Policy Research Working Paper 9242 Determinants of Social Distancing and Economic Activity during COVID-19 A Global View William Maloney Temel Taskin Equitable Growth, Finance and Institutions Practice Group Office of the Chief Economist May 2020 Policy Research Working Paper 9242 Abstract The paper uses Google mobility data to identify the deter- than repressive measures per se. These results are consis- minants of social distancing during the 2020 COVID-19 tent across country income groups, with only the poorest outbreak. The findings for the United States indicate that countries showing limited effect of non-pharmaceutical much of the decrease in mobility is voluntary, driven by interventions and no voluntary component, consistent with the number of COVID-19 cases and proxying for greater resistance to abandon sources of livelihood. The paper also awareness of risk. Non-pharmaceutical interventions such confirms the direct impact of the voluntary component on as closing nonessential businesses, sheltering in place, and economic activity, by showing that the majority of the fall school closings are also effective, although with a total in restaurant reservations in the United States and movie contribution dwarfed by the voluntary actions. This sug- spending in Sweden occurred before the imposition of any gests that much social distancing will happen regardless non-pharmaceutical interventions. Widespread voluntary of the presence of non-pharmaceutical interventions and de-mobilization implies that releasing constraints may not that restrictions may often function more like a coordi- yield a V-shaped recovery if the reduction in COVID risk nating device among increasingly predisposed individuals is not credible. This paper is a product of the Office of the Chief Economist, Equitable Growth, Finance and Institutions Practice Group. 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 wmaloney@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 Determinants of Social Distancing and Economic Activity during COVID-19: A Global View 1 William Maloney 2 and Temel Taskin 3 JEL Classification: F00; I15; I18; O10. Keywords: COVID-19, Pandemic, Epidemic, Social distancing, Non-pharmaceutical interventions, Mobility, Economic crises. 1 The opinions are those of the authors and do not represent the official position of the World Bank. Our thanks to Richard Baldwin, Robert Beyer, Nick Bloom, Xavi Cirera, Tito Cordella, Aart Kraay, Pravin Krishna, Norman Loayza, Hideaki Matsuoka, Cedric Okou, Leonardo Pio Perez, Cevdet Cagdas Unal, and Shu Yu for excellent comments. 2 Chief Economist, Equitable Growth, Finance, and Institutions, The World Bank. Corresponding author, wmaloney@worldbank.org 3 Economist, Prospects Group, Equitable Growth, Finance, and Institutions, The World Bank. I. Introduction Understanding the determinants of social distancing is central to addressing both the medical and economic aspects of COVID-19. 4 On the one hand, reducing interactions among people is critical to reducing the propagation and a variety of non-pharmaceutical interventions (NPIs), such as closure of nonessential businesses, stay-at-home orders, or school closings, have been put in place to this end, with some success. 5 While there is controversy around whether this should be the goal in developing countries as well (Barnett- Howell and Mobarak 2020, Loayza 2020), there is also concern about whether such measures would work: government capabilities to enforce may be weaker, and resistance may be higher since the trade-off with livelihood is harsher. At the other extreme of the cycle - where the debate is when to loosen NPIs as it is in several advanced countries – preliminary evidence from Wuhan suggests that when opened, mobility and economic activity may not respond quickly. 6 Similarly, recent polls suggesting that 58% of Americans are concerned that restrictions will be lifted too soon raise the question of how much of an impact opening will have in practice and hence the shape of the recovery, whether V or U. 7 This paper uses Google mobility data to explore which factors are proving important during the 2020 COVID- 19 outbreak in the United States and globally. In all but the poorest countries, it confirms that NPIs can be effective, but that voluntary de-mobilization on the part of the population is much more important,-driven by fear or perhaps a sense of social responsibility. This suggests that much social distancing will happen regardless of the presence of restrictions and that NPIs may often function more like a coordinating device among increasingly predisposed individuals than repressive measures per se. We also confirm a more direct link of this voluntary effect using data on restaurant reservations in the United States and movie releases and revenues in Sweden and show that these, too, experience most of their fall before any imposition of NPIs. Overall, the 4 There are three margins upon which societies can work to reduce the death toll. 1. Detect and quarantine so the disease never gets a foothold. 2. Once established, reduce social mobility to mitigate the spread (reduce the R factor). 