Contrasting drivers of diversification rates on islands and continents across three passerine families

Abstract

Diversification rates vary greatly among taxa. Understanding how species-specific traits influence speciation rates will help elucidate mechanisms driving biodiversity over broad spatio-temporal scales. Ecological specialization and range size are two hypothesized drivers of speciation rates, yet each mechanism predicts both increases and decreases in speciation. We constructed a continuous index of specialization using avian bill morphology to determine the relative effect of specialization and range size and shape on speciation rates across 559 species within the Emberizoidea superfamily, a morphologically diverse New World clade. We found a significant positive correlation between specialization and speciation rate and a negative correlation with range size. Only the effect of specialization persisted after removing island endemics, however, suggesting that ecological specialization is an important driver of diversity across large macroevolutionary scales, and the relative importance of specific drivers may differ between islands and continents.

1. Introduction

Biodiversity varies widely among clades, and speciation rates vary over time, among regions and across taxa [1–5]. The recent availability of well-resolved phylogenies has heightened interest in identifying factors that drive variation in diversification rates over macroevolutionary scales (e.g. [4–6]). Both geographical and ecological mechanisms can drive speciation. Geographical vicariance can divide a species range into genetically isolated populations that can influence speciation [7–9]. New ecological opportunities or innovations that allow species to invade new niches can also drive divergence and produce clades of specialized taxa [10].

While specialization following expansion into novel niche space plays a role in adaptive radiations [3,11,12], most examples come from islands or lakes (but see [13]), which have a particular geographical arrangement. While island radiations provide insights into rapid diversification, geographical isolation is often confounded with ecological opportunities in these settings [14]. Thus, it remains unclear how important specialization and geography are for driving diversification over broad spatial scales.

Moreover, geographical range size and ecological specialization can both positively and negatively affect speciation rates [15–18]. Ecological divergence can promote rapid diversification by partitioning a single ancestral niche into multiple, smaller niches occupied by genetically isolated populations, resulting in clades of specialized taxa [10,19–21]. If specialization partitions niche space, however, speciation rates may slow as species accumulate and niches fill [1,3,22,23]. Moreover, specialized taxa with narrow ecological niches may become constrained in their ability to adapt to new ecological opportunities [15,24] and may have smaller population sizes, which together increase extinction risk and decrease the probability of advantageous mutations for further adaptation [25,26]. Thus, specialization could explain increases in the diversification rate early during adaptive radiation and slower diversification rates later on.

Range size and shape are also predicted to influence speciation rates in contradictory directions [7,16–18,27]. Larger or more elongated ranges can increase speciation potential owing to a higher probability of encountering new habitats or dispersal barriers that isolate populations [16,17,27–29]. Large ranges also allow for larger population sizes, increasing the probability of mutations driving divergent adaptation in different environments [26]. However, range size may also be correlated with dispersal ability, as longer dispersers are more likely to cross habitat gaps. High dispersal ability should reduce divergence by maintaining the gene flow [6,9,30,31]. Similar to specialization, geographical partitioning may slow diversification as species accumulate. High rates of geographical partitioning during speciation may produce numerous small-range taxa that experience higher extinction rates [32] and have smaller population sizes with lower adaptive capacities. Thus, range partitioning during speciation could also explain increased diversification early during radiation and slower rates later on.

Our objective is to determine the effect of specialization and range size on speciation rates and determine if the processes are similar on islands and continents. We investigate these questions across three avian families within Emberizoidea (Icteridae, Passerellidae and Thraupidae). This group is ecologically diverse and contains multiple independent origins of specialized bill shapes, from long curved bills of nectivores (e.g. Diglossa, Cyanerpes, Coereba) to short, stout, seed-crushing bills of granivores (e.g. Geospiza, Oryzoborous) [33,34]. Species within this clade occur across the Americas and associated islands and include small-range endemics and species whose ranges span continents.

While specialization has multiple definitions [35], avian bill morphology is a well-known functional trait, and the relationship between morphological diversification and ecological specialization is well documented (e.g. [12,36–38]). Although recent studies indicate that diet alone is a poor predictor of variation in bill morphology [39,40], passerine radiations frequently involve bill characteristics that are functionally related to use of food resources [2,3,12]. We developed a continuous metric of bill specialization that reflects the outcome of selection and avoids relying on subjective categorical variables. We use a quantitative morphometric approach with higher statistical power to detect shape differences than linear measurements [41,42]. By combining our specialization index with species geographical distributions, we aim to determine the relative effect of ecological specialization and geographical distribution on speciation rates across islands and continents.

