Phylogenetically informed predictions outperform predictive equations in real and simulated data

Abstract

Inferring unknown trait values is ubiquitous across biological sciences—whether for reconstructing the past, imputing missing values for further analysis, or understanding evolution. Models explicitly incorporating shared ancestry amongst species with both known and unknown values (phylogenetically informed prediction) provide accurate reconstructions. However, 25 years after the introduction of such models, it remains common practice to simply use predictive equations derived from phylogenetic generalised least squares or ordinary least squares regression models to calculate unknown values. Here, we use a comprehensive set of simulations to demonstrate two- to three-fold improvement in the performance of phylogenetically informed predictions compared to both ordinary least squares and phylogenetic generalised least squares predictive equations. We found that phylogenetically informed prediction using the relationship between two weakly correlated (r = 0.25) traits was roughly equivalent to (or even better than) predictive equations for strongly correlated traits (r = 0.75). A critique and comparison of four published predictive analyses showcase real-world examples of phylogenetically informed prediction. We also highlight the importance of prediction intervals, which increase with increasing phylogenetic branch length. Finally, we offer guidelines to making phylogenetically informed predictions across diverse fields such as ecology, epidemiology, evolution, oncology, and palaeontology.