A cross-sectional analysis of meteorological factors and SARS-CoV-2 transmission in 409 cities across 26 countries

The study investigates whether meteorological conditions influenced SARS-CoV-2 transmission during the early phase of the COVID-19 pandemic. Early commentary drew analogies with seasonal respiratory viruses like influenza that peak in colder, drier conditions, but mechanisms and the role of weather versus susceptibility and behavior were unclear. Prior studies produced conflicting results and often did not account for key confounders such as government interventions, socio-economic factors, population density, or age structure. The authors aim to estimate the association between weather variables and the effective reproduction number (R_e) across 409 cities in 26 countries, explicitly controlling for non-pharmaceutical interventions (NPIs) and city-level covariates, focusing on an early window of local epidemics to minimize confounding.

Prior research on weather effects on COVID-19 transmission has been mixed: studies reported positive, negative, non-linear, or null associations between temperature and COVID-19 outcomes. Many had methodological issues, including inadequate control for NPIs and socio-demographic confounding. Some large-scale analyses found modest negative associations of temperature with incidence or no significant association after adjusting for interventions, with stronger impacts attributed to NPIs such as social distancing and school closures. Non-linear associations in US data suggested minima around 11 °C with increases up to 20 °C and declines above, and specific humidity ranges (about 7.6–11.4 g/m³) have been linked to growth rates. Overall, the literature underscores the challenge of disentangling seasonal signals from meteorological factors given short time series and variable interventions across regions.

Design: Two-stage ecological analysis across 409 cities/small regions in 26 countries, covering early epidemic windows between 11 January and 28 April 2020. Stage 1 (Estimation of transmission): For each city, the effective reproduction number (R_t/R_e) was estimated over a city-specific early window (10–20 days), starting after at least 10 confirmed cases in a 10-day period and excluding days when the OxCGRT Government Response Index exceeded 70. R_e was estimated using EpiNow2 (Bayesian latent variable renewal model): latent infections were inferred via a gamma-distributed generation time (mean 3.6 days, SD 3.1), mapped to observed cases through log-normal incubation and onset-to-report delays, with a negative binomial observation model and day-of-week effect. R_t was piecewise constant with an initial 3-day breakpoint; estimates from the remainder of the window were used. Four MCMC chains (1000 warmup, 4000 post-warmup) with convergence checks were applied. Exposures and covariates: Meteorological variables were derived from ERA5 at 0.25° resolution: daily mean temperature (°C), relative humidity (%), absolute humidity (g/m³, derived from temperature and dew point), surface solar radiation downwards (J/m²), wind speed (m/s from 10 m u,v), and total precipitation (m). City exposures were averaged over the city-specific window by assigning grid cells containing city centroids. Socio-economic/demographic covariates included total population, population density, GDP per capita, proportion of population >65 years, and PM2.5 (µg/m³) from CAMS near-real-time data; skewed variables were log-transformed. The OxCGRT Government Response Index (0–100), lagged by 10 days and taken at the last day of the city window, was used to capture NPIs. Stage 2 (Cross-sectional association): A two-level multilevel meta-regression (cities nested within countries) related city-specific R_e to each meteorological variable, allowing for the variance of R_e estimates. Non-linear associations were modeled using natural splines (0–5 internal knots; model selection by AIC). Models adjusted for OxCGRT Government Response Index, population, population density, GDP per capita, % >65, and PM2.5. Country random effects accounted for between-country differences; city random effects accounted for within-country variability. Effect sizes were derived from predicted curves as differences between maxima and minima of R_e for meteorological variables (or 5th–95th percentiles when appropriate) and for OxCGRT between the 5th and 95th percentile values. Model fit and explanatory power were summarized using likelihood ratio R² statistics (ΔR²_L) relative to a null random-effects model and to a base covariate model. Sensitivity analyses: Tested robustness including country fixed effects, restricting to weaker interventions (OxCGRT <60), excluding cities with R_e <1, excluding China and Brazil, excluding tropical or southern hemisphere cities, and applying 10-day lags to meteorological variables. Interaction between mean temperature and RH (≤65% vs >65%) was examined.

