Lockdown Strategies, Mobility Patterns and COVID-19
LLockdown Strategies, Mobility Patterns and COVID-19 ∗ Nikos Askitas
IZA - Institute of Labor Economics and CESifo
Konstantinos Tatsiramos
University of Luxembourg, LISER, IZA and CESifo
Bertrand Verheyden
Luxembourg Institute of Socio-Economic Research (LISER)
May 27, 2020
Abstract
We develop a multiple-events model and exploit within and between country variationin the timing, type and level of intensity of various public policies to study their dy-namic effects on the daily incidence of COVID-19 and on population mobility patternsacross 135 countries. We remove concurrent policy bias by taking into account thecontemporaneous presence of multiple interventions. The main result of the paper isthat cancelling public events and imposing restrictions on private gatherings followedby school closures have quantitatively the most pronounced effects on reducing thedaily incidence of COVID-19. They are followed by workplace as well as stay-at-homerequirements, whose statistical significance and levels of effect are not as pronounced.Instead, we find no effects for international travel controls, public transport closuresand restrictions on movements across cities and regions. We establish that these find-ings are mediated by their effect on population mobility patterns in a manner consistentwith time-use and epidemiological factors.
Keywords:
COVID-19, public policies, non-pharmaceutical interventions, mul-tiple events, mobility
JEL codes:
I12, I18, G14 ∗ We thank an anonymous reviewer, Jan van Ours and Adrian Nieto Castro for discussions. • Askitas:[email protected] • Tatsiramos: [email protected] • Verheyden: [email protected] a r X i v : . [ ec on . E M ] M a y Introduction
In December 2019, the COVID-19 outbreak was registered in Wuhan China. The WorldHealth Organization declared it a “Public Health Emergency of International Concern” onJanuary 30, 2020 and escalated it to a pandemic on March 11, 2020. The disease hasbeen recorded in over 200 countries and territories with several millions of confirmed casesand a case mortality rate of around seven percent. In the early stages of the outbreak,attempts were made to trace every infection back to its origin. Tracing back to the “index”case on an international level soon became impossible and most countries responded byimposing restrictions on international travel. In the later stages of the epidemic, a numberof non - pharmaceutical interventions (henceforth referred to as NPIs or public policies) wereundertaken, which were of a domestic nature revolving around the idea of “social distancing”.The aim of these interventions was to slow down the pandemic by restricting mobility sothat it does not overwhelm health system capacities.This paper studies how lockdown policies affect the daily incidence of COVID-19 andpopulation mobility patterns across 135 countries based on several data sources. Under-standing the effectiveness of these policies is important as policy makers and the society atlarge seek to achieve an optimal health outcome in the fight against the pandemic at thelowest economic cost.We exploit between and within country variation in the type, timing, and level of intensityof lockdown policies in a multiple events study approach, which aims at disentangling theeffect of each intervention on COVID-19 incidence and mobility patterns, while controlling forthe presence of concurrent policy measures during the event window of the policy of interest, COVID-19 Dashboard by the CSSE at Johns Hopkins University (JHU). The data sources include: i) Coded government response data obtained from Hale et al. [2020], ii)prevalence data from European Centre for Disease Prevention and Control (ECDC) and iii) populationmobility data from Google Community Mobility Reports. The analysis includes the latest data up to thiswriting.
1s well as for time fixed effects, day of week fixed effects, lagged COVID-19 prevalence, regionfixed effects and time-varying country-specific characteristics.The main contributions of the paper are the following. First, we develop a multiple eventsmodel which allows us to identify the dynamic effects of each intervention while taking intoaccount the presence of other concurrent interventions at each event time. Accounting forconfounding policies is important because it allows us to avoid attributing the effect of otherinterventions to the policy of interest, and in addition to establish that it is policies thataffect mobility patterns and not that policies ex-post respond to changing mobility patternsin the population.Second, we consider a wide range of interventions across 135 countries, which vary intheir type, intensity, and timing. The policy responses in focus are i) international travelcontrols, ii) public transport closures, iii) cancelation of public events, iv) restrictions onprivate gatherings, v) school closures, vi) workplace closures, vii) stay-at-home requirementsand viii) internal mobility restrictions (across cities and regions).Third, we link policy interventions to mobility patterns by studying not only the impactof these policies on the incidence of COVID-19, but also on the time spent in a numberof types of places such as i) retail and recreation, ii) grocery and pharmacy, iii) parks, iv)transit stations, v) the workplace and vi) residential areas. Each of these types of places ischaracterized by different epidemiological features and, therefore, has a different potentialfor viral transmission. The mobility data can then also be viewed as a measure of complianceto the policies introduced and a mediator between policies and the spread of the disease.The main result of the paper is that cancelling public events and imposing restrictionson private gatherings followed by school closures have quantitatively the most pronouncedeffects. They are followed by workplace as well as stay-at-home requirements, whose statis-tical significance and levels of effect are not as pronounced. Instead, we find no effects forinternational travel controls, public transport closures and restrictions on movements across2ities and regions. We thus establish i) the order in which public policies help curb the pan-demic and ii) that these effects are mediated by the way they change population mobilitypatterns in a manner consistent with time-use and epidemiological factors.The rest of the paper is structured as follows. Section 2 contains a literature review,while Section 3 discusses the data and presents summary statistics about NPIs and mobilitypatterns. Section 4 describes the model and identification issues. The results are presentedin Section 5, which contains a discussion linking the evidence on COVID-19 incidence withmobility patterns. Section 6 concludes by summarizing the findings and discussing futureresearch.
