key: cord-322787-dbtc0bo3
authors: Runkle, Jennifer D.; Sugg, Margaret M.; Leeper, Ronald D.; Rao, Yuhan; Mathews, Jessica L.; Rennie, Jared J.
title: Short-term effects of weather parameters on COVID-19 morbidity in select US cities
date: 2020-06-09
journal: Sci Total Environ
DOI: 10.1016/j.scitotenv.2020.140093
sha: 
doc_id: 322787
cord_uid: dbtc0bo3

Abstract Little is known about the environmental conditions that drive the spatiotemporal patterns of SARS-CoV-2, and preliminary research suggests an association with weather parameters. However, the relationship with temperature and humidity is not yet apparent for COVID-19 cases in US cities first impacted. The objective of this study is to evaluate the association between COVID-19 cases and weather parameters in select US cities. A case-crossover design with a distributed lag nonlinear model was used to evaluate the contribution of ambient temperature and specific humidity on COVID-19 cases in select US cities. The case-crossover examines each COVID case as its own control at different time periods (before and after transmission occurred). We modeled the effect of temperature and humidity on COVID-19 transmission using a lag period of 7 days. A subset of 8 cities were evaluated for the relationship with weather parameters and 5 cities were evaluated in detail. Short-term exposure to humidity was positively associated with COVID-19 transmission in 4 cities. The associations were small with ¾ cities exhibiting higher COVID19 transmission with specific humidity that ranged from 6 to 9 g/kg. Our results suggest that weather should be considered in infectious disease modeling efforts and future work is needed over a longer time period and across different locations to clearly establish the weather-COVID19 relationship.

Experimental and observational studies demonstrate the influence of meteorological parameters on the seasonal transmission of influenza, human coronavirus (HCoV), and human respiratory syncytial virus (RSV), which are often characterized by distinct increases in incident cases and detection frequency in the winter months (Lowen and J o u r n a l P r e -p r o o f 3 in the US. As the US begins its public health response to COVID-19, the implementation of extensive public health interventions are needed at appropriate time scales to mitigate the public health and economic impacts of the COVID-19 pandemic. Research on the seasonality and influence of meteorological parameters on COVID-19, such as temperature and specific humidity, can be used to inform the timing of effective interventions to mitigate SARS-CoV-2/COVID-19 transmission at the local scale to save countless lives and resources.

The objective of this research is to examine the association between meteorological variables and COVID-19 in US cities. Unlike previous studies, we will use a highresolution spatiotemporal meteorological dataset to answer the following: Is there an association with meteorological parameters and COVID-19? If so, which meteorological parameters predict COVID 19 transmission? Is the association stronger after accounting for locally implemented social distancing measures? How does this relationship vary spatially across the US? By answering these questions, the knowledge gained on the contribution of environmental factors like temperature and humidity on transmission can be paired with other nonpharmaceutical interventions related to behavioral (e.g., wearing face mask, washing hands) factors that boost immunity or the timing of social distancing measures with seasonal spikes in influential environmental parameters to reduce transmission.

This retrospective case-crossover study examined the nonlinear and delayed association between environmental factors and COVID-19 J o u r n a l P r e -p r o o f 4 transmission. We selected the following US locations that exhibited high relative caseloads of COVID-19 in the early stages of the pandemic for their underlying populations: Seattle, Washington; New York, NY; Albany GA; New Orleans, LA.;

Bridgeport-Stamford-Norwalk, Conn.; Pittsfield, Mass; Detroit, MI; Chicago, IL. Figure 1 is a map of the 8 study locations. We assumed at least a median incubation period of 5.2 days (Lauer et al 2020) . Case counts were log-transformed, and time series were created when cities had more than 2 new daily cases. Because deidentified and anonymized data on case morbidity were obtained from a publicly accessible data portal this research did not involve participant consent and institutional review was not warranted.

Environmental Parameters. Meteorological data were derived from the European Center for Median Range Weather Forecast (ECMWF) atmospheric reanalysis dataset (ERA-5) (C3S, 2017). ERA-5 provides a suite of hourly weather parameters that may affect local COVID-19 transmissions at a 30-km spatial resolution. While not commonly used in environmental health studies, the advantage that ERA-5 data provides over individual weather station data is that spatial heterogeneity is more representative and the estimation of health effects of temperature and humidity can be derived in locations far J o u r n a l P r e -p r o o f 5 from weather stations or without any station. Previous research has shown that reanalysis data and weather station data show similar health risk estimates (Roye et al 2020). Daily average near-surface air temperature, specific humidity, and solar radiation were extracted from ERA-5 for each study location by a simple spatial average. Because relative humidity (RH) is highly correlated with temperature, we chose to instead include specific humidity as a predictor variable in the analysis.

