ANCESTRAL STATE ESTIMATION OF COMPLEX CHARACTERS: THE CASE OF FEATHER EVOLUTION

Ancestral state estimation is a powerful tool to explore the evolution of phenotypic complexity, a key goal of evolutionary biology. These statistical methods rely on three major components: a phylogenetic tree, the observed character states at the tips, and an evolutionary model describing how a trait changes through time. Estimating ancestral states from discrete character data remains poorly studied. This is especially true for complex characters, i.e. characters with dependent relationships between states. Analysing such characters poses many methodological challenges: Which coding strategy is the most appropriate? Which a posteriori time scaling method should be used? How should character dependencies be accounted for? Which model should be favoured? Here, we use feather evolution as a case study to explore different statistical methods in estimating ancestral states of complex characters. This is a suitable case study as data are known from both living and fossil taxa, and there are multiple ways in which character states can be distinguished and coded. The origin and early evolution of feathers is subjected to an intense debate. There are competing hypotheses concerning the homology of feathers and whether they evolved just once, or multiple times along the avian stem-lineage.

We studied the effect of diverse coding strategies, a posteriori time scaling methods, outgroups and models. Our results suggest that all of these, on their own or in association, influence the ancestral likelihoods and substantively change interpretation of feather evolution. The coding strategy providing the maximum amount of information on the trait studied should be identified. The outgroup and branch lengths should be carefully considered as well. The simplest model, equal-rates unordered, provides here results associated with the lowest uncertainty and one of the highest amounts of information. This model is in disagreement with the recent proposed hypothesis of a single point of origin of feathers, around 240–250 Ma. It is worth noting though, that its Akaike Information criterion (AIC) is not the lowest among the models tested. It is thus possible that the AIC may not be appropriate for model selection in ancestral state estimation. Ultimately, our results inform best practices for modelling complex discrete characters.