Research on infectious diseases focuses on vaccinations and drugs but it also aims at curbingthe spread of the diseases by understanding and predicting their spatiotemporal dynamics,especially in the event of a new virus outbreak. Recent epidemiological models have beenenriched to incorporate the impact of NPIs on these dynamics, which are at the core ofthis paper. The canonical model used in epidemiology is the so-called SIR model (Kermacket al. [1927]). It provides a simple and relevant representation of the mechanics of viruspropagation with three categories of individuals: i) the people who are susceptible to becomeinfected (the S subpopulation), ii) the infected who can transmit the disease (the I) and iii)those who have recovered and cannot infect anymore (the R). A crucial concept in theSIR model is the R
0, which is the average number of people that a sick person infectsbefore she recovers. While the R0 is often considered as a biological characteristic of thevirus transmissibility, it is also affected by environmental, behavioral, and social dimensions,including NPIs.In its basic form, the SIR compartmental model assumes that the population of interest3s homogeneous in terms of exposure, immunity and chances of recovery. This assumption,however, is not realistic as in practice these factors have proven to be key in guiding policyinterventions (Auchincloss et al. [2012]). Relaxing this assumption gave rise to extensions ofthis model which aim at capturing the multiple dimensions of heterogeneity by partitioningthe population into groups based on age or location. Pushed to the extreme, such partitionslead to the individual-based models (Eubank et al. [2004]), which require data on the intensityof contacts between individuals of different age groups to calibrate person-to-person contactrates, for instance via social mixing matrices. Using this approach, Jarvis et al. [2020] find forthe UK that lockdown policies reduced the average number of daily contacts by 73 percent,resulting in a drop of the R . .
62, while Singh and Adhikari [2020] show forIndia that lockdown policies are unlikely to be effective if applied for 3 weeks or less.Beyond the partitioning of the population by age groups or communities, the compart-mental model has been extended to take into account important specificities of the disease,such as the incubation period, the duration of the acquired immunity, or the challenges itpresents given the current state of knowledge. In the context of the COVID-19 pandemic,several variants of the compartmental model have been used. The SEIR model takes intoaccount the group of exposed individuals who can be asymptomatic carriers during the in-cubation period (Karin et al. [2020], Pang [2020]; Roy [2020], Lyra et al. [2020], Lai et al.[2020]). In the SIRS model, recovery only provides a short-lived immunity, so that the Rgroup moves back to the S group after some time (Ng and Gui [2020]). The SIOR modelconsiders that only a fraction of the infected group is detected, or observed, by healthcareservices (Scala et al. [2020]). While these various models capture different important featuresof the COVID-19 pandemic and provide predictions on the impact of NPIs on the spread ofthe disease, their results are usually simulation-based and rely on structural assumptions.In the field of economics, recent contributions depart from simple versions of the SIRmodel and introduce confinement policies as well as economic concepts (e.g. incentives, eco-4omic cost of lockdown, value of life). These papers highlight through calibrated simulationsthe tradeoffs between the mortality induced by the excess demand for healthcare servicesand the economic losses induced by confinement policies. Gonzalez-Eiras and Niepelt [2020]develop a SIR model in which policies take into account, among others, the rate of timepreference, the learning of healthcare services and the severity of output losses. Garibaldiet al. [2020] depart from the observation that the SIR model treats transitions from S to Ias exogenous. In other words, the SIR model does not take into account individuals decisionto reduce the intensity of their contacts and their exposure to the virus. The authors bor-row concepts from the search and matching model (Pissarides [2000]) to introduce a contactfunction into the SIR model with forward-looking agents. They show that the decentralizedepidemic equilibrium is likely to be suboptimal due to the presence of externalities: whileindividuals care about the private benefits of distancing, they neglect its social benefits andthe fact that it reduces the risk of hospital congestion; on the other hand, from a dynamicperspective, they do not take into account the benefits of herd immunity. Greenstone andNigam [2020] develop a method to quantify the economic benefits of social distancing mea-sures in terms of lives saved. They find that 1.7 million lives could be saved by applyingmild social distancing for 3 to 4 months, which they estimate to be worth 8 trillion dollarsaccruing for 90 percent to the population above 50 years of age. Barro [2020] studies theimpact of NPIs in the US during the Great Influenza Pandemic at the end of 1918 findingthat even though NPIs reduced deaths peaks, and thereby reduced the stress imposed onhealthcare services, they failed to significantly decrease overall mortality, which is likely dueto the short application of the NPIs, with an average duration of around one month.The papers related to this study are Chen and Qiu [2020], Gao et al. [2020], Engleet al. [2020] and Huber and Langen [2020]. Chen and Qiu [2020] focus on the reproductionnumber, which depends on the timing of NPI’s with a parametric time lag effect, and predictfor 9 countries the transmission dynamics under various sets of NPI’s showing that country5ifferences lead to different optimal policies with heterogeneous tradeoffs between health andeconomic costs. By combining geographic information systems and daily mobility patterns inUS counties, derived from smartphone location big data, Gao et al. [2020] show that in manycounties in which mobility restrictions were only recommended but not imposed, mobilitydid not decrease. Engle et al. [2020] using US county-level GPS and COVID-19 cases, studythe impact of local disease prevalence and confinement orders on mobility solving a utilitymaximization problem after splitting the utility derived from traveling a unit of distanceinto costs independent from the epidemic and costs related to perceived risk of contractingthe disease. They find substantial effects of local infection rates, while official confinementorders lead to a mobility reduction of less than 8 percent. Huber and Langen [2020] exploitregional variation in Germany and Switzerland to assess the impact of the timing of COVID-19 response measures finding that a relatively later exposure to the measures entails highercumulative hospitalization and death rates.We differ from these papers in the following ways: i) we develop a multiple events modelexploiting the timing, type and level of intensity of several public policies with the advantageof flexibility in the non-parametric estimation of their dynamic impacts, taking into accountthe contemporaneous presence of multiple interventions; ii) we consider as outcomes boththe COVID-19 cases and various mobility patterns, with the latter capturing how often andhow long certain public places or one’s residence is frequented; and iii) we analyze a paneldataset of 135 countries.
