key: cord-0780009-fybrukdc authors: Landier, J.; Paireau, J.; Rebaudet, S.; Legendre, E.; Le Hot, L.; Fontanet, A.; CAUCHEMEZ, S.; Gaudart, J. title: Colder and drier winter conditions are associated with greater SARS-CoV-2 transmission: a regional study of the first epidemic wave in north-west hemisphere countries. date: 2021-01-26 journal: nan DOI: 10.1101/2021.01.26.21250475 sha: aa1916d3fa1abd53fa3387ca6f60bcffea062764 doc_id: 780009 cord_uid: fybrukdc Higher transmissibility of SARS-CoV-2 in cold and dry weather conditions has been hypothesized since the onset of the COVID-19 pandemic but the level of epidemiological evidence remains low. During the first wave of the pandemic for Spain, Italy, France, Portugal, Canada and USA presented an early spread, a heavy COVID-19 burden, and low initial public health response until lockdowns. We used regional death counts as a proxy for infections while diagnostic tests remained limited and calculated a basic reproduction number (R0) in 63 regions. After adjusting for population density, early spread of the epidemic, and aged population, temperature and humidity were negatively associated to SARS-CoV-2 transmissibility. A -1g/m3 lower mean absolute humidity was associated with a 0.15-unit higher R0. Below 10{degrees}C, a 1{degrees}C-higher temperature was associated with a 0.16-unit lower R0. Our results confirm a strong dependency of SARS-CoV-2 transmissibility to weather conditions in the absence of control measures. The initiation of the second wave in north-west hemisphere countries was likely triggered by the transition from summer- to winter-like conditions. High levels of restrictions to social activities should be maintained until spring to avoid or limit a third wave. During the first wave of the pandemic for Spain, Italy, France, Portugal, Canada and USA presented an 22 early spread, a heavy COVID-19 burden, and low initial public health response until lockdowns. We 23 used regional death counts as a proxy for infections while diagnostic tests remained limited and 24 calculated a basic reproduction number (R0) in 63 regions. After adjusting for population density, early 25 spread of the epidemic, and aged population, temperature and humidity were negatively associated 26 to SARS-CoV-2 transmissibility. A -1g/m3 lower mean absolute humidity was associated with a 0.15-27 unit higher R0. Below 10°C, a 1°C-higher temperature was associated with a 0.16-unit lower R0. 28 Our results confirm a strong dependency of SARS-CoV-2 transmissibility to weather conditions in the 29 absence of control measures. The initiation of the second wave in north-west hemisphere countries 30 was likely triggered by the transition from summer-to winter-like conditions. High levels of restrictions 31 to social activities should be maintained until spring to avoid or limit a third wave. 32 The spread of SARS-CoV-2 in February-March 2020 caught the majority of European and North 3 American countries unprepared. The spread of the virus was largely uncontrolled until movement 4 restriction and social distancing policies were put in place at national or regional level with various 5 degrees of intensity [1] . Most regions experienced a peak in hospital admissions, approximately 2 6 weeks after lockdown was imposed, corresponding to infections acquired around the date of 7 lockdown [2]. 8 The spread of this first wave was largely heterogeneous between countries. Various determinants 9 were proposed to explain the differential spread of the virus. Asian countries (Japan, South Korea, 10 Vietnam and Thailand) and isolated countries (Australia, New Zealand, Iceland) experienced only 11 limited transmission and stood out for high levels of preparedness and efficient response strategies. 12 Even within South West European and North American countries where no effective response was 13 deployed before lockdowns, SARS-CoV-2 virus spread heterogeneously at the regional level, as 14 observed from hospitalization, death counts, and confirmed by serological surveys [3, 4] . 15 Epidemic spread is characterized by the basic reproduction number, or R0. R0 expresses the number 16 of secondary cases resulting from a given case in the context of a naive population with uniform 17 probability of contact. 