3. Increase the capability to treat the sick. On the third, Favero (2020) notes that limitations on ICU beds led to the extremely high death rate in Lombardy. In practice, developing countries have far less capability to treat- 10 African countries have no respirators.https://www.nytimes.com/2020/04/18/world/africa/africa-coronavirus- ventilators.html?referringSource=articleShare If northern Italy could not ramp up sufficiently enough along this dimension, it is highly unlikely that most poor countries can. On the first, many advanced countries have missed the window to detect and quarantine and again, this may be more challenging in the developing world. 5 See Chen and Qiu (2020), Gonzalez-Eira and Niepelt (2020) for conceptual treatments of optimal shutdown policies. Hartl et al (2020) find for Germany that growth rates of Covid-19 cases fell 50% as a result of German restrictions to shut down schools, stadiums and eventually many restaurants and shops. See Baldwin and Weder de Mauro (2020) for a compilation of recent thinking on Covid Economics. 6 https://www.bloomberg.com/news/articles/2020-04-15/wuhan-s-life-after-lockdown-isn-t-business-as-usual? 7 The NBC News-Wall Street Journal poll was conducted between April 13 and April 15 among a sample of 900 registered voters. 2 evidence suggests that moves to unfreeze the economy will fail unless there is confidence that, in fact, the risk has passed. Several recent papers suggest that NPIs have had an impact in the United States. Engle at al (2020) use daily average changes in distance traveled in every U.S. county as a proxy for reduction in exposure to COVID-19 and find that an official stay-at-home restriction order corresponds to reducing mobility by 7.87%. Brzezinski et al (2020), also using cell phone data, find that a lockdown increases the percentage of people who stay at home by 8% across U.S. counties. Painter and Qiu (2020) show that the introduction of shelter-in-place policies is associated with a 5.1 percentage point increase in the probability of staying home (see also Andersen (2020)). However, voluntary de-mobilizing behavior that intensifies with prevalence of the disease is also an important driver and affects the effectiveness of official measures. Auld (2006), for example, finds that during the AIDS epidemic, an average respondent decreased risky behavior by about 5% in response to a 10% increase in AIDS prevalence. Further, the 1918 Spanish Flu epidemic suggests that the predisposition of the population to demobilize drove both the incidence of official restrictions and their effectiveness. On the one hand, as Crosby (2003) details, that restrictions were binding is revealed by the fact that in San Francisco “The places of amusement opened first, to huge crowds starved for entertainment (p. 99)” and in Philadelphia “The long thirst was over, and arrests on drunken and disorderly charges bounded back up to and beyond normal levels” (p. 85). However, it is also true that while the San Francisco Department of Health could request that people to smother coughs and sneezes, only when enough fatalities were registered were “San Franciscans…scared enough to accept drastic measures to control the epidemic” (p.95)—and ex post, “Fear had been the enforcer of the Board of Health’s policies.”(p. 108) not the authorities themselves. When schools in San Francisco were opened, many parents kept their children home out of continuing fear. This resonates with the reports from Wuhan today of the anemic rebounding of the small restaurant sector when restrictions were released. Viewed through this lens, restrictions may often function more like a coordinating device among increasingly predisposed individuals than a repressive measure- if we are all working from home, then I will not be viewed badly if I do; whether schools are online or in person requires a decision that individual concerned parents cannot effect. This, in turn, raises the question of the whether the impact of lock-down measures per se and their subsequent removal is overstated. II. Data Mobility and Economic Activity: Using data from the Maps application on smartphones, Google generates COVID-19 Community Mobility Reports 8 that use aggregated, anonymized data to construct an index of how visits and length of stay at different places change compared to a baseline. They can then follow 8 https://www.google.com/covid19/mobility/ 3 movement trends over time by geography, and across different high-level categories of places such as workplaces, retail and recreation, groceries and pharmacies, parks, transit stations, and residential. These measures are explicitly considered proxies for social distancing and we focus on the first, workplace-related mobility, as the most relevant to economic activity and most prominent in the policy debate. The reports consist of per country downloads (with 131 countries covered initially), further broken down into regions/counties in some cases. Because location accuracy and the understanding of categorized places varies from region to region, Google does not recommend using these data to compare changes across countries or regions with different characteristics. To address this, our empirics rely only on within-area variation across time and reporting or categorization differences are absorbed in included fixed effects. This measure is limited by the degree to which coverage of smart phones offers a representative sample of the population. As Annex 1 shows, few developing