key: cord-0800738-33y8ihtr
authors: Zhang, Haoran; Li, Peiran; Zhang, Zhiwen; Li, Wenjing; Chen, Jinyu; Song, Xuan; Shibasaki, Ryosuke; Yan, Jinyue
title: Epidemic versus economic performances of the COVID-19 lockdown in Japan: A Mobility Data Analysis
date: 2021-10-22
journal: Cities
DOI: 10.1016/j.cities.2021.103502
sha: e65fec2545ef76f0ba80905004b5b33b278b3d96
doc_id: 800738
cord_uid: 33y8ihtr

Lockdown measures have been a “panacea” for pandemic control but also a violent “poison” for economies. Lockdown policies strongly restrict human mobility but mobility reduce does harm to economics. Governments meet a thorny problem in balancing the pros and cons of lockdown policies, but lack comprehensive and quantified guides. Based on millions of financial transaction records, and billions of mobility data, we tracked spatio-temporal business networks and human daily mobility, then proposed a high-resolution two-sided framework to assess the epidemiological performance and economic damage of different lockdown policies. We found that the pandemic duration under the strictest lockdown is less about two months than that under the lightest lockdown, which makes the strictest lockdown characterize both epidemiologically and economically efficient. Moreover, based on the two-sided model, we explored the spatial lockdown strategy. We argue that cutting off intercity commuting is significant in both epidemiological and economical aspects, and finally helped governments figure out the Pareto optimal solution set of lockdown strategy.

individual-level activity, instead of a macro statistical number, which can afford detailed and spatialized simulation result; 2. took the human mobility as a bridge, precisely simulate both economic loss propagation and epidemical transmission. 3. can work out a Pareto optimal strategy, consisting of a series of policy-set, affording more flexibility than the past parameterized approaches. Our results found that the pandemic duration under the strictest lockdown is less about two months than that under the lightest lockdown, which makes the strictest lockdown characterize both epidemiologically and economically efficient. Moreover, based on the twosided model, we explored the spatial lockdown strategy. We argue that cutting off intercity commuting is significant in epidemiological and economic aspects and finally helped governments figure out the Pareto optimal solution set of Lockdown strategy. This study will offer fundamental support for guiding regional and national governments in designing health, social, and economic policies.

Interdisciplinary models are critical tools for anticipating, predicting, and responding to this pandemic crisis-including its biological and economic aspects-and also give governments theoretical support for coping with this thorny problem (Barton et al., 2020; Dong et al., 2020) .

Most of the past related studies usually began with an in-depth interpretation of lockdown policies from a one-sided dedicated perspective -merely from epidemics or economics.

From the perspective of epidemics, most of these studies employed macro-scale statistical data, province/city-level mobility data, and infected numbers to simulate/evaluate the effectiveness of government J o u r n a l P r e -p r o o f 4 Kucharski et al. 2020 To analysis the effect of government interventions (travel restrictions) on epidemical transmission A modified SEIR model Macro statistical data Country-level mobility data Country-level infected number × √ × Wells et al. 2020 To analysis the effect of government interventions (travel restrictions) on epidemical transmission

Monte Carlo simulation Macro statistical data Country-level mobility data Country-level infected number × √ × Bonaccorsi et al. 2020 To analysis the effect of government interventions (travel restrictions) on economical loss

Macro statistical data (People's situation provided by Facebook)

Country-level infected number √ × × Huang et al., 2020 To analysis the effect of government interventions (travel restrictions) on economical loss However, for the governments, the possibilities of pandemic scenarios are challenging to estimate, making it hard to extract academic support from one-sided analyses for policymaking directly. Therefore, it is essential to further reveal research gaps in our understanding of the dual character of lockdown policy to develop novel solutions to manage the spread of disease and lessen its impacts. Some researchers tried to solve this bi-sided complex by taking mobility as the bridge between pandemic scenarios and economic cost/loss. Aiming to find an optimal lockdown policy for a planner who wants to control the fatalities of a pandemic while minimizing the output costs of the lockdown, Alvarez (2020) employed a SIR epidemiology model and a linear economy model to formalize a simple optimal control model. Taking the entire country as a whole, they conclude with a parameterized optimal solution for lockdown measures. Karin et al.

