-------- Original Message --------
Subject: RE: Fwd: CVA versus MANOVA
Date: Wed, 23 Mar 2011 09:24:22 -0400
From: Sheets, H David <[email protected]>
To: [email protected] <[email protected]>
You always want to run cross-validation (or jack-knifing) with CVA. The
plots you get of CVA scores are called "resubstitution" values, from
scoring the same specimens you used to create the CVA axes. These
scores always overstate the effectiveness of the CVA, you really need to
look at the jack-knife rates, particularly at small sample sizes.
There is a jack-knife assignment command in IMP CVAGen, which you really
need to use to determine the effectiveness of the CVA.
-Dave
H. David Sheets, PhD
Chair and Professor
Dept. of Physics
Canisius College
2001 Main St
Buffalo, NY 14208
________________________________________
From: morphmet [[email protected]]
Sent: Wednesday, March 23, 2011 7:51 AM
To: morphmet
Subject: RE: Fwd: CVA versus MANOVA
-------- Original Message --------
Subject: RE: Fwd: CVA versus MANOVA
Date: Tue, 22 Mar 2011 17:34:50 -0700
From: Sarah Degroot <[email protected]>
To: <[email protected]>
In SPSS you can use the "leave one out classification" option for
cross-validation.
Frequently I also randomize my data and re-run the analysis with the
leave-one-out classification, just to see if the actual data groups
cases better than random data.
Sarah De Groot
Graduate Student, Rancho Santa Ana Botanic Garden and Claremont Graduate
University
________________________________
From: morphmet [mailto:[email protected]]
Sent: Mon 21/03/2011 5:44 AM
To: morphmet
Subject: Re: Fwd: CVA versus MANOVA
-------- Original Message --------
Subject: Re: Fwd: CVA versus MANOVA
Date: Mon, 21 Mar 2011 08:03:00 -0400
From: Øyvind Hammer <[email protected]>
To: <[email protected]>
Hi, this is normal ... if you include a large number of variables
compared with cases in the CVA, you will see a strong separation even
for completely random data with no real groups. The MANOVA, on the other
hand, will adjust for the number of variables, and (correctly) report
non-significance.
In such cases, you will see that the seemingly successful
classification breaks down completely if you run a cross-validation
(jack-knifing) on the CVA.
A CVA should always be accompanied by a MANOVA to check that the groups
are "real".
Or something like that.
Øyvind Hammer
Natural History Museum
University of Oslo
Hi,
I am analyzing a dataset with 21 landmarks and 2 groups. I have run
CVA
on both IMP and MorphoJ and it suggests a very strong grouping. The
deformation plot in IMP shows 3 areas of deformation while the
MorphoJ
plot doesn't show any obvious deformations on CV1 (the significant
axis
as displayed by IMP). However when I ran a univariate ANOVA on the
centroid sizes and a MANOVA on the partial warps and relative warps
with
SPSS there was no significant difference between groups.
I am new to this type of analysis and I was wondering if perhaps I am
missing something?
Any suggestions would be greatly appreciated.
Thanks,
Michelle
