Felsenstein versus Ockham?
tdib at UMICH.EDU
Mon Jan 25 14:35:45 CST 1999
R. Zander wrote:
>Scientists in general are interested in prediction of testable hypotheses. Cladists are interested in retrodiction of a
>single, complex phenomenon.
Yes, we have a historical science. But that does not mean that we do not develop testable hypotheses. We develop
testable hypotheses of homology between character states, and test those hypotheses for congruence against
> Simplicity is justified in prediction when selecting hypotheses to test, since you have a testable
>hypothesis and you can always get more data.
I think there is more to it than that. Simplicity is inherint in the scientific approach. If one does not attempt to
formulate the simplest explanation, then one might as well rest content with a separate explanation for each
phenomenon (the most complex explanation); i.e. no science.
And we can always get more data too; ie more charcters.
> Simplicity applies to evolution to the extent that observation has told us (inference) that like produces like, that
>much convergence is not associated with most observed evolutionary processes.
I disagree, That is not the justification for simplicity per se in systematics. It may be, in a sense, a justification for
doing systematics in general, but only to the extent that if convergence was ubiquitous, then one would have no hope
of ever discovering historical relationships. We do assume that there is historical "signal" in character distributions.
That gives liscense to our science in general. Simplicity tags along, since it is an inherint part of science.
>Thus, optimality criteria can be used in eliminated unreasonable hypotheses, or in the case of cladistics,
>hypotheses of homology associated with long trees. Homology as observed and theorized in evolutionary processes
>is mostly maximized. Mostly, but not all.
The homology we hypothesize is prior to evolutionary theory (as is systematics). Evolutionary theory must explain our
discoveries of homology, they are not the source of it.
> There is no evidence a priori that a data set with species A and B sharing characters 1-5 must be more closely
>related than species A and C sharing characters 6-9.
What do you mean? That *is* the evidence. Character one has been studied and found to be the *same thing* in both
taxon A and B. Thus it directly indicates relationship between them. The same is true for characters 2-5. The others
indicate a different relationship. Thus our final hypothesis will have to "invent" some unseen factors (ad hoc
hypotheses of homoplasy) in order to square our empirical observations with our underlying assumptions (a single
branching hierarchy). Just as with the Ptolomeic epicycles, we follow the discipline of minimizing the number of
unseen factors we must invoke in order to explain the data. It is the discipline of empiricism; we try to square our
theories with empirical evidence as much as possible, and thus minimize invented band-aids (homplasy). That is what
makes the parsimonious explanation the "best"; it has nothing to do with an assuption that the world is simple, nor
does it have anything to do with a probabilistic assessment.
>Your use of the term "prefer" demonstrates a fatal weakness (associated with optimality criteria) in cladistics as
>generally practiced. Philosophers "prefer," "select" or "choose" theories, scientists act upon them by basing other
>analyses on them:biogeographic, other phylogenetic analyses, etc.
I dont follow the logic here. Scientists as well as philosophers, select theories. And we as cladists not only
select a theory but act on it as well.
>A probabilistic (or actionable or dependable) hypothesis should have more going for it than "preference."
What do you suggest as "more... than preference", certainty?
But in any case, ours are not probabilistic hypotheses. There is no way to calculate the probability that hair arose
one time or 4000 times. There is no homogenous reference class against which to calculate the probabilites of
evolutionary events of character transformation. At a higher level, there is no way to characterize the probabilities of
various topologies. The evolutionary process happanded only one time; I dont see from what perspective you can
calculate probabilites of various outcomes.
>In the case of parsimony, I recommend a high Bremer support value for a clade before a cladistic hypothesis can be
>seen as a genuine probabilistic reconstruction that is better than any other method of grouping taxa by characters
>influenced by evolutionary processes.
I have no argument with the notion that better-supported trees are a good thing to have. And I agree that Bremer
support is a relevant and useful metric. And I agree that a weakly supported tree is not one on which anyone should
base susbequent dependant studies. But I dont think many people do that anyway. And I dont think it has anything to
do with probabilism.
> We all like phylogenetic reconstruction methods that model evolution
>(parsimony, likelihood, Bayesian)
Cladistic parsimony does NOT model evolution, and "we all" do not necessarily like reconstruction methods
that do. In fact, cladistic parsimony methods explicitly shun evolutionary models, based on the understanding that
models are attempts to *explain* pattern, and must come after the pattern has been determined. To model the
process as a means of determining the pattern is to put the cart before the horse and sets up a situation in which
one's model is untestable, since the product of the model (the phylogeny) is, at once, both the only potential test of
the model, and also a dependent outcome of the model.
>And I said, in a recent publication, "...a corollary to Occam's Razor is that explanations must remain multiple when
>no one of them is probabilistically adequate.
And I disagree that this corollary has relvance to these issues. One cannot probabilistically assess the adequacy of a
phylogeny based on presumed knoweldge of evolutionary processes, since knowledge of evolutionary processes
cannot be determined except in the context of a given phylogeny.
>> This says nothing about how complex or parsimonious the real world is. One must explain whatever
>> complexity one discovers,
>No, I don't think so. One does not have to explain whatever complexity one discoveres, most especially when the
>data is contradictory and more than one reasonable hypothesis is possible. In the context of action, you are arguing
>that, of a lineup of criminals, one must find some one of them guilty, namely the one with the most evidence against
>him/her, even if there is reasonable doubt.
But you will agree, will you not, that one of the alternative toplogies for a given set of taxa MUST BE guilty, no?
>>The only optimality critereon in cladistic parsimony is that one report only
>> those evolutionary transformations that are indicated by the data
>But you select those that you report. Autapomorphies are not reported. Contrary evolutionary transformations in a
>tree one step longer are not reported.
Autapomorphies can certainly be reported. But they need not be, since they contain no information bearing on
relationships. And yes, of course we select what we report. We report those evolutionary events that are directly
indicated by the data, plus a minimal amount of non-empirically based "ghost" transformations necessary to explain
the data in light of our assumptions of a singular historical hierarchy.
More information about the Taxacom