The analysis combines information from multiple data sources on (i) the non-pharmaceuticalinterventions implemented by governments, (ii) the daily number of infections, (iii) the evo-lution of population’s mobility patterns, and (iv) various country characteristics.6on-pharmaceutical interventions are collected by the Oxford COVID-19 GovernmentResponse Tracker (henceforth OxCGRT) for most countries of the world. The OxCGRTgathers publicly available information on several indicators of public policies aiming at miti-gating the propagation of the virus. We focus on the following interventions: i) internationaltravel controls, ii) closure of public transport, iii) cancelation of public events, iv) restrictionson private gatherings, v) closure of schools, vi) closure of workplaces, viii) restrictions oninternal movement and viii) stay-at-home requirements. For each of these policies, we ex-ploit information on the dates of introduction as well as qualitative time-varying informationon their intensity. Intensity is measured in a scale from 1 to 6, which reflects whether theintervention is (i) recommended, (ii) mandatory with some flexibility, and (iii) mandatorywith no flexibility, and whether it is geographically targeted or applied to the entire country.Recommended policies which are targeted obtain a value of 1, while mandatory policies withno flexibility applied to the entire country obtain a value of 6, with values in between refer-ring to combinations of the policy stringency and its geographic scope. We use a sample of135 countries for the estimation of NPIs, which is the set of countries for which we also haveinformation on country characteristics (for a complete list see Appendix B).Figure 1 presents the distribution of the number of days it took for each policy to beintroduced after the first COVID-19 case averaged across countries. The distribution for theinternational travel controls is bimodal with the first mode well ahead of the first case. Allpolicies have a main mode close to zero, with cancelation of public events and school closuresenacted earlier, followed by restrictions on private gatherings and workplace closures, stay-at-home requirements, internal mobility restrictions and public transportation restrictions, To fix ideas, when the schooling policy receives an intensity score of 4, it means that it was not mademandatory in all schools or in all education levels, but it was applied to the entire country. A score of 5means that it was made mandatory to all schools and education levels, but only in some areas of the country.A score of 6 means that it is mandatory for all schools and areas of the country. The average intensitylevel across countries is 2.9 for international travel controls, 3.8 for public transport closures, 5.4 for schoolclosures, 3.8 for workplace closures, 4.7 for cancelling public events, 4 for restrictions on private gatherings,3.2 for stay-at-home requirements and 4.2 for internal movement restrictions. To study how mobility patterns have evolved worldwide, we resort to Google’s Commu-nity Mobility Reports. The Google mobility data are created with “aggregated, anonymizedsets of data from users who have turned on the Location History setting” on their phone andshow how “visits and length of stay” at different types of places change compared to themedian value, for the corresponding day of the week, during the 5-week period from January3, 2020 to February 6, 2020. Google’s ability to accurately locate phones and to correctlycategorize places varies both across countries as well as within (urban vs. rural areas). Thesedata contain information on various epidemiologically relevant categories of places such as: i)retail and recreation, ii) grocery and pharmacy, iii) parks, iv) workplaces, v) transit stationsand vi) residential areas. Retail and recreation covers visits to restaurants, cafes, shoppingcenters, theme parks, museums, libraries, and movie theaters. Grocery and pharmacy coversgrocery markets, food warehouses, farmers markets, specialty food shops, drug stores, andpharmacies. Parks encompass national parks, public beaches, marinas, dog parks, plazas, Using the 3-day moving average helps visualization without changing the main findings. We alsoconsidered a 7-day moving average, which similarly maintains the main findings while removing reportingidiosyncrasies but additionally flatten features of the data which might be of interest. We opted for the 3-daymoving average as the middle ground. We also conduct our analysis without the 3-day moving average asdiscussed below. COVID-19 Community Mobility Reports. In terms of densities, retail and recreation, grocery and pharmacy, transitstations as well as workplaces are somewhat similar, while parks and residential areas are ontheir own on opposite sides.
We follow an event-study approach around the time of policy implementation, which weextend to account for multiple events. The single event study (e.g. Kleven et al. [2019]) canbe expressed with the following equation: Y c,t = (cid:88) j (cid:54) = − α j I [ j = t − t πc ] + (cid:88) l γ l I [ l = t ]+ (cid:88) d δ d I [ d c ( t ) = d ] + (cid:88) r ρ r R r + φZ c,t − + θX c + (cid:15) c,t , (1) Considering that staying at home is by far the most time intensive activity, according to time-usestudies, this value is quite large. We expand on this point in Section 5.3. Y c,t denotes the outcome in country c at event time t . The first term, on the right handside, is a set of event time dummies for the intervention of interest π , where t πc denotes itsimplementation day in country c . We consider the outcome in the window starting 20 daysbefore the intervention up to 35 days after its implementation, so the event time runs from −
20 to +35. We omit the event time dummy at j = −
20 so that the event time coefficientsof interest α j measure the impact of intervention π at time j relative to the twenty daysbefore the policy was implemented.The second term, on the right hand side of equation (1), is a set of dummies whichcontrol non-parametrically for trends in the time since the first-observed COVID-19 case.Identifying the coefficients of the event time dummies conditional on these time effects ispossible because the timing of NPIs differs across countries. The third term, is a set of day-of-week dummies controlling for potential day-specific differences both in terms of reportingof new cases and of mobility patterns (here d c ( t ) returns the day of the week for event time t in country c ). The fourth term, is a set of dummies for the following regions: Europe, Asia,Middle East, North America, South America, Oceania and Africa. The fifth term, is the logvalue of the total number of confirmed cases at t − c . Including this variableallows us to capture the size of the pool of infected people, which is a crucial factor both whenthe outcome is the incidence of new cases, in line with the SIR framework, as well as when itis mobility