-------- Original Message --------
Subject: Re: MorphoJ - PLS: Blocks within a single configuration --
bugs and MorphoJ update
Date: Tue, 27 Sep 2011 14:48:20 -0400
From: Chris Klingenberg <[email protected]>
Reply-To: [email protected]
Organization: University of Manchester
To: [email protected]
Dear Jana
Thank you very much for pointing out those things. They are two bugs in
the program (ooops, sorry!).
One bug, for PLS within a configuration of landmarks, resulted in
erroneous data in the Procrustes re-fit for the permutation test (all
statistics: RV coefficient, singular values, correlations of PLS
scores). The most obvious consequence was on the singular values, as
Jana pointed out.
The other bug was in the two-block PLS for separate blocks, and
concerned only the P-values for the correlations between PLS scores. The
P-values were correct for the RV coefficient and singular values, but
not for the correlations. I was aware of this for some time but
compounded things by mistakenly "explaining away" the problem (ooops,
sorry, again!).
Both bugs are now fixed and I have uploaded the new version 1.03d to the
MorphoJ web page today:
http://www.flywings.org.uk/MorphoJ_page.htm
Thanks to Jana and the participants of the workshop in Sabadell, a
couple of weeks ago, for spotting these and a few other bugs.
***
When using and comparing the P-values for the singular values and the
correlations between PLS scores, a good deal of caution is still
advisable. It is not entirely straightforward to decide what null
hypothesis to use, particularly for the correlation between PLS scores.
MorphoJ consistently uses the null hypothesis that there the two blocks
are independent (for PLS within configurations: that they are
independent except for the effects of the joint Procrustes fit). With
the permutation test, there can be some ambiguities over the identities
of PLS axes that are associated with singular values of similar
magnitudes (this is analogous to problems with resampling principal
components when two or more eigenvalues are not very different).
The test for the overall association, along with the RV coefficient that
indicates the strength of this association, is relatively simple to
interpret. When it comes to the singular values and correlations one by
one, I'd be cautious. For the choice of how many PLS axes you want to
consider in an analysis, the magnitudes of the singular values may be
more informative than their statistical significance.
I hope this helps.
Best wishes,
Chris
On 9/26/2011 6:23 PM, morphmet wrote:
-------- Original Message --------
Subject: MorphoJ - PLS: Blocks within a single configuration
Date: Sun, 25 Sep 2011 19:26:24 -0400
From: Jana Makedonska<[email protected]>
To: [email protected]
Dear All,
I am a relatively new MorphoJ user studying craniofacial integration. I
used to apply a 2-block PLS approach to the study of the architecture of
the cranium, yet I have realized that the relative orientation of units
belonging to a functional whole should definitely be taken into account,
and thankfully, Dr. Klingenberg has implemented such a test in MorphoJ.
I have a question regarding the permutation tests of the singular axes
in a PLS analysis within a single configuration. I am providing an
example below:
Overall strength of association between blocks:
RV coefficient: 0.4261
Permutation test against the null hypothesis of independence
Number of randomization rounds: 10000
P-value:<.0001
Singular values and pairwise correlations of PLS scores between blocks:
Singular value P-value (perm.) % total covar.
Correlation P-value (perm.)
PLS1 0.00018668 1.0000 61.637 0.92015
<.0001
PLS2 0.00012496 1.0000 27.618 0.92618
<.0001
PLS3 0.00005060 1.0000 4.528
0.45438 0.8520
PLS4 0.00004219 1.0000 3.148
0.40476 0.8485
PLS5 0.00002855 1.0000 1.442
0.49857 0.1652
PLS6 0.00002183 1.0000 0.843
0.35042 0.6877
PLS7 0.00001383 1.0000 0.338
0.29619 0.7267
PLS8 0.00001321 1.0000 0.309
0.30962 0.3313
PLS9 0.00000881 1.0000 0.137
0.15853 0.8138
My question is: Why is the P-value of the singular axes always equal to 1?
Also, why are the p-values of the singular axes and the correlations
numerically identical when a 2-block PLS is run? (example pasted below)
I do have the impression that when a 2-block PLS is run in one of the
softwares of the IMP package, these p-values differed.
Overall strength of association between blocks:
RV coefficient: 0.1233
Permutation test against the null hypothesis of independence
Number of randomization rounds: 10000
P-value: 0.0148
Singular values and pairwise correlations of PLS scores between blocks:
Singular value P-value (perm.) % total covar.
Correlation P-value (perm.)
PLS1 0.00012760 0.0556 43.048
0.55104 0.0556
PLS2 0.00009617 0.0373 24.452
0.46313 0.0373
PLS3 0.00006845 0.1563 12.389
0.48423 0.1563
PLS4 0.00005697 0.0741 8.581
0.36577 0.0741
PLS5 0.00003937 0.3700 4.098
0.28668 0.3700
PLS6 0.00003259 0.2263 2.809
0.28990 0.2263
PLS7 0.00002723 0.1072 1.960
0.29277 0.1072
PLS8 0.00002195 0.0608 1.274
0.25144 0.0608
PLS9 0.00001644 0.0841 0.715
0.26427 0.0841
PLS10 0.00001275 0.0433 0.430
0.19014 0.0433
PLS11 0.00000960 0.0195 0.244
0.16680 0.0195
Any clue will be highly appreciated!
Thanks!
Jana
([email protected]<mailto:[email protected]>)
--
***************************************************************
Christian Peter Klingenberg
Faculty of Life Sciences
The University of Manchester
Michael Smith Building
Oxford Road
Manchester M13 9PT
United Kingdom
Telephone: +44 161 275 3899
Fax: +44 161 275 5082
E-mail: [email protected]
Web: http://www.flywings.org.uk
Skype: chris_klingenberg
***************************************************************
