Infectiousness of places – Impact of multiscale human activity places in the transmission of COVID-19

COVID-19 raises attention to epidemic transmission in various places. This study analyzes the transmission risks associated with human activity places at multiple scales, including different types of settlements and eleven types of specific establishments (restaurants, bars, etc.), using COVID-19 data in 906 urban areas across four continents. Through a difference-in-difference approach, we identify the causal effects of activities at various places on epidemic transmission. We find that at the micro-scale, though the transmission risks at different establishments differ across countries, sports, entertainment, and catering establishments are generally more infectious. At the macro-scale, contradicting common beliefs, it is consistent across countries that transmission does not increase with settlement size and density. It is also consistent that specific establishments play a lesser role in transmission in larger settlements, suggesting more transmission happening elsewhere. These findings contribute to building a system of knowledge on the linkage between places, human activities, and disease transmission.

gatherings. Information on the closure of outdoor sports grounds is incomplete thus excluded from further analysis.

Spatial units and characteristic data in the United Kingdom. Although the Office for
National Statistics of the United Kingdom delineates 'built-up areas', which is consistent with our concept of settlements, they cannot be linked with the infection case data. Because to compute the daily infection cases in 'built-up areas', infection case data at the Middle layer Super Output Areas (MSOA) is needed, which can then be aggregated to the 'built-up area' level. Though the U.K. government does publish case data at the MSOA level, only latest data can be downloaded from the official website, so that we cannot compile the historical daily infection cases in the 'built-up areas'. Instead, we use the Local Authority Districts (LAD) as the spatial unit of analysis. There are 290 LADs with population larger than 100,000 in 2020; however, these include 32 London boroughs which are combined, resulting in 259 spatial units 42 . The sources for other demographic and economic characteristics of LADs are: population age profile from the mid-2020 edition of population estimates 42 , ethnicity profile from Census 2011 43 , and gross domestic product per head from the 2019 edition of regional gross domestic product estimates 44 (we do not find personal income data at the LAD level).

COVID-19 infection data in the United
Kingdom. The U.K. government publishes  infection case data in fine-grained units 45 . We request the data at the lower tier local authority level and aggregate the case numbers to LADs and London.
Place closure and other intervention data in the United Kingdom. The government interventions in COVID-19 are decided by the four countries in UK, thus are collected from the official websites of the corresponding governments except for Northern Ireland, for which the information is not complete. Since the interventions are mostly uniform within the countries, four groups of interventions are highly simultaneous (Kendall's  > 0.95) thus combined: closing non-essential retail and banning large-size indoor gatherings; closing restaurants, cultural venues and banning small-size indoor gatherings; closing entertainment venues and outdoor sports grounds; and banning small-size and large-size indoor gatherings. After cleaning missing information, we end up with 234 LAD spatial units.

Spatial units and characteristic data in the United States. We choose Metropolitan Statistical
Areas (MSAs) as the spatial unit of analysis in the United States, which are regions with a relatively high population density at the core and socio-economically linked communities in the surroundings, delineated by the U.S. Office of Management and Budget. Since continuous built-up areas often spread beyond administrative borders, MSAs are more suitable for the analysis, although they are larger than a single continuous built-up area and contain rural areas in many cases. There are 363 MSAs with a population larger than 100,000 according to the 2019 estimates 46 . To more accurately represent the population and density of major settlements in MSAs, we link MSAs with another statistical unit-urban areas, and compute the sum of urban population and urban land area in each MSA 47  Place closure and other intervention data in the United States. We collect the dates of the interventions at the state level from the websites of state governments, which is the major level of government in charge of intervention policy making in COVID-19. By doing so, we do not account for the few cases in which county or local governments take alternative actions, which are relatively few 34 . We do not find clear information on closing offices, outdoor sports grounds and banning indoor gatherings thus these interventions are not included in further analysis. We end up with 308 MSA spatial units after cleaning missing information.
Spatial units and characteristic data in Brazil. We use the second-level administrative division-municipalities, as the units of study in Brazil. In 2020, there are 326 municipalities with more than 100,000 population in Brazil, which are taken as the subjects of analysis 54 . According to Brazil's last census, the proportion of urban population is larger than 80% in most of these municipalities (median = 96%), indicating that related demographic and economic statistics would mainly reflect the conditions of the dense urban settlements in the municipalities 55 . The population age profile is also from the 2010 Census 55 ; the data for personal income (2019 data, available only at the state level) and gross domestic product per head is from the Brazilian Institute of Geography and Statistics (2018 data) 56,57 . Place closure and other intervention data in Brazil. The interventions are mainly decided by the state governments in Brazil, therefore collected from the state government websites. Three pairs of interventions are implemented and relaxed simultaneously during the study period (Kendall's  = 1), which are the closures of schools and childcare centers, the bans on small-size indoor gatherings and small-size outdoor gatherings, and the bans on large-size indoor gatherings and large-size outdoor gatherings. We end up with 319 spatial units after cleaning missing information. Table 1 describes the closure of eleven types of establishments and five other interventions in this study. The interventions are coded 0, 0.5 or 1: 0 indicates that an intervention is not in place, 1 indicates a full restriction, and 0.5 indicates a partial restriction, including recommending instead of enforcing closures, or only closing establishments larger than a certain size, or shortening opening hours. Note that all interventions in Japan are non-compulsory, but since they are reported to have similar effects as full restrictions, we code them as 1 60 .