- Across 409 cities, mean R_e during the observation window was 1.4 (range 0.7–2.1), with all but 10 cities >1. - Temperature: A modest non-linear association with R_e. R_e rose to a peak at 10.2 °C and fell to a trough at 20 °C, with R_e at the trough 0.087 lower than at the peak (95% CI: 0.025; 0.148; p=0.014). Mean temperature explained 2.4% of R_e variability (ΔR²_L ≈ +2.5). - Absolute humidity: Similar non-linear pattern with a maximum difference of 0.061 (95% CI: 0.011; 0.111; p=0.036) between 6.6 g/m³ (peak) and 11 g/m³ (trough), explaining ~2.0% of R_e variability (ΔR²_L ≈ +2.0). - Relative humidity: Weak evidence (effect size 0.043; 95% CI: −0.001; 0.087; p=0.058; ΔR²_L ≈ +1.5). - Solar radiation, wind speed, and precipitation: No clear associations with R_e after adjustment. - Non-pharmaceutical interventions (NPIs): The OxCGRT Government Response Index (lagged 10 days; capped <70 in windows) showed a strong association with R_e (p<0.0001). Increasing OxCGRT from 21 (5th percentile) to 66 (95th percentile) was associated with an estimated reduction in R_e of 0.285 (95% CI: 0.223; 0.347), explaining 13.8% of the variability—about six times the variation explained by mean temperature. - City-level socio-demographic covariates explained 1.4% of R_e variability. - Sensitivity analyses: Main results robust under various specifications. Temperature association persisted across analyses; AH association was less robust (e.g., attenuated when excluding certain regions or applying meteorological lags). In some restricted analyses (e.g., OxCGRT <60), wind and precipitation showed associations. No evidence of interaction between temperature and RH categories (p=0.428).

The study addressed whether weather influenced early SARS-CoV-2 transmission by integrating high-resolution city-level transmission estimates with meteorological, socio-demographic, and intervention data. Findings indicate that while mean temperature and absolute humidity have modest, non-linear associations with R_e, their explanatory power is small relative to the effect of NPIs, as captured by the OxCGRT Government Response Index. The high correlation between temperature and absolute humidity suggests intertwined effects that are difficult to disentangle. Potential mechanisms for the observed meteorological associations include laboratory-demonstrated effects of temperature, humidity, and solar radiation on viral stability; seasonal modulation of innate and adaptive immunity; and weather-driven changes in human behavior and indoor crowding. Comparisons with prior studies reveal consistency with analyses showing limited or modest weather effects and strong impacts of interventions on epidemic growth. The results underscore that population behavior and policy interventions were the dominant drivers of transmission during the early pandemic phase, with no meteorological conditions found that obviate the need for precautions.

This global, city-level analysis using a two-stage modeling framework and rigorous control for NPIs and socio-demographic confounders finds that meteorological factors, particularly temperature and absolute humidity, have only modest non-linear associations with early-phase SARS-CoV-2 transmission. Government interventions had substantially larger effects on R_e than weather variables. The study contributes robust evidence that mitigation measures are critical irrespective of prevailing weather. Future research should leverage longer time series spanning multiple seasons to better characterize potential seasonality, further investigate indoor vs outdoor conditions (especially absolute humidity), refine measures of NPIs and behavior, and explore heterogeneity across climate zones with increased representation of tropical and southern hemisphere cities.

- Geographic representation skewed toward the northern hemisphere; fewer tropical and southern hemisphere cities reduced power in those sub-analyses. - Short early-epidemic windows limit seasonal coverage, making it difficult to separate seasonality from meteorological effects. - Case data uncertainties: varying case definitions, under-ascertainment, high proportion of asymptomatic infections, and reporting delays; although R_e estimation attempted to account for these, residual biases may remain. - Strong correlation between temperature and absolute humidity precluded mutual adjustment, limiting causal separation of their effects. - Use of city-level aggregates may introduce exposure misclassification; ERA5 grid assignment via centroids may not capture intra-urban variability. - OxCGRT index summarizes diverse NPIs and was capped <70 within windows; residual confounding by interventions and behavioral changes may persist. - Model assumptions (e.g., generation time, delay distributions, piecewise-constant R_t) and choice of 10–20-day windows may affect estimates.