The empirical results provide strong evidence for slowdown in speciation rate in large clades. The magnitude of slowdown seems to be quite large. For example, among the 22 clades with 15 or more lineages at 2 Mya, the median value for γ = −1.77, and the median clade size is 29, consistent with a slowing of speciation rate in the later stages of a radiation to 10%–50% of the initial rate (Protocol S1
). Our conclusions are predicated on the assumption that the molecular phylogenies accurately reconstruct the timing of speciation events. In particular, if saturation is present in the molecular data, deep branch lengths will be consistently underestimated, leading to a bias toward a negative γ [29
]. Over the time scales of this investigation—and given our use of the complex GTR + I + G model—this seems unlikely to be a problem. Further, if nucleotide saturation is driving patterns, we expect to see greater slowdown in older clades, the opposite of what is observed. Nor is the negative correlation of γ with clade size (A) expected to arise as a consequence of nucleotide saturation. Estimates of speciation patterns based on gene trees add error to the estimate of speciation times (e.g., [30
]), but this source of error should make slowdowns more difficult to detect, rather than introducing a bias. The observed negative relationship between clade size and γ is not limited to this study. Indeed, two earlier studies based partly on subsets of the data analyzed here reported a similar trend [18
Two predictions regarding temporal patterns of speciation arise from adaptive radiation models [2
]. First, clades should show a slowdown in speciation toward the present as niche space is filled. This prediction rests on the assumption that members of a clade, in our case usually a genus, experience competition over niche space. Second, given density-dependent speciation, slowdowns should be particularly evident in large clades [18
]. We found strong support for both these predictions. However, the same predictions also arise under a simple model where speciation rates are constant across time (and this also applies to models which allow for low levels of extinction, see Table S1
). This is because small clades are likely to have diversified, by chance, slowly early on, and large clades are likely to have diversified quickly early on, both followed by regression to the mean. Thus a bias on the part of researchers toward studying large clades leads to the expectation that those clades that are studied will show the pattern of slowdown. This bias probably affects tests of other questions [25
]. For example, very few phylogenetic studies have identified an extinction rate greater than zero, which is paradoxical given estimates of extinction derived from fossils [32
]. A signal of extinction is identified from reconstructed phylogenies, because, under a constant rate model, extinction leads to an increase in the observed branching rate toward the tips of the reconstructed tree for extant species [10
]. Given that we expect to find a decrease in diversification rate toward the present in large young phylogenies (if b
are relatively constant through time and d
is low), then a bias toward testing for extinction in such large young clades introduces a strong bias against detecting extinction.
Under the adaptive radiation model, a negative correlation between clade age and γ is expected, because rapid initial diversification followed by a slowdown should result in more-pronounced slowdowns in older clades. The reverse is expected under constant rate models: after controlling for clade size, a positive correlation between clade age and γ (i.e., younger clades tend to show the strongest slowdowns) is predicted (B). This is because clades that quickly attain a given size are likely to have experienced above-average rates of initial diversification. As lineages accumulate, the overall diversification rate will approach the underlying mean, resulting in slowdowns. Thus our finding of a tendency toward a positive association (albeit marginally nonsignificant) between clade age and γ () is more in accord with the constant speciation model than the adaptive radiation model.
Both the adaptive radiation model and constant-rate, stochastic model predict negative γ in large clades, whereas the marginally nonsignificant positive correlation between clade age and γ is more consistent with the constant rate null model. However, constant rate models cannot explain the very high prevalence of significant slowdowns observed in large clades. In particular, our simulations show that if the actual extinction rate across bird lineages over the past 20 million years has approached the speciation rate, then the probability that the strong slowdowns observed in large clades could have arisen under a constant-rate birth–death model becomes vanishingly small. We thus conclude there is strong evidence for density-dependent cladogenesis in large clades.
Speciation rates across a whole clade may slow through time because a few ecologically unusual and/or geographically restricted lineages persist for a long period of time without speciating, even as the rest of the clade continues to follow a constant birth–death model [23
]. This should create strong tree imbalance. However, the tree-imbalance parameter we used explains only a small proportion of the variance in γ (partial r2
= 0.08, n
= 45 clades), and it is likely that slowdowns are the result of more general ecological interactions. For example, in the Old World Leaf Warblers (Phyllscopus
), related sympatric species in the Himalayas are old and occupy different habitats, which are presumed to have arisen in association with mountain building or climate change 8–10 Mya [33
]. Even Leaf Warbler allospecies, with abutting geographical ranges, are typically separated by millions of years [33
]. We suggest that limited ecological space in this and other groups has restricted the ease of range expansions, and consequently further allopatric speciation. Similarly, Ricklefs [24
] found a negative correlation between number of species in a clade and age of the clade across passerine birds, and he interpreted this finding in terms of niche-filling, as we do here. Some alternative explanations for slowdowns have been suggested, including nonrandom extinction [14
] and episodic appearances of multiple barriers [14
], but these seem less likely to produce such a general pattern.
In conclusion, we find that two factors contribute to the prevalence of slowdowns reported in large phylogenies. First, the strong signal of slowdown supports an adaptive radiation model, where speciation is accelerated in empty environments and slows as niches get filled. Second, speciation events happening randomly within clades through time may also result in the presence of a slowdown in large young clades. Randomness does not mean that speciation is completely unpredictable, but rather that multiple independent causes are likely to contribute [35
]. Speciation may be promoted by factors such as occasional extinctions creating new ecological opportunities, appearance of habitat that can be exploited by multiple lineages (rather than a single lineage that rapidly diversifies), the strength of barriers, chance dispersal events, and the occasional evolution of traits within lineages that affect speciation probability. The overall importance of random processes as causes of slowdowns depends on the true extinction rate. If extinction rates are low, the importance of stochastic factors in generating slowdowns may have been underestimated. If, as seems likely, extinction rates approach the speciation rate [36
], then constant birth–death models on their own cannot explain slowdowns. Instead, our findings of strong slowdowns provide support for nonrandom processes of species diversification through time.