patterns, as populations may react to the perceived threat of contamination. Thesixth term, is a set of country specific variables controlling for differences across countries,such as per capita GDP, population density and the urbanization rate, followed by the errorterm. When evaluating the effect of the intervention of interest π , it is important to take into Adjusting the total number of confirmed cases by the number of deaths does not affect our main results. When the outcome is the incidence of COVID-19, these controls are epidemiologically relevant, whereas,when we consider Google mobility types as our outcome they help control for differences in Google’s abilityto geo-locate phones and detect types of places. π with multiple events is more challenging thanin the single-event case, especially when the multiple events fully overlap during the eventwindow of the policy of interest. Concurrent NPIs, denoted by π (cid:48) , can be controlled for byintroducing in equation (1) a new term, F π (cid:48) [ j = t − t πc ], which is a set of dummies - onedummy for each event time of the policy of interest π - which are equal to one if any otherinterventions π (cid:48) are in effect at event time j for country c . The multiple events regressionequation can then be written as follows: Y c,t = (cid:88) j (cid:54) = − α j I [ j = t − t πc ] + (cid:88) j β j F π (cid:48) [ j = t − t πc ] + (cid:88) l γ l I [ l = t ]+ (cid:88) d δ d I [ d c ( t ) = d ] + (cid:88) r ρ r R r + φZ c,t − + θX c + (cid:15) c,t . (2)The identification problem in the multiple-event case emerges as soon as other policieshave been introduced before the start of event window of policy π . This would result ina complete overlap of policies within the event window, making it impossible to separatelyidentify the effect of the event of interest from the other contemporaneous events. To achieve identification in the multiple events case, we use the level of intensity of eachpolicy which varies both within and across policies, as well as across countries. This variationof policy intensity allows us to identify separately the effect of the policy of interest π , whiletaking into account other concurrent NPIs, π (cid:48) . The extended multiple events regression When other policies are enacted within the event window, then the two set of event dummies are notperfectly collinear so the coefficient estimates α j and β j can be separately identified, but at the cost of highvariance because of multicollinearity. Y c,t = (cid:88) j (cid:54) = − α j S π [ j = t − t πc ] + (cid:88) j β j ¯ S π (cid:48) [ j = t − t πc ] + (cid:88) l γ l I [ l = t ]+ (cid:88) d δ d I [ d c ( t ) = d ] + (cid:88) r ρ r R r + φZ c,t − + θX c + (cid:15) c,t , (3)where the first term, S π [ j = t − t πc ], is taking the value of the level of intensity of thepolicy of interest π in country c at event time j and zero otherwise, while the second term,¯ S π (cid:48) [ j = t − t πc ], is equal to the average level of intensity of all other contemporaneous policies π (cid:48) of country c at the event time j , and zero if there are no other policies active at that eventtime. That is, we extend equation (2) in two ways: 1) we multiply the event dummies forpolicy π with the intensity level of the policy at event time j - in other words, I [ j = t − t πc ]in equation (2) generalizes to S π [ j = t − t πc ] in equation (3); and 2) we multiply the dummiescontrolling for the presence of any other policies π (cid:48) - at event time j for policy π - with theiraverage intensity at event time j - in other words, F π (cid:48) in equation (2) generalizes to ¯ S π (cid:48) c,t inequation (3).Our identification relies on the variation in the timing and intensity of various inter-ventions both within and across countries, conditional on the prevalence of COVID-19, timeeffects since the first observed case, day effects, country-specific characteristics and continenteffects. This variation allows to separately identify the effect of intervention π from that ofother concurrent NPIs, π (cid:48) . The coefficient estimates a j in equation 3 measure the unit levelintensity effect of policy π at event time j on the outcome. This section contains the results split in three subsections. Subsection 5.1 contains resultson the effect of NPIs on the incidence of COVID-19, whereas Subsection 5.2 contains results12n the effect of NPIs on population mobility patterns from Google’s Community MobilityReports. Finally, Subsection 5.3 provides a consolidated view on the link between the impactof NPIs on new cases through their effect on mobility.
We start by comparing the estimates for the dynamic effects of each intervention obtainedfrom the two versions of the model: i) ignoring concurrent interventions, i.e. estimatingequation (3) without the second term, and ii) controlling for concurrent interventions, i.e.estimating equation (3). Comparing the two sets of estimates, it becomes apparent that ignoring the presenceof other interventions leads to biased estimates. Specifically, the results of Figure 3, whichreport the estimates ignoring concurrent policies, suggest that all policies tend to have asignificant impact following a similar pattern. That is, in the days preceding the introductionof the policy, the incidence of COVID-19 increases until it reaches a peak after few a daysfollowing its introduction. Then, the number of new cases per day start to decrease, andwithin a month they become significantly lower than the reference event time (20 days beforethe policy).The analysis without controlling for other concurrent policies seems to suggest that allinterventions were successful in containing new infections. However, the estimates in Figure4, which are obtained after controlling for concurrent policies, convey a different message.This is especially true for the two transport related interventions, i.e. international travelcontrols and public transport closure , and for restrictions on internal movement , which havealmost no impact on new cases once other interventions are controlled for.The two policies with the largest effects, which are robust to confounding by other policies, We focus on the results where the dependent variable is the 3-day moving average of confirmed newcases. The estimates with the number of confirmed cases are reported in Figures A.1 and A.2 in AppendixA. cancelling of public events and restrictions on private gatherings . These are policies whichaim to reduce massive contacts. For both, we observe a drop in the incidence of COVID-19starting about one week after implementation, which becomes significantly different thanzero within two weeks. Around the end of the event window, a unit increase in the intensityof the policy of interest leads to a 20% decrease in the number of new infections in the case ofpublic events cancelation, and a decrease of about 12% in the case of restrictions on privategatherings, compared to the reference event time.
