Tokyo's COVID-19: An urban perspective on factors influencing infection rates in a global city

Abstract

This research investigates the relationship between COVID-19 and urban factors in Tokyo. To understand the spread dynamics of COVID-19, the study examined 53 urban variables (including population density, socio-economic status, housing conditions, transportation, and land use) in 53 municipalities of Tokyo prefecture. Using spatial models, the study analysed the patterns and predictors of COVID-19 infection rates. The findings revealed that COVID-19 cases were concentrated in central Tokyo, with clustering levels decreasing after the outbreaks. COVID-19 infection rates were higher in areas with a greater density of retail stores, restaurants, health facilities, workers in those sectors, public transit use, and telecommuting. However, household crowding was negatively associated. The study also found that telecommuting rate and housing crowding were the strongest predictors of COVID-19 infection rates in Tokyo, according to the regression model with time-fixed effects, which had the best validation and stability. This study's results could be useful for researchers and policymakers, particularly because Japan and Tokyo have unique circumstances, as there was no mandatory lockdown during the pandemic.

1. Introduction

After almost three years of COVID-19 pandemic and more than 6.5 million deaths (as of February 2023), according to the World Health Organization (WHO), it is likely that the COVID-19 pandemic will come to an end soon (Adam, 2023). However, COVID-19 dramatically changed cities. It had major impacts on how people interact in cities, how policy makers understand the dynamics of urbanization, and how the quality of life is defined ontologically and epistemologically (Joiner et al., 2022). Early media outlets pointed out that urban features such as public transport, density, and recreation spaces are the main influential factors of COVID-19 spread (Florida and Mellander, 2022). Urbanization and compact development were extensively blamed as key risk factors for the spread of COVID-19 (Hamidi et al., 2020). Accordingly, many countries faced urban degradation and population loss in large cities (like London and New York) during the first months of the pandemic (Naud´e and Nagler, 2022).

COVID-19 was not the first pandemic that changed cities. Beyond the initial opinions and declarations concerning the impact of urban features on COVID-19, there are inquiries being raised about the ability of cities to prepare for and withstand pandemics. These concerns require more deliberate attention and resolution from planning researchers and policy makers. Non-pharmaceutical interventions have been advocated to reduce the pressure and worriedness during the pandemic (Lertsakornsiri et al., 2022; Hassankhani et al., 2021). Also, several literature reviews are conducted to explain how the built environment, socio-economic, health, and environmental factors have influenced the pandemic (Alidadi and Sharifi, 2022; Alam and Sultana, 2021; Teller, 2021).

Previous studies have shown that some built environment factors like land use, housing conditions, density, transport, access to facilities, and health infrastructures are critical to understanding the dynamics of COVID-19 infection in cities (Khavarian-Garmsir et al., 2021; Li et al., 2020; You et al., 2020; Kamble and Bahadure, 2021). Regarding land use and mobility, it has been discussed that a mix of different land uses increases local accessibility to daily and weekly needs, which reduces the need for travel, thereby facilitating better resilience to pandemics. However, the high density of facilities like clinics and hospitals attracts people from different parts of the city, and may increase the chance of the virus transmission (Lak et al., 2021). Open and green spaces, and their distribution patterns,are also found to be important for controlling the spread of the virus. Such spaces allow for better compliance with protection measures (Sharifi, 2022). On the micro-scale, housing conditions, including the size of the house, the number of rooms, and the per capita space, are positive predictors of COVID-19 spread (Consolazio et al., 2021; Wheaton and Kinsella Thompson, 2020). Overcrowding in housing increases the vulnerability when one person is infected (Alidadi and Sharifi, 2022). Additionally, some human factors like occupation, income, household structure, travel behaviour, age, education level, and social class were among the influential factors in explaining the variability and intensity of COVID-19 spread in cities (Kashem et al., 2021; Ali et al., 2021; Liao et al., 2021). For example, people engaged in health, transport, recreation, retail, and life-related services are often more vulnerable than others due to their physical presence and close proximity to others (Wang et al., 2021). However, some studies have shown that cities' context, method, political circumstances, and socio-cultural background are very important and affect the spread and control patterns of the pandemic. While, in the last three years, different aspects of cities and COVID-19 are investigated, empirical studies from different cases around the world can still enrich the literature.

