Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Contents lists available at ScienceDirect Spatial and Spatio-temporal Epidemiology journal homepage: www.elsevier.com/locate/sste Original research Macro-level drivers of SARS-CoV-2 transmission: A data-driven analysis of factors contributing to epidemic growth during the first wave of outbreaks in the United States Matthew J. Watts 1 Institute of Environmental Science and Technology (ICTA), Autonomous University of Barcelona (UAB), Bellaterra, Spain ARTICLE INFO ABSTRACT Dataset link: https://doi.org/10.5281/zenodo.4 Background: Many questions remain unanswered about how SARS-CoV-2 transmission is influenced by aspects 994110 of the economy, environment, and health. A better understanding of how these factors interact can help us to Keywords: design early health prevention and control strategies, and develop better predictive models for public health SARS-CoV-2 risk management of SARS-CoV-2. This study examines the associations between COVID-19 epidemic growth COVID-19 and macro-level determinants of transmission such as demographic, socio-economic, climate and health factors, Epidemic-growth during the first wave of outbreaks in the United States. Doubling-time Methods: A spatial–temporal data-set was created from a variety of relevant data sources. A unique data-driven United States study design was implemented to assess the relationship between COVID-19 infection and death epidemic doubling times and explanatory variables using a Generalized Additive Model (GAM). Results: The main factors associated with infection doubling times are higher population density, home overcrowding, manufacturing, and recreation industries. Poverty was also an important predictor of faster epidemic growth perhaps because of factors associated with in-work poverty-related conditions, although poverty is also a predictor of poor population health which is likely driving infection and death reporting. Air pollution and diabetes were other important drivers of infection reporting. Warmer temperatures are associated with slower epidemic growth, which is most likely explained by human behaviors associated with warmer locations i.e. ventilating homes and workplaces, and socializing outdoors. The main factors associated with death doubling times were population density, poverty, older age, diabetes, and air pollution. Temperature was also slightly significant slowing death doubling times. Conclusions: Such findings help underpin current understanding of the disease epidemiology and also supports current policy and advice recommending ventilation of homes, work-spaces, and schools, along with social distancing and mask-wearing. Given the strong associations between doubling times and the stringency index, it is likely that those states that responded to the virus more quickly by implementing a range of measures such as school closing, workplace closing, restrictions on gatherings, close public transport, restrictions on internal movement, international travel controls, and public information campaigns, did have some success slowing the spread of the virus. 1. Introduction (US), the country that currently stands with the highest number of fatalities. The current COVID-19 pandemic is posing severe challenges to In the US and in other European countries like the United Kingdom, health systems, societies, and economies worldwide. At the time of governments and public health systems were initially caught off guard writing, the SARS-CoV-2 virus has already infected more than 175 by the sudden and rapid spread of the virus. This was partly due to a million people globally and caused 3.7 M deaths. In addition, the long- lack of political preparedness and a coherent strategy; lack of public term health impacts on those who have recovered from the SARS-CoV-2 health resources after years of cuts to public health budgets; or to infection are still unknown (Mahase, 2020). Approximately a sixth of the adoption of the wrong or no policy in terms of mask-wearing, the total deaths – more than 600,000 – occurred in the United States contact tracing, border controls, or lack of testing to detect community E-mail address: matthew.watts@uab.cat. 1 Equal contributor. https://doi.org/10.1016/j.sste.2022.100539 Received 5 July 2021; Received in revised form 9 February 2022; Accepted 13 September 2022 Available online 27 September 2022 1877-5845/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 transmission (Altman, 2020; Lee et al., 2021; Dyer, 2020; Ham, 2021; 1.1.1. Demographic/socio-economic factors Nowroozpoor et al., 2020; Lewis(a), 2020). Furthermore, the scientific Given the transmission pathways of SARS-CoV-2, as a priori, we community took some time to reach a general consensus regarding the would expect to see more infections occur in locations with higher modes of transmission of the virus; in particular, airborne dispersal population densities (e.g., metropolitan areas, cities) with high public was not considered a major pathway at the beginning of the pandemic, transport usage, overcrowded living spaces, and industries where busi- and this inhibited control and containment strategies (Lewis(b), 2020). ness takes place indoors—all of which naturally bring people into closer Even though thousands of papers have been written on COVID-19 contact, allowing airborne transmission to take place. To represent this related topics in the past year or so, many questions still remain in the models, variables were selected representing population density, unanswered, especially in terms of how SARS-CoV-2 transmission is public transport usage and household overcrowding. We would also influenced by aspects of the economy, environment, and health. A expect areas with a higher number of new residents arriving from better understanding of how these factors interact can help us to design abroad or out of state, to have had a larger number of outbreaks during the early stages of the pandemic through importation of the virus from timely health prevention and control strategies, and to develop better infected areas. To represent this in the model, a variable was built that predictive models for public health risk management of SARS-CoV-2 captured the annual rate of new residents arriving to a county from and other novel coronaviruses (Barouki et al., 2021). abroad or a different state. This study explores how some of the macro-level drivers of epidemic At the beginning of the pandemic, it took some time before a growth in the United States are associated with COVID-19 infection and consensus was reached about airborne transmission (Zhang et al., 2020; death doubling times during the first wave of the pandemic (in early Altman, 2020; Leung, 2021; Lewis(b), 2020; Tang et al., 2021), which 2020). The reason for selecting the United States is not only that it had major implications for early policy and practice, like improving is one of the hardest-hit countries, but also that it provides us with a ventilation in workspaces and adoption of behavioral changes like unique opportunity to study this phenomenon at a macro-scale, since it mask-wearing. We would expect the adoption of self-protective health encompasses a diverse range of climate types over a vast geographical behaviors (e.g., social distancing, work from home) that can reduce area, with a somewhat homogeneous political system, allowing us to the chance of contracting and spreading the virus (Papageorge et al., disentangle the effects of the environment from demographic and socio- 2021; Mongey et al., 2020; Fana et al., 2020) to be harder for low economic factors. Furthermore, the scientific institutions of the United skilled workers or those working in specific economic sectors (like States offer a vast quantity of high-quality data which allows us to manufacturing). Moreover, the inability to self protect may be accen- investigate our research question rigorously. By focusing on the first tuated for those who suffer from in-work poverty or precariousness wave of the pandemic, it is possible to better isolate the effects of since they may also be obliged to work, even when suffering with demographic, socio-economic and environmental factors, since it took symptoms, because of a lack of sick pay, fear of losing a day’s salary some time for the population to adopt self-protective behaviors like vac- and top-down pressures (Whitehead et al., 2021; Patel et al., 2020; cination, social distancing and mask-wearing; it also took some time for Finch and Hernández Finch, 2020). These factors are represented in state governments to apply containment measures, like school closures, the empirical models using variables that capture unemployment rates, limits on gathering and non-essential business closures (Lewis(a), 2020; employment levels in key economic sectors, education of the labor Papageorge et al., 2021; Margraf et al., 2020). force, and poverty. The empirical strategy for this study relies on county-level mor- bidity and mortality data as the main unit of analysis, which consists 1.1.2. Environmental factors Meteorological factors may affect SARS-CoV-2 transmission by al- of counts of individual infections and deaths, aggregated per county. tering human behavior; a basic assumption is people are likely to The use of data aggregated at the county