2. Methods

(a). Taxon sampling

We sampled taxa from three families (Passerellidae, Icteridae and Thraupidae) within Emberizoidea, an ecologically and morphologically diverse clade that is distributed across the Americas and associated islands. We sampled adult males and females and included specimens from different subspecies and across the breeding range of each species when possible to account for intraspecific variation.

(b). Quantifying specialization using morphometrics

We photographed 2831 specimens (1–10 individuals per species, mean = 5) of 565 species by positioning each study skin with the bill laid laterally against a ruler under a tripod-mounted camera. We digitized specimens using TPSdig2, v. 2.30 [43], used the ruler to set the scale and outlined bill shapes by placing five homologous landmarks and three curves, each containing eight semi-landmarks (electronic supplementary material, figure S1). We included the maxilla and mandible to provide a more complete measure of bill shape. Although specimen preparation or desiccation could introduce distortion that may bias our measurements, we attempted to minimize this by visual inspection when selecting specimens to photograph.

We analysed the landmark data in the R package Geomorph v. 3.0.6 [44,45]. We converted the landmark data to shape information using generalized Procrustes analysis, using the gpagen function. The shape variables are residuals after size, position and orientation are removed from the shape coordinates. We used the distance-minimizing approach for curve sliding, which is less likely to introduce variance than the bending-energy approach [46]. We obtained the least-squares mean Procrustes distance for each species using the advanced.procD.lm function to account for differences in sample size and used the morphol.disparity function to estimate a proxy of morphological specialization. This index estimates the Procrustes variance using residuals of a linear model fit. Procrustes variance is the average Procrustes distance of each species' shape relative to the mean shape of measured specimens [47], representing the position of each species in multidimensional morphological space relative to the grand mean consensus shape of species included in the analysis. This creates a continuous proxy for specialization, where bill shapes farthest from the grand consensus mean along any morphological axis are the most specialized. We assume that specialization in one morphological direction comes at a cost to specialization in another direction. Thus, those farthest from the centre of morphological space may be more restricted in their functional abilities, and the grand mean consensus shape is an estimate of the shape of the most generalized bill representing a compromise among functional attributes of bill morphology. While this relative measure depends on species included in the analysis, we create a continuous index that measures the outcome of multiple selection pressures and avoids relying on arbitrarily defined dietary categories. We visualized major axes of morphological variation to confirm that we captured ecologically relevant specialization by plotting each species mean in tangent space and examining shape change along principal component (PC) axes.

(c). Range size

We obtained geographical distributions of our focal species from BirdLife International [48]. We obtained geographical ranges for 559 of the 565 species we digitized. These 559 species form the basis of all our trait analyses. We included only breeding and resident ranges to avoid introducing bias between migratory versus non-migratory species and only used the data from within native ranges to avoid including areas where species were introduced. We calculated area and perimeter in ArcGIS. We defined range shape as the total range area divided by the total perimeter for each species. Species with smaller values have more disjunct or elongated ranges with greater potential for decreased gene flow among isolated or edge populations. We denoted island endemics as species that only occur on islands and those occurring on both continents and islands as continental species. We defined islands broadly, by including oceanic and continental shelf islands. Although distance from the mainland could influence speciation dynamics, our sample size did not allow us to adequately test for differences among island groups.

(d). Speciation rate estimation

We used time-calibrated trees from a published phylogeny of Emberizoidea [34] to estimate speciation rates using two methods across the entire tree (including all 795 tree tips). First, we calculated species-specific rates using the inverse of the equal splits measure, or diversification rate (DR) statistic, which estimates tip-specific rates of diversification without a formal parametric model [4,49]. The equal splits measure is the sum of the edge lengths from each tip to the root of the tree, with each consecutive edge towards the root weighted by a factor of 1/2. The inverse is interpreted as the speciation rate. We calculated rates across all species using the maximum clade credibility (MCC) tree and average rates across 100 trees sampled from the pseudoposterior of the previous study [50].

