James Francis Lyons-Weiler
weiler at ERS.UNR.EDU
Sat Mar 8 13:46:43 CST 1997
Tom D commented, after me:
> and how can you tell the
> difference a lot of tress generated by random data, and
> a lot of trees generated by screened data.
>Personally this has never been a major concern of mine, and I still
>dont really understand why it should be. I dont use random data, I
>use "screened" data. My tree calculation, as a logical combination of
>hypotheses, is meant to tell me how those hypotheses can be logically
>combined. Thats all. I believe in using a 'total evidence" approach,
>such that the tree I end up with is the logical combination of all
>the evidence out there for the group. We cant really do much better
>than that, until we gather more evidence. So the tree stands as the
>best logical combination of available evidence no matter how much or
>how little it differs from "randomness". I am not saying that I would
>be untroubled by a demonstration that my tree is no better supported
>than it would be given random data,,,,but in any case, the only thing
>I could do, or would do, is go out and gather more evidence.
This response really begs the question: you claim from the beginning
that your don't use random data. So a test against randomness should
NEVER find a cladistic data set where the distribution of character
states among taxa (or binary re-coded character states) is not
different from random, right?
There are two reasons to test for signal, whether one uses the parsimony
criterion, or compatibility methods. First, your screening tests may be
better than mine are; i.e., I may be working on sets of putative
homologies for which corrborative evidence is difficult to define (I'm
talking about morphology here). Second, sometimes one may generate
hypotheses of homology which might as well have been generated at random.
Imagine two scenarios: in the first, you're quite confident about your
hypotheses of homology, and you don't test for signal. In the second,
you're quite confidence about your hypotheses of homology, yet you DO test
for signal. Nothing is lost by testing, if you pass; there is plenty to
gain by testing if the matrix fails. The test itself, you see, provides
another critical test of the sets of hypotheses of homology. Obviously,
if one has so much confidence in their perfectly logical (?!) method of
combining hypotheses of homology, that they decide never to test against
randomess, they'll never know.
Now, tests of signal pertain to the matrix as it has been constructed; it
matters little how the matrix came to be, it only matters that it be
sufficiently testing, and to be allowed to kick and scream. The absence
of signal is a symptom that something is wrong somewhere. The next
question is "where"? Is it a particular taxon that is messing things up?
Is it a particular character? A set of characters? Can the signal be
exposed by removing noise?
i think all of this falls in line with your thinking as well as the
neo-cladistic interpretation of homology, trees, and estimates of
phylogeny. I see very little difference, except for the use of
trees as critical tests. Imagine that all of your hypotheses of homology
are correct; the mpt won't tell you anything you didn't already know;
again, where are the bold statements?
\ / / \ / JAMES LYONS-WEILER ______________
\/ / \/ |..............|
\ / / |..............|
\/ / DOCTORAL PROGRAM IN |..............|
\ / ECOLOGY, EVOLUTION, AND |...***........|
\ / CONSERVATION BIOLOGY |..*****.......|
\ / |.******.......|
\/ 1000 VALLEY ROAD/186 |********......|
______________ THE UNIVERSITY OF --------------
| will perform | NEVADA, RENO
| statistical | RENO, NEVADA 89512-0013
| phylogenetic |
| analyses for | "(Biology) is not religion; if it were, we'd
| food | have a much easier time raising money."
-------------- -Leon Lederman
More information about the Taxacom