level means we cannot stay indoors on days with very low or very high temperatures, and/or make individual level inferences and adjust for individual-level risk high rainfall. Furthermore, as a priori, we would expect people to bet- factors e.g. age, gender, and occupation. Nevertheless, this type of ter ventilate their homes/workspaces in places with warmer climates empirical investigation maintains high merit, as it enables a quick (e.g., leave their windows open, use wall and ceiling fans), which could exploration of geographic associations between the disease and the have an observable overall effect on disease transmission. To represent predictor variables, which can instigate further debate on this topic these factors in the models, variables were selected representing av- and may trigger more refined channels of research. The next subsection erage rainfall, temperate and relative humidity. Meteorological factors presents a short analytical framework, explaining how demographic, can also change the transmission potential and decay rate of the virus socio-economic, climate and health factors, as well as containment in air and on surfaces by altering its stability. (Schuit et al., 2020; Chan measures, are expected to influence the spread of the disease, and et al., 2011). Strong UV light can also inactivate SARS-CoV-2; however, describes the variables selected to measure such factors. this was not considered a significant predictor for COVID-19 infections and mortality since most transmission takes place indoors (de Oliveira 1.1. Analytical framework et al., 2021). 1.1.3. Health factors SARS-CoV-2 transmission takes place through 4 major pathways Initial reports from the ECDC (ECDC, 2020), the WHO (WHO, 2020) including exchange of saliva and mucus through human to human phys- and the CDC (CDC, 2020a) suggest that those most at risk of serious ical contact, indirect contact via fomites, or inhalation of large droplets morbidity and mortality are older people and people with underlying and fine aerosols (Leung, 2021; Lewis(b), 2020). Social distancing can health conditions such as diabetes, obesity, respiratory diseases, cancer, be one of the most effective measures to limit transmission, but this and cardiovascular diseases, poverty is a major risk factor of poor can be rendered ineffective in closed spaces with poor ventilation since population health and is correlated with such conditions (CDC, 2021; the virus can transmit through long-distance airborne dispersal (Zhang Drewnowski and Specter, 2004; Hawkins et al., 2012; Addo et al., 2012; et al., 2020; Nissen et al., 2020; Lewis(b), 2020). This study emphasizes Ward et al., 2004). As a priori, we would expect locations with higher demographic, socio-economic and climate factors can influence human proportions of residents with underlying health conditions to report to human contact and proximity, and can therefore modulate SARS- more infections and deaths. To represent this in the models, variables CoV-2 transmission (CDC, 2020b). Data on government containment were selected that capture the age structure of the population, poverty measures will also be analyzed since they can moderate SARS-CoV-2 rates, long term air pollution to proxy underlying pulmonary health transmission, and therefore morbidity and mortality reporting. conditions and the prevalence of diabetes. 2 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 1.1.4. Containment measures 2.1.3. Environmental data State governments implemented a wide range of measures to tackle Temperature (◦ C), precipitation (1/100′′ ), and relative humidity COVID-19 outbreaks such as school closures, workplace closures, re- data were sourced from the Global Surface Summary of the Day (GSOD) strictions on gatherings, close public transport, stay at home require- data provided by the US National Climatic Data Center (NCDC) (NCEI, ments, restrictions on internal movement, international travel controls 2021). This data-set provides daily GPS observations from all weather and public information campaigns, all of which could have had some stations situated in the US. To join county data with the GSOD weather success in suppressing the spread of the disease (Hale et al., 2021), such observations, centroids were created for each county using R’s SF containment measures would moderate the effects of the risk factors package and the United States Census Bureau’s county shape-files. The and drivers of disease transmission. To account for this in the models K-nearest neighbor join function in R’s SF package was used to create a ‘‘Stringency Index’’ variable was selected that reflects the level of a spatial join between the weather stations (GPS coordinates) and the a state government’s response to COVID-19 outbreaks, by quantifying county centroids. Mean climate values were created for a county-based how many measures were implemented and to what degree they were on data from a maximum of 10 nearest weather stations within a applied. The equations used to construct the ‘‘Stringency Index’’ will 100 km radius of each county centroid. be further explained in the next section. Compulsory stay at home Data on air quality was sourced from the United States Environ- orders (lock-downs) were not included in the ‘‘Stringency Index’’, since mental Protection Agency (EPA, 2020). Annual maximum reported Air they were used to determine the temporal cut off points of the study Quality Index (AQI) values were used, taken over a 20-year average. window, this is also explained in the section. This indicator was derived from data from EPA’s AQS (Air Quality System) database. The EPA establishes an AQI based on five major 2. Methods air pollutants including ground-level ozone (O3 ), particle pollution (also known as particulate matter, including PM2.5 and PM10), carbon All data were aggregated at the county level, apart from some data monoxide CO sulfur dioxide (SO2 ) and nitrogen dioxide (NO2 ). The U.S. on containment measures which are presented at state level. Below a AQI index runs from 0 to 500. The higher the AQI value, the greater the detailed description of the data sources. level of air pollution and the greater the health concern. The AQI is di- vided into 6 categories, each corresponding to a different level of health 2.1. Data collection and processing concern; generally, they represent 0 to 50—good; 51 to 100—moderate; 51 to 100—unhealthy for sensitive groups; 151 to 200—unhealthy; 201 2.1.1. Morbidity and mortality data to 300—very unhealthy; 301 and higher—hazardous. SARS-CoV-2 morbidity and mortality data were sourced from Johns Hopkins University’s Centre for Systems Science and Engineering’s 2.1.4. Containment measures (CSSE) GitHub repository (CSSE, 2020). In general, during the first Data on county-level stay-at-home orders (lock-down) were ex- wave of outbreaks in the US, testing was conducted only on those tracted from the CDC’s ‘‘U.S. State, Territorial, and County Stay-At- reporting more serious symptoms (see Additional file 1—COVID policy Home Orders’’ dataset (DATA.CDC.GOV, 2020). This dataset provides tracker). Almost all diagnostic testing for COVID-19 was done with the information on county-level executive orders, administrative orders, PCR-based methods, using nasopharyngeal or oropharyngeal specimens resolutions, and proclamations and can be used to determine the date (nose or throat swabs). of county-level stay-at-home orders (lock-down). Data on state-level containment measures were sourced from the 2.1.2. Demographic, socio-economic and population health data Data on county population, public transport usage, population age Oxford COVID-19 Government Response Tracker (OxCGRT) data set structure, health insurance coverage, immigration, disabilities, and (OxCGRT, 2020). The ‘‘Stringency Index’’ variable from this dataset was household overcrowding were sourced from the United States Census used to account for the application of state-level containment measures Bureau using 2015–2019 ACS 5-year estimates (USCB, 2020). To stan- in the final models. The composite time-series measure, ranging from dardize data across counties, all appropriate variables were converted 0 to 100 (100 = strictest) is based on 9 response indicators including to percentages/averages of the total county population. A household data on school closing, workplace closing, restrictions on gatherings, was considered overcrowded if the number of rooms was less than close public transport, stay at home requirements, restrictions on in- the number of inhabitants (above 1.01 people per room), this figure ternal movement, international travel controls, and public information included all rooms in a household (not just bedrooms). The disabilities campaigns. The indicator reflects the level of a state government’s measure captured various health conditions such as difficulty seeing response to COVID-19 outbreaks and quantifies how many measures or hearing, restricted movement, learning disabilities, cerebral palsy were implemented, and to what degree they were implemented. The or other developmental disabilities, or intellectual or mental health index cannot ascertain whether a government’s policy has been imple- disabilities (Taylor, 2014). mented effectively nor the effectiveness of an individual measure (Hale Population