|Date: ||Wed, 16 Jul 1997 13:26:54 +1200|
|Sender: ||"SPSSX(r) Discussion" <SPSSX-L@UGA.CC.UGA.EDU>|
|From: ||A Dharma <dharma@WAIKATO.AC.NZ>|
|Subject: ||changing ref category in log regression|
|Content-Type: ||text/plain; charset="us-ascii"|
I have the following estimated odds ratios from logistic regression (I don't
have access to the orginal data).
Secondary schooling--1.00 (ref category)
The estimates are sig at p<.05 (again no access to actual std errors).
To use the above results, I need to change the refe category: to No
schooling. I have done it as follows:
No schooling--.32/.32=1.00 (ref category)
Now I have two questons:
Is this conversion technically ok? Is there any way to test the statistical
sig of the estimates?
For some reason I decided to collapse the two categories (prim and
secondary) into one. To get a combined odds ratio I weighted the individual
odds ratios, and the weights being the proportion of cases with primary and
secondary education. For example, if a is the proportion with primary educ
and b is the proportion with secondary educa (such that a+b=1.0), then the
weighted estimate is: (1.31 times a)+(3.125 times b).
Of course the weighted estimate cannot exactly be the same as the one we
would get if we did the collapse in the original data and did the analysis.
My question is would this weighted one would be closer to the one we would
get if we manipulated the orignal data rather than manipulating the estimates?
Thank you very much for your time.