(2020) employed SEIR models and stochastic network-based models for epidemics, an experience-based analysis for economics, and concluded that a cyclic schedule of 4-day work and issues based on macro statistical data instead of a detailed investigation of a metropolitan area.

Second, for the current studies, which aim to understand the dual character of epidemics and economics, they only tried to find the relationship by regression fitting instead of simulation in physics. Lastly, although there have been studies trying to employ individual-level mobility data (mainly collected from part of mobile phone users), no attempt has been conducted to map the virtual data (part of population) to the real world (total population).

Intensity and duration characterize the lockdown's impacts on epidemiological performance This study mainly focused on dissecting epidemiological performances and economic impacts of different lockdown policies and devising the optimal policy mix for the government to balance lockdown's pros and cons. Figure 1a displays the modeling mechanism. The framework has four essential parts: lockdown scenario setting, epidemiological performance assessment, economic damage evaluation, and policy mix suggestion. Concerning the lockdown scenario setting, we figured out three lockdown clusters according to the lockdown intensities of metropolitan areas worldwide. Moreover, to comprehensively illustrate the temporal impacts of different lockdown policies on metropolitan areas, we set two lockdown duration determination method scenarios:

fixed schedule and dynamic adjusting. The main inputs of the model are individual mobility data and financial transaction data. Concerning the epidemiological performance assessment, from the individual mobility data, we can observe the individual daily trajectories and then detect their travel purposes by fusing trajectories data with the urban point of interest information (POI) data.

Based on the real-world mobility changes in different travel purposes under different lockdown policies, we adopted a sampling method to generate individual mobility data and construct an individual physical contact network under different lockdown policy cases. Then, we utilized a compartmental model in epidemiology, a grid-based SEIR model to simulate the coronavirus spread process following human mobility. It also pointed out the case number 

This study employed various datasets: big GPS record data for human mobility tracing, metropolitan POI data for travel purpose labeling, financial transaction records and firm information data for economic modeling, and some other data helping analysis:

Human Mobility Data. To model real-world human mobility used for epidemic simulation, we collected a GPS log dataset anonymously from about 1.6 million real mobile-phone users in Japan over three years (from August 1st, 2010 to July 31st, 2013). This dataset contains about 30

billion GPS records, more than 1.5 TeraBytes. The data collection was conducted by a mobile operator (i.e., NTT DoCoMo, Inc.) and a private company (i.e., ZENRIN DataCom Co., Ltd.) under the consent of mobile phone users. These data were processed collectively and statistically in order to conceal private information such as gender or age. By default, the positioning function on the users' mobile phones is activated every 5 minutes, so their positioning data (i.e., latitude and longitude) are uploaded onto the server. However, several factors, such as loss of signal or low battery power, are affected by data acquisition. In addition, when a mobile phone user stops at a location, the positioning function of his mobile phone is automatically turned off to save power. This study selected Greater Tokyo Area (including Tokyo City, Kanagawa Prefecture, Chiba Prefecture, and Saitama Prefecture) as the target area of epidemic simulation. A user ID will be selected as our experimental data if 80% of that user's trajectory points locate in Greater Tokyo Area. After this, we can obtain 145,507 users' trajectories in total. 

Sampling Trajectory Under Different Lockdown Scenarios. To simulate mobility change under possible three situations during Covid-19, we simulated the human mobility data to fit the mobility before April 8th, after April 8th, and under a fully restricted situation. In every scenario, we used the Google community mobility report dataset of COVID-19 in Japan to acquire the reduction rate of each type of activity (home places, workplaces, park places, and other activities). Moreover, we randomly selected different ratios of mobile users and every mobile user's different workdays and replaced them with home places to simulate more approximated situations with these three scenarios by Gibbs sampling.

Under a specific lockdown strategy, people's mobility behavior changes correspondingly. Two aspects could simulate this change: On one hand, we employed an existing grid-based epidemic model to capture the epidemic transmission and to scale it up by 'smallworld' model (Lin et al., 2021) , on the other hand, a propagation-tracking model captures the economic loss. Combining these two results (in epidemics and economics), we can evaluate the strategy and finally find the Pareto optimal solution.