countries show coverage of smart phones above 50% and Ethiopia, Nigeria, Sudan, Bangladesh, and Pakistan hold up the bottom of the top 50 countries with rates under 20% of coverage. This said, several developing countries also have reasonable coverage when we adjust for the share of adults in the population: the United Kingdom: 100%, Sweden: 96%, the United States: 95%, Italy 67%, Japan 63%, Brazil 52%, and South Africa 50%. While clearly not representative, the differences between Italy and Japan on the one hand and Brazil and South Africa on the other are not so large as to justify throwing out the possible information on how developing countries may differ. Further, while we may miss the mobility of for instance, micro firm owners without smartphones, many of their customers will have them and the shutting down of the firm will be partially registered. Data on restaurant reservations in the United States are taken from OpenTable. 9 Movie release and theater revenue data for Sweden are from International Movie Database. 10 COVID-19 Cases: Though there may be several mechanisms through which cases translate into lower mobility, we interpret this as a signal to individuals about the likelihood of a serious negative health outcome. National cases can inform about the overall evolution of the disease, while local numbers finetune the proximate threat. We standardize by the corresponding population in the figures. In some regressions, we can expand the sample by using log (cases) and the population scaling is absorbed in the corresponding fixed effect. Global data are drawn from the Johns Hopkins Coronavirus Resource Center. Country-specific regional data come from national sources: the United States: Johns Hopkins Coronavirus Resource Center; Brazil, Italy, Japan, South Africa, Sweden, and the United Kingdom: from national sources (see Annex II). 9 www.opentable.com 10 https://www.boxofficemojo.com/ 4 Non-Pharmaceutical Interventions (NPIs): We use mandatory closures of nonessential businesses as both most relevant to the issue of economic mobility and figuring most prominently in the policy debate. State-level data for the United States are collected from Raifman et al (2020) and NPIs enter as indicator variables taking a value of 1 if a given NPI is implemented and 0 otherwise. Globally, we employ information on national NPIs available from the Blavatnik School of Government at Oxford University. For select countries for which we employ subnational mobility data to explore the impact of local case incidence, we use national data on the nationally implemented NPIs as controls. The exception is Brazil for which NPIs are established by states, and we collect data at that level. III. Results: United States Figure 1 plots the level of mobility against the log of the number of cases per capita by U.S. state for the United States. It further divides the sample by whether the states are covered by restrictions on nonessential businesses (red) or not (blue). Two drivers appear as potentially important. First, the data are consistent with restrictions leading to lower levels of mobility. However, more strikingly, there is a clear downward sloping relationship between reported cases and mobility independent of such restrictions. Figure 1: Mobility, COVID Cases and Official Restrictions, United States US Workplace mobility (% deviation from baseline) -60 -40 -20 0 20 -5 0 5 10 Log(case per 1 million people) No shutdown Shutdown Notes: Workplace mobility is Google measure of work-related mobility index. See text for sources. Table 1 more formally tests this relationship by estimating = 0+ 1 −1 + 2 −1 + 3 −1 + + + (1) Where mobility is the Google measure, Cases is the log incidence, Aggregate Cases is the national analogue, NPI are non-pharmaceutical intervention(s), and ui are subnational (state) fixed effects that also effectively put cases in per capita terms, and vt, time fixed effects. There are clear issues of bi-directional causality here. 5 Lower mobility, in theory, lowers the number of cases and may also possibly affect the likelihood of imposing restrictions. This should induce a downward bias to both coefficients on the right-hand side and our results should be taken as a lower bound. As we are working with a larger group of countries, we do not attempt to instrument which would not be feasible in most, but we lag both explanatory variables 1 period. The results change modestly in magnitude, with even more lags, but the overall patterns remain consistent. Table 1: Mobility, COVID Cases and NPIs, United States (1) (2) (3) (4) (5) (6) Workplace Workplace Workplace Workplace Residential Residential Close N.E. business -4.373*** -5.281*** -2.071 -3.075*** 2.047*** 0.830* (1.235) (0.689) (2.006) (1.051) (0.356) (0.463) Log cases -4.502*** -1.291*** -2.904*** -1.284*** 0.551*** 0.577*** (1.153) (0.437) (0.915) (0.385) (0.185) (0.161) Log national cases -2.671** -3.038*** -2.193** -2.837*** 0.957*** 0.875*** (1.063) (0.425) (0.860) (0.383) (0.225) (0.177) Close K-12 -11.975*** -0.866 -0.092 (1.704) (1.169) (0.407) Stay home/SIP -3.289 -3.855*** 2.144*** (2.630) (1.134) (0.485) Constant 24.030*** 10.503*** 18.981*** 9.574*** -4.472*** -3.986*** (5.191) (1.756) (4.526) (1.509) (1.250) (0.968) Time FE No Yes No Yes Yes Yes Day of the week FE Yes No Yes No No No State