I don't believe that Mantel-Haenszel is appropriate for either of the tests that the original correspondent wanted to address. By way of recall, the problem was one in which a response measured at two time points (pre- and post-treatment) was obtained on case and control subjects. Thus, we have correlated responses. The correspondent wanted to test whether there was a difference between pre- and post-treatment responses within each of the case/control groups. McNemar's test is appropriate for the test of association when the responses are correlated as they are when examining the response vs pre-/post-treatment by group.

The second test that the correspondent wanted to make was to see if the change between pre- and post-treatment responses was different for cases and controls. This might be where a Mantel-Haenszel statistic comes into play. We do now have a third (stratification) factor to consider. But Mantel-Haenszel is used to examine whether there is an association between X and Y across all strata defined by Z. That is not what is wanted for this third test. Moreover, Mantel-Haenszel does not allow for correlated responses. The Breslow-Day test is used to examine whether the odds ratio is homogeneous across strata. So, one might consider that statistic rather than the Mantel-Haenszel statistic. However, to the best of my knowledge, Breslow-Day does not allow for correlated responses, either. Thus, I firmly believe that the second question "Is the change from pre-test to post-test different between case and control?" can only be examined through a model which takes into account the correlated responses. GEE and a generalized linear mixed model are two such methods.

Dale

--- "David L. Cassell" <Cassell.David@EPAMAIL.EPA.GOV> wrote: > "Dennis G. Fisher" <dfisher@CSULB.EDU> replied: > > I am not sure if a Mantel Hansel approach would work for you or > not. > > That's Mantel-Haenszel, also known as Cochran-Mantel-Haenszel, > although some people make a distinction based on the test. Often, > the 1-degree-of-freedom correlation test is called the > Mantel-Haenszel > test, while the 'ANOVA' statistic and the 'general association' > statistic (together with the first-mentioned test) are often > collectively called Cochran-Mantel-Haenszel tests. > PROC FREQ does all three CMH tests (when appropriate). And > yes, I agree with you that this approach may be reasonable. > However, Dale's suggestion may provide more power. I don't know > of a reference off the top of my head which would provide a > recommendation either way... > > David > -- > David Cassell, CSC > Cassell.David@epa.gov > Senior computing specialist > mathematical statistician

===== --------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@fhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------

__________________________________________________ Do you Yahoo!? HotJobs - Search new jobs daily now http://hotjobs.yahoo.com/