Results

On average across our sample, confirmed COVID-19 cases grow at a rate of 13.2%/d (SD of 24.4%), equivalent to a doubling time of 5.2 d. Growth rates generally decreased over the first months of the outbreak, with the sample average growth rate falling from 15.7% during the first month of a location’s outbreak to 2.3% in subsequent months (Fig. 2B). These declines are consistent with the epidemiological dynamics of the virus (SI Appendix, section A.1) and with strengthened containment efforts over time, although they may also reflect other factors.

Applying a panel-regression model to growth rates (Materials and Methods and SI Appendix, section A.2) demonstrates a statistically significant effect whereby increases in surface UV intensity lower subsequent COVID-19 growth rates. In our primary specification, we estimate that a 1 $$ \mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$ increase in local UV reduces local COVID-19 growth rates by 0.09 ($$ \pm $$0.04, $$ P=0.01$$, range indicates $$ \pm $$ 1 SD) percentage points over the ensuing 17 d (Fig. 3A and column (col.) 3 of SI Appendix, Table S1). The effects associated with UV are consistently negative across lags and peak in magnitude after 9 to 11 d (Fig. 3B and SI Appendix, Fig. S5). This delay between UV exposure and changes in the COVID-19 growth rate is consistent with the reported time frame between exposure to the virus and its detection (36, 47, 48). The estimated UV effects imply that a sample SD increase in UV (10.73 $$ \mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$, accounting for semiparametric controls) reduces growth by 0.97 ($$ \pm $$0.38) percentage points or 7% of the sample average growth rate of 13.2%. This amounts to an increase in doubling time of COVID-19 cases from 5.2 d—at the average growth rate—to 5.7 d ($$ \pm $$0.2). In contrast, the effects of higher temperatures and higher levels of specific humidity are of less consistent sign, with cumulative effects over the 17-d interval being statistically insignificant and of opposite sign to that of the lag with the greatest magnitude (Fig. 3 A and B and SI Appendix, Table S1). A SD increase in temperature (2.84 °C) and specific humidity (0.13%) leads to small and uncertain estimated effects on the growth rate of 0.49 and 0.29 percentage points, respectively.

Fig. 3.

Empirical estimates of the relationship between COVID-19 and local environmental conditions. (A) The estimated cumulative effect of a change in daily UV, temperature, or specific humidity on subsequent daily COVID-19 growth rates over the following 17 d. All estimates are standardized, such that effect sizes represent percentage point changes per SD ($$ \sigma $$). Our central estimate (SI Appendix, Eq. S1) is shown by the large circle, along with vertical lines representing the 95% confidence interval, which is calculated allowing for arbitrary correlation of model errors within administrative units over time. Smaller circles show estimates of the cumulative effect from alternative plausible statistical model formulations that, among other things, vary the stringency of the spatiotemporal controls or additionally control for the local timing of COVID-19 outbreaks, testing regimes, or COVID-19 containment policies (SI Appendix, Table S1, cols. 1–7 and Figs. S5–S7). (B) The temporal dynamics of each estimated lagged effect of UV, temperature, and specific humidity on the daily COVID-19 growth rate over 17 d for our central estimate (thick lines, one for each coefficient in SI Appendix, Eq. S1) and alternative model formulations (thin lines, same alternative models as in A). Coefficients have been divided by three to show per-day effects. The displayed curve is a fit to the estimated lag coefficients from our central estimate (SI Appendix, section A.3). C replicates the cumulative effect of each weather variable on daily COVID-19 growth rates from the primary specification in A in gold (UV), maroon (temperature), and green (specific humidity). In purple, treatment effects are reported for the period before an administrative unit imposed any social distancing measures (large purple diamond) and after such measures were put in place (small purple diamond). Similarly, in light green, treatment effects of each weather variable are reported for the first 30 d of the location-specific outbreak (large green square) and for all dates after the first 30 d (small green square). Vertical lines indicate 95% confidence intervals. Arrows indicate where confidence intervals have been truncated for display. Effects of social distancing policies and outbreak duration on individual lag coefficients for all three weather variables are shown in SI Appendix, Fig. S8.

