The role of weather conditions in COVID-19 transmission: A study of a global panel of 1236 regions.

It is believed that weather conditions such as temperature and humidity have effects on COVID-19 transmission. However, these effects are not clear due to the limited observations and difficulties in separating impact of social distancing. COVID-19 data and social-economic features of 1236 regions in the world (1112 regions at the provincial level and 124 countries with the small land area) were collected. Large-scale satellite data was combined with these data with a regression analysis model to explore the effects of temperature and relative humidity on COVID-19 spreading, as well as the possible transmission risk by seasonal cycles. The result shows that temperature and relative humidity are negatively correlated with COVID-19 transmission throughout the world. Government intervention (e.g. lockdown policies) and lower population movement contributed to decrease the new daily case ratio. Weather conditions are not the decisive factor in COVID-19 transmission, in that government intervention as well as public awareness, could contribute to the mitigation of the spreading of the virus. So, it deserves a dynamic government policy to mitigate COVID-19 transmission in winter.

Introduction

Cases of Severe Acute Respiratory Syndrome Coronavirus Disease-2019 (COVID-19) have been extensively reported since December 2019. As this disease spread worldwide rapidly, it poses great challenges not only to human health but also to social-economic development (Amankwah-Amoah, 2020; Chen et al., 2021; Hossain, 2020; Hsiang et al., 2020). As of 26, September 2020, COVID-19 has resulted in more than 32 million confirmed cases and a death toll of 1 million globally. Currently, the lack of an effective vaccine for COVID-19 has resulted in a global spread that has varied in speed across regions, and thus it is arguably an urgent task to explore the determining factors for virus transmission.

Importantly, epidemiological studies (Barreca and Shimshack, 2012; Casanova et al., 2010; Chan et al., 2011; Shaman and Kohn, 2009; van Doremalen et al., 2013) revealed that seasonal and geographic climatic variation (i.e., low air temperature and low humidity) modulate respiratory pathogens transmission and most respiratory pathogens exhibit prevalence peaks in temperate regions in winter. They were focused on Influenza, Coronavirus, SARS-COV , and MERS-COV. Few could provide information about COVID-19 virus because it is much different from those known viruses in pathogenicity and transmission. COVID-19 is epidemiologically similar to the influenza virus since both are highly transmissible by the respiratory route and cause acute infection (Cobey, 2020). Some studies explored the weather effect on COVID-19 transmission at the country level (Iqbal et al., 2020; Y. Wu et al., 2020) or using limited regional level datasets (Sun et al., 2020; Wang et al., 2020; Pan et al., 2021; Thu et al., 2020). However, country-level studies cannot capture regional diversity in weather among countries with large areas and uneven population distribution, such as the USA, China, and Brazil. They ignored the weather heterogeneity within countries and relied on variations across large continents as well as multiple climate zones. Thus, current studies likely misestimated the weather effect.

Moreover, the effect of weather conditions on the transmission is likely to be sensitive to some possible confounding factors in quantitative studies, where social and economic conditions, including government intervention, are the dominant ones among them. Governments are taking a wide range of measures in response to the COVID-19 outbreak, including, but not limited to, school closings, travel restrictions, bans on public gatherings, and contact tracing (Hale et al., 2020). For example, previous quantitative conclusions about the weather-transmission relationship ignored the effect of the incubation period and omitted some key variables (ex: active cases and susceptive population). Furthermore, the evidence (Dalziel et al., 2018; Jia et al., 2020) indicated that population concentration and economic condition (including social distancing) may also shape the transmission intensity. Up to now, the roles which social-economic conditions play in the spread of COVID-19 are still not clear. Moreover, dynamic transmission models for COVID-19 (i.e., SIER model), which adopt non-empirical parameters (Baker et al., 2020; Kissler et al., 2020) from other coronaviruses, cannot precisely separate the weather contribution in shaping the potential dynamic route. Therefore, current studies likely offer biased estimations about the role of weather in COVID-19 transmission.

Theoretically, there are several methodologies to explore the role of climate conditions in COVID-19 transmission. We recognize that COVID-19 transmission is complicated and involves several important factors other than the climate condition. However, given the current situation, the regression-based approaches are relatively effective in attributing the role of climate to COVID-19 by controlling for all other potentially confounding factors. In contrast, laboratory experiments are generally based on limited samples and simulate environments with no consideration of social-economic factors.

To overcome the disadvantages discussed above, we collected global provincial data to investigate the effects of temperature and relative humidity on COVID-19 transmission. Due to unavailable control experiments concerning the weather effect, we applied a multi-variables regression model with lagged variables to control the incubation period. Additionally, a set of socio-economic control variables from multi-source and government intervention to take account of these confounding factors were adopted when estimating the weather effect on COVID-19 transmission. An additional novelty in our study is that the weather factor is merged into an available dynamic transmission model by providing more reliable parameters to model the dynamic transmission route. To the best of our knowledge, this is the first study conducted at the provincial level on a global scale. This work may provide a reference point for a flexible government response to COVID-19 transmission during seasonal cycles.