J o u r n a l P r e -p r o o f 

where , are the 0 to 1 values representing the scale of firm and (we use to represent the resident customers of each firm); is the amount of transactions between them before the iteration, and ′ is the decreased transaction volume.

Firm scale shrinkage. In return, the scale of a firm is recalculated according to the decreased revenue, which is formulated as

where ′ is the shrunken scale of firm . When the firm-scale is lower than a certain threshold and with the same ratio reduction in their salary (however, according to the labor law, the salary reduction rate cannot be higher than 40%). Based on the average wage of each industry, , we calculated the average salary of the employee of firm i , :

Then we could update the income level of each administrative region r, . Based on the survey results of consumption structure in different income-level groups, we reconstructed the consumption structure of each region. We figured out the resident consumption reduction on each industry's product. Finally, we could estimate out the of the residential customers, and take this into the iteration process of tracking the propagation of the economic impact of lockdown. When the iteration time excesses the time threshold , we output the result.

For each prefecture of Greater The retail & recreation line represents mobility trends for places like restaurants, cafes, shopping centers, theme parks, museums, libraries, and movie theaters. Groceries line represents mobility trends for places like grocery markets, food warehouses, farmers markets, specialty food shops, drug stores, and pharmacies. The Parks line represents mobility trends for places like national parks, public beaches, marinas, dog parks, plazas, and public gardens. The transit stations line represents mobility trends for places like public transport hubs such as subway, bus, and train stations. Workplaces line represents changes in mobility trends for places of work. Furthermore, the Residential line represents changes in mobility trends for places of residence. The transparent squares represent the confidence intervals. The blue and orange light lines represent the time when the retail and groceries mobilities reduce. The dotted lines represent the average decreased values. The transparent boxes represent the confidence intervals.

We found that metropolitan activity characteristics are quite different under different lockdown (Figure 3b ). We found the significant difference between the medium and hard lockdowns is that hard lockdown more significantly limits activity in the grocery category, implying worse economic damage and a more significant decrease in mobility. Additionally, we pointed out the highly similar activity change characteristics in different metropolitan areas when local governments implement medium or hard lockdown policies. The results show that we can confidently utilize the mean value of mobility changes in medium or hard lockdown to approximately simulate the same situation of a metropolitan area that hard or medium policies have never controlled. However, when it comes to soft-lockdown policy, there are huge differences between different metropolitan areas. Since softlockdown policies do not adopt many mandatory constraints, the mobility change will be affected more by different cultures, residents' habits, etc., compared with medium and hard-lockdown policies. The evidence behind this statement is our finding that the metropolitan areas which adopt soft-lockdown policies in the same country show highly correspondence with human mobility changes. In contrast, metropolitan areas in different countries show much more difference in human mobility behavior.

Here, we quantified huge gaps in epidemiological performances of different lockdown policies.

We mainly focused on a representative metropolitan case-the Greater Tokyo Area, the most populous metropolitan area and the largest metropolitan economy globally-to dissect the relationship among lockdown implementation, mobility reduction, and epidemiological performance. Based on COVID-19 Community Mobility Reports from Google, POI data, and 1.3 billion metropolitan trajectory data, we simulated human mobile behaviors under different lockdown policies, as shown in Figure 4a . It can be seen that the stricter the lockdown policy, the less the human mobility, especially in the metropolitan center area and the main commuting channel areas. The results show that a lockdown policy can effectively reduce mobility intensity, cut people's physical contact network, and theoretically mitigate the virus spread rate. The dark color represents the high mobility intensity, and the light represents low intensity. The number in the legend represents the percentage of people who have ever stayed at (or passed by) a specific cell. For example, when one cell's color is the darkest one, that means more than 21% of the total population have ever stayed at (or passed by) this cell during the lockdown period. If the value is less than 0.1%, we will not show the cell's color anymore. The blue circle represents the Greater Tokyo Area center area where also has a high resident population density. The three yellow circles represent the main commuting channel areas from the surrounding cities to the Tokyo center. It is clear that with the increase of the strengthening of lockdown policy, the mobility in the center area of the Greater Tokyo Area trends to be sparser. b. Epidemiological performances of different lockdown policies. Different colors represent different lockdown policies or baselines. The ending color point of each curve represents the pandemic ending time under the lockdown policy when the TCN does not increase anymore. The value in the bracket and nearby the ending color point represents the ending time and TCN of each policy. For example, the value of the medium lockdown (87, 8.68) means the pandemic duration is 87 days, and TCN is 8.68 thousand people.