The effect of UV radiation on the COVID-19 growth rate (Fig. 3 A and B) is robust to a range of alternative statistical models, including controls for days since initial outbreak of the virus in each location (col. 1 in SI Appendix, Table S1); linear, country-specific time trends (col. 2 in SI Appendix, Table S1); controls for future weather as a test of reverse causality (col. 4 in SI Appendix, Table S1); controls for temporally and spatially varying policies, such as work-from-home policies, school closures, and event cancellations (col. 5 in SI Appendix, Table S1 and col. 2 in SI Appendix, Fig. S5); controls for the extent of COVID-19 testing availability (col. 6 in SI Appendix, Table S1); semiparametric controls that allow for within-country differential seasonality, in addition to country-wide seasonal patterns (col. 7 in SI Appendix, Table S1 and col. 3 in SI Appendix, Fig. S5); removal of country-level data and addition of semiparametric controls that allow for country-by-day specific shocks (col. 8 in SI Appendix, Table S1 and col. 4 in SI Appendix, Fig. S5); analysis using a Poisson regression model (SI Appendix, Eq. S2) in place of ordinary least squares (col. 9 in SI Appendix, Table S1 and col. 5 in SI Appendix, Fig. S5); replacing specific humidity with relative humidity (SI Appendix, Fig. S6); and changing the number of lags included in the estimating equation (SI Appendix, Fig. S7). We further show our estimates are insensitive to outliers using a procedure whereby we reestimate our cumulative effect after systematically dropping each of our 3,235 geospatial units (SI Appendix, Fig. S9). Finally, we estimate an alternative model that allows for nonlinearities between weather conditions and COVID-19 growth rates and find that the UV effect exhibits strong linearity (SI Appendix, Fig. S10 and Eqs. S3–S4). Whereas the significance and magnitude of the cumulative UV effect are stable across the different model specifications, the cumulative effects of temperature and humidity are insignificant across all model specifications and have inconsistent sign (Fig. 3A and SI Appendix, Table S1).

In contrast to UV estimates being insensitive to the addition or modification of controls, omitting location and time fixed effects or omitting temporal trends leads to substantially biased estimates of the environmental determinants of transmission compared to our primary specification. In Fig. 2C we show three sets of response functions relating daily average UV exposure to its cumulative effect on subsequent daily COVID-19 growth rates, where responses are centered relative to the in-sample average UV value of 50 $$ \mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$. When all semiparametric controls are omitted (teal line in Fig. 2C, corresponding to data shown in Fig. 2A, Left), the estimated effect of UV is of the opposite sign to that hypothesized in prior literature and estimated in our primary specification (gold line in Fig. 2C, corresponding to data shown in Fig. 2A, Right). Similarly, omission of temporal controls (brown line in Fig. 2C, corresponding to data shown in Fig. 2A, Center) overestimates the negative effect of UV on COVID-19 growth rates, as time-trending unobservable factors, such as changes in reporting and the general progression of the disease, correlate with gradual changes in UV exposure. These results highlight the empirical importance of adequately removing the influence of key confounding factors that have to date limited the ability to determine whether and how environmental conditions constrain the evolution of COVID-19 (13, 14).

The cumulative lagged effect of weather conditions on COVID-19 growth rates reflects the average treatment effect over all geospatial units and over the course of the observed pandemic (Fig. 3A). It can be inferred, however, that effective social distancing policies will reduce any relationship between UV exposure and transmission of COVID-19. Transmission must be nonnegative; thus, any single reduction limits the scope for further reduction. Consistent with this, we find suggestive evidence that social distancing policies such as school closures, mandatory work from home orders, and large event cancellation regulations weaken the link between COVID-19 and weather conditions. Specifically, using a binary policy variable indicating whether an administrative unit has any one of a set of social distancing measures in place (SI Appendix, section B.3), we find that the cumulative effect of 1 $$ \mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$ increase in UV falls from −0.09 ($$ \pm $$0.04, $$ P$$ = 0.02) before policies are introduced to $$ -0.04$$ ($$ \pm $$0.06, $$ P$$ = 0.52) after policies are put in place (Fig. 3C).

