Meeting Abstract
Phylogenetic linear models (e.g., regressions, ANOVA, or ANCOVA) provide a statistically rigorous framework for comparative studies of phenotypic traits across taxa. However, the development of their multivariate counterparts is still lagging behind because of the computational challenges encountered with multidimensional datasets. In particular, when the number of traits p approach or exceed the number of taxa n, the conventional statistical machinery is limited, and we have to rely on alternative methods that are approximate and restricted to the Brownian motion model of trait evolution. Here we developed more flexible multivariate phylogenetic linear models (e.g., multivariate regressions, MANOVA, MANCOVA) to deal with the high-dimensionality of modern high-throughput comparative datasets. We used intensive simulations to assess the performances of the proposed approaches to various level of phylogenetic signal, of correlations between the traits, and distributions of phenotypic changes in the multivariate space. We show that the proposed approaches outperform conventional ones when p, and current alternative when p>n. We further show that current available approaches to deal with high-dimensional datasets lack the power to detects differences in multivariate datasets and may have high type I error rates. Finally, we provide an empirical test of our phylogenetic MANOVA on a geometric-morphometric dataset describing the mandible morphology in phyllostomid bats along with data on their diet preferences. Overall our results show significant differences between ecological groups while accounting for the mild phylogenetic signal of these ecomorphological data. We provide some guidance on the use of multivariate statistics for comparative analysis and discuss some recent concerns about the use of phylogenetic comparative methods.