I think at least part of the reason that quartet puzzling tends to give
low values is that each p-tree is built without further optimization
following addition of the last taxon. It is analogous to a parsimony,
likelihood, or Fitch-Margoliash distance analysis in which multiple trees
are built by stepwise addition of randomly ordered sequences but in each
case with no further exploration of tree space to find the topology that
best fits the data. The resulting set of unoptimzed trees is then used to
compute a 50% majority rule consensus tree.
If you were to conduct a Kishino-Hasegawa test on the set of p-trees, I
fully expect you would find the majority are significantly different from
the best of the lot, which itself is liable to be significantly different
from a fully optimized tree, yet all p-trees are given equal weight in the
computation of the quartet puzzling values.
Sean Turner, Ph.D.
Department of Biology
Jordan Hall 142
1001 E. Third Street
Bloomington, Indiana 47405-3700
E-mail: sturner at bio.indiana.edu