18 R0 depends on individual susceptibility to infection when in contact with an infective individual, and 19 on the probability of an infectious contact. Environmental parameters affect R0: population density 20 increases the probability of contacts between individuals and weather conditions may affect the 21 survival of the virus or the individual susceptibility to an infection. Most known respiratory viruses 22 spread during the cold season in the temperate Northern hemisphere [5] . Weather conditions in 23 winter can also affect individual susceptibility to infection through irritation of the nasal mucosa, but 24 also influence the behaviour of individuals towards conditions prone to transmission (living or 25 gathering in closed, heated spaces with a dry atmosphere) [5] . In addition, temperature, humidity, 26 and UV, might directly affect the virus survival and modify infectiousness [6, 7] . Individual preventive 27 behaviours (masks), collective strategies reducing mobility and contacts (lockdowns) or limiting the 28 duration of the infectious period (detection and isolation) modify the number of secondary cases and 29 the effective reproduction number can be calculated, which accounts for these alterations in the 30 "natural" history of transmission. 31 In spite of a large number of studies, the evidence regarding the link between weather conditions and 32 SARS-CoV-2 transmission remains limited. At date of 15 May 2020, a systematic review retained 61 33 studies analysing the relationship between COVID-19 epidemic and environmental factors [8] . 34 Methodological issues included the lack of controlling for confounding factors such as population 35 density [8] . Inappropriate epidemiological and statistical methods were also pointed out [8, 9] . 36 Comparison between countries with different counter epidemic responses, testing strategy, or delayed 37 onset of the epidemic might also have led to inconsistent results [8] . The third study was at the global scale for 203 states (USA, Canada, Australia and 2 China) or countries and identified a negative association between UV light exposure and SARS-CoV-2 3 growth. Temperature was negatively associated with growth only after adjustment for UV light 4 exposure [13] . All three studies identified a positive association between population density and 5 epidemic growth. 6 Overall, multiple studies described a negative relationship between temperature and COVID-19 7 outcomes but the majority were unadjusted ecological correlation studies with strong risk of bias 8 bringing low quality evidence [8] , and evidence from multivariable growth studies remains ambiguous. 9 A modelling study defined the range of possible dependency between SARS-CoV-2 transmission and 10 absolute humidity, based on two known coronaviruses and influenza [14] , and recent models still do 11 not include specific SARS-CoV-2 data [15] . More precise estimations of the effect of meteorological 12 conditions on the spread of SARS-CoV-2 are required to better anticipate and inform policies regarding 13 seasonal adjustments [16] . 14 The objective of this study was to evaluate the contribution of weather parameters in the transmission 15 of SARS-CoV-2, by analysing their effect on SARS-CoV-2 basic reproduction number in a context of low 16 public health response during the early phase of the first wave in 6 north-western hemisphere 17 countries. 18 In order to compare the drivers of the epidemic dynamics between regions accurately, we calculated 22 the basic reproduction number (R0) of the virus for each region affected by the first wave of COVID-19 23 epidemic in six countries, using the dynamics of the daily death counts. These six countries were 24 located in the western part of the northern hemisphere, approximately between 25 and 50° of latitude 25 ( Figure S2 ). These countries experienced winter conditions and underwent significant SARS-CoV-2 26 transmission in a context of low public health response during the first wave of the epidemic. 27 This analysis was conducted at the first administrative subdivision country, here referred to as 28 "region". Regions corresponded to States in USA and Canada, autonomous communities in Spain, and 29 regions in Italy (regioni), Portugal (região) and France (regions). 30 Confirmed case counts were insufficiently reliable due to overall lack of tests and different testing 31 strategies between countries, between regions of the same country and between periods for the same 32 region. Hospitalization counts were not available consistently at regional level across countries. 33 Overall, COVID-19 deaths were preferred as they were less likely to have different definitions within 34 the same region or country and to undergo significant changes over the study period. R0 is an indicator 35 of the speed of progression of the outbreak, and was therefore less likely to be biased compared to 36 indicators based on cumulative counts or cumulative incidence rates. Estimating R0 values based on 37 deaths counts relies on the minimal assumption that the infection-fatality rate was constant over the 38 study period (~1 month) for a given region, which defined a proportional relationship between 39 infections and deaths. 40 41 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint Daily reporting of COVID-19 deaths was not initiated immediately after the first recorded deaths, and 2 often started in a staggered manner in the different regions of a country. Analyses started on the date 3 when 10 cumulative deaths were reached in a given region to account for these limitations of available 4 data and to limit stochasticity in the occurrence of the first deaths. 