Re: [MORPHMET] Doubt Scaling photos

[Anderson Feijo]

Dear Murat and Carmelo,

Thank you for your suggestions and important comments.  I will add other
specimens from the same locality and explore the variation at different
distances to assess how meaningful is the effect.

All the Best,

Anderson

On Thu, Jan 4, 2018 at 6:30 PM, Carmelo Fruciano 
wrote:

> Hi Anderson,
>
> I concur with Murat about the usefulness of having actually multiple
> specimens. Analyses on a single specimen, although in principle attractive
> (as one "isolates" error from biological variation), are not necessarily
> informative on a practical level. Also there are two problems, one of
> magnitude (e.g., how large is variation due to distance/presentation
> relative to biological variation), one of direction (e.g., if one
> distance/presentation gets consistently digitised differently from the
> other, that is, you have non-random error/bias). You can find discussions
> of these in my recent review which Murat very kindly suggested.
>
>
> I would add that, in addition to the digitization process, differences
> might have been caused by factors. However, another thing that I noticed in
> your TPS file is that you have first all the repetitions at one distance,
> then all the others. So I assume you digitized them in that order. This
> suggests that, as already pointed out by Murat, the differences might be
> due to "learning" or changes over time in the way you digitize.
>
>
> Most likely, the final error will come from a combination of various
> sources, as it's generally the case.
>
>
> In general, I would suggest you to run a larger preliminary analysis with:
>
> - multiple biological specimens
>
> - multiple distances
>
> - multiple digitizations
>
> - digitizing in random order across repetitions (there is a very useful
> function for this in tpsUtil by F.J. Rohlf)
>
> - using multiple statistical tools to gauge how large is the error
> relative to biological variation and if there are non-random patterns
>
>
> I hope this helps,
>
> Carmelo
>
>
>
> Il 2/01/2018 7:28 PM, Murat Maga ha scritto:
>
> Hi Anderson,
>
>
>
> I don’t think PCA is the tool you want to use in this context. Procrustes
> anova already tells you that your two magnification levels are done
> differently. Since your question is ‘can I combine data from different
> magnification level in the same analysis?’, the answer to that in your
> current study design (single sample measured five times at two different
> scales) is a very qualified no.
>
>
>
> But, what is more relevant, (or what I would have liked to know) the
> magnitude of this error in context of actual biological variation.
>
> For that, you have to have a study design where you have a few samples of
> different sizes measured at different magnification scales a couple times
> (not just measuring the same object repeatedly in different
> magnifications), or something along those lines.
>
>
>
> There is a really a lot of literature on this.  But you need to use a
> statistics based framework, not EDA tools like PCA to answer this.
>
>
>
> M
>
>
>
>
>
> *From:* Anderson Feijo [mailto:andefe...@gmail.com ]
> *Sent:* Monday, January 1, 2018 11:02 PM
> *To:* Murat Maga  
> *Cc:* MORPHMET  
> *Subject:* Re: [MORPHMET] Doubt Scaling photos
>
>
>
> Dear Murat,
>
>
>
> Thank you very much for your email and time for explore my dataset. To
> perform this simple experiment, given the aim was to explore the
> magnification issue, I chose only three landmarks (as you could see in the
> tps file) easily to replicate. The reason I am trying to combine two
> magnification is that I will work with groups with different sizes; for
> example, species with 10 cm and others with 3 cm of total length of skull.
> Place the camera at a good distance to avoid parallax for the large group
> will lead a low resolution images of the small groups. Thus, my initial
> idea was to have two standardized distances and combine later after
> accounting the scale effects.
>
>
>
> In my short experiment, what intrigues me is the clear groups in PCA (and
> other exploratory analyses) based on the two magnifications (see below). At
> the beginning, I was expecting a great overlap between the two distances
> and minimum differences due to landmarks-placement error as seen within
> each of the group.
>
>
>
> Best,
>
> Anderson
>
>
>
> [image: Inline image 1]
>
>
>
>
>
> On Sat, Dec 30, 2017 at 3:57 PM, Murat Maga  wrote:
>
> Dear Anderson,
>
>
>
> It has nothing to do with the scale of your data (or least not directly).
>
>
>
> As you can see, you actually performed those two digitization attempts
> differently. It is maybe 1-2 pixels off, but in a very consistent way
> (greens are from one scale, red is the other scale, and cross is the
> consensus shape). This type of systematic error commonly occur in
> digitization process, as one learns better or 