Similarly, the effect of UV exposure on transmission of COVID-19 is likely to decline over the course of the pandemic, as social distancing policies are enacted and individuals gain more awareness of and information about the virus. Indeed, we find suggestive evidence that the cumulative effect of UV on the growth rate of COVID-19 cases is stronger during the first month of a population’s outbreak (−0.14 $$ \pm $$ 0.04) than in subsequent months (−0.08 $$ \pm $$ 0.04). The pattern of effect attenuation shown in Fig. 3C is also observed for the individual lags of temperature and specific humidity that have the largest magnitude (SI Appendix, Fig. S8), although cumulative effects of temperature and specific humidity are statistically indistinguishable from zero both with and without public health policies in place (Fig. 3C).

The estimated effect of UV on the COVID-19 growth rate has seasonal implications (Fig. 4). To illustrate the role of changing UV in the evolution of the disease over the year, we use the cumulative effect of UV recovered in Fig. 3A, along with the local seasonal climatology of surface UV intensity between January and June, to quantify how seasonality in UV altered daily COVID-19 growth rates across the world (SI Appendix, section A.4). This period, besides encompassing our entire data period, also covers the full range of seasonal UV exposure experienced in any location, as shown in Fig. 4A. Between January and June, the increase in seasonal UV exposure ($$ +81.0\hspace{0.17em}\mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$) lowered extratropical Northern Hemisphere (above $$ \mathrm{23}°$$ north) COVID-19 growth rates by 7.4 ($$ \pm $$2.9) percentage points. Over the same period, the seasonal decline in UV ($$ -80.7\hspace{0.17em}\mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$) raised growth rates by 7.3 ($$ \pm $$2.9) percentage points in the extratropical Southern Hemisphere (below $$ \mathrm{23}°$$ south) (Fig. 4 B and C). This seasonal change amounts to an increase in the doubling time from an average of 5.2 to 11.8 d in the extratropical Northern Hemisphere and a corresponding decrease to 3.4 d in the extratropical Southern Hemisphere. Seasonality in UV in the coming boreal winter reverses this pattern. Between June and December, our estimates imply that COVID-19 growth rates increase by 7.8 ($$ \pm $$3.0) percentage points in the extratropical Northern Hemisphere and fall by 7.7 ($$ \pm $$3.0) percentage points in the extratropical Southern Hemisphere from changes in UV (SI Appendix, Fig. S11). These changes in COVID-19 growth correspond to lowering the average doubling time to 3.3 d in northern latitudes and raising it to 12.6 d in southern latitudes.

Fig. 4.

Seasonality in the simulated COVID-19 growth rate. (A) The average seasonal cycle of UV, temperature, and specific humidity for northern (above $$ \mathrm{23}°$$ north) and southern (below $$ \mathrm{23}°$$ south) latitudes. (B) The individual impacts of seasonal changes in UV (gold), temperature (maroon), and specific humidity (green), as well as their combined effect (black), from January to June. Points indicate average simulated impacts for northern latitudes, the tropics ($$ \mathrm{23}°$$ south to $$ \mathrm{23}°$$ north), and southern latitudes. Horizontal lines show 95% confidence intervals, which account for uncertainty in statistical parameters. (C) Map of the influence of expected seasonal changes in UV alone on the COVID-19 growth rate from January to June.

As a whole, the tropics display moderate seasonal changes driven by UV, with our simulations generating an increase of 0.57 ($$ \pm $$0.22) percentage points between January and June due to an decrease in UV exposure of $$ 6.3\hspace{0.17em}\mathrm{k}\mathrm{J}\cdot {\mathrm{m}}^{-2}\cdot {\mathrm{h}}^{-1}$$. A notable regional exception is that the onset of the South Asian monsoon causes decreased surface UV regionally in June, thus raising summer COVID-19 risks. We emphasize that these simulations are merely illustrations of the potential seasonal influence of UV. Changes in population immunity rates, genetic mutations of the virus, and public health policies, among many other factors, could alter the sensitivity of COVID-19 to environmental conditions, causing future seasonal implications to differ from those derived over our sample period.

