Weights
Tom DiBenedetto
tdib at UMICH.EDU
Wed Mar 5 15:37:35 CST 1997
Guy Hoelzer wrote:
>
>I think you agreed in other posts that choosing different weighting schemes can
>change the tree (your conclusion). I think that this fact, combined with the
>arbitrariness of any weighting scheme, demonstrates that you are making
>assertions (assumptions) about past processes in your group when you apply equal
>weighting.
I think that you are conflating two issues here; weighting and the
probability of particular processes. This is how I understand our
dispute:
We both make use of a parsimony algorithm. It is possible to
introduce a weighting factor into the algorithm, one which can
influence the results. One can choose to allow this weighting factor
to bear any consideration one wants. But the question which is
immediatly begged, is does one wish to introduce an a priori
consideration which has the power to influence one's results or not?
And, as you emphasize, is the decision to forgo such an input in need
of as much justification as the decision to make use of it?
After this issue is dealt with, and one (you) decide to make use of
such a factor, a second question arises,,,what consideration shall
the factor bear, and is one willing to be limited to generating
results which are contingent on this factor.
I think we disagree on both issues, and it would probably be best to
consider them separatly.
As to the question of simply using such an apriori factor or not,,I
choose not to. My justification makes no reference to process
considerations, because my approach does not attempt to include
presumptive knowledge of these factors to begin with. I am dealing
with a set of homology hypotheses which are being combined in the
parsimony algorithm into a logical hierarchy. That is the only
purpose for which I use the algorithm. Whatever considerations I may
deem relevant to a determination of whether a particular character
distribution accuratly indicates a set of homologus states (and hence
indicates homologous transformations) have been dealt with
beforehand. The matrix I use is composed of characters I have
determined to be homologus to the best of my ability, and I now
simply wish to combine them into a parsimonious hierarchy. I do not
recognize any inherent inequality in this set of grouping hypotheses,
and I see no justification for imposing variables which can structure
the outcome in accord with my preconcieved notions.
There is a solid methodological logic which undergirds this approach;
one which has developed from a centuries old approach to studying
character diversity and homology. The use to which we put the
parsimony algorithm is defined in this context, and has proven
extremely successful in practical application. The potential
introduction of weighting factors is not an issue in this approach.
It strikes me that you take a very different approach, and see the
apriori addition of influencing factors to be justified. Within the
context of an approach which uses the algorithm for very different
purposes, as an arena in which process considerations are intended
to be introduced, then the use of the weighting variable seems
justified. I am not convinced that the general notion of using
process considerations is wise, but within that context, I can
understand your use of weights.
And then we reach the decision to use the variable to carry a
probability score. My objections here are, as I think you percieve,
that specific probability scores calculated from process
generalizations should
not be expected to apply in new situations; that
whether they do or not is something we wish to learn from the tree,
and that allowing probabilites to structure the tree would be
circular. The tree becomes contingent upon the general expectations,
and
cannot speak to the issue of whether those expectations are valid in
this case.
> Your justification in the last sentence seems to me
>completely lacking in logic:
>
>1. You do not consider "that factor" relevant
>therefore
>2. equal weighting does not even consider such factors and makes no assertions
>whatsoever on that score.
>How does one follow the other? Whether or not equal weighting makes such
>assertions surely does not depend on what you personally consider relevant.
I have an apporach in which process generalization factors are not
deemed relevant. The fact that I equally weight makes perfect sense,
and is completely justified by the conceptual framework within which
I operate, and this justification makes no reference to process
matters at all. Were I to operate in your concpetual framework; were
I to be trying to discover a tree which is probable given a
consideration of process generalizations, then yes, equal weighting
would represent an assertion on that score. But that is not what I am
doing at all. I am using the algorithm to logically combine a set of
grouping hypotheses, each of which has an inherently equal standing
as a corroborated homology. Thats all.
> if I am correct that you acknowledge that different weighting schemes,
>which reflect our expectations of relative process frequencies in the past, lead
>to different trees, then they must be relevant. You are simply unwilling to
>apply unequal weighting because you think that practice requires unavailable
>knowledge of the history of characters in your group.
Its deeper than just that. My approach does not attempt to derive
trees which are contingent on presumed processs knowledge. I am
logically combining pattern statements, not process hypotheses.
Pattern statements (grouping hypotheses) reflect the factual
distribution of states which I observe and corroborate. Probability
weighting has no
relavance to the task I am asking my algorithm to perform.
>I think that equal weighting is no different. If in fact the frequencies of particular
>character-state transformations was strongly biased in the history of your
>group, then your tree will be mislead by your assumption of equal weighting.
You are simply endlessly trapped within your own
conceptual framework. I am not asking you to leave it, just to
realize it is not the only approach. Within your framework, equal
weighting is in need of as much justification as any other scheme. I
heard you the first time, and I certainly understand.
> Are you saying that you essentially do compatibility analysis and dispose of all >characters showing any signs of homoplasy? Is it always your goal to end up with a >clique of completely compatible characters? If so, then there would be no
> distinction for you between weighting characters as 1 or 0, and in weighting events > as 1 and 0 because every mutation in every character you retain would provide a
>synapomorphy on your tree.
Huh? No. Where did you get the compatability stuff from? I was making
a different point about how we test homology hypotheses before
arriving at matrix, the tests do not entail assigning probabilites to
hypotheses, but merely refuting them or corroborating them. The
congruence test does the same, and does so not only for the character
as whole, but the implied transformations.
>>This is a form of verificationism; your trees will,
>>to some extent, reflect the assumptions you pump into the process;
>>they will serve to verify previous knowledge rather than challange it.
>I would not use such trees to test the same processes that I assumed in building
>the tree. But I agree that my conclusions follow from my assumptions. So do
>yours.
Well, good, I am glad you see that your trees are disqualified as a
test of the process generalizations,,,so how then do you ever propose
to test those process generalizations?
And yes, your conclusions follow from your assumptions. You have a
tree which is probable given your generalizations. We are attempting
to approach things in a way which avoids such limitations. We are not
assigning equal probabilties to various evolutioanry transformations;
we are assigning equal weights to homologous grouping hypotheses.
>Again, I am not using the tree to test those process assumptions. Nevertheless,
>it is critical that we all own up to our assumptions, be aware that they affect
>our conclusions and be open to alternative approaches to data analysis that make
>different sets of assumptions.
And it would help if we could stretch our minds to concieve of
different approaches which do not necessarily revolve around the
process concerns that you have. It shouldnt be too hard; this is how
non-molecular systematics has been done for quite some time.
More information about the Taxacom
mailing list