We have shown that the genetic population structure of influenza A (H3N2) arises in part from global migration dynamics, with the most important contributions from China and Southeast Asia, but nonetheless significant contributions from temperate regions outside Asia (). In contrast to the prevailing source-sink model, we find evidence of significant migration of viruses from temperate regions to tropic regions, and that lineages may exist outside of Asia for several seasons, persisting through dynamic migration between regions of different seasonality (). Additionally, we find that China, Southeast Asia and the USA all contribute to the trunk of the influenza genealogy (), and thus mutations occurring within these regions have shaped the global flu population. The evolution of H3N2 influenza over the past 10 years thus reflects the dynamics of a global metapopulation, rather than a metapopulation restricted to East and Southeast Asia.
Our use of the structured coalescent model to analyze influenza evolution represents a significant step forward over previous techniques. The tree constructed by Russell et al. 
using phylogenetic methods, is a single estimate of the HA genealogy. We use a Bayesian sampling technique to analyze a large number of trees concordant with the genetic data 
. More importantly, our coalescent method explicitly incorporates sampling date, sampling location and also an underlying model of the demographic process. These details provide substantially more context, and thus allow for more accurate reconstruction. Rambaut et al. 
use a similar Bayesian coalescent approach; however, their technique did not take into account population structure.
By analyzing a large number of sampled trees and through resampled replicates, we establish the degree of uncertainty of our estimates. Each migration rate has a confidence interval attached to it (Table S3
). Additionally, our estimates of the trunk location over time have a degree of confidence associated with them. Our statistical model strongly suggests that the 1998–1999 USA epidemic forms the trunk of the influenza genealogy (). Consistent with this hypothesis, we observe that samples from the USA during this period coalesce to the trunk of the genealogy rapidly in absolute terms, not just relative to other samples (). From 2000 to 2002, Chinese samples are closer to the trunk than samples from the USA and Oceania, but are not close to the trunk in absolute terms (). Because of this, there is considerable uncertainty as to whether the trunk of the genealogy resides in China or in Southeast Asia (). This particular result is especially supportive of our method, as we lack samples in Southeast Asia prior to 2002 (Figure S2
), yet still we infer that Southeast Asia may be contributing to the trunk of the genealogy. There are other time periods (e.g. 2006) in which samples are far from the trunk, suggesting the possibility that the trunk may be located outside of sampled regions.
Regardless of methodological differences and differences of interpretation, our results are compatible with the results of Russell et al. 
and Rambaut et al. 
. In their analysis of mean distance to the trunk, Russell et al. find that the USA is behind China, Taiwan, Hong Kong and South Korea, but ahead of every other country sampled, including all of the Southeast Asian countries. This itself should suggest that the USA plays an important role in the global migration dynamics. Additionally, the inference of the 1998–1999 USA epidemic as the trunk of the genealogy is congruent with the findings of Rambaut et al. In a genealogy produced from only USA sequences (their Supplementary ), it is clearly seen that while most USA epidemics occur as side branches, the 1998–1999 epidemic is distinctly part of the trunk of the genealogy. Russell et al. state: “the tree does show evidence for bidirectional seeding but no evidence for non-E&SE Asian strains contributing to long-term evolution of the virus during the study period.” We suggest that if Russell et al. had analyzed samples from 1998–1999, they would have obtained different results.
It is possible, for example that some of the strength of the USA's contribution to the migration dynamics () comes from its proximity to the Central American tropics. In this scenario, gene flow back and forth across the Pacific would be attributable to strains of influenza circulating in Central America that pass through the USA in their spread to the rest of the world. However, if this scenario were true, we might expect that the Central American influence would extend to South America in addition to the USA. We do not see evidence of this; South America contributes the least among studied regions to the global migration dynamics.
Additional evidence for a temperate contribution to the migration network and to the trunk of the genealogy comes from epidemiological simulations (see Materials and Methods
). In simulations with equal rates of contact between hosts in a northern population, a tropical population and a southern population we observe that despite strong seasonality in the temperate regions, substantial emigration occurs out of the temperate populations (). In this scenario, we find that although the trunk of the genealogy resides predominantly in the tropics, it often passes through the temperate populations during the course of a seasonal epidemic ().
Analysis of epidemiological simulations for a source-sink model (i) and an equal contacts model (ii) of spatial structure.
Examining the influenza genealogy (), it is apparent that regional outbreaks often result from very few immigration events, consistent with previous results 
. For example, the 2003 epidemic in Oceania appears almost completely monophyletic and can trace its history to a single migration event (or perhaps multiple migration events of identical strains) in early 2003. Thus, even if there are millions of infected individuals at the peak an epidemic, the genetic diversity of the virus will be bottlenecked at the beginning of the outbreak 
. The bottlenecking effect of low migration may have contributed to the observed pattern of restricted genetic diversity in influenza. We observe similar effective population sizes across all regions (Table S4
), consistent with the hypothesis that epidemic influenza is passed from one region to another, persisting nowhere. If there were a reservoir of endemic influenza in E-SE Asia (or elsewhere) that repeatedly seeded epidemics in the rest of the world, then the coalescence of E-SE Asian lineages would not be bottlenecked to the same extent, in which case we would observe significantly deeper coalescence events in this region and a correspondingly larger effective population size. This is not, however, the pattern we observe, reinforcing the idea that a metapopulation structure underlies influenza's persistence, even in the tropics 
The global dynamics of influenza virus population influence a variety of public health decisions. Because influenza frequently migrates out of the USA, seeding epidemics in other parts of world, actions taken to combat the disease in the USA can have global impacts. For instance, the use of antivirals in the USA may promote the evolution of drug-resistant strains, which could then spread to the rest of the world. And conversely, vaccination programs outside of E-SE Asia have the potential to curb the global spread of the disease. Additionally, with increased knowledge of the patterns of flu migration, it may be possible to tailor vaccine design to particular areas of the globe. For instance, we observe that most influenza in South America arrives from the USA. This suggests that vaccines used in South America should be preferentially constructed from the USA strains of the previous season.
Our research suggests that the majority of historically relevant evolution of the influenza virus occurs in China, Southeast Asia and the USA, with other regions of the globe playing significant, but relatively minor roles. This conclusion is, to some extent, contingent upon the restricted temporal and spatial patterns of viral sampling. There may be other regions of the world, such as Africa, Central America and India, that act as important sources in the worldwide influenza migration network. Increased worldwide sampling of the influenza virus would further clarify the complex migration dynamics of the virus.