FE Yes Yes Yes Yes Yes Yes # of States 51 51 51 51 51 51 Obs. 1189 1189 1189 1189 1189 1189 R-squared 0.836 0.963 0.875 0.964 0.956 0.959 Notes: Regression of Google measure of work/residential related mobility on NPIs, the log of cases, the log of national cases, state, days of the week/time fixed effects. Robust clustered errors are in parenthesis. *** p<0.01, ** p<0.05, *<0.1 Table 1 suggests that both effects are at work although with surprising relative contributions. Columns 1-2 present the impact on mobility of just business closure restrictions, the log of local cases and the log of national cases with and without time fixed effects. Of the roughly 60-point decline in mobility seen in Figure 1, roughly 5 points appear due to official workplace closures. This is of the order of magnitude identified in previous studies on other measures of mobility. However, the component due to case incidence, both national and local appears to be able to account for much of the fall in mobility by itself. For instance, with the 10-log point increase in local cases in Figure 1, roughly 43 points (2/3) of the fall in mobility are accounted for, and more without FE by “voluntary” self-restriction. Columns 3 and 4 introduce two other NPIs- school closures for K-12 and stay-at-home/shelter in place orders. The impact of imposing restrictions on businesses falls significantly, suggesting that, as expected, it was picking up the effects of other correlated measures. The three together can account for almost 8 points of the fall in mobility. This remains dwarfed by the roughly 40% arising from the number of local and national cases whose impact stays roughly the same. Hence, it appears that in the United States, the largest effect is due to protective 6 measures taken by individuals as they learn more about the prevalence of the disease. The question then arises, will the effect of removing those restrictions in fact lead to the hoped-for rejuvenating effect on the economy if case numbers remain high? As a confirmatory test on the complement to workplace mobility, columns 5 and 6 show that increased NPIs and case incidences lead to a rise in residential mobility. IV. Results: Global Sample Figure 2 plots the same relationship for six countries of potential interest: Italy, Japan, Sweden and the United Kingdom and two upper-middle-income countries, Brazil and South Africa, for which we have reasonable smart phone coverage. In every case, the figures show evidence of decreased mobility with the increase in case numbers. Table 2 formalizes the graphs by running subnational mobility against subnational and national COVID case incidence, including NPIs appropriate to the country case. NPIs are at the country level with the exception of Brazil where they are set at the state level and the data are therefore subnational. Three findings emerge. First, in Brazil, Italy, South Africa, Sweden, and the United Kingdom, the combined semi-elasticities of mobility with respect to case incidence are comparable to those found in the United States, ranging between 5 and 7. Hence, in Sweden, a country with notable fewer NPIs until end-March, mobility falls 60 points or almost that seen in the United States. The sharp contrast often depicted between Sweden and more interventionist countries appears overdrawn. Second, our estimates suggest that some NPIs have large effects in Italy, South Africa (some with unexpected sign however), and the United Kingdom. For Brazil, Italy, South Africa, Sweden, and the United Kingdom, however, the “voluntary” component still contributes the largest share. Third, Japan shows an unusually low semi elasticity of around 3 and poses something of a puzzle given that it is a country with an informed public, effective governance and high social capital. Press coverage points to NPIs as important coordination mechanisms. Although schools were closed and large events were cancelled since early March, business continued as normal until early April 7 when the State of Emergency (SoE) was declared. But even under the SoE, governors could only request that people stay home and that businesses close. Tokyo’s governor asked that people not go out at night but said restaurants and bars could remain open until 8 PM. These tepid measures faced strong headwinds in other social norms. For instance, there is resistance rooted in the country’s work culture where employees fear being seen as slackers if they do not appear 7 for work in person. 11 Unless everyone is sent home, everybody goes to work. The lack of a stronger coordination mechanism through official measures is a plausible explanation for both for the absence of much of an impact of formal measures, as well as limited self-motivated reductions in mobility. Figures 2a-f: Workplace Mobility vs. Cases and Closure of Non-Essential Businesses Brazil Italy Workplace mobility (% deviation from baseline) Workplace mobility (% deviation from baseline) 20 0 0 -20 -20 -40 -40 -60 -80 -60 -80 -4 -2 0 2 4 6 -2 0 2 4 6 8 Log(case per 1 million people) Log(case per 1 million people) No shutdown Shutdown No shutdown Shutdown Japan South Africa Workplace mobility (% deviation from baseline) Workplace mobility (% deviation from baseline) 20 0 20 0 -20 -20 -40 -40 -60 -60 -80 -2 0 2 4 6 -2 0 2 4 Log(case per 1 million people) Log(case per 1 million people) No shutdown Shutdown No shutdown Shutdown Sweden UK Workplace mobility (% deviation from baseline) Workplace mobility (% deviation from baseline) 20 0 20 0 -20 -20 -40 -40 -60 -60 -80 -80 -2 0 2 4 6 8 -6 -4 -2 0 Log(case per 1 million people) Log(case per 1 million people) No shutdown Shutdown No shutdown Shutdown Notes: Workplace mobility is Google measure of work-related mobility index. See Annex II for country-specific sources. 