5 This study aimed at estimating SARS-CoV-2 R0 prior to implementation of major interventions in 6 different regions. It was necessary to avoid the influence of the lockdown measures on the 7 transmission. Reports indicated that symptoms occurred in median 5 days after infection (IQR=2-7), 8 and hospitalization occurred in median 7 days after the onset of symptoms (IQR=3-9) [17, 18] . Reports 9 of delay between hospitalization and death ranged from 5 days in New York to 12 days in Italy for 10 patients admitted in ICU [19, 20] . In Spain and Italy, a median of 11 days (IQR=7-17), respectively 12 11 days (IQR 7-20), were reported between onset of symptoms and death [21, 22] , while an early study 12 in China indicated a median of 18 days from onset of symptoms to death [23]. These figures lead to an 13 overall estimation that deaths occurred in median 17-18 days after infection, with a broad range of 14 variations. In France, mortality peaked at week 14 when lockdown occurred at the beginning of week 15 12, i.e. for infections likely acquired during week 11. As a result, we initially assumed a 3-week delay 16 between infection and death and excluded all data beyond 28 days from lockdown date to limit this 17 study to deaths caused by pre-lockdown infections. We also conducted a more restrictive sensitivity 18 analysis excluding all data beyond 18 days after lockdown date, which corresponded to the estimated 19 median delay between infection and deaths. 20 21 Regions where the smoothed daily death count did not rise above 5 deaths/day during the study period 23 were not included in the analysis. Regions where the smoothed daily death count did not rise above 24 10 deaths/day during the linear growth phase were excluded. This limited the study to regions having 25 displayed a clear exponential growth phase. 26 One Italian region was excluded because there was no weather station data available except for 2 27 stations >2000m altitude. 28 29 Deaths from COVID-19 31 Regional level data on COVID-19 deaths were retrieved from data shared by national health ministries 32 and/or public health agencies of Spain, Italy, Portugal, France, United States of America, and Canada 33 (Table S1 ). Deaths were reported as daily new death counts or as a cumulative number. 34 35 Population structure by one-or five-year age groups, sex and region was retrieved from open access 37 data shared by national institutes or administrations responsible for national statistics or demography 38 (Table S1 ). The percentage of the region population aged >70 or >80 years was calculated in order to 39 adjust for differences in way of life (e.g. rural regions have older population than metropolitan areas 40 in Europe). The percentage of elderly inhabitants also represented an adjustment in case of a 41 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint differential spread of the epidemic among the elderly population (which would lead to a non-1 proportional increase in deaths compared to cases. 2 Region shapefiles by country were obtained from national geographical authorities, open source 3 datasets or as provided in the coronavirus open data packages proposed by several national health 4 agencies (Table S1 ). The region surface was either obtained from the geographical layer (land surface 5 area), or calculated from the polygon extent. We estimated the percentage of each region surface with 6 >5 inhabitants per km 2 based on WorldPop 2020 raster dataset [24] . Population density was estimated 7 as the population in the region divided by the surface after excluding areas <5 inhabitants/km 2 in order 8 to limit the underestimation of population density when high heterogeneity existed between urban 9 centers and sparsely populated territories within the same region (deserts, mountains, polar regions). 10 11 Weather/climate 12 Weather station reports since 1 January 2020 were obtained from US National Oceanic and 13 Atmospheric Administration (NOAA) through the R-package {worldmet}. We extracted hourly 14 temperature, relative humidity (RH), dew point (DP), precipitation and windspeed observations for 15 each station. Hourly absolute humidity (AH, in g/m 3 ) was calculated from RH and temperature based 16 on the Clausius-Clapeyron formula [25] . Daily minimum, maximum and mean values were calculated 17 for temperature, AH, RH, dew point, as well as cumulative sum of precipitation and mean wind speed. 18 Days with >5 missing hourly record (no observation of any parameter) were excluded. 