Heat Mapping. Preliminary studies have suggested that the combination of humidity and air temperature could affect the transmission of local COVID-19 cases (e.g., Sajadi et al., 2020 , Lou et al., 2020 , Oliveriros et al., 2020 . We examined the association between local COVID-19 cases and air temperature and specific humidity using the density heatmap. To construct the density heatmap, the daily confirmed COVID-19 case reports were first separated based on their corresponding daily mean air temperature (every 1°C) and mean specific humidity (every 0.5 g/kg). All daily confirmed case counts were classified into the same air temperature and specific humidity conditions (e.g., 0 °C < T air < 1 °C and 1 g/kg < Q < 1.5 g/kg) and evaluated together as a density measurement. This explanatory analysis was intended to demonstrate the association of COVID-19 cases with the combined effect of air temperature and specific humidity.

The heatmap could identify the range of optimal meteorological conditions for local transmissions. Considering the incubation period of COVID-19, we applied the analysis J o u r n a l P r e -p r o o f 6 Nonlinear model (DLNM). This approach is more flexible than conditional logistic regression (Armstrong et al 2014) in that it allows for overdispersion. The application of the DLNM to the case-crossover design provides a means to assess the nonlinear and delayed effects, as well as the cumulative exposure-response between short-term daily average exposure to meteorological parameters and daily counts for COVID-19 cases.

We performed separate analyses for our primary health outcome --COVID-19 morbidity --and each meteorological parameter relative to the median and quartiles (i.e., 50th versus 75th). This approach is suitable for studying the effects of time-varying exposures (e.g., intermittent changes in meteorological) on a rare, acute condition (i.e., COVID-19 transmission) (Armstrong et al., 2019 , Malig et al., 2016 , Guo et al., 2011 .

We relied on the following equation:

where t is the day of the observation; Y t is the observed daily case counts on day t; α is the intercept; T t,l is a matrix obtained by applying the DLNM to temperature or humidity, β is a vector of coefficients for T t,l , and l is the lag days. Strata t is a categorical variable of day (30 day time period) used to control for trends, and λ is a vector of coefficients.

SD t is a binary variable that is "1" if day t was a social distancing order, and υ is the coefficient. Our model was adapted from similar work by Guo et al 2011 who also employed a case-crossover design and DLNM to investigate the effects of temperature on mortality (Guo et al 2011). Given that the incubation period between exposure and symptom occurrence is 2 to 14 days (Linton et al 2020), we used a maximum 14-day lag period to explore the potential delayed association of temperature and humidity in our model for approximating the pre-and post-infection period for each case. Sensitivity Analysis. A sensitivity analysis was conducted to select degrees of freedom for the lag polynomial (2-8 degrees of freedom) and the response polynomial (2-8 degrees of freedom) for New Orleans, LA (data not shown). In addition, we changed the maximum lag to 14 and 20 days, which gave similar results (data not shown). Prior research has examined a 0 day, 3 day to 5 day lags for COVID-19 transmission (Ma et al 2020 all the way to a lag period extending 7 to 14 days for meteorological parameters (Islam et al 2020). For our initial examination of meteorological parameters independently, we compared the best model fit using quasi-Akaike Information Criterion (q-AIC) to determine the optimal degrees of freedom and lag periods. q-AIC is a well-established technique for sensitivity analysis and was used to compare DLNM-only models and DLNM+Case-Crossover models to confirm the final model selection (Guo et al. 2011) . Models were also examined for the adjustment for trends, such as the day of the week. However, due to the short time series, differences were minimal or resulted in a poor fit (high q-AIC) and thus we selected the most J o u r n a l P r e -p r o o f 8 parsimonious model. The "dlnm" package was used to create the DLNM model (Gasparrini 2011) using R statistical software (R Core team 2020). We adopted the rare-disease assumption where our study hypothesis tested the association between weather exposure and a disease (i.e., COVID-19) characterized by low prevalence. Therefore, we assumed the odds ratio to approximate the relative risk. All relative risks (RR) were presented with corresponding 95% confidence intervals (95% CIs).

Attributable Burden of COVID-19 Transmission Due to Weather. In epidemiology, measures of potential impact are used to examine the expected impact of changing the distribution of one or more risk factors in a particular population (Kleinbaum et al 1982 , Szklo et al 2014 . For example, the attributable risk, also known as the etiologic fraction, is used to examine the proportion of all new cases in a given time frame that is attributable (or causally associated) to the exposure of interests (Szklo et al 2014) .