Other seasonally varying climate variables may have also influenced COVID-19 cases during the first 6 mo of infection, including temperature and specific humidity (Fig. 4A), whose effects on COVID-19 transmission rates remain uncertain. Indeed, similar exercises for northern and southern latitudes using only January to June seasonality in temperature or specific humidity do not yield changes in daily COVID-19 growth rates during these first 6 mo that are statistically distinguishable from zero because the cumulative effect of each variable is uncertain (maroon and green bars in Fig. 4B). The uncertain contribution of these variables renders the total effect of seasonality across all three variables uncertain, with 95% confidence intervals for the cumulative effect containing zero for northern and southern latitudes (black bars in Fig. 4B). In the tropics, seasonality is smaller and more complex but the total effect is significant between January to June because UV, temperature, and specific humidity influences align.

Discussion

Using a global, harmonized dataset of daily COVID-19 cases, we find that the daily growth rate of confirmed COVID-19 cases responds negatively to increased UV. Importantly, variations in the COVID-19 growth rate lag variations in UV by up to 2.5 wk, consistent with times required for incubation, testing, and reporting. The UV response is robust to a range of model specifications and controls. However, the influence of other seasonally varying environmental conditions is not precisely estimated. We hope that the distributed-lag, panel-regression framework employed here may underlie further analyses of the influence of environmental conditions on COVID-19 transmission, particularly as COVID-19 data availability improves.

Our findings are consistent with the hypothesis that UV radiation alters COVID-19 transmission rates by more rapidly deactivating the SARS-CoV-2 virus residing on surfaces or in aerosol form, as suggested by recent laboratory studies (12). We cannot rule out, however, that UV may also influence the incubation period of SARS-CoV-2, testing rates, behavior such as time spent indoors or socializing (49), or other disease-transmission or monitoring properties. Our findings also indicate that climate has a modest effect on viral transmission relative to social distancing. Measurements of the effect of social distancing policies on COVID-19 growth early in the epidemic (16) are 3 to 6 times larger than the influence of UV seasonality that we estimate for the extratropical Southern and Northern Hemispheres. Due to minimal seasonality in the tropics, social distancing policy effect sizes are 35 to 85 times larger than the effect of UV seasonality on COVID-19 growth that we estimate in this region (SI Appendix, section A.4). Because factors such as social distancing policies have a larger influence than seasonal variation in UV, and high susceptibility to date among the global population permits for rapid transmission, COVID-19 growth is unlikely to exhibit substantial seasonality, at least in the near term (50, 51). If COVID-19 becomes widely established, environmental influences may become more important for inducing seasonal variations in the growth of infections (50).

Our study has a number of important limitations. First, as is true in any empirical study of disease, we can only observe cases that are confirmed. The fact that confirmed cases of COVID-19 are likely well below the actual number of infections (52) would not, of itself, affect estimates of the growth rate if confirmed cases were a constant ratio of the actual value. However, other factors such as variations over time in the rates of testing or testing procedures (53, 54) could alter observed growth rates. In some settings, the bias in growth rates due to such time-varying underreporting has been found to be quantitatively small (16). Moreover, our research design ensures that such imperfect reporting does not systematically bias our estimated effects of environmental conditions, provided that testing procedures or reporting practices are uncorrelated with climatological variables (SI Appendix, section A.2). We additionally address this concern statistically by accounting for location-specific trends in confirmed COVID-19 cases and by controlling for the availability of COVID-19 testing at the country level (SI Appendix, section B and Table S1), although reporting issues could remain.

Second, it is possible that the behavioral response to random day-to-day fluctuations in UV (and other environmental variables) differs from the behavioral response to expected seasonal changes. For example, an anomalously sunny day in March may elicit different human behavior than a day in July with the same UV exposure. It may be possible to estimate such state-dependent effects after the accrual of multiple seasons of data. There is also a potential concern that the slow response of the dynamic system of the disease would damp the amplitude of the response to high-frequency day-to-day environmental fluctuations. We find using stochastic simulations of the SEIR model, however, that simulated UV-induced variations in transmission are generally insensitive to the frequency of perturbations (SI Appendix, section A.1 and Fig. S13G).