Here, we analyzed the economic damage of one-month lockdown scenarios at different policy intensity levels (Figure 5a ). We found an approximate arithmetic progression relationship among the degrees of economic damage associated with different lockdown policies. Due to local store closures and inactive consumption, a locked-down metropolitan area suffers severe economic J o u r n a l P r e -p r o o f damage. In the Greater Tokyo Area, value-added loss reaches 1.9%, 6.2%, and 9.7% under the soft, medium, and hard lockdown policies. Additionally, as the center of the national economy, the shutdown of the metropolitan area also brings a huge negative effect on other economic regions. Therefore, regarding the whole country, the economic damage could reach a high level.

Specifically, the Greater Tokyo Area lockdown causes 1.1% value-added loss to Japan as a whole under soft lockdown, 3.7% under medium lockdown, and 5.9% under hard lockdown. In addition to the overall loss, we also indicated the economic damage in different industries (Figure 5b) . The results show that most of the metropolitan economies experience a downturn. The service industry, real-estate business, and transportation/communication industry are the top three most severely affected sectors. Additionally, as the intensity of lockdown measures increases, upstream industries are gradually impacted by economic chain reactions.

Considering that the government may extend or diminish the lockdown duration according to changes in the TCN increase, we assumed that the government would lift the lockdown policy once the TCN does not increase anymore. Under this assumption, we simulated economic damage again under a dynamic adjusting scenario (Figure 5c ). The results show that, although the hard lockdown hits the economy most severely when considering the lockdown duration, hard lockdown is the most economical measure, compared with the soft and medium lockdown policies. The lockdown duration trends to be more sensitive to the value-added loss than lockdown intensity since the economic impacts' propagation grows exponentially during the medium-term of the damage,. Therefore, with a shorter implementation duration, hard lockdown caused lighter economic damage (a 1.7% annual value-added loss) than soft lockdown, which caused up to 11.2% annual value-added loss. Moreover, suppose the government makes a risky decision to lift the lockdown policy as early as the TCN increase trends to less than 100 per day.

In that case, stricter lockdown policies also have a larger economy saving space than do lighter policies.

J o u r n a l P r e -p r o o f 

The above result proves that hard lockdown is the optimal measure in epidemiological and economic aspects when the government dynamically adjusts the lockdown duration. Meanwhile, in addition to adjusting the duration, the government also can optimize the policy mix by implementing different levels of lockdown in different regions. Here, as shown in Figure 6a , we figured out specific strategies concerning the regional lockdown policy mix for the Greater Tokyo Area (there are four administrative regions, with one central city and three surrounding satellite city groups, which are Tokyo city, Chiba prefecture, Saitama prefecture, and Kanagawa prefecture). We found the Pareto optimal solution set of lockdown strategy includes two policy mixes: the first one is a whole-hard-lockdown mode, which has the least number of cases (H-H Although the "besieging" model caused 1.6 thousand cases more than did the whole-hard-J o u r n a l P r e -p r o o f lockdown mode (Figure 6b ), the former strategy minimizes the economic damage from 1.7% to 1.1% (Figure 6c ). On the contrary, the whole-soft-lockdown mode is the worst strategy in terms of epidemiological performance (S-S-S-S in Figure 6a ); and the "volcano" mode, i.e.

implementing medium lockdown in Tokyo city and implementing soft lockdown in the surrounding satellite cities, causes the most significant economic damage (M-S-S-S in Figure 6a ).