Because the evidence-base linking COVID-19 transmission and weather is new and evolving, it is too early to assume a causal association. Therefore, we relied on the excess fraction (EF) as an analogous, but alternative measure to the attributable risks in our analysis, to approximate the excess caseload due to exposure. To examine the attributable burden of transmission for COVID-19 due to weather we calculated the percent excess fraction for humidity and temperature for individual cities. We used the following equation: % EF = b x (RR i -1) / b x (RR -1.0 i ) + 1.0), where b is the point prevalence of COVID-19 for each city. Point prevalence was calculated as the number of cases over the study period divided by the total population in a specific city. We Table   1 ). The crude rate of COVID-19 per location was highest for New Orleans, LA (374 daily cases per 100,000 people), followed by New York City, NY (51 daily cases per 100,000 people), Albany, GA (42 daily cases per 100,000 people), and Bridgeport, CT (25 daily cases per 100,000 people). The lowest rates of COVID-19 cases were in Seattle, WA (4 daily cases per 100,000 people), and Pittsfield, MA (8 cases per 100,000 people). The highest crude death rates were observed in New York City, NY (6 daily deaths per 100,000 people), Albany GA (3 daily deaths per 100,000 people), New

Orleans, LA (2 daily deaths per 100,000 people) and Detroit, MI (2 daily deaths per 100,000 people).

The density heatmap ( Figure 2 ) presents a descriptive explanatory analysis of the combined association of temperature and specific humidity on COVID-19 cases for the selected cities. Based on the heatmap, COVID-19 cases were more common in low specific humidity (2 -6 g/kg) and low temperature (2 -11 °C) conditions. This association was consistent when we consider different incubation times (lag 0-14 days).

All Locations. Table 2 shows the goodness of fit (qAIC) values across model types for all locations and parameters, which is a common validation and sensitivity technique (e.g, Gasparrini et al. 2010 , Guo et al. 2011 . In general, the humidity was the strongest J o u r n a l P r e -p r o o f 10 predictor for COVID-19 cases, with a better model performance for humidity than temperature across all model types and study locations. Case crossover models performed higher in Seattle, WA, New York City, NY, Chicago, IL, and New Orleans, LA. The variation in the dose-response profile for humidity was negligible before and after adding temperature as a predictor into the model, indicating that humidity exhibited a robust association. Model performance was poor (indicated by high qAIC values) for Detroit, MI, Pittsfield, MA, and Bridgeport, CT. Results for these cities were insignificant and therefore not reported in the final results (Supplemental Figures 1,2) . Overall, the case-crossover + DLNM model outperformed the DLNM only model. However, select locations had better model fit for DLNM only (e.g., Albany, GA), although marginally better. Results were presented for the following cities: New Orleans, LA, Albany, GA, and Seattle, WA, and models were selected based on qAIC values. DLNM and casecrossover models were also constructed for these locations to analyze the effect of solar radiation (Watts/M 2 ) on COVID incidence rates.

New Orleans, LA. The relative risk for COVID-19 exhibited a U-shaped relationship with increases in cases at high and low humidity in New Orleans. With reference to the median humidity, relative risk peaked at minimum (5 g/kg, RR: 1.98, CI: 1.07-3.66) and maximum (16 g/kg, RR: 2.18, CI: 1.28-3.72) values. Similarly, temperature exhibited a U-shaped relationship with reference to the median and a significant relative risk at 16-17°C (RR: 1.17-1.23; CI: 1.03-2.24) and at the maximum observed temperatures (23°C; RR: 1.75, CI: 1.13-2.44). Solar values exhibited an inverted U-shaped relationship with a higher relative risk from 5200 -6300 (Watts/M 2 ) (Figure 3 ).

Albany, GA. Temperature and solar radiation were not significant predictors of COVID-19 cases. With reference to the median humidity, significant relative risk is observed from 6 to 9 g/kg (RR: 1.23-1.47, CI: 1.06-1.94). Due to a lower qAIC value and more robust results, unlike other cities, a DLNM-only model was applied to the humidity and the COVID-19 relationship for Albany, GA (Figure 4) . incidence that revealed a protective effect from 9 to 10°C (RR: 0.60-0.69, CI:0.39-0.95), whereas no relationship was observed between humidity and solar radiation and COVID-19 cases in NYC ( Figure 6) .

Seattle, WA. The temperature was significant from 3 to 5°C (RR: 1.59-1.95, CI: 1.22-2.65). However, the humidity was significant at the lowest values with an increased risk of transmission occurring at less than 3 g/kg (RR: 1.44, CI: 1.16-1.80). Lower risk of transmission was observed for the lowest values of solar radiation (i.e., less than 3200 Watts/M 2 , RR: 0.93, CI: 0.89-0.97) (Figure 7) .

The excess burden of new COVID-19 cases due to weather. Overall, the attributable burden of excess COVID-19 cases associated with exposure to humidity and temperature was low for each city (Table 3 ). The excess fraction was the highest for J o u r n a l P r e -p r o o f New Orleans, with 3.7 to 4.5% of new cases occurring within the humidity range of 5g/kg to 16g/kg and 6.8 to 9.1% occurring within the temperature range of 16(°C) to 23(°C).