Third, some studies suggest a relationship between air pollution and COVID-19 transmission and mortality (555657–58). Although it is theoretically possible that the negative effect of UV that we recover is partially explained by air pollution attenuating UV and increasing COVID transmission, this is unlikely to be the case, given that day-to-day variation in UV is driven primarily by changes in cloud cover, with much smaller contributions coming from variations in ozone, aerosols, and water vapor (59).

Finally, although we know of only one publicly available laboratory study of the UV–COVID-19 transmission relationship (12), we view our approach as complementary. Although laboratory studies isolate the biology of virus transmission, our statistical approach using observed COVID-19 cases captures those channels as well as behavioral adjustments individuals make in response to short-term UV fluctuations, such as decisions to spend time indoors or outside, to exercise, or to attend social gatherings, and other activities and health investments (49, 60). As public health officials grapple with the costs and benefits of a range of possible responses to the current pandemic, quantifying the influence of both biologically direct and behaviorally induced modification channels is essential to building appropriate policies.

Materials and Methods

To construct a harmonized global dataset of geolocated daily confirmed COVID-19 cases, we assemble publicly available data from national governments, subnational authorities, and newspapers. Subnational sources are described in SI Appendix, Table S2; for all countries for which subnational records were not publicly available at the time of writing, we use national-level records provided by the Johns Hopkins University Center for Systems Science and Engineering (61). No statistical methods were used to predetermine sample size.

We link COVID-19 case data to gridded daily weather data from the European Centre for Medium-Range Weather Forecasts Reanalysis 5th (ERA5) product (17) by calculating the population-weighted average UV, temperature, specific humidity, and precipitation for each day across all grid cells within each administrative unit (SI Appendix, section B). We additionally combine case records with data on location-specific containment policies and testing regimes from refs. (15, 16).

UV radiation is represented as including wavelengths from 200 to 440 nm. Limitations associated with the representation of radiative transfer associated with ERA5 reanalysis prevent us from distinguishing between UVa (400 to 315 nm), UVb (315 to 280 nm), and UVc (280 to 100 nm) effects. Higher-energy UVc and UVb radiation may more rapidly deactivate SARS-CoV-2 than UVa but is also more readily absorbed in the atmosphere (12, 62). A more detailed analysis of the transmission and reflection of different types of UV in association with destruction of the SARS-CoV-2 is a fruitful area for future research.

We statistically estimate the effect of weather on the daily growth rate of confirmed COVID-19 cases using a longitudinal (i.e., panel) regression model. Daily COVID-19 growth rates are estimated as a linear function of UV, temperature, specific humidity, and precipitation exposure over the preceding 17 d in a model allowing the effect of each environmental variable to differ across lag intervals of 3 d (SI Appendix, Eq. S1). We include indicator variables (i.e., fixed effects) for each subnational or national administrative unit, for each day of the sample, and for each country by week in the sample. We calculate standard errors accounting for serial correlation across days within each administrative unit. Clustering at the country level, which further allows for spatial correlation of arbitrary form within administrative units from the same country, does not discernibly change the precision of our estimates. All P values are calculated using two-sided tests. We test for heterogeneity in the estimated effect of weather conditions on COVID-19 growth rates by interacting the lagged weather variables with binary variables indicating whether containment policies are in place and whether the observation is in the first month of the outbreak (SI Appendix, section A.2).

Seasonal simulations (Fig. 4) use the daily seasonal climatology of UV, temperature, and specific humidity, which we calculate by averaging daily data from the ERA5 reanalysis product over the years 2015 to 2019 (SI Appendix, section A.4).

Throughout the analysis daily growth rates of confirmed COVID-19 cases are shown using units of percent. For example, a growth rate of 0.13 is approximately equivalent to a growth rate of 13%. Changes in growth rates are given in units of percentage points. For example, a reduction of the growth rate by 1 percentage point would change the growth rate from 13% to 12%.