Date: Mon, 26 Aug 2002 15:12:39 -0700
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: "David L. Cassell" <Cassell.David@EPAMAIL.EPA.GOV>
Subject: Re: Stratified Non-parametric analysis / confidence intervals
Content-type: text/plain; charset=us-ascii
Jones John <John.Jones@UCB-GROUP.COM> replied:
> I would recommend some variation on the Cochran-Mantel-Hanzel test.
> statistic is well documented in the FREQ procedure.
While I agree with John's excellent (of course!) advice, I just want to
point out one issue. The original poster wrote:
>> The van Elteren test could be an alternative, but according to a text
>> book, it is only applicable, if the results per center are not
>> contradicting each other (i.e., in one center treatment A is better
>> B and in an other center B is better than A).
The CMH tests also are at a disadvantage when the results by 'stratum'
not all aligned similarly. In fact, CMH tests are notorious for their
low power for detecting a significant association when the patterns for
some strata run opposite to those of other strata. So the poster needs
to check for this when reporting his results, because a nonsigificant
result *may* mean that no association was found, or it *may* mean that
there is no dominant pattern of association, while significant
can still exist within separate strata.
Ahh, the joys of real data...
David Cassell, CSC
Senior computing specialist