seem to fit the bill.  I vaguely remember a macro
> to do this reported in a SUGI Proceedings from the early 1980's.
>
> Any ideas would be appreciated....
>