density per km2 was calculated using R’s SF package et al., 2021). To get an estimate of a government’s response leading up and the United States Census Bureau Cartographic county-level shape- to the first lock-down (compulsory stay at home order), the average files. Because the range of population density values was very wide, all stringency index value was calculated using a time window: from the values above 2500 km2 were capped to this value. This modification day the first 5 infections were reported the day before the first lock- was tested in the final models and did not affect the results and allowed down. Arkansas, Iowa, Nebraska, North Dakota, and South Dakota did for better interpretability of the results. not implement state-wide lock-downs. In these states, the average score County-level data on unemployment (%), median household in- was calculated from the day the first infections were reported to the come ($), and poverty % were sourced from the USDA Economic last lock-down date in our sample (2020-07-04) to make this value Research Service (ERS, 2020). The ‘‘Poverty %’’ indicator represents comparable to other states. the percentage of people/families whose earnings are less than the threshold designated by the Census Bureau’s set of money income 2.2. Study design thresholds. Data on diabetes prevalence were sourced from the CDC’s diabetes atlas (CDC, 2019). Economic dependence of a county was The spread of the disease (epidemic growth) is modeled by calcu- represented using the ERS county-level typology data-set (ERS, 2020), lating COVID-19 infection and death doubling times; these measures this breaks down a county into one of 6 major economic typologies: were then used as dependent variables to explore associations between farming, mining, manufacturing, federal/state government, recreation, epidemic growth, demographic, socio-economic, environmental, and and non-specialized. health factors. Doubling times capture exponential growth, in this 3 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 mortality doubling times, the doubling times may be less stable than that of the case data-set. Again, any county that did not meet this requirement was excluded from the study. To calculate the infection and death doubling times for each county, the following formulas were applied: 𝐸𝑒𝑛𝑑 − 𝐸𝑠𝑡𝑎𝑟𝑡 𝑟= × 100 𝐸𝑠𝑡𝑎𝑟𝑡 where: 𝑟 = growth rate; 𝐸𝑠𝑡𝑎𝑟𝑡 = Start of the event—when the 50 infections/20 deaths are detected 𝐸𝑒𝑛𝑑 = End of the event—cumulative infections/deaths per county at the lock-down date; Next, the doubling time is calculated using the following formula: 𝑙𝑛(2) 𝑇𝑑 = 𝑡 𝑟 𝑙𝑛(1 + 100 ) where: 𝑇𝑑 = doubling time in days 𝑡 = time in days (Estart to Eend) 𝑟 = growth rate Arkansas, Iowa, Nebraska, North Dakota, and South Dakota did not implement a state-wide lock-down (stay at home order), so an artificial date was set to calculate doubling times, mirroring the latest lock-down date in our sample (2020-07-04). 2.2.2. Time lags—disease progression Disease progression was also considered when calculating the dou- bling times, a time lag was applied to account for the discrepancy Fig. 1. Study design: capturing epidemic growth. between the date an event was reported (an infection or death) and the date the transmission event is likely to have took place. For data on confirmed COVID-19 infections, a lag of 21 days was set which considers a maximum 14-day incubation period based on instance, the number of days taken for cases and deaths to double. This findings from cohort studies by Lauer 2020 (Lauer et al., 2020), with measure has several advantages: first, it provides a way of standardizing an extra 7 days to account for any reporting delays. The implication differences in sampling effort between different locations and health here is that infection data for anything up to 21 days post lock-down authorities; second, because it provides us with a time determinant was used to calculate doubling times. measure to facilitate understanding of the spread of the virus. In other For the mortality data set, a lag of 42 days was set days which words, this metric not only has the advantage of accounting for popula- includes the maximum 14-day incubation period based on findings tion size but also incorporates a time dimension. Therefore, COVID-19 from cohort studies by Lauer et al. (2020) and a maximum of 21 days transmission is measured by calculating doubling times for infections from the first onset of symptoms to death based on findings from cohort and deaths, at the county level (Kröger and Schlickeiser, 2020; Lurie studies by Verity et al. (2020), plus an extra 7 days to account for et al., 2020; Muniz-Rodriguez et al., 2020; Pellis et al., 2021). any reporting delays. The implication here is that mortality data for anything up to 42 days post lock-down was used to calculate doubling 2.2.1. Calculating infection and death doubling times times. Doubling times were calculated by capturing a window of infection Data on environmental factors were also joined to the lagged county opportunity, which started on the date a minimum number of infec- doubling time variables, meaning that they were linked to the date tions/deaths were detected in a county, to the date of the first major when a disease event is likely to have took place, rather than when state or county level intervention was implemented i.e. compulsory reported. stay-at-home orders, otherwise known as a lock-down (see Fig. 1). A time lag was also applied to the doubling times in order to account for 2.2.3. General additive regression model to assess the impact of independent the time infection or mortality events took place, since there is a lag variables on doubling times at the county level between the date an event is reported (a case or death) and the date One of the main issues with the data-set is that it did not meet the transmission event took place. Therefore all infection and mortality some basic assumptions for statistical inference, that is the data are not data was lagged by a maximum incubation period (onset of symptoms) independent and identically distributed random variables (iid). More or a maximum time from final infection to death, these are further specifically, observations cannot be considered independent because of described below. spillover effects from neighboring counties, therefore an appropriate For the calculation of the infection doubling times, the count was statistical design was needed to control for a lack of independence started when the county reached a minimum of 50 confirmed infec- between neighboring counties. A Generalized Additive Model (GAM) tions, over a minimum 7 day reporting period. Any county that did not using R’s Mgcv statistical package because of its versatility and ability meet this requirement was excluded from the study. to fit complex models that would converge even with low numbers Since the mortality data-set contained fewer observations than the of observations and could capture potential complex non-linear rela- infections data set, the count was set when the county reported a min- tionships. One of the advantages of GAMs is that we do not need to imum of 20 deaths over a minimum 7 day reporting period. Although determine the functional form of the relationship beforehand. In gen- these values yielded enough observations to carry out the study on eral, such models transform the mean response to an additive form so 4 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Table 1 Infections data-set - summary statistics. N = 640 (number of counties selected for the study which met the inclusion criteria laid out in study design section). Statistic Min Max Mean St. Dev. Confirmed infections 54 35,571 950.9 2953.7 Incidence per 100,000 15.8 5611.3 277.5 434.1 Study period days 7 40 18.7 7.5 Growth rate (over period) 8 62,305 1503.5 5040.1 Infection doubling times 2.5 72.0 11.6 8.4 County population 5861 10,081,570 373,339.6 657,728.9 Population density per km2 2.0 2500.0 245.3 415.8 Stringency index 11.0 64.1 34.1 10.5 Public transport usage (pop %) 0.0 32.0 1.3 3.1 Median household income ($) 26,348 151,806 65,642.1 18,915.3 Unemployment % 1.8 12.0 3.8 1.2 Population % 65+ 7.9 41.1 16.5 4.2 Poverty % 2.7 36.6 13.1 5.9 Health insurance coverage % 62.2 99.9 98.0 3.0 Annual new arrivals/by pop (%) 0.0 3.1 0.5 0.4 Population % with disabilities 5.0 25.7 13.1 3.3 Population % with diabetes 2.2 23.1 10.1 2.9 Household overcrowding % 0.1 7.0 1.3 1.0 Population % with degree or higher 3.6 29.0 12.6 4.6 Temperature (◦ C) −3.4 24.7 12.7 6.4 Precipitation (1/100′′ ) 0.0 9.2 2.1 1.6 Relative humidity 30.7 86.1 67.0 8.6 Air quality index (AQI) 32.5 347.4 130.4 31.3 that additive components are smooth functions (e.g., splines) of the co- 3.0.1. Descriptive statistics variates, in which functions themselves are expressed as basis-function Two sources of information were analyzed, data on confirmed in- expansions. The spatial auto-correlation in the GAM was approximated fection and deaths. Tables 1 and 2 provide summary statistics for our by a Markov random field (MRF) smoother, which represents the spatial final data-sets. dependence structure in the data. R’s Spdep package was used to To calculate doubling times, counties were only selected that had create a queen neighbors list (adjacency matrix) based