In this letter we put forth topological bias as a potential problem in interpreting simulations comparing different phylogeny reconstruction algorithms. In particular, a long-branch-attracts bias can have effects even when there are not two obviously long branches present. This phenomenon can be used to explain previous results by Yang (M.B.E. 14:105-108, 1997) which appeared to show ML using the wrong model to outperform ML using the correct model. We show that ML using the wrong (overly homogeneous) model is quantitatively biased toward joining long branches on unresolved four-taxon trees.
Full length postscript is here.