Phylogeny and trait data

We derived a maximum clade credibility phylogenetic tree for the entire tanager clade (n = 355 spp.) [49] from the posterior distribution of trees constructed in BEAST [50] using two mitochondrial markers (Cyt-b, ND2) and four nuclear markers (RAG1, FGB-I5, MB-I2, and AC01-I10). We also used the posterior distribution of trees to assess the robustness of our results to topological and dating uncertainty (S8 Fig). To simplify biogeographic reconstructions (see more details in the Subsetting likely interactions section below), and to focus on continental-scale processes, we excluded insular taxa and taxa lacking biogeographic data (n = 32), leaving a dataset restricted to continental taxa (n = 323).

We compiled measurements of ecological and social traits for as many species as possible. For resource-use traits, we collected data on beak length, width, and depth, as well as wing, tarsus, and tail length from museum specimens of tanagers (n = 319 species) using established procedures [30,51]. We also included published data on body mass (n = 322) [29]. For social traits, we compiled data on both visual signals (plumage coloration) and acoustic signals (song). Variation in plumage coloration were extracted from reflectance spectra measured separately on males (n = 312) and females (n = 314), with species differences calculated by comparing between overall coloration patterns as well as between particular plumage patches [31,52]. Acoustic data included song tempo (n = 294), song frequency (n = 294), and the temporal changes in frequency throughout songs (n = 287) [32,53]. All traits were measured on several individuals of each species where possible (mean ± standard deviation: resource-use traits = 5.35 ± 7.69 individuals, male plumage coloration = 4.27 ± 1.17, female plumage = 3.84 ± 1.32, song = 8.53 ± 7.24).

We constructed phylogenetic principal component axes in phytools [54,55]. Apart from two song variables (frequency and amplitude slopes), all data were ln-transformed prior to construction of the axes, and trait data for all taxa, including island taxa and taxa without biogeographic data, were included in the construction of the pPC axes. Rather than compiling all data into a single set of pPC axes, we grouped variables that describe functional variation in birds. In total, we constructed nine sets of pPC axes describing variation in resource-use (axes for beak morphology and locomotory traits), plumage (axes for male and female plumage coloration and complexity), and song (song tempo, song frequency, and note frequency) (S1 Table, S2 Table). We used pPCs, rather than untransformed raw trait data, as previous work has shown that these character axes describe functionally relevant traits [30]. Although we acknowledge the circularity of using pPCs in comparative analysis of traits, we note that using phylogenetically informed principal components and including all of the resulting axes should minimize the impact of known statistical issues arising from using PCs in comparative analyses [56]. In addition to the pPC axes, we also ran analyses on ln-transformed mass, which is a commonly used proxy for resource use in comparative studies [57].

Subsetting likely interactions

Because only species that coexist can interact, models that incorporate species interactions require identifying subsets of sympatric species [28]. To assign each lineage to a biogeographic region, we began by determining the region for each lineage as defined by ref. [37]. To ensure that the ancestral range estimation models were not too complex and thus too large to fit, we then collapsed these regions with >65% overlapping species and removed island taxa (Galápagos and other equatorial Pacific islands, south Atlantic islands, and Caribbean islands; n = 30), as well as two species endemic to the Madrean highlands, resulting in 10 final biogeographic regions (see Fig 1). In total, there are 323 species for which we had both phylogenetic and regional data. In BioGeoBears [33], we fit the dispersal-extinction-cladogenesis model of range evolution to the observed ranges of species. In this model, ranges change anagenetically as a function of dispersal and local extinction parameters and cladogenetically as a function of several range inheritance scenarios (see additional details in refs. [33,58]). Any attempt at ancestral range estimation over large macroevolutionary and spatial scales is bound to return a crude estimate of lineage presence and absence through time. To incorporate uncertainty in these reconstructions, we used the maximum likelihood parameter values for the DEC model to generate a bank of stochastic maps [59] using the function runBSM, each representing a hypothesis about where each lineage existed at a given point in time. These stochastic maps are largely congruent in which lineages they assign as sympatric (>75% congruent in the most recent 35 My) and thus largely represent uncertainty at much deeper timescales (S1 Fig). We also note that there is no reason why error in ancestral range estimation would impact distinct traits differently.