Lineage-specific measures capture subtle rate shifts but may show rate heterogeneity when there is no variation in diversification rates [51]. Thus, we also used a tree-wide, model-based approach. Bayesian analysis of macroevolutionary mixtures (BAMM) [52] uses a reversible-jump Markov chain Monte Carlo (MCMC) approach to estimate the diversification rate shifts across branches of a phylogenetic tree and produces accurate estimates with low error [51]. We performed four runs for 50 million generations, allowing for time-heterogeneous speciation rates. We accounted for incomplete taxon sampling by supplying a global sampling fraction of 95%, as reported in the study by Barker et al. [34]. We estimated priors using setBAMMpriors in BAMMtools [53]. We discarded the first 10% as burn-in and verified convergence by inspecting stability in log-likelihood scores and the location and number of rate shifts across multiple runs. We also confirmed the effective sample sizes for the log-likelihoods and shifts exceeded 1000 using the CODA package [54]. We extracted mean speciation rates for each species using the getTipRates function and summarized outputs using BAMMtools v. 2.1 [53].

(e). Testing for trait-dependent speciation

We treated our measure of specialization, range size, range shape and island endemism as traits. We assessed covariation among traits using the pgls function in the caper package. We log-transformed continuous traits to improve normality. We used several methods to assess whether our trait values covaried with speciation rates for the subset of 559 species for which we had the trait data. First, we used two semiparametric simulation-based methods (ES-sim and fast, intuitive state-dependent speciation-extinction (FiSSE)) to test for a significant correlation between our trait values and speciation rates. Second, we used a model-based approach (multistate hidden-state speciation and extinction, MuHiSSE) that jointly estimates rates of trait evolution, speciation and extinction to determine if speciation rates differed between specialists and generalists and between island and continental taxa. We omitted range size in these last analyses because ancestral state reconstructions and assumptions of Brownian motion built into the SSE models do not account for the process of range splitting during the speciation process.

ES-sim is a non-parametric method that tests for correlations between each trait and the DR statistic by determining the extent to which the correlation deviates from a simulated null distribution [55]. This method is robust to pseudoreplication, has high power and has a low false positive rate [55]. To construct the null distribution, we simulated trait evolution 1000 times across the tree using root state and diffusion rate (σ²) parameters from the maximum-likelihood fit of an Ornstein–Uhlenbeck (OU) model to the trait data. Model comparison using the fitContinuous function in the Geiger package [56] showed an OU model fit our trait data better than Brownian motion. We performed these analyses on the entire dataset and separately for island and continental species to determine the influence of insularity on the associations between speciation rates with geographical range and specialization.

For our binary trait of island endemism, we used FiSSE analysis to test for significant differences in speciation rates between islands and continents. FiSSE is a non-parametric test that compares the distribution of branch lengths with and without a binary trait and compares the value of the test statistic to a simulated null distribution [57]. We did not test for trait-dependent diversification using BAMM estimates, because existing methods are limited by the number of rate regimes detected [58], and we only detected five rate shifts (electronic supplementary material, figure S3).

We used MuHiSSE [59] as a model-based approach to further evaluate the effect of specialization and island endemism on speciation rates and to determine whether the effect of specialization varied between islands and continents. MuHiSSE estimates differences in speciation, extinction and transition rates as a function of multiple discrete character states and allows for inclusion of unobserved hidden states to account for variation in diversification rates owing to factors other than the focal traits [59–61]. Thus, the null model incorporates variable rates that are independent of the trait of interest and an equal number of parameters as a trait-dependent model (i.e. similar complexity [61]). To test for trait-dependent speciation, and interactions among focal traits, we constructed four trait-dependent models. We allowed turnover to vary: (i) only among island and continental taxa, (ii) only among specialists and generalists, (iii) among both observed states, and (iv) among both observed and hidden states. We constructed null models allowing rates to vary only among two, four or eight hidden states, matching the number of speciation rate parameters in our trait-dependent models. Because extinction is difficult to infer from phylogenetic data [62], we kept extinction fractions constant and allowed transition rates to vary. We estimated confidence intervals from the two log-likelihood support regions around our estimates using the function SupportRegionMuHiSSE. Because MuHiSSE models only allow for binary traits, we defined specialists (1) as those above the 60th percentile in our specialization index and all others as generalists (0). We compared fits using Akaike information criterion (AIC) [63].

We used two additional approaches to investigate if a particular bill shape, rather than specialization per se, influenced speciation rates. This tests whether any relationships between diversification and specialization in the first modelling steps may be explained by the influence of a particular type of specialization. In this case, we would expect speciation to be a linear function of a particular axis of shape variation. We used quantitative state speciation and extinction (QuaSSE) [64] to test for evidence of a linear or unimodal function using our first PC axis, which identified variation in bill length and thickness (short stout versus long thin bills). QuaSSE evaluates trait evolution and speciation simultaneously, where speciation rates vary as a function of a continuous trait. We quarter-power transformed our PC values to improve normality and constructed six likelihood functions where speciation is a constant, linear or unimodal function of bill shape as described by the first PC axis. We included the drift parameter to account for directionality of trait evolution in our data. We included a vector of variance estimates for each species and a sampling fraction of 0.95.

