Policy Research Working Paper 10189 When the Lights Go Out The Economic Impacts of Covid-19 on Cities Globally Amjad Muhammad Khan Hogeun Park Mark Roberts Putu Sanjiwacika Wibisana Urban, Disaster Risk Management, Resilience and Land Global Practice & Water Global Practice September 2022 Policy Research Working Paper 10189 Abstract This paper uses high-frequency nighttime time lights data also observed. Cities in low- and middle-income countries to estimate the impacts of the Covid-19 crisis on economic faced a significantly larger overall loss of economic activity activity during the first year of the pandemic for a global compared to those in high-income countries. Additionally, sample of 2,800 cities, covering a total population of 2.5 cities with higher population densities are found to be more billion people. Activity is found to be negatively affected resilient in the face of the global shock as compared to less by both the spread of the virus and the imposition of non- dense ones, but this difference is only observed in low- and pharmaceutical interventions, but the negative impacts of middle-income countries. Taken together, the findings sug- the spread are large compared to those of nonpharmaceu- gest that the Covid-19 crisis gave rise to divergence in urban tical interventions. Large differences in city trajectories are economic trajectories, both across and within countries. This paper is a product of the Urban, Disaster Risk Management, Resilience and Land Global Practice and the Water Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http:// www.worldbank.org/prwp. The authors may be contacted at akhan31@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team When the Lights Go Out: The Economic Impacts of Covid-19 on Cities Globally Amjad Muhammad Khan, Hogeun Park, Mark Roberts and Putu Sanjiwacika Wibisana1 JEL codes: O18, R11, R12 Keywords: Covid-19, nighttime lights, non-pharmaceutical interventions, cities, urban resilience 1All authors are affiliated with the World Bank. Amjad Khan is affiliated with the World Bank’s Water Global Practice; and Hogeun Park, Mark Roberts, and Putu Sanjiwacika Wibisana with the World Bank’s Urban, Disaster Risk Management, Resilience, and Land Global Practice. Corresponding author: Amjad Khan (akhan31@worldbank.org). This paper was produced as a background paper to the World Bank report Vibrant Cities: Priorities for Green, Resilient, and Inclusive Urban Development. The authors thank Somik Lall and Forhad Shilpi for helpful comments and suggestions. They also thank Pui Shen Yoong and Ran Goldblatt for early discussions related to this work, and Andres Elizondo for his help with the processing of the nighttime lights data. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. 1. Introduction The sudden, global onset of the Covid-19 pandemic in March 2020 brought major disruptions to human activity worldwide. As the pandemic unfolded, governments around the world applied a variety of countermeasures to contain the spread of the virus as well as the economic damage it caused. The pandemic impacted urban economic activity in particular, given the central place of cities in global transport networks as well as the increased susceptibility to contagion that accompanies their high densities. Indeed, whether the concentration of economic activity in large cities would survive the pandemic was one of the most hotly debated issues during the start of the pandemic. The role of population density during such a crisis is ex-ante ambiguous as there are two opposing forces at play: “health- preserving behavior” and a “resilience effect”. On the one hand, denser cities are more prone to contagion risks during an outbreak,2 which nudges households and firms to curb in-person economic activities more in denser areas to avoid health risks. Such health-preserving behavior includes both state-sanctioned and voluntary reductions of activities which require movement outside the home and interactions with others. Bisin and Moro (2022) show that incorporating a spatial dimension into the now well-known Susceptible, Infectious, and/or Recovered (SIR) model reveals that city density is a crucial determinant of epidemic diffusion – since density is proportional to an affected agent’s number of contacts. This is especially the case in densely populated cities in low- and middle-income countries where density is often associated with crowding in the form of slums and social distancing is difficult (Bhardwaj et al., 2020; Jedwab et al., 2021; Lall and Wahba, 2021). On the other hand, historical experience suggests that denser cities may be more resilient to large negative shocks (city resilience). Numerous studies have demonstrated long-run persistence in the spatial distribution of economic activity, even in the face of very large adverse shocks. This persistence can arise as a result of various forces, including locational fundamentals (i.e., natural advantages); sunk investments in durable capital, including transport and utility networks and buildings (i.e., man-made advantages); agglomeration economies; and institutions (Davis and Weinstein, 2002; Bleakley and Lin, 2012; Henderson et al., 2018; Dell and Olken, 2019). In addition, cities can benefit from higher density if it entails higher levels of human capital, diversification of economic bases and higher capacities to provide public goods and services (Duranton and Puga 2004, 2014, 2020). In lower-income countries, more densely populated cities also tend to have higher levels of access to water, sanitation and hygiene (WASH) and health services (Henderson et al., 2019; World Bank, 2009). If forward-looking economic agents believe that denser cities are more resilient, this may result in both households and firms locating in these cities instead of less dense cities when confronted by a large “symmetric” shock that affects all cities. For these reasons, density may also play a positive role in limiting the adverse economic impacts of Covid-19. However, relatively little is known regarding the magnitudes of the pandemic’s impacts on cities or their determinants. Standard official statistics, such as for gross domestic product (GDP), lack the spatial and temporal granularity necessary for a robust analysis of such impacts. Hence, even for high-income countries, GDP data is rarely available for cities and, when it is, it tends to be at an annual frequency. Standard official statistics are also invariably published with a lag, which undermines the possibilities for “near real-time” analysis. This study circumvents the data availability issue by using high-frequency (monthly) nighttime lights data derived from the Visible Infrared Imaging 2The higher risk of spread of contagious diseases in more densely populated cities has been cited by Glaeser as one of the “demons of density” – to quote, “If two people are close enough to exchange an idea face to face they are also close enough to exchange a contagious disease… ” (https://www.cfr.org/event/poor-world-cities- conversation-edward-glaeser). 