Additionally, competition is only likely between species that use similar resources or that come into close contact within their range. To estimate the likelihood of competition, we delineated three sets of potential competitors. We defined one set of species based on their basic habitat preferences [36] (S2 Fig). Specifically, each lineage was classified as occurring in dense (n = 128), semi-open (n = 143), or open habitats (n = 49). We defined another set based on diet using the EltonTraits database [29]. Each species was assigned to one of four of the following diet categories: invertebrate specialist (≥60% invertebrate diet, n = 80), frugivore (≥60% fruit diet, n = 59), granivore (≥60% seed diet, n = 66), or omnivore (<60% of either category, n = 115) (S3 Fig). We removed two species of nectarivores (Coereba flaveola and Diglossa gloriosissima) prior to running analyses. Following Tobias et al.[36], we defined a third set of species that defend permanent (year-round) territories (n = 35) (S4 Fig), on the grounds that these lineages are more aggressive, and thus more likely to interact in agonistic contexts than lineages defending only breeding territories. For each category, we fit Mk models of discrete character evolution to the tree, allowing transitions between each subcategory to occur at different rates. We then used the maximum likelihood Q matrix for each category to construct a bank of 50 stochastic maps using make.simmap in phytools [55], in which each stochastic map is a hypothesis about which guild each lineage belonged to at any given moment in evolutionary history.

Stochastic maps of biogeography were combined with stochastic maps of set membership to create a series of interaction matrices (A) defining which lineages are able to interact at each point along the phylogeny (S5 Fig). Thus, for a given time (t), if lineages i and j are both estimated to occur in the same region and to belong to the same set (e.g., insectivores), A(t)_i,j = 1. Otherwise, A(t)_i,j = 0. We then fit datasets to these trimmed trees by specifying that only lineages that are both sympatric and belong to the focal subgrouping (e.g., insectivores) can interact (lineages of nonfocal subgroupings evolve as if in allopatry, i.e., via BM). After calculating the expected variance-covariance matrix under this model, we then trimmed any remaining tips that do not belong to the focal subgrouping for subsequent likelihood calculation.

Although different partitions varied in the number of taxa with trait data (range = 32–143 spp.), there was no effect of sample size on the mean relative support for a model including interspecific competition (as defined in Fig 2; linear regression: effect of sample size on relative support index: t = −0.65, p = 0.52).

Model fitting

Rather than isolate a subset of pPC axes on which to run analyses, which can bias trait evolution models [60], we ran our analyses on all 26 pPC axes and body mass data. We ran the following two models that did not account for interspecific competition: the BM drift model [34] and the OU model of adaptation to an optimum trait value [35]. We also ran the following three models that accounted for interspecific competition: the MC model, in which the trait values of competing lineages are repelled from one another [28,61], and two diversity dependent models (DD_lin and DD_exp), in which evolutionary rates vary as a positive or negative linear or exponential function of the number of lineages in the reconstructed phylogeny [20,28]. In clades with high levels of extinction, inferences made on contemporary species can be misleading. Diversity-dependent models, in particular, can generate biased inferences when the number of reconstructed lineages does not accurately reflect the true historical diversity (e.g., if there are many extinct lineages the magnitude of slope parameters are underestimated [28]). Thus, we first verified that model estimates of diversity through time did not deviate from the observed number of reconstructed lineages through time (see Diversification analyses, below). We fitted BM and OU models using fitContinuous in geiger [62], and MC and DD models with interaction matrices (described above) using fit_t_comp in RPANDA [63]. We constrained the S parameter of the MC model to be negative to stabilize likelihood optimization, but we allowed the rate parameters of the DD models to be either positive or negative (reflecting rate increases [positive DD] or rate decreases [negative DD] with increasing species richness, respectively). For each species category, we selected the best model as the model with lowest corrected Akaike Information Criterion (AICc). The model selection procedure used here has reasonable power and Type I error rates [28], and simulations under a model of character displacement show that both MC and DD_exp can detect signatures of character displacement [18]. We also computed the Akaike weight of each model (denoted wi) and measured the relative support for a model with interspecific competition (MC, DD_lin, or DD_exp) as the relative Akaike weight of the best model with competition when compared to the best model without competition: max(MC_wi, DD_{lin_wi}, DD_{exp_wi})/((max(BM_wi,OU_wi)+max(MC_wi, DD_{lin_wi}, DD_{exp_wi})).

