Re: [dinosaur] Phylogenetics in general was Re: Placental Mammal Diversification Across the K-Pg Boundary (free pdf)

 El dom., 1 de dic. de 2019 a la(s) 15:23, David Marjanovic (
[email protected]) escribió:

That said, I wonder how expensive it really is. It costs a lot more in
> person-hours; but the work itself is much cheaper: no expensive machines,
> no expensive chemicals, no unbelievably expensive electrophoresis gel every
> week, no ultrafreezers, not even liquid nitrogen. Just pay someone's costs
> of living and a few travels, and the work will get done.
>

That's only true for some types of data collection; high-throughput CT
scanning, for example, is not exactly cheap, and judging from a couple of
recent papers, the next big thing in morphological data collection may well
be neutron tomography, where that holds to an even greater extent.

On the other hand, calling sequencing supplies "unbelievably expensive"
does a disservice to just how insanely cheap sequencing has gotten over the
last 20 years. That whole thing about how it cost $3 billion to sequence a
whole human genome back in 2003 (when the HGP was completed) but about $200
now has become a cliché, but the point is undeniable. The amount of data
stored in GenBank has been growing at a superexponential rate, which is
simply not true for morpho phylogenetic datasets, where even a well-studied
group like the tetraodontiforms can have over a decade's worth of studies
recycling the same ~200-character matrix.


> Does anybody understand what this "ascertainment bias" is, and if it
> applies to anything but maximum likelihood?
>

It applies to all model-based phylogenetics (ML, Bayesian inference, and
distance methods insofar as they use models to correct the raw, observed
distances) and it refers to a situation where constant characters –
characters for which every taxon has the same state – are not included in
the dataset for some reason, usually because the very nature of the
characters makes the notion ill-defined. Model-based analyses need constant
characters to estimate branch lengths (in non-clock analyses, which
confound rates and times and treat their products as i.i.d. variables) or
branch rates; if they don't get them, these parameters will be
overestimated, and the bias might then propagate to other parameters of
interest (topology, divergence times).

This is also known as "acquisition bias", and it first came up for
restriction sites, where it was dealt with by Felsenstein (1992) using
conditional likelihoods. Of course, it's also inherent in analyzing
discrete morphological data: even if people were in principle willing to
score constant characters for their datasets, there would be no way of
knowing when they scored the right number of them, which is an issue that
just doesn't arise for sequence data. Therefore, starting with Lewis's
(2001) paper, most attempts to estimate phylogenies from discrete
morphological characters in a model-based framework have relied on
corrections implemented using Felsenstein's method. RAxML, MrBayes, BEAST
2, and RevBayes all do it that way, sometimes with the extra option of
specifying whether to condition on variable characters or on
parsimony-informative characters only.

*Refs:*

Felsenstein J 1992 Phylogenies from restriction sites: a maximum-likelihood
approach. Evolution 46(1): 159–173

Lewis PO 2001 A likelihood approach to estimating phylogeny from discrete
morphological character data. Syst. Biol. 50(6): 913–25


Best,
-- 
David Černý