Japan is a peculiar case study to investigate the dynamics of COVID-19. The government has implemented several measures to control the spread of COVID-19. To restrict vehicle movement, it declared a state of emergency in several areas, urging people to stay at home and avoid unnecessary travel. Educational institutes were also shut down to prevent mass gatherings. Border restrictions were implemented to limit the entry of foreign travellers. Vaccination campaigns have been rolled out, with priority given to healthcare workers and the elderly. Public awareness campaigns have emphasized the use of masks, avoiding crowded places, frequent handwashing, and other preventative measures.

On the other hand, the government never implemented a lockdown (Fukuchi et al., 2022). To avoid social and mental health consequences, even the emergency states did not take too long as these situations could have increased worriedness (Lertsakornsiri et al., 2022). Additionally, Tokyo is characterized by a high population and employment density, a high share of public transport, crowding in public spaces, and a high share of small apartments. According to the literature mentioned above, these characteristics may increase the vulnerability to the pandemic. Therefore, understanding the dynamics behind the spread of the virus in Tokyo and how urban features are associated with the severity of the pandemic could bring valuable contributions to the current literature. Furthermore, elderly people are highly concentrated in the Tokyo prefecture. The population over 65 is 3.12 million as of September 2023, which is still growing (source: https://www.metro.tokyo.lg.jp/english/index.html). Considering their vulnerability, it is an urgent task to reduce COVID-19-related risk in Tokyo.

Against this backdrop, this study has three main objectives. Initially, to find how spatially and temporally COVID-19 spread in Tokyo. Secondly, to find how urban variables are associated with COVID-19 infection rate. Finally, to investigate and explore the predictors of COVID-19 spread through different regression models. Accordingly, the research will have some methodological, theoretical, and practical contributions. Methodologically, different quantitative and spatial models are employed to go beyond the simple correlation analysis and find a validated method for variable selection. We have applied a step-wise regression model to select explanatory variables based on statistical tests, not arbitrarily. Theoretically, there are some contradictions regarding interactions between urban features and COVID-19, and our result contributes to this area. Practically, Tokyo is among the most populated, dense, transit-based, and diverse cities in the world. Our results can assist policy makers in making more resilient cities by providing them with unique insights from a global city.

2. Materials and methods

2.1. Data and variables

This study uses three main types of data to analyse the effect of built environment factors on the spread of COVID-19 in Tokyo. Built environment and socio-economic data and COVID-19 data by municipality were collected from different organizations. COVID-19 data was the daily number of positive cases in Tokyo municipalities, which the Ministry of Health, Labour and Welfare (MHLW) collected. Then we aggregated the data and converted them to monthly cases. These data cover 19 months of COVID-19 diffusion from April 2020 to October 2021 in 53 municipalities of Tokyo prefecture. The built environment and socio-economic data for our analysis were also on the municipality scale. The sources of data are mentioned in Table 1 .

Table 1.

Description of variables and their data sources.