2 Radiometer Suite (VIIRS) instrument to derive a proxy measure of economic activity. This proxy measure of economic activity is used to examine the impacts of the Covid-19 crisis for a global sample of 2,830 cities drawn from 162 countries. Our analysis first documents large, but heterogeneous, declines in light intensity in cities around the globe during the first 15 months following the emergence of Covid-19, starting in January 2020, which is when the World Health Organization (WHO) first declared the disease a “Public Health Emergency of International Concern”.3 A difference in the decline in lights in cities in high- and lower- income countries is observed. The majority of the global loss in lights – i.e., the total light “lost” over the period – is concentrated in low- and middle-income countries. Losses are also higher for the median city in lower-income countries than for the median city in high-income countries. We then proceed to analyze what may have determined the size of these declines in lights. Panel data estimates exploiting within-country variation in Covid-19-related deaths and NPIs suggest that a more severe loss of lights is associated with both the spread of the disease and imposition of NPIs, mainly driven by the experience of cities in lower-income countries. The negative impact of Covid- 19-related deaths appears to be much larger than those of NPIs in relative terms. Our estimates suggest that a one standard deviation increase in the natural log of deaths reduces a city’s economic activity, as measured by the natural log of lights, by 2.5 times more than an equivalent increase in the stringency of NPIs, as captured by the Oxford Covid-19 Government Response Tracker (OxCGRT) stringency index (Hale et al., 2022). In lower-income countries, elasticities are estimated to be less negative for cities with higher population densities, implying less adverse impacts of the Covid-19 crisis on these cities as compared to less densely populated ones. Higher population densities are also associated with lower total loss of lights for cities in lower-income countries. These results suggest that, for lower-income countries, the tendency of dense agglomerations to persist outweighs the negative impacts of density during an epidemic outbreak – i.e., the city resilience effect tends to outweigh the health-preserving behavior effect. By contrast, for cities in high-income countries, population density is uncorrelated with the loss of lights. This likely reflects differences in the characteristics of lower density locations between rich and poor countries. For instance, compared to similar cities in poorer countries, low-density cities in richer countries have higher capacities to provide public goods and services that build resilience to shocks (World Bank, forthcoming). Methodologically, our paper builds on a well-established and growing literature in economics that uses nighttime lights to proxy economic activity (Henderson et al., 2012; Bleakley and Lin, 2012; Storeygard, 2016; Jedwab et al., 2021). However, with only a few exceptions, most of these studies have relied on data from Operational Linescan System (OLS) satellite sensors from the US government’s Defense Meteorological Satellite Program (DMSP). While these data have the advantage of having a relatively long annual time series spanning 1992–2013, they do not cover more recent years. They also suffer from several drawbacks that undermine their ability to proxy economic activity at finer geographic resolutions. These drawbacks include the inability to detect nighttime light variations for the brightest urban centers due to top-coding, the misattribution of lights across pixels due to overglow (or “blooming”) and the absence of calibration both across different OLS sensors and over time (Doll, 2008; Gibson et al., 2021). By contrast, this paper uses nighttime light data from the VIIRS satellite Day-Night Band (DNB) sensor. Not only are data available for more 3After the first cases of Covid-19 were identified in China in December 2019, the WHO declared Covid-19 a “Public Health Emergency of International Concern” on January 30, 2020, and as a global pandemic on March 11, 2020. 3 recent years, but they do not suffer from the aforementioned limitations of the DMSP-OLS data.4 They are furthermore available at more granular time scales (i.e., monthly composites), making them better suited for detecting variation in economic activity at the city level over shorter periods of time (Gibson et al., 2021; Roberts, 2021).5 Finally, this study also contributes to an emerging literature on the economic impacts of pandemics, galvanized by the onset of the Covid-19 crisis. Miguel and Mobarak (2022) provide a survey of the rapidly growing empirical research documenting the current pandemic's many adverse impacts on living standards, human capital, and inequality in low- and middle-income countries. Many studies have also focused on the medium- to long-run economic impacts of historic pandemics, often focusing on the redistributive impacts across labor and capital (Almond, 2006; Alfani, 2022; Jedwab et al., 2021; Jedwab et al., 2022; Jordà et al., 2022). To the best of our knowledge, our study is the first to examine the more immediate, short-run response of city trajectories to the outbreak, and how these vary across cities, at a global scale. In doing so, it builds on country-specific applications of nighttime lights data to study the economic impacts of the Covid-19 crisis for China (Elvidge et al., 2020; Liu et al., 2020), India (Beyer et al., 2021), and Morocco (Roberts, 2021). It also contributes to the debate on the efficacy of NPIs in containing the negative health and economic impacts of pandemics (Correia et al., 2020; Demirguc-Kunt et al., 2020; Barro, 2020). The evidence from our analysis suggests that, in terms of economic activity, the negative impacts of Covid-19-induced deaths outweigh the costs of NPIs.6 The next section of the paper, Section 2, introduces the data used and presents some descriptive statistics regarding the evolution of lights following the emergence of Covid-19. Section 3 then investigates the determinants of the local economic impacts of the Covid-19 pandemic using three empirical approaches – a panel data approach that exploits both cross-country and temporal variation in Covid-19 deaths and the stringency of NPIs; a cross-sectional approach that estimates the strength of the correlation between various city-level characteristics, including density, and deviations in nighttime light intensity from a city’s own pre-Covid-19 growth path following the onset of the crisis; and an event study approach that examines impacts on more versus less densely populated cities. Section 4 concludes. 2. Data Description This study combines various spatial data sets to measure aggregate economic activity in cities, the progress of Covid-19, the policy measures implemented in response, and cities’ socio-economic characteristics. The primary data is VIIRS DNB monthly composites obtained from the Earth Observation Group (EOG) at the Colorado School of Mines. This data covers the period from April 2012 through March 2021. Each monthly data point represents a pixel’s average nighttime light intensity over all cloud-free nights in that month measured in nanowatts/cm2/sr. For each month, we proxy a city’s overall level of economic activity by its “sum of lights” (SOL) – i.e., by the sum of the nighttime light intensities across all 15 arc-second (approximately 460 m2 at the equator) pixels that fall within a city’s extent. For global consistency, we use data on “urban centers” from the Global 4 While overglow still occurs in the VIIRS data, it is a much less severe problem than for the DMSP-OLS data (Small, 2019). 5 Other studies have exploited alternative proxies for economic activity derived from credit card transactions and private sector payroll firms (Chetty et al., 2020), Google Trends search data (Woloszko, 2020), and data on outdoor air pollution levels (Masaki et al., 2020). 6 Given that, for any given month, the VIIRS nighttime lights composite becomes available within a few weeks of the month’s end under the Colorado School of Mines’ Earth Observation Group (EOG) subscription service, the use of nighttime lights also has potential to contribute to the “near real-time” monitoring of economic activity in response to shocks. 4 Human Settlement – Urban Center Database (GHS-UCDB) to measure city extents, so that all cities are defined following the European Commission’s “degree of urbanization” methodology (Dijkstra and Poelman, 2014; Dijsktra et al., 2021).7 We focus on urban centers with a population of at least 200,000 in 2015. This gives a total sample of 2,830 cities drawn from 162 countries, shown in Figure 1. The minimum population threshold ensures that large rural towns are excluded from our analysis. Nonetheless, our sample is large enough to cover a total population of 2.5 billion urban dwellers. For select cities, Figure 2 shows the deviation of the SOL from their long-term trend paths, where these paths are estimated using the lights data for the full-sample period of April 2012 – March 2021.8 Specifically, the figure shows residual plots from a regression of the natural log of SOL on a full set of city fixed effects, country-specific linear time trends and country-specific month-of-year fixed effects, which are included to account for seasonality. Panel A shows the trajectories of four cities from four different countries – Beijing, China; Milan, Italy; Dhaka, Bangladesh; and Harare, Zimbabwe – while Panel B illustrates the heterogeneous trajectories cities followed within two countries – the Arab Republic of Egypt and Saudi Arabia. The graphs illustrate the substantial heterogeneity in city trajectories, both within and across countries. Figure 1: Distribution of the Global Sample of 2830 cities 7 The “degree of urbanization” is a globally consistent approach to defining cities, towns, semi -dense and rural areas that was endorsed by the United Nations’ Statistical Commission in 2020. Under this methodology, a city or “urban center” is composed of contiguous grid cells of 1 km2 having density of at least 1,500 inhabitants per km2, and total population of at least 50,000. More details on the methodology can be found on the European Commission’s Joint Research Center website: https://ghsl.jrc.ec.europa.eu/ghs_stat_ucdb2015mt_r2019a.php . 