---------------------------- Original Message ----------------------------
Subject: Re: Quantifying landmark variability
From:    "F. James Rohlf" <[EMAIL PROTECTED]>
Date:    Sat, November 5, 2005 6:30 am
To:      "Morphmet" <[email protected]>
--------------------------------------------------------------------------

Note that the covariance matrix will be singular. 4 zero eigenvalues. This
standard discriminatin methods cannot be used on the alligned data
directly. You also cannot interpret the landmark displacements so
directly. The apparent displacements at one landmark are also a function
of all of the other landmarks.

Jim Rohlf
-------
Sent remotely by F. James Rohlf

-----Original Message-----
From: morphmet <[EMAIL PROTECTED]>
Date: Fri, 04 Nov 2005 05:02:05
To:morphmet <[email protected]>
Subject: RE: Quantifying landmark variability

Hi Antigoni,
Regarding (a):
        If you accept that the Procrustes algorithm is finding the true
relative rotations and translations of each lizard, then the registered
landmarks can be treated as if they are the true raw positions of the
landmark coordinates.  That is, the registered landmarks can be treated as
if they had not been arbitrarily rotated and translated.  Hence, Sx + Sy,
the standard deviations of the landmarks in the x and y directions, will
give you the amount that each landmark "moves".  Interpreting Sx and Sy
separately is not sensible because the x and y axes are arbitrary.

Regarding (c):
        If are treating the registered landmarks as unrotated and
untranslated data, as above, then linear discriminant analysis (also
called canonical discriminant analysis) will give you a nice function
highlighting the group differences in each landmark.  The loadings of the
function will tell you which landmark is most different across groups.
        A similar approach is to use logistic regression, with the group as
the binary response and the registered landmark coordinates as the
continuous predictor variables.  Logistic regression is more robust to
non-normality than linear discriminant analysis and the coefficients of
the logistic regression function are interpreted similarly to linear
regression coefficients (look it up).  If you have more than two groups,
ordinal or Poisson regression is the corresponding extension to logistic
regression.
        Both linear discriminant analysis and logistic regression are
available in most standard statistics packages.  There are other methods,
for example, regression trees, but their results may not be so easy to
interpret.  A good intro to the methods I have mentioned can be found at
http://www.statsoftinc.com/textbook/stathome.html.

Both the methods I describe here assume that it is valid to analysis the
covariance matrix of the Procrustes-registered data.  How acceptable is
this?
        Ben Flood
        Ben dot Flood at ul dot ie.


-----Original Message-----
From: morphmet [mailto:[EMAIL PROTECTED]
Sent: 03 November 2005 12:45
To: morphmet
Subject: Quantifying landmark variability


Good morning morphometricians

I am a PhD student studying morphology of lacertid lizards in the
Iberian Peninsula and for the last year I've been trying to apply GM for
my investigation. Although I've passed a year trying to figure out how GM
methods could be applied to my lizards, there are still some issues that
I'm not very sure about, no matter how preliminary they might seem. I hope
someone here might help me.

I was wondering if there is some way of putting into numbers variability
from landmark data. Intuitively I would say there must be because that's
what they were developed to do, but I'm not sure which of the parameters
calculated during the analysis would be a measure of that. In detail: a)
Is there a way of putting into numbers and testing which of the landmarks
in a sample present more global variablility ("move" more)? Would the
variances (S, Sx, Sy) around landmarks given by the report of tpsRelw be a
measure of that?

b) When comparing groups within a sample, is it correct to evaluate local
differences by examining partial warps independently? I know PWs are
correlated and should not be separated for statistical analyses, but is
there some way of quantifying and examining very local differences?

c) Again when comparing groups within a sample, is there some way of
evaluating which are the landmarks that contribure more to the observed
differences? For example, when comparing male to female lizards I have run
a canonical analysis on PWs and calculated canonical scores. Then I
regressed shape variables to the canonical scores in tpsRegr to view the
shape changes. When selecting to view them as vectors instead of
deformation grids, I can see that some vectors are notably "longer" than
others, but what is the variable that expresses the length of these
vectors? I suppose it must be something related to the mean Procrustes
Distance for each group (sex in my case), but I'm not completely sure
about it.

I'm sorry if this sounds too elemental, but passing from mathematics to
the biological application of GM seems to be quite a step to take after
all.
I'd be very thankful for any suggestions on the above problems.
Cheers
Antigoni
-- 


Antigoni Kaliontzopoulou
Centro de Investigação em Biodiversidade
e Recursos Genéticos (CIBIO/UP)
Campus Agrário de Vairão
4485-661 Vairão
PORTUGAL
Herpetologia, Dep. Biologia Animal (Vertebrats)
Fac. de Biologia, Univ. de Barcelona
Av. Diagonal 645
08028 Barcelona
SPAIN

tel: +351 91 3086188
mail to: [EMAIL PROTECTED]
-- 
Replies will be sent to the list.
For more information visit http://www.morphometrics.org
.




Reply via email to