The “long-branch attraction” (LBA) phenomenon in phylogeny reconstruction is well cited but its causes have been poorly characterized. In this article, we point out that different biological processes can lead to similar forms of long-branch attraction. That is, although sequences generated by different processes look similar “through the eyes of parsimony,” the ensemble of sequence site patterns (not just the parsimony sites) can distinguish between these processes. In 1978, Felsenstein described an evolutionary scenario under which unequal amounts of change in nonadjacent lineages would mislead tree-building methods based on parsimony (or on uncorrected distances). Other authors have since shown that when substitution models are misspecified, maximum likelihood and distancebased methods can be similarly misled (e.g., Hillis et al., 1994; Lockhart et al., 1996; Bruno and Halpern, 1999; Swofford et al., 2001; Sullivan and Swofford, 2001; Ho and Jermiin, 2004). LBA problems may also arise because of sparse and/or unbalanced taxon sampling (Hendy and Penny, 1989; Holland et al. 2003; Lockhart and Penny, 2005) and/or because of lineage-specific differences in rates or processes of evolution (e.g., Hasagawa and Hashimoto, 1993; Steel et al., 1993, 2000). Felsenstein (1978) assumed that whilst evolutionary rates varied across a tree, individual sites in sequences could be ascribed a rate of change that was the same at other sites in the same sequence. That is, if a lineage was fast (or slow) evolving, then the evolution of sites in a sequence belonging to that lineage was also fast (or slow). A site position that has evolved under this scenario can be seen as a special case of “heterotachy.” This is a property of individual sequence positions, which literally means different speeds. It is the concept of sequence evolution at a given site undergoing substitution at different rates in different parts of the tree (Lopez et al., 2002). Interestingly, Simon et al. (1996) have also described this phenomenon and referred to it as “mosaic evolution.” It is important to note that variation in the substitution rate of a site throughout the tree is distinct from rate variation across sites (as modeled, for example, by a gamma distribution). In the latter case there is a site-specific substitution rate that varies randomly across the sites, but at any site it applies equally to all the branch lengths of the tree (the branch lengths at the site are all multiplied by the site-specific rate). Consequently, the ratio of substitution rates on two different branches is constant across sites in such models, even when one allows both rate variation in the tree (as in Felsenstein’s scenario) as well as an independent process of rate variation (e.g., gamma distribution) across sites. In contrast, with more general forms of heterotachy, the ratio of substitution rates on different branches of the tree may vary across sites. The special case of heterotachy assumed by Felsenstein (1978) is different from another very special type of heterotachy recently explored in simulations by Kolaczkowski and Thornton (2004). These authors envisaged a four-taxon tree, for which the external lineages evolved in such a way that sites in the sequences accumulated substitutions at one of two rates, either slow or fast. As Spencer et al. (2005) point out, the frequencies of patterns expected under the simulation model studied by Kolaczkowski and Thornton (2004) are a small snapshot of the full range of possibilities when all possible combinations of short and long branches are considered, and most of these do not cause LBA (further concerns regarding the findings of Kolaczkowski and Thornton [2004] have been discussed by Steel, 2005). Thus, the patterns that Kolaczkowski and Thornton (2004) studied are different from those expected under the standard stationary covarion (or “covarion drift”) models, which have been the subject of much recent study (e.g., Tuffley and Steel, 1998; Penny et al., 2001; Gaucher et al., 2001; Huelsenbeck, 2002; Galtier, 2001; Misof et al., 2002; Inagaki et al., 2004; Ane et al., 2005; Guindon et al., 2004). These standard covarion models have reversible stationary substitution rates among character states that are switched “on” (variable), and a reversible stationary process between the state of “off” (invariable) and “on.” The latter condition will maintain