Cold and dry winter conditions are associated with greater SARS-CoV-2 transmission at regional level in western countries during the first epidemic wave

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, Spain, Italy, France, Portugal, Canada and USA presented an early spread, a heavy COVID-19 burden, and low initial public health response until lockdowns. In a context when testing was limited, we calculated the basic reproduction number (R0) in 63 regions from the growth in regional death counts. After adjusting for population density, early spread of the epidemic, and age structure, temperature and humidity were negatively associated with SARS-CoV-2 transmissibility. A reduction of mean absolute humidity by 1 g/m3 was associated with a 0.15-unit increase of R0. Below 10 °C, a temperature reduction of 1 °C was associated with a 0.16-unit increase of R0. Our results confirm a dependency of SARS-CoV-2 transmissibility to weather conditions in the absence of control measures during the first wave. The transition from summer to winter, corresponding to drop in temperature associated with an overall decrease in absolute humidity, likely contributed to the intensification of the second wave in north-west hemisphere countries. Non-pharmaceutical interventions must be adjusted to account for increased transmissibility in winter conditions.


Systematic review of available evidence
Systematic reviews addressed the issue until 15 May 2020 (Briz-Redon A, et al, Prog Phy Geogr 2020& Mecenas P et al, PLoS One, 2020. Between 15 May 2020 and 15 December 2020, we searched PubMed database using the request: "(((COVID-19 OR SARS-COV-2)) AND (("2020/05/16" [Date -Publication]: "2020-12-15"[Date -Publication]))) AND (humidity [MeSH Terms] OR temperature [MeSH Terms] OR weather [MeSH Terms]OR climate[MeSH Terms])". Of 464 references, we identified 82 research articles relative to the relationship between weather parameters and SARS-CoV-2 transmission. Of these, 67 (82%) analysed case, death or hospitalization counts, of which 11 presented the analysis of a single time series and 56 the analysis of cumulative or longitudinal counts in multiple locations ranging from regions to countries. Of 67, 52 (78%) presented only univariate analyses of weather variables, or the analysis of multiple correlated weather variables without evaluation of potential confounders. Only 15 (18%) studies analysed the growth rate or the reproduction number of SARS-CoV-2, of which only 4 studies did a multivariate analysis.
In spite of their large numbers, univariate studies and studies based on counts suffered from multiple biases and, according to systematic reviews, provided only low-grade evidence of a negative relationship between SARS-CoV-2 and temperature.   Biaising factor Causal path Other connection

Figure S3: delay to peak in daily deaths according to the reduction of population mobility measured in transit stations (A) or in workplaces (B). The death count at the peak (maximal death count) is figured by the dot size (0, 100, 400).
The correlation between the delay to reaching the peak in daily death counts and the reduction of population mobility was strong, with respective Spearman correlation coefficient values of 0.636 (p<10^-5) for transit stations and 0.631 (p<10^-5) for workplaces (p<10^-5).

Figure S4: R0 calculation periods by country
Vertical green lines indicate the window considered for R0 calculation, starting at the date of 10 cumulative deaths and ending 28 days after lockdown. Vertical blue line indicate the end of the window considered for R0 calculation in the sensitivity analysis, 18 days after lockdown. The portion of the epidemic curve defined as displaying linear trend on the log-scale and used to extract the growth rate and perform the R0 calculation is highlighted in red, and the corresponding linear regression is displayed in orange. (See Results section for median and IQR values for calculation periods).

France
Italy USA (continued)  Based on the results of table S2 and S3, the log10 of population density was included as a linear covariate, due to a small gain in AIC with a limited loss in deviance explained. Distance to the first region affected was included as a spline to avoid issues in log transformation of 0 values (for regions first affected). Weather/climate parameters were included as non-linear predictors. Table S4: Multivariable results for the relationship between R0 and weather parameters, under the assumption of a linear relationship obtained with the hierarchical generalized additive model. Weather parameters are temperature, absolute humidity, and dew point temperature, adjusted for distance to the first region affected, population density, and elderly population. The linear approximation appears valid for dew point temperature: the spline had a linear shape ( Figure 6) and the linear approximation does not modify the percentage of deviance explained. The linear approximation results in a minor decrease in deviance explained for absolute humidity. It appears however not relevant for temperature since it results in a strong decrease in deviance explained.  Lag=0 (green line) indicates the model presented in Figure 5, lag=-1 indicates a weather summary period 7 days later, lag=1 indicates a weather summary period 7 days earlier…

Figure S6
Sensitivity analysis for mean weather variable splines in the multivariate model.
Lag=0 indicates the final model, as shown in Figure 6, lag=-1 indicates a weather summary period 7 days later, lag=1 indicates a weather summary period 7 days earlier etc.
The effects of log10(population density) and percentage of population aged>80 were stable across lags.

Figure S8: Schematic representation of time periods used in the study
The study period (green) was defined as the period where the growth of death counts was considered to occur without early stochasticity (cumulative death count>10) and with limited influence from lockdown (date<lockdown+28 days).
Within the study period, the exponential growth period (in blue) was defined as the linear portion of the log(daily death count)=f(t) curve (black), and R0 was estimated on this period.
The transmission period (in red) was defined as the period when infections corresponding to the exponential growth period were acquired, i.e. 3 weeks earlier, over the same duration. Weather variable summary values were calculated over the transmission period.