----- Forwarded message from Øyvind Hammer -----
Date: Sun, 29 Apr 2012 08:01:27 -0400
From: Øyvind Hammer
Reply-To: Øyvind Hammer
Subject: Re: MorphoJ and CVA
To: [email protected]
Hello, your multivariate test is "correct", the univariate one is
"incorrect".
Picking out only the single most discriminatory axis from CVA for
statistical testing is "cheating". Especially if you have many variables
compared with cases, the first CVA axis can show fairly "good"
discrimination even for completely random data (try it), and a t test on
the first axis scores can come out significant.
If you do a "parametric" MANOVA based on a statistic such as Wilks
lambda, it will adjust for the degrees of freedom. A multivariate
permutation test does not have to be adjusted; the p value magically
comes out right "by itself".
Sorry for all the "quotation marks" above :)
Oyvind Hammer
Natural History Museum
University of Oslo
On Wed, 25 Apr 2012 17:01:37 -0400, morphmet
<[email protected]> wrote:
> -------- Original Message --------
> Subject: MorphoJ and CVA
> Date: Wed, 25 Apr 2012 13:46:30 -0400
> From: Robert Ward <[email protected]>
> To: [email protected]
>
> Please forgive me if I'm making an obvious error. I'm trying to
> reconcile different outputs relating to CVA results in MorphoJ.
>
> The analysis is looking for sex differences in a 6-landmark shape and
> the output from MorphoJ is shown below (I left out the Procrustes
> tests and canonical coefficients). The MorphoJ output seems to
> suggest
> that CVA was not able to find a set of shape features to make a
> highly
> reliable discrimination, p=.06, not terrible but not great.
>
> On the other hand, if I export the CV1 scores from that analysis, and
> compare the male and female scores, then I get highly significant
> differences with either a two-sample t-test, t(85)=3.8, p=.0002, or a
> two-sample permutation test, Z=3.55, p=.0004.
>
> So I reckon I am misunderstanding something somewhere. If CVA is
> finding a vector through shape space that best discriminates the two
> groups, and the CV1 score reflects position on this vector, then
> shouldn't it be fine to test for a sex difference by a two-sample
> test
> of some kind? If so, then I wonder why the big discrepancy between
> the
> MorphoJ results, and tests using the CV1 scores?
>
> Thanks for your help,
> Rob
>
> Canonical Variate Analysis: CVA x6 ... Sex
> Dataset: mouthx6
> Classification criterion: Sex
> Groups Observations
> 1. F 40
> 2. M 48
>
> Variation among groups, scaled by the inverse of the within-group
> variation
> Eigenvalues % Variance Cumulative %
> 1. 0.16891938 100.000 100.000
>
> Mahalanobis distances among groups:
> F
> M 0.8160
>
> P-values from permutation tests (10000 permutation rounds) for
> Mahalanobis distances among groups:
> F
> M 0.0621
----- End forwarded message -----
