Hi Seth,

Good to hear from you!

In and of itself, I do not think that PCA is going to tell you too much about 
how the shapes changed. The exception might be if the shape trajectory (i.e., 
temporally ordered set of shapes) closely aligns with the first PC. In this 
instance, you could argue that temporal changes are the shapes' primary source 
of variation.

However, I think what your questions are really getting at are whether we can 
represent shapes in a trajectory using a few PC scores without losing relevant 
information. We always lose some information when we reduce the dimensions on a 
sample space, be that a tangent projection off a shape space or not. The answer 
to whether the lost information matters will depend on what you want to know 
about the trajectory and on the specifics of your trajectory.

In my dissertation work, I used vector autoregressive models to represent shape 
trajectories. The goal was to represent the trajectory in a manner that 
preserved information about the autocorrelation between shapes. I did use PC 
scores to represent the shapes in the trajectory, and I had some success with 
this strategy.

However, the econometrics literature has looked at a similar problem (i.e., 
dimension reduction techniques for multivariate time series that preserve the 
autocorrelative structure) and has concluded that PCA is suboptimal for this 
purpose. Instead, a second technique, called dynamic factor analysis, usually 
performs better. You can see the method description here:

Engle, R and Watson, M. 1981. A one-factor multivariate time series model of 
metropolitan wage rates. Journal of the American Statistical Association 76: 
774-781.

I have used dynamic factor analysis to reduce the dimensionality of shape 
trajectories, and I did get some favorable results in a pilot study. I wrote a 
poster abstract on the topic:

Soda KJ. 2020. Dynamic factor analysis as a dimension reduction technique for 
shape trajectory data. American Journal of Physical Anthropology 171: 269.

However, the pandemic caused a shift in my research agenda, and I had to 
sideline the project. As such, I cannot say much more than that. One point to 
keep in mind, though: Dynamic factor analysis is pretty data intensive, so you 
would need a fairly long trajectory to make it work; motion capture data would 
be ideal.

If the autocorrelative structure is not of much interest to you, though, then 
PCA may still work. Really, all PCA does is rotate the axes on your tangent 
space. If you GPA aligned the shapes in the trajectory, they arguably are 
already in the same tangent space, so I cannot see any harm in implementing a 
PCA on this space. It's more of a question of what you want to do with the data 
afterwards.

Hope something in this excited rant is of use to you; I just love shape 
trajectories!

K. James Soda (He/Him/His)
##############################################
Assistant Professor of Mathematics and Data Science
Department of Mathematics and Statistics
Quinnipiac University
Hamden, CT
##############################################
________________________________
From: [email protected] <[email protected]> on behalf of Seth 
Boren <[email protected]>
Sent: Wednesday, April 12, 2023 8:43 AM
To: Morphmet <[email protected]>
Subject: [MORPHMET2] Question about longitudinal change over time and PCA

CAUTION: This email originated from outside of the organization. Do not follow 
guidance, click links, or open attachments unless you recognize the sender and 
know the content is safe.

Morphomet mailing list,

I am doing a longitudinal analysis of shape change (how shape of changes over 
time) and I have some questions.

Is a PCA a good option for analyzing shape change over time? All the "shapes" 
in my sample have 30 landmarks a piece and are aligned via GPA. Would I still 
be able to adequately analyze the how shape is changing over time with reduced 
dimensionality? Can PCs be directly comparable between time points?

As I understand it, PCA performed within each time point can reduce the 
dimensionality of the data for each time point independently, but does not 
guarantee that the underlying data would be captured in each PC.

Thank your for your time and consideration,

Seth Boren

--
You received this message because you are subscribed to the Google Groups 
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to 
[email protected]<mailto:[email protected]>.
To view this discussion on the web visit 
https://groups.google.com/d/msgid/morphmet2/2c624bf9-772a-4346-b541-a08c99fcc901n%40googlegroups.com<https://groups.google.com/d/msgid/morphmet2/2c624bf9-772a-4346-b541-a08c99fcc901n%40googlegroups.com?utm_medium=email&utm_source=footer>.

-- 
You received this message because you are subscribed to the Google Groups 
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/morphmet2/SA2PR10MB4506E7467A054BDA40112ED5FF989%40SA2PR10MB4506.namprd10.prod.outlook.com.

Reply via email to