We constructed multi-state speciation and extinction (MuSSE) models to test for the same relationship in a categorical framework, allowing for complex relationships between speciation rate and bill shape described by the first PC. We divided taxa into three groups using the first PC axis: species with scores in the lowest 20th percentile (long thin bills), the middle 60th percentile (generalists) and the highest 20th percentile (short stout bills). We repeated this analysis using the 70th percentile to define our generalist category. We obtained speciation rate estimates for generalists and the two types of specialists to determine whether particular bill shapes disproportionately affect speciation rates among specialist taxa. We used Bayesian MCMC methods to examine uncertainty in model parameters. We implemented QuaSSE and MuSSE models in the diversitree package in R [65].

3. Results

(a). Quantitative traits

Our index of specialization ranged from 0.0007 to 0.1220. Figure 1 shows locations of representative bill shapes along the first two PC axes, which account for 70.2% and 12.5% of the variation, respectively. Species with high specialization values included those with large stout bills that specialize on large seeds (e.g. Melopyrrha nigra, Sicalis taczanowskii, Geospiza magnirostris and the genus Oryzoborous), long curved bills of nectar specialists in the genus Cyanerpes and recurved flower-piercing bills in the genus Diglossa. Smaller finch-like and sparrow-like bills with more generalist diets, such as members of Poospiza, Aimophila and Atlapetes, had lower specialization values and were located in the centre of the morphospace. Geographical range size varied from 2.5 km² to 1.5 × 10⁷ km². Many of the smallest-range species are island endemics (e.g. Nesospiza wilkinsi), while the largest range species, Volatina jacarina, is distributed across Central and South America and the Caribbean Islands.

Figure 1.

Locations of species in morphospace along PC axes 1 and 2 with warp grids showing major axes of variation in bill shape. (Online version in colour.)

(b). Speciation rate

We found heterogeneity in speciation rates across all 795 species in the MCC tree and also among our 559 study species in three families within Emberizoidea (figure 2; electronic supplementary material, figure S2). Speciation rates across the full MCC tree based on the DR statistic ranged from 0.05 to 6.26 species Myr⁻¹ (median = 0.26, mean = 0.42 species Myr⁻¹). We obtained similar results when averaged across 100 trees sampled from the posterior distribution (range 0.05–7.04, median = 0.26, mean = 0.42). The effective sample size of the log-likelihood and number of shifts in each sample confirmed BAMM convergence. The maximum posterior probability inferred five rate shifts (electronic supplementary material, figure S3). The mean tip speciation rates inferred from BAMM ranged from 0.12 to 1.81 (mean = 0.25; median = 0.12). Speciation rate estimates from BAMM and the DR statistic across all 795 species in the tree, and the subset of 559 species with the trait data, were correlated (all species: r = 0.66, p < 0.001; study species: r = 0.70, p < 0.001; electronic supplementary material, figure S2).

Figure 2.

Phylogenetic tree of 559 study species, with branch colours representing the log of the inverse splits, or DR statistic. Bars represent range size (inner ring) and specialization (outer ring). Illustrations by Gabrielle Cilfone. (Online version in colour.)

(c). Comparative analysis

Our specialization index was not correlated with range size (R² < 0.01, p = 0.60) or shape (R² < 0.01, p = 0.23). Island taxa had smaller ranges (R² = 0.22, p < 0.01), but were not more specialized than continental taxa (R² < 0.01, p = 0.59).

ES-sim results suggest specialization and geographical distributions influence speciation rates across broad spatial scales (table 1). We found a significant positive correlation between specialization and the DR statistic (r = 0.17, p = 0.03) and a significant negative correlation with speciation and both range size (r = −0.19, p = 0.001) and shape (r = −0.13, p = 0.03). These results suggest that faster speciation rates are associated with modern taxa with a higher degree of specialization, smaller range size, and more elongated or disjunct spatial distributions. We also found that speciation rates were higher on islands than continents (r = 0.27, p < 0.001). FiSSE results confirmed different speciation rates for island and mainland species (p = 0.01). After removing island endemics, specialization was still positively correlated with speciation rates (r = 0.15, p = 0.05), but range size and shape were no longer significant. We did not detect significant correlations with any of the three variables among island endemics, but our sample size was small (n = 43 species).