Coding of interventions. Supplementary
Estimating Rt. The instantaneous reproduction number, Rt, is the average number of secondary infected cases caused by a primary infected individual at time t, reflecting the speed of virus spread in a certain area, which is taken as the outcome of interest when analyzing the impacts of various types of places. Rt is estimated using the method developed by Cori et al., which is widely used in epistemology 61 . The method requires parameters of serial interval, which is set to be constant with mean=7.5 days and standard deviation=3.4 days following previous epidemiological investigation of  . Further, to smooth out daily fluctuations, a 7-day sliding window is applied on the daily cases. Finally, Rt estimates with a coefficient of variation larger than 0.3, indicating insufficient cases in the time window to generate reliable estimates, are excluded from further analysis.
Testing for parallel pre-trend. To acquire reliable estimates from DiD analysis, the data need to meet the assumption of parallel trend, meaning that the outcome of interest should move in parallel trend in all units, absence of intervention. Since we cannot directly observe the counterfactual trend in units that have implemented a closure, we examine this assumption on the pre-intervention periods. This is implemented by an event study, adding terms indicating preintervention and post-intervention periods to the basic two-way fixed-effect model, specified as follows (4) where xc,i,s,t,t+b (b  5) denotes the change in the status of establishment s in unit i from day t to day t+b and the same applies to xc,s,i,t+b,t+5, xc,s,i,t-j-1,t-j and xc,s,i,1,t-a-1. Correspondingly, c,s,pre denotes the estimate of pre-trend b days before a change, which is the coefficient of concern.
c,s,pre', c,s,j and c,s,posta are to control for the effects in other periods: c,s,pre' denotes the estimate of pre-trend b to 5 days before a change (if b=5, then the estimate is NA); c,s,j is the effect j days (0  j  a) after an intervention; c,s,posta is the effect a days after a change, which should all be interpreted as compared to the period preceding to 5 days before a change. We test the pre-trend in 1 to 5 days before an intervention, since establishment closure and reopening tend to be quick decisions. The choice of a does not significantly affect the estimates, for which we choose 14, equaling to two weeks. X'c,i,t denotes the status of the other ten types of establishments in unit i on day t and c denotes the coefficients. The rest of the notations are the same as in Eq. (1). The inclusion of pre-intervention variable can be considered as a placebo test, which should be insignificant if the parallel trend assumption is satisfied.

Robustness check.
To examine the robustness of the results, we conduct a series of experiments to test how alternative settings would affect our conclusion. First, we test whether the estimates are strongly affected by certain units in the data by re-running the models withholding part of the data. Second, we explore the confounding of possible missing variables. Last, we examine whether the interaction between establishment closures and settlement characteristics is sensitive to the indicators of settlement size and density.
To test the sensitivity of our results to specific units, we re-estimate Eq. (1) by withholding one spatial unit at a time for k times, where k is the number of spatial units in each country.
Supplementary Figure 5 shows the distribution of point estimates in these alternative settings, which are mostly close to the estimates in the default setting. In the few cases where an alternative estimate departs from the default, the sign of the estimate does not change unless the default estimate is close to zero and not statistically significant.
We control the status of five other interventions in estimating the impact of closing establishments. However, Rt may also be influenced by other variables missed from our analysis.
To explore whether our estimated impacts of closing establishments could be affected by missing variables, we assess (1) how the estimates change when we include additional confounding variables-in this case, requirements of wearing masks, and (2) how they change when we exclude existing variables. The results are shown in Supplementary Figure 6, which demonstrates that the significance and magnitude of the estimates do not change much in these experimental conditions, lending more confidence to the robustness of our estimates to potentially missing variables.
Since we use spatial units larger than continuously built-up settlements, there can be multiple indicators to represent the settlement conditions in a spatial unit. For example, the spatial unit in Japan-prefecture is the first-level administrative division and usually contains more than one large settlement. Statistics are available on the total population and land area of second-level divisions (cities, districts, villages), as well as those of densely inhabited areas in second-level divisions. In the main analysis, we use the population and density in the densely inhabited area of the largest city in a prefecture to represent the settlement conditions, since they reflect the conditions of the most populated area of the prefecture. However, there can be other plausible indicators, including the total population of the largest city in a prefecture and the total population of all densely inhabited areas across a prefecture as indicators of settlement size, and the average density of all densely inhabited areas across a prefecture as the indicator of settlement density. The spatial units analyzed in other countries are more similar to the extents of continuous built-up areas, but may also contain separate small settlements or non-built-up areas.
We conduct similar robustness check on the United States, where we acquire data on both the population and density in the entire areas of MSAs and those in the urban areas. The latter are used in the main analysis and we test the sensitivity with the other indicators. Supplementary   Figure 7 shows that the results are generally robust to the choice of indicators.