19 Each region was attributed stations based on geographic location. All available observations in weather 20 stations of the region contributed to the regional daily average. Weather stations can be located in 21 mountains or inhabited locations where they record weather conditions that differ strongly from 22 actual populated areas of the same region. To avoid this bias, observations were assigned population-23 based weights: we estimated the population located within 10 km of each weather station using 24 WorldPop 2020 data; and for each day and each region, we calculated the total population within 10km 25 of any station reporting data for that day. Daily station observations were weighted in proportion of 26 the population around each station relative to the total population. Stations located within 10-km of 27 each other were included in a single buffer and each station was assigned equal weight within the 28 buffer. 29 For each weather parameter, the regional summary value was calculated over the assumed 30 transmission period. The assumed transmission period corresponded to the duration of the R0 31 calculation period, lagged by 3 weeks. Sensitivity analyses included lags ranging from 2 to 8 weeks from 32 R0 calculation period, i.e. -1 to +5 weeks from assumed transmission period. The longer lags were likely 33 irrelevant for actual transmission but aimed at identifying climate rather than weather aspects. 34 Using this approach, the weather parameters averaged over the assumed transmission period and over 35 each region included: minimum, mean and maximum average values for temperature, relative 36 humidity, absolute humidity, dew point, average cumulative rainfall per day and average wind speed. 37 38 Lockdown date was defined using Google mobility data as the date when a decrease >25% in workplace 40 localization was reported and sustained over 3 days in the region [26] . All references to a « date of 41 lockdown » hereafter refer to this definition. This definition matched national lockdowns in European 42 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint countries. This simplification was necessary due to the heterogeneity in social distancing measures 1 taken at regional (state) level in the USA and Canada. The objective was to exclude periods where 2 transmission would start to slow down due to these measures. 3 4 Distance to first region with 10 cumulative deaths 5 For each of the four European countries considered, the first region with 10 cumulative deaths was 6 defined. In Portugal, two regions reached >10 deaths on the same day, and the region with the highest 7 count (14 versus 12) was selected. Within each country, the euclidean distance in kilometers between 8 the main city in each region and the main city in the first region above 10 cumulative deaths was 9 calculated to reflect spatial autocorrelation due to proximity in the spread of the epidemic. In USA and 10 Canada, distance to the first region above 10 cumulative deaths were calculated separately for East 11 Coast and West Coast, using the limit between Central and Mountain time zones. This was necessary 12 due to the early start of the epidemic in the state of Washington, which reached 10 deaths by 2 March 13 2020, 16 days before the next state (New York on 18 March). 14 15 Statistical analyses were performed using R version 4.0. Maps were produced using ArcGIS 10.7.1. 17 Daily death counts were smoothed using a 5-day moving average filter to account for irregularities in 18 data transmission and publication. 19 For each region, the exponential growth period was estimated using a log(deaths)=f(time) 20 representation and r (the exponential growth rate) was extracted as the coefficient of a Poisson 21 regression. R0 was calculated for each region using the generation time method assuming a gamma 22 distribution with parameters 7 and 5.2 for SARS-CoV-2 generation time [27, 28] . In order to improve 23 the adjustment of the regression, the start and end dates of calculation period were allowed to shift 24 by up to 2 days (+/-1 day or +1/+2 days for start date if calculation period began at the date of 10 25 cumulative deaths) using the built-in function sensitivity.analysis() of the package {R0} [29] . 26 A directed acyclic graph (DAG) was constructed using Dagitty v3.0 web-based application 27 (http://www.dagitty.net/dags.html) in order to visualise the relationships between R0 and the 28 explanatory covariates ( Figure S1 ). Dependence and independence assumptions were verified using 29 Spearman correlation coefficient. 30 Relative humidity values are temperature-dependent and absolute humidity or dew point 31 temperatures are also strongly correlated to temperature [30] . The different weather covariates 32 observed at the same 3-week lag from R0 calculation period were included separately in the models. 