I have compared fits between SAS phreg and Splus coxph on scores of
data sets. Both procedures are reliable, and their answers agree. A
query about "why are they different" gets sent to me once or twice a year.
In order of frequency the usual reasons are
a. phreg was run with ties=breslow (its default) and coxph with
ties=efron (its default).
b. the SAS and Splus data sets differ.
c. trivial differences due to convergence.
As to the last one: phreg has a smaller "epsilon" for convergence than coxph,
and often does one more iteration. I am opinionated on this, if the se of
beta is .1, say, then iterating until the MLE is correct to the .001 digit
is silly -- a really smart program wouldn't even print those digits.
It is possible that you have found a test case in which one or the other
of the programs is in error. I am always interested in such potential cases,
but need more details, and probably a copy of the data, to follow up.
My favorite "differences" email to date was the one that began "I have
found a bug in your coxph program; it gives a different answer than SAS".
Terry M. Therneau, Ph.D. (507) 284-3694
Head, Section of Biostatistics (507) 284-9542 FAX
Mayo Clinic therneau.terry@mayo.edu
Rochester, Minn 55905
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|