8 For Figure 2, long-term trends paths are estimated with 2021 data included, for illustration purposes. In the cross-section analysis in section 3.3 below, long-term trend paths are estimated by only using data from the pre-Covid-19 period (i.e. before 2021) to avoid biasing estimates. 5 An important point to note is that, for cities in certain regions of the world, high cloud coverage during certain periods (during monsoon seasons, for example) constrains data availability, as indicated by the circular markers in panel A. These reflect cloudy months. Therefore, in the regression analysis of Section 3, city-month observations for which nighttime lights data are missing (i.e., for which it is not possible to calculate an SOL value) are dropped. Relatedly, all regressions also include a full set of country-specific month-of-year fixed effects to account for seasonal nighttime light trends at the country level, which may be related to cloud coverage, and also, potentially, atmospheric distortions.9 Meanwhile, disease progression statistics (number of reported Covid-19 cases and deaths), and measures of both the stringency of NPIs (the stringency index) and economic support measures (the economic support index) were retrieved from The Oxford Covid-19 Government Response Tracker (OxCGRT) (Hale et al., 2022). The stringency index measures the strictness of mobility limitation resulting from government-imposed measures, while the economic support index captures the strength of economic support policies (including cash assistance, tax cut, or debt relief) implemented to combat the negative economic impacts of the crisis. Both the NPI stringency and economic support indices track policy responses starting from January 2020. It is important to note that, while this data has time variation, it is mostly only available at the country level.10 For select countries, Figure 3 shows the within-country variation in monthly Covid-19-related deaths, the stringency and economic support indices. Returning to Figure 2, it can be seen that cities’ nighttime light trajectories also appear to be correlated with the trajectories of Covid-19 deaths in the corresponding countries. Broadly, periods of higher death rates at the national level are associated with sharp negative deviations of cities’ nighttime lights from their long-term trend growth paths. The sudden implementation of NPIs early in the pandemic is also associated with falls in economic activity in China and Italy but not in poorer economies like Bangladesh and Zimbabwe. This likely reflects differences in enforcement capacity across the countries. Panel B of Figure 2 shows that these patterns also broadly hold for Egypt and Saudi Arabia, but there is variation in the impact across cities within countries as well. Using data from the GHS-UCDB, we also compiled variables related to various city characteristics, including the total land area, population, estimated GDP, and built-up area. For each city, an index of compactness — the Polsby-Popper Ratio (PPR) — was calculated based on its spatial extent to capture a city’s urban form. The length of roads within cities was derived from Open Street Map (OSM) data to capture a city’s infrastructure development. Summary statistics for all variables included in the analysis are given in Table A1 in the appendix. 9 In additional results reported in the appendix, we also show the robustness of our results to weighting observations based on the number of pixels within a city’s area that had more than one cloud -free day to address the related measurement error. In other words, these regressions give less weight to observations that are more contaminated by cloud coverage and which, therefore, are of lower quality. 10 During the earlier stages of the pandemic, most NPI and economic support measures were implemented at the national level, thereby making the absence of city-specific data less of an issue. For example, though the UK attempted locally targeted lockdowns between July and October of 2020, their efficacy varied by location and authorities reverted to nationwide lockdowns from November 2020 as cases rose. Similarly, France also reverted to a national state of emergency after a brief period of lifting the emergency between July and October 2020. This makes the measurement error implied by the use of the national NPI and economic support indices less of an issue during the first year of the pandemic. Subnational data is available for five countries (Brazil, Canada, China, United Kingdom and United States) at the state or equivalent level. We analyze this data as part of robustness checks reported in the appendix. 6 Figure 2: Different Cities, Different Trajectories Panel A: Heterogenous city trajectories across countries for (April 2012 – March 2021) (Figure 2 continues on next page) 7 Figure 2 (continued): Different Cities, Different Trajectories Panel B: Heterogenous city trajectories within countries, (April 2017– March 2021) Note: For select cities, the plots show the monthly deviation of SOL from a city’s long-term trend growth path, as measured by the residuals from a regression of the ln(SOL) on a full set of city fixed effects, country-specific linear time trends and country-specific month-of-year fixed effects to account for seasonality. Circular markers represent missing observations (i.e. city-month observations for which the percentage of cloud-free pixels is less than 1), that, for presentational purposes, have been filled using linear interpolation. Figure 3: Monthly variation in Covid-related deaths and NPIs for select countries, January 2020 – March 2021 Note: For select countries, the plots show the cross-country variation in monthly values of Covid-related deaths, and the stringency and economic support indices. Each line in each plot represents a different country. 8 3. Empirical findings 3.1 Roles of disease progression, NPIs and recovery packages in explaining the evolution of lights Panel A of Figure 4 shows the total monthly deaths for the countries in our sample, with the sum of monthly SOL residuals plotted for high-income countries and lower-income countries separately.11, 12 In the case of lower-income countries, higher levels of monthly deaths are associated with lower levels of economic activity, as measured by deviations of city lights from their trajectories (correlation coefficient = -0.35). However, this negative correlation between deaths and economic activity is less obvious in high-income countries. In fact, for high-income countries, there is a weak positive correlation (correlation coefficient = 0.05) between deaths and economic activity. This could be due to the presence of reverse causality, which we discuss further below. Some other differences between high- and lower-income countries are also worth noting. Monthly deaths on the right axes show that high-income countries had much higher levels of deaths than lower-income countries – at their peak Covid-19 deaths were three times higher in high-income countries. Two important reasons for this difference exist: (i) lower levels of reporting of Covid-19 deaths and cases in lower- income countries; and (ii) a higher share of the elderly in the populations of high-income countries who are more at risk from the disease.13 