33 A generalized additive mixed model (gamm) regression was used to evaluate the effects of climate, 34 population and other determinants on the value of R0, using a Gaussian distribution and the identity 35 link function (package {mgcv}). A country-level random effect was included to account for within-36 country correlations. Canada presented with only 2 regions and was grouped with USA for random 37 effect, while the single region included from Portugal was grouped with Spain. Univariate analyses 38 were conducted assuming linear and non-linear effects, using B-splines to model non-linear effect of 39 covariates. Models were compared using the percentage of deviance explained and Akaike's 40 Information Criterion. 41 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in First, we assessed the importance of the 3-week delay for weather variables (corresponding to the 5 delay between transmission period and R0 calculation period based on death count exponential 6 growth). For this, we tested a variety of lags from -1 to 5 weeks from that estimated transmission 7 period (i.e. 2 to 8 weeks from the R0 calculation period) and included them as linear explanatory 8 variables in the univariable hierarchical model or as non-linear explanatory variables in the 9 multivariable hierarchical generalised additive model. 10 Second, we assessed the effect of the 28-day window to define the exponential growth period by using 11 a narrower window ending 18 days after date of lockdown. We recalculated R0 for regions where the 12 linear growth period retained for the main analysis extended beyond the 18-day limit, and followed 13 the same plan as the main analysis. 14 Third, we assessed possible continent specific effects by fitting continent-specific splines for weather 15 variables in the final multivariate models. 16 The six countries (United States, Canada, Spain, Italy, France and Portugal) included 128 regions/states. 20 Overseas regions (n=11) and regions which had experienced <10 cumulative deaths 28 days after 21 lockdown (n=15) were not included ( Figure 1 ). Likewise, regions with a maximal daily mortality <5 22 deaths (n=19) within 28 days after lockdown were not included. Overall, 83 regions were assessed for 23 exponential growth period, and R0 was calculated for 64 regions with sufficient exponential growth 24 ( Figure note, larger reductions in human mobility patterns after lockdown as assessed from google mobility 31 led to shorter delay (Spearman correlation coefficient = 0.63, p<0.0001, Figure S3 ). This delay was 32 reduced to 19 days in median (IQR=17-27) for South European countries with nationwide lockdowns 33 and strong mobility reductions (>60% in average, Figure S3 ). 34 R0 were estimated over a median exponential growth period of 11 days (IQR=9-14, range=5-19 ). In 35 median, R0 estimation period started 5 days (IQR=0-8.5) and ended 16 days after the date of lockdown 36 (IQR=11-21, range 1-27). Figure S4 presents details calculation periods. 37 The median R0 value was 2.58 (IQR=2.08-2.66). R0 estimates were lowest (<1.5) in two regions in 38 France and one in Spain, USA and Italy (respectively in Centre-Val de Loire, Nouvelle Aquitaine, La 39 Rioja, Alabama and Abruzzo). R0 were highest (>4.0) in New York (USA), Lombardia and Piemonte 40 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint (Italy), Castilla-La Mancha (Spain) and Ontario (Canada) (Figure 3A and B). Spain and Canada had overall 1 higher R0 values compared to Italy, France and USA ( Figure 4A ). R0 values exhibited significant spatial 2 autocorrelation (Moran's I=0.20, p=0.0087). 3 Covariates were heterogeneous between countries (Figure 3 and 4) . Population density ranged 4 between 82 and 1056 inhabitants/km 2 (median=271, Figure 3C and D) and was overall higher in Italy 5 ( Figure 4B ). Mean absolute humidity during the estimated transmission period was 4.98 in median 6 (IQR=4.29-5.99, range=2.26-11.32, Figure 3E and F) corresponding to a median dew point of 3.7°C 7 (IQR=0.9-6.3, range=-7.7 -+15.3). Mean temperature was 9.8°C in median (IQR=7.1-11.5, range=-2.0-8 19.9 ). Distance to the first region with 10 cumulative deaths was 406km in median (IQR=228-794) and 9 much larger in the USA compared to European countries ( Figure 4F ). 10 11 Factors associated with R0 12 Each variable was included in a univariate model assuming linear ( Figure 5, Mean temperature led to the model with the largest deviance explained compared to models including 21 minimum or maximum temperature, or any AH or dew point temperature (Table 1 ). In this model, a 22 10-fold (+1 log10 unit) increase in population density was associated with a +0.67 R0 unit increase 23 (95%CI=0.05-1.28). The proportion of population aged 80 years and older was not significantly 24 associated with R0. The relationship between mean temperature and R0 was not linear. A strong, 25 nearly linear drop of approximately 1.0 R0 unit was observed between 2.5 and 12°C, and a plateau 26 beyond 12°C (Figure 6 and 7) . The residuals from this model did not exhibit significant spatial 27 autocorrelation (Moran's I=0.054, p=0.215). 