Re: odds ratios in JMP vs SAS - why are they different?

["Davern, Sean" <sdavern@amgen.com>]

Sam,

  This question might best be addressed by sending your email to support@jmp.com.

Best of luck,

Sean

 

Sean Davern
Amgen
1201 Amgen Court West
Mail Stop AW2/D2152
Seattle, WA 98119-3105
Phone: (206) 265-7074

-----Original Message-----
From: jmp-l-owner@lists.biostat.wustl.edu [mailto:jmp-l-owner@lists.biostat.wustl.edu]On Behalf Of S. Brar
Sent: Friday, January 13, 2006 2:40 PM
To: jmp-l@lists.biostat.wustl.edu
Subject: [jmp-l] odds ratios in JMP vs SAS - why are they different?

I am having difficulty 
reconciling the differences in the odds ratios from the same data set while 
performing a logistic regression with JMP (v5) vs. SAS (v9). 

When the 
variable is categorical with only two response levels both JMP and SAS calculate 
the same odds ratios. When the categorical variable has 3 or more response 
levels the ORs differ between JMP and SAS. Changing to reference coding from 
effects in SAS doesn't change the ORs (the estimates do change). I'm not sure 
how JMP is coding the dummy variables and why the outputs differ. Following is a 
summary of my results:

Model: female = ses read. Female (0/1, probability 
modeled is female=0). ses (three responses: 1, 2, 3). Read is a continuous 
variable (min value 28, max value 76).

JMP Logistic Regression 
Analysis

    Term        
Estimate       
SE        
ChiSq       OR      
ORLower    ORUpper

Intercept     -0.486         0.781       0.39
ses[1]        -0.482         0.243       3.94        0.382     0.143           0.971
ses[2]         0.241         0.193       1.55        1.619     0.761           3.474
read            0.004         0.015       0.09        1.232     0.308           4.939

SAS output

Intercept     -0.486         0.781       0.39
ses 1         -0.482          0.243       3.94        0.486      0.211            1.118
ses 2          0.241          0.193       1.55        1.000      0.513            1.952
read            0.004          0.015       0.09        1.004      0.976            1.034

My SAS statement is as follows:

proc logistic data= "">    class ses PARAM = REF;
    model female =  ses read / CLODDS = BOTH;
run;

The "Whole Model 
Test" generated by JMP is the same as the Likelihood ratio Chi-Square in SAS for 
"Test Global Null Hypothesis."

Can someone explain the reasons for these 
differences in both the categorical variable with >2 response levels (such as 
'ses' in this example) and why the continuous variable 'read' is different? 
Since the estimates are the same it must be a difference in how the two packages 
calculate the odds ratios. 

-- 
Sam Brar, MD
Kaiser 
Permanente/UCLA
Fellow, Division of Cardiology
Office: 
323-783-4915
Fax: 323-783-5509