Dear MorphMetters,
Some of you may have been in the auditorium in the Department of Botany,
University of Vienna, back in March when Philipp Mitteroecker and I were
the two scheduled discussants for the conference "GMAustria19" on
applications of geometric morphometrics. Several of the papers delivered
there used between-group principal components analysis (bgPCA), and after
each of those papers I mentioned in the course of my commentary that bgPCA
was fatally flawed in applications to most GMM data sets and should NEVER
be used here. In my keynote address, which closed the meeting, I had one
cryptic slide about this assertion, with an example that flashed on the
screen but was immediately replaced by the next slide.
The typical response to both my own talk and my criticism of the talks of
others, as far as bgPCA was concerned, was along the lines of "Hunh?" or
sometimes "What are you blathering about this time? Isn't bgPCA in the
standard toolkit?" I answered that the Bookstein paper they should read was
just then being written, as one of a pair jointly arising from
conversations with Andrea Cardini, Jim Rohlf, and Paul O'Higgins following
an original hunch of Cardini's, and that my argument would be pretty
convincing once it was actually written down. The claim isn't that people
are using bgPCA incorrectly. They're using it according to the published
formulas, yes, but the method itself yields biological nonsense much too
often.
That was March. In April, two different articles in Nature (one by Detroit
et al., one by Chen et al.) buttressed claims about sister species of Homo
sapiens using the bgPCA method, and so suddenly it became clear that we
authors had to do something quickly lest this become an epidemic of bad
biometrics. So we accelerated our writing. My paper was the first to be
finished, probably because it is a single-authored item by an emeritus with
no other obligations, and it seemed like a good idea to upload the final
draft to https://www.biorxiv.org even before submitting the paper, so that
any letter to the editors of Nature could include a link to the argument
as to exactly WHY bgPCA is nearly always unsound and its inferences invalid
for applications in contemporary GMM.
That is the draft that has just appeared as
https://www.biorxiv.org/content/10.1101/627448v1
For those of you who were at the March meeting, this is the argument
(complete with formulas) defending my stern condemnation there. I won't try
to summarize it in this morphmet note -- if you're interested, just read
the abstract on page 1 of the link. For those of you who have already
published bgPCA analyses, you know who you are -- my paper argues strongly
that you need to go back and revisit the inferences of those papers in a
mood of much more intense multivariate skepticism. For the rest of you,
please consider this draft manuscript to be a wake-up call. A technique
that has appeared in dozens of papers and that was, alas, specifically
praised by Mitteroecker and Bookstein personally (back in 2011) could
nevertheless, when examined closely (for the first time!), turn out to be
algebraic garbage when applied to data sets where there are far more shape
coordinates than specimens. But isn't that the usual situation in GMM these
days?
As always, I welcome all responses, both positive and negative. The biorxiv
posting is permanent, but there is plenty of time for me to make changes
before the paper is published (at present it has not yet even been
submitted anywhere), so feel free to try to find the flaws in my argument.
But I hope you will want to try some of these simulations on your own
before you argue against mine. You will also want to study the companion
piece by Cardini, O'Higgins, and Rohlf that should likewise be available
for download before too long.
Fred Bookstein
--
MORPHMET may be accessed via its webpage at http://www.morphometrics.org
---
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].