on counties reported at least 50 infections or 20 deaths over a minimum 7-day with contiguous boundaries i.e., those sharing one or more boundary period before the first lock-down. Both sources of information were points. The local Markov property assumes that a county is condition- chosen as they allow us to explore and compare different features and ally independent of all other counties unless they share a boundary. characteristics of the epidemic. Figs. 1 and 2 map the geographical This feature allows us to model the correlation between geographical distribution for infection and death doubling times in counties that met neighbors and smooth over contiguous spatial areas, summarizing the our inclusion criteria (colored from red to yellow). Major cities with trend of the response variable as a function of the predictors (Wood, populations >250,000 people are highlighted on each map. The coun- 2017). Models were fit using a gamma distribution, after inspecting ties first affected by SARS-CoV-2 during the first wave of the epidemic tended to be located around major cities and metropolitan areas on the data, it was concluded that the gamma distribution worked well the east coast, mid west, and south of the United States, with high with the shape of our response variable, which was positively skewed population density and presumably higher numbers of international (i.e., non-normal, with a long tail on the right). The gamma distribution and domestic travelers. is a two-parameter distribution, where the parameters are traditionally known as shape and rate. Its density function is: 3.0.2. Regression results 1 It was not possible to explore the individual impact of all the 𝑓 (𝑥) = 𝑥𝛼−1 𝑒−𝑥∕𝛽 , 𝛽 𝛼 𝛤 (𝛼) variables in our data-set because of collinearity issues (see Additional where 𝛼 is the shape parameter and 𝛽 − 1 is the rate parameter file 2). Public transport was positively correlated with population den- (alternatively, 𝛽 is known as the scale parameter). sity so therefore removed from the analysis. Median income was also The empirical model can then be written as: removed from the analysis because it was positively correlated with education, and negatively correlated with poverty, disabilities and 𝐸(𝑌 ) = 𝑓1 (X𝑖 ) + 𝑓𝑚 (County𝑖 ) diabetes. Tables 3 and 4 show the results of the statistical analysis for both where the 𝑓 (.) stands for smooth functions; 𝐸(𝑌 )𝑖 is equal to infection data sets and summarize the relevant statistics (AIC, Deviance, Adjusted or death doubling time in county 𝑖, which we assume to be gamma- R squared (𝑅2 and so on) to compare the different specifications. distributed; 𝑋𝑖 - is a vector of demographic, socio-economic, climate, Both statistical models were built in a step-wise fashion using the health and containment variables (as described in the previous section). lowest Akaike Information Criterion (AIC) and 𝑅2 to help us assess 𝑐𝑜𝑢𝑛𝑡𝑦𝑖 represents neighborhood structure of the county. the different specifications. Variables were included in each specifi- Analysis of model diagnostic tests did not reveal any major issues, in cation according to their category i.e., spatial, socio-economic, and general residuals appeared to be randomly distributed. For robustness, environmental. All variables were included in the final specification models were also fit using the Gaussian and Tweedie distributions, and to ascertain the contribution of each driver or risk factor, all else also fit using a non-additive-GLM (see Additional file 2). equal. Note that, as we are not estimating a standard regression model, the figures reported should not be read as coefficients, but degrees 3. Results of freedom of the smooth terms. Given that we cannot interpret the coefficients to infer the sign and magnitude of the relationship, we To carry out the empirical analysis, a unique spatial data-set was visualize it by plot. Figs. 3–11 plot the partial effects—the relationship compiled that captured potential drivers of human-to-human SARS- between a change in each of the covariates and a change in the fitted CoV-2 transmission and risk factors of serious infections and mortality values in the full model. Standard errors on the plots show the 95% due to COVID-19 in US counties. confidence interval for the mean shape of the effect. 5 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Table 2 Mortality dataset—summary statistics. N = 263 (number of counties selected for the study which met the inclusion criteria laid out in study design section). Statistic Min Max Mean St. Dev. Confirmed infections 22 4902 233.4 541.3 Incidence per 100,000 1.4 284.6 39.9 45.4 Study period days 7 56 24.4 10.2 Growth rate (over period) 9.5 24,410.0 906.5 2346.6 Infection doubling time 5.7 80.0 18.5 12.5 County population 8737 10,081,570 677,179.5 929,606.3 Population density per km2 2.9 2500.0 442.2 570.3 Stringency index 11.0 64.1 32.4 11.6 Public transport usage (pop %) 0.0 32.0 2.4 4.5 Median household income ($) 36,894 151,806 71,804.3 20,582.3 Unemployment % 1.8 9.6 3.7 1.0 Population % 65+ 9.5 41.1 16.1 4.0 Poverty % 2.7 30.7 12.0 5.1 Health insurance coverage % 89.9 99.6 98.7 1.2 Annual new arrivals/by pop (%) 0.01 2.2 0.6 0.4 Population % with disabilities 5.8 23.5 12.1 2.9 Population % with diabetes 5.2 22.3 9.3 2.3 Household overcrowding % 0 6 1.4 1.0 Population % with degree or higher 3.9 27.6 14.2 4.4 Temperature (◦ C) 0.2 25.3 11.6 5.9 Precipitation (1/100′′ ) 0.0 7.4 2.0 1.4 Relative humidity 25.5 81.6 64.9 8.8 Air quality index (AQI) 41.5 347.4 143.8 31.5 Fig. 2. COVID-19 infection doubling times in US Counties and major cities with over 250,000 people. Counties highlighted in white not selected for study (Date range: 2020-03-05 to 2020-04-29, data source: Johns Hopkins University). 3.0.3. Infection data model times although the confidence intervals are fairly large so the sample Table 3 and Figs. 3–7 show the results of the model fit using infec- does not provide a precise representation of the population mean. tion data. The ‘‘Spatial’’ model was fit first to estimate the contribution The stringency index variable, which captures the number of contain- of the spatial lag component against the other specifications. A high ment measures adopted by states, and the degree to which they were proportion of the variance is explained just by controlling for spatial implemented, is also statistically significant (p < 0.05), and has a correlation between counties (𝑅2 0.35). The ‘‘Full model’’ has the best positive relationship with the infection doubling times, suggesting that fit in terms of the AIC and adjusted 𝑅2 , followed by the socio-economic measures had some success in suppressing the virus. Human population model, and finally the environmental model. The adjusted 𝑅2 in the density per km2 is highly significant (p < 0.001), higher densities are as- final model is 0.56, indicating that 56% of the variance in our model sociated with faster infection doubling times, although the relationship is explained by the explanatory variables. is not linear and flattens out at higher population densities. Although As for the contribution of individual variables on infection doubling the slope is gentle, ‘‘Poverty %’’ is a highly significant (p < 0.01) times, counties with manufacturing and recreation as their predomi- predictor of infection doubling times, the relationship is negative which nant economic activity were associated with faster infection doubling means doubling times are faster with higher levels of poverty (in other 6 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Fig. 3. COVID-19 death doubling times in US Counties and major cities with over 250,000 people. Counties highlighted in white not selected for study (Date range: 2020-03-10 to 2020-05-19, data source: Johns Hopkins University). Fig. 4. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 infections. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower infection doubling times. words, the infection spreads faster). On the contrary, the variable ‘‘Pop people per room) is highly significant (<0.01) and is associated with % with disabilities’’ (p < 0.01) has a positive relationship with infection faster infection doubling times. Temperature is also a good predictor of doubling times, meaning it is a predictor of slower doubling times. The infection doubling times; higher temperatures appear to slow infection prevalence of diabetes (Pop % with diabetes) in a county, an indicator doubling times. (p < 0.01), although this relationship breaks down at that not only represents the disease itself, but also a range of other con- lower temperatures given there are few observations, the confidence ditions such as obesity, poor diet, lack of exercise was also a significant intervals are much larger meaning the results are less accurate. ‘‘Max (p < 0.01) predictor of faster infection doubling times. ‘‘Population % AQI’’, which represents the maximum air quality index values averaged home overcrowding’’, which represents the percentage of households over 20 years, is also highly significant and is associated with faster in a county where there is less than one room per inhabitant (>1.01 infection doubling times in locations with poor air quality (p < 0.01). 7 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Fig. 5. