-------- Original Message --------
Subject: RE: regression of Procrustes coordinates on classifiers
Date: Fri, 27 Feb 2009 12:35:59 -0800 (PST)
From: F. James Rohlf <[email protected]>
Reply-To: <[email protected]>
Organization: Stony Brook University
To: <[email protected]>
References: <[email protected]>

Not quite correct.

Using Hotelling's generalized T^2 statistic would be like using a
squared t-test statistic. Using Mahalanobis D^2 (i.e., D squared) is
like using the number of standard deviations by which two means
differ. That makes it a reasonable squared distance to use in some
applications. For example, if you assume a simple equal-rate
Brownian-motion random walk as your evolutionary model then time since
separation should be proportional to D^2. Using a generalized distance
is appropriate here because one assumes there would be more rapid
evolution in directions in which there is more variation within
populations.

On the other hand, if you want an absolute measure of the amount of
shape difference between two species then you should use Procrustes
distance (not squared).

I agree that any "correcting" should be done using linear models on
your shape variables in the tangent space (shape coordinates, partial
warp scores, etc.).

=========================
F. James Rohlf
Distinguished Professor, Stony Brook University
http://life.bio.sunysb.edu/ee/rohlf

-----Original Message-----
From: morphmet [mailto:[email protected]]
Sent: Friday, February 27, 2009 1:05 PM
To: morphmet
Subject: Re: regression of Procrustes coordinates on classifiers



-------- Original Message --------
Subject: Re: regression of Procrustes coordinates on classifiers
Date: Fri, 27 Feb 2009 07:29:36 -0800 (PST)
From: Joseph Kunkel <[email protected]>
To: [email protected]
CC: Joseph Kunkel <[email protected]>
References: <[email protected]>

Louis,

My first comment is about using Mahalanobis D as a 'distance'.  That
is really a misnomer.  Mahalanobis D is a sum of independent F-tests
when you look at how it is calculated.  The true distance is perhaps
the mean Euclidean distance and Mahalanobis D is the test of whether
that Euclidean Distance is actually significantly different from
zero.   Using Mahalanobis D as a distance is like using a t-test
value
as a measure of a difference-of-interest.  If that is what you want,
an abstraction of how many standard deviations you are away from
zero
that is fine.  But remember that F-tests explode (they are ratios of
squares so if the denominator is unstable the explosion is greater)
when the the null hypothesis is not true!    Sorry for sawing an old
horse, or am I wrong in a more basic way in this instance?

I would export the aligned coordinates, analyze then in a GLM
factorial design, remove your 'desert vs forest "factor"' before and
after calculating your distance and testing for their significance.
To be pedantic for the sake of novice readers of this list:  As you
suggest, in the General Linear Model, Y = XB, your factors belong on
the right in the design matrix X.  They are corrected for as
factors.
Continuous variables belong on the left as columns of the Y matrix
and
should be eliminated by the generalized test of additional
information. (Rao, 1965, Linear Statistical Inference and its
Applications)

I would avoid the perhaps easier route of analysis and suggest that
if
MorphoJ wants to include correcting for factors within a comparison
group that it be done in a formally correct way.   Doing things in
convenient ways will take us down the same road traveled by the
Windows operating System.

Joe


On Feb 27, 2009, at 9:01 AM, morphmet wrote:

>
>
> -------- Original Message --------
> Subject:   regression of Procrustes coordinates on classifiers
> Date:      Fri, 27 Feb 2009 05:55:11 -0800 (PST)
> From:      Louis Boell <[email protected]>
> To:        <[email protected]>
>
>
>
> Dear colleagues,
>
> I wish to investigate how strongly the fact that a sample of
specimens
> belongs to a given class, say, samples from desert vs samples from
> forest, influences the Mahalanobis distances between my samples.
This
> amounts to "correcting" for the desert vs forest "factor"
> and checking how much smaller the Mahalanobis distances between
desert
> and forest samples get when calculated from the residuals of the
> "correction".
>
> I have about 10 samples per class (each sample in itself
consisting of
> enough specimens given the number of landmarks) from each class).
>
> In this situation, I tried "correction" by using a dummy-coded
> regression (desert=1, forest=2) of Procrustes coordinates on my
factor
> in MorphoJ. The results are appealing. Now I know that for a
> noncontinoous independent variable, you´d prefer to use logistic
> regression instead of simple regression, because the fit will be
more
> appropriate in this case.
>
> My question is: is there a statistical reason not to use this
> procedure?
> Or caution about the interpretation of the results?
> I´d be grateful for any advice
>
> Best wishes
>
> Louis
>
> Louis Boell
> MPI für Evolutionsbiologie
> August-Thienemannstr.2
> 24306 Plön
> [email protected]
> [email protected]
>
>
>
> ------------------------------------------------------------------
------
> Sicher, schnell, übersichtlich - der Internet Browser vom
Marktführer!
> <http://redirect.gimas.net/?n=M0902IE8beta>
>
> --
> Replies will be sent to the list.
> For more information visit http://www.morphometrics.org

-·.  .· ·.  .><((((º>·.  .· ·.  .><((((º>·.  .· ·. .><((((º> .··.·
>=-       =º}}}}}><
Joseph G. Kunkel, Professor
Biology Department
University of Massachusetts Amherst
Amherst MA 01003
http://www.bio.umass.edu/biology/kunkel/




--
Replies will be sent to the list.
For more information visit http://www.morphometrics.org




--
Replies will be sent to the list.
For more information visit http://www.morphometrics.org

Reply via email to