corroboration

[R. Zander]

Doug Yanega wrote:

> John Trueman wrote:
>
> >2.  that *after* the hypothesis was constructed we subjected it to
> >such-&-such *critical* tests, and in so far as it did not fail those tests
> >it is corroborated.
> >
> >Here, the key words are 'after' and 'critical'.
> >
> >Re After:   A hypothesis cannot be corroborated using the very data from
> >which it was constructed.  In this I disagree with certain well-known
> >cladists who see 'corroboration' of a parsimonious tree merely in the fact
> >it is parsimonious. These cladists argue that "the most parsimonious tree
> >is the least falsified tree", meaning it is the tree least falsified by
> >homoplasy.  Out of the set of all possible trees it is the tree least in
> >need of protection from falsification by the addition of ad-hoc hypotheses.
>
> As I have always understood it, the tree is NOT the hypothesis being tested
> in cladistics.

Trees as hypotheses were assumed in my contributions to this thread, too, Doug.
Yes, I acknowledge your correction. I test trees, cladistics does not. I test
them by looking for data sets that may change their reasonability for the better
or worse. In likelihood, that means a higher posterior probability, and in
parsimony studies that can only mean data that support no reasonable
alternatives.


> Each set of character states is a separate hypothesis of
> homology. When one performs a cladistic analysis, all of the hypotheses of
> homology are compared to one another simultaneously, and the most
> parsimonious tree is the resulting hierarchy which results in the smallest
> total number of rejected hypotheses.

This is an optimality criterion. The smallest number of rejected hypotheses
determines a heirarchy that selects one (or a small set) of trees from a large or
very large set of reasonable trees without justification. Occam's razor works
only because, psychologically, we try to test the simplest hypotheses first and
nature is generally parsimonious but not optimally so, and Occam's razor only
works with prediction, not retrodiction.


> Trees are not falsifiable nor
> corroborated in any sense, only the individual hypotheses of homology upon
> which the tree is based, and they are only rejected *in* the context of
> constructing the tree, not *after* the tree is constructed. In this sense,
> your argument about corroboration appears to be a red herring. At the
> least, one cannot assign probabilities of being right or wrong to these
> individual hypotheses either before or after the analysis.

No, one cannot assign probabilities in parsimony analysis, one can only rejoice
in a vaguely defined pool of reasonable trees, which is why statistical phylogeny
is so attractive.


>
>
> >Do changes to our character set change the tree?  We might try a character
> >jackknife.  We might try resampling from the available characters as in a
> >nonparametric bootstrap: which of the nodes are supported and which have no
> >support (which are corroborated and which are refuted) by this test?  If we
> >have an explicit model we may try a parametric bootstrap technique.
>
> To quote Wenzel's recent paper, after he gives a list of known (published)
> limitations and problems with the bootstrap, he concludes; "It is also
> important to note that the ordinary application of bootstrapping procedures
> is to derive a confidence interval for the estimate of a parameter of a
> distribution (Manly, 1991), and that statistical mean and variance do not
> apply to unique historical identity of the phylogeny (Wenzel & Carpenter,
> 1994). So, although the bootstrap values mean something, it is not clear
> what they mean exactly, and it is clear that they do not mean the same
> thing when compared across different trees. It would seem that these
> objections would suffice to exile the bootstrap from its current place of
> honor." He suggests that Bremer support values are more informative, but
> notes; "...these are not intended to serve as statistical confidence
> tools."
>
> I think it would be fair to say that cladists do not view ANY of this
> process as having or requiring statistical properties, and therefore view
> attempts to include probabilistic or statistical approaches in phylogeny
> reconstruction as misguided at best, and misleading at worst, stemming from
> a preoccupation with *process*-based models (e.g. max-like) which are NOT
> compatible with parsimony-based cladistics. Parsimony is not a process
> model - it is Occam's Razor applied to a nested set of hypotheses of
> homology. The kinds of tests you want to apply appear to be inappropriate
> for cladistic analyses.

Cladistics confuses classification, which can be based on anything arbitrary,
with reconstruction, which requires an expectation (read statistical probability)
of being correct. Cladistic analysis ends with elimination of grossly
unreasonable trees. Likelihood analysis begins by evaluating these trees
statistically using models and molecular data. Likelihood analysis fails,
however, when it stops with optimality criteria like maximum likelihood. One must
calculate posterior probability, which must be very high for a scientifically
acceptable reconstruction.


>
>
> Peace,
>
> Doug Yanega    Depto. de Biologia Geral, Instituto de Ciencias Biologicas,
> Univ. Fed. de Minas Gerais, Cx.P. 486, 30.161-970 Belo Horizonte, MG   BRAZIL
> phone: 031-449-2579, fax: 031-441-5481  (from U.S., prefix 011-55)
>                   http://www.icb.ufmg.br/~dyanega/
>   "There are some enterprises in which a careful disorderliness
>         is the true method" - Herman Melville, Moby Dick, Chap. 82



--


Richard H. Zander
Curator of Botany, Buffalo Museum of Science
1020 Humboldt Pkwy, Buffalo, NY 14211 USA
bryo at paradox.net   voice: 716-896-5200 ext. 351