Biology Reference
In-Depth Information
If phylogenetic analysis is performed with each of the weighting
schemes and the same tree is produced each time, a high degree of con-
fidence can be attributed to the result. On the other hand, if changing
a single weight completely alters the outcome, care should be taken
when interpreting the results.
How then is this random weighting done? The simplest technique
is called jackknife. For a dataset containing
n
characters,
n
trees are
calculated by eliminating each character in turn from the original
dataset. In other words, the weight of a single character is set to 0 in
each dataset and all other characters are given identical weights. The
fraction of jackknife trees that are identical to the global tree solution
is given.
A slightly more complicated technique, taking more calculating time
but producing more convincing estimates of robustness, is called the
bootstrap. Here, a large number of artificial datasets are randomly cre-
ated, typically 100 or more. Each artificial or bootstrap dataset is created
by drawing
n
characters randomly from the original dataset of size
n
,
with replacement. One can picture a box containing the
n
characters:
after a character is drawn from the box, it is put back in for the next draw.
This way, some characters may be drawn several times, while others are
not drawn at all. Subsequently, a bootstrap tree is calculated for each
bootstrap dataset. Then, for each branch of the original tree in turn
(the one obtained with the original dataset), the bootstrap proportion
of that branch is calculated as the percentage of bootstrap trees which
contain the branch. The bootstrap proportion may be considered as a
measure of confidence in the branch. The confidence level can therefore
vary throughout the tree.
There are other methods of testing the robustness of individual
branches, but they are specific to a certain reconstruction technique.
MCMC methods, for instance, test many intermediate trees during their
search phase. It is possible to benefit from these intermediate trees in order
to obtain confidence levels for individual branches of the final phylogenetic
tree. However, since the sample of intermediate trees is strongly autocorre-
lated, much larger samples are required than for bootstrap methods.
9
Similarly, in CMP and CC, a method called branch-decay analysis can
be used. As with Bayesian methods, large numbers of trees are calculated