Variable	Abb	Indicator	Definition/unit	Mean	Max	Min	Source
Density	d1	Population density	Number of people per km2	10,949.54	22,067.56	608.44	SSDSE
	d2	Residential density	Number of people per residential area (km2)	29,777.48	325,350.98	3907.53	SSDSE
	d3	FAR	Total floor area to residential area	1.13	11.68	0.15	HLS
	d4	Activity density	Number of employees per km2	9945.47	80,818.10	733.13	SSDSE
	d5	Density of schools	Number of schools to total area	3.04	9.21	0.21	SSDSE
	d6	Density of retail stores	Number per area	94.24	346.82	6.13	SSDSE
	d7	Density of restaurants	Number of restaurants per area	81.94	490.01	1.78	SSDSE
	d8	Density of all medical facilities	Clinics and hospitals per area	13.37	56.81	0.57	SSDSE
	d9	agriculture, forestry workers density	Number of occupants (of the job) per area	2.63	17.43	0.00	SSDSE
	d10	fishery workers density	Number of occupants (of the job) per area	0.10	3.72	0.00	SSDSE
	d11	mineral extraction, quarrying, sand mining workers density	Number of occupants (of the job) per area	2.02	55.23	0	SSDSE
	d12	construction workers density	Number of occupants (of the job) per area	458.22	3509.40	17.39	SSDSE
	d13	manufacturing workers density	Number of occupants (of the job) per area	595.57	3937.74	43.81	SSDSE
	d14	electricity, gas, heat, water workers density	Number of occupants (of the job) per area	24.78	258.17	0	SSDSE
	d15	information and communication workers density	Number of occupants (of the job) per area	1006.30	10,382.25	0.18	SSDSE
	d16	transportation, postal service workers density	Number of occupants (of the job) per area	366.66	2159.16	10.03	SSDSE
	d17	wholesale, retail workers density	Number of occupants (of the job) per area	2234.71	21,444.37	57.07	SSDSE
	d18	finance, insurance workers density	Number of occupants (of the job) per area	547.76	11,005.57	1.39	SSDSE
	d19	real estate, goods leasing workers density	Number of occupants (of the job) per area	401.33	3360.63	3.42	SSDSE
	d20	academic, professional, technical workers density	Number of occupants (of the job) per area	603.28	8302.83	5.24	SSDSE
	d21	accommodation, food service workers density	Number of occupants (of the job) per area	931.77	5548.09	26.64	SSDSE
	d22	life related services, entertainment workers density	Number of occupants (of the job) per area	336.43	1852.89	9.51	SSDSE
	d23	education workers density	Number of occupants (of the job) per area	369.25	2478.12	1.71	SSDSE
	d24	medical, welfare workers density	Number of occupants (of the job) per area	720.55	2025.66	59.41	SSDSE
	d25	complex service business workers density	Number of occupants (of the job) per area	31.68	296.91	1.25	SSDSE
	d26	Other services workers density	Number of people in other services rather than abovementioned per area	1226.77	12,937.31	12.26	SSDSE
Land use	lu1	Residential land use	Share of residential to total area	61.71	94.59	5.67	HLS
	lu2	Commercial land use	Share of commercial to total area	13.67	75.89	0.50	HLS
	lu3	Industrial land use	Share of industrial to total area	12.65	64.27	0	HLS
	lu4	Share of green land uses	Share of forests, wastelands, parks and paddy lands to total area	12.75	77.13	0.17	HLS
	lu5	Share of open spaces	Share of non-built-up areas to total area	0.18	0.82	0.01	HLS
Housing conditions	h1	Housing crowding	Number of rooms per residences	3.46	4.87	2.50	HLS
	h2	Residence shares of housing	Total floor area per residence	69.48	97.49	51.85	HLS
	h3	Household crowding	Number of people per household	2.08	2.60	1.61	HLS
	h4	Share of rooms	Number of people per room	0.62	0.72	0.52	HLS
	h5	Housing type - detached	Share of detached houses (1&2 story) to total houses	47.36	95.74	3.93	HLS
	H6	Housing type – 6 story and more	Share of units in 6-story and more buildings to total houses	38.83	95.39	4.47	HLS
Transport features	t1	Mobility length	Minutes of time moved	2,000,102	6,890,175	84,075	NLNI
	t2	Train trips	Share of trips by trains	34.00	76.62	7.46	NLNI
	t3	Bus trips	Share of trips by buses	2.95	6.29	0.67	NLNI
	t4	Car trips	Share of trips by cars	20.33	69.49	6.02	NLNI
	t5	Bicycle trips	Share of trips by bicycles	19.64	31.67	2.18	NLNI
	t6	Walk trips	Share of trips by walking	23.08	28.06	12.47	NLNI
	t7	Train station density	Number of trains to total area	18.27	66	0	NLNI
Socio-economic features	s1	Share of population under 15	Number of under 15 population to total population	11.68	14.94	8.30	SSDSE
	s2	Share of population 15–64	Number of 15–64 population to total population	65.10	71.95	51.01	SSDSE
	s3	Share of population over 65	Number of over 65 population to total population	23.22	36.11	16.12	SSDSE
	s4	Mean income	Average income of households in each zone	4052.78	9017.47	2935.60
	s5	Share of single families	Share of single families to total households	46.11	67.29	21.57	SSDSE
	s6	Unemployment rate	Share unemployed population to 14–65 population	2.85	4.14	1.16	SSDSE
	s7	Ratio of university degree	Share of population with university degree to total population	41.16	62.51	13.80	National census 2020
	s8	Day to night population ratio	Day population to night population	141.47	1460.58	73.78	National census 2020
	s9	Telecommuting rate	Share of remote workers	0.17	0.30	0.10	National census 2020

