[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Subject Index][Author Index]

Re: Benton et al.'s Supertree (longer)

My 2p worth. Not sure if this was meant to be an on or off-list post, but
I've scribbled some thoughts down about this weighting issue. If it was an
accidental post, forgive the intrusion.

Mickey_Mortimer111@msn.com wrote:

>> How is valuing someone's undefended opinion of a relationship
>> equally to
>> someone else's result of a 300+ character phylogenetic analysis
>> useful?  ... An earlier tree with less
>> characters and taxa available can only be "bad" compared to a modern
>> analysis.

I think we're all agreed that at a gut level we know some methods of
phylogeny reconstruction are better than others. Because true phylogeny is
unknowable, we can rank methods, but we cannot quantitatively say how much
more 'correct' the results of one method are than another, even when applied
to the same data. We all do this when we decide to use HKY rather than
2-parameter ML models, or parsimony rather than phenetics (or cladistics
rather than dowsing). The problem comes when we 'admit' in a supertree
(henceforth MRP) analysis that this difference in likelihood of
correspondence to the true tree exists, because we suggest that some
character sets a priori contain the most correct signal we can get, whilst
other sets contain less correct signal. We then weight according to our
ranking scheme if we wish. In morphological cladistics, this is equivalent
to choosing and coding characters so that a particular topology will be
favoured in the end - it's a cardinal sin, and invalidates parsimony as an
optimality criterion because the final topology does not represent the
minimal amount of feature (not character) change experienced by the group.
The ideal strict cladistic analysis will include the entire universe of
possible informative features, coded as characters in an unbiased way, so
that global parsimony sorts the winning signal from the losing (homoplasy),
with no a priori amplification,

Obviously this is why strict cladists get into such frightful arguments -
because no analysis is ever going to be like this, signal amplification
always occurs, and the result is never objective (even if parsimony is an
accurate reflection of reality in the first place). But in MRP we're hugely
privileged because we know that global parsimony is absolutely the correct
criterion ('cos we chose it to be so), and that we have completely sampled
the universe of possible characters ('cos there are only so many source tree
nodes, and we've coded all of them) but that means we have to play by
parsimony's rules. That means we assume a priori equality of character
'truth content' and find the globally parsimonious solution, rather than one
distorted by well-intentioned weighting: we put in all the alternatives
without suggesting that some are better than others because the optimality
criterion is the judge of that. I'm not advocating equality because it
doesn't involve weighting, but precisely because it weights source tree
statements equally. We must to avoid deciding which our 'best' characters
are before the analysis at all costs, if the resulting tree is to have any
meaning at all. 

It's this problem of the final tree 'meaning' something that causes all the
problems. We cannot know how close our source trees, and therefore any MRP
product of them, is to reality. We can differentially weight in an attempt
to make the MRP tree closer to what we suspect is the truth, but this can be
nothing more than a suspicion, and as I hope I've just suggested it
invalidates the use of parsimony as an optimality criterion. In my opinion,
far better to present an MRP tree as a strictly parsimonious consensus of
source tree statements and if one's happy with that then maybe "an
algorithmic consensus of phylogenetic opinion over a period of time".

Would like to hear people's opinions on this - I'm finalising a thesis
submission at some point!