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 infections. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower infection doubling times. Fig. 6. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 infections. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower case doubling times. 3.0.4. Mortality model (<0.001) is also an important predictor: generally, higher population Table 4 and Figs. 8–11 show the results of our model fit using density is associated with faster death doubling times, however, this mortality data. A high proportion of the variance is explained just trend reverses at around 1400 inhabitants per km2 and levels off. by controlling for spatial correlation between counties (𝑅2 0.22). The ‘‘Population % 65+’’ (<0.001) is highly significant; higher values are ‘‘Full model’’ has the best fit in terms of the AIC and adjusted 𝑅2 associated with faster death doubling times. Again, as with the infection 0.48, followed by the socio-economic model (0.44) and the environ- data analysis, ‘‘Poverty %’’ is also a highly significant predictor of death mental model (0.31). The ‘‘Stringency index’’ indicator is statistically doubling times (<0.001), that is higher levels of poverty are associated significant (p < 0.05) and is associated with slower death doubling with mortality. ‘‘Pop % with disabilities’’ (<0.01) is also highly sig- times; that is more stringent containment measures are associated with nificant; as with the infection data model, this predictor is associated slower COVID-19 death doubling times. ‘‘Population density per km2 ’’ with slower death doubling times. The prevalence of diabetes (Pop % 8 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Fig. 7. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 infections. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower infection doubling times. Bottom left: Categorical variables—dashed horizontal lines on the categorical variables represent the confidence intervals and solid red lines represents the mean value. Fig. 8. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 deaths. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower death doubling times. with diabetes) in a county is also a significant predictor (p < 0.05) of 4. Discussion faster death doubling times, as is the air quality index (‘‘Max AQI’’), which is highly statistically significant (p < 0.01). Temperature and In this study, I examined which demographic, socio-economic, precipitation are slightly significant (p < 0.1). and appear to slow down and environmental factors are associated with SARS-CoV-2 epidemic death doubling times at higher values. growth. To explain biases in reporting, I included health risk factors 9 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Fig. 9. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 deaths. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower death doubling times. Fig. 10. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 deaths. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent slower death doubling times. that can contribute to serious SARS CoV-2 infections and deaths. We variation in viral traits/strains. Furthermore, no vaccines were yet in would expect infection reporting to be a function of all these factors circulation. since testing policy during this phase of the epidemic was aimed at those with symptoms (see Additional file 1—COVID-19 policy tracker). 4.1. Containment measures to reduce disease spread We can also assume that, during this wave of the epidemic in the US, only one strain of SARS-CoV-2 (although always evolving) During the first wave of the epidemic in the US, governments, and was in circulation and therefore the variation in infection and death public health systems were initially caught off guard by the rapid rates across space can be attributed to external factors i.e., testing spread of the virus. Some of the states did apply more rigorous control differences, aspects of the population and environment, rather than measures than others, attempting to suppress the spread of the virus 10 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Fig. 11. Generalized additive model (GAM) plots showing the partial effects of the explanatory variables on the doubling times of COVID-19 deaths. The tick marks on the 𝑥-axis are observed data points. The 𝑦-axis represents the partial effect of each variable. The dots represent partial residuals. The shaded areas indicate the 95% confidence intervals. Higher values on the 𝑦-axis represent longer death doubling times. Bottom left: Categorical variables—dashed horizontal lines on the categorical variables represent the confidence intervals and solid red lines represents the mean value. early on e.g., by restricting gatherings, closure of public spaces, creating doubling time. However, these variables are only significant in the public awareness campaigns and contact tracing (see Additional file infection model and not the mortality model. One possible explanation 1). Stringency index scores in both our models are associated with is that they represent transmission among younger people of working slower doubling times and can be interpreted as, the more stringent age, students, and younger families, who are less likely to die from the measures applied by state governments early on, the more success COVID-19. they had in suppressing the virus. In terms of age population structure, having a higher proportion over 65-year-old’s was also a significant predictor of faster death dou- 4.2. Demographic, socio-economic, factors bling times, concurrent with the literature and common understanding about the disease; age is one of the major risk factors. Major outbreaks Results show that human population density is one of the strongest have occurred in care homes (Leclerc et al., 2020) suggesting that some predictors of infection and death doubling times, the relationship is of the counties most affected by COVID-19 in the first wave of the negatively linear to a point, where higher population densities are epidemic was in locations with a higher proportion of retirees and care associated with faster doubling times, but this trend tends to level off at homes. population densities of above 400 people per km2 , and reverses slightly In terms of other socio-economic factors affecting the disease, for death doubling times at densities above 1000 people per km2 . Per- poverty was also a significant predictor of faster doubling times in haps because of features relating to the built environment i.e. building both infection and mortality models. As mentioned in the conceptual types, age structures, demographic or socio-economic conditions associ- framework, this can be explained since those who suffer from in-work ated with wealthier city dwellers. However, in general, the relationship poverty are likely to be doing jobs where it is difficult to work from between population density and COVID-19 transmission is logical given home or adopt self-protective health behaviors such as social distancing the virus mainly transmits when humans are in close proximity to (Papageorge et al., 2021). Furthermore, even when suffering from one another. Human population density also captures other important symptoms, many low skilled workers and precarious workers may have features of the built environment; for example, locations with high been obliged to work because of a lack of sick pay, fear of losing a population density are cities or metropolitan areas, usually with high day’s salary and pressures from bosses (Whitehead et al., 2021; Patel public transport usage, more recreational businesses like restaurants et al., 2020; Finch and Hernández Finch, 2020). Poverty is also a risk and bars, and indoor work-spaces like offices. All of which naturally factor of poor population health and is correlated with a multitude of bring people into closer contact and encourages airborne transmission underlying health conditions believed to lead to adverse outcomes for of SARS-CoV-2. those suffering from COVID-19 (CDC, 2021). This is further supported Results also show that counties that rely on manufacturing or by the results of our final models; higher diabetes prevalence is also recreation as their main economic activity, also tend to have faster associated with faster infection and death doubling times. Again, those infection doubling times. Again, this is likely due to aspects of the work suffering from diabetes are likely to suffer from comorbidities such environment like the lack of proper physical distancing and ventila- as obesity and heart problems (Jelinek et al., 2017). These results tion. These findings are corroborated by studies (Leclerc et al., 2020; are also concurrent with work conducted by Williamson et al. (2020), Middleton et al., 2020) that report many SARS-CoV-2 clusters were who found that greater age, deprivation, diabetes, severe asthma, and linked to a variety of indoor settings including households, hospitals, various other medical conditions were at higher risk of death due to elderly care homes, and food processing plants (classed as factories). COVID-19 infection. For both data-sets ‘‘Pop % with disabilities’’ tended This concept is also further supported by our indicator representing to be correlated with slower doubling times. Although this group may household