A series of other trait evolution models have been developed in the literature (reviewed in [64]), including models with evolutionary rates that vary continuously with time [57] or variation in environmental variables [65], with discrete shifts [66], and more complex OU models in which large clades are given their own rates and optima [24,67]. We did not consider the former because we wanted to focus on process-based models rather than phenomenological models [28], nor the latter because they serve to test other types of hypotheses, such as large-scale convergent evolution [23,24].

Diversification analyses

To assess whether model estimates of diversity through time and the number of lineages present in the reconstructed phylogeny are concordant, we fit several diversification models to the MCC tree using fit_bd in RPANDA [63]. Specifically, we fit (1) a pure-birth model with constant speciation and no extinction, (2) a pure-birth model with exponential variation in speciation and no extinction, and (3) a birth-death model with exponential variation in speciation and constant extinction. Both models with exponential variation in speciation outperformed the pure-birth constant rate model (model 2 versus 1, likelihood ratio = 40.97, p < 0.001; model 3 versus 1, likelihood ratio = 44.17, p < 0.001), but the model with extinction did not outperform the simpler model without extinction (model 3 versus 2, likelihood ratio = 3.20, p = 0.07). Thus, because extinction is estimated to be negligible, we conclude that standard diversity-dependent models that rely on the number of lineages in the reconstructed phylogeny are adequate for our analyses. We acknowledge that the lack of significant evidence for extinction likely represents the well-known low power to detect extinction in reconstructed phylogenies rather than a true absence of extinction [68] (see also discussion in [69]). However, there is no reason to expect that this would systematically bias inferences for one trait type (e.g., resource-use traits) or guild (e.g., frugivores) and not others.

Regression-based analyses

Although interspecific competition should drive diversification of resource-use traits, exploitative competition between species can actually result in adaptive convergence in signal traits; if the benefits to an individual of excluding heterospecifics from their territory are similar to the benefits of excluding conspecifics, then selection may favor convergence in traits that are used to mediate territorial interactions [17,70]. However, the process-based models of interspecific competition are not designed to detect trait convergence resulting from interspecific interactions [18]. Therefore, to test whether interspecific competition leads to trait convergence in competing lineages, we coupled linear regressions with phylogenetic simulations to test for a pattern of sympatric convergence [18,71].

In these models, we tested for convergence in (1) female plumage coloration pPC axes, (2) male plumage coloration pPC axes, and (3) song pPC axes. For each of these three groups of traits, we created a dissimilarity index by calculating the square-root transformed Euclidean phenotypic distance between all relevant pPC axes. To quantify sympatry, we calculated between-species range overlap using range data from BirdLife International and NatureServe [72]. We calculated the Szymkiewicz-Simpson coefficient (the area of overlap divided by the area occupied by the species with the smaller range) and created four indices of sympatry (one each for Szymkiewicz-Simpson coefficient threshold values of 5%, 20%, 50%, and 80%) [73]. Because convergence in sympatry can also result from selection favoring particular phenotypes in similar environments, we also included predictor variables indicating whether species pairs were found in the same habitat type or had the same diet. If, after controlling for similarity in habitat or diet, sympatry was negatively correlated with the dissimilarity index, this would be consistent with character convergence resulting from interspecific interactions.

