In article <86qkp2$4ae$1 at mercury.hgmp.mrc.ac.uk>,
Brice Quenoville <quenovib at si.edu> wrote:
[If you could keep your lines to 70 characters it would be
much easier to respond to your messages.]
>1/ I am analyzing nuclear data sequences for which I have some sites
>coded as ambiguities (following a IUB code). I am wondering how
>the last Paup version exactly treats such positions during a ML search.
I only know in detail how PHYLIP treats them, but I suspect PAUP*
is the same.
Normally the likelihood of an observed site is 1.0 for the base
observed, and 0.0 for the three bases not observed. An ambigous
site has a likelihood of 1.0 for each of the bases it could be.
If you'd like to think of this in a parsimony framework, it's
analogous to saying that we'll assume whichever of A or G allows
the more parsimonious solution; if we have several taxa who
are ambiguous for this site, we'll assume A or G for each
one separately to get the most parsimonious solution.
There is still information in such a site, and it is worth including
in the analysis.
>2/The last updated version of Paup provides a table including
>branch lengths, their standard errors and a LRT test under the null
>hypothesis that a branch has zero length. For some branches I have a
>standard error that is weakly higher or equal to the branch length
>itself. However, the LRT tests still tells me that these branches
>are significantly different from zero and thus statistically do exist.
These are two quite different statistical tests, making different
assumptions and approximations. They frequently disagree on
borderline cases. As far as I know there is no clear rule for
preferring one test over the other. I'd say that a branch on which
the tests disagree is weakly supported at best.
PHYLIP shows the same behavior--I inadvertantly used such a case
as a teaching example once, and the students really grilled me on
Can't help with your other two questions, sorry.
Mary Kuhner mkkuhner at eskimo.com