28 Mean AH and mean dew point values exhibited similar profiles with increasing values associate with 29 decreasing R0 values (Table 1, Figure 6 ). In spite of the narrower range of values, it seemed that the 30 relationship between R0 and AH or dew point temperature did not reach a plateau (Figure 7 ). Assuming 31 a linear relationship, a 1 g/m3 higher absolute humidity translated in a 0.15 unit lower regional R0, 32 respectively a 1°C higher dew point temperature translated in a 0.08 unit lower R0 (Table S5) . 33 34 First, using lagged weather summary values as linear univariate predictors, our statistical model found 36 similar relationships between R0 and temperature variables, but 0-, 1-and 5-week lags led to the 37 strongest effect for mean AH and mean DP ( Figure S5 ). Using lagged weather summary values as non-38 linear predictors in the multivariate model, the shapes of the relationships between temperature, AH 39 and DP remained similar, with a nearly linear drop reaching a plateau at values corresponding to milder 40 winter weather/climate ( Figure S6 ). Overall, correlations were strong between weather summary 41 observations at the different lags ( Figure S7 ). Correlations of lagged temperature and humidity 42 observations was highest between -1, 0, 1 and 5-week lagged observations compared to other lags. 43 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in Second, when limiting the R0 calculation window to 18-day after date of lockdown, 10 additional 1 regions were excluded and analysis was conducted on 53 regions. After adjusting for population 2 density, population over 80 and distance to the first region affected, the non-linear negative 3 relationship between R0 and temperature remained unchanged, reaching a plateau around 10°C 4 (spline p-value=0.0475). The shape of the relationship between AH, respectively DP, and R0 remained 5 similar but did not reach statistical significance (p-value=0. 19) . Later AH or DP summary values, 6 corresponding to 1 week later than the estimated transmission period (i.e. 2-week lag from R0 7 calculation period), restored the negative relationship with R0 (p=0.0661, respectively p=0.0667). 8 Third, fitting continent-specific splines for weather covariates, also resulted in low statistical power: 9 only 37 regions in South Europe and 26 in North America. After adjusting for population density, 10 population over 80 and distance to the first region affected, humidity variables (AH or dew point) 11 retained a negative association with R0 (continent-specific spline p-values were 0.0915 (North 12 America) and 0.0988 (South Europe) for AH, respectively 0.0506 and 0.1078 for dew point 13 temperature). The relationship between R0 and mean temperature was markedly different between 14 North America (negative association, p=0.005) and South Europe (p=0.17). 15 In this study, we analysed SARS-CoV-2 propagation parameter R0 during the first wave of the pandemic 19 in 63 regions of 6 north western countries. We showed that R0 values were influenced by population 20 density (+0.6 for a 10-fold increase in density), by proximity with the first epidemic focus of the country 21 or coast for USA (-0.3 for a 10-fold increase in distance to the first region to record 10 COVID-19 22 deaths), and by weather or climate conditions. For regions with mean temperatures below 10°C during 23 the transmission period, a linear association was observed with R0 values: a difference between 24 regions of +1°C was associated with a 0.16-unit decrease in R0. A difference in mean absolute humidity 25 of +1 g/m 3 was associated with a 0.15-unit decrease in R0. Similar results were obtained with dew 26 point temperatures, with a difference of 1°C was associated with a 0.08-unit decrease in R0 (Table S5) . 27 After adjusting for major confounders and spatial autocorrelation, our results indicate that weather 28 conditions brought a significant contribution to drive the magnitude of the first wave, even if it was 29 limited by an initial heterogeneous spread of the virus, which protected regions located furthest away 30 from the first foci. 31 32 This study relied on a regional scale analysis and accounted for different dynamics within the same 34 country. The overall epidemic wave at country level was actually the sum of diverse dynamics, as is 35 obvious for large countries but also true for Spain, Italy or France, where regions were heterogeneously 36 affected, as confirmed by serological studies [3, 31]. Likewise, weather or climate heterogeneity 37 between regions of a given country was large. By analyzing distinct spatial units with heterogeneous 38 population density and weather, we were able to assess the effect of parameters that may be 39 otherwise confounded by country-level parameters such as response strategy, timing of the analysis 40 period compared to the progression of the epidemic, but also age structure of the population. 