Assumptions and reliability of solutions

[Karl Magnacca]

On 14 Mar 2005 at 16:20, Richard.Zander at MOBOT.ORG wrote:
> I've gathered together, for an article I'm writing, about 35 assumptions
> commonly ignored in modern analyses. Molecular studies are correct and
> reliable only if there is unambiguous sequence alignment, results robust to
> all reasonable evolutionary models, all software implementations agree, no
> change in results with different outgroups or ingroups or genes or amount of
> data,

These are not assumptions, but products of the data.  The sequence
alignment is either unambiguous or has some potential error, all
software implementations either agree or they don't (or some aren't
checked), etc.

> no sample error, identifications are all correct, reagents are not
> contaminated, gap costs are fully justified

*These* are assumptions.

> Suppose of the 35 assumptions, 10 are relevant (or less than totally
> correct) and each is 0.999 correct (one out of a 1000 that a branch
> arrangement of interest (as some solution to a taxonomic or evolutionary
> problem) will be different if the assumption is not true. The reliability
> measure given in the cladogram must be then reduced by 1 percent (10 times
> 1/1000).

I would be very interested to hear how you would assign probabilities to
whether determinations are correct.  After all, if we could do that for
all the things you list above (including the ones that aren't really
assumptions), we would have a simple measure for telling how good our
trees are.  People have been looking for that for 30 years, and still
haven't found a good one.

> Suppose you have a branch arrangment presented by the software as 100
> percent reliability (say, given 100% Bayesian posterior probability). Up to
> ten assumptions at 1/200 chance of being incorrect reduce the reliability to
> up to 95 percent. More than 10 assumptions at 1/200, or if 10 assumptions
> are more than 1/200 chance of being incorrect, reduce the reliability to
> less than 95%.

This is a strawman.  Bayesian probabilities are not absolute
probabilities of the arrangement being correct.  Like bootstraps,
they're simply a way of assigning some kind of accuracy number to
something that can't really be given a probability.

> The difference between the cladist and the non-cladist is an apprehension of
> the former that the product of the probability of all assumptions being
> correct is always more than 95 percent (more towards 100 percent), and the
> decision of the latter that it is less than 95 percent.

Actually, this sounds like the difference between the cladist and
someone who doesn't understand basic cladistic principles.

Anyway, the whole argument that has been going on was about cladists in
the sense of those who favor strict cladistic classification vs. those
who prefer to allow paraphyletic groups, not cladists in the sense of
people who do cladistic phylogenetics.  I assume by non-cladists or
"traditional systematists" you're talking about what might be termed
strict taxonomists (in terms of their activities, not ideology).

> I think the anger of non-cladists is that the onus of dealing with
> problematic assumptions is laid on them.

In what sense do you mean?  Having to deal with changing classifications
created by poorly-supported phylogenies?

Karl
=====================
Karl Magnacca, USGS-BRD
PO Box 11, Hawaii Natl. Park, HI 96718
"Democracy used to be a good thing, but now it has
gotten into the wrong hands."   --Sen. Jesse Helms