Re: Mantel Haenszel test in very unbalanced situation

On 21 Sep 2005 05:32:29 -0700, billy.pilgrim@xxxxxxxxxxxxxx wrote:

> I have read that Mantel-Haenszel test does not require a large sample
> size for each stratum but only for the overall sample.
>
> However, in my problem I have the largest stratum with 60 units and the
> smallest one with only 8 units.
>
> Watching at the Mantel-Haenszel formula (a ratio of sum), I think that
> is is sensible to small sizes. In other words, it seems to me that
> moving a units from a category to the other one in the smallest stratum
> can change considerably the p-value of the test.
>
> Can someone tell me if my thinking is right?

MH uses the sum of the Expectations, and also the sum
of the Observed. If you combine categories that are
similar in "what happened", Expected versus Observed,
the MH should not change.

It *must* make a difference, in general, where the
cases are observed, or there would be no purpose in
having strata. - The only relevance I see in "small N"
is that it is easier to change the discrepancy between
Expected and Observed by changing a couple of cases,
when the N is smaller.

I hope that I am referring to the same MH test that you
are asking about. There are several tests by that name.


--
Rich Ulrich, wpilib@xxxxxxxx
http://www.pitt.edu/~wpilib/index.html
.