"Power analyses" are approximations, which are useful to apply for
the sake of your main hypotheses. Hypotheses should be stated
as clearly and plainly as possible. At times, what is convenient and
most convincing for a power analysis is a simpler variation of the
intended test -- for instance, describing the pre-post change from
beginning to end, instead of the test on a trend line.
I have seen and done a lot of power analyses, and I've never seen
one where I would want to use Kendall's correlations or gamma.
Kendall's will be hard to interpret, and gamma is heavily dependent
on the marginals, which makes it both hard to predict and hard to
interpret. I think it would be hard to display - and hard to defend -
a power analysis based on those statistics. (If you can, indeed,
imagine how to display your eventual power analysis, then you are
already close to designing a Monte-carlo experiment to do that
power analysis.)
So I'm saying, even if you find specific references for using them,
it is probably a bad idea. If you can't improve the design and proposed
measurements, simplify the test (for the power analysis) to a test that
uses something more self-evident, like a contingency test on a 2x2 table.
--
Rich Ulrich
Date: Wed, 28 Sep 2011 10:50:58 +0100
From:
[hidden email]Subject: minimum sample size
To:
[hidden email]
Hi
I am looking for a reference for minimum sample size for non-parametric tests including Kendall tau-b& c and Gamma
any suggestions?
thanks in advance
Muir
Locked
