[Taxacom] cladistics (was: clique analysis in textbooks)

[Sergio Vargas]

Servus,

I've noticed some interesting points, made a day ago, one day too late:

 >I am not tied to any particular analysis. I've used parsimony because 
it's widely accepted readily accessible and presents a straightforward 
clustering procedure (with >caveats). In principle I would be happy to 
use three-item analysis, although right now I do not have access to a 
program to do that.

I'm not sure one could/should call Maximum Parsimony (MP), or Maximum 
Likelihood (ML) and derivatives, clustering procedures. There are many 
fundamental differences between a clustering algorithm and what 
tree-search algorithms do. I would accept Neighbour Joining (NJ) to be 
called a clustering algorithm because this method actually produces a 
cluster based on a (overall) similarity index calculated for pairs of 
taxa. MP and ML do not work this way. In a NJ analysis the tree space is 
never searched because the leaves of the tree are clustered together 
using the NJ criterion whereas in a MP or ML analysis the algorithm 
actively searches the tree space to find the best tree (best not equal 
to the true tree). This means that each new tree is evaluated to check 
whether it is a new best hit. The leaves of the tree are never clustered 
together, using a pair-wise measure of similarity NJ-style, during the 
analysis. In this respect MP and ML are not clustering techniques even 
though they produce a cluster-like graph as output.

 >Don't understand what is meant about how the retained data set will 
mechanically change without biological reasons with respect to larger or 
narrower ingroup.

say you want to evaluate the relationships of the following taxa, 
((Chimp,Human,Gorilla,Orang),Monkeys=Outgroup).

What Pierre is pointing out is that if you constraint your data matrix 
in a way that only characters that appear to be "shared derived" in the 
ingroup are used for the analysis, the data matrix (lets call it A) you 
end up using is (necessarily) a proper subset of a bigger matrix (lets 
call it B) composed by the characters matching your pre-selection 
criterion + the character not matching it. After the analysis of matrix 
A you find ((((Orang,Human),Gorilla),Chimp),Outgroup) as the MP tree, so 
far so good.

But say you now want to evaluate the relationships of 
((Chimp,Human,Gorilla,Orang,Monkeys),Whatever_appropriate_taxa=Outgroup). It 
is very likely that applying the same pre-selection method, viz. only 
shared derived characters in the ingroup will be included, you will end 
up having a matrix (say C) that is either very similar to B or is B 
itself, but either way must include A too because the way you pre-select 
your characters. After the analysis, you find that the new MP topology 
is (((((Chimp,Human),Gorilla),Orang),Monkeys),Outgroup). This happens 
simply because of your pre-selection criterion, has no biological basis, 
and would not have happened if you had included all characters from the 
very beginning. I would add is sufficient to reject the pre-selection 
criterion ad portas and refute the results obtained using it as 
logically inconsistent.

 >Where this is done to include all characters, whether or not they are 
restricted to the ingroup, to my mind results in an analysis of overall 
similarity rather than >cladistic derivation.

I think the use of "overall similarity" here is misleading because it 
sound as if MP and NJ are exactly the same, and they are not (see also 
above). The overall similarity John Grehan is referring to, seems to be 
NJ-like, that is: character vectors are compared pair-wise without 
reference to other character vector during this comparison, and the more 
similar taxa are grouped together. MP analysis does not work this way. 
In a MP analysis the total cost of a transformation series is 
effectively determined by the entire set of characters with the larger 
set of congruent characters dominating the analysis (and topology), and 
setting a higher total cost for non-congruent transformation series. The 
analysis of all characters is only the analysis of overall similarity if 
one uses a method that does this, as I have said before in this thread, 
the method of analysis can be phenetic not the character matrix. The 
character optimizations resulting from the analysis of a non-preselected 
character matrix using MP are "cladistic derivation" not overall similarity.

cheers

sergio


-- 
Sergio Vargas R., M.Sc.
Dept. of Earth&  Environmental Sciences
Palaeontology&  Geobiology
Ludwig-Maximilians-Universität München
Richard-Wagner-Str. 10
80333 München
Germany
tel. +49 89 2180 17929
s.vargas at lrz.uni-muenchen.de
sevra at marinemolecularevolution.org

check my webpage:
http://www.marinemolecularevolution.org

check my research ID:
http://www.researcherid.com/rid/A-5678-2011