[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Subject Index][Author Index]
Re: SICB Report
> In a message dated 1/19/99 7:58:09 AM EST, firstname.lastname@example.org writes:
> << Some of George's statements
> about being 18 percent wrong or something are not entirely correct - >>
> The 18% wrong ratio comes from the "stratocladistics" study presented at the
> SVP meeting in Chicago in 1997 (last SVP I was at). The authors (don't have
> their names handy, but their paper is abstracted in the JVP supplement for
> that year) claimed that incorporating stratigraphic information improved this
> ratio to about 10% wrong. At least, that's what I gathered from the talk, and
> it didn't seem unreasonable to me.
I didn't see that talk (unfortunately), and eagerly await its
publication in peer-reviewed media. But whether the results are
reasonable or unreasonable will depend very, very heavily on many
things: the taxon to character ratio, the initial assumptions of
character evolution, how they simulated character incompleteness, and
how they resolved stratigraphic sampling issues. This will also impinge
on how widely we can apply their simulations to real data - it might be
of general importance, or it might not.
If you want an example of how the initial assumptions and subsequent
protocols can color the ultimate results of a simulation, have a look at
Lamboy, W.F. 1994. The accuracy of maximum parsimony method for
phylogeny reconstruction with morphological characters. Systematic
Wiens, J.J., and D. Hillis. 1996. Accuracy of parsimony analysis using
morphological data: a reappraisal. Systematic Botany, 21:237-253.
(This also includes a reply from Lamboy.)
Basically, Lamboy designed a set of simulations that, in his eyes,
painted a very dismal portrait of morphology-based parsimony analysis.
In fact, his simulations were very poorly designed, and his methods for
comparing his results with real data sets were based on very
poorly-chosen metrics. Wiens and Hillis' review is important for
highlighting how a simulation can be set up to yield bogus results.
(Lamboy's reply is worth reading, even though he clearly misunderstood
what Wiens and Hillis really said.)
One thing I will stand by - we cannot fix any specific percentage to
general phylogenetic accuracy, because no two sets of problems will be
the same. 18% inaccuracy (based on number of incorrect nodes) may be
the best we can do for some groups, far better for others, and not even
close to the best we can do for others. That's why I cited the
Huelsenbeck 1995 paper - it addresses a whole suite of scenarios,
applicable to morphology (even though couched in molecular terms).