Impact of errors on cladistic inference: Simulation-based comparison between parsimony and three-taxon analysis

Abstract

Simulation-based and experimental studies are crucial to produce factual arguments to solve theoretical and methodological debates in phylogenetics. However, despite the large number of works that tested the relative efficiency of phylogenetic methods with various evolutionary models, the capacity of methods to manage various sources of error and homoplasy has almost never been studied. By applying ordered and unordered methods to datasets with iterative addition of errors in the ordering scheme, we show that unordered coding in parsimony is not a more cautious option. A second debate concerns how to handle reversals, especially when they are regarded as possible synapomorphies. By comparing analyses of reversible and irreversible characters, we show empirically that three-taxon analysis (3ta) manages reversals better than parsimony. For Brownian motion data, we highlight that 3ta is also more efficient than parsimony in managing random errors, which might result from taphonomic problems or any homoplasy generating events that do not follow the dichotomy reversal/convergence, such as lateral gene transfer. We show parsimony to be more efficient with numerous character states (more than four), and 3ta to be more efficient with binary characters, both methods being equally efficient with four states per character. We finally compare methods using two empirical cases of known evolution.