11 https://www.nytimes.com/2020/04/19/world/asia/tokyo-japan-coronavirus.html?smid=em-share 8 Table 2: Mobility, COVID Cases and NPIs, Select Countries (1) (2) (3) (4) (5) (6) Brazil Italy Japan S. Africa Sweden UK Close N.E. business 2.996 -28.781*** 3.054 -5.871** -20.337*** (2.375) (0.836) (2.190) (2.166) (0.322) K-12 closure -2.135 -13.583*** -12.670*** (1.680) (2.275) (0.462) Cancel public events -1.697 10.798*** -7.837*** (1.842) (2.150) (2.039) Close public transport. 4.102* (1.782) Public info. camp. 46.285*** 12.420*** (7.338) (1.794) Restr. on internal mov. -37.443*** (0.924) Log cases -1.413** -2.775*** 0.166 -1.294 -4.499** 0.719 (0.595) (0.865) (0.561) (1.982) (1.796) (0.517) Log national cases -3.544*** -3.157** -3.229*** -4.371** -2.601 -6.994*** (0.464) (1.134) (0.553) (1.711) (2.290) (0.566) Constant 9.550*** 22.787*** 3.909* 25.710*** 18.885* 39.349*** (1.982) (6.500) (1.976) (5.624) (9.309) (2.783) Time FE Yes No No No No No Day of the week FE Yes Yes Yes Yes Yes Yes State FE Yes Yes Yes Yes Yes Yes # of States 27 20 46 7 21 95 Obs. 762 865 2361 169 758 2566 R-squared 0.811 0.945 0.484 0.956 0.637 0.956 Notes: Regression of Google measure of work-related mobility on NPIs, the log of cases, the log of national cases. Mobility, Cases and National Cases at subnational level. NPIs at national level with the exception of Brazil for which all data is at the subnational level. Robust clustered errors are in parenthesis. *** p<0.01, ** p<0.05, *<0.1 Global Sample Figure 3a-c groups the global sample of countries which have national data on mobility and NPI. Figure 3a divides the sample into those with and without restrictive orders. As in the individual case, there appears to be evidence for both the impact of restrictions and of the relationship with cases incidence. Figure 3b breaks the data apart into four income categories, Low Income Countries (LIC), Lower Middle Income (LMIC), Upper Middle Income (UMIC) and High Income (HIC) which include primarily the wealthier OECD countries (see Annex III for categorization). Figure 3c is the same, but only for country/periods when official restrictions on non-essential businesses are in place. In both cases, the downward slope appears across all income categories. Table 3 largely confirms previous findings. Each specification is presented with and without time fixed effects which, in some categories, consume substantial degrees of freedom. Preliminary explorations suggest that world COVID case incidence does not enter and we drop that term. This makes sense if we think that citizens of a country may pay attention to national trends, as was the case in the United States, but maybe less cases 9 across the ocean. The semi-elasticity on home case incidence appears both of larger magnitude than in the United States and very similar across LMICs and HICs at around 4.3. Without time fixed effects, UMICs are of similar magnitude, and LICs are a third to a half below that found in the other groups. However, with them, the UMICs fall by more than half and become insignificant and the LICs coefficient disappears completely. A monotonic story in income is thus not clean, but it is consistent with the argument that in very poor countries, people cannot afford not to work and hence they will continue to do so. Figures 3a-c: Mobility, COVID Cases and NPIs, Global Sample Notes: Workplace mobility is Google measure of work-related mobility index. LIC, LMIC, UMIC, and HIC stand for Low Income Countries, Lower Middle-Income Countries, Upper Middle-Income Countries, and High Income Countries, respectively. See Table AIII for income group classification. The impact of NPIs themselves is mixed. Workplace closures are most clearly significant in LMICs, accounting for almost 9 points of reduced mobility which in UMICs and HICs, the point estimate is roughly half that and becomes insignificant with the inclusion of time fixed effects. School closures are robustly significant and account for 10 points in HICs, suggesting that having to school children at home is a limitation on job-related mobility. For UMICs, the coefficient is similar without time fixed effects, but falls to 6.6 pts and becomes insignificant with their inclusion. For LICs and LMICs, the point estimates are insignificant. This monotonic 10 increase with lower incomes is consistent with children playing a different role, perhaps helping in a business with less regard to human capital accumulation foregone. Again, the sampling for the LIC and LMIC samples for sure are not representative and what we may be finding is simply that people who can afford smart phones behave similarly around the world. Still, either LMIC governments have the capability to, at least, coral the elites, or, again, are simply providing a coordination mechanism. Cancelling public events never enters significantly with full time fixed