Supplementary Notes -Government interventions in first-wave COVID-19 in the sample countries
Japan The first reported case of COVID-19 in Japan was confirmed on January 16 2020. The daily new cases were at two digits in February and March and rose to three digits in April. The central government started to declare state of emergency in April, which conferred relevant regional governments the power to implement partial or full lockdown. However, lockdown in Japan was distinct from many other countries in that government could only request residents to stay at home or businesses to close but could not force it, which relied on peer pressure and people's deference to authority. On April 7 2020, the central government declared a state of emergency for seven prefectures and extended the emergency state to nationwide nine days later. From late March to mid-April, the 47 prefectures gradually requested stay-at-home and closure of non-essential facilities. Certain prefectures, such as Tokyo and Hokkaido, issued the requests even before the central government granted the power through the emergency declaration.
Despite of the lenient approach, it was reported that there was fairly extensive voluntary compliance. The daily new cases fell back to two digits by mid-May, and remained at that low level till July. Correspondingly, prefectures started to release the interventions from late May, which however was followed by a stronger resurgence of infections in July and August (1,000+ new cases a day at the peak). Wales continued to require people to stay within five miles from home till July 6, while people in England were free to leave their home from June 1.

United States
The first reported case of COVID-19 in the United States was confirmed on January 20. The situation aggravated to thousands of new cases a day in March and more than 100,000 cases in total by late March. From March to April, the federal, state and local governments launched a number of orders and guidelines on social distancing, including ordering stay-at-home. Similar to Brazil, the power of ordering mass quarantine and closure lies primarily with the states, while the federal government provides recommendations and guidelines. Though all states imposed place-related interventions, the timetable and stringency varied a lot. Eight states did not mandate state-wide stay-at-home in our study period, while the dates of imposing stay-at-home in the other states ranged from March 19 to Apr 7. There were also up to more than 20 days differences in the closure of various businesses and facilities, if ever closed. In May and June, many states began to loosen the restrictions though cases were still increasing at a rate of more than 20,000 a day, which was followed by a resurgence to more than 60,000 cases a day in July. By the end of our study period, the United States became the country with the highest number of cases in the world.

Brazil
The first reported case of COVID-19 in Brazil was confirmed on February 25, 2020, which was also the first confirmed case in the South America. The number of infections rose up to more than 1000 by mid-to-late March, when state and municipal governments started to declare an emergency situation and take measures such as closing non-essential businesses and suspending public activities. However, the interventions launched by state and local governments were undermined by the federal government especially the president, who kept dismissing the danger of the pandemic even issuing orders to expand the classification of essential businesses.
Due to the autonomy of the states and the lack of federal-level coordination, heterogeneous interventions were observed across the country, e.g. the size of gathering banned by state governments in March varied from five people to hundreds.
Despite of the measures, the number of infections did not turn to a downward trajectory for months. By May 2020, Brazil became a new epicenter with more than 10,000 new cases  Figure 6. Robustness of the estimates to omitting or including additional variables. The first bar in each group (dark grey) is the estimate in the default setting, followed by bars indicating estimates from alternative settings shown in the legend (except for the one that excludes the variable being examined). The error bars represent 95% confidence intervals.

Supplementary Figure 7. Interaction between alternative indicators of settlement size and density and the impacts of establishment closures.
The ribbons represent 95% confidence intervals. The blue lines are estimates from low-value groups and red lines are from high-value groups. Supplementary Table 1

Place closures
Closing schools Including all types of schools. If face-to-face teaching is allowed only for certain types of schools or classes, then this is considered as a partial restriction.
Closing childcare centers Note that the service for essential workers usually remain.

Closing offices
Closing non-essential retails Including all retail establishments other than those considered essential, such as grocery stores, supermarkets and pharmacies. The definition of essential retails may differ across countries and regions.
Closing restaurants Including restaurants, cafes, canteens, etc. Note that food delivery and take-away service usually remain.
Closing cultural venues Including museums, libraries, galleries, etc.
Closing religious venues Including worship places for various religions.
Closing indoor sports venues Including gyms, dance studios, indoor sports fields, etc.
Closing outdoor sports grounds Including playgrounds, sports fields, etc.

Other interventions
Ordering stay-at-home Requiring citizens to stay at home unless going out for essential purposes such as working in the essential sectors, shopping for necessities and seeing doctors. There can be small variations in full restriction in certain cases, for example people are allowed to travel within five miles from home in Wales in early June 2020, in which case the intervention is still coded as 1.
Banning small-size outdoor gatherings The size limit is at most 10 people.
Banning large-size outdoor gatherings The size limit is larger than 10.
Banning small-size indoor gatherings The size limit is at most 10 people.
Banning large-size indoor gatherings The size limit is larger than 10.