Probabilities on Phylogenetic Trees

Curtis Clark jcclark at CSUPOMONA.EDU
Thu Sep 11 07:50:49 CDT 1997

At 07:40 AM 9/11/97 -0700, Richard Zander wrote:
> Simply put, I'd like a
>measure of when my shortest tree has more evidence for than against. I
>think there is a mathematical measure, though not exactly sure what it

Aha! I've been puzzling over this thread, knowing that there was something
important here, but not clear enough on the discussion to know what was
going on. Now maybe I do.

>  Also, if my shortest tree isn't well-supported, I'd like to identify
>portions of the cladogram that have more evidence for than against. (I
>think this is what people mean by "we don't think our cladograms are
>necessarily exactly right, but they are probably mostly right."

I wonder if you are maybe not asking the right questions. Certainly knowing
whether a tree has more evidence for than against must be dependent on the
size of the tree. If we look at phylogenetic information in a
"signal-to-noise" fashion, we might postulate as a null hypothesis that
noise is evenly distributed, but a lot of evidence would lead us to believe
otherwise. So any tree will include "noisy" areas as well as "signally"
areas. Any measure of the probability of the tree ignores that.

Identifying the portions of the cladogram that have the most support seems
to be more relevant to what we are trying to do. As an analogy, let's say
you want to determine if there is any pattern in the first appearance of
the sun each day. You have a very accurate clock, and you note each day the
time at which the full disk of the sun is apparent. But you live in a
cloudy climate. Sometimes the sun appears from the horizon, sometimes from
a cloud. Let's say that you know the difference between the horizon and a
cloud, but for some reason or another you are not able to identify which is
which. What you want to know is the portion of the data that corresponds to
the sun appearing over the horizon. The probability of all the data
representing that is very low, and that is not the right way to look at the
data. What you want instead is an estimate of which part of the data is
more reliable.

That's what bootstrap, jackknife, decay indices, and the like are supposed
to do. For reasons that are not yet clear to me, there are detractors of
all these methods, but it seems to me that people who make these methods
are on the right track.

I've recently and somewhat tongue-in-cheek suggested that nodes in a
cladogram without at least 50% bootstrap support should be collapsed into
polytomies, since in a sense they don't really exist. If bootstrap analysis
doesn't do what it is supposed to do, so be it, but if we find a method
that does, my technique at least focuses us on what we *know*.

On a perhaps more humorous note, I've finally put on the web a spoof I did
some years ago on the Statistics Cops:

Curtis Clark       
Biological Sciences Department                     Voice: (909) 869-4062
California State Polytechnic University, Pomona    FAX:   (909) 869-4078
Pomona CA 91768-4032  USA                          jcclark at

More information about the Taxacom mailing list