overcrowding, which is another strong predictor of infection be vulnerable to COVID-19 infections, they can often suffer from social 11 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Table 3 COVID-19 Infection model—Generalized additive regression model for assessing associations between the demographic, socio- economic, climate and health factors on county level infection doubling times. Note that as we are not estimating a standard regression model, the figures reported should not be read as coefficients, but degrees of freedom of the smooth terms. Given that we cannot interpret the coefficients to infer the sign and magnitude of the relationship, we visualize it by plot. Spatial Socio-econ Envir Full Intercept 2.34*** 2.45*** 2.34*** 2.48*** (0.02) (0.12) (0.02) (0.12) Industry: Manufacturing −0.27* −0.32* (0.13) (0.13) Industry: Government −0.02 −0.04 (0.13) (0.12) Industry: Recreational −0.22 −0.23. (0.14) (0.13) Industry: Non-specialized −0.13 −0.17 (0.12) (0.12) Industry: Agricultural 0.09 0.04 (0.24) (0.23) Stringency index 1.00* 1.00* (1.00) (1.00) Pop density per km2 6.30*** 5.81*** (7.56) (7.01) Unemployment % 1.61 1.31 (1.82) (1.50) Population % 65+ 1.00* 1.00 (1.00) (1.00) Poverty % 1.00* 1.00** (1.00) (1.00) Health insurance coverage % 1.62 1.57 (1.83) (1.80) New arrivals into county population % 1.40 1.55 (1.63) (1.78) Population % with disabilities 1.00** 1.00** (1.00) (1.00) Population % with diabetes 1.00* 1.00** (1.00) (1.00) Population % living in overcrowded homes 1.00* 1.00** (1.00) (1.00) Temperature ◦ C 3.54** 3.71** (4.10) (4.25) Precipitation 2.21 1.00 (2.75) (1.00) Relative humidity 1.00 3.54. (1.00) (4.15) Air quality index (AQI) 1.00*** 1.67** (1.00) (1.87) County 137.13*** 135.30*** 132.98*** 135.55*** (169.51) (167.57) (165.05) (167.11) AIC 3772.29 3499.61 3705.13 3459.43 BIC 4393.02 4210.00 4341.94 4212.15 Log Likelihood −1747.01 −1590.58 −1709.83 −1561.00 Deviance 91.15 56.41 81.36 51.50 Deviance explained 0.61 0.76 0.65 0.78 Dispersion 0.21 0.13 0.19 0.11 R2 0.35 0.54 0.41 0.56 GCV score 0.23 0.16 0.21 0.15 Num. obs. 640 640 640 640 Num. smooth terms 1 11 5 15 *𝑝 < 0.05. **𝑝 < 0.01. ***𝑝 < 0.001. . 𝑝 < 0.1. isolation which provides some explanation. Furthermore, these groups observational studies (Cole et al., 2020; Travaglio et al., 2021). Some are more likely to self-isolate (Macdonald et al., 2018; Emerson et al., authors propose that air pollution increases infectivity, as SARS-CoV-2 2021) to avoid infections. binds with airborne particulate matter (Nor et al., 2021; Lolli et al., 2020; Solimini et al., 2021) allowing the disease to persist for longer 4.2.1. Environmental factors in the air. Although this should not be ruled out, as mentioned, air Although a broad measure, the air quality index (‘‘Max AQI’’) quality indicators also tend to proxy poor pulmonary health, which provides us with a way to proxy for counties with poor air quality may increase death and infection reporting, that is people with lung and population-level pulmonary health conditions, caused by long- problems induced by air pollution are more likely to have symptomatic term exposure to harmful pollutants such as PM 2.5, PM10, NO2 , infections. It is well documented that long term exposure to certain SO2 and NO𝑥 . This indicator is strongly correlated with COVID-19 pollutants has knock-on effects for people suffering from pulmonary infections and death doubling times, where higher AQI tends to speed viral infections (Chauhan and Johnston, 2003; Croft et al., 2018; Grigg, up infection and death reporting. This result is consistent with other 2018; Kirwa et al., 2021). For example, a study by Becker and Soukup 12 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Table 4 COVID-19 mortality model—Generalized additive regression model for assessing associations between the demographic, socio- economic, climate and population health factors on county level death doubling times. Note that as we are not estimating a standard regression model, the figures reported should not be read as coefficients, but degrees of freedom of the smooth terms. Given that we cannot interpret the coefficients to infer the sign and magnitude of the relationship, we visualize it by plot. Spatial Socio-econ Envir Full Intercept 2.83*** 2.65*** 2.81*** 2.78*** (0.03) (0.25) (0.03) (0.25) Industry: Manufacturing 0.14 −0.02 (0.28) (0.28) Industry: Government 0.18 0.03 (0.26) (0.26) Industry: Recreational 0.18 0.13 (0.28) (0.28) Industry: Non-specialized 0.13 −0.02 (0.26) (0.26) Stringency index 1.00* 1.00* (1.00) (1.00) Population density per km2 7.00*** 5.84*** (8.35) (7.05) Unemployment % 1.00 1.00 (1.00) (1.00) Population % 65+ 1.00* 1.00*** (1.00) (1.00) Poverty % 1.00* 1.00** (1.00) (1.00) Health insurance coverage % 1.00 1.00 (1.00) (1.00) New arrivals into county population % 1.00 1.00 (1.00) (1.00) Population % with disabilities 1.00. 1.00** (1.00) (1.00) Population % with diabetes 1.00 1.00* (1.00) (1.00) Population % living in overcrowded homes 1.00 1.00 (1.00) (1.00) Temperature ◦ C 2.51 2.71. (3.07) (3.28) Precipitation 3.03. 4.61. (3.66) (4.91) Relative humidity 1.00 1.00 (1.00) (1.00) Air quality index (AQI) 1.04** 1.00* (1.07) (1.00) County 56.42*** 56.15*** 59.31* 57.32*** (67.05) (66.65) (69.50) (67.52) AIC 1854.52 1742.72 1824.06 1724.01 BIC 2063.22 2021.91 2070.15 2036.51 Log Likelihood −868.84 −793.20 −843.14 −774.52 Deviance 46.09 26.26 38.10 22.83 Deviance explained 0.50 0.72 0.59 0.75 Dispersion 0.27 0.15 0.22 0.13 R2 0.23 0.44 0.31 0.48 GCV score 0.29 0.20 0.26 0.19 Num. obs. 263 263 263 263 Num. smooth terms 1 11 5 15 *𝑝 < 0.05. **𝑝 < 0.01. ***𝑝 < 0.001. . 𝑝 < 0.1. (1999) found that regulated inflammatory responses to viral infections place over short distances in closed spaces), the influence of weather are altered by exposure to PM10, potentially increasing the spread on human behavior is likely one of the major drivers of SARS-CoV- of infection and therefore increasing viral pneumonia-related hospital 2 transmission. Weather is widely considered to influence people’s admissions. behavior (de Freitas, 2014) but research on this topic is surprisingly In general, infection and death reporting doubling times were scant. According to (Daniel, 2018), people living in warmer/hotter negatively associated with temperature. There is increasing evidence locations, or during periods of warmer weather are more likely to that COVID-19 is a seasonal disease (Kaplin et al., 2021; Choi et al., employ a range of adaptive behaviors in response to warm and hot 2021), especially in temperate climates where there are distinct sea- sonal phases i.e. summer and winter, with distinct temperature ranges, conditions i.e., keeping windows and doors open, use of wall and distinct levels of ultraviolet radiation (UV) and seasonal differences in ceiling fans, air conditioning, which in turn may initiate a range of self- air moisture carrying capacity. Although, it is important not to rule out protective behaviors against SARS-CoV-2 transmission. Furthermore, physical factors influencing transmission, especially for long-distance warmer weather is also associated with recreational time spent out- transmission, given the nature of the disease (transmission mainly takes doors (Bélanger et al., 2009) where SARS-CoV-2 transmission risk is 13 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 likely to be lower. Although temperature also exhibited similar patterns et al., 2021; Allen et al., 2017; Wardeh et al., 2021) will continue for the death data model, it was only weakly statistically significant. to present risks globally until better controls and regulations can be implemented (Dobson et al., 2020). If new coronaviruses emerge, with 4.3. Limitations similar modes of transmission, we should hope that governments can quickly apply top-down measures to suppress the virus before more Some of the limitations of the study are as follows. Since the study sophisticated measures can be implemented i.e. rapid community test- is limited to using aggregated data at the county level, we cannot make ing to isolate the infected. I hope this work will contribute to the inferences about individual-level associations and cannot not adjust scholarly debate and can shed light on some of the environmental and for individual-level risk factors e.g. age, gender, race, and occupa- socio-economic factors driving SAR-COV-2 transmission. tion. However, that would be outside the scope of this study, since I was interested in understanding macro socio-economic and ecological 5.1. Abbreviations drivers. Additionally, we cannot draw causal inference as the applied methodology only reveals adjusted correlations. Therefore, results were GDP: Gross Domestic Product; US: United States of America. carefully evaluated from individual-level and clinical-based studies to draw conclusions. The use of further explanatory variables would have Declaration of competing interest surely improved the study i.e. on homelessness, availability of Intensive Care Units (ICU), quality of medical facilities, and ratio of medical The authors declare that they have no known competing finan- staff per person, but these data were not available. It is also important cial interests or personal relationships that could have appeared to to note that given the unprecedented