These results reveal that cutting off intercity commuting from surrounding satellite cities to the central city is significant in mitigating the virus pandemic. When considering reducing economic damage, moderately easing the lockdown degree in the central city while strictly controlling the external input is also an optimal strategy. The exhaustive examples of the policy mix. Each point represents a policy mix plan, and the data label of the point shows the description of the policy mix. S represents soft lockdown, M is for medium lockdown, and H is for hard lockdown. The first letter represents the policy level in Tokyo city, the second is for Chiba prefecture, the third is for Saitama prefecture, and the last is for Kanagawa prefecture. For example, H-M-S-S means Tokyo city implements hard lockdown, Chiba prefecture implements medium lockdown, Saitama and Kanagawa implement soft lockdown. The orange boxes show the confidence intervals for both epidemiological performance and economic damage. The red dotted line connects the Pareto optimal policy mix solutions. The maps show the implementations of the optimal policy mixes. As the same, red is for hard lockdown, orange is for medium lockdown, and green is for soft lockdown. b.

Epidemiological performances of different lockdown policies in the Pareto optimal solution set. Different colors represent different lockdown policies. c. The annual value-added losses under dynamic adjusting lockdown policies. but high-level in surround) mode are two optimal lockdown policy strategies in terms of optimal policy. The fundamental mechanism of the above phenomenonthe hard lockdown is the most economical measure compared with the soft and medium policies-is the duration period, since the hard lockdown can quickly terminate the pandemic. Alvarez et al. (2020) employed an SIR epidemiology model and a linear economy model to formalize a simple optimal control model. Taking America as a case study, they concluded that the optimal policy prescribes a severe lockdown beginning two weeks after the outbreak, covers 60% of the population after a month, and is gradually withdrawn, covering 20% of the population after three months. Welfare under the optimal policy with testing is higher, equivalent to a onetime payment of 2% of GDP. As a comparison, our optimal lockdown strategy can reach a 1.1% economic loss. Although we took different regions as the case study, we believe that these results are comparable since the total population of Japan and America is of the same order of magnitude. Besides, Karin et al. (2020) employed SEIR models and stochastic network-based models for epidemics, concluded that a cyclic schedule of 4-day work and 10-day lockdown, or similar variants, can prevent a resurgence of the epidemic while providing part-time employment, which will subside the economic loss. However, they did not give any quantitively economic analysis, only analyzed it based on experiences.

In general, our proposed method shows reliability -we examined our results by ground-truth data, and it achieves high accuracy with a 5.67% relative mean deviation ( Figure 7) ; spatialization -we employed big and detailed data (including 1.3 billion metropolitan trajectory data and 1.2 million financial transaction data, etc.) for individual-level activity tracking, instead of a macro statistical number; and realityemployed the 'small-world' model (Lin et al., 2021) to map the virtual data (part of population) to the real world (total population).

J o u r n a l P r e -p r o o f Figure 7 . Ground-truth Study of the Proposed Model. We validate our epidemic simulation result with the ground-truth accumulated case number from January 24th to May 8th, 2020, in the Greater Tokyo area. There is a soft lockdown policy in Greater Tokyo Metropolis before April 8th, when the Japanese government announced the state of emergency (orange points). After April 8th, the policy came to medium lockdown (blue points). Thus, we conduct the simulation corresponding to reality. The result reaches high accuracy with a 5.67% relative mean deviation.

How to balance epidemiological performance versus economic damage caused by lockdown policies is one of the thorniest puzzles for governments during the COVID-19 pandemic and in future similar crises, if, unfortunately, they occur. Lockdown policies strongly restrict human mobility, which hence results in economic damage. A comprehensive quantificational analysis of the pros and cons of lockdown policies is enormously instructive to help governments cope with crises, save people's lives, and mitigate economic damage, but significant research gaps exist.

Based on the global COVID-19 community mobility data, we figured out three types of metropolitan lockdown policies. Then we focused on how epidemiological performance versus economic damage differs under these different policies. We dissected a case study representative in the global context, the Greater Tokyo Area, basing that study on various datasets, including 1.3 billion metropolitan trajectory data, 1.2 million financial transaction data, etc. Firstly, we simulated human mobile behavior under lockdown and then tracked the spatio-temporal virus spread process and figured out the TCN under different lockdown policies.

