[sharkthread]
Sender: [EMAIL PROTECTED]
Precedence: bulk
Reply-To: [EMAIL PROTECTED]
My response to Fred...
I purposely placed the term "designed morphospace" in my comment to
see if it would draw fire. It did. Though I am rethinking my initial
willingness to be a whipping boy - we need this debate if papers validly
using MANOVA are getting rejected for publication.
Fred may have been right that MANCOVA subverts Procrustes
distance. But I believe he is wrong that this is "for no good reason".
He argued that the MANCOVA approach fails to improve explanation and
that it destroys any possibility for shape interpretation. Below I
demonstrate why he is wrong.
Semantics - In my view a canonical axis built to explain systematic
variance in partial warp scores, between groups or along a gradient, is
a legitimate technique (see also Bookstein's Combining chapter in the
White Book). I see no reason to avoid calling this a morphospace,
though I dare not say shape space in mixed company.
Semantics aside, MANCOVA is simple, effective, practical,
philosophically sound and as far as I can fathom, is perfectly rigorous
as well. The first three issues (simplicity, efficacy, practicality)
are important but often neglected in this field.
_Simplicity, efficacy, practicality_
MANCOVA provides an accessible means for all practitioners to
visualize shape change using their current skills in statistics.
MANCOVA (regressing PWs against predictor variables and interactions)
generates axes describing shape differences between groups or over
gradients. It does nothing different than tpsRegr, except it offers the
flexibility to work in a stats package. Thus one can analyze more
complicated MANCOVAs than one can code in tpsRegr. Furthermore, in a
stats package one can readily enjoy all the best multivariate delights
such as centroids (multivariate least square means), confidence
ellipsoids, measures of effect size (partial eta-squared) and centered
polynomials.
Perhaps most importantly, my experience says visualizations from
this approach accurately depict the nature of shape differences due to
main effects of interest. Once I create a visualization of an axis, I
can train my eye on shape differences between groups and then classify
unknowns to their correct groups with great accuracy. Being a
naturalist, that is what sold me on Fred's and Jim's methods in the
first place. I am a strong advocate of geometric morphometrics,
especially in the MANCOVA context, because it works - it powerfully
discriminates groups (or ends of a gradient) AND the canonical axes meet
my intuitions, based on functional principles, about the manner in which
shape SHOULD vary.
_Philosophy - What do you want from your morphometrics?_
What many of us want from our morphometrics is to define an axis
that tells how shape varies across groups or gradients. That is, we
want a "designed morphospace". An elegant design must be simple, such
as a linear gradient of shape change related to something of theoretical
interest about organisms. That axis describes an EFFECT of interest,
and it rightly should reflect those aspects of shape that vary most
across groups or gradients. The rest is dross.
I want a morphospace designed based on variation in organismal shape
variance. For example, suppose PW1Y fully describes the difference
between fish from populations having or lacking predators. If that PW
is related to morphology in a way that suggests greater escape ability
in fish from sites with predators, then my morphospace should rightly be
built along this gradient (i.e. it should weight PW1Y heavily).
_Rigor_
Here I am less opinionated and decidedly liberal. Fred's methods
sound great if one can do it that way, and teach it to one's graduate
students, and still have an easy luxury of visualization, hypothesis
testing, and simplicity of philosophy. I get all these things quickly
from a MANCOVA using JMP and visualization in tpsRegr.
I don't believe MANCOVA lacks rigor. I believe it rigorously
captures THE MOST RELEVANT aspects of phenotypic variation. In my world
view a designed morphospace should demonstrate shape differences
explicit among groups chosen for study based on a priori hypotheses.
-Thom
________________><()()(>________________
Dr. Thomas J. DeWitt, Assistant Professor
Department of Wildlife & Fisheries Sciences
& Program in Bioenvironmental Sciences
Texas A&M University
2258 TAMU
College Station, TX 77843-2258
Tel. (979) 458-1684 (office)
Tel. (979) 845-7522 (lab)
Fax (979) 845-4096
E-mail [EMAIL PROTECTED]
Web http://wfscnet.tamu.edu/faculty/tdewitt/webpage.htm
TAMU Map to DeWitt lab & office:
http://www.tamu.edu/map/gifs/detail/FGHB.gif
==
Replies will be sent to list.
For more information see http://life.bio.sunysb.edu/morph/morphmet.html.