Table 1.

ES-sim results from the MCC tree (results were similar averaged over 100 trees sampled from the pseudoposterior). (We log-transformed continuous traits prior to analysis. The values for trait-rate correlations are Pearson's r, and p-values are one-tailed tests for a significant trait-rate correlation.)

		Pearson's r	p-value
all species	specialization	0.17	0.03
	range area	−0.19	0.001
	range shape	−0.13	0.03
	island	0.27	<0.001
continental species	specialization	0.15	0.05
	range area	−0.05	0.20
	range shape	0.01	0.44
island endemics	specialization	−0.01	0.47
	range area	−0.16	0.23
	range shape	−0.24	0.13

MuHiSSE results suggest that both specialization and island endemism are important drivers of speciation. The best-supported model suggests an interaction between island endemism and specialization (table 2). Speciation rates on continents were similar or higher among specialists than generalists, but on islands rates were similar or higher for generalists (figure 3; electronic supplementary material, table S1).

Table 2.

MuHiSSE model selection results. (ΔAIC is the difference in AIC values. Parameters indicate the number of diversification parameters. We defined specialists as greater than the 60th percentile in our specialization index.)

model	parameters	ΔAICa
island + specialization + hidden	8	0
island	2	16.9
island + specialization	4	30.4
null (four hidden states)	4	39.4
specialization	2	44.9
null (two hidden states)	2	59.4
null (eight hidden states)	8	64.1

^aThe lowest AIC value was 3283.4.

Figure 3.

Distribution of speciation rate estimates from MuHiSSE averaged across hidden states. Circles with error bars represent the mean and s.d., while carpet points indicate tip-specific rates, shown as waiting times in millions of years. I, island endemism; C, continental species; S, specialists; G, generalists. Adapted from Nakov et al. [59].

Our best-supported QuaSSE model suggested speciation was a unimodal function of bill shape described by the first PC axis (table 3). This supports the role of specialization, rather than a specific bill shape in driving increased speciation rates. MuSSE parameter estimates agree, showing higher rates in both specialist bill categories than among generalists, although taxa with short stout bills had slightly higher speciation rates than taxa with long curved bills (figure 4; electronic supplementary material, figure S4).

Table 3.

QuaSSE model selection showing the top two competing models with speciation rate as a function of the first PC axis describing bill shape. (ΔAIC is the difference in AIC values, d.f. indicates the number of parameters and w is the Akaike weight. Full results included in the electronic supplementary material, table S2.)

model	d.f.	ΔAICa	w
unimodal with drift	7.00	0.00	0.999
linear with drift	5.00	28.4	<0.001

^aThe lowest AIC value was −453.7.

Figure 4.

Posterior distribution of speciation rate estimates from MuSSE models examining the role of bill shape described by the first PC axis. Generalists represent the central 70th percentiles of scores along PC1, specialists with long curved bills had negative scores on PC1, and specialists with short stout bills had positive scores on PC1. The 60th percentile provided similar results (electronic supplementary material, figure S4). Illustrations by Gabrielle Cilfone. (Online version in colour.)

4. Discussion

We examined the effect of specialization, range size and range shape on speciation rates across a morphologically diverse and broadly distributed clade to determine the relative importance of ecological and geographical processes across macroevolutionary scales. We found higher speciation rates among taxa with a higher degree of specialization and ranges that were smaller and more disjunct or elongated (table 1). We also found differences in the strength of each mechanism between the island and continental taxa.

The positive correlation between ecological specialization and speciation rate persisted after removing island endemics (table 1), which includes the radiation of Darwin's finches. We also show higher speciation rates among continental specialists over generalists (figure 3; electronic supplementary material, table S1), suggesting that ecological speciation may be an important driver of diversity across diverse clades and large geographical scales. Moreover, bill morphology captures just a single axis of specialization. Niche partitioning could occur through phenotypic changes that do not influence bill shape (e.g. body and skull size) [39,40], yet, bill specialization alone explained as much variation in speciation rates as range size. Including specialization along additional niche dimensions could explain more variation in diversification rates.

