[This follows up on discussion in bionet.software but we want to move it to
bionet.molbio.evolution. I have posted a note to bionet.software that
this reply is to be found here. My personal e-mail address is incorrectly
generated if you Reply to this, so in that case use the one at the bottom of
the posting.]
In article <Cquwxy.FwK at zoo.toronto.edu> mes at zoo.toronto.edu (Mark Siddall) writes:
>In article <schwarze-030694194428 at fennel.bio.caltech.edu> schwarze at starbase1.caltech.edu (Erich Schwarz) writes:
>>>> I was thinking of situations where there is in fact one single most
>>parsimonious tree. As I understand the debate, one either can consider
>>that tree to have a probability of effectively 100% (which I think is the
>>cladistics position) or try to assign it some other probability (which I
>>understand to be Dr. Felsenstein's position.)
>>[Mark Sidall replies:]
>Cladists do not assign probabilities to their most parsimonious tree(s)
>except that they contend that by being most parsimonious, they are more
>likely than those that are less parsimonious.
>The most ardent of us would suggest that ALL equally parsimonious tree
>are equally well supported.
...
>I have yet to hear a compelling argument for something with a worse fit
>to the data having a better p-value than something with a better fit but
>then I don't claim to fully (or perhaps even partially) understand
>maximum liklihood.
"Cladists" (phylogenetic systematists of the Willi Hennig Society persuasion)
usually say you should look only at the most parsimonious tree or trees. But
they acknowledge that these don't have a 100% probability of including the
true tree. I can't quite put this together unless they believe that
a statistical approach would be valid, but that existing ones are not, so one
should avoid looking at the confidence intervals they suggest. However I
don't hear that from cladists, but rather a complete rejection of the framework
of statistics instead. Perhaps I miscontstrue.
If one is using the trees for some secondary analysis such as looking at
host-parasite coevolution, and one concentrates only on most parsimonious
trees, it would seem that if a statistical framework is allowed even in
principle, then one is effectively assuming a 100% probability for the
set of most parsimonious trees.
I think all statistical types agree that the best-fitting trees are those that
are most probable. I don't know anyone who argues for preferring a less
well-fitting tree. But that may or may not be the same thing as a less
parsimonious tree, as parsimony may well not be the best measure of goodness
of fit.
[This is a somewhat separate issue, so I have extracted and isolated it.]
>We start getting into grey area here when one considers the criterion of
>choice by cladists to be more like goodness of fit, and, unless I am
>mistaken, the criterion of maxlik (Felsenstein) to be based more on
>confidence than g-o-f.
Certainly for discrete data such as molecular sequences, one can in fact show
that the maximum likelihood tree is also the best g-o-f tree, using the
G measure of goodness of fit which is a logarithmic analog to chi-square.
So I don't think there is that distinction here, or if it is it works the
other way.
If one believes that degree of parsimony is by definition always the best
measure of goodness of fit, then of course ML would not always be g-o-f,
but I would argue against that way of defining g-o-f.
I'm delighted to find some discussion of this -- as far as I can see
cladists are currently extremely reluctant to openly discuss the logical
foundations of their approach. See my book review in Cladistics, volume
8 pages 191-196, 1993 and the lack so far of the discussion it pointedly
calls for. Mark is to commended for rising to the occasion.
-----
Joe Felsenstein, Dept. of Genetics, Univ. of Washington, Seattle, WA 98195
Internet: joe at genetics.washington.edu (IP No. 128.95.12.41)