Fisher's test vs. other?
On Thu, 13 Jul 2000 12:49:24 GMT, Mithat Gonen <[EMAIL PROTECTED]>
wrote:
> I don't think there is an agreement in the statistical community. In the
> clinical trials circles conditional tests are highly regarded and the
> most convincing evidence I have seen so far is Yates (1984, J. of Royal
> Stat Soc, A147:426-463).
< ... >
That is a whole set of articles, and they are worth reading. The
Yates cited as first author (1984) is the same Yates who devised the
2x2 correction term in c. 1935.
(a) There seems to be agreement that the 2x2 test with Yates's
correction does a fine job of approximating or estimating Fisher's
Exact test (2x2).
(b) There is agreement that this test is appropriate when "both sets
of marginals are fixed." For instance, the Group marginals are fixed
if you are comparing two regular groups. The <other var> marginals
are fixed if - for instance - you split on the Median in order to tab.
(c) Here is a cute, extra fact. Yates points out that the 2x2 test,
uncorrected, was devised while *assuming* that both marginals are
fixed. It is only a happenstance that it gives a 2x2 test that some
people prefer. Okay. I still prefer it, because it works.
When I did my own randomization testing, using margins that were
random proportions (not fixed), there were only 3%-3.5% rejections by
Fisher's or by the corrected test, and 5% by the raw test. So I like
the raw test. And the other arguments have never gotten through to
me. Except, of course, you do the test that the journal will accept.
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
