Striking a balance, weighting and Cladistics

[Thomas DiBenedetto]

-----Original Message-----
From: Richard Zander [mailto:rzander at SCIENCEBUFF.ORG]

>Those who reject statistics but use philosophy to guide their choices must
>often make Type I errors. They almost always get publishable results,
>because an exact solution is (philosophically) always acceptable (if only
>"poorly supported").

Richard,
I am a little confused as to the relevance of your point. Let me try to
respond with a few points of my own.
1. The focus of my argument was on the method for selecting a preferred
tree. Your point seems to address the issue of the "strength" of the result,
an issue that seems to have a general relevance, not limited to any
particular procedure for arriving at the result.
2. Cladistic results are always accompanied by a set of indices or metrics
of the "strength" of the results; most notably the decay index. So
cladistics certainly recognizes that some results are weakly supported, and
supplies the numbers by which that can be ascertained. I think that most
everyone in the field gives the appropriate amount of respect to a weakly
supported result.
3. It is also happens very frequently that the result of a cladistic
analysis is a set of equally most parsimonius trees, so I don think that it
is fair to say that the method always produces some specific result.
4. Ironically, it is the most common statistical method (max-like, the
method I was discussing) that tends to always generate a single result (the
tree of maximum likelihood) no matter how weakly supported. As a statistical
method, the procedure is certainly amenable to generating a confidence set,
but how often do we see this? How often do we see a max-like result with
collapsed branches, representing a consensus of all trees not significantly
different from the tree of maximum likelihood? I recognize that it is
standard to approach this issue after the analysis, by use of a bootstrap,
but it doesnt seem to me that that obviates the need to represent the
results of the max-like analysis itself in a responsible manner.
5. Your use of statistical concepts to describe non-statistical methods is a
bit confusing. The 'null hypothesis" of no support is not part and parcel of
a cladistic analysis. We are simply trying to select one or more patterns
from a finite set of patterns, and reporting explicitly the evidentiary base
for that choice. As I said above, this does not preclude a result that
effectivly accepts your null hypothesis (by indicating that all patterns
have equal levels of support - i.e. a bush).