Re: [MORPHMET] Doubt Scaling photos

[Carmelo Fruciano]

Hi Anderson,

I concur with Murat about the usefulness of having actually multiple 
specimens. Analyses on a single specimen, although in principle 
attractive (as one "isolates" error from biological variation), are not 
necessarily informative on a practical level. Also there are two 
problems, one of magnitude (e.g., how large is variation due to 
distance/presentation relative to biological variation), one of 
direction (e.g., if one distance/presentation gets consistently 
digitised differently from the other, that is, you have non-random 
error/bias). You can find discussions of these in my recent review which 
Murat very kindly suggested.



I would add that, in addition to the digitization process, differences 
might have been caused by factors. However, another thing that I noticed 
in your TPS file is that you have first all the repetitions at one 
distance, then all the others. So I assume you digitized them in that 
order. This suggests that, as already pointed out by Murat, the 
differences might be due to "learning" or changes over time in the way 
you digitize.



Most likely, the final error will come from a combination of various 
sources, as it's generally the case.



In general, I would suggest you to run a larger preliminary analysis with:

- multiple biological specimens

- multiple distances

- multiple digitizations

- digitizing in random order across repetitions (there is a very useful 
function for this in tpsUtil by F.J. Rohlf)


- using multiple statistical tools to gauge how large is the error 
relative to biological variation and if there are non-random patterns



I hope this helps,

Carmelo




Il 2/01/2018 7:28 PM, Murat Maga ha scritto:


Hi Anderson,

I don’t think PCA is the tool you want to use in this context. 
Procrustes anova already tells you that your two magnification levels 
are done differently. Since your question is ‘can I combine data from 
different magnification level in the same analysis?’, the answer to 
that in your current study design (single sample measured five times 
at two different scales) is a very qualified no.


But, what is more relevant, (or what I would have liked to know) the 
magnitude of this error in context of actual biological variation.


For that, you have to have a study design where you have a few samples 
of different sizes measured at different magnification scales a couple 
times (not just measuring the same object repeatedly in different 
magnifications), or something along those lines.


There is a really a lot of literature on this.  But you need to use a 
statistics based framework, not EDA tools like PCA to answer this.


M

*From:* Anderson Feijo [mailto:andefe...@gmail.com]
*Sent:* Monday, January 1, 2018 11:02 PM
*To:* Murat Maga 
*Cc:* MORPHMET 
*Subject:* Re: [MORPHMET] Doubt Scaling photos

Dear Murat,

Thank you very much for your email and time for explore my dataset. To 
perform this simple experiment, given the aim was to explore the 
magnification issue, I chose only three landmarks (as you could see in 
the tps file) easily to replicate. The reason I am trying to combine 
two magnification is that I will work with groups with different 
sizes; for example, species with 10 cm and others with 3 cm of total 
length of skull. Place the camera at a good distance to avoid parallax 
for the large group will lead a low resolution images of the small 
groups. Thus, my initial idea was to have two standardized distances 
and combine later after accounting the scale effects.


In my short experiment, what intrigues me is the clear groups in PCA 
(and other exploratory analyses) based on the two magnifications (see 
below). At the beginning, I was expecting a great overlap between the 
two distances and minimum differences due to landmarks-placement error 
as seen within each of the group.


Best,

Anderson

Inline image 1

On Sat, Dec 30, 2017 at 3:57 PM, Murat Maga > wrote:


Dear Anderson,

It has nothing to do with the scale of your data (or least not
directly).

As you can see, you actually performed those two digitization
attempts differently. It is maybe 1-2 pixels off, but in a very
consistent way (greens are from one scale, red is the other scale,
and cross is the consensus shape). This type of systematic error
commonly occur in digitization process, as one learns better or
alters the way landmarks are digitized/defined along the process.
In your case, perhaps higher resolution image gave a better
definition (imaging sense) of landmarks, so you captured them very
slightly differently than low-res image (but in a very consistent
way).

If you are concerned, the first step perhaps is to do the same
thing (i.e. digitization of same sample with different
magnifications several times) with several real specimens and get
a sense of the magnitude of this error in context of the