The success of 16 methods of phylogenetic inference was examined using
consistency and simulation analysis. Success-the frequency with which
a tree-making method correctly identified the true phylogeny-was exam
ined for an unrooted four-taxon tree. In this study, tree-making metho
ds were examined under a large number of branch-length conditions and
under three models of sequence evolution. The results are plotted to f
acilitate comparisons among the methods. The consistency analysis indi
cated which methods converge on the correct tree given infinite sample
size. General parsimony, transversion parsimony, and weighted parsimo
ny are inconsistent over portions of the graph space examined, althoug
h the area of inconsistency varied. Lake's method of invariants consis
tently estimated phylogeny over all of the graph space when the model
of sequence evolution matched the assumptions of the invariants method
. However, when one of the assumptions of the invariants method was vi
olated, Lake's method of invariants became inconsistent over a large p
ortion of the graph space. In general, the distance methods (neighbor
joining, weighted least squares, and unweighted least squares) consist
ently estimated phylogeny over all of the graph space examined when th
e assumptions of the distance correction matched the model of evolutio
n used to generate the model trees. When the assumptions of the distan
ce methods were violated, the methods became inconsistent over portion
s of the graph space. UPGMA was inconsistent over a large area of the
graph space, no matter which distance was used. The simulation analysi
s showed how tree-making methods perform given limited numbers of char
acter data. In some instances, the simulation results differed quantit
atively from the consistency analysis. The consistency analysis indica
ted that Lake's method of invariants was consistent over all of the gr
aph space under some conditions, whereas the simulation analysis showe
d that Lake's method of invariants performs poorly over most of the gr
aph space for up to 500 variable characters. Parsimony, neighbor-joini
ng, and the least-squares methods performed well under conditions of l
imited amount of character change and branch-length variation. By weig
hting the more slowly evolving characters or using distances that corr
ect for multiple substitution events, the area in which tree-making me
thods are misleading can be reduced. Good performance at high rates of
change was obtained only by giving increased weight to slowly evolvin
g characters (e.g., transversion parsimony, weighted parsimony). UPGMA
performed well only when branch lengths were close in length.