Mammalian omnivores are a broad and diverse group of mammals that are often lumped together in ecological studies. As a result, there are open questions about what macroevolutionary and macroecological information can be gained when omnivorous dietary differences are investigated in ecological and phylogenetic comparative studies. In this study, we investigate the frequency at which vertebrate prey, invertebrate prey, fibrous plant material, and non-fibrous plant material co-occur in the diets of omnivorous species. We quantify the body size distributions and phylogenetic signal of terrestrial mammals that consume different omnivorous diets and using multistate reversible jump MCMC, we assess the transition rates between mammalian diet strategies on the mammalian phylogenetic tree. We find that omnivores that consume all four food types are rare and most omnivorous mammals consume only invertebrate prey and non-fibrous plants. We also find that omnivores that only consume invertebrate prey are on average smaller than omnivores that incorporate vertebrate prey as many are from within Rodentia. Our transition rate models show that there are high transition rates from invertivorous omnivory to herbivory, and from vertebrate predation to prey mixing and ultimately invertivory. Our results suggest that prey type is an important aspect of omnivore macroevolution and macroecology, as it is correlated with body mass, evolutionary history, and diet-related evolutionary transition rates.

Dataset and phylogenetic tree

Using previously published datasets, we compiled diet data and body masses for 1437 extant terrestrial mammals (about a quarter of all mammal species) (Supplemental data 1). Aquatic mammals, dependent on food webs with a dramatically different structure than their terrestrial counterparts, are expected to experience different ecological and evolutionary pressures than terrestrial mammals and have been excluded from this analysis. We chose to use a dataset compiled from primary literature sources generated previously by some of the authors (Price et al. 2012). This dataset did not use dietary inferences based on phylogeny or morphology, only using studies that reported stomach contents or cheek-pouch contents, the contents of food stores, direct behavioral observations, or fecal analysis. While this dataset contains fewer species, and uses a categorical schema, we use it because of the high diet resolution and because it allows us to directly compare our results to those previously published by Price et al. 2012. The dataset has broad taxonomic and phylogenetic representation and good coverage of all the mammalian orders (Supplemental data 1). As it was compiled, they prioritized species from underrepresented families in the diet literature, as opposed to obtaining exhaustive samples of well-represented genera. The data in Price et al. 2012 were analyzed using the three basic trophic categories, however, it was originally collected using more detailed categories and dietary descriptions. Food materials were split into four food categories: invertebrate prey, vertebrate prey, fibrous plant parts (mature leaves, stems, wood, and bark), and non-fibrous plant parts (any other parts of plants). A category was scored as present in the diet of the organism if, in the primary literature, it constituted at least 5% of the food consumed by volume, weight, or feeding time. Body masses for omnivorous species were gathered from the PanTHERIA database. For all phylogenetically-informed analyses, we used a fully resolved set of phylogenetic trees from Faurby and Svenning 2015.

Phylogenetic signal

We calculated the phylogenetic signal of each diet category, treating each diet category as a binary trait over ten randomly selected trees with the phylo.d function in the caper package in R. This method calculates a D statistic which is close to 1 if the observed trait has a phylogenetically random distribution or 0 if the observed trait is dispersed on the tree in a way that is consistent with a threshold model of Brownian motion evolution. The trait distribution for the Brownian motion model is calculated by simulating a continuous trait along the phylogeny, defining a threshold value that ensures that the number of tips with each character state remains the same as in the original dataset, then defining the character state at each tip using the threshold value of the continuous trait. Values lower than 0 indicate phylogenetic clustering beyond what is expected by the Brownian motion threshold model. The phylo.d function also tests for significant departure from both a phylogenetically random distribution and the phylogenetic distribution generated under the threshold model.

Transition rates

We calculated transition rates between the seven dietary guilds using Bayesian Markov Chain Monte Carlo (MCMC) methods in the program BayesTraits. Specifically, a multistate reversible jump MCMC was used to estimate transition rates without assuming a single model of trait evolution. Reversible jump MCMC explores all possible models and generates a posterior distribution of models and parameter estimates by setting each transition rate parameter to either a unique value, equal to one or more of the other transition rates, or zero. This process allows for the exploration of loglikelihood especially when there is a large number of possible models. Because this is computationally intensive, all BayesTraits analyses were run on the University of Oregon Talapas High Performance Computing cluster.

To take into account variability in tree topology we ran independent chains on 100 randomly selected fully resolved trees. We used uniform hyperpriors to seed the exponential prior on the parameters, seeding the mean of the exponential prior from a uniform distribution on the interval 0 to 10 or 0 to 2. To ensure stationarity was reached each chain was run for 1 billion iterations with a sampling interval of 300,000 and a burn-in of 100,000 iterations. We examined the effective sample sizes, autocorrelation, and convergence using packages coda and btw in R (Supplemental data 23). We also checked the autotuning mechanism by examining schedule files to make sure the chains were mixing appropriately. The medians and interquartile ranges (IQR) were then calculated for each transition rate along with the frequency with which a transition rate was reconstructed as zero (% Z). To investigate whether the differences in transition rates were meaningful, we ran the same analyses on a tree with randomly reassigned dietary categories. We produced a random dataset in R using the sample function on our existing data to guarantee the same number of individuals in each dietary guild. We then used the same reversible jump MCMC procedure in BayesTraits to calculate median transition rates, % Z, and model posterior distribution. This allowed us to determine whether our observed results differed from those expected when there is no phylogenetic signal in dietary guilds.

Additionally, to accurately compare our transition rates to those estimated by Price et al. 2012, which used a different phylogeny and methods, we used the same reversible jump MCMC procedure in BayesTraits to estimate transition rates between the three main trophic categories (herbivory, carnivory, omnivory). We ran ten independent chains using 10 randomly selected trees. Furthermore, following the procedures of Price et al. 2012, we conducted analyses using the “Multiple State Speciation Extinction” (MuSSE) model in the diversitree package on our seven dietary guilds. We estimated model fit using maximum likelihood instead of Bayesian MCMC methods because of computational constraints. We ran two models one with the speciation and extinction rates constrained to those reported in Price et al. 2012 and one with no constraints. We incorporated the sampling frequency to be 0.224 of all Mammalia as the current reported number of mammalian species is 6,495. We ran each model on 10 randomly selected trees.