We explored the association between influenza seasonality and climate in a representative sample of 78 global sites, spanning an absolute latitudinal range between 1° and 60°. Our analyses revealed two distinct types of climatic conditions associated with influenza seasons observed globally: “cold-dry” and “humid-rainy”. In general, sites that experienced low levels of specific humidity and temperature (less than 11–12 g/kg and 18–21°C) for at least one month during the year were characterized by seasonal influenza activity during the months with minimal levels of specific humidity and temperature. In contrast, sites that maintained high levels of specific humidity and temperature were generally characterized by influenza epidemics during the most humid and rainy months of the year. The predictions of our climate-based models compared favorably to influenza epidemiological information collected independently of the dataset used for the model-building exercise.
The bimodal nature of the relationship, in both relative and absolute terms, between specific humidity and influenza peaks, and its strong relationship to other climate variables such as temperature and precipitation, makes specific humidity a useful gauge of the environmental favorability of influenza activity across all latitudes (). However, although the specific humidity models were significant predictors of influenza peaks globally, this was primarily due to their performance in high latitudes. In low latitudes, precipitation was a stronger predictor of the timing of influenza activity, with peaks typically occurring in months with average precipitation greater than 150 mm ().
Overall, although precipitation was strongly associated with influenza peaks in low latitudes, the timing of influenza peaks in this region was more difficult to predict than in high-latitude sites. Several sites in this region were not characterized by well-defined influenza season; rather, influenza activity was present year-round likely due to the limited seasonal environmental variation that characterizes much of the region. For example, equatorial sites such as Iquitos, Peru, and Singapore—where influenza seasonality is weak 
— experience limited fluctuations in precipitation, with monthly averages constrained to a narrow range of 150–300 mm year-round. In contrast, middle and low-latitude sites such as Fortaleza, Brazil and Yangon, Myanmar —which are noted for their well-defined influenza seasons 
—are characterized by large amplitude range in average monthly precipitation from 25 mm in the dry season to over 300 mm and 600 mm in the rainy season, respectively.
Model performances were particularly poor in a number of middle latitudes sites. Predicting influenza peaks in these sites may be complicated by large seasonal swings in climate that characterize the region, generating both cold-dry and humid-rainy seasons that are equally favorable for seasonal influenza activity, such as in Senegal (). For these sites other factors might play a critical role in determining the timing of influenza activity, including population mixing (i.e., travel) with regions that do experience well-defined influenza seasons 
, or the seasonal phasing with school cycles 
. Moreover, the presence of both cold-dry and humid-rainy seasons could explain the occurrence of semi-annual influenza epidemics in some of these middle-latitude sites. For example, Hong Kong has a primary influenza peak in winter when average monthly specific humidity and temperature are less than 8 g/kg and 17°C, and a secondary influenza peak in summer when average monthly precipitation is near 400 mm.
Temperature was a strong predictor of influenza seasonality in high latitudes, suggesting that cold temperatures may drive seasonal epidemics in these regions. However, previous analyses of laboratory experiments have indicated that specific humidity is a more parsimonious predictor of virus survival and transmission than temperature 
. Furthermore, individuals in temperate regions spend a vast majority of their time indoors where temperature is managed and does not correlate well with outdoor temperatures. Yet, temperature may affect the timing of influenza epidemics through mechanisms independent of virus survival; for example, low outdoor temperatures may promote indoor crowding, thereby increasing person-to-person contact rates 
. It is also possible that even limited exposure to cold outdoor temperatures can have long-lasting physiological effects on hosts that make them more susceptible to infection or affect viral shedding 
. Additional experimental and observational work is needed to disentangle the contribution of specific humidity and temperature on influenza seasonality; epidemiological information from Central Africa would be particularly useful in this respect as our climate-based predictive models disagreed in this region.
The findings that both cold-dry and humid-rainy conditions are associated with influenza peaks could be used to support the hypothesis that two distinct mechanisms account for influenza seasonality in temperate and tropical climates, perhaps due to changes in the dominant mode of transmission 
. For example, specific humidity may drive the timing of influenza epidemics in high latitudes by increasing virus survival and enabling aerosol transmission; whereas direct transmission or transmission by fomites may dominate in low-latitude sites where rainy conditions favor indoor crowding. Middle latitudes may be a transition zone where influenza seasons are driven by low specific humidity or high levels of precipitation depending on local climate. Another intriguing possibility is that the relationship between specific humidity and virus survival underlies influenza transmission across all latitudes. For example, a few experimental studies have indicated a U-shaped relationship between relative humidity and influenza virus survival, suggesting a similar relationship for specific humidity given that experiments were held at constant temperature 
. Other laboratory studies, however, have indicated that virus survival and transmission increase monotonically as specific humidity decreases 
. Further, the hypothesis that specific humidity drives influenza transmission globally is inconsistent with the low predictive power of this climatic variable in middle and low-latitude sites in our study.