effects although the point estimates are often in the -6 to -10 range. The restriction that most robustly reduces mobility among the LICs is closing of public transport, accounting for a massive 16.5 points. In UMICs, and arguably in HICs, the value is a third of that. This would seem the most potent tool of control in the poorest countries. Public information campaigns curiously enter positively and significantly in LMICs and almost in UMICs with coefficients of roughly 7-10. The intuition is not clear, but it may be the case that guidance on washing hands and wearing masks makes individuals feel more in control and protected and hence, the net impact is to increase mobility. Restrictions on internal movement have large and significant effects (12, 14.3) in LMICs and UMICs, with much less impact in HICs and virtually none in LICs. In the latter case, this may testify to difficulty in enforcing such shelter-in-place ordinances relative to, for instance, shutting down public transport. In sum, in HICs, and LMICs, the voluntary component is still as or more important as NPIs. UMICs look quite similar to HICs with the exception of anomalous lack of impact of case incidence, and the large impact of restrictions on internal movement which it shares with LMICs. It may be that in fact, LMICs and UMICs are more effective in enforcing such measures. Overall, for LICs the voluntary component is absent and the only NPI that appears to have any effect is closing public transportation. Again, with the caveat that cell phone coverage in such countries is around or under 20% of the population, this is consistent, again, with limited state capability and more resistance from the population to stop working. Again, Annex IV presents the complementary regressions on residential mobility and finds patterns that mirror those presented above. 11 Table 3: Workplace Mobility, COVID Cases, and NPIs, Global Sample (1) (2) (3) (4) (5) (6) (7) (8) LIC LIC LMIC LMIC UMIC UMIC HIC HIC K-12 closure 3.13 0.04 1.24 0.64 -6.62 -10.60** -10.20*** -13.32*** (4.83) (3.03) (4.61) (5.11) (4.80) (3.90) (3.16) (3.85) Close N.E. business 1.00 -0.80 -8.83* -9.30 -3.96 -8.59** -4.73 -8.75*** (7.40) (4.45) (5.01) (5.61) (3.37) (4.09) (2.84) (2.90) Cancel public events -9.77 -6.37 -5.26 -6.66* -1.49 -5.66 -2.32 -6.35* (5.27) (4.46) (3.88) (3.75) (5.96) (4.45) (3.04) (3.16) Close public transp. -16.51* -16.17* -2.20 -5.35 -5.37* -4.93 -5.06 -6.44** (8.37) (7.18) (4.93) (5.02) (2.86) (3.64) (3.03) (2.71) Public info. camp. 0.77 -0.40 9.90*** 10.47*** 7.32 8.99** 4.71* 5.59** (3.23) (3.35) (2.89) (2.31) (4.91) (4.07) (2.62) (2.70) Restr. on internal mov. -1.21 -1.85 -12.03*** -10.52** -14.32*** -16.81*** -2.72 -5.53** (3.57) (3.13) (2.98) (3.81) (3.78) (4.46) (2.04) (2.18) Log cases -0.03 -2.43* -4.30*** -5.57*** -1.50 -3.85*** -4.61*** -3.42*** (1.89) (1.17) (1.13) (0.56) (1.63) (0.80) (0.97) (0.75) Constant 3.76 14.08** -5.82 6.46 -0.50 8.68* -1.73 10.41*** (3.41) (5.84) (4.46) (4.02) (5.05) (4.75) (2.58) (2.49) Time FE Yes No Yes No Yes No Yes No Day of the week FE No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes # of Countries 8 8 24 24 29 29 40 40 Obs. 193 193 720 720 945 945 1777 1777 R-squared 0.69 0.62 0.77 0.73 0.85 0.80 0.86 0.80 Notes: Regression of Google measure of work-related mobility on NPIs, the log of national cases, country, and days of the week/time fixed effects. Robust clustered errors are in parenthesis. *** p<0.01, ** p<0.05, *<0.1. LIC, LMIC, UMIC, and HIC stand for Low Income Countries, Lower Middle Income Countries, Upper Middle Income Countries, and High Income Countries, respectively. See Table AIII for income group classification. V. Mapping to Economic Activity Do these voluntary declines in Google mobility in fact map to economic activity? Preliminary evidence from the United States and Sweden suggests they do. Figure 4 presents restaurant reservations by state against COVID incidence for the United States. What is immediately clear is that the fall in reservations predated the closing of nonessential businesses. This is confirmed by Table 4 which suggests a combined elasticity of over 10 and virtually no impact of business closing measures. That is, the entire fall can be accounted for by the increase in cases. The results suggest that what slowed economic activity was not the NPIs, but rather voluntary demobilization as evidence of the magnitude of the threat accumulated. In the same vein, Figure 2b presents preliminary national data from movie theater releases and revenues in Sweden, which imposed a 50-person limit on gatherings and suggested closing of non-essential businesses only on March 30th. By March 17th, revenues had fallen to almost zero while releases continue unaffected. Since the data are at the national level, we cannot document these trends more formally. 