Looking at the left axis on each graph also reveals that the loss of lights experienced, for all cities in aggregate, in lower-income countries is much larger than that experienced in high-income countries. Panel B of Figure 4 plots the six-month moving average of the median of the monthly SOL residuals, and the trends confirm that the average lower-income country city experienced a larger deviation in its growth trajectory as compared to the average high-income country city. This may be a result of cities in lower-income countries having different characteristics. For instance, the average population density of cities in high-income countries is 3,100 persons per square km (n = 406), while for cities in lower-income countries it is 10,900 persons per square km (n = 2,435). The regressions below estimate the impact of the Covid-19 crisis on cities globally, while the next subsection returns to the question of the role of population density in determining the economic impacts of the crisis. Lastly, a sharp rebound in aggregate activity in high-income countries is observed after December 2020. Though the reason for this is unclear, it may be a response of economic activity to the development of effective Covid-19 vaccines.14 11 Lower-income countries consist of all countries that the World Bank does not define as high-income (see https://datahelpdesk.worldbank.org/knowledgebase/articles/906519-world-bank-country-and-lending- groups). 12 Ideally, in Panel A of Figure 4, we would like to show total monthly deaths in cities. However, data for total monthly deaths is only available at the national level for countries globally. 13 Jedwab, Pereira and Roberts (2021) present data on the historical evolution of age dependency ratios, defined as the ratio of the number of people aged 65 or older to the working age population (aged 15-64), for a global sample of cities. They show that whereas, on average, age dependency ratios for cities in high-income countries stand at around 0.2 today (or around 4-5 working age adults per senior), for cities in lower-income countries they stand at below 0.1 (i.e., more than 10 working age adults per senior). 14 The US Food and Drug Administration (FDA) approved the Pfizer vaccine for emergency use by December 11, 2020. 9 Figure 4: Differences between high- and lower-income counties in the aggregate and median deviation in urban-economic activity, March 2017 – March 2021. Panel A: Total deviation from trend paths Panel B: Median of deviations from trend paths (six-month moving average) Note: Panel A shows the total aggregate deviation of all cities’ sum of lights from their long -term trend growth paths for countries grouped by income levels, alongside monthly Covid-related deaths. Panel B shows the six-month moving average of the median of deviations. Deviations are measured as the residuals from a regression of the sum of lights on a full set of city fixed effects, country-specific linear time trends and country-specific month-of-year fixed effects to account for seasonality. Note the difference in y-axis scales within and between the graphs in panel A. 10 An additional issue to address is the endogeneity of lockdowns and other NPI measures. Countries that experienced more deaths may have introduced stricter NPI measures, which may have been driving economic losses. To better disentangle the roles played by the spread of Covid-19 itself (as captured by the monthly number of reported deaths), and both NPI and economic support measures in driving the evolution of nighttime lights, panel data methods are employed. In particular, the following regression is estimated: ( ) = + + + + ℎ + [1] where for city , in country in month , ( ) is the (natural) log-transformed sum of nighttime lights level, includes a measure of deaths (as ln(deaths + 1)), and is a vector of policy variables in the form of both NPI and economic support measures. Estimates are obtained by using monthly data from January 2015 to March 2021.15 The specification also includes city fixed effects ( ) to control for unobserved time-invariant city characteristics that may be correlated with both the outcome and , and allows for differential trends of lights across countries ( ). The country- specific month-of-year effects ( ℎ ) control for potential seasonal effects at the national level, including those associated with atmospheric distortions of the lights data and any predictable monthly variation in cloud coverage. Results from the baseline regression are shown in Table 1. Column 1 shows that disease progression, as measured by the natural log of number of deaths, has a significant negative relationship with a city’s nighttime light intensity, as expected. Likewise, the estimated coefficients on the stringency and economic support indices have the a priori expected negative and positive signs, respectively. The estimates suggest that a doubling of Covid-19-related deaths reduces a city’s nighttime lights by 0.85 percent, while a jump in the stringency of NPIs from 0 to 100 – i.e., going from no NPIs whatsoever to a stringent national lockdown – reduces them by 3 percent. In relative terms, the negative impact of deaths on nighttime lights is large compared to the negative impact of NPIs. A one standard deviation change in ln(deaths) (S.D. = 3.4) and the stringency index (S.D. = 28.5) are associated with a reduction in () of 2.9 and 1.1, respectively. Columns 3 and 5 illustrate important differences between cities in high- and lower-income countries. The estimated negative coefficient on Covid-19-related deaths is driven mainly by the lower-income country subsample. The estimated coefficient on NPI stringency is found to be more negative in high- income countries, likely reflecting their greater capacity to enforce lockdown measures. On the other hand, for high-income countries, the estimated coefficient on deaths is puzzlingly positive. Similarly, the estimated coefficient on economic support is counter-intuitively negative for the lower-income country sample. As discussed further below, this could reflect reverse causality, where the continuation of economic activity resulted in higher death rates or a Covid-19 crisis induced recession led to the provision of economic support. 15Starting with January 2015 helps us to examine interactions with characteristics measured in 2015 in the next section. 11 Table 1: Covid Deaths, NPIs and Interaction with Pop Density Lower-income ALL (Low- and Middle-Income) High-income (1) (2) (3) (4) (5) (6) ln_sol ln_sol ln_sol ln_sol ln_sol ln_sol (log) Number of Death -0.00852*** 0.0674*** -0.0108*** 0.0510** 0.0138*** -0.0114 (0.000875) (0.0184) (0.000929) (0.0208) (0.00193) (0.0332) … × (log) Population -0.00727*** -0.00517*** 0.00319 (0.00163) (0.00185) (0.00373) … × (log) Area 0.00465*** 0.00145 -0.00336 (0.00103) (0.00123) (0.00368) Stringency Index -0.000371*** -0.00238 -0.000196* -0.00145 -0.000670** 0.0107 (0.000105) (0.00202) (0.000110) (0.00214) (0.000312) (0.00663) … × (log) Population 0.000355* 0.000223 -0.00148* (0.000196) (0.000208) (0.000772) … × (log) Area -0.000626*** -0.000402** 0.00161* (0.000172) (0.000184) (0.000858) Economic Support Index 0.000230* -0.00291 0.0000999 -0.00303 -0.00135*** -0.00931 (0.000119) (0.00275) (0.000128) (0.00306) (0.000369) (0.00846) … × (log) Population 0.000149 0.000188 0.00107 (0.000267) (0.000298) (0.000978) … × (log) Area 0.000252 0.000151 -0.00120 (0.000219) (0.000247) (0.00102) N 198582 198582 174061 174061 24521 24521 Cities 2841 2841 2435 2435 406 406 r2 0.962 0.963 0.956 0.956 0.937 0.937 r2_within 0.0556 0.0561 0.0593 0.0598 0.0116 0.0118 City FE Y Y Y Y Y Y Country x Month-of-year FE Y Y Y Y Y Y Country-specific linear time trend Y Y Y Y Y Y Note: Controls for city fixed effects, country-specific time trend and country-month-of-year fixed effects. Standard errors in parentheses clustered at the city-level. * p<0.1, ** p<0.05, *** p<0.01. Estimates obtained using monthly data covering the period January 2015 – March 2021. 