-------- Original Message --------
Subject: Re: Splitting up a data set versus combining it for CVA
Date: Wed, 10 Aug 2011 07:05:19 -0400
From: Carmelo Fruciano <[email protected]>
To: [email protected]

morphmet <[email protected]> ha scritto:



-------- Original Message --------
Subject:        Splitting up a data set versus combining it for CVA
Date:   Mon, 8 Aug 2011 08:00:31 -0400
From:   Alexandra Wegmann <[email protected]>
Reply-To:       Alexandra Wegmann <[email protected]>
To:     [email protected] <[email protected]>



Dear Morphometricians
I am trying to publish the results of my master thesis. In the thesis I
did a separate canonical variates analyses for each question including
only the procrusted data of the specimens of interest (i.e. one data set
just contained the adults, another one only the juveniles, another one
just the laboratory reared specimens etc). A paper should however be as
short as possible, therefore I am considering combining all the data
into one data set, running the analysis once and then just discussing
the groups in question in each chapter (like this I only have to explain
the morphological changes associated with each axis once in the paper).

Dear Alexandra,
I wasn't able to understand the exact meaning of your question.
You first write about separate CVAs "for each question", then you
think about pooling all observations.

My point is: is there a single "question" or more?

A typical example would be if you were interested in the difference
between two (or more) groups of individuals (say, sampled at two
geographical locations) but in those two groups there is variation due
to other causes such as growth-related shape changes. In this case a
typical approach (limitations may apply, of course) would be to run a
MANCOVA using as covariate size (as proxy for age) to control (in the
example I made) for growth-related variation and testing for
differences between groups (after growth-related variation has been
controlled). General linear models can incorporate both categorical
and continuous predictors so you could use (within reason) lab
reared/non lab reared, size and so on as predictors controlling for
the others (and also probably testing for the interaction terms).
Given a specific problem there could be limitations to this approach,
but the general idea is widely used.

Now, if the "questions" you ask about your groups are fundamentally
distinct and depend on the group (for instance, you test for variation
between sampling sites in wild specimens and the effect of temperature
in lab-reared specimens), I wonder why you would want to pool all
observations.

Well, I hope that this is of even remote help...
Carmelo


--
Carmelo Fruciano
Dipartimento di Biologia
University of Catania
Tel. +39 095 7306023
Cell. +39 349 5822831
e-mail [email protected]
http://www.fruciano.it/research/

----------------------------------------------------------------
Universita' di Catania - A.P.Se.Ma.
Servizio di Posta Elettronica



Reply via email to