bilat.symmetry without replicates question

[Ashly Romero]

Hello all,

I'm trying to use the bilat.symmetry function on a small sample of my data to do a preliminary test, but I want to remove replicates from the model because I'm using landmarks that were placed automatically using the SlicerMorph ALPACA package. When I run the bilat.symmetry function without replicates, I don't see a p-value for the ind:side interaction (FA) in the Procrustes ANOVA output. Output here:

Call:
bilat.symmetry(A = LMs, ind = Classifiers.batch$Individual, object.sym = TRUE,  
    land.pairs = LM.pairs.batch)

Symmetry (data) type: Object

Type I (Sequential) Sums of Squares and Cross-products
Randomized Residual Permutation Procedure Used
1000 Permutations

Shape ANOVA
         Df      SS      MS     Rsq        F      Z Pr(>F)  
ind      24 0.05579 0.00232 0.11156   4.3707 4.0962  0.001 **
side      1 0.43151 0.43151 0.86292 811.4024 3.9471  0.001 **
ind:side 24 0.01276 0.00053 0.02552                          
Total    49 0.50005                                          
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> 

Is there something I can adjust in my code to get a p-value for this interaction effect? Or maybe the way this function is written, it only works with replicates? Any help is greatly appreciated. I've attached files with my LMs, LM pairs, and classifiers in case this is helpful.

All the best,

Ashly 

---

Ashly Romero

PhD Candidate

Biological Anthropology

University of Arkansas

[Adams, Dean [EEOB]]

Ashly,

 

By definition, there is no test of significance of the interaction of the bilateral symmetry model without the replicates. Put more statistically, the interaction term is tested against replicate variation (see basic theory papers for bilateral symmetry designs for more detail).

Dean

 

Dr. Dean C. Adams (he/him)

Distinguished Professor of Evolutionary Biology

Director of Graduate Education, EEB Program

Department of Ecology, Evolution, and Organismal Biology

Iowa State University

https://faculty.sites.iastate.edu/dcadams/

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/5ef072f6-d513-4550-a372-4e34cd377d82n%40googlegroups.com.

[Ashly Romero]

Thank you for the reply! So this means that you can't use automatically placed landmarks in the Procrustes ANOVA model, correct? When I place them a second time they're in exactly the same position (of course), so the replicate seems basically useless. 

[Adams, Dean [EEOB]]

Correct. You must digitize individuals multiple times to have variation in the ‘replicate’ term, if one wishes to statistically interrogate the ind:side interaction term. A careful reading of the original theory papers will explain why.

 

What that likely implies is that IF automatic landmark placement procedures do so without error (or with consistent error for the same object), these cannot be used if one wishes to investigate fluctuating asymmetry.

 

Perhaps this fact has not been considered properly by those who have provided the automatic landmark placement software, or at least they have not discussed this issue with their users.

To view this discussion on the web, visit https://groups.google.com/d/msgid/geomorph-r-package/783101ca-c4e4-4c5d-8f24-02649c5d627fn%40googlegroups.com.

[mura...@gmail.com]

I would like to chime a few things on this line.

Most automated methods can replicate things, if you change your starting parameters. However, this more of replication of the pipeline, not just the final step of digitization, which bring my second point: If the goal to replication, I would argue that it is important to replicate the whole process, not putting points on digital pictures. The way you mount your specimen under the microscope will generate variation, or you derive your 3D mesh will cause variation. Thus, replicating only the landmarking part of it, is not a true replication in this sense. I suspect if you take two pictures of a specimen under the same photo stand/microscope back to back, import into R and sum the absolute value of the difference of pixel values, you will not get zero, meaning there is variability even at the sensor level. If you replicate only the final step of the pipeline (which is putting points), you are bound to get much less variation compared something like whole pipeline replication. So not all replicates are equal. 

Also, I think one can still calculate FA scores without replications, at least that's what the ?bilat.symmetry snippet shows for the example for the scallops 3D data. Of course, one will not be able to assess whether that small L/R variance is indeed more or less than the measurement error or not.  

[Adams, Dean [EEOB]]

Yes one can calculate FA scores without replication. However, as I stated and you repeated, one cannot statistically evaluate the degree to which FA varies relative to measurement error.

Dean

 

Dr. Dean C. Adams (he/him)

Distinguished Professor of Evolutionary Biology

Director of Graduate Education, EEB Program

Department of Ecology, Evolution, and Organismal Biology

Iowa State University

https://faculty.sites.iastate.edu/dcadams/

phone: 515-294-3834

 

To view this discussion on the web, visit https://groups.google.com/d/msgid/geomorph-r-package/1921f1e4-cf28-4742-997c-9a8e870ec494n%40googlegroups.com.

[mura...@gmail.com]

Isn't that true for any procrustes anova methodology?

[Adams, Dean [EEOB]]

Nope. As with all ANOVA-based linear models, how model effects are statistically evaluated depends entirely upon the error structure specified for the various terms of the model (note: both geomorph and RRPP allow for different error structures to be specified, as required by the statistical design).

 

For example, with fixed effects factorial models, all model terms (including the interaction) are tested against the total error term. For models containing random or nested effects, error terms for each model term can vary, depending upon the structure.

 

For symmetry analyses, some terms are of the nested/random flavor (e.g., the interaction is tested against the replicate term). This is why I referred the prior user to the original theory papers, as this is described more fully there (and why such terms are tested in what manner).

 

In general, one should consult a statistical model design text to ensure that they are performing ANOVA (or Procrustes anova) properly.

To view this discussion on the web, visit https://groups.google.com/d/msgid/geomorph-r-package/52beed3b-0a38-412a-b517-9fa81aebd799n%40googlegroups.com.