Chronogram or phylogram for ancestral state estimation? Model-fit statistics indicate the branch lengths underlying a binary character's evolution

Modern methods of ancestral state estimation (ASE) incorporate branch length information, and it has been demonstrated that ASEs are more accurate when conducted on the branch lengths most correlated with a character's evolution; however, a reliable method for choosing between alternate branch length sets for discrete characters has not yet been proposed. In this study, we simulate paired chronograms and phylograms, and generate binary characters that evolve in correlation with one of these. We then investigate (a) the effect of alternate branch lengths on ASE error and (b) whether phylogenetic signal statistics and/or model-fit statistic can be used to select the branch lengths most correlated with a binary character. In agreement with previous studies, we find that ASEs are more accurate when conducted on the branch lengths most correlated with the character. Phylogenetic signal statistics show limited utility for selecting the correct branch lengths, but model-fit statistics are found to be more accurate, with the correct branch lengths generally returning greater model-fit (lower AICc and BIC values). Using this method to choose between alternate branch length sets is more accurate when tree and character properties are more favourable for model optimization, and when shape differences between alternate phylogenies are greater. Our results indicate that researchers conducting ASEs on discrete characters should carefully consider which branch lengths are appropriate, and, in the absence of other evidence, we suggest estimating model-fit values over alternate branch length sets and evolutionary models and choosing the branch length/model combination that returns better model-fit.

Automated summary

Research suggests that selecting branch lengths most correlated with a character can enhance accuracy in ancestral state estimation. Phylogenetic signal statistics have limited utility in choosing the correct branch lengths, while model-fit statistics are more accurate.