Relative humidity was a strong predictor of influenza peaks in high and low latitudes, but a poor predictor in middle-latitude. In high-latitude regions, relative humidity can vary significantly between indoor and outdoor environments, and it is typically minimal indoors during the winter when building air is heated. Our analysis relied on outdoor humidity and hence we cannot rule out that winter influenza epidemics in high latitudes could be related to low indoor relative humidity and associated changes to host physiology, such as reduced mucociliary clearance 
. In low latitudes it is possible that relative humidity is confounding precipitation in our analysis. Disentangling these two factors will require more highly-resolved epidemiological data from equatorial regions, and further experimental and observational studies.
Solar radiation was also a significant predictor of influenza peaks in high latitudes suggesting that it may also underlie influenza seasonality in these regions, perhaps through variation in vitamin D intake 
. However, solar radiation was not as strong a predictor of influenza peaks as were specific humidity and temperature. This corroborates recent studies indicating that specific humidity is a stronger predictor of seasonal influenza activity than solar radiation and vitamin D variability in the U.S. 
. Still, the potential seasonal forcing of solar radiation on influenza transmission warrants further experimental and observational investigation.
The power of this study was rooted in the large number of spatially diverse sites used to develop the epidemiological and climatic databases and associated models. However, the challenge of describing seasonal influenza activity consistently across a variety of data sources required a crude epidemiological measure, such as the average month of peak influenza activity. This measure of influenza activity has several key drawbacks. Foremost, all months with the exception of the peak influenza months were considered equal, whether they had substantial influenza activity or not. Second, the month of peak influenza activity may not be contemporaneous with the month in which transmission is under the most environmentally favorable conditions, since non-environmental factors such as viral seeding, population susceptibility, and person-to-person contact rates likely play a role in the timing of influenza epidemics 
. In this respect, it is reassuring that a 1-month lag maximized the association between influenza peak and most of the climatic variables, which is broadly consistent with the time scale of the ascending phase of a local epidemic. Third, we could not assess putative geographical variation in the transmission potential or intensity of influenza epidemics. For example, we may expect locations that have the most favorable environmental conditions to experience the greatest influenza annual attack rates and reproduction numbers, holding all other relevant variables equal. A further limitation relates to between-year variability in influenza timing and the limited temporal sampling of our dataset, which may have resulted in imprecise estimates of the average influenza peak in some sites, especially sites that had only one year of influenza data. However, sensitivity analyses limited to multi-year studies revealed similar relationships between climate predictors and influenza peaks, confirming the robustness of our results. Finally, we were unable to check whether between-year fluctuations in climatic variables may result in departures from average influenza seasonal characteristics in specific years. This question could be an interesting area for future research with more temporally refined epidemiological datasets.
A number of follow-up studies could help refine our understanding of the small and large-scale processes underlying influenza seasonality. For example, experimental infections in humans under controlled temperature and humidity conditions could determine which environmentally-mediated mechanisms are most important for human-to-human transmission. However, there are several ethical and methodological hurdles to overcome in such studies 
. Seasonal fluctuation in contact rates could be monitored by wireless sensor technology, which has recently proved successful in estimating dynamic contact patterns in schools and at conferences 
. On a broader spatial scale, determining regional differences in influenza transmission dynamics and attack rates would be most informative. A recent study has suggested that the reproduction number of seasonal epidemics was lower on average in Brazil than in temperate countries, which could be mediated by environmental factors 
. Modeling of long-term influenza time series data could help assess the transmission impact of seasonal fluctuations in population mixing in different regions, such as those associated with school cycles 
and transportation networks 
. For example, epidemiological evidence indicates that influenza circulation was weakly seasonal in Iceland prior to the 1930s, presumably because of low connectivity with other populations, and epidemics only became fully synchronized with those in Europe and the USA following a dramatic increase in international travel in the 1990s 
. Hence, efforts to collate multiyear influenza epidemiological information retrospectively and prospectively in various regions of the globe, especially from middle and low latitude regions, will be of tremendous help to further elucidate the environmental and population drivers of seasonality.
In conclusion, our study broadens our understanding of the relationships between seasonal influenza epidemics and environmental factors and provides a synthesis of epidemiological and climatic characteristics across temperate, subtropical and tropical regions. We have highlighted the importance of thresholds in specific humidity, temperature and precipitation that are associated with the epidemiology (and potentially the modes of transmission) of influenza. The results of this study could help improve existing influenza transmission models by providing a more accurate estimate of the environmental forcing on transmission processes, particularly in low and middle latitudes 
. Further, our models could be used to predict the seasonal timing of influenza activity in locations with little or no observational data on influenza activity, and help target surveillance efforts and optimize the timing of seasonal vaccine delivery, 
. More broadly, we hope that our work will generate interest in testing the association between climatic patterns and infectious disease across a wide range of diseases and latitudes, particularly for respiratory and enteric pathogens that display marked seasonality 
. A better understanding of the environmental, demographic and social drivers of infectious disease seasonality is key for improving transmission models and optimizing interventions