12 An important caveat to again note here is that the numbers of Covid-19 cases and deaths, the stringency of NPIs and the strength of support measures are all measured at the national level, 16 meaning that we lack within-country variation across cities in our analysis. The empirical specification here leverages country-time variation in these variables to estimate the elasticity of city lights to national-level variation in the independent variables.17 Although this allows the analyses to cover a broad global sample, it also leads to the possibility of bias. Since changes in city-level economic activity within a country are likely correlated with subnational differences in the spread of Covid-19 or the degree to which national-level measures are implemented across a country. More broadly, reverse causality is a concern in general. The continuation of economic activity in a country’s cities may have resulted in higher death rates or the reduction in activity may have led to the provision of economic support at the national level. Given these issues, the estimated relationships should not be interpreted causally, but as associations. Next, we turn to examining how city characteristics affected the economic impacts of the Covid-19 crisis. On average, has density been associated with larger or smaller economic impacts of the Covid- 19 crisis across cities globally? Or is it other correlated city characteristics that matter? To investigate these questions, we use three different sets of econometric methods. First, we expand the panel data regressions above to include interactions with city characteristics. Second, cross-section regression analysis is used to identify determinants of economic trajectories across cities in response to the Covid-19 crisis, leveraging variation across cities within countries. Finally, difference-in-difference techniques are applied to estimate whether the impacts of the crisis have differed between high- and low-density cities. 3.2 Does density interact with Covid-19 dynamics to influence economic impacts? First, to test whether a city’s population density and other characteristics affect the strength of the observed economic impacts, we extend the regression specification in equation [1] to include interactions of these characteristics with our measure of the spread of Covid-19, and both the NPI stringency and economic support indices. Specifically, we interact and with city-specific characteristics as below: ( ) = + + ( × ) + ( × ) + + + ℎ + [2] where, in addition to the terms described before, is a vector of city-specific characteristics, which includes the (natural) log of population size in 2015 and the (natural) log of area (sq km). Taken together, these terms capture the impacts of population density. When the interactions with the city-level characteristics are added (column 2, Table 1), population density appears to amplify the negative economic impacts of disease spread. This follows from the findings that the negative economic impacts of deaths are estimated to be magnified for cities that 16 Early in the pandemic, most NPI and economic support measures were implemented at the national level (e.g., national lockdowns), arguably making our use of these variables less contentious. 17 In additional results in the appendix, we examine subnational variation for selected countries (Brazil, Canada, China, United Kingdom and United States), finding similar results. The appendix also shows that the reported results are robust to alternate specifications that utilize the number of Covid-19 cases as an alternative measure of disease spread to the number of Covid-19 deaths, and to the use of first differences of the dependent and independent variables. We also confirm the robustness of the results when observations are weighted using percentage of cloud-free pixels for each city-month observation. This means that higher weight is given to nighttime light observations that are less distorted by cloud coverage. 13 have larger populations and smaller areas (i.e., higher population density). On the other hand, higher population density is found to attenuate the negative impacts of stricter NPIs, as evidenced by the positive (negative) estimated coefficient on the interaction of the NPI measure with population (area). Columns 4 and 6 show that these interaction effects are driven by the lower-income country subsample, and the interaction terms for cities in high-income countries are of opposite signs.18 These opposing effects lead to the further question of whether, on net, higher population density is associated with larger or smaller economic impacts of the Covid-19 pandemic across cities. 3.3 Examining between-city differences in economic trajectories To understand the effect of population density (and other city characteristics) on city trajectories in the cross-section, we first develop a conceptual measurement framework for quantifying losses, both at various points in a city’s trajectory following the onset of the pandemic and in aggregate (i.e., the total loss of economic activity or total loss of light).19 In this framework, we start from the assumption that, in the absence of the Covid-19 shock, the intensity of a city’s (seasonally adjusted) nighttime lights would have continued to follow its own pre-Covid-19 trend growth path, as defined using the monthly lights data for the period April 2012 – December 2019. Given this, we then compare the observed post-Covid-19 onset, seasonally adjusted, value of a city’s nighttime lights against this counterfactual growth path. For any given month, a negative disparity between a city’s actual and counterfactual light intensity is then defined as the loss of light associated with the pandemic for that month. Based on this measure, we define the initial impact of the pandemic (I) as the loss of light experienced during the first month where the number of Covid-19 cases is greater than zero, depth (D) as the largest absolute loss of light for a month experienced within the time span between the arrival of Covid-19 in a country and December 2020 (when an effective vaccine was developed), and current (C) as the loss of lights in December 2020.20 Hence, I, D, and C capture impacts at different points in a city’s post-Covid-19 onset trajectory. Finally, we also calculate the total loss of lights (TLL), defined as the sum of lost lights, relative to the counterfactual growth path, across all months from the first month of the disease to December 2020. This provides a measure of the cumulative economic impact of the Covid-19 crisis. Based on this framework, we then estimate the correlation between population density and I, D, and C. Dividing our global sample into cities in high-income and lower-income countries, we find that population density is positively correlated with less loss of lights (i.e., smaller cumulative economic impact), but only in cities in lower-income countries. The correlations in Figure 5 show that more densely populated cities tended to suffer smaller impacts of the Covid-19 crisis in terms of TLL, the depth of the shock, and the city’s position in December 2020. Prima facie, this suggests that, for cities in lower-income countries, high population densities have attenuated the impacts of the crisis. 18 The results in table 1 also reveal that the effect of the economic support index on a city’s nighttime lights is not significantly affected by either its population or area. And this is the case also for both the high- and lower- income country subsamples of cities. 19 This conceptual framework builds on that used by Roberts (2021). 20 We limit the cross-sectional analysis in this subsection to the period before the development of effective vaccines in December 2020 to exclude the impacts of revised expectations of recovery. For example, economic activity may have reacted on beliefs that richer regions would obtain the vaccine first, thus influencing recovery. This is consistent with the sharp uptick in lights for cities in high-income countries seen in panel A of figure 7. 