Method

Statistical analysis

We built a multivariate regression model (see Eq. (1)) to explore the weather condition effect on transmission: where i indexes a region, t a day, and a lag day. We consider the new daily cases fraction () and the basic reproductive number (R0) as dependent variables (Y) in Eq. (1). R 0 is calculated as in Eqs. (3), (4), where the mean serial interval V, exponential growth rate λ of the cumulative number of cases (confirmed), and the ratio of the infectious period to the serial interval h are set following Lipsitch et al. (2003).

Accordingly, daily average temperature (Tmean, in Celsius degree) and relative humidity (%, RH) are our variables of interest. Here, refers to flexible functional forms of temperature, including higher degree polynomials, thereby allowing for a possible non-linear relationship. RH is a control for evaporation and affects a droplet’s size and its chemical microenvironment (Marr et al., 2019). Therefore, it is RH (but not absolute humidity) that acts as a determinant factor for virus survival in aerosols.

To the best of the authors’ awareness of the current state of knowledge, individuals who get infected are likely to experience an incubation period before onset. Current evidence (Li et al., 2020; Wu, et al, 2020a) suggests that the incubation period may vary between 6 and 8 days. Accordingly, we focused on the effects of temperature and relative humidity with a 6-day lag and further examined the same effects with 5-day to 14-day lags for control groups.

One should note that the specification of the starting date (day = 1) in the study period is vital for estimating the temperature effect. In this study, the day when the first case is confirmed is not chosen as the starting day due to the following reasons. Firstly, at the early stage of the outbreak, the population flow intensity, which is the dominant factor for transmission (Fauver et al., 2020), varies across regions globally. Evidence from the USA suggests that the risk of domestic importation at present far outweighs that of international introductions. However, due to data limitation, we have no information about the population inflow/outflow to exclude its impact. Secondly, due to the insufficient focus on the epidemic at the early stage, there were no effective measures in medical supplies or public management policies. Accordingly, these data are not timely released and might probably lead to deviations. Finally, after the early stage, inter-regional population movements turned out to be the most determinant factor for transmission. One must therefore take into account both the epidemic scale and the observation size. On this basis, we set the starting date as the time when the total regional confirmed cases reach 100.

Additionally, we add a time fixed effects as a control for factors that are common to all countries, such as the global virus prevention materials supply (e.g., ethanol, mask, and protective suit) and public awareness of COVID-19 at different stages.

To further explore the effect of weather conditions on transmission by income group, we divided the samples by GDP per capita per group in accordance with the World Bank criterion, and construct a dummy variable high with high = 1 indicating high-income regions and high = 0 low-income regions. This dummy variable and its interactions with Tmean and RH are both added to the models. Afterward, we merged the weather variables into the SIER model.

Control variables

There are a number of obvious confounding factors (e.g., active case fraction, economic development, population concentration (Dalziel et al., 2018), age structure (Geard et al., 2015; Ioannidis et al., 2020), geographic conditions (Tian et al., 2017), and government intervention (Giordano et al., 2020), that affect the transmission of an epidemic and thus should be controlled for in the regression analysis. On this basis, the control variables should include gross regional product per capita (GRPper), regional population concentration (PopCon), government response (Lockdown), elevation, and suspected population. The last one is composed of the working population (aged 15–64) ratio ( and school-age group (aged 6–15) ratio (. The role of GRP in COVID-19 transmission is a priori unclear. A higher GRP per capita means closer social distance and more frequent population movement while it also denotes higher education attainment and better cognition on COVID-19. Hospital conditions are conventionally used to measure medical quality on disease cure but can infer little about prevention. Hospital conditions may have a significant impact on death cases but little on new cases. Therefore, we do not include hospital condition as a control variable. Some of the employed control variables (work age and school-age population) are related to social distancing.

As for population concentration, a higher population concentration means that individuals in short social distancing will face a higher risk of getting infected by droplet transmission (Moriyama et al., 2020), leading to a high infected rate. Geographic factors, such as elevation, are highly associated with the weather type and indirectly affect air pressure (which controls the virus transmission rate in aerosol) (Tian et al., 2017). The air pressure in high-elevation regions will limit virus transmission in the aerosol.

Age structure is also an important factor, as current evidence (Glynn, 2020) shows a very strong age-dependence. Note that the COVID-19 transmission is also connected to close contact of the susceptible population. Therefore, we control for the school-age and labor group to exclude the effect of susceptive population on COVID-19.

It is widely accepted that government response is a vital factor for COVID-19 transmission (Giordano et al., 2020; Prem et al., 2020), of which local and trans-regional transmission are two possible outbreak channels. Note that the measures taken by governments across the globe have affected public movement greatly, such as border controls, teleworking from home, social distancing, and limiting the sizes of gatherings. The heterogeneity of government intervention is difficult to take account of since all these measures are generally implemented simultaneously and no available data is available to measure their outcome. Therefore, we employ a time fixed effect to absorb the unobserved factors that vary in time that are related to containment, economic, and health measures. Correspondingly, we added a variable Lockdown, which controls for government intervention in local and trans-regional COVID-19 transmission through social distancing and trans-border flow. However, due to data limitations, it is difficult to accurately evaluate the contribution of government response to COVID-19 transmission. Here, we assumed that its contribution would increase as the policy continues to take effect. Nevertheless, it is not likely to increase infinitely, i.e., the contribution rate will slow down when approaching its peak. Under this hypothesis, we considered a logistic transforming function (Eq. (5)) to evaluate the government response in transmission.where is the initial value of at T = 0. T denotes the lasting weeks of government response. for example, if the government policy regarding COVID-19 has continued 19 days in day t, T should be equal to 2.7 (ratio of lasting days divided by 7). In our setting, before government response, this value should be as close to zero as possible because it denotes the initial value of the response. Thus, the value of determines the initial intensity of government response. Here, we set  = 0.001 to assume that a slight contribution in COVID-19 at an early stage. The shape of the logistical function under different values of can be found in Figure S1. Alternatively, we examined the temperature and relative humidity effects on COVID-19 transmission by setting to values of 0.01, 0.03, 0.07, and 0.1, respectively.