SSDSE: Standardized Statistical Data Set for Education (https://www.nstac.go.jp/use/literacy/ssdse/)

NLNI: National Land Numerical Information download service (https://nlftp.mlit.go.jp/ksj/)

HLS: Household and Land Survey (https://www.stat.go.jp/data/jyutaku/index.html)

NLNI: National Land Numerical Information download service (https://www.stat.go.jp/data/kokusei/2020/kekka.html)

2.2. Methods

Previous studies have shown that the association between urban features and COVID-19 is sensitive to methods (RazaviTermeh et al., 2021; Yao et al., 2021). Therefore, in this study, we employed different quantitative and spatial models to explain this relationship. To understand the dynamics of COVID-19 dispersion in Tokyo, we first applied Moran's I statistics to find the spatial clusters. Then, after correlation analysis, we performed regression analysis considering spatial and temporal patterns to find the predictors of COVID-19 spread. We chose the infection rate (instead of the cumulative number of cases) as it was very common in the literature. Also, the infection rate is standardized by the population size of each municipality. The number of cases without such a standardization typically confounds with population size. Therefore, it is often difficult to distinguish if regression coefficients really explain the number of cases or if they just explain population size. Because of these reasons, we chose the infection rate. This section explains the indicators and models used in these analyses.

2.2.1. Global Moran's I

To understand the spatial distribution of COVID-19 cases in Tokyo, we have evaluated the strength of spatial autocorrelation using Moran's I statistics (Anselin, 1995). The global Moran's I shows the level of clustering/dispersion (Dadashpoor and Alidadi, 2017). The global Moran's I is defined as below:

$$I=\frac{n}{{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}}\frac{{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}$$ (1)

where $x_i$ is COVID-19 rate in municipality i, and $\overline{x}$ is the average, w_ij quantifies spatial connectivity that will be 1.0 if the municipalities i and j share the border while zero otherwise ($w_{ij}=0$ if $i=j$). n is the total number of zones in the city. The value is −1 ≤ I ≤ 1, where a negative value means negative spatial autocorrelation, suggesting that nearby observations are dissimilar (i.e., checkerboard pattern), value of 0 means randomly distributed, and a positive value means positive spatial autocorrelation, suggesting that nearby observations are similar (i.e., clustered pattern).

2.2.2. Local Moran's I

To understand the location of clusters/dispersion, a local indicator is needed. Local Moran's I, defined as below, maps the spatial clusters or outliers of COVID-19 rate:

$$I_i=\frac{\left(x-\overline{x}\right)}{\frac{{\sum }_{i=1}^n\left(x_i-\overline{x}\right)}{n-1}}\sum _{j=1,j\ne i}^n{w}_{ij}\left(x_j-\overline{x}\right)$$ (2)

$I_i$ takes a large positive value if positive (or negative) values are clustered around the site i, while negative if positive observations are clustered around the site which has negative observation.

