apologies for not continuing this thread immediately, I've been doing
the proverbial chicken without a head thing recently...
...the above also pertains to Eric's concerns regarding lack of
follow-up... (in article 1759 of this group)...
In article <2sri9k$eh4 at news.u.washington.edu> joe at evolution.u.washington.edu (Joe Felsenstein) writes:
>>>"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.
I agree with Joe here that there is some strong resistance to the notion]
of pursueing a statistical framework in cladistics (ASIDE: I actually
like the word cladist... and am an ardent one :-) ...); witness
Carpenter's diatribe against all-things-random in the journal _Cladistics_
about 2 yr ago. Though defintitely a cladist myself, I don't fall into
that group that eshews issues of probability and have even taken a stab
at some of this myself (see below). Perhaps the perspective is that
there is a feeling that one cannot ever know the "true" phylogeny so
how does one go about looking to empirically measure the performance of
a statistical approach to phylogeny reconstruction?
Parsimony is seen by many (like myself) to be a logical "path of least
resistance" approach to phylogeny reconstruction. That is, why
propose widespread convergence (for example) when there's a simpler
explanation.
I think that one would likely find that those that resist a
probabilistic approach have their grounding in morphology and not in
sequence data.
Whereas an argument can (and has) been made to think of the phylogenetic
signal in sequence data to be a very jumbled one and in need of some
filtering, perhaps by a maximum liklihood approach, or by transversional
weighting or whatever, the equivalent can not really be said of
the evolution of a femur. Mind you, the latter may well be revisited
at differing taxonomic levels; what may be clear at one level may not
be at others.
I confess, once again, to not really understand max-like and cannot
therefore criticize it. Then again, my work to date has been largely
morphological (it's changeing, though) and thus I have dug deep into
a very comfortable parsimony pit, one that I believe in strongly.
Of the "statistical" approaches to cladograms, I believe that there is
great merit to investigating the sensitivity of a data set and the
resulting tree(s) to disturbances. The bootstrap tests a particular
type of disturbance (i.e., can we simulate what parts of the tree(s)
might be unstable to further character data). Mind you, whereas I
agree with the use of a bootstrap to investigate this "variance" I do
not buy into using it to construct some other tree (as proposed by Joe
in his seminal 1985 paper). One reason for this is that by analogy,
where statisticians may use a bootstrap to get a better handle on variance
in a small or complete population, I have never seen it used to reject
the mean as the best-available estimate.
My other reason, related, is that I believe that the bootstrap, though
it can tell you how well a clade IS supported, cannot tell you how well
it is NOT supported. It, thus, IMHO, allows you to accept certain
clades but does not allow you to reject clades in your most parsimonious
trees. Thus, constructing a bootstrap tree, which by excluding them,
rejects certain clades in those most parsimonious, is not a valid
approach. Assigning bootstrap values to clades in most parsimonious
tree(s) is.
>>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.
>
This IS disturbing, I agree. The above, however, assumes that the
coevolutionary biologist is concerned with confidence in their
coevolutionary hypothesis. Rarely are they. Rod Page is. So am I.
The extent to which I have taken it (submitted... and crossing my fingers)
is to ask can I get a fit of the host and parasite cladograms as good
or better when I randomize the observed associations of host(s) and
parasite(s). The upshot of this is, that where the answer is:
observed is no better than random... then less than 100% confidence
in the contributing cladograms isn't going to make it any better.
Where the answer is: non-random association... one could make the argument
that this is only a partial probability of the system.
I do not profess to have a solution to this.
>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.
>
Agreed. Like all issues of fitting a function to a set of data points,
things are not what they appear. Take a set of corresponding values.
Measure the performance of a linear, logistic and logarithmic
function to the data... find that logistic fits best (and has a
significant p-value)... one would choose the logistic. BUT... there
is an infinite number of better fitting uninvestigated functions!
It all comes down to what one considers to be (hopefully a priori)
the suite of defensible approaches to the data. So, the cladists,
like myself vis morphology, contend that the ONLY defensible approach
is parsimony. This tends to put a deadening tone to the argument.
Parsimony finds the best fit given parsimony, thus all equally
parsimonious trees are equally well supported. One can't reject them
any more than one can reject an observed mean as the best available
estimate of the parametric mean given the availability of data.
However, being my sort of cladist involves an underlying conviction
that would be like a conviction that the mean is the best estimate
of central tendency. And, of course, make certain assumptions about
the nature and structure of the data. I am not afraid of such things...
I wish my fellow cladists would more often own up to it though.
That last comments approaches addressing the following:
>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.
>
Thanks to Joe for the kind words. I too am glad of some discussion.
I would invite Joe to (as his time permits) submit a synopsis, in laymans
terms of course for those who like me may not make it past the
formulae, ... that ran on some, sorry... a synopsis of maximum liklihood...
I would definitely be interested!
If only to arm myself against it ;-)
a
Mark
That was long - ouch - I'll try for brevity in the future.
--
Mark E. Siddall "I don't mind a parasite...
mes at vims.edu I object to a cut-rate one"
Virginia Inst. Marine Sci. - Rick
Gloucester Point, VA, 23062