[Fwd: Re: Probabilities on Phylogenetic Trees]

[James Francis Lyons-Weiler]

On Tue, 16 Sep 1997, Tom DiBenedetto wrote:

In response to Richard Zander

> Wow,,,this is a strange perspective. If you find the bloody knife in
> the butler's locker, and it has his fingerprints on it, and the
> victim was known to be blackkmailing him, and he has no alibi for the
> time of death, perhaps you may bet your career on his guilt. But
> given that there are 6 billion potential murderers in this world, at
> some scale you could consider that to be a low-probability result.
> But of course, you go out to collect evidence precisely for the
> purpose of narrowing the scale of your possibilities. Instead of 6
> billion, you narrow the scope down to those who had no alibi, those
> who were being blackmailed, ultimately, to those whose fingerprints
> match those on the knife.

        How can you narrow the scope down to those that have no
        alibi, those that were being blackmailed, etc. unless
        you have checked all 6 billion?  You can't really.  So
        you really rely on the covariance of the evidence, and
        the implicit (hidden; read as the opposite of explicit)
        probabilistic measure of "surprise" that we actually
        found one individual that fits the improbabilistic bill.

        Statistics merely formalizes the process you describe.
        Our guesses include that the victim was not a rampant
        blackmailer, for example (a Bayesian prior, whether
        you admit it or not).

> In systematics, we go out to collect
> evidence. A matrix is the result of our investigations. The
> probability of a result may be low in the face of no evidence, but is
> high in the context of the evidence.

        Huh?  The probability of the evidence is high in the
        context of the evidence?  No kidding...

        It takes a healthly dose of positivism to say that the
        tree we get from the evidence we collect then has a
        high probability... You've lost the game of science from
        the get-go unless you concern yourself with a critical
        test of the (apparent) informativeness of the evidence
        you have collected.  Dick's point is entirely salient
        to asking questions about the joint probability density
        functions of the data, the algorithm, and the tree(s).
        Just because he's posing a complex question doesn't
        mean that it's not worth asking.

In fact, for the shortest tree,
> it is maximized, relative to the evidence.

        Huh, again?  The probability of a tree is maximized
        relative to evidence if it is the shortest tree?
        What does that mean? What are you saying?  That
        given the shortest tree, the shortest tree becomes
        a highly probable event?

        Others rampantly disagree.  The LESS probable a result
        is, the more surprised we are, and the greater the
        corrorborated the result is.  You can't discuss the
        objective application of probability theory if
        one never bothers to measure the probability of an
        event.  A roll a die.  It lands on six.  What was the
        P(6)?  I look at the evidence.  P(6) = 1.0?  No.
        P(Tom will get an F on a probability exam) = ????

        (For the uninitiated, Tom is immune to real criticism,
        so my flames are not ever really felt by him).
        .
> Ultimate probabilites for unique historical events (did they really
> happen?) are either 0 or 1. Intermediate probabilites can only be
> calculated relative to a base of evidence, and are reflective of our
> sense of the reliability of the evidence. Character matches are
> statements of homology which are conclusions based on the sum of our
> biological knowledge. A shortest tree has the
> highest probability *given the evidence* of everything we think we
> have learned about the biology of the character system. within of
> course, the
> underlying assumptions. I cant imagine that you can do better than
> that.

        Again, most statisticians would disagree.  We use
        parsimony to estimate the population mean; in fact,
        the sample mean is the value with minimum error around
        it, and is the maximum likelihood estimate of the sample
        mean when the proper assumptions apply. The fact that
        we try to be precise in our measurements is a given.

        The degree of confidence in a tree or group should be
        a function of the low probability of that group given
        an appropriate null.  Why?  Because the probability
        that one or more shortest trees exist for any matrix
        with variable character states is ca. 1.0, and the probability'
        that those trees will denote groups can be very high, regardless
        of the process that generated the matrix.

>
> >> Are you basing that [this talk of the probability of the true tree] on some b
> elief >>that a particular model is an accurate mirror of evolutionary process?
>
> >Exactly. Assuming that the particular model and particular regularity
> >assumptions you are using are indeed true,*mathematically* can you
> >evaluate the probability that you have retrieved the true tree?
>
> If your process model is TRUE, then why wouldnt your tree be TRUE?
> The whole problem arises from the fact that we KNOW that our process
> models are not true,,and the only way to make them approach the truth
> is to formulate them with reference to well-corroborated observable
> regularities (patterns). We learn about processes through the
> procedure of devising explanations for patterns, not the other way
> around.

        Sounds like a justification for maximum likelihood parameter
        estimation to me.  Also, cladistics in vitro was process-
        oriented... evidence of shared geneaological descent
        invokes a multitude of processes, among them inheritance,
        geneaology, birth, death... and the evidence is
        taken directly from the pattern of character state changes
        on a parsimony tree (or so they thought).  So what that the
        pattern is really not an observation, but rather is an
        inference, a guess?  So what if the degree to which we
        might expect character state distributions to be hierarchically
        distributed appears to require knowledge we can't ever have?

        Your process position is a straw man.  The degree to which
        we might expect to see something is FAR different that the
        degree to which we do in fact see something.  the question
        is, are we observing something (an amount of pattern, a
        short tree) that deserves a PHYLOGENETIC explanation, or
        is it a result that could have happened by chance alone?
        If the result (the degree of covariation in the implied
'       hierarchy found in a matrix) can be easily dismissed as
        a chance event, it (in total) doesn't require a geneaological
        explanation.  Sure, genealogy and inheritance may have been
        ongoing, but those processes and the processes of character
        evolution may have interacted in ways that DESTROY or
        mask the evidence of geneaology you expect to find on the
        mpt. To say otherwise is to claim sufficient knowledge of
        evolutionary processes.

        J. L-W.