Dear All,
While, as far as I am concerned, I promised myself to put an end to the
"oligochaete" discussion, I feel necessary to add a few comments to L.
Vogt's message.
>If those two taxa are not closely related, MP will
>probably cluster them incorrectly. But if they are closely related (which
>would not be too improbable to be unrealistic, since they both share
>relatively high evolution-rates) ML will probably separate them
>incorrectly.
This is exactly what I did not observe with my data. For instance, leeches
and branchiodbellidans are long-branched taxa. Leeches cluster together
with branchiobdellidans as a sister (homogeneous) group. All the problem is
to know whether this sister relationship is due to a spurious attraction or
not. In this case, neither MP nor ML separate them.
More importantly, I know that Siddall asserted that "certain methods (ML -
my note -) are found to behave inaccurately in a portion of the parameter
space where the two-edge rate is proportionally large. This phenomenon, to
which parsimony is immune, is termed "long-branch repulsion" and the region
of poor performance is called the Farris Zone. Maximum likelihood methods
are shown to be particularly prone to failure when closely related taxa
have long branches..." (Siddall M. E., 1998. Success of Parsimony in the
Four-Taxon Case: Long-Branch Repulsion by Likelihood in the Farris Zone.
Cladistics 14: 209-220).
I think that this assertion is wrong. Siddall made computer simulations
using a four-taxon model tree in which the two-edge rate relates to
branches that are sister taxa. If you know that MP is prone to group
long-branched taxa, it is not surprising that MP performs better than ML if
it happens that these taxa are true sister groups... In fact, you are
making a comparison between a biased "method" with a theoretically unbiased
method (ML with a correct model).
You will interestingly read a reply that Felsenstein made me about this
problem I addressed him a few weeks ago:
PM:
> According to my limited understanding of the ML method, I feel that Siddall
> is wrong, at least about the "long-branch repulsion" assertion. The fact
> that the MP method performs better than the ML method when closely related
> taxa have long branches does not reflect a better efficiency, but rather
> results from a comparison of a biased, sometimes inconsistent method (MP)
> with a nearly unbiased, consistent method (ML with an adequate model).
JF: You are quite correct. The case where MP does better (one that was
actually first discovered by Ziheng Yang) is a case where the bias
happens to point in the right direction. ML is not misbehaving. In fact
you can show that if the internal branch that defines the topology is
vanishingly small (say 10^(-100) in length, MP still gets the right topology
most frequently. The reason is that it is not responding to the (vanishingly
faint) signal but is responding to its bias. There is no symmetry between
the Farris Zone and the Felsenstein Zone -- in the former case ML will be
consistent, in the latter case MP will not be.
I have been heard that Felsenstein and Swofford are working on a reply to
Siddall. Wait and see.
That's all this time !
Patrick MARTIN
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Dr. Patrick MARTIN
Institut royal des Sciences naturelles de Belgique
Biologie des Eaux douces
29, rue Vautier
B-1000 Bruxelles
Belgium
Tel : +32/2/627.43.17
Fax : +32/2/627.41.13
Email : martin at kbinirsnb.behttp://www.kbinirsnb.behttp://eea.eionet.eu.int/ec-chm/cbro/country/Belgium.htm#127
-- ANNELIDA
Discuss = annelida at net.bio.net = talk to all members
Server = biosci-server at net.bio.net = un/subscribes
Archives = http://www.bio.net/hypermail/annelida/
Resources = http://biodiversity.uno.edu/~worms/annelid.html
--