12 Figure 4: Decline in Restaurant Figure 5: Decline in Movie Theater Reservations vs. COVID Cases Revenues and Releases vs. COVID Cases US Sweden March 30 50 50 Restaurant reservations (Y/Y percent) Gross revenues and movie releases 0 0 -50 -50 -100 -100 -2 0 2 4 6 -5 0 5 10 Log(case per 1 million people) Log(case per 1 million people) Revenues, Y/Y percent Releases, level No shutdown Shutdown Notes: U.S. Restaurant reservations against COVID incidence. Sweden: Movie releases and theater revenues against COVID incidence. See text for sources. In both the cases of restaurant reservations in the United States and theater demand in Sweden, demand has fallen sharply and independent of NPIs. This suggests that, as in Wuhan, it is likely that release of NPIs will have little effect unless individuals are confident that the risk has diminished. VI. Conclusion Several key findings thus emerge. First, clearly, the pattern of demobilization varies across countries with the political choices made. The United States and Japan have radically different degrees of demobilization. Second, decreased mobility seems more driven by “voluntary” individual response to increased local and national COVID-19 case incidence, proxying for awareness or fear or social responsibility, rather than formal measures. For all except the poorest countries (LICs), the response of mobility with respect to cases is of similar orders of magnitude and can explain most of the reduction in mobility, dwarfing the effect of NPIs. Third, that said, there is evidence that less affluent countries were also able to implement NPIs. LMICs and UMICs appear to have been able to engineer as much or more of a fall in mobility through NPIs as some HICs. Fourth, our global data suggest that other measures beyond closing nonessential workplaces have important impacts-school closures, restrictions on internal mobility/shut-down of public transportation. Counterintuitively, public information campaigns appear to raise mobility- information on protective measures may make individuals feel more confident moving about. 13 Table 4: Restaurant Reservations, COVID Cases, and NPIs, United States (1) Restaurant reservations Close N.E. business 0.818 (1.381) Close K-12 2.349 (1.720) Stay home/SIP 0.952 (1.139) Log cases -0.678 (1.125) Log national cases -9.775*** (0.884) Constant 31.251*** (6.388) Time FE Yes State FE Yes # of States 49 Obs. 1877 R-squared 0.958 Notes: Regression of restaurant reservations (Y/Y percent change) from OpenTable, on NPIs, the log of cases, the log of national cases, state, time fixed effects. Robust clustered errors are in parenthesis. *** p<0.01, ** p<0.05, *<0.1 Fifth, the dominant contribution of voluntary self-restraint along with historical and anecdotal evidence suggests that formal NPIs may be as much coordination mechanisms as repressive measures. For instance, no parent may want to send his/her child to school, but only when schools force all students online can continued safe learning at a distance be realized. As in Japan, no one may want to be seen as the slacker by not showing up at work, but if the government signals that this is the safe thing to do, then all can work from home without stigma. Sixth, these findings offer both good and bad news. First, they imply that for many countries in the world, self- enforcing dynamics and NPIs can reduce mobility and business activity substantially. That mobility fell almost as much in Sweden, with no NPIs, as the United States dramatically illustrates this point and suggests that the focus on government NPI policy in explaining Sweden’s mortality rate may be justified. The finding that only shutting down public transport has any effect in LICs is consistent with arguments that government capacity may be generally low, and resistance to demobilizing is high where it implies lost livelihoods. 14 Seventh, the potentially bad news is that releasing constraints may not, as appears to be the case in Wuhan, have the economically rejuvenating effect that was expected if people are not convinced that, in fact, the coast is clear. Given this, we are more likely to be facing a U-shaped recovery rather than a V propelled by the release of constraints. 15 VII. References Anderson, Roy, Hans Heesterbeek, Don Klinkenberg, and Déirdre Hollingsworth (2020). “How will country- based mitigation measures influence the course of the COVID-19 epidemic?” The Lancet 395 (10228), 931–934. Auld, M. C. (2006). “Estimating behavioral response to the AIDS epidemic.” The BE Journal of Economic Analysis & Policy 5(1). Baldwin, R. and B. Weder de Mauro (2020). Economics in the Time of COVID-19, https://voxeu.org/content/economics-time-covid-19. Barnett-Howell, Z and A.M Mobarak (2020). “Should Low-Income Countries Impose the Same Social Distancing Guidelines as Europe and North America to Halt the Spread of COVID-19?” Yale School of Management and Y-RISE. Brzezinski, A., G Deiana, V. Kecht and D. Van Dijcke (2020). “The COVID-19 pandemic: Government versus community action across the United States,” Covid Economics, Issue 7. Chen, X. and Z.Qiu (2020). “Issue Scenario analysis of non-pharmaceutical interventions on global COVID- 19 transmissions,” Covid Economics, Volume 7. Crosby, Alfred W. (2003). America’s Forgotten Pandemic: The Influenza of 1918. Cambridge University Press. Engle,S J. Stromme and Z. Anson (2020) “Staying at home: Mobility effects of Covid-19” Covid Economics, Issue 4. Favero, C. (2020). “Why is Covid-19 mortality in Lombardy so high? Evidence from the simulation of a SEIHCR model,” Covid Economics, Issue 4. Gonzalez-Eiras, M. and D. Niepelt (2020). “On the optimal 'lockdown' during an epidemic,” Covid Economics, Issue 7. Hale, T., A. Petherick, T. Phillips, S. Webster (2020a) “Variation in government responses to COVID-19” BSG Working