School and workplace closures aim to control contacts in large groups, but unlike publicevents and private gatherings are easier to monitor and regulate as well as trace wheneverinfections do occur. We find that new infections start declining a few days after school clo-sures, with the effect becoming negative and significant about 25 days after implementation.Around the end of the event window, a unit increase in the intensity of school closures leadsto about a 15% drop of new infections compared to the reference event time. For workplaceclosures, we find that new infections start declining starting from the second week afterimplementation and the effect becomes negative and significant only towards the end of theevent window, with a unit increase in the policy intensity leading to about a 10% drop ofnew infections.Finally, stay-at-home requirements aim to impose mobility constraints at the individuallevel, which is arguably the most extreme of all measures and was generally introduced wheninfections were reaching alarming growth rates. This is captured in Figure 4 which showsthat, around the date of introduction of the policy, there were on average 20% more newcases every day compared to the reference event time, with the policy reversing that trendimmediately. By the end of the window, the coefficient estimates are statistically significantlylower from those around the policy implementation day, although they are not statisticallydifferent from zero. See the discussion in Section 5.3. .2 Lockdown Policies on Google Mobility Patterns Google mobility patterns are observed as percentage point deviations from a reference cal-endar period before the onset of COVID-19. As a result, the coefficient estimates of interest- first term of equation (3) - measure the percentage point change in mobility patterns fora unit level of intensity of each intervention compared to the reference point before imple-mentation. Similar to COVID-19 confirmed new cases, we obtain estimates both with, aswell as without controls for other ongoing interventions. The estimates with controls forconcurrent policies are presented in Figures 5, 6, 7 and 8, while those without controllingfor confounding policies can be found in Figures A.3, A.4, A.5 and A.6 in Appendix A.We find that, when we do not control for confounding policies, right after the day ofpolicy implementation there is a general pattern of sharp and large drops in all mobilitypatterns related to activities undertaken outside residential areas, and an increase in theamount of time spent in the place of residence. However, once we control for other con-current NPIs, many of these effects are either much smaller, or sometimes not significantlydifferent from zero. For example, the estimates for international travel controls withoutaccounting for confounders, shown in panel (a) of Figure A.3, suggest a significant declinein movements immediately after the policy implementation across most places (retail andrecreation, grocery and pharmacy, parks, transit stations, workplaces) and increases in stay-ing home. However, after controlling for other interventions present around the same time,we find in panel (a) of Figure 5 that restrictions in international travel have a much smallerimpact on all types of movement.After accounting for multiple events, panel (b) of Figure 5 shows that restrictions on public transportation lead to a sharp discontinuity at the day of the intervention with lowermovements outside home. Interestingly, the strength of this decrease in mobility tends toweaken with time. The limited - and not very persistent - reductions in mobility patternsobserved for international travel controls and closure of public transportation are consistent15ith the small effects of these policies on the incidence of new cases reported in Section 5.1.The cancelation of public events and restrictions on private gatherings , which led tothe most important reductions in new infections, as reported in Section 5.1, also exhibitlarge and persistent negative impacts on retail and recreation, transit stations, workplacesand to a lesser extent grocery and pharmacy (panels (a) and (b) of Figure 6). For bothpolicies, the magnitude of these drops is around 5 percentage points per unit level of policyintensity. These findings are consistent with the fact that attending public events and privategatherings generate spillover effects on various activities outside the homeplace. Conversely,these policies have significantly increased time spent at home.As reported in Subsection 5.1, the set of interventions with the second strongest reduc-tions on subsequent infections were school and workplace closures . Figure 7 shows thatthese policies do change the mobility trends associated with crowded places, such as retailand recreation, transit stations and workplaces. Again, this can be explained by the fact thatclosing schools and workplaces generate spillover effects on other activities. The sharpnessand the magnitude of the mobility decreases is much stronger in workplace than in schoolclosures, with a unit level increase in the intensity of the policy leading to a stable declineof up to about 7-8 and 2-3 percentage points, respectively. This difference is also consistentwith the following observations. First, workers generally have access to more mobility pat-terns and activities than pupils. Second, while pupils staying at home might constrain themobility of one parent, closing workplaces affects the mobility of all adults working in thehousehold.
Stay-at-home requirements (panel (a) of Figure 8) result in large drops in all populationmobility patterns at the time when they were introduced, a fact which is consistent with thereversal of the increasing trend of new cases reported in Subsection 5.1. Finally, in line withthe results obtained for infection cases, internal mobility restrictions (panel (b) of Figure8) have a similar impact on all mobility patterns as home confinement, though neither as16harp, nor as strong. The magnitude of these effects are about half the size compared tostay-at-home requirements.We conclude this section with three remarks. First, mobility patterns do not exhibitmuch in the way of anticipation effects once we control for confounding NPIs. This suggeststhat it is policies affecting mobility patterns and not that policies ex-post responding to defacto changing mobility patterns in the population. It is worth noting that estimates ignoringconcurrent NPIs would have led to a completely different conclusion; as shown in FiguresA.3 to A.6 in Appendix A, for several interventions mobility patterns seem to respond beforepolicies are in place. Second, we observe a spike in movements to groceries and pharmaciesprior to the introductions of several NPIs, such as public transport and workplace closure,as well as stay-at-home and internal movement restrictions. This is consistent with thewidely reported runs on the shelves in anticipation of lockdowns, concerns about imminentshortages, as well as with inadvertent signaling from these interventions about the threatlevel of the pandemic. Again, we are able to detect these mobility patterns only when weaccount for confounding policies. In light of the fact that we control for the state of theepidemic by using lags of total confirmed cases, this result is robust and shows the strengthof our model. Third, it appears that the decline in mobility patterns is stable towards thelast days of analysis, suggesting that compliance does not decline over time, at least withinthe 35-day window of our study.
