s-news
[Top] [All Lists]

Conditional Logistic Regression

To: s-news@wubios.wustl.edu
Subject: Conditional Logistic Regression
From: Carlos Alzola <calzola@apa.com>
Date: Thu, 14 Dec 2000 17:27:22 -0500 (EST)
Reply-to: Carlos Alzola <calzola@apa.com>
Good afternoon,

We are running a matched logistic regression where some of the cases are 
matched 
1:1, but we also have 1:2, 1:3 and 1:4 matches.

The analyst originally ran it in SAS. He used PROC PHREG with cases coded as 1, 
controls as 0. Follow-up was defined to be 2 for controls and 1 for cases, 
stratifying in the case-control group. All this is explained in one of the SAS 
technical reports. We get coefficients in the order of 15 and -15 and all the 
standard errors had the same value of 1519. I don't have experience with 
conditional logistic regression, but these numbers made me think that the model 
didn't converge and SAS was silent about it.

So, I tried it in S-Plus (3.4, Unix, cph in Design library) and I got the 
message "log likelihood converged before ... beta may be infinite". The 
parameter estimates weren't quite as high as those of SAS (in the order of 2), 
but when I increased the number of interations the coefficients increased to 
values in the order of six. I observed the same pattern with standard errors 
(all 2.58 or something like that)

My questions:

Is this a case where the parameter estimates are infinity? Everything seems to 
indicate so.

Coxph requires to code survival time as a constant. Why does SAS do it 
differently?

Thanks in advance

Carlos Alzola




<Prev in Thread] Current Thread [Next in Thread>