A comparison of ancestral state reconstruction methods for quantitative characters
Choosing an ancestral state reconstruction method among the alternatives available for quantita- tive characters may be puzzling. We present here a comparison of five of them, namely the maximum likelihood, restricted maximum likelihood, generalized least squares, phylogenetic independent con- trasts and squared parsimony methods. A review of the relations between these methods shows that the first three ones infer the same ancestral states and can only be distinguished by the distributions accounting for the reconstruction uncertainty which they provide. The respective accuracy of the methods is assessed over character evolution simulated under a Brownian motion with (and without) drift. We start by giving the general form of ancestral state distributions conditioned on leaf states under the simulation model. Ancestral distributions are used first, to give a theoretical lower bound of the expected recon- struction error, and second, to develop an original evaluation scheme which is more efficient than comparing the reconstructed and the simulated states. Our simulations show that: (i) the methods do not perform well as the evolution drift increases; (ii) the maximum likelihood method is generally the most accurate and (iii) not all the distributions of the reconstruction uncertainty provided by the methods are equally relevant.