Paper Series BSG-WP-2020/031 Version 4.0 April 2020. Hale, T., S. Webster, A. Petherick, T. Phillips, and B. Kira (2020b). Oxford COVID-19 Government Response Tracker, Blavatnik School of Government: https://www.bsg.ox.ac.uk/research/research- projects/coronavirus-government-response-tracker Hartl, T., K. Wälde and E. Weber (2020). “Measuring the impact of the German public shutdown on the spread of COVID-19,” Covid Economics, Issue 1. Loayza, N. (2020). “Costs and Trade-Offs in the Fight against COVID-19,” World Bank, Washington, DC. Painter, M. O. and T. Qiu (2020). “Political belief affect compliance with COVID-19 social distancing orders,” Covid Economics, Issue 4. Raifman J., Nocka K., Jones D., Bor J., Lipson S., Jay J., and P. Chan (2020). "COVID-19 US state policy database,” available at: www.tinyurl.com/statepolicies 16 Annex I. Smartphone Coverage Country Smartphone penetration United Kingdom 82.20% 1 Netherlands 79.30% 2 Sweden 78.80% 3 Germany 78.80% 4 United States 77.00% 5 Belgium 76.60% 6 France 76.00% 7 Spain 72.50% 8 Canada 72.10% 9 Australia 68.60% 10 Korea, Rep. 68.00% 11 Kazakhstan 64.90% 12 Poland 64.00% 13 Russian Federation 63.80% 14 Taiwan, China 60.00% 15 Italy 58.00% 16 Malaysia 57.50% 17 Japan 55.30% 18 China 55.30% 19 Romania 53.80% 20 Ukraine 48.30% 21 Argentina 46.90% 22 Saudi Arabia 46.00% 23 Mexico 45.60% 24 Philippines 44.90% 25 Chile 44.20% 26 Thailand 43.70% 27 Brazil 41.30% 28 Venezuela, RB 40.80% 29 Colombia 39.80% 30 Morocco 37.90% 31 Turkey 37.90% 32 Vietnam 37.70% 33 South Africa 35.50% 34 Iran, Islamic Rep. 64.60% 35 Peru 32.10% 36 Uzbekistan 31.30% 37 Algeria 29.10% 38 Egypt, Arab Rep. 28.00% 39 India 27.70% 40 17 Indonesia 27.40% 41 Ghana 24.00% 42 Myanmar 21.80% 43 Kenya 20.90% 44 Sudan 19.70% 45 Bangladesh 16.10% 46 Uganda 15.60% 47 Pakistan 13.80% 48 Nigeria 13.00% 49 Ethiopia 11.20% 50 Source: Newzoo's Global Mobile Market Report (2018) as cited at https://en.wikipedia.org/wiki/List_of_countries_by_smartphone_penetration 18 Annex II. Subnational Data Sources Brazil: Official state websites, Platforma COVID Brazil by the Government of Brazil: https://covid19br.wcota.me/ Italy: Dipartimento della Protezione Civile: https://github.com/pcm-dpc/COVID-19 Japan: Japan COVID-19 Data Repository: https://github.com/sanpei3/covid19jp South Africa: Department of Health: https://github.com/dsfsi/covid19za Sweden: https://www.boxofficemojo.com/weekend/by-year/2020/?area=SE UK: Department of Health and Social Care: https://github.com/tomwhite/covid-19-uk-data 19 Annex III. Income Groups LIC LMIC UMIC HIC Afghanistan Angola Argentina Australia Burkina Faso Bangladesh Belize Austria Mali Bolivia Bosnia and Herzegovina Belgium Mozambique Cameroon Botswana Canada Niger Cabo Verde Brazil Chile Rwanda Egypt, Arab Rep. Bulgaria Croatia Tanzania El Salvador Colombia Czechia Uganda Ghana Costa Rica Denmark Honduras Dominican Republic Estonia India Ecuador Finland Indonesia Guatemala France Kenya Iraq Germany Kyrgyzstan Jamaica Greece Hong Kong SAR, Lao PDR Jordan China Mongolia Kazakhstan Hungary Myanmar (Burma) Lebanon Ireland Nicaragua Libya Israel Nigeria Malaysia Italy Pakistan Mauritius Japan Papua New Guinea Mexico Luxembourg Philippines Namibia Netherlands Vietnam Paraguay New Zealand Zambia Peru Norway Zimbabwe Romania Panama South Africa Poland Sri Lanka Portugal Thailand Puerto Rico Turkey Saudi Arabia Venezuela, RB Singapore Slovak Republic Slovenia Korea, Rep. Spain Sweden Switzerland Trinidad and Tobago United Arab Emirates United Kingdom United States Uruguay 20 Annex IV. Table A4: Residential mobility, global sample (1) (2) (3) (4) (5) (6) (7) (8) LIC LIC LMIC LMIC UMIC UMIC HIC HIC K-12 closure -1.59 -1.81 1.78 1.86 4.34** 5.67*** 3.71** 5.18*** (3.57) (2.00) (2.33) (2.57) (2.04) (1.74) (1.39) (1.57) Close N.E. business 0.84 0.49 4.37** 4.63* 2.69* 4.68*** 1.65 3.10** (1.75) (2.16) (2.06) (2.35) (1.31) (1.46) (1.39) (1.34) Cancel public events 7.34*** 4.03 0.67 1.39 1.39 2.71 0.87 2.83** (1.61) (2.58) (1.71) (1.82) (2.54) (1.98) (1.38) (1.31) Close public transp. 2.74 4.78 -0.07 1.25 0.42 0.30 3.25** 3.19** (2.26) (2.93) (2.07) (2.07) (1.61) (1.77) (1.38) (1.21) Public info. camp. -2.71** -2.34 -5.94*** -6.16*** -5.27** -5.22*** -2.32 -2.41* (0.96) (2.37) (1.89) (1.45) (2.28) (1.72) (1.43) (1.34) Restr. on internal mov. 2.46 3.06 6.35*** 5.83*** 7.90*** 9.26*** 0.69 1.34 (1.68) (1.63) (1.23) (1.67) (1.74) (1.77) (1.01) (0.98) Log cases 0.84 1.37* 1.68*** 2.20*** 0.12 1.25*** 1.99*** 1.55*** (0.83) (0.60) (0.42) (0.27) (0.74) (0.38) (0.55) (0.36) Constant 7.28*** 1.43 5.60** 2.52 1.35 -1.25 1.01 -4.48*** (1.22) (3.99) (2.28) (2.18) (2.52) (1.88) (1.35) (1.18) Time FE Yes No Yes No Yes No Yes No Day of the week FE No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes # of Countries 8 8 24 24 29 29 40 40 Obs. 193 193 711 711 942 942 1775 1775 R-squared 0.78 0.71 0.80 0.77 0.83 0.80 0.85 0.79 Notes: Regression of Google measure of residential mobility on NPIs, the log of national cases, country, and days of the week/time fixed effects. Robust clustered errors are in parenthesis. *** p<0.01, ** p<0.05, *<0.1. LIC, LMIC, UMIC, and HIC stand for Low Income Countries, Lower Middle Income Countries, Upper Middle Income Countries, and High Income Countries, respectively. See Table AIII for income group classification. 21