Therefore, they should be controlled for the regression analysis, namely active case intensity, economic development, population concentration (Jia et al., 2020) and age structure (Geard et al., 2015), geographic conditions, and government intervention (Giordano et al., 2020). As a result, the control variables include Gross Regional Product per capita (GRPper), Regional Population Concentration (PopCon), Government Response (Lockdown), Elevation, and Susceptive Population. The last one is composed of the working age (15–64) ratio ( and school-age (6–15) ratio (.

Robustness checks

To reduce the possibility of selective bias on some key variables, we conduct three robustness checks for the weather-transmission relationship: (1) The selection of a threshold of total regional confirmed cases for observations is vital to the estimation. Here, we examined the relationships by increasing the threshold of total case numbers to 200 and 300, respectively. (2) Considering that daily temperature differences between the maximum and minimum temperature vary across regions globally, we substituted average temperature by its maximum and minimum counterparts, separately. (3) Multi-initial values of lockdown in logistic functions were applied to prove that its initial value can affect the weather-transmission relationship.

Non-linear effect of temperature on COVID-19 transmission

We try to explore the possible non-linear effect of temperature on COVID-19 transmission by setting temperature function as higher-degree polynomials. Based on the incubation period, we focus on a 6-day lag variable of average temperature and relative humidity. Besides, a partial relationship between temperature and transmission that filtered for other explanatory variables was estimated with reference to Barro (1991).

Estimation of temperature and relative humidity effects based on SIER model

We simulatethe SEIR epidemic model (Wu, et al, 2020b) to estimate the effects of temperature and relative humidity on the infection rate. The initial SEIR model is expressed as follows (Eq. (6) to (11)): where (incubation period), (duration period), and are defined according to Prem et al. (2020).

The temperature and relative humidity effects were incorporated into the SIER model as follows (Eq. (12) and (13)): where and are the estimation coefficients of T and RH, respectively, as listed in Supplementary Table S2.

Project transmission risk due to temperature and relative humidity effects

To assess the maximum possible risk of transmission due to seasonal temperature variations, we calculate the transmission risk attributable to temperature in winter and summer. To facilitate the comparison, we set 6-day lagged variables of average temperature () and relative humidity (), which take the day with the maximum growth rate of confirmed cases as the benchmark value. The average temperature and relative humidity in July and January of 2020 were assumed to be the same as those in 2019. The risk of temperature in summer or winter, which is calculated by Eq. (14), denotes changes in new daily cases fraction compared with and .

Source of data

We manually collecte the new daily cases, cured, and deaths in 1236 regions in the world as of 31 May 2020, which were extracted from the COVID-19 epidemic information released by publicly available daily COVID-19 reports from the official health department of countries. To deal with small countries that lack sub-national case data, whose average land areas are about 185,000 km2 and among which the largest one is Algeria (2,382,000 km2), we selecte alternative country-level data from the COVID-19 Data Repository established by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University. Finally, our sample covers 5,926,622 confirmed cases and 7.4 billion of the global population, which are equal to 98.7% of global confirmed cases and 98.2% of the global population, respectively. Our study area is comprised of 1112 sub-national regions (in 54 countries) and 124 countries (Fig. 1 ). Our subnational samples cover 54 countries of which 13 are Asian countries, 27 are European, 3 are North American, 4 are South American, 5 are African, and 2 are Ocean. The sources of remote sensing satellite data, as well as weather and social-economic features, are listed in Table 1 .

Fig. 1

Confirmed cases per million inhabitants by subnational region. Note: The data were collected and calculated by authors own calculation as of 31, May 2020. The observations are classified into 10 groups by every 10th quantiles of confirmed cases per million population. The map division is only a schematic diagram and does not indicate an accurate administrative area. Map data is from https://gadm.org/

Table 1

Data source of the variables.

