Date:         Mon, 23 Jun 2008 21:40:06 -0400
Reply-To:     Peter Flom <peterflomconsulting@mindspring.com>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject:      Re: ZIP results so confusing?
Content-Type: text/plain; charset=UTF-8

I wrote:

>> A couple points:
>> 1) SAS is notorious for the odd way it codes the LOGISTIC
>> models.  Often, SAS is doing the opposite of what one might
>> expect.  So, the first thing is to check that this is
>> correct.  This has had me scratching my head many times
>>
>

Dale replied....

>Good point!  However, in this case it would not be SAS that might have
>coded the model such that the parameters run counter to what one might
>think.  The ZIP model is constructed in NLMIXED, I believe.  One can
>always examine the code to ensure that zero values are what is being
>modeled.  The easy way to do this is to construct a binary response
>variable Y_alt as
>
>  if Y>=0 then Y_alt = (Y>0);
>
>Note that
>
>  Y_alt=0  if Y=0
>  Y_alt=1  if Y>0
>
>Now if Y_alt is passed as the response variable for PROC LOGISTIC with
>no DESCending option, then the LOGISTIC procedure should model the
>probability of Y_alt=0.  The parameters of this regression should be
>similar (though not exactly the same) as the parameters of the portion
>of the ZIP model which models the zero-inflation process.
>

Oh yeah, of course.
I've got LOGISTIC on the brain, as I am finishing up a NESUG paper called PROC LOGISTIC: Traps fro the unwary. :-)

But you're right, there's no LOGISTIC in the ZIP.  Unless she is doing it 'by hand' with a LOGISTIC and a Poisson.... in which case, she needs to search the archives and find your code about ZIP -- which is not only excellent code, but came with great commentary.


Another point that occurs to me, is that, in my experience, the Poisson is rarely right, and you usually want the Negative Binomial.  But the ZINB model is hard to estimate, especially with more than a couple IVs.

Peter

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