Date:         Thu, 23 May 1996 00:12:02 GMT
Reply-To:     Netnews Server <NETNEWS@AMERICAN.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
Comments:     RFC822 error: <W> Incorrect or incomplete address field found and
              ignored.
Comments:     RFC822 error: <W> Incorrect or incomplete address field found and
              ignored.
Comments:     RFC822 error: <W> Incorrect or incomplete address field found and
              ignored.
From:         Netnews Server <NETNEWS@AMERICAN.EDU>
Organization: Fred Hutchinson Cancer Research Center
Subject:      Re: PROC LOGISTIC warning message

In <4nvnod$4gu@babbage.ece.uc.edu>, MOORDI@ucunix.san.uc.edu writes:
>
>A while back Michel Sylvain <michel.sylvain@HEC.CA> wrote:
>> =========================================================================
>> From SAS-L_Archives Thu Apr 25   10:25:00 1996
>> Reply-To:     Michel Sylvain <michel.sylvain@HEC.CA>
>> Sender:       "SAS(r) Discussion" <SAS-L@VM.MARIST.EDU>
>> From:         Michel Sylvain <michel.sylvain@HEC.CA>
>> Subject:      warning message
>>
>> I am using the proc logistic procedure (link=normit) and I would like to
>> know what does this warning means: "There is possibly a quasicomplete
>> separation in the sample points. The maximum likehood estimate may not
>> exist".  What should I do to correct this error.
>
>A user here has also received this message, but no one seems to
>know exactly what it means.  If anyone has a clue regarding this
>warning, could you please explain it to me.  I would definitely
>appreciate it.  (FWIW, the link option doesn't seem to make any
>difference, since we used a logistic link function.)
>
>
>Thanks,
>    D. Moore
 
 
David,
 
Suppose that you have a predictor variable which takes on only two values.
Complete separation exists if you have the situation
 
                                  Predictor
 
                                  0   |   1
                              -----------------
                              |       |       |
                Response  0   |  25   |   0   |
                              |       |       |
                              -----------------
                              |       |       |
                          1   |   0   |  25   |
                              |       |       |
                              -----------------
 
That is, there are no successes when the value of the predictor variable
is 0, and there are no failures when the value of the predictor variable
is 1.  Obviously, this represents very good classification, but you cannot
get maximum likelihood estimates of the odds ratio as the value p/(1-p)
is not defined for one of the levels of the predictor variables.
 
If one of the off-diagonal cells is nonzero, then you get quasi-complete
separation.  Depending on how the data are set up, you may not be able to
compute the odds ratio [p1/(1-p1)]/[p2/(1-p2)].  I will let you play with
the 2x2 model to see in what situations you can and cannot compute an odds
ratio.
 
 
Dale McLerran
Fred Hutchinson Cancer Research Center
1124 Columbia Street
Seattle, WA 98104