In this section, we expand on how the observed variation in the effects of NPIs on theincidence of COVID-19, reported in Section 5.1, can be understood by the way in whichthey affect various mobility patterns across places, reported in Section 5.2, which differ ina number of characteristics as they pertain to epidemiology, as well as in their time-useintensity. We thus provide a consistent framework for our results.17irst, the degree to which restricting mobility to different places is expected to affect newinfections depends on several characteristics of these places, where the most important are thefollowing: i) numerosity, ii) density, iii) social norms, iv) geographical range and v) trackingability. For example, more numerous and dense places, such as large private gatherings andpublic events, are more likely to contribute to new infections because the two-meter safesocial-distancing rule is more likely to be violated there than say in parks. However, placeswith similar density can be conducive to different behavior types due to social norms; forexample, in a soccer game, where there are large numbers of people densely brought together,there are different norms of accepted behavior compared to the regulated environment of aworkplace. Furthermore, places such as schools vs. transit stations, or public events, canhave different epidemiological range. For example, an infection at school has a range ofperhaps a couple of kilometers (students reside close to their schools), while in the case of asoccer game it might be several kilometers and even cross country borders. Finally, placesdiffer in terms of how easy it is to trace an infection back whenever it occurs, which isimportant because tracking contains the spread of the virus. For example, an incident ata workplace can be announced immediately to employees and an ad hoc lockdown can beprobably enforced at the same time, while an infection which occurs at a transit station isimpossible to trace back or treat with a local lockdown.Second, places differ from a time-use perspective. Based on time-use surveys on howpeople spend their time in everyday life, for example, European adults in selected countriesbetween the ages of 20 and 74 years old, spend on average on a daily basis: i) 15 hours athome preparing meals, sleeping, and on household activities, ii) a little less than 3 hoursat work, which has mostly a large workplace component, iii) a little more than 1 hourtraveling and commuting and iv) about 4 and a half hours on other activities includingleisure (recreation, parks, home) and shopping (retail, groceries and pharmacies). These https://ec.europa.eu/eurostat/documents/3930297/5953614/KS-58-04-998-EN.PDF cancellation of public events , and to a lesser extent restrictions onprivate gatherings , which are seen to lead to a large reduction in new infections, are inter-ventions that reduce exposure to numerous and dense locations, where contact tracing isdifficult, and can have a large epidemiological range within and across countries. Similarly, stay-at-home requirements , workplace and school closures reduce activities away from homeand lead to significant reductions in the incidence of new infections, which nevertheless arenot as large as for public events and private gatherings, possibly because of the differencesin numerosity, density and ability to trace new infections in these environments.On the other hand, although restrictions on internal movements reduce mobility acrosscities and regions, they impact the spread of the disease in a less pronounced way. Thisis consistent with the fact that these restrictions are not clearly linked to places with highdensity, and their potential to slow down new infections by restricting geographical mobility isreduced, once other policies such as workplace closures and restrictions on private gatheringsare in place. Furthermore, public transport closures were introduced on average at a timewhere demand for traveling and commuting has declined due to other restrictions in placesuch as workplace closures, which can explain both their limited impact on mobility and onreducing new infected cases.Finally, the limited impact of international travel controls , although they were imposedrelatively early by many countries, is likely explained by the lack of stringency of the con-19rols. If countries have banned all international travel soon after the outbreak in China, itwould have certainly be an effective measure to seal the country from the virus. However,because most countries did not introduce such bans before the virus has started spreadingdomestically, or they did introduce some restrictions but not complete bans, those restric-tions had a limited impact on mobility and could only reduce new imported infections butnot contain the spread of the virus. The COVID-19 pandemic impacts societies and economies in multiple and dramatic ways.The exact extent of this impact in economic and social terms is certainly going to remaina topic of interest in the time ahead. In this paper, we develop a multiple events modelto study the effect of lockdown policies on the incidence of new infections and on mobilitypatterns exploiting variation in the type, timing and intensity of confinement policies across135 countries. The key contributions of the paper are twofold: i) we model the dynamiceffects of each policy on the incidence of new infections accounting for concurrent policies,while in line with the standard SIR model, we specify future infections (incidence) as afunction of past cases (prevalence), as well as a number of risk related characteristics, suchas GDP per capita, population, population density and urbanization rates, all of which enrichthe exposure to risk of infection with heterogeneity within and between countries and ii) welink the effect of NPIs on new infections through their impact on mobility patterns.Our findings establish that cancelling public events and enforcing restrictions on privategatherings followed by school closures, which reduce mobility patterns in numerous anddense locations, each with their own particular behavioral norms, have the largest effecton curbing the pandemic in terms of statistical significance and levels of effect. They arefollowed by workplace and stay-at-home requirements, which also reduce activities away from20ome and lead to significant reductions in the incidence of COVID-19, which neverthelessare not as large as for public events, private gatherings and school closures, possibly becauseof the differences in numerosity, density and the ability to trace new infections in theseenvironments. Instead, restrictions on internal movement, public transport closures andinternational travel controls do not lead to a significant reduction of new infections. Thelimited impact of travel controls, although imposed relatively early in many countries, islikely explained by their lack of stringency allowing the virus to cross borders.Our econometric framework is suitable for the study of dynamic effects with multipleevents, which can be applied in many settings. A natural one is the upcoming exit strategiesfrom the lockdowns, which we will turn to next.
References
Amy H Auchincloss, Samson Y Gebreab, Christina Mair, and Ana V Diez Roux. A review ofspatial methods in epidemiology, 2000–2010.
Annual review of public health , 33:107–122,2012.Robert J Barro. Non-pharmaceutical interventions and mortality in us cities during the greatinfluenza pandemic, 1918-1919. Technical report, National Bureau of Economic Research,2020.Xiaohui Chen and Ziyi Qiu. Scenario analysis of non-pharmaceutical interventions on globalcovid-19 transmissions. arXiv preprint arXiv:2004.04529 , 2020.Samuel Engle, John Stromme, and Anson Zhou. Staying at home: mobility effects of covid-19.