nature and scale of COVID-19 influence the work reported in this paper. outbreaks, data quality issues arise owing to the under-reporting of infections i.e., through under-diagnosis, lack of diagnostic tests and Data and code availability a lack of resources/time to carry out and implement mass testing. If data collection methods remained constant across counties over the An R project containing all the data and code that support the time frame of this study, the calculation of doubling times can be findings of this study is available in .Rdata format from https://doi. a reliable measure. However, doubling times can be inflated by im- org/10.5281/zenodo.4994110. proving testing procedures i.e., better detection and reporting through the availability of better diagnostic tests, better sampling techniques, Acknowledgments resource allocation, and increased awareness of the disease. Thanks to Panagiota Kotsila, Graham Mortyn, Victor Sarto Monteys, 5. Conclusions Cesira Urzi Brancati and Matthew D. Peavy who provided advice and feedback on the research. Panagiota Kotsila, Graham Mortyn and Victor This paper investigated drivers of epidemic growth during the first Sarto Monteys also supervised the project and are responsible for wave of outbreaks in US counties, by assessing the association between formulating ideas for the umbrella project Impacts of Climate Change COVID-19 epidemic doubling times with demographic, socio-economic, (CC) on Human Health (HH) at ICTA-UAB: Integrating socio-economic environmental and health factors, with state government containment and policy studies with natural science studies to enhance consilience measures. Results suggest that the main drivers of new infections are of climate policy science. higher population density, home overcrowding, manufacturing and This research was funded by ICTA’s Maria de Maeztu Unit of Ex- recreation industries and poverty. By contrast, warmer temperatures cellence, awarded by the Spanish Ministry of Economy and Competi- slowed epidemic growth which was likely to be the result of human tiveness. The award is the highest institutional recognition of scientific behavioral responses to temperature. The main factors associated with research in Spain. Thanks also to Patrizia Ziveri and Pedro Manuel death doubling times were age, poverty, air pollution and diabetes Gonzalez Hernandez for supporting the project. prevalence. Such findings help underpin current understanding of the disease epidemiology and also support current policy and advice rec- Appendix A. Supplementary data ommending ventilation of homes, work-spaces and schools, along with social distancing and mask-wearing. Supplementary material related to this article can be found online The results also suggest that states which adopted more stringent at https://doi.org/10.1016/j.sste.2022.100539. containment measures early on, did have some success at slowing the spread of the virus. There are numerous reports that there were References huge failures at local level i.e. in care homes and business owners Addo, J., Ayerbe, L., Mohan, K., Crichton, S., Sheldenkar, A., Chen, R., Wolfe, C., failing to protect residents and staff, by acting too slow or failing to McKevitt, C., 2012. Socioeconomic status and stroke. Stroke 43 (4), 1186–1191. implement control measures such as mask wearing and creating better Allen, T., Murray, K., Zambrana-Torrelio, C., Morse, S., Rondinini, C., Di Marco, M., ventilation in closed spaces (O’Neill, 2020; Chapman and Harrington, Breit, N., Olival, K., Daszak, P., 2017. Global hotspots and correlates of emerging 2020; Grabowski and Mor, 2020). The results also show that those zoonotic diseases. Nature Commun. 8 (1). counties with the highest percentages of people with certain underlying Altman, D., 2020. Understanding the US failure on coronavirus—an essay by Drew Altman. BMJ m3417. health conditions, age, and poverty were also those which had higher Barouki, R., Kogevinas, M., Audouze, K., Belesova, K., Bergman, A., Birnbaum, L., death doubling times. Protecting these groups early on with income Boekhold, S., Denys, S., Desseille, C., Drakvik, E., Frumkin, H., Garric, J., support schemes could have allowed the working vulnerable to stay at Destoumieux-Garzon, D., Haines, A., Huss, A., Jensen, G., Karakitsios, S., home and avoid infection (Dasgupta et al., 2020; Yang et al., 2020). Klanova, J., Koskela, I., Laden, F., Marano, F., Franziska Matthies-Wiesler, E., Mor- ris, G., Nowacki, J., Paloniemi, R., Pearce, N., Peters, A., Rekola, A., Sarigiannis, D., Furthermore, home overcrowding was also a very important factor in Šebková, K., Slama, R., Staatsen, B., Tonne, C., Vermeulen, R., Vineis, P., 2021. The infection doubling times and a voluntary policy of providing quarantine COVID-19 pandemic and global environmental change: Emerging research needs. locations for those infected with SAR-CoV-2 would have surely slowed Environ. Int. 146, 106272. epidemic growth (Haroon et al., 2020). Becker, S., Soukup, J.M., 1999. Exposure to urban air particulates alters the Finally, while it is not clear where the next threat will come from, macrophage-mediated inflammatory response to respiratory viral infection. J. Toxicol. Environ. Health A 57 (7), 445–457. anthropogenic activity like deforestation, wildlife trade, and intensive Bélanger, M., Gray-Donald, K., O’loughlin, J., Paradis, G., Hanley, J., 2009. Influence of animal rearing, that encourages spillover from wild reservoirs, and in- weather conditions and season on physical activity in adolescents. Ann. Epidemiol. fluences the emergence and evolution of novel coronaviruses (Barouki 19 (3), 180–186. http://dx.doi.org/10.1016/j.annepidem.2008.12.008. 14 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 CDC, 2019. Diabetes data and statistics. Kirwa, K., Eckert, C., Vedal, S., Hajat, A., Kaufman, J., 2021. Ambient air pollution and CDC, 2020a. Assessing risk factors for severe COVID-19 illness. risk of respiratory infection among adults: evidence from the multiethnic study of CDC, 2020b. Social-distancing. atherosclerosis (MESA). BMJ Open Respir. Res. 8 (1), e000866. CDC, 2021. People with certain medical conditions. Kröger, M., Schlickeiser, R., 2020. Gaussian doubling times and reproduction factors of Chan, K., Peiris, J.M., Lam, S., Poon, L., Yuen, K., Seto, W., 2011. The effects of the COVID-19 pandemic disease. Front. Phys. 8, http://dx.doi.org/10.3389/fphy. temperature and relative humidity on the viability of the SARS coronavirus. Adv. 2020.00276. Virol. 2011, 1–7. Lauer, S.A., Grantz, K., Bi, Q., Jones, F., Zheng, Q., Meredith, H., Azman, A., Reich, N., Chapman, S., Harrington, C., 2020. Policies matter! factors contributing to nursing Lessler, J., 2020. The incubation period of coronavirus disease 2019 (COVID-19) home outbreaks during the COVID-19 pandemic. Policy Politics Nurs. Pract. 21 from publicly reported confirmed cases: Estimation and application. Ann. Internal (4), 191–192. http://dx.doi.org/10.1177/1527154420959021. Med. 172 (9), 577–582. Chauhan, A., Johnston, S.L., 2003. Air pollution and infection in respiratory illness. Br. Leclerc, Q., Fuller, N., Knight, L., Funk, S., Knight, G., 2020. What settings have been Med. Bull. 68 (1), 95–112. linked to SARS-CoV-2 transmission clusters? Wellcome Open Res. 5, 83. Choi, Y., Tuel, A., Eltahir, E., 2021. On the environmental determinants of COVID-19 Lee, A., English, P., Pankhania, B., Morling, J., 2021. Where England’s pandemic seasonality. GeoHealth 5 (6), http://dx.doi.org/10.1029/2021gh000413. response to COVID-19 went wrong. Public Health 192, 45–48. Cole, M., Ozgen, C., Strobl, E., 2020. Air pollution exposure and Covid-19 in dutch Leung, N., 2021. Transmissibility and transmission of respiratory viruses. Nat. Rev. municipalities. Environ. Resour. Econ. 76 (4), 581–610. Microbiol.. Croft, D., Zhang, W., Lin, S., Thurston, S., Hopke, P., Masiol, M., Squizzato, S., van Lewis(a), D., 2020. Why many countries failed at COVID contact-tracing — but some Wijngaarden, E., Utell, M., Rich, D., 2018. The association between respiratory got it right. Nature 588 (7838), 384–387. infection and air pollution in the setting of air quality policy and economic change. Lewis(b), D., 2020. Is the coronavirus airborne? Experts can’t agree. Nature 580 (7802), Ann. Am. Thorac. Soc.. 175. CSSE, 2020. Johns Hopkins University’s Centre for Systems Science and Engineering’s Lolli, S., Chen, Y., Wang, S., Vivone, G., 2020. Impact of meteorological conditions (CSSE) GitHub repository. and air pollution on COVID-19 pandemic transmission in Italy. Sci. Rep. 10 (1), Daniel, L., 2018. ‘we like to live in the weather’: Cooling practices in naturally http://dx.doi.org/10.1038/s41598-020-73197-8. ventilated dwellings in Darwin, Australia. Energy Build. 