Including procrustes fit for each permutation step in integration.test and modularity.test?

33 views
Skip to first unread message

Austin Ashbaugh

unread,
May 4, 2020, 12:19:39 PM5/4/20
to geomorph R package
Hello all!

Thank you so much for adding me to the discussion forum! 

First off, I want to say that I am just starting in geometric morphometrics in comparison to others here, so please tell me if I am misunderstanding some aspect of GM in my question (or future ones). 

I am currently looking to test the integration and modularity of a landmark set including 3 a priori landmark partitions using integration.test and modularity.test. However, I came across some literature that was saying to be wary of using permutation functions without adding a procrustes fit step to each permutation to prevent landmark dependance from developing through the permutation. Do the functions integration.test and modularity.test already account for this? Or is this something that I will have to add to the function myself? 

Again, thanks for letting me into the forum and I look forward to discussing more about geomorph!

Kindest regards

Austin

Adams, Dean [EEOB]

unread,
May 4, 2020, 3:47:20 PM5/4/20
to geomorph-...@googlegroups.com

Austin,

 

I’m not sure what you mean by ‘prevent landmark dependance from developing through the permutation’.  Once the shapes are aligned via GPA, they are in a common coordinate system. That does not change in the permutation step, and as such the functions you mentioned do not include an additional GPA anywhere. The input data should have been aligned via GPA first.

 

Dean

 

Dr. Dean C. Adams

Director of Graduate Education, EEB Program

Professor

Department of Ecology, Evolution, and Organismal Biology

Iowa State University

https://www.eeob.iastate.edu/faculty/adams/

phone: 515-294-3834

--
You received this message because you are subscribed to the Google Groups "geomorph R package" group.
To unsubscribe from this group and stop receiving emails from it, send an email to geomorph-r-pack...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/geomorph-r-package/6626f02d-fc7c-4164-9461-2a1cc1435eea%40googlegroups.com.

Ian Dworkin

unread,
May 4, 2020, 5:32:25 PM5/4/20
to geomorph R package
 Austin,

What paper are you thinking of? I think we might not be quite sure of what you are expressing here. There is a wide variety of methods (some with drawbacks that need to be considered) for both modularity and integration. geomorph has functionality for some of these, but for other ones, there are other R libraries (or self contained software).

Ian


To view this discussion on the web, visit https://groups.google.com/d/msgid/geomorph-r-package/CH2PR04MB661720201E2CEA98B4A68410A2A60%40CH2PR04MB6617.namprd04.prod.outlook.com.


--
Ian Dworkin
Department of Biology
McMaster University

David K

unread,
May 4, 2020, 7:32:33 PM5/4/20
to geomorph R package
Austin is talking about, e.g., Klingenberg 2009 (DOI: 10.1111/j.1525-142X.2009.00347.x). The caution mentioned there is that when using a single Procrustes fit for landmark data that consists of two or more modules, the superimposition generates its own covariance across modules (among their respective coordinates). So, for each permuted set of associations between modules, you perform a new superimposition, generating a new set of induced covariances. Not necessary when the blocks are superimposed separately because there is no induced covariance between them. 

(Correct me if I'm wrong but...) modularity.test wouldn't superimpose per permutation because the covariance ratio function shuffles assignments of landmarks (or phenotypic values) to blocks, not blocks to specimens. Integration.test... I don't remember. Does the function behave one way if all the phenotype data is in "A" and another if the phenotype data is divided between "A" and "A2"?

To view this discussion on the web, visit https://groups.google.com/d/msgid/geomorph-r-package/CAGudrjk_yKN6PHi3FSii01XH4fRF%2B0%3D4RC6JGv%3DyJejPBGpsEA%40mail.gmail.com.