14 However, population density is also correlated with many other city characteristics, which may also explain the economic impacts of the pandemic. As shown in Figure 6, a city’s population density is significantly positively correlated with both its total GDP and its level of infrastructure development, as proxied by the length of its road network.21 A weak partial correlation is found with a measure of a city’s compactness, the PPR. However, as evidenced by the t-statistic, this relationship is not significant at conventional levels. Meanwhile, a city’s population density is significantly negatively correlated with its built-up area. Given these relationships, we move on to a regression framework to estimate the relationship between a city’s population density and the (I, D, C, and TLL) impacts of the Covid-19 crisis, while controlling for other (observable) city characteristics. Hence, we regress each of I, D, C, and TLL on a city’s population and area, while also controlling for a city’s economic scale (GDP in 2015), built-up area, the total length of its road network, and its compactness as measured by the PPR. We also include country fixed effects in our regressions to control for unobservable differences between cities that are driven by cross-country differences. This implies that the estimated relationship between the (I, D, C, and TLL) impacts of the crisis and population density is identified based on the variation between cities within countries. More specifically, to estimate the cross-sectional relationship between city trajectories and city-level characteristics, we estimate the following equation: = + + [3] Where, for city in country , is one of the following loss-of-light measures (as described above): initial (I), depth (D), current (C), and TLL. All terms are represented as relative to average of October- November-December 2019. The average of the last three months (i.e., of October – December) of 2019 as opposed to, say, December 2019, is used as the baseline against which I, D, C and TLL are measured to minimize the impacts of any temporary measurement error in the baseline level of nighttime lights on our results. is a vector of city-specific characteristics, which include population in 2015 (ln), city area in square kilometers (ln), city GDP in 2015 (ln), built-up area in square kilometers (ln), length of roads in kilometers (ln), and the PPR compactness measure (ln). is the country-specific fixed effect. The results, which are reported in full in Table 2, suggest that a higher population and a smaller area (i.e., a higher population density) are associated with smaller I, D and TLL, but not C, impacts of the Covid-19 crisis. As before, this result appears to be driven by cities in lower-income countries, as can be seen from comparing the results from the lower-income country sub-sample with those from the high-income country sub-sample in Table 2. 21The scatterplots shown in figure 6 are partial scatterplots which control both for other city characteristics and country fixed effects. 15 Figure 5. Correlation between population density and Covid impacts on city trajectories. Note: Higher population density is correlated with smaller impacts of the Covid-19 crisis in terms of the total loss of lights, the depth of the shock, and the city’s position in Dec 2020, but only in lower-income (i.e., developing) countries. Negative values on the y-axes represent adverse impacts (i.e., losses) Figure 6. Conditional Correlation between Population Density and other City Characteristics Note: Plot for each variable (GDP: top left, Built-up area: top right, road length: bottom left, compactness: bottom right) is obtained after controlling for other city-level characteristics, including country fixed effects 16 Table 2. Cross-Sectional Regression of Trajectories with City Characteristics Lower-income High-income All (Low- and Middle-Income) (1) (2) (3) (4) ( 5) (6) (7) (8) (9) (10) (11) (12) Initial Position TLL Initial Position TLL Initial Position TLL Change Depth Dec ‘20 in 2020 Change Depth Dec ‘20 in 2020 Change Depth Dec ‘20 in 2020 lpop 4.353*** 5.229*** -0.261 51.50*** 4.257** 5.506*** -0.607 50.73*** -1.856 3.221 5.838 46.15 (1.663) (1.616) (1.688) (18.23) (1.716) (1.676) (1.755) (18.83) (3.443) (5.195) (4.580) (30.23) larea -8.328*** -4.704** -1.973 -61.97** -8.494*** -4.900** -2.020 -65.08** 4.765 11.93 6.121 91.25 (2.382) (2.242) (2.654) (27.66) (2.466) (2.327) (2.754) (28.81) (5.988) (8.189) (7.090) (58.23) lgdp -0.452 -1.305 -2.545* -20.41 -0.474 -1.135 -2.511 -19.12 1.696 -6.019* -3.845 -53.41*** (1.046) (1.127) (1.535) (16.05) (1.067) (1.148) (1.569) (16.40) (2.163) (3.181) (3.597) (19.25) lbuiltup 2.977** 1.118 -0.530 8.807 3.170** 1.056 -0.531 9.293 -7.059** -0.996 0.312 -4.772 (1.422) (1.222) (1.381) (16.37) (1.437) (1.238) (1.401) (16.58) (3.509) (5.259) (4.548) (41.22) lroad -0.122 0.430 3.955** 12.66 -0.245 0.569 4.242** 14.05 3.243 -8.160 -10.23** -76.82** (1.721) (1.217) (1.646) (16.07) (1.750) (1.233) (1.674) (16.32) (3.422) (5.010) (4.593) (35.87) lppr 2.289 3.205 1.664 48.07* 2.629 3.827 2.012 50.70 1.835 0.506 0.193 44.43* (2.077) (2.323) (2.682) (27.13) (2.369) (2.629) (3.060) (31.10) (2.633) (3.491) (3.510) (24.17) N 2710 2733 2701 2753 2330 2353 2321 2371 380 380 380 382 R-sq 0.388 0.355 0.292 0.415 0.389 0.340 0.288 0.416 0.304 0.409 0.406 0.368 Country FE Y Y Y Y Y Y Y Y Y Y Y Y Note: All columns control for country FE. Robust standard error in parentheses. * p<0.1, ** p<0.05, *** p<0.01 17 3.4 Event-study analysis: Do high-density cities fare better? In our final analysis, we deploy a difference-in-difference approach to estimate whether the impacts of the Covid-19 crisis differed between high- and low-density cities, based on a comparison of their differences before and after the onset of the crisis. First, we define a city as high-density if it has a population density that was in the top quartile of city population densities within the country in which it is located, while all remaining cities are defined as low-density.22 A comparison of the TLL for these two groups of cities shows that low-density cities have experienced larger TLL, even though the aggregate GDP generated by the two groups in 2015 was roughly the same (Figure 7). Second, we interact the density classification with a dummy that equals 1 for all months from March 2020 onwards (i.e., the post-Covid-19 onset period) and 0 for all prior months (i.e., the pre-Covid-19 onset period), noting that the WHO declared Covid-19 to be a global pandemic on March 11, 2020. The following equation is then estimated: ln ( ) = + ( × ) + + + + + ℎ + [4] where: ( ) is the (natural) log-transformed nighttime sum of lights level; represents the post-Covid-19 dummy, which is equal to 1 if time is March 2020 onwards (and zero otherwise); is a dummy for high-density cities (top quartile within its country); is a measure of Covid-19 deaths (natural log-transformed); and are policies in the form of NPIs and economic support measures. Following the same notation as in equations 1 and 2, the specification also includes city fixed effects (to control for time-invariant city characteristics) and allows for differential trends of lights across countries as well as country-specific month-of-year effects to control for potential seasonal effects (possibly associated with atmospheric and other distortions of the lights data, including any predictable monthly variation in cloud coverage at the national level). In robustness checks, country-specific post-Covid-19 dummies are introduced to allow for national impacts of Covid-19 to vary by country (leaving between-year variation within each country after the onset of Covid-19). The results are reported in Table 3 for the full global sample of cities, as well as for cities in just lower- income countries and cities in just high-income countries. The results suggest that while, on average, the Covid-19 crisis has negatively impacted lights in cities, the effect has been smaller for high- density cities (see results in columns 1 and 2).23 This result is robust to using country-specific post- Covid-19 dummies (column 3). Comparing the results in columns 4 and 5 with those in columns 7 and 8, respectively, this time, the estimated coefficients are similar across cities in high- and lower- income countries. They are not, however, statistically significant for cities in high-income countries, likely due, at least in part, to the smaller sample. The estimates suggest that after the onset of the pandemic, the loss in light intensity in high-density cities was around 2 percentage points less than for lower density cities. For comparison, the estimates in column 4 suggest that less dense cities in lower-income countries experienced a loss of 10 percentage points in their nighttime lights intensity following the emergence of Covid-19. Lastly, some differences in the estimated coefficient on the post-Covid-19 dummy are observed for the lower-income and high-income country subsamples. For the lower-income country subsample, the inclusion of Covid-19-related deaths and NPIs in column 5 absorbs the negative effect of the post- Covid-19 dummy in column 4. By contrast, for cities in high-income countries, the magnitude and 22 Hence, we define a city as being high density if its population is dense relative to other cities in the same country. The classification of cities into high- and low-density is based on their 2015 populations. 