A cautionary note on the use of Ornstein Uhlenbeck models in macroevolutionary studies

Abstract

Phylogenetic comparative methods are increasingly used to give new insights into the dynamics of trait evolution in deep time. For continuous traits the core of these methods is a suite of models that attempt to capture evolutionary patterns by extending the Brownian constant variance model. However, the properties of these models are often poorly understood, which can lead to the misinterpretation of results. Here we focus on one of these models – the Ornstein Uhlenbeck (OU) model. We show that the OU model is frequently incorrectly favoured over simpler models when using Likelihood ratio tests, and that many studies fitting this model use datasets that are small and prone to this problem. We also show that very small amounts of error in datasets can have profound effects on the inferences derived from OU models. Our results suggest that simulating fitted models and comparing with empirical results is critical when fitting OU and other extensions of the Brownian model. We conclude by making recommendations for best practice in fitting OU models in phylogenetic comparative analyses, and for interpreting the parameters of the OU model.

The impact of phylogenetic dating method on interpreting trait evolution: a case study of Cretaceous–Palaeogene eutherian body-size evolution

“…Some authors have noted a tendency to erroneously recover strong support for OU models in simulated data where the true model is unbounded BM [22]. However, these effects are strongest in small trees and are indicated by low alpha parameters, neither of which occurs here.…”

Massive increase in visual range preceded the origin of terrestrial vertebrates

“…Bayesian OU instead agnostically infers whether changes in the selective regime of a trait occurred through evolutionary time and if so, along which branches in the phylogeny these changes likely happened. It is also less prone to error for small phylogenetic comparative datasets (N < 100) than other methods (79). We multiplied the traits (i.e., residuals from PGLS) by 10 to avoid computational issues, because the combination of the small-valued residuals and the timespan of more than ≈100 My frequently yielded infinite likelihoods.…”

Anyone with a Long-Face? Craniofacial Evolutionary Allometry (CREA) in a Family of Short-Faced Mammals, the Felidae

“…Some of these may require very large numbers of species to be accurate (Boettiger et al 2012;Cooper et al 2016) and indeed it has been suggested (Jhwueng 2013) that, for less than 100 species, at least in univariate analyses, there could be little reason to explore alternatives to BMs, as they tend to perform equally well and produce congruent results. However, to provide a simple preliminary assessment of the sensitivity of results to models other than BM, we repeated PGLS regressions after modifying branch lengths either by setting all branch lengths to unit (equivalent to a punctuated equilibrium model, where change occurs only during speciation events), or, following the example of Díaz-Uriarte & Garland, (1998), by changing Grafen's rho.…”