2.2.3. Correlation coefficient

Correlation analysis is one of the most fundamental methods to understand the association between the built environment and human factors of COVID-19 spread in cities (Kadi and Khelfaoui, 2020; Tieskens et al., 2021; Mehmood et al., 2021; Kamble and Bahadure, 2021). We have employed Pearson correlation for two reasons: First, to understand the association between selected factors and COVID-19 rate spread. Second, to understand to what extent explanatory variables are correlated with each other. The Pearson correlation between variables $x_i$ and $y_i$ are defined as below:

$$r=\frac{{\sum }_{i=1}^n{\sum }_{j=1}^n\left(x_i-\overline{x}\right)\left(y_j-\overline{y}\right)}{\sqrt{{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}\sqrt{{\sum }_{j=1}^n{\left(y_j-\overline{y}\right)}^2}}$$ (3)

$\overline{x}$ and $\overline{y}$ are the means of $x_i$ and $y_i$ respectively

2.2.4. Regression models

A multivariate regression model is applied to understand the relative importance of COVID-19 ratio predictors. There are tens of explanatory variables which can explain the dynamics of COVID-19 in urban areas (Alidadi and Sharifi, 2022). However, the inclusion of independent variables should be methodologically viable. We have applied a step-wise multivariate regression model to select the most relevant factors. The multivariate linear regression comes as below:

$$Y_i={\beta }_0+\sum _{k=1}^K{X}_{i,k}{\beta }_k+{\epsilon }_i,{\epsilon }_i\sim N\left(0,{\sigma }^2\right)$$ (4)

where $Y_i$ is the dependent variable at municipality i, $X_{i,k}$ (k = 1,2,3,..., K) is explanatory variable, β ₀ is constant, β_k is regression coefficient, ε represents residual/noise with variance ${\sigma }^2$.

We initially selected the most relevant variables based on data availability and the literature. Then, we employed a bidirectional step-wise regression model to select the most significant factors that explain our dependent variable. To select variables, we used Bayesian Information Criterion (BIC) to find the best model. As most explanatory variables were highly correlated, collinearity was one of the issues considering variable selection. We used Variance Inflation Factor (VIF) to find the most relevant and significant factors. While there is no consensus regarding the VIF threshold, 5 is one common criterion that we also relied on (Russette et al., 2021; Li et al., 2021b).

Although residuals are assumed independent in Eq. (4), they can be spatially autocorrelated in the presence of omitted variables, which are missed in the regression model, that leads to Type I error (i.e., over-estimation of statistical significance). To deal with this issue and eliminate the spatial dependency in the residual, spatial error model (SEM) and spatial lag model (SLM) were applied. The equation for the spatial error model comes as below:

$$Y_i={\beta }_0+\sum _{k=1}^K{X}_{i,k}{\beta }_k+{\epsilon }_i,{\epsilon }_i=\lambda \sum _{j=1}^N{w}_{i,j}{\epsilon }_j+u_i,u_i\sim N\left(0,{\sigma }^2\right)$$ (5)

The model assumes that the residuals ${\epsilon }_i$ are spatially correlated, and the spatial autocorrelation captured by the spatial weight $w_{i,j}$,describing the connectivity among municipalities, and the parameter λ control the strength of the spatial correlation.