COVID Economics , (2020-4), 2020.Stephen Eubank, Hasan Guclu, VS Anil Kumar, Madhav V Marathe, Aravind Srinivasan,21oltan Toroczkai, and Nan Wang. Modelling disease outbreaks in realistic urban socialnetworks.
Nature , 429(6988):180–184, 2004.Song Gao, Jinmeng Rao, Yuhao Kang, Yunlei Liang, and Jake Kruse. Mapping county-levelmobility pattern changes in the united states in response to covid-19.
Available at SSRN3570145 , 2020.Pietro Garibaldi, Espen R Moen, and Christopher A Pissarides. Modelling contacts andtransitions in the sir epidemics model.
Covid Economics Vetted and Real-Time Papers,CEPR , 2020.M Gonzalez-Eiras and D. Niepelt. On the optimal ”lockdown” during an epidemic.
CovidEconomics Vetted and Real-Time Papers, CEPR , 2020.Michael Greenstone and Vishan Nigam. Does social distancing matter?
University ofChicago, Becker Friedman Institute for Economics Working Paper , (2020-26), 2020.Thomas Hale, Samuel Webster, Anna Petherick, Toby Phillips, and Beatriz Kira. Oxfordcovid-19 government response tracker.
Blavatnik School of Government , 2020.Martin Huber and Henrika Langen. The impact of response measures on covid-19-relatedhospitalization and death rates in germany and switzerland, 2020.Christopher I Jarvis, Kevin Van Zandvoort, Amy Gimma, Kiesha Prem, Petra Klepac,G James Rubin, W John Edmunds, CMMID COVID-19 working group, et al. Quantifyingthe impact of physical distance measures on the transmission of covid-19 in the uk.
BMCMed , 18(124), 2020.Omer Karin, Yinon M Bar-On, Tomer Milo, Itay Katzir, Avi Mayo, Yael Korem, BoazDudovich, Eran Yashiv, Amos J Zehavi, Nadav Davidovich, et al. Adaptive cyclic exitstrategies from lockdown to suppress covid-19 and allow economic activity. medRxiv , 2020.22illiam Ogilvy Kermack, A. G. McKendrick, and Gilbert Thomas Walker. A contributionto the mathematical theory of epidemics.
Proceedings of the Royal Society of London.Series A, Containing Papers of a Mathematical and Physical Character , 115(772):700–721, 1927. doi: 10.1098/rspa.1927.0118. URL https://royalsocietypublishing.org/doi/abs/10.1098/rspa.1927.0118 .Henrik Kleven, Camille Landais, and Jakob Egholt Søgaard. Children and gender inequality:Evidence from denmark.
American Economic Journal: Applied Economics , 11(4):181–209,2019.Shengjie Lai, Nick W Ruktanonchai, Liangcai Zhou, Olivia Prosper, Wei Luo, Jessica RFloyd, Amy Wesolowski, Chi Zhang, Xiangjun Du, Hongjie Yu, et al. Effect of non-pharmaceutical interventions for containing the covid-19 outbreak: an observational andmodelling study. medRxiv , 2020.Wladimir Lyra, Jose Dias do Nascimento, Jaber Belkhiria, Leandro de Almeida, Pedro PauloChrispim, and Ion de Andrade. Covid-19 pandemics modeling with seir (+ caqh), socialdistancing, and age stratification. the effect of vertical confinement and release in brazil. medRxiv , 2020.Kok Yew Ng and Meei Mei Gui. Covid-19: Development of a robust mathematical modeland simulation package with consideration for ageing population and time delay for controlaction and resusceptibility, 2020.Weijie Pang. Public health policy: Covid-19 epidemic and seir model with asymptomaticviral carriers, 2020.Christopher A Pissarides.
Equilibrium unemployment theory . MIT press, 2000.Shovonlal Roy. Covid-19 pandemic: Impact of lockdown, contact and non-contact transmis-sions on infection dynamics. medRxiv , 2020.23ntonio Scala, Andrea Flori, Alessandro Spelta, Emanuele Brugnoli, Matteo Cinelli, WalterQuattrociocchi, and Fabio Pammolli. Between geography and demography: Key interde-pendencies and exit mechanisms for covid-19.