158, 549–557. http://dx. Lurie, M., Silva, J., Yorlets, R., Tao, J., Chan, P., 2020. Coronavirus disease 2019 doi.org/10.1016/j.enbuild.2017.10.031. epidemic doubling time in the United States before and during stay-at-home Dasgupta, S., Bowen, V., Leidner, A., Fletcher, K., Musial, T., Rose, C., Cha, A., Kang, G., restrictions. J. Infect. Dis. 222 (10), 1601–1606. http://dx.doi.org/10.1093/infdis/ Dirlikov, E., Pevzner, E., Rose, D., Ritchey, M., Villanueva, J., Philip, C., Liburd, L., jiaa491. Oster, A., 2020. Association between social vulnerability and a county’s risk for Macdonald, S., Deacon, L., Nixon, J., Akintola, A., Gillingham, A., Kent, J., Ellis, G., becoming a COVID-19 hotspot — United States, June 1–July 25, 2020. MMWR Mathews, D., Ismail, A., Sullivan, S., Dore, S., Highmore, L., 2018. ‘the invisible Morb. Mortal. Wkly. Rep. 69 (42), 1535–1541. http://dx.doi.org/10.15585/mmwr. enemy’: disability, loneliness and isolation. Disabil. Soc. 33 (7), 1138–1159. mm6942a3. Mahase, E., 2020. Covid-19: What do we know about “long covid”? BMJ m2815. DATA.CDC.GOV, 2020. U.S. state, territorial, and county stay-at-home orders: March Margraf, J., Brailovskaia, J., Schneider, S., 2020. Behavioral measures to fight COVID- 15-may 5 by county by day. 19: An 8-country study of perceived usefulness, adherence and their predictors. de Freitas, C., 2014. Weather and place-based human behavior: recreational preferences PLOS ONE 15 (12), e0243523. and sensitivity. Int. J. Biometeorol. 59 (1), 55–63. Middleton, J., Reintjes, R., Lopes, H., 2020. Meat plants—a new front line in the de Oliveira, P., Mesquita, L., Gkantonas, S., Giusti, A., Mastorakos, E., 2021. Evolution covid-19 pandemic. BMJ m2716. of spray and aerosol from respiratory releases: theoretical estimates for insight on Mongey, S., Pilossoph, L., Weinberg, A., 2020. Which Workers Bear the Burden of viral transmission. Proc. R. Soc. A: Math. Phys. Eng. Sci. 477 (2245), 20200584. Social Distancing? Working Paper 27085, National Bureau of Economic Research, Dobson, A., Pimm, S., Hannah, L., Kaufman, L., Ahumada, J., Ando, A., Bernstein, A., http://www.nber.org/papers/w27085. Busch, J., Daszak, P., Engelmann, J., Kinnaird, M., Li, B., Loch-Temzelides, T., Muniz-Rodriguez, K., Chowell, G., Cheung, C., Jia, D., Lai, P., Lee, Y., Liu, M., Ofori, S., Lovejoy, T., Nowak, K., Roehrdanz, P., Vale, M., 2020. Ecology and economics for Roosa, K., Simonsen, L., Viboud, C., Fung, I., 2020. Doubling time of the COVID-19 pandemic prevention. Science 369 (6502), 379–381. epidemic by Chinese Province. http://dx.doi.org/10.1101/2020.02.05.20020750. Drewnowski, A., Specter, S., 2004. Poverty and obesity: the role of energy density and NCEI, 2021. NOAA’s National Centers for Environmental Information (NCEI). energy costs. Am. J. Clin. Nutr. 79 (1), 6–16. Nissen, K., Krambrich, J., Akaberi, D., Hoffman, T., Ling, J., Lundkvist, Å., Svensson, L., Dyer, C., 2020. Covid-19: UK government response was overcentralised and poorly Salaneck, E., 2020. Long-distance airborne dispersal of SARS-CoV-2 in COVID-19 communicated, say peers. BMJ m4445. wards. Sci. Rep. 10 (1). ECDC, 2020. Risk factors and risk groups. Nor, N., Yip, C., Ibrahim, N., Jaafar, M., Rashid, Z., Mustafa, N., Hamid, H., Emerson, E., Fortune, N., Llewellyn, G., Stancliffe, R., 2021. Loneliness, social support, Chandru, K., Latif, M., Saw, P., Lin, C., Alhasa, K., Hashim, J., Nadzir, M., 2021. social isolation and wellbeing among working age adults with and without Particulate matter (PM2.5) as a potential SARS-CoV-2 carrier. Sci. Rep. 11 (1), disability: Cross-sectional study. Disabil. Health J. 14 (1), 100965. EPA, 2020. Air data: Air quality data collected at outdoor monitors across the US. http://dx.doi.org/10.1038/s41598-021-81935-9. ERS, 2020. U.s. Department of Agriculture. Nowroozpoor, A., Choo, E., Faust, J., 2020. Why the United States failed to contain Fana, M., Torrejón Pérez, S., Fernández-Macías, E., 2020. Employment impact of Covid- COVID-19. J. Am. Coll. Emerg. Phys. Open 1 (4), 686–688. 19 crisis: from short term effects to long terms prospects. J. Ind. Bus. Econ. 47 (3), O’Neill, D., 2020. Covid-19 in care homes: the many determinants of this perfect storm. 391–410. BMJ m2096. http://dx.doi.org/10.1136/bmj.m2096. Finch, W., Hernández Finch, M., 2020. Poverty and Covid-19: Rates of incidence and OxCGRT, 2020. Covid policy tracker. deaths in the united states during the first 10 weeks of the pandemic. Front. Sociol. Papageorge, N., Zahn, M., Belot, M., van den Broek-Altenburg, E., Choi, S., Jami- 5. son, J., Tripodi, E., 2021. Socio-demographic factors associated with self-protecting Grabowski, D., Mor, V., 2020. Nursing home care in crisis in the wake of COVID-19. behavior during the Covid-19 pandemic. J. Popul. Econ. 34 (2), 691–738. JAMA 324 (1), 23. http://dx.doi.org/10.1001/jama.2020.8524. Patel, J., Nielsen, F., Badiani, A., Assi, S., Unadkat, V., Patel, B., Ravindrane, R., Grigg, J., 2018. Air pollution and respiratory infection: An emerging and troubling Wardle, H., 2020. Poverty, inequality and COVID-19: the forgotten vulnerable. association. Am. J. Respir. Crit. Care Med. 198 (6), 700–701. Public Health 183, 110–111. Hale, T., Angrist, N., Goldszmidt, R., Kira, B., Petherick, A., Phillips, T., Webster, S., Pellis, L., Scarabel, F., Stage, H., Overton, C., Chappell, L., Fearon, E., Bennett, E., Cameron-Blake, E., Hallas, L., Majumdar, S., Tatlow, H., 2021. A global panel Lythgoe, K., House, T., Hall, I., 2021. Challenges in control of COVID-19: short database of pandemic policies (Oxford COVID-19 Government Response Tracker). doubling time and long delay to effect of interventions. Philos. Trans. R. Soc. B Nat. Hum. Behav. 5 (4), 529–538. 376 (1829), 20200264. http://dx.doi.org/10.1098/rstb.2020.0264. Ham, C., 2021. The UK’s poor record on covid-19 is a failure of policy learning. BMJ Schuit, M., Ratnesar-Shumate, S., Yolitz, J., Williams, G., Weaver, W., Green, B., n284. Miller, D., Krause, M., Beck, K., Wood, S., Holland, B., Bohannon, J., Freeburger, D., Haroon, S., Chandan, J., Middleton, J., Cheng, K., 2020. Covid-19: breaking the chain Hooper, I., Biryukov, J., Altamura, L., Wahl, V., Hevey, M., Dabisch, P., 2020. of household transmission. BMJ m3181. http://dx.doi.org/10.1136/bmj.m3181. Airborne SARS-CoV-2 is rapidly inactivated by simulated sunlight. J. Infect. Dis. Hawkins, N.M., Jhund, P.S., McMurray, J.J., Capewell, S., 2012. Heart failure and 222 (4), 564–571. socioeconomic status: accumulating evidence of inequality. Eur. J. Heart Fail. 14 Solimini, A., Filipponi, F., Fegatelli, D.A., Caputo, B., De Marco, C.M., Spagnoli, A., (2), 138–146. http://dx.doi.org/10.1093/eurjhf/hfr168. Vestri, A.R., 2021. A global association between Covid-19 cases and airborne Jelinek, H., Osman, W., Khandoker, A., Khalaf, K., Lee, S., Almahmeed, W., Alsafar, H., particulate matter at regional level. Sci. Rep. 11 (1), http://dx.doi.org/10.1038/ 2017. Clinical profiles, comorbidities and complications of type 2 diabetes mellitus s41598-021-85751-z. in patients from United Arab Emirates. BMJ Open Diabetes Res. Care 5 (1), Tang, J., Marr, L., Li, Y., Dancer, S., 2021. Covid-19 has redefined airborne e000427. transmission. BMJ n913. Kaplin, A., Junker, C., Kumar, A., Ribeiro, M., Yu, E., Wang, M., Smith, T., Rai, S., Taylor, D., 2014. Americans with disabilities: 2014. Household Econ. Stud.. Bhatnagar, A., 2021. Evidence and magnitude of the effects of meteorological Travaglio, M., Yu, Y., Popovic, R., Selley, L., Leal, N., Martins, L., 2021. Links between changes on SARS-CoV-2 transmission. PLOS ONE 16 (2), e0246167. http://dx.doi. air pollution and COVID-19 in England. Environ. Pollut. 268, 115859. org/10.1371/journal.pone.0246167. USCB, 2020. United States Census Bureau. 15 M.J. Watts Spatial and Spatio-temporal Epidemiology 43 (2022) 100539 Verity, R., Okell, L., Dorigatti, I., Winskill, P., Whittaker, C., Imai, N., Cuomo- Williamson, E., Walker, A., Bhaskaran, K., Bacon, S., Bates, C., Morton, C., Curtis, H., Dannenburg, G., Thompson, H., Walker, P., Fu, H., Dighe, A., Griffin, J., Mehrkar, A., Evans, D., Inglesby, P., Cockburn, J., McDonald, H., MacKenna, B., Baguelin, M., Bhatia, S., Boonyasiri, A., Cori, A., Cucunubá, Z., FitzJohn, R., Tomlinson, L., Douglas, I., Rentsch, C., Mathur, R., Wong, A., Grieve, R., Har- Gaythorpe, K., Green, W., Hamlet, A., Hinsley, W., Laydon, D., Nedjati-Gilani, G., rison, D., Forbes, H., Schultze, A., Croker, R., Parry, J., Hester, F., Harper, S., Riley, S., van Elsland, S., Volz, E., Wang, H., Wang, Y., Xi, X., Donnelly, C., Perera, R., Evans, S., Smeeth, L., Goldacre, B., 2020. Factors associated with Ghani, A., Ferguson, N., 2020. Estimates of the severity of coronavirus disease COVID-19-related death using OpenSAFELY. Nature 584 (7821), 430–436. 2019: a model-based analysis. Lancet Infect. Dis. 20 (6), 669–677. Wood, S.N., 2017. Generalized Additive Models: An Introduction with R. CRC Press. Ward, E., Jemal, A., Cokkinides, V., Singh, G., Cardinez, C., Ghafoor, A., Thun, M., Yang, C., Briar-Lawson, K., Urbaeva, J., 2020. Employment and income support policies 2004. Cancer disparities by race/ethnicity and socioeconomic status. CA: Cancer J. during the early phases of COVID-19: Lessons from the U.S., Denmark, and Taiwan. Clin. 54 (2), 78–93. Greenwich Soc. Work Rev. 1 (2), 97–108. http://dx.doi.org/10.21100/gswr.v1i2. Wardeh, M., Baylis, M., Blagrove, M., 2021. Predicting mammalian hosts in which novel 1167. coronaviruses can be generated. Nature Commun. 12 (1). Zhang, R., Li, Y., Zhang, A., Wang, Y., Molina, M., 2020. Identifying airborne Whitehead, M., Taylor-Robinson, D., Barr, B., 2021. Poverty, health, and covid-19. BMJ transmission as the dominant route for the spread of COVID-19. Proc. Natl. Acad. n376. Sci. 117 (26), 14857–14863. WHO, 2020. COVID-19 and NCD risk factors. 16