Variables	Data Source
COVID-19 cases (as of 31 May 2020)	Subnational: COVID-19 website and situation report from Department of Health by countries.
	National: John Hopkins GitHub repositories. https://github.com/CSSEGISandData/COVID-19
Air & Dew point temperature (as of 31 May 2020)	Fifth generation ECMWF atmospheric reanalysis of the global climate assimilation system (ERA5). https://cds.climate.copernicus.eu
Relative humidity	Calculated by Air & Dew point temperature
GRP per capita	1.Subnation region: EuroStat (Europe Union members): https://ec.europa.eu/eurostat/web/regions/data/databaseOECD stat database (OECD countries): http://stats.oecd.org/Department of Statistics: (Asia, Africa and South America countries)2. Country level: World Development Indicators Database, World Bank: https://databank.worldbank.org/source/world-development-indicators
Population concentration	Gridded Population of the World (GPW), v4 from Socioeconomic Data and Applications Center, Columbia. https://sedac.ciesin.columbia.edu/data/collection/gpw-v4
Elevation	Altimeter Corrected Elevations (ACE2), v2 (1994–2005) Digital Elevation Model, Socioeconomic Data and Applications Center, Columbia. https://sedac.ciesin.columbia.edu/mapping/ace2/
School population ratio	Gridded Population of the World (GPW), v4 from Socioeconomic Data and Applications Center, Columbia.
Labor population ratio	Gridded Population of the World (GPW), v4 from Socioeconomic Data and Applications Center, Columbia.
NOx density	Aura OMI satellite, OMINO2D level3 daily data file. https://disc.gsfc.nasa.gov/datasets/OMNO2d_003/summary
Lockdown	Oxford COVID-19 Government Response Tracker. Blavatnik School of Government. https://www.bsg.ox.ac.uk/research/research-projects/coronavirus-government-response-tracker2020.

Weather variables

Weather data: The meteorological data of selected regions and countries are from the fifth-generation ECMWF atmospheric Reanalysis of the global climate Assimilation system (ERA5), from which we extract the hourly variable of ‘2 m temperature’ and ‘2 m dew-point temperature’ from ERA5.The data file is assembled in the resolution of 0.25° × 0.25°. In transforming from grid data to province level data, for regions of which the area is bigger than the minimum size of the grid, the climatic variables were calculated by averaging the values of the grid that the region is covering. If this was not the case, the climate variables were set equal to the values of the nearest grid to the region’s geological centroid. The dDaily average (maximum and minimum) temperature was calculated by averaging all pixels in a region with the Spatial Analyst tool of ArcGIS. Relative Humidity is obtained from Eq. 15–17 of the World Meteorological Organization (WMO, 2010), where Tc denotes air temperature (in Celsius) and Td denotes dew point temperature (in Celsius).

Social-economic variables

Values for the gross regional product (GRP) per capita (the price of 2017 international dollar, Purchasing Power Parity (PPP)) are taken from four type of sources: Annual statistic reports of department of statistics by countries, the Eurostat database, the OECD regional economics database, and the World Development Index Database. For those countries which were not covered by these data, country-level GDP per capita from the World Bank Development Indicators was used. GRP per capita data covers 40 countries (1126 sub-nation regions) and country-level GDP covers 142 countries. We convert GRP into real values of 2017 international dollars PPP by using the economic indicators from the International Comparison Program (ICP), World Bank, released in May 2020.

Elevation data were obtained from Altimeter Corrected Elevations dataset (ACE2), v2 (1994–2005) Digital Elevation Model captured by Shuttle Radar Topography Mission (DEM-SRTM), National Aeronautics and Space Administration (NASA), which provides information at 30 arc-seconds for the range of 60°N to 60°S.

Labor-age (15–64) and School-age (6–15) population data were taken from the Gridded Population of the World (GPW), v4 from Socioeconomic Data and Applications Center (SEDAC) Columbia (available at a 1 km × 1 km resolution). The dataset was constructed from the latest digital population census data of countries. The labor and school age ratio are measured by the ratio of the corresponding groups to the regional total population. All raster and grid data were processed by ArcMap 10.6 and Python 3.6.

Population concentration is measured by the Herfindahl-Hirschman Index (HHI) (see Eq. (18)), where G denotes the 1 km × 1 km-grid number located in region i and is equal to the area of a region in square kilometers. A simple index of population per square kilometer cannot account for the population distribution in space. HHI (Chakravarty et al., 2020) is a commonly accepted measure of market concentration, the value of which is much more sensitive to the number of agents. To make HHI comparable among different regions, we modified it to generate the index of population concentration. In this way, the population concentration will reach a maximum when the population is located in certain small regions (e.g., more people live in the metropolitan areas).

For measures of government intervention in COVID-19 we collected the earliest execution dates of national-wide government measures in public gatherings, border controls, etc, from the dataset Oxford COVID-19 Government Response Tracker. The variable of Lockdown is generated by Eqs. (19), (20), where and are the dates when the measures take effect and begin to deregulate, respectively.

Finally, NOx density in the troposphere is defined as a proxy variable, as shown in Eq. (21), to dynamically measure the population movement intensity. The troposphere NOx column density data are calculated from the OMINO2D level-3 products from remote sensing satellite Aura OMI, where the daily data file is assembled into HDF5 formation with a resolution of 0.25° × 0.25°. Cars, trucks, power plants, and other industrial facilities emit nitrogen oxides (NOx) as a product of burning fossil fuels (Ogen, 2020; Wang et al., 2011). Therefore, NOx levels will decrease when businesses and factories are closed, or when there are few vehicles on the road. Similarly, a decrease in NOx level suggests a lower movement intensity.

