[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Subject Index][Author Index]
Re: [dinosaur] Dinosauria reclassification joins Ornithischia and Theropoda in Ornithoscelida
Gesendet: Freitag, 24. März 2017 um 10:40 Uhr
Von: "David Černý" <email@example.com>
> Mickey Mortimer
> <firstname.lastname@example.org[mailto:email@example.com]> wrote:
> > I'm not a phylogenetics expert when it comes to the algorithms used and
> >such, but I believe what Marjanovic meant was that a morphological dataset
> >like this one that tries to maximize taxonomic coverage will have a lot of
> >highly incomplete taxa.
> David M. is obviously the best-qualified person to clarify what he meant, but
> it didn't occur to me to read his comment on the difference between
> morphology and molecules as a statement about incompleteness. Sparse
> supermatrices with wildly varying levels of coverage are common in densely
> sampled molecular analyses as well; it's not a problem unique to fossil
I'm sorry I wasn't clearer. I mixed up several things:
First, what Mickey said. Even more generally than that, wildcard taxa – often
more incomplete than usual, but the correlation is weaker than most people seem
to assume – severely depress bootstrap values. See the tree figures in my
preprint with Michel Laurin*
for plenty of examples; note that the bootstrap trees are separate figures.
Second, character samples in morphological matrices are small. (Most or all of
them could be bigger, often much bigger, but getting them there is a lot more
work than for molecular data.) This exacerbates the above problem _and_
depresses synapomorphy counts _and_ can cause accidental sampling bias. In the
preprint I ran analyses with 102 and 150 OTUs but the same character sample and
watched almost all bootstrap values plummet from the former to the latter;
compare its bootstrap tree figures.
Third, yes, morphological analyses are generally quite a bit less reliable than
molecular ones today; this could change as morphological matrices increase in
size and improve in quality (as character correlation becomes better understood
and as typos and similar unsystematic mistakes are eliminated), but we're not
close. A bootstrap value of 66% on something so unexpected is, relatively
speaking, a cause to celebrate.
Fourth, I conflated these issues with the comparison of bootstrap values and
Bayesian posterior probabilities... thanks for the reference, I only knew its
conclusion as a rumor. :-)
* First revision of a manuscript submitted to PeerJ. The second round of
reviews arrived a while ago, and I've almost finished the second revision.