-------- Original Message --------
Subject: Re: Prediction of landmark coordinates from other landmarks
Date: Mon, 17 Jan 2011 02:13:43 -0500
From: Pierre Guyomarc'h <[email protected]>
To: [email protected]
Thank you Stefan! I didn't think about that, I'll try and compare
errors. I still would prefer a prediction method as my reference sample
is not representative of ALL the existing shapes; and as the skull
correlates (moderately ok...) with the skin, prediction may be more
objective and less sample-specific. But the error rates may be better,
I'll keep you posted about that!
--
Pierre Guyomarc'h (PhD student)
Université Bordeaux 1 - UMR 5199 PACEA (CNRS)
/Anthropologie des Populations Passées et Présentes/ (A3P)
Av. des Facultés, Bât B8 - 33405 Talence cedex
On Sat, Jan 15, 2011 at 8:59 AM, Stefan Schlager
<[email protected]
<mailto:[email protected]>> wrote:
Hi Pierre,
you can use a nearestest neighbour method by generating a weighted
(by Procrustes or Mahalanobisdistance) mean of the nearest
neighbours - this avoids improbable prediction results. It works
pretty good on facial estimation.
If you are interested in R scripts, drop me aline or two.
stefan
--
Stefan Schlager M.A.
Anthropologie
Medizinische Fakultät der der Albert Ludwigs- Universität Freiburg
Hebelstr. 29
79104 Freiburg
Anthropology
Faculty of Medicine, Albert-Ludwigs-University Freiburg
Hebelstr. 29
D- 79104 Freiburg
phone +49 (0)761 203-5522
fax +49 (0)761 203-6898
On 14.01.2011 21:11, morphmet wrote:
-------- Original Message --------
Subject: Prediction of landmark coordinates from other
landmarks
Date: Fri, 14 Jan 2011 12:30:21 -0500
From: Pierre Guyomarc'h <[email protected]
<mailto:[email protected]>>
To: [email protected] <mailto:[email protected]>
Dear morphometricians,
I’m trying to use geometric morphometrics to predict the
coordinates of
a group of landmarks from another group of landmarks (of the same
individual). The goal is to predict human facial skin features
from bony
morphology. As I’m no mathematician or statistician, I have some
difficulties to evaluate the integrity of my methods. After short
discussion with some morphometricians, I heard different
opinions and
advices. That’s why I’m submitting my questions more largely by
throwing
this bottle to the sea! What do you think of:
-Method 1: predict each 3D coordinate of the unknown landmarks
(group 2)
from the PCs of a PCA ran on the known landmarks (group 1) through
multivariate regressions.
-Method 2: use best covariating groups of landmarks through PLS
analysis
and use the scores and PCA to predict the 3D coordinates of
group 2.
This methodology has been proposed at the 4th Meeting of Junior
Scientists in Anthropology (Freiburg im Breisgau, March 2010). The
proceedings are available at
http://www.freidok.uni-freiburg.de/volltexte/7603/ and my
contribution
is at p.84. Feel free to consult it if you have time. A more
complete
description of this method is exposed.
-Method 3: I would be really grateful if you can find me a third
method!
I’m open to all comments, even negative ones since they are
constructive…
Thanks!