On the other hand, we tracked the capital flows on the end-consumer and business networks in all industries. We then simulated the negative propagation of the economic impacts of different lockdown policies. We found that the lockdown policies effectively reduce the number of cases;

specifically, lockdown policies decrease the case number by 35.29%, 87.35%, and 98.60%, respectively, under the soft, medium, and hard implementation levels. At the same time, the onemonth lockdown under these three policies causes 1.1%, 3.7%, and 5.9% value-added loss in Japan. However, it is surprising that the hard lockdown is the most economical measure when considering pandemic duration compared with the soft and medium policies since the hard J o u r n a l P r e -p r o o f lockdown can quickly terminate the pandemic. Finally, we helped the government figure out two Pareto optimal solutions for lockdown strategy. From the results, we argued that: 1. Cutting off intercity commuting from surrounding satellite cities to the central city is significant in epidemiological and economic aspects; 2. Moderately easing the lockdown degree in the central city while strictly controlling the external input is also an optimal strategy to reduce economic damage while bringing the pandemic under control.

With the support of reliable and fine-grained datasets and the proposed comprehensive analysis framework, we believe our proposed findings can offer fundamental support for guiding governments in designing economic and social policies in the context of this particular stage.

Additionally, the proposed methods are not limited to the studied COVID-19 lockdown case but are also practical for analyzing global cases of potential future waves of COVID-19 and similar crises.

Despite those strengths, shortcomings remain. Due to the limitation of data, what we have done is all about the situation in Japan. It is still unknown whether and to what extent regional differences may influence the optimal strategy. Besides, this study ignored imported people from abroad, focusing on the domestic population, ignoring some errors. In the future, we plan to expand the epidemic-economic model to a disaster-economic model and evaluate the method by data in more cities around the world.

• Employed millions of financial transaction records and billions of human trajectory records.

• Proposed a two-sided model, taking mobility as the bridge to connect epidemics and economics.

• Assess the epidemiological performance and economic damage by different lockdown policies 

A Simple Planning Problem for COVID-19 Lockdown A Simple Planning Problem for

The coronavirus and the great influenza pandemic: Lessons from the "spanish flu" for the coronavirus's potential effects on mortality and economic activity

Call for transparency of COVID-19 models

Economic and social consequences of human mobility restrictions under COVID-19

Mobile phone data analytics against the COVID-19 epidemics in Italy: flow diversity and local job markets during the national lockdown

h3: Hexagonal hierarchical geospatial indexing system

The impact of the COVID-19 pandemic on consumption: Learning from high frequency transaction data

The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak

An interactive web-based dashboard to track COVID-19 in real time

Impact of the coronavirus pandemic on the global economy -Statistics & Facts. Statista

Human mobility restrictions and the spread of the novel coronavirus (2019-ncov) in china

Transmission dynamics of the COVID-19 outbreak and effectiveness of government interventions: A data-driven analysis

Economic effects of coronavirus outbreak (COVID-19) on the world economy. Available at SSRN 3557504

Quantifying the economic impact of COVID-19 in mainland China using human mobility data

The propagation of the economic impact through supply chains: The case of a mega-city lockdown against the spread of COVID-19

Modeling the control of COVID-19: Impact of policy interventions and meteorological factors

Population flow drives spatio-temporal distribution of COVID-19 in China

Adaptive cyclic exit strategies from lockdown to suppress COVID-19 and allow economic activity

Early dynamics of transmission and control of COVID-19: a mathematical modelling study

The positive impact of lockdown in Wuhan on containing the COVID-19 outbreak in China

Small World Model for scaling up prediction result based on SEIR model

The global macroeconomic impacts of COVID-19: Seven scenarios

The Socio-Economic Implications of the Coronavirus and COVID-19 Pandemic: A Review

Impact of international travel and border control measures on the global spread of the novel 2019 coronavirus outbreak

Coronavirus disease (COVID-2019) situation reports

Large Scale Mobility Analysis: Extracting Significant Places Using Hadoop/Hive and Spatial Processing. International Conference on Knowledge, Information, and Creativity Support Systems

Modified SEIR and AI prediction of the epidemics trend of COVID-19 in China under public health interventions

EpiMob: Interactive Visual Analytics of Citywide Human Mobility Restrictions for Epidemic Control. ArXiv

Author contributions H.Z. and J.Y. designed the study. H.Z., Z.Z., and P. L. developed the algorithms and performed the analysis. H.Z., W.L., J.Y., X.S. and R.S. led the writing of the paper. H.Z. and P.L. revised the manuscript.J o u r n a l P r e -p r o o f