23 Some care of interpretation is required here, however, given that the estimated coefficient on the interaction dummy (i.e., the post-Covid-19 onset dummy  high-density dummy) is only significant at the 10 percent level. 18 significance of this estimated coefficient on the post-COVID-19 dummy increase with the inclusion of these variables as we move from columns 7 to 8. 19 Table 3. Event-Study Analysis using Diff-in-Diff Setup All Lower-income countries High-Income countries (1) (2) (3) (4) (5) (6) (7) (8) (9) ln_SOL ln_SOL ln_SOL ln_SOL ln_SOL ln_SOL ln_SOL ln_SOL ln_SOL Post x High Density 0.0206* 0.0213* 0.0199* 0.0218* 0.0229* 0.0203 0.0180 0.0120 0.0169 (0.0116) (0.0116) (0.0114) (0.0129) (0.0130) (0.0127) (0.0149) (0.0149) (0.0149) Post-Covid Dummy -0.0895*** -0.0187* -0.10*** 0.00113 -0.0161* -0.0615*** (0.00517) (0.0103) (0.00574) (0.0118) (0.00899) (0.0154) Ln(death) -0.00756*** -0.0152*** -0.0113*** -0.0175*** 0.0183*** 0.0167*** (0.00116) (0.00117) (0.00129) (0.00128) (0.00241) (0.00395) NPI Stringency -0.00035*** 0.000103 -0.000198* 0.000156 -0.00081** -0.00070** (0.000101) (0.0000992) (0.000106) (0.000105) (0.000317) (0.000278) Econ support 0.000289** 0.000291** 0.0000706 0.000216 -0.00106*** -0.0012*** (0.000137) (0.000138) (0.000153) (0.000156) (0.000363) (0.000412) N 198823 198582 198582 174302 174061 174061 24521 24521 24521 Cities 2841 2841 2841 2435 2435 2435 406 406 406 r2 0.963 0.962 0.963 0.956 0.956 0.956 0.937 0.937 0.937 r2_within 0.00326 0.00377 0.000757 0.00383 0.00462 0.000917 0.000200 0.00240 0.000762 City FE Y Y Y Y Y Y Y Y Y Country x Y Y Y Y Y Y Y Y Y Month-of-year FE Country-specific Y Y Y Y Y Y Y Y Y linear time trend Country--specific N N Y N N Y N N Y Covid dummy Note: All columns control for city FE, country-specific time trend and country-month-of-year FE. Columns 3, 6 and 9 include a country-specific dummy indicating the months after the first Covid case was detected. High-density cities are those that have population densities above the 75 th percentile within a country. Standard error in parentheses clustered at the city-level. * p<0.1, ** p<0.05, *** p<0.01 20 Figure 7. Low population-density cities experienced worse loss of lights in aggregate 4. Conclusion This paper has documented the impact of the spread of Covid-19 and the imposition of NPIs on economic activity in cities, as proxied by monthly nighttime lights. Combining this data allows us to examine the short-term impacts of the Covid-19 crisis on economic activity across a global sample of 2,830 cities drawn from 162 countries. Additionally, it focused on the first year of the pandemic, i.e., the period before the mass roll-out of Covid-19 vaccines across the globe. The findings suggest that the early months of the pandemic gave rise to a divergence in (urban) economic growth trajectories across cities globally. In general, the negative economic impacts of Covid-19 were found to be starker in cities in lower-income (i.e., low- and middle-income) countries and in denser cities. This can have important implications for long-run growth trajectories as well, since the effects of temporary shocks to urban systems can have effects that potentially persist over long periods of time (Hanlon, 2017). Building on this, we also investigated the correlation between city-level characteristics and the differential economic impacts of the Covid-19 crisis across cities globally, with a special focus on the role of population density. Our findings suggest that density plays an important role in determining city resilience, especially in lower-income countries, to a global shock. Even though denser cities have higher health risks that impair economic activity, on net, the positive effects of density in terms of bolstering resilience offset these adverse effects, suggesting a net benefit from density when it comes to withstanding the economic impacts of the Covid-19 pandemic. Finally, we reiterate the limitations of our analysis. Our results cannot be interpreted causally because of the lack of sufficient data on Covid-19-related deaths and NPIs at the subnational level, as well as potential reverse causality arising from increases in economic activity leading to faster spread of the disease and implementation of lockdowns and economic support measures. This is a trade-off of maintaining the broad geographic scope of this study. Future research can build on this approach by incorporating sufficient subnational data on Covid-19-related variables and identifying plausibly exogenous sources of variation in Covid-19-related variables at the city level. Nonetheless, the spatial and temporal granularity of the data used in the analyses here and the rich specifications they allow for help estimate refined correlations and elasticities that are informative for policy design, as they narrow down the factors correlated with city resilience. 21 References Alfani, G., 2022. Epidemics, inequality, and poverty in preindustrial and early industrial times. Journal of Economic Literature, 60(1), pp.3-40. Almond, D., 2006. Is the 1918 influenza pandemic over? Long-term effects of in utero influenza exposure in the post-1940 US population. Journal of political Economy, 114(4), pp.672-712. Barro, R.J., 2020. Non-pharmaceutical interventions and mortality in US cities during the great Influenza pandemic, 1918-1919 (No. w27049). national bureau of economic Research. Bisin, A. and Moro, A., 2022. JUE insight: Learning epidemiology by doing: The empirical implications of a Spatial-SIR model with behavioral responses. Journal of Urban Economics, 127, p.103368. Bhardwaj, G., Esch, T., Lall, S.V., Marconcini, M., Soppelsa, M.E. and Wahba, S., 2020. Cities, Crowding, and the Coronavirus. Bleakley, Hoyt, and Jeffrey Lin. "Portage and path dependence." The quarterly journal of economics 127, no. 2 (2012): 587-644. Chetty, R., Friedman, J.N., Hendren, N. and Stepner, M., 2020. Real-time economics: A new platform to track the impacts of Covid-19 on people, businesses, and communities using private sector data. NBER Working Paper, 27431. Correia, Sergio, Stephen Luck and Emil Verner, 2020. “Pandemics Depress the Economy, Public Health Interventions Do Not: Evidence from the 1918 Flu,” working paper, 10 April Davis, D.R. and Weinstein, D.E., 2002. Bones, bombs, and break points: the geography of economic activity. American economic review, 92(5), pp.1269-1289. Demirguc-Kunt, Asli; Lokshin, Michael; Torre, Ivan. 2020. The Sooner, the Better : The Early Economic Impact of Non-Pharmaceutical Interventions during the Covid-19 Pandemic. Policy Research Working Paper;No. 9257. World Bank, Washington, DC. © World Bank. https://openknowledge.worldbank.org/handle/10986/33820 License: CC BY 3.0 IGO. Dell, M. and Olken, B.A., 2020. The development effects of the extractive colonial economy: The dutch cultivation system in java. The Review of Economic Studies, 87(1), pp.164-203. Duranton, G. and Puga, D., 2004. Micro-foundations of urban agglomeration economies. In Handbook of regional and urban economics (Vol. 4, pp. 2063-2117). Elsevier. Duranton, G. and Puga, D., 2014. The growth of cities. Handbook of economic growth, 2, pp.781-853. Duranton, G. and Puga, D., 2020. The economics of urban density. Journal of Economic Perspectives, 34(3), pp.3-26. Gibson, J., Olivia, S., Boe-Gibson, G. and Li, C., 2021. Which night lights data should we use in economics, and where?. Journal of Development Economics, 149, p.102602. Hale, T., Angrist, N., Kira, B., Petherick, A., Phillips, T. and Webster, S., 2020. Variation in government responses to Covid-19. 