In this study, we examined whether daily meteorological patterns in humidity, temperature, and solar radiation were associated with the transmission of COVID-19 in U.S. cities that emerged as early hot spots for infection. We applied the DLNM to a case-crossover design to assess the nonlinear and delayed effects of meteorological parameters on COVID-19 incident cases. To our knowledge, this study is the first to assess the effects of meteorological variables on COVID-19 morbidity using a robust distributed lag nonlinear model and case-crossover design. We observed a weak but statistically significant relationship between COVID and meteorological parameters for (Sajadi et al., 2020) .

Humidity was observed as the best predictor for the coronavirus outbreak followed by temperature and solar radiation. The majority of cities included in this study demonstrated a nonlinear dose-response relationship between a range of specific humidity conditions and sustained COVID-19 transmission. More specifically, 3 of the 4 cities were characterized by a significant relationship between COVID-19 transmission and humidity (e.g., Albany, GA, New Orleans, LA, and Chicago, IL). Humidity in the range of 6 to 9 g/Kg (analogous to an Absolute Humidity range of 7.56-11.37 g/m 3 ) was a significant predictor of COVID-19 cases and resulted in an up to two-fold increased Temperature and solar radiation did not exhibit a strong association with COVID19 incidence in our study locations. Our results for New York City, NY support and extend previous research on COVID19 and meteorological parameters in New York City that found a significant association with temperature using simple correlation coefficients (Bashir 2020). Bashir et al. observed a direct association with higher temperatures predicting higher COVID-19 cases (2020). Conversely, our research found a protective effect at higher temperatures and is corroborated by earlier studies (Qi et al., 2020, Unlike previous studies examining the influence of meteorological factors on COVID19 transmission, one strength of our study is the adjustment for social distancing measures (Sajadi et al 2020) . Most environmental health research includes either a variable for relative humidity (RH) and/or absolute humidity (AH). Specific humidity is more conservative and less susceptible to changes in pressure and temperature compared to AH. Further, in addition to the confounding influence of humidity and temperature, RH is typically not useful as a stand-alone humidity variable in environmental health or epidemiological research. Our results are only comparable to a few recent studies examining the association between COVID-19 and specific humidity (e.g., Ma et al., 2020 , Sajadi et al., 2020 .

Recent research has demonstrated the linkage between poor air quality and COVID-19 mortality (Wu et al. 2020) , and in our study, we did not adjust for background air quality measures as a potential confounding factor. While our modeling strategy did adjust for social distancing measures, our estimates do not account for underreporting of case counts (Lachmann 2020) , demographic data on cases, changes in testing capacity, or the date of onset of COVID19 symptoms. This study did not include information on the 

Newly discovered coronavirus as the primary cause of severe acute respiratory syndrome

Correcting under-reported COVID-19 case numbers. medRxiv

The frequency and seasonality of influenza and other respiratory viruses in T Tennessee: two influenza seasons of surveillance data

The incubation period of coronavirus disease 2019 (COVID-19) from publicly reported confirmed cases: estimation and application

YangY, et alIncubation period and other epidemiological characteristics of 2019 novel coronavirus infections with right truncation: a statistical analysis of publicly available case data

Defining the epidemiology of Covid-19-studies needed

Roles of humidity and temperature in shaping influenza seasonality

The role of absolute humidity on transmission rates of the COVID-19 outbreak

Effects of temperature variation and humidity on the death of COVID-19 in Wuhan

A time-stratified case-crossover study of ambient ozone exposure and emergency department visits for specific respiratory diagnoses in California

Determining the seasonality of respiratory syncytial virus in the United States: the impact of increased molecular testing

Seasonal variations of respiratory viruses and etiology of human rhinovirus infection in children

Seasonality of Respiratory Viral Infections. Annual Review of Virology

Coronavirus infections-more than just the common cold

Coronavirus as a possible cause of severe acute respiratory syndrome

COVID-19 transmission in Mainland China is associated with temperature and humidity: a timeseries analysis. Science of The Total Environment, 138778. R Core Team (2020). R: A language and environment for statistical computing. R Foundation for Statistical Computing

Temperature and latitude analysis to predict potential spread and seasonality for COVID-19

The impact of temperature and absolute humidity on the coronavirus disease 2019 (COVID-19) outbreak-evidence from China

Potential factors influencing repeated SARS outbreaks in China

Epidemiology: beyond the basics

Global influenza seasonality: reconciling patterns across temperate and tropical regions

Environmental predictors of seasonal influenza epidemics across temperate and tropical climates

High temperature and high humidity reduce the transmission of COVID-19

Exposure to air pollution and COVID-19 mortality in the United States

The relationship between air quality and respiratory pathogens among children in Suzhou City