Date: Tue, 3 Jun 2008 17:12:18 -0500
Reply-To: Robin R High <rhigh@UNMC.EDU>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Robin R High <rhigh@UNMC.EDU>
Subject: Re: Zero Successes in Logistic Issue Creates Pairwise Problems
In-Reply-To: <43C07A163F7E764A8B27F6DAE2B126BF1AAF7D33@tpwd-mx0.tpwd.state.tx.us>
Content-Type: text/plain; charset="US-ASCII"
Warren,
Look at this paper by Paul Allison from the most recent SGF in San Antonio
http://www2.sas.com/proceedings/forum2008/360-2008.pdf
You should not apply PROC GENMOD in this situation; Proc Logistic has the
exact option, and you can also try the penalized approach with a SAS macro
available at:
http://www.meduniwien.ac.at/msi/biometrie/programme/fl/
I've had success with this latter approach with a zero count in n trials
from one of several group, though you need to dummy code data, including
the response (0/1); in some circumstances for reasons unknown, it doesn't
work.
Robin High
UNMC
Warren Schlechte <Warren.Schlechte@TPWD.STATE.TX.US>
Sent by: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
06/03/2008 03:35 PM
Please respond to
Warren Schlechte <Warren.Schlechte@TPWD.STATE.TX.US>
To
SAS-L@LISTSERV.UGA.EDU
cc
Subject
Zero Successes in Logistic Issue Creates Pairwise Problems
Suppose I have the following data (for illustrative purposes, these are
a simplification of data which are causing the problem):
Successes Trials Group
1 5 A
0 5 B
20 100 C
0 100 D
For groups A and C, my estimate of p-hat = 0.20, for B and D, p-hat =
0.00. Under the binomial approach, var(B)=var(D)=0.
However, if I were to compare A and B, it seems likely they could both
arise from the same underlying mechanism, but low sample size means I
might observe 0 successes; whereas the same comparison for C versus D
would suggest these likely arrive from two different mechanisms. A
Bayesian approach, using a uniform prior [0,1] would suggest what logic
suggests.
The question I have is, how do I get reasonable estimates from SAS of
the variance (and associated CI) for the p-hats when I have observed
zero successes? I am using data of this type within the Proc Genmod
framework and getting very unreasonable CI for my groups B and D.
Because of this, I cannot seem to get reasonable pairwise tests of my
groups (my SE are much too large). Do I need to do my SE calculations
and comparisons by hand afterwards or is there a SAS fix?
Thanks,
Warren Schlechte
HOH Fisheries Science Center
5103 Junction Hwy
Ingram, TX 78025
Phone 830.866.3356 x.214
Fax 830.866.3549