22 Hanlon, W.W., 2017. Temporary shocks and persistent effects in urban economies: Evidence from British cities after the US Civil War. Review of Economics and Statistics, 99(1), pp.67-79. Henderson, J.V., Squires, T., Storeygard, A. and Weil, D., 2018. The global distribution of economic activity: nature, history, and the role of trade. The Quarterly Journal of Economics, 133(1), pp.357- 406. Henderson, J. V., V. Liu, C. Peng, and A. Storeygard. 2019. Demographic and health outcomes by degree of urbanization. Perspectives from a new classification of urban areas. Brussels: European Commission. Jordà, Ò., Singh, S.R. and Taylor, A.M., 2022. Longer-run economic consequences of pandemics. Review of Economics and Statistics, 104(1), pp.166-175. Jedwab, R., Khan, A.M., Russ, J. and Zaveri, E.D., 2021. Epidemics, pandemics, and social conflict: Lessons from the past and possible scenarios for Covid-19. World Development, 147, p.105629. Jedwab, R., Johnson, N.D. and Koyama, M., 2022. The economic impact of the Black Death. Journal of Economic Literature, 60(1), pp.132-78. Jedwab, R., Loungani, P. and Yezer, A., 2021. Comparing cities in developed and developing countries: Population, land area, building height and crowding. Regional Science and Urban Economics, 86, p.103609. Jedwab, R., D. Pereira and M. Roberts., 2021. Cities of workers, children or seniors? Stylized facts and possible implications for growth in a global sample of cities. Regional Science and Urban Economics, 87. Lall, S.V. and Wahba, S., 2021. Crowded cities: New methodology in Covid-19 risk assessment. Sustainability, 13. Masaki, T., Nakamura, S. and Newhouse, D., 2020. How is the Covid-19 crisis affecting Nitrogen Dioxide emissions in Sub-Saharan Africa? Miguel, E. and Mobarak, A.M., 2021. The economics of the COVID-19 pandemic in poor countries. Annual Review of Economics, 14. Ramani, A. and Bloom, N., 2021. The Donut effect of Covid-19 on cities (No. w28876). National Bureau of Economic Research. Roberts, M., 2021. Tracking economic activity in response to the Covid-19 crisis using nighttime lights–The case of Morocco. Development Engineering, 6, p.100067. Small, C., 2019. Multisensor characterization of urban morphology and network structure. Remote Sensing, 11 (18), 2162. Woloszko, N., 2020. A Weekly Tracker of activity based on machine learning and Google Trends. OECD Economics Department Working Papers 1634. World Bank. 2009. Reshaping Economic Geography. Washington, D.C.: The World Bank. World Bank. Forthcoming. From Suriving to Thriving: Making Cities Green, Resilient, and Inclusive. Washington, D.C.: The World Bank. 23 Appendix Table A1. Summary Statistics of Impact Measures and Core City Characteristics All Cities Lower-income Countries High Income Countries (N = 2841) (N = 2435) (N = 406) Mean Median SD Mean Median SD Mean Median SD Initial Impact -0.03 -0.39 45.92 0.54 0.15 49.21 -3.44 -2.66 15.45 Depth -43.38 -32.32 37.50 -45.75 -35.14 39.00 -29.23 -22.61 22.06 Position in December 2020 -0.48 -2.94 50.11 -0.51 -3.27 53.75 -0.31 -1.65 15.98 Total Loss of Lights (TLL) 6.24 0.98 446.40 2.87 0.00 478.78 26.45 18.83 138.59 Population, log of 13.07 12.81 0.86 13.04 12.79 0.84 13.29 13.00 0.98 Area, log of 4.29 4.27 1.12 4.13 4.12 1.07 5.30 5.18 0.89 GDP, log of 21.64 21.61 1.60 21.35 21.38 1.48 23.36 23.07 1.18 Built-up area, log of 3.06 3.26 1.69 2.79 3.05 1.63 4.72 4.60 0.98 Length of road (km), log of 12.10 12.10 1.20 11.92 11.91 1.13 13.27 13.11 0.95 Compactness measure (PPR), log of -1.17 -1.12 0.42 -1.13 -1.08 0.41 -1.40 -1.34 0.42 24 Table A2.A: Covid Deaths, NPIs and Interaction with Pop Density – Weighted using Cloud-free Percentage. Lower-income ALL (Low- and Middle-Income) High-income (1) (2) (3) (4) (5) (6) ln_sol ln_sol ln_sol ln_sol ln_sol ln_sol (log) Number of Death -0.00334*** 0.0575*** -0.00493*** 0.0541*** 0.0120*** -0.0116 (0.000622) (0.0105) (0.000662) (0.0121) (0.00134) (0.0239) (log) Number of Death * (log) Population -0.00653*** -0.00597*** 0.00315 (0.000964) (0.00113) (0.00279) (log) Number of Death * (log) Area 0.00598*** 0.00477*** -0.00332 (0.000765) (0.000918) (0.00289) Stringency Index -0.000415*** -0.00178 -0.000282*** -0.00223 -0.000757*** 0.00148 (0.0000784) (0.00155) (0.0000832) (0.00170) (0.000157) (0.00308) Stringency Index * (log) Population 0.000316** 0.000336* -0.000147 (0.000159) (0.000174) (0.000349) Stringency Index * (log) Area -0.000657*** -0.000598*** -0.0000563 (0.000156) (0.000170) (0.000376) Economic Support Index 0.000124* -0.000108 0.0000138 0.00125 -0.000829*** 0.000583 (0.0000699) (0.00153) (0.0000778) (0.00184) (0.000148) (0.00290) Economic Support Index * (log) Population -0.0000843 -0.000189 -0.000310 (0.000164) (0.000196) (0.000347) Economic Support Index * (log) Area 0.000278 0.000278 0.000505 (0.000169) (0.000204) (0.000369) N 195046 195046 170799 170799 24247 24247 Cities 2841 2841 2435 2435 406 406 r2 0.986 0.986 0.984 0.984 0.982 0.982 r2_within 0.153 0.155 0.160 0.161 0.0448 0.0450 City FE Y Y Y Y Y Y Country x Month-of-year FE Y Y Y Y Y Y Country-specific linear time trend Y Y Y Y Y Y * p<0.05, ** p<0.01, *** p<0.001 Note: Standard errors in parentheses clustered at the city-level. Controls for city FE, city-specific time trend and country-month FE and Covid-19 first month. 25 Table A2.B: Covid Deaths, NPIs and Interaction with Pop Density Subnational-Level Regression on Five Countries (Brazil, Canada, China, UK and USA) (1) ln_sol (log) Number of Death -0.00259 (0.00185) Stringency Index -0.00184*** (0.000229) Economic Support Index 0.00126*** (0.000269) N 42860 Cities 676 r2 0.942 r2_within 0.0895 City FE Y State x Month-of-year FE Y State-specific linear time trend Y Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 Controls for city FE, state-specific time trend and state-month FE and Covid- 19 first month. 26 Table A2.C: Covid Death, NPIs – First Differences Lower-income ALL (Low- and Middle-Income) High-Income (1) (2) (3) Δ ln_sol Δ ln_sol Δ ln_sol Δ (log) Number of Death -0.0131*** -0.0191*** 0.0172*** (0.00211) (0.00233) (0.00477) Δ Stringency Index 0.000172 0.000336* -0.000618 (0.000158) (0.000172) (0.000395) Δ Economic Support Index 0.000357 0.000234 -0.000486 (0.000229) (0.000253) (0.000386) N 192911 170246 22665 Cities 2841 2435 406 r2 0.0734 0.0690 0.140 r2_within 0.000491 0.000496 0.00227 City FE Y Y Y Country x Month-of-year FE Y Y Y Country-specific linear time trend Y Y Y Note: Standard errors in parentheses clustered at the city-level. Controls for city FE, city-specific time trend and country-month FE 27 Table A2.D: Covid Cases, NPIs and Interaction with Pop Density – Using Cases instead of Deaths to measure disease spread Lower-income ALL (Low- and Middle-Income) High-Income (1) (2) (3) (4) (5) (6) ln_sol ln_sol ln_sol ln_sol ln_sol ln_sol (log) Number of Cases -0.00839*** 0.0489*** -0.00953*** 0.0434** 0.00255 -0.00921 (0.000751) (0.0173) (0.000795) (0.0193) (0.00165) (0.0298) (log) Number of Cases * (log) Population -0.00500*** -0.00413** 0.00125 (0.00151) (0.00169) (0.00322) (log) Number of Cases * (log) Area 0.00195** 0.000234 -0.00144 (0.000863) (0.00102) (0.00298) Stringency Index -0.0000325 -0.00315 0.000111 -0.00285 0.000153 0.0165** (0.000114) (0.00227) (0.000119) (0.00242) (0.000364) (0.00761) Stringency Index * (log) Population 0.000414* 0.000338 -0.00217** (0.000212) (0.000227) (0.000861) Stringency Index * (log) Area -0.000538*** -0.000346* 0.00243*** (0.000173) (0.000187) (0.000926) Economic Support Index 0.000273** -0.00208 0.0000917 -0.00273 -0.000798** -0.0120 (0.000119) (0.00281) (0.000128) (0.00312) (0.000356) (0.00839) Economic Support Index * (log) Population 0.0000295 0.000135 0.00154 (0.000271) (0.000303) (0.000965) Economic Support Index * (log) Area 0.000438** 0.000245 -0.00175* (0.000219) (0.000248) (0.000997) N 198582 198582 174061 174061 24521 24521 Cities 2841 2841 2435 2435 406 406 r2 0.962 0.963 0.956 0.956 0.937 0.937 r2_within 0.0558 0.0563 0.0594 0.0600 0.0101 0.0108 City FE Y Y Y Y Y Y Country x Month-of-year FE Y Y Y Y Y Y Country-specific linear time trend Y Y Y Y Y Y * p<0.05, ** p<0.01, *** p<0.001 Note: Standard error in parentheses clustered at the city-level. Controls for city FE, city-specific time trend and country-month FE. 28