To correct for phylogenetic nonindependence in these pairwise datasets, we conducted a phylogenetic simulation, which is more reliable than other methods for controlling for phylogenetic structure [18]. We simulated 5,000 datasets for each of the 27 traits using the ML parameter estimates for both (1) the BM model and (2) the OU model. We then conducted the same regression models on Euclidean distances between each of these simulated datasets to produce a phylogenetically informed null distribution of test statistics against which to compare our nonphylogenetic test statistic. We then calculated a two-tailed p-value using this distribution, which we subsequently corrected for multiple testing (using a critical value of 0.0125 because we conducted analyses on four sympatry indices [i.e., 0.05/4]).

Tempo of trait evolution

To estimate rates of evolution, we fitted BM models to the MCC tree and 100 posterior trees and the most complete trait datasets available (i.e., incorporating island and continental taxa, n = 313–353, see S2 Table). Because BM rate parameters are sensitive to changes in the scale of the trait values, we transformed all variables into z-scores so that all traits were on the same scale. We then ran ANOVAs to test whether the trait type (i.e., resource-use, female plumage, male plumage, and song) explained variation in the maximum likelihood estimate (MLE) of the BM rate parameter (σ²) on each of the 100 trees.

In addition to calculating the rates of trait evolution as the MLE of σ² in BM models, we calculated two other indices of evolutionary rates: (1) “felsens,” a measurement of divergence between sister taxa scaled by their divergence time [74], and (2) BM fits to untransformed data, incorporating ME as the standard error of each individual trait measurement where available (unknown values were then estimated directly as an additional parameter), in geiger [62]. We then fit ANOVAs to the ln-transformed felsen value and the ln-transformed estimate of the BM rate parameter (σ²), as above.

We further examined the tempo of trait evolution by analyzing how disparity accumulates through time for each trait [75]. Using the dtt function in geiger [62] with the average squared Euclidean distance among z-transformed trait values, we simulated 1,000 BM datasets to create a null distribution for the MCC tree and each of 100 posterior trees. We then calculated the MDI as the area between the null and observed disparity-through-time curves for each phylogeny, such that large MDI values indicate rapid and sustained accumulation of within-clade disparity greater than that expected under BM. As with rates of trait evolution, we ran ANOVAs on each posterior tree to test for an effect of trait type on the MDI value.

Measurement error

To identify the impact of ME on model fitting, we first compiled estimates of ME for each species and trait value.

For body mass, we obtained the mean and standard deviation following ref. [65]. Specifically, we obtained these values from the source data [76] where possible, calculating a standard deviation from the range of mass values when standard deviations were not reported (following [77]) or using the averaged standard deviation across species when a sample size and mean were the only values reported in ref. [76] or for cases where measurements were only made on a single individual [78] (n = 15). We then calculated the ME for each species using the formulae presented in ref. [65].

For plumage pPC axes, the data used in main analyses were calculated from reflectance spectra averaged across individuals, so we first recalculated the spectral indices used in our main analyses on each individual. For both plumage and song pPC axes, we projected individual-level data into pPC scores [54] and directly calculated the ME on these pPC scores (i.e., as |(x_individual-x_species.mean)/x_species.mean|) and then calculated the average ME for each species X trait combination. For downstream analyses of the impact of ME, we removed trait measurements with only one measurement (male plumage n = 16, female plumage n = 24, song n = 37).

For the remaining pPC axes describing resource-use traits, we did not have individual measurements, so we first simulated datasets comprising the same number of measurements as in the empirical dataset by sampling from normal distributions with means and standard deviations of empirical values. We then calculated the pPC scores and ME as above. Likewise, for downstream analyses, we removed trait measurements with only one measurement (n = 19).

To test for a potential effect of ME on our results, we computed the median ME for the subset of species and traits corresponding to each of our analyses. We then tested for a relationship between this median ME and the relative index of support for competition models using multiple linear regression analyses. Because increasing rates models may receive inflated support with high ME [48], we also analyzed the effect of ME on the slope parameter of the DD_exp model, which was commonly the best-fit model in our analyses.