Results

Baseline results

To avoid the potential multicollinearity, we conduct a VIF test. The results show that the VIF values for all the variables are below 2.5, which is much smaller than the threshold value (10). The effects of temperature and relative humidity on COVID-19 transmission are captured using Eq. (1), as shown in Fig. 2 . The average daily temperature is significantly negatively correlated with the new daily cases fraction and R0 (Fig. 2(1) and (2)). Given an average incubation period of 6 days, every degree Celsius increase in daily average temperature with a 6-day lag results in a 2.88% (95% C.I.: [−3.12%, −2.64%], p-value < 0.0001) decrease in new daily case fraction (supplementary Table S1) and 0.62 percent point decrease (95% C.I.: [−0.68, 0.56], p-value < 0.0001) in R0 (supplementary Table S2). In comparison, every one percent point increase in relative humidity with 6-day lag causes a 0.19% decrease (95% C.I.: [−0.29, −0.10], p-value = 0.0093) in new daily case fraction and 0.022 percent point decrease (95% C.I.: [−0.045, 0.00123], p-value = 0.063) in R0. Considering that the channels of COVID-19 transmission are through direct contact, and droplet and possible aerosol transmissions, a higher temperature and a higher relative humidity could decrease virus stability and viability on the surface of containments in the public (Chia et al., 2020), thereby indirectly decreasing its transmission efficiency on its host. Weather conditions, which are probably not the determinant factor, can indeed modulate the transmission to some extent. Compared with similar studies on other respiratory pathogens, our results are partially consistent with the evidence concerning SARS (Chan et al., 2011) and Influenza (Lowen et al., 2007; Shaman and Kohn, 2009).

Fig. 2

Effects of temperature and relative humidity on COVID-19 transmission. Note: Average temperature effect on the natural logarithm of (ln) new cases fraction (1) and R0 (2). Relative humidity effect on ln new cases fraction (3) and R0 (4). The points and error bar are the estimated value with 95% C.I. 5–10 day lagged variables of average temperature and relative humidity are added in the linear form separately. Besides, Fig. 2(1) and (2) control GRP per capita, population concentration, elder population ratio, elevation, government intervention, and active case fraction while positive case fraction is excluded in Fig. 2(2) and (4). The observation selection criterion is that when total cases exceeding 100. Time fixed effect is included in the model. The regression table of the model with 6-day lag can be found in Supplementary Table S2 and S3.

We estimate the model by specifying a quadratic polynomial of temperature to verify whether there is a non-linear relationship between temperature and the transmission rate. The partial relationships between temperature and transmission that filtered the effect of other explanatory variables can be found in Fig. 3 (1) and (2). In Fig. 3(1) the quadratic polynomial curve is compared with a linear one (see the blue line), while in Fig. 3(2), the response curve estimated by the quadratic polynomial is similar to the linear curve for R0. Both curves indicate a decreasing marginal temperature effect on transmission, and the effect of high temperature is weaker than that of low temperature. For the current samples, the results show that there is no evidence of a strong non-linear relationship between temperature and transmission rate. A reasonable explanation is that most observations of average temperature are below 25 °C due to the period of the sample, i.e., as of 31, May 2020, so that limited high-temperature observations are found in the sample, thereby leading to uncertainty around the effect of high temperature.

Fig. 3

Temperature effects (partial relation) on COVID-19 transmission in linear and quadratic polynomials. Note: Average temperature effect on the natural logarithm of (ln) new cases fraction (1) and R0 (2). Marginal temperature effect on ln new cases fraction (3) and R0 (4). 6-day lagged variable of average temperature and relative humidity are added to the model by fitting Eq. (1). Other specifications are consistent with Fig. 2. In Fig. 3 (1) and (2), dependent variables were filtered for the estimated effect of the explanatory variables other than temperature. The filtered values were then normalized to have zero mean. The regression table of the model with 6-day lag can be found in Supplementary Table S4 and S5.

The effects of weather conditions with 6-day lag on transmission are simulated using a SIER dynamic transmission model (Fig. 4 ). Compared with the baseline scenario, a 2- and a 5-degree Celsius increase in ambient temperature would delay the peak day of new daily cases by 10 and 30 days, respectively, while a 2- and 5-degree Celsius decrease in ambient temperature would bring the peak forward by 8 and 18 days, respectively (Fig. 4(1)). However, when relative humidity changes from a 30% decrease to a 30% increase, the deviation of peak day is much smaller (at most 10 days) than that of temperature change (Fig. 4(2)). The shapes of the infected growth curve are not visually different. As for the total infection fraction, a lower temperature results in a much higher fraction of infection compared with the baseline scenario (Fig. 4(3)), while a 30% increase in relative humidity does not result in visually different curves (Fig. 4(4)).

Fig. 4

Simulation of temperature and relative humidity effect on SIER model. Note: Dynamic daily infected fraction under different ambient temperatures (1) and relative humidity (2). Total infected fraction under different ambient temperatures (3) and relative humidity (4). The figures are simulated based on the result of Fig. 2 (1) and (2). The number in parentheses denotes the difference in peak day of a daily infected fraction compared with the baseline scenario.