The dependent variable ($Y_i$) in our analysis is the rate of COVID-19. Therefore, step-wise regression model with time-fixed effects is employed to assess the effect of time. The time-fixed effects model finds the unobserved variables which might have changed through time. The model comes as below:

$$Y_{i,t}={\beta }_0+\sum _{k=1}^K{X}_{i,t,k}{\beta }_k+\sum _{t=1}^{T-1}{D}_{i,t,k}{\gamma }_t+u_{i,t},u_{i,t}\sim N\left(0,{\sigma }^2\right)$$ (6)

In this equation, Y_it represents the response variable for observation i at time t, X_i,t,k represents the k predictor variables for observation i at time t, D_{i , t , k} represents the dummy variable for the t-th time periods, and u_it is the error term for observation i at time t. The coefficients β ₁ to β_k represent the effect of each predictor variable on the response variable, while the coefficients γ ₁ to γ_{T −1} represent the effect of each time period on the response variable. All the calculations for data cleaning, aggregation, analysis, and mapping are done in RStudio 4.2.2 version and Arc-Map 10.3, and spatial analyses are done by two packages, “spdep” version 1.2.7 and “spatialreg” version 1.2.6. `

2.3. Case study

The Tokyo prefecture covers Tokyo city and the western wards, comprising around 13.2 million people over 1778 square kilometers. The prefecture is one of the most densely populated areas in the world, with the average density being 6306 people/km², which increases to 15,250 people/km² inside the Tokyo city. The prefecture has more than 9 million employment opportunities, most of which are located within Tokyo city and the remainder in the surrounding areas (see Fig. 1 ). It has an extensive public transportation system, including trains, subways, and buses. Offices and commercial buildings are concentrated in the center of the Tokyo city, and many people commute to the center.

Fig. 1.

The case study area and the spatial distribution of its population and employment (Authors calculation based on SSDSE data).

Japan reported its first COVID-19 case reported in February 2020. It experienced several waves of COVID-19 peaks that forced the government to recommend strict restrictions. During the studied period (April 2020 to October 2021), the Japanese government tried to introduce some measures to control the virus. Specifically, Japan declared a state of emergency in April 2020 for several prefectures, including Tokyo. Most schools were closed during the period, and many office workers shifted to remote work. The emergency declaration was lifted in May 2020, but was re-imposed in January, April, and July 2021 due to a surge in cases. Japan has also imposed border restrictions on travelers from over 100 countries since April 2020. In March 2021, it temporarily suspended the entry of all non-resident foreign nationals due to the rise in COVID-19 cases. During the study period, Japan has strongly encouraged people to wear masks in public places.

State of emergency, closure of some specific businesses, international border closure and Three Cs (closed spaces, crowded places, and close-contact settings) were among the most influential factors to avoid COVID-19 outbreaks In Japan. Japanese government declared four emergency states (Fukuchi et al., 2022; Song and Karako, 2021)

Tokyo had the highest rate of infection throughout the country (348,011 positive cases during the covered period). While it comprises 11% of Japan's population, around 22% of positive cases were reported in this prefecture. Around 350 thousand positive cases were reported in Tokyo prefecture during the studied period. Fig. 2 reveals that infection changes dramatically throughout the covered period. The lowest number of cases was in June 2020 (935), while the highest was in August 2021 (116,798). Population density (7442 per km²) and employment density (5046 per km²) in this prefecture, specifically Tokyo city (23 wards), are among the highest in the world. Tokyo and Japan both had the same patterns, four main waves of COVID-19 cases; its worst one (in terms of morbidity and mortality) occurred in August 2021. It is worth mentioning that Tokyo Olympics was held between 23 July to 5 August 2021. However, the government came up with a state of emergency (supplemented by some mobility restrictions and massive vaccination), contributing to a surprisingly very low level of COVID-19 infection at the end of October 2021 (Song and Karako, 2021).

Fig. 2.

COVID-19 cases and their spatial distribution in Tokyo prefecture during the covered period (Authors calculation based on SSDSE and MHLW data).