41 The 6 countries were selected due to their homogeneous location in the northern hemisphere 42 between 25 and 50 degrees of latitude and the low efficacy of their counter epidemic responses until 43 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint a lockdown was decreed. We are therefore as close as possible of conditions allowing R0 estimation. 1 This 4-parameter hierarchical generalized additive model provides an explanation of up to 45% of R0 2 variability across 63 affected regions of the northern hemisphere, thus providing useful insights on the 3 drivers of the first wave, and allowing to estimate the contribution of population density and weather 4 conditions for the next waves, when the effect of local introduction is no longer relevant. 5 We acknowledge several limits. The first one is the necessity to rely on death counts to estimate R0 6 during this wave due to the lack of information on actual infections from limited testing and the lack 7 of consistency between countries for hospitalization counts. This assumption is similar to usual 8 assumptions for R0 estimation based on diagnosed cases, since true number of infections remains 9 unknown and detection occurs at variable delays after infection but it could have led to an 10 underestimation of the R0. Indeed, while the average delay between infection and death was 3 weeks, 11 deaths occurring from infections contracted the same day may in fact occur over a larger window 12 extending several weeks beyond this delay. 13 The second one is the use of a single summary weather observation over the assumed transmission 14 period, which may fail to express the dynamical aspects of weather. The choice of fixed, 1-week 15 increments to study the effect variations in the weather summary window may be too coarse to 16 capture effects for narrow exponential growth periods. The sensitivity analysis showed a slight increase 17 in effect when considering weather summary values calculated 1 week earlier than the estimated 18 transmission period, but the 18-day sensitivity analysis showed a stronger negative relationship 19 between R0 and AH summary values corresponding to 1-week later than the estimated transmission 20 period. The role of the weather conditions might be more important during the beginning of the 21 exponential growth period, until a sufficiently large number of persons becomes infected and 22 parameters such as population density become increasingly important. This study could also only 23 evaluate the effect of a limited range of weather conditions on SARS-CoV-2 transmission, since the 24 number of observations with a mean temperature below 2.5°C or above 15°C was low. This prevents 25 from analyzing the effect of higher temperatures such as during autumn. This restriction was however 26 necessary to achieve a minimal homogeneity of the studied regions in terms of their exposure and 27 response to the pandemic. Finally, this analysis includes only 63 regions and may lack statistical power. 28 The necessity to ensure that the epidemic growth of deaths was sufficient led to the exclusion of 29 regions that may be affected, but not enough for daily deaths counts to reach 10 deaths/day. 30 Finally, this study showed an association between SARS-CoV-2 R0 and temperature/absolute humidity, 31 but due to the strong correlation between absolute humidity and temperature in the seasonal 32 conditions analysed here, it is difficult to determine which parameter is more important and they could 33 not be analysed in combination [30] . The continent-specific analysis suggests that the relationship 34 between absolute humidity and R0 was more stable than that of temperature, which was strong in the 35 US but less so in South Europe. We could not conclude whether the relationship results from a direct 36 role of weather on individual susceptibility to viral invasion (e.g. dry nasal mucosa from indoors heating 37 and outdoors cold) or on viral persistence/survival, or from an indirect role of climate on human 38 behaviors (e.g. regions with cold winter favor more indoors living conditions and lead to bigger 39 infection opportunities). 40 Our study shows an important dependency of SARS-CoV-2 transmission to weather/climate, with a 42 0.16-unit increase in R0 for a 1°C difference in mean regional temperature below 10°C, or a 0.15-unit 43 increase for a -1g/m3 decrease in absolute humidity. Northern hemisphere countries experienced a 44 second wave of SARS-CoV-2 infections during autumnal transition from summer to winter, while still 45 actively maintaining control strategies. Public health strategies need to account for a further increase 46 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. in transmissibility when/where cold and dry winter conditions are reached. In the regions of the north-1 west hemisphere, lifting restrictions on social activities during the winter would likely lead to faster 2 increases than observed during the second wave, when autumnal conditions were prevalent. 