Contribution of economic condition and government intervention in weather effects

Estimates of the temperature and humidity effects on transmission by economic level are shown in Fig. 5 . It can be seen that after other variables are controlled, the new daily cases in the low-income group would increase by 3.90% (95% C.I. [3.50, 4.20]), compared with a 2.60% (95% C.I.: [2.35, 2.85], p-value < 0.0001) increase in the high-income group when the temperature falls by 1 °C (Fig. 5(1)). However, the humidity effect in high-income countries is greater than that in their low-income counterparts. The point estimates suggest that when relative humidity decreases by 1 percentage point, the new daily cases in the high-income group would increase by 0.36% (95% C.I. [0.25, 0.46], p-value < 0.0001) more than that in low-income countries. Nevertheless, since the 95% confidence interval of the difference in humidity effect by low-income contains the zero value (Fig. 5(2)), we cannot decisively reject the null hypothesis that there is no significant humidity effect in the low-income group. Our findings also revealed that different income levels can make a significant difference in the weather effects on virus spreading and that lower-income regions will face a higher risk when the temperature falls. A possible explanation for this might be that households with higher income may adjust their indoor temperature via air conditioner or heating devices, and may also pay more attention to precautionary measures based on better media communication and higher educational attainment.

Fig. 5

The economic condition in the weather-COVID-19-transmission relationship. Note:Effect of average temperature (1) and relative humidity (2) on the natural logarithm (ln) new daily case fraction. The points and error bars are the estimated value with 95% C.I. Other specifications are consistent with Fig. 2.

Governments play a crucial role in the COVID-19 outbreak (Table 2 ). A one percent point increase in government intervention intensity leads to a 0.54% decrease (95% C.I.: [−0.61, −0.48], p-value < 0.0001) in new daily cases and a 0.34 percent point decrease (95% C.I.: [−0.36, −0.33], p-value < 0.0001) in R0. Additionally, it is found that population movement is positively associated with the COVID-19 outbreak (Table 2). Population movement is positively associated with the COVID-19 outbreak (Table 2). NOx density in the troposphere, which is highly correlated with transportation activities, is defined as a proxy variable to measure the population movement intensity. Here, we find that 1 unit (1015 molec/cm2) increase in troposphere NOx density would increase 1.7% of new daily confirmed cases and 0.52 percent point of R0. This is possible that a strict government policy in public health and social distancing would limit the possibility of transmission by direct contacting when no active precautionary measure is implemented.

Table 2

Effects of population movement and government intervention effect in temperature/humidity-transmission.

Variables	(1)	(2)	(3)	(4)	(5)	(6)
Tmean Lag 6 day	−0.0294 (p < 0.0001)	−0.0288 (p < 0.0001)	−0.0296 (p < 0.0001)	−0.0074 (p < 0.0001)	−0.0062 (p < 0.0001)	−0.0072 (p < 0.0001)
RH Lag 6 day	−0.0032 (p < 0.0001)	−0.0019 (p < 0.0001)	−0.0032 (p < 0.0001)	−0.0003 (0.017)	−0.0002 (0.063)	−0.0004 (0.0019)
Lockdown		−0.5445 (p < 0.0001)			−0.3414 (p < 0.0001)
NOX			−0.0169 (0.0045)			−0.0052 (0.00015)
Control	Yes	Yes	Yes	Yes	Yes	Yes
Time Fixed Effect	Yes	Yes	Yes	Yes	Yes	Yes
N	21180	21180	21154	41419	41209	41148

Note: The dependent variable is ln new daily case fraction from Column (1) to (3) and basic reproductive number (R0) from Column (4) to (6). In Column (1) to (6), 6-day lagged variables of temperature and relative humidity are added to the model. Besides, it controls GRP per capita, population concentration, school and labor-age population ratio, elevation, government intervention, and active case fraction in column (1) to (3) while it excludes positive case fraction in column (4) to (6). The observation selection criterion is when total cases exceeding 100. Time fixed effect is included in the model. p-values (two-tailed) in parentheses.

Robustness checks

Our robustness checks for the weather-transmission relationship can be found in Fig. 6 . We used extreme temperature indexes rather than average value to examine the relationship. The temperature effects are significantly negatively associated with both new daily cases and R0. When one increases the threshold to 200 and 300, respectively, the corresponding results are consistent with what we obtained above, suggesting that our result is robust with respect to the selection of maximum and minimum temperature and the starting date. Alternatively, we examined the temperature and humidity effect on COVID-19 transmission by setting as 0.01, 0.03, 0.07, and 0.1. The results showed a significant negative relationship between weather and new daily cases (Table 3 ), and the results for R0 are similar (Table 3), indicating that our specification of the initial values of lockdown does not affect the robustness of the relationship.

Fig. 6

Robustness checks for temperature-transmission relationship. Note: Effect of maximum/minimum temperature on the natural logarithm (ln) new daily case fraction (1) and R0 (3). Effect of average temperature on ln new daily case fraction (2) and R0 (4) with the threshold is equal to 200 or 300. The points and error bars are the estimated value with 95% C.I. Other specifications are consistent with Fig. 2.