Available at SSRN 3572141 , 2020.Rajesh Singh and R Adhikari. Age-structured impact of social distancing on the covid-19epidemic in india. arXiv preprint arXiv:2003.12055 , 2020.24 igures . . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
International travel controls . . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
Cancel public events . . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
School closure . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
Workplace . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
Restrictions on internal movement . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
Public transport closure . . . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
Stay-at-home requirements . . . D en s i t y -80 -60 -40 -20 0 20 40 60 80 100 Country means
Restrictions on gatherings
Figure 1: Lockdown Policies - Days after first COVID-19 case each policy was introduced.25 . . . . . a ft e r . . . be f o r e -80 -60 -40 -20 0 20 40 60 80 country means before after Retail and Recreation . . . . . a ft e r . . . . . be f o r e -80 -60 -40 -20 0 20 40 60 80 country means before after Grocery and Pharmacy . . . . a ft e r . . . be f o r e -80 -60 -40 -20 0 20 40 60 80 country means before after Parks . . . a ft e r . . . . be f o r e -80 -60 -40 -20 0 20 40 60 80 country means before after Transit Stations . . . a ft e r . . . . be f o r e -80 -60 -40 -20 0 20 40 60 80 country means before after Workplaces . . . . a ft e r . . . be f o r e -80 -60 -40 -20 0 20 40 60 80 country means before after Residential Areas
Figure 2: Google Mobility Patterns - Densities before and after first COVID-19 case.26 . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
International travel controls - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Public transport closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Cancel public events - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on gatherings - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
School closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplace closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Stay-at-home requirements - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on internal movement
Note: Data from Hale et al. (2020), European CDC and own calculations
Figure 3: Effects of lockdown policies on
COVID-19 confirmed new cases (3-day movingaverage, in logs) without concurrent policy controls.27 . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
International travel controls - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Public transport closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Cancel public events - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on gatherings - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
School closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplace closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Stay-at-home requirements - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on internal movement
Note: Data from Hale et al. (2020), European CDC and own calculations
Figure 4: Effects of lockdown policies on
COVID-19 confirmed new cases (3-day movingaverage, in logs) with controls for concurrent policies.28 - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) International travel controls - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Public transport closure
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure 5: Effects of international travel controls (panel a) and closure of public trans-portation (panel b) on Google mobility patterns.29 - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) Cancel public events - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Restrictions on gatherings
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure 6: Effects of public events cancellations (panel a) and restrictions on gather-ings (panel b) on Google mobility patterns. 30 - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) School closure - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Workplace closure
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure 7: Effects of school (panel a) and workplace (panel b) closures on Google mobilitypatterns. 31 - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) Stay-at-home requirements - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Restrictions on internal movement
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure 8: Effects of stay-at-home requirements (panel a) and restrictions on internalmobility (panel b) on Google mobility patterns.32
Appendix - Figures - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
International travel controls - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Public transport closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Cancel public events - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on gatherings - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
School closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplace closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Stay-at-home requirements - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on internal movement
Note: Data from Hale et al. (2020), European CDC and own calculations
Figure A.1: Effects of lockdown policies on
COVID-19 confirmed new cases (in logs) withoutcontrolling for concurrent policies. 33 . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
International travel controls - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Public transport closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Cancel public events - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on gatherings - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
School closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplace closure - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Stay-at-home requirements - . - . - . - . - . . . . P r opo r t i ona t e c hange o f ne w c a s e s -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Restrictions on internal movement
Note: Data from Hale et al. (2020), European CDC and own calculations
Figure A.2: Effects of lockdown policies on
COVID-19 confirmed new cases (in logs) con-trolling for concurrent policies. 34 - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) International travel controls - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Public transport closure
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure A.3: Effects of international travel controls (panel a) and public transportationclosure (panel b) on Google mobility patterns without concurrent policy controls.35 - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) Cancel public events - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Restrictions on gatherings
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure A.4: The effect of public events cancellations (panel a) and restrictions ongatherings (panel b) on Google mobility patterns without concurrent policy controls.36 - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) School closure - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Workplace closure
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure A.5: Effects of school (panel a) and workplace (panel b) closures on Google mobilitypatterns without concurrent policy controls. 37 - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (a) Stay-at-home requirements - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Retail and Recreation - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Grocery and Pharmacy - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Parks - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Transit Stations - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Workplaces - - - - - - - - - - P e r c en t age po i n t c hange -20 -15 -10 -5 0 5 10 15 20 25 30 35 Days since policy
Residential Areas (b) Restrictions on internal movement
Note: Data from Hale et al. (2020), Google Community Mobility Reports and own calculations
Figure A.6: Effects of stay-at-home requirements (panel a) and restrictions on inter-nal mobility (panel b) on Google mobility patterns without concurrent policy controls.38
Appendix - Sample of countries
Estimations for COVID-19 cases are based on a sample of 135 countries • Afghanistan, Angola, Argentina, Australia, Austria, Bahrain, Bangladesh, Barbados, Belgium, Be-lize, Bolivia, Bosnia and Herzegovina, Botswana, Brazil, Bulgaria, Burkina Faso, Cameroon, Canada,Cape Verde, Chile, Colombia, Costa Rica, Croatia, Czech Republic, Denmark, Dominican Republic,Ecuador, Egypt, El Salvador, Estonia, Finland, France, Gabon, Germany, Ghana, Greece, Guatemala,Honduras, Hong Kong, Hungary, India, Indonesia, Iraq, Ireland, Israel, Italy, Jamaica, Japan, Jordan,Kazakhstan, Kenya, Kuwait, Kyrgyz Republic, Laos, Lebanon, Libya, Luxembourg, Malaysia, Mali,Mauritius, Mexico, Moldova, Mongolia, Mozambique, Myanmar, Namibia, Netherlands, New Zealand,Nicaragua, Niger, Nigeria, Norway, Oman, Pakistan, Panama, Papua New Guinea, Paraguay, Peru,Philippines, Poland, Portugal, Puerto Rico, Qatar, Romania, Rwanda, Saudi Arabia, Serbia, Singa-pore, Slovak Republic, Slovenia, South Africa, South Korea, Spain, Sri Lanka, Sweden, Switzerland,Tanzania, Thailand, Trinidad and Tobago, Turkey, Uganda, United Arab Emirates, United Kingdom,United States, Uruguay, Vietnam, Zambia, Zimbabwe. • No school intervention: Burundi, Nicaragua. • No workplace intervention: Brunei, Burundi, Eswatini, Mozambique, Nicaragua, Niger, Tanzania. • No events intervention: Burundi, Nicaragua, Sweden. • No transport intervention: Australia, Brunei, Bulgaria, Burundi, Canada, Chile, Czech Republic, Do-minica, Estonia, Germany, Hong Kong, Iceland, Japan, Malawi, Malaysia, Mali, Mauritania, Mozam-bique, Namibia, Nicaragua, Niger, Panama, South Korea, Sweden, Switzerland, Tanzania, Zambia. • No mobility intervention: Burundi, Hong Kong, Iceland, Malawi, Mozambique, Nicaragua, Tanzania. • No travel intervention: Luxembourg, United Kingdom. • No home intervention: Brunei, Burundi, Cameroon, Iceland, Nicaragua, Norway, Sweden, Tanzania.