dthulman <[email protected]> ha scritto:
Hello All,
I recently reviewed a paper in which the authors were analyzing artifact
shapes using landmarks-based geometric morphometrics. They performed a DFA
and permutted the data 1000 times. They then modified the critical values
before comparing the results using a correction to control for false
discovery rate.The formula is in:
Benjamini Y, Yekutieli D (2001) The control of the false discovery rate in
multiple testing under dependancy. Ann Stat 29:1165–1188.
I've not seen this used before and was wondering if it is appropriate in
DFA's for morphometric data. I've been evaluating the effectiveness of a
DFA using bootstrapping and a misallocation table and not relying on the
p-values.
Dear Dave,
that method (and others, related) is used widely in fields where one
does a really large quantity of tests (microarray data is the
classical example, with thousands of tests on per gene). The reason
for this popularity is that it is much less conservative than other
classical methods, such as the classical Bonferroni or the sequential
Bonferroni (introduced by Holm 1979 Scand J Stat and later on
popularized in evolutionary biology by Rice 1989 Evolution). With such
large numbers of tests, using (sequential) Bonferroni adjustements
would leave very few significant tests (if any).
Although I guess there should be no special problem in using the
Benjamini Hochberg and related methods with morphometric data,
normally morphometric studies do not perform so many tests (wether the
p-value is obtained through resampling or not).
Just a couple of other comments:
- not everyone agrees on the necessity of corrections for multiple
tests (see Perneger 1998 BMJ for a brief list of the reasons) and
actually controlling for false discovery rate (as in Benjamini
Hochberg and related methods) has been suggested as a possible way to
reconcile different views about the necessity of corrections for
multiple tests (see Garcia 2004 Oikos for a comment more or less along
these lines)
- I guess that (cross-validated) classification rates in discriminant
analysis and multivariate tests for difference in means, although
somehow related, can still be thought to serve two different purposes
so I personally don't think one should be abandoned in favour of the
other
Best,
Carmelo
--
Carmelo Fruciano
Marie Curie Fellow - University of Konstanz - Konstanz, Germany
Honorary Fellow - University of Catania - Catania, Italy
e-mail [email protected]
http://www.fruciano.it/research/
--
MORPHMET may be accessed via its webpage at http://www.morphometrics.org
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