3. Results

3.1. Hot spots of COVID-19 in Tokyo

As explained in Sections 2.2.1 and 2.2.2, the global and local spatial autocorrelation statistics are applied to understand the level of clustering. The result of the global spatial autocorrelation analysis is presented in Fig. 3 . All of the clustering coefficients are statistically significant at a 99.9% confidence level. The clustering dramatically changed over time; varying from its lowest level of 0.17 (June 2020) to its highest level of 0.81 (July 2021). Beyond these two values, there are some other peaks and valleys that could result from some interventions or changes in the virus's nature. The clustering peaks are found in April 2020, October 2020, December 2020, April 2021, and July 2021. All of these peaks were followed by low levels of clustering in the following months. While more evidence is needed, it seems that the lowest level of clustering is in the first month after the state of emergency was declared. The most tangible drop in clustering coefficient was found from July 2021 (0.8) to October 2021 (0.2), which reduced to one-fourth of its original.

Fig. 3.

Changes in global Moran's I spatial autocorrelation of COVID-19 infection rate in Tokyo prefecture (Authors calculation based on MHLW data).

Global spatial autocorrelation for the cumulative rate of infections in Tokyo was 0.76, which shows a very strong level of spatial clustering/positive spatial autocorrelation. It was statistically significant (p-value 0.000). This index for the cumulative number of cases was 0.44, which shows a lower level of clustering in comparison to the infection rate. Local spatial autocorrelation is presented in Fig. 4 . This result reveals a high level of clustering in the central municipalities of Tokyo. Out of 53 municipalities, 18 are included in high-high clusters (which means they have a high value of COVID-19 infection rate and are surrounded by the same ones). However, when it comes to the cumulative number of cases, there is another pattern. Out of the 53 municipalities, eight are in the high-high clusters, which shows a lower level of clustering as the global index was also lower than the infection rate (these are Shinagawa-Ku, Meguro-Ku, Shibuya-Ku, Nakano-Ku, Suginami-Ku, Toshima-Ku, Kita, and Itabashi-Ku). More importantly, the low-low clustering (which means concentration of low values) is not in the central areas, they are in the western and northern suburbs of Tokyo.

Fig. 4.

Local Moran's I clusters of cumulative number and infection rate of COVID-19 in Tokyo prefecture (Authors calculation based on MHLW data).

3.2. Correlation analysis

We have employed the Pearson correlation, which was explained in Section 2.2.3, to understand the association between selected variables and the spread of COVID-19 in Tokyo. In the next step, the correlation between explanatory variables was also calculated to know the level of independence. All of the calculations were done on a 95% confidence level.

As shown in Table 2 , the rate of COVID-19 infection in Tokyo is associated with some variables of density, housing condition, transport, socio-economic factors, and land use. Except for the density of agriculture, forestry, and fishery worker jobs, the significance level of correlation between COVID-19 and density variables is less than 0.05. As was expected, this association for all variables is positive. Interestingly, the highest level of correlation is reported for areas with a high density of restaurants, retail stores, and medical facilities. Additionally, the zones with higher rates of people who work in accommodation and food services, life related services and entertainment, and medical and welfare services have higher rates of COVID-19 infection (it should be noted that this result is just based on a simple correlation not causality analysis).

Table 2.

Pearson correlation coefficient of COVID-19 infection rate and explanatory variables.