The negative correlation between speciation rates and range size, however, did not persist after removing island endemics. This suggests that the ability of geographical range to predict diversification may be captured by differences between islands and continents. Allopatric speciation events across archipelagos that create multiple island endemics from a single widespread colonizer could produce this pattern. If these speciation events were associated with niche shifts, this is the basic mechanism proposed by taxon cycling [66], although island vicariance following a reduction in dispersal ability after colonization could also explain this pattern [67]. This hypothesis predicts higher speciation among generalists, which are more likely to colonize new habitats [68] and fragment into populations that produce ecologically similar species. Indeed, island generalists can exhibit higher speciation rates than specialists (electronic supplementary material, table S1). Alternatively, the negative correlation between contemporary range size and speciation rates, and the lack of such a correlation with island taxa removed, could be driven by an island effect, simply because islands have smaller areas than continents. A larger sample of island species would more robustly test whether geography plays a larger role on islands than continents. Geographical processes may still influence speciation on continents, but our range metrics do not explain as much variation in speciation rates as a single, narrowly defined metric of ecological specialization within this clade.

The positive correlation between specialization and speciation rates was not ubiquitous across lineages. We detected high speciation rates in lineages showing neither specialized bills nor small ranges. Models with specialization and hidden states improved model fit over specialization alone (table 2), suggesting other factors, such as climate and sexual selection [69–71], may explain additional variation in diversification. We did not attempt to explain the relative importance of all mechanisms of diversification, but compared the relative importance of two mechanisms in birds (specialization and range characteristics) on islands versus continents.

We also found highly specialized species with long branch lengths and low speciation rates. Catamblyrhynchus diadema, a large-billed species that specializes on foraging in bamboo, and species within the genus Cyanerpes that specialize on nectar, show high degrees of specialization, but lower species diversity than similarly specialized clades such as Diglossa flowerpiercers and large-billed finches. The Cyanerpes radiation occurred much longer ago than other specialist clades with higher speciation rates. The DR statistic estimates lower divergence rates for radiations that occur closer to the root and then slow or stop. Clades that specialized closer to the root of the tree could also erode their ability to diversify further due to low genetic variation, trade-offs with other lifestyles, or increases in extinction rate [15]. If specialists have higher extinction rates [32], early radiations would have more time to erode and may possess more extinct species than more recent radiations. If this is true, we expect members of older radiations to have lower diversification rates based on the DR statistic, and disproportionately older splits with sister taxon relative to specialists with higher diversification rates or generalists with lower, but more consistent, diversification. This hypothesis might be more robustly tested across the entire superfamily.

Although we show higher speciation rates for species on islands, with small-range size and greater degrees of specialization, available methods inhibit our ability to determine causation or separate the effects of each variable. Ecological specialization may occur following speciation via vicariance or parapatry where resulting species experience different environmental conditions. Islands may also facilitate specialization, leading to lower variation in the degree of specialization among island endemics, which could preclude our ability to detect an effect of ecological specialization on islands. However, we do not find evidence for greater specialization for island taxa among our study species. In addition, extinction rates may be higher for island taxa [14], which may bias speciation rate estimates.

Our results contrast with a recent study that showed no association between specialization and diversification across 11 avian orders [72]. This suggests that the relative importance of diversification mechanisms may vary across taxonomic scales. While our specialization index focused on bill morphology, our morphometric approach may capture subtle shape differences, such as the recurved bills of Diglossa flowerpiercers. This may account for more fine-scale niche partitioning among species than linear measurements can capture, which may have increased our ability to detect a pattern. Additional analyses at multiple taxonomic scales may provide insights into the relative importance of mechanisms driving diversification among different clades.

We confirmed heterogeneity in speciation rates within Emberizoidea, consistent with other studies [33,73]. We found support for a single metric of ecological specialization as an important driver of speciation across a diverse clade at continental scales. We also found higher speciation rates on islands and differences in the relative effect of specialization on islands versus continents, suggesting that the relative importance of mechanisms may vary across different geographical contexts. Further investigation into variation in diversification processes across different biogeographical regions and taxa will help elucidate which mechanisms are important at different scales. Including intraspecific variation may also provide insights into the diversification process [74] and determine mechanisms driving evolutionary lability.

Data accessibility

Data and code to replicate analyses are available on the Dryad Digital Repository: https://doi.org/10.5061/dryad.465b258 [75].

Authors' contributions

M.C. and B.J.O. conceived of and designed the study; M.C. collected data, conducted the analysis and drafted the manuscript, with input from B.J.O.; M.C. and B.J.O. edited and wrote the final version; both authors gave final approval for publication and agree to be held accountable for the work performed therein.

Competing interests

We declare we have no competing interests.

Funding

The Maine Agricultural and Forest Experimental Station (MAFES), the University of Maine and the USDA National Institute of Food and Agriculture McIntire-Stennis Project Number ME0-21710 provided funding.