--
David C. Katz, Ph.D.
Postdoctoral Fellow
Benedikt Hallgrimsson Lab
University of Calgary

Research Associate
Department of Anthropology
University of California, Davis

Adams, Dean [EEOB]

unread,
May 4, 2020, 7:37:01 PM5/4/20
to geomorph-...@googlegroups.com
Yes that was stated. Never empirically demonstrated by simulation. There are still real covariances among landmarks between modules.

Dean

From: geomorph-...@googlegroups.com <geomorph-...@googlegroups.com> on behalf of David K <katz...@gmail.com>
Sent: Monday, May 4, 2020 6:32:05 PM

To: geomorph R package <geomorph-...@googlegroups.com>
Subject: Re: [geomorph-r-package] Including procrustes fit for each permutation step in integration.test and modularity.test?
 

Austin Ashbaugh

unread,
May 4, 2020, 8:31:17 PM5/4/20
to geomorph R package
Hi all, 

Ah okay, I was pretty sure that was the case with modularity.test but I wasn't sure with integration.test. I was concerned because of the argument mentioned in Klingenberg 2009 using simultaneous procrustes fit before permutation, as mentioned by Dr. Katz. 

Thank you all for your rapid response.

Austin

Adams, Dean [EEOB]

unread,
May 4, 2020, 9:17:22 PM5/4/20
to geomorph-...@googlegroups.com

Folks,

 

Just a point of clarification, as this topic comes up rather frequently.

 

Yes, such a claim was stated in 2009: that separate vs. simultaneous GPA could make a difference. And yes, a single dataset example showed different point-estimates for the RV coefficient, as shown in that paper. But a single empirical example does not analytical theory make.

First, no mathematical proof (i.e., analytically-derived mathematical theory) was proposed to describe what one should do in such cases, and which procedure was preferable to the other. Second, there was no demonstration via simulation to show either the problem or the preferred solution. Thus, the field has languished in this ‘recommendation’ of either/or separate-combined GPA for 10 years without any real theoretical solution. Until proper mathematical or simulation-based evidence is presented, one should follow the simpler course, and perform a single GPA and then perform modularity/integration analyses (that is, if those structures are really sub-structures of a single configuration).

 

On a related topic there is a new paper just accepted in Evolutionary Biology (Collyer, Davis, Adams) that finally explores the related topic of ‘combining’ landmark datasets across configurations with respect to how one scales them for size. The connection is that here too, a single empirical example ‘proved’ that one approach to scaling might be useful, whereas a full simulation-based investigation showed conclusively that this was not at all the case. Now, whether or not the same may be said of the resulting shape variables for integration and modularity is premature, but I suspect the ‘combine or not combine for GPA’ question will fall in line with the results of the paper above. Time (and importantly simulations) will tell.

 

Regardless, I urge empiricists not to simply take the word of a publication on how to perform morphometric analyses without the demonstration by simulation or mathematical proof of what that pipeline generates and provides. Math, analytics, and simulations, will ultimately dictate our decisions in this regard.

 

Dean

 

Dr. Dean C. Adams

Director of Graduate Education, EEB Program

Professor

Department of Ecology, Evolution, and Organismal Biology

Iowa State University

https://www.eeob.iastate.edu/faculty/adams/

phone: 515-294-3834

 


Sent: Monday, May 4, 2020 7:31 PM
To: geomorph R package <geomorph-...@googlegroups.com>

--

You received this message because you are subscribed to the Google Groups "geomorph R package" group.
To unsubscribe from this group and stop receiving emails from it, send an email to geomorph-r-pack...@googlegroups.com.

Ariella Rink

unread,
May 5, 2020, 2:58:46 AM5/5/20
to geomorph-...@googlegroups.com
Dear Dean and Michael

I'm very interested in the findings of your newly accepted paper: I have been concerned with whether or not to scale by centroid size the separate configurations I have for the different anatomical components of an articulated structure (femur, tibia, tarsus, tarsal claw, of beetle legs). 

What approach did the simulations support?

Kind regards
Ariella Rink


Reply all
Reply to author
Forward
0 new messages