Table 3

Temperature/Humidity-COVID transmission under Different Initial Values of F

Variable	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Tmean Lag 6 day	−0.0288 (p < 0.0001)	−0.0294 (p < 0.0001)	−0.0297 (p < 0.0001)	−0.0300 (p < 0.0001)	−0.0062 (P < 0.0001)	−0.0063 (P < 0.0001)	−0.0065 (P < 0.0001)	−0.0067 (P < 0.0001)
RH Lag 6 day	−0.0019 (p < 0.0001)	−0.0022 (p < 0.0001)	−0.0023 (p < 0.0001)	−0.0025 (p < 0.0001)	−0.0002 (0.063)	−0.0002 (0.054)	−0.0002 (0.055)	−0.0002 (0.065)
Lockdown (F0 = 0.01)	−0.5445 (p < 0.0001)				−0.3414 (P < 0.0001)
Lockdown (F0 = 0.03)		−0.4648 (p < 0.0001)				−0.3751 (P < 0.0001)
Lockdown (F0 = 0.05)			−0.4186 (p < 0.0001)				−0.3916 (P < 0.0001)
Lockdown (F0 = 0.1)				−0.3508 (p < 0.0001)				−0.4152 (P < 0.0001)
Control	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time Fixed Effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	21180	21180	21180	21180	41209	41209	41209	41209

Note: The dependent variable is ln new daily cases fraction from Column (1) to (4) and basic reproductive number (R0) from Column (5) to (8). 6-day lagged variables of temperature and relative humidity are added to the model. Other specifications are consistent with Table 2p-values (two-tailed) in parentheses.

Prediction of temperature and humidity effect on COVID-19 transmission

The risks attributable to temperature and humidity on transmission in winter and summer were forecasted to assess the maximum possible transmission risk resulting from weather conditions in terms of seasonal cycles (Fig. 7 ). It can be seen that the spread of coronavirus slows down in summer, while a lower temperature accelerates its spread in other seasons. In July the value of average risk will increase by 45% (mean value, [10%, 79%]) in Australia, Southern America, and Southern Africa, indicating that the confirmed fraction would increase by 45% compared with the benchmark condition. In January the North American region and the northern region of the Euro-Asia continent will face a greater risk, with an average risk value of 87% (Mean value, [34%, 140%]). Considering the higher population concentration, we predict that the northern hemisphere will be at a greater risk of transmission in winter. Furthermore, poor regions will be likely exposed to a higher risk driven by weather conditions.

Fig. 7

Regional projection of temperature and relative humidity effect in summer and winter. Note: The colors denote the effect of temperature and humidity on peak new daily case compared with the benchmark weather conditions. The risk in winter and summer are calculated based on the historical average temperature and relative humidity in July and January 2019. The map division is only a schematic diagram and does not indicate an accurate administrative area. Map data is from https://gadm.org/.

Discussion

Based on the data from 1236 regions in the world as of 31 May 2020, we examined the role of temperature and relative humidity in COVID-19 transmission at the global sub-national level. By explicitly controlling for social-economics variables and government interventions, we found the negative effects of temperature and humidity on the COVID-19 transmission. Several points are worthy to discuss.

As for the data period, since the COVID-19 in many European and Asian countries was controlled by effective government intervention at the end of May 2020, regional cases data in several countries are not updated timely. This point limits our data coverage. Further, the border among countries has been reopened in July, 2020. As a result, the role of population movement, which would be the potential factor to COVID-19 transmission, was not well measured since the unavailable data for trans-border traffic. A biased in population movement could lead to a biased estimation of the weather effect on COVID-19 transmission. Although our data terminates in May, it is likely that the climate condition especially for temperature, plays a vital role in the transmission while most countries implement strict custom control and boarding regulation, making the trans-countries transmission a low possibility. As of 15 October 2020, 6 southern hemisphere countries (Brazil, Argentina, Colombia, Peru, South Africa, and Chile) are listed among the top 15 countries by confirmed cases. These southern hemisphere countries have suffered a rapid increase in confirmed cases in the last 5 months in the winter. Although our data misses a full seasonal cycle, it does contain regional samples covering both summer and winter from both the southern and northern hemispheres. We thus believe that our conclusion is reliable enough to consider making the public aware of the possibility that the COVID-19 transmission could worsen during winter.

Furthermore, we conducted several robust checks regarding the relationship between the weather condition and COVID-19 transmission. First, both maximum and minimum temperature were used separately to replace the average temperature to examine the relationship. Second, we experimented with different settings in measuring government intervention to examine whether the relationship is stable with regard to the choice of function. Thirdly, considering the data delay at the early stage, the relationships were examined under different thresholds of starting points. Our conclusions were shown to be robust to these specifications of the empirical model.

The evaluation of the effectiveness of government response is also a vital point. Currently, the numbers of total COVID-19 cases still keep growing among some developed countries in the warming northern hemisphere, which is likely due to ineffective government responses and overwhelmed health systems. The variance in COVID-19 spreading across regions is thus not only attributable to weather conditions, but also social-economic situations and government intervention to a significant extent. It is no doubt that government responses to COVID-19, including contact tracing, quarantine, and social distancing contributed substantially to this (Giordano et al., 2020; Kraemer et al., 2020; Wu and McGoogan, 2020). Moreover, some studies (Iqbal et al., 2020; Pan et al., 2021; Sobral et al., 2020) concerned with the government response (lockdown) ignored the dynamic effect of government intervention, thereby probably overestimate the weather effect. To address this issue we constructed two proxies of government intervention by combining multi-source data currently available from large-scale remote sensing satellite and grid data, thus providing an effective and robust estimation of the weather-transmission relationship.

We also modified the dynamic transmission model by adopting more reliable parameters of weather conditions. The current dynamic transmission models assume that the weather effect is invariant, which leads to a significant deviation in forecasting. In fact, weather conditions frequently vary daily. With the proposed model in this study, we are able to estimate the temperature effect more accurately by explicitly controlling for the underlying factors listed above, and conclude that weather conditions can significantly shape the transmission curve and alter the peak prevalence. More precisely, the lower the ambient temperature, the earlier the transmission rate peak will appear. This poses a great challenge to all regions globally, with the specific extent of risk depending on the regional social-economic conditions. According to our study, COVID-19 is more likely to occur again in regions with high latitudes in the northern hemisphere in winter.