3 4 Authors' contribution 5 JL, JG, SC and AF designed the study. JL conducted the analysis with methodological contributions of 6 JP, EL, and LL. JL JP EL LL AF SC JG contributed to interpreting the results and editing the manuscript. 7 8 Acknowledgements 9 The authors would like to thank Laurax Simgiane Ferbliegas from Hab' lab Marseille and L. Jae for 10 help and support in retrieving multi-country regional data. 11 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in Tables 1 Table 1: Multivariable model results for the relationship between R0 and weather parameters, 2 obtained with the hierarchical generalized additive model. Weather parameters are temperature, 3 absolute humidity, and dew point temperature, adjusted for distance to the first region affected, 4 population density, and elderly population. Non-linear effects are presented in Figure 6 . Models 5 assuming linear effects for weather covariates are presented in Table S5 perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in lockdown by each country, and the thin red line the median date +28 days. 2 3 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint Figure 3 : Map of regional values for R0 and selected covariates, panels are presented by continent. 1 A,B: R0 ; C,D: population density (inhabitants per km2) ; D,E: mean temperature ; F,G: mean absolute 2 humidity. 3 4 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint Figure 6 : Non-linear effects in the multivariable model for weather parameters (see Table 1 to first region affected, model 3. 4 5 6 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in percentage of inhabitants >80 years (5.6%), and corresponding to the first region first affected 3 (distance=0km). 4 5 6 7 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 26, 2021. ; https://doi.org/10.1101/2021.01.26.21250475 doi: medRxiv preprint Estimating the effects of 2 non-pharmaceutical interventions on COVID-19 in Europe Lockdown impact on COVID-19 epidemics 4 in regions across metropolitan France Prevalence of SARS-CoV-2 in Spain (ENE-COVID): a nationwide, population-based seroepidemiological 7 study Population-based seroprevalence surveys of anti-SARS-CoV-2 antibody : An 9 up-to-date review Seasonality of Respiratory Viral Infections Effect of 13 Environmental Conditions on SARS-CoV-2 Stability in Human Nasal Mucus and Sputum. Emerg Infect 14 Dis Mechanistic 16 theory predicts the effects of temperature and humidity on inactivation of SARS-CoV-2 and 2 other 17 enveloped viruses The effect of climate on the spread of the COVID-19 pandemic: A 19 review of findings, and statistical and modelling techniques Comment on A. 21 annua and A. afra infusions vs. Artesunate-amodiaquine (ASAQ) in treating Plasmodium falciparum 22 malaria in a large scale, double blind, randomized clinical trial Effects of 24 temperature and humidity on the spread of COVID-19: A systematic review Impact of climate and ambient air pollution on the epidemic 27 growth during COVID-19 outbreak in Japan Population Density, and Temperature With the Instantaneous Reproduction Number of SARS-CoV-2 30 in Counties Across the United States Seasonality and uncertainty in global COVID-19 growth rates Susceptible supply limits the role of 34 climate in the early SARS-CoV-2 pandemic. Science (80-) Immune life history, 36 vaccination, and the dynamics of SARS-CoV-2 over the next 5 years. Science (80-) A framework for 39 research linking weather, climate and COVID-19 COVID-19): A Review cases and 1286 of their close contacts in Shenzhen, China: a retrospective cohort study Presenting 3 Characteristics, Comorbidities, and Outcomes among 5700 Patients Hospitalized with COVID-19 in 4 the New York City Area Risk Factors Associated 6 With Mortality Among Patients With COVID-19 in Intensive Care Units in Lombardy Informe sobre la situación de COVID-19 en España Superiore di Sanità. Characteristics of SARS-CoV-2 patients dying in Italy Report based on 10 available data on Clinical course and risk factors for mortality of adult 12 inpatients with COVID-19 in Wuhan, China: a retrospective cohort study Global High Resolution Population 15 Denominators Project Impact of meteorological conditions and air pollution on 18 COVID-19 pandemic transmission in Italy Google LLC. Google COVID-19 Community Mobility Reports Estimating the burden of 22 SARS-CoV-2 in France To cite this version How generation intervals shape the relationship between growth rates and 24 reproductive numbers The R0 package : a toolbox to estimate reproduction numbers for 26 epidemic outbreaks Use of Weather Variables in SARS-CoV-2 Transmission Studies 4.5% 30 of population in metropolitan France developped antibodies against SARS-CoV-2: first results from 31 the national survey EpiCov