Our model assumes that there is no vaccine available. Theoretically, A candidate vaccine against SARS-CoV-2 might act against infection, disease, or transmission, and a vaccine capable of reducing any of these elements could contribute to disease control. As Hodgson et al. (2020), Peiris and Leung (2020) mention, the transmission would not be terminated if effective vaccines are available. The first concern is that a SARS-CoV-2 vaccine could reduce the severity of the diseases but lead to prolonged shedding of infectious virus, which could have important consequences for public health if shedding resulted in increased transmission. The second is COVID-19 vaccine allocation strategy and affordability. As is seen, vaccine of COVID-19 will be expensive and supply-limited if released in the early stage. How to allocate the vaccine to reduce the risk of transmission still calls for huge consideration. In these possible situations, our research concludes that the weather condition would play a greater role in the transmission and the health inequality would be increased around the world.

In addition, our results can provide several valuable pieces of information to help epidemic prevention and public health intervention. Firstly, the less developed and developing countries are likely to suffer more from COVID-19 when the temperature drops. Given the diversity in both weather and social-economic conditions, the transmission risk will vary globally and possibly amplify the existing global health inequality. Therefore, more attention should be paid to low-income regions, especially for the case of Africa (Gilbert et al., 2020; Nkengasong and Mankoula, 2020). Secondly, COVID-19 transmission will accelerate in the winter. In the northern hemisphere, especially in the temperate and sub-cold zones, the possibility of COVID-19 recurrence in the winter deserves special attention. As for other regions, there is still a long-term need to deal with the importation of risk via travelers from high-risk areas. In the summer higher temperatures may help control the disease spreading in the northern hemisphere. Therefore, a second-wave pandemic is likely to occur in the winter again. Effective government intervention and public awareness about COVID-19 are necessary to mitigate transmission (Jayaweera et al., 2020). As the seasonal cycles vary between the northern and southern hemispheres, government intervention in the spread of COVID-19 ought to be dynamically adapted.

Conclusion

We quantitatively analyze the role of weather conditions in the spread of COVID-19 in a global context by controlling for a rich set of social-economic variables. We found that higher temperature and higher humidity could reduce transmission. Additionally, there was no evidence of a strong non-linear relationship between temperature and transmission rate in our current sample. Arguably our research provided more reliable conclusions with regard to the temperature and relative humidity effects. From this aspect, the application of merging multi-source data from remote sensing, statistic indexes, and gridded-visualized data can provide a powerful tool and new information in environmental evaluation, allowing for more flexible statistical methodologies with higher dimensional observations and giving more reliable conclusions at a low cost.

Our estimates provide more practical parameters to identify the possible risk over the post pandemic period and forecasted the tendencies in the future. Temperature and relative humidity are shown to be negatively correlated with COVID-19 transmission throughout the world. Weather conditions are not the decisive factor in COVID-19 transmission, in that government intervention as well as public awareness, could contribute to the mitigation of the spreading of the virus. As temperature drops in the winter, the transmission will possibly speed up again., and thus deserves a dynamic government policy to mitigate COVID-19 transmission.

Some limitations of the present study should also be pointed out. First, our conclusions are based on observations over certain periods, More specifically, our conclusions are drawn under the assumption of no available COVID-19 vaccine and a consistent and positive government response to the virus. How one would indeed separate the contribution of social-distancing from endogenous immune drivers presents a formidable challenge. Secondly, our data were obtained from daily reports, in which the individual clinical information (e.g., channels of infection, age, and burden of chronic diseases) was missing. Therefore, the heterogeneity in individuals was not considered. Thirdly, our conclusions are drawn on statistical models, but it still requires epidemiological analysis or random control experiment to explore the effect of weather more precisely. Finally, we hope to explore the underlying non-linear effect of temperature on COVID-19 transmission in the future with more available datasets.

CRediT authorship contribution statement

Chen Zhang: Data curation, Formal analysis, Writing - original draft, designed the study. collected the data. analyzed the data. drafted the manuscript. contributed to the revision of the manuscript. All authors contributed to the interpretation of the results and approved the final version. Hua Liao: contributed to the, Supervision, Data curation, Formal analysis, designed the study. collected the data. analyzed the data. contributed to the revision of the manuscript. All authors contributed to the interpretation of the results and approved the final version. Eric Strobl: Data curation, Formal analysis, analyzed the data. contributed to the revision of the manuscript. All authors contributed to the interpretation of the results and approved the final version. Hui Li: Data curation, collected the data. contributed to the revision of the manuscript. All authors contributed to the interpretation of the results and approved the final version. Ru Li: Data curation, collected the data. All authors contributed to the interpretation of the results and approved the final version. Steen Solvang Jensen: contributed to the revision of the manuscript. All authors contributed to the interpretation of the results and approved the final version. Ying Zhang: conceived the idea of this work. contributed to the, Supervision, designed the study. contributed to the revision of the manuscript. All authors contributed to the interpretation of the results and approved the final version.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.