Dimension	Indicator	Correlation	Dimension	Indicator	Correlation
Density	d1	0.715 (0.000)	Density	d14	0.558 (0.000)
	d2	0.293 (0.036)		d15	0.633 (0.000)
	d3	0.318 (0.023)		d16	0.614 (0.000)
	d4	0.650 (0.000)		d17	0.621 (0.000)
	d5	0.684 (0.000)		d18	0.398 (0.004)
	d6	0.823 (0.000)		d19	0.694 (0.000)
	d7	0.806 (0.000)		d20	0.543 (0.000)
	d8	0.850 (0.000)		d21	0.768 (0.000)
	d9	0.211 (0.136)		d22	0.821 (0.000)
	d10	0.270 (0.055)		d23	0.612 (0.000)
	d11	0.301 (0.032)		d24	0.818 (0.000)
	d12	0.712 (0.000)		d25	0.389 (0.005)
	d13	0.547 (0.000)		d26	0.618 (0.000)
Housing conditions	h1	-0.835 (0.000)	Socio-economic conditions	s1	-0.691 (0.000)
	h2	-0.758 (0.000)		s2	0.788 (0.000)
	h3	-0.857 (0.000)		s3	-0.626 (0.000)
	h4	0.753 (0.000)		s4	0.622 (0.000)
	h5	0.098 (0.492)		s5	0.857 (0.000)
	h6	0.521 (0.000)		s6	-0.652 (0.000)
Transport factors	t1	0.213 (0.133)		s7	0.606 (0.000)
	t2	0.824 (0.000)		s8	0.337 (0.016)
	t3	0.147 (0.303)		s9	0.817 (0.000)
	t4	-0.680 (0.000)	Land use factors	lu1	-0.197 (0.165)
	t5	-0.438 (0.001)		lu2	0.686 (0.000)
	t6	0.075 (0.600)		lu3	0.104 (0.466)
	t7	0.751 (0.000)		lu4	-0.525 (0.000)
p-value presented in parenthesis		lu5	-0.541 (0.000)

The correlations between the COVID-19 rate and all socio-economic variables are statistically significant. However, all of them are not in the same direction. The share of people under 15 and over 65, and the unemployment rate are negatively associated with the rate of COVID-19 infection (population aged 15–65 are positively and significantly associated with the infection rate in Tokyo). It shows that the younger population are more vulnerable to the virus as they probably spend more time in high-risk places like workplace, public spaces, and bars. Housing condition factors, as expected, are highly correlated with the rate of infection, specifically overcrowding, which is measured by the number of rooms per residence (−0.83) and share of rooms (0.75).

Transport variables also were highly correlated with infection rate. The share of trips by train (0.82), train station density (0.75), and car trips (−0.68) are significantly correlated with the rate of infection. The lowest levels of correlation were found for land use factors. The share of commercial land use is associated with a higher rate of infection. On the other hand, the share of open and green spaces is negatively correlated with the COVID-19 infection rate.

Fig. 5 shows the correlation between the explanatory variables themselves. For our purpose, we have presented the absolute value of the correlation coefficient; the direction of the correlation is not considered. The results reveal that density variables are highly correlated with each other. They are also correlated with socio-economic and urban transport factors in Tokyo. Additionally, socio-economic variables correlate with most selected factors in the analysis. Residential density, FAR, housing type, some specific occupations’ density (like agriculture and fishery), mobility length, the share of walking trips, and land use factors were less correlated to other explanatory variables. However, Fig. 5 is just for a better explanation, and the correlation between these variables will be investigated through the step-wise regression modeling. It is worth mentioning that the correlation between explanatory variables reduces the reliability of regression models.

Fig. 5.

Absolute value of correlation coefficient of explanatory variables themselves (Authors calculation based on various data sources).

3.3. Predictors of COVID-19 cases in Tokyo

We examined the literature and evaluated available data in Tokyo to clarify how COVID-19 spreads in the prefecture. As outlined in the methodology section, a Step-wise regression approach is utilized to determine the most appropriate variables for our model. Several factors need to be considered when developing the global linear regression model. Firstly, it was important to ensure that the explanatory variables were not correlated with one another (i.e., collinearity). Secondly, the dependent variable needed to have a distribution that closely resembled a normal distribution (i.e., normality). Thirdly, it was necessary to avoid over- or under-fitting of the model through the inclusion of appropriate variables. Fourthly, we sought to eliminate spatial dependency in the residuals to the greatest extent possible.

Based on the requirements mentioned above we transformed the infection rate variable to a logarithmic form (log(r): logarithmic transformation of infection rate) as it was more closely aligned with normal distribution. A total of 53 variables (in main areas of density, land use, socio-economic conditions, housing conditions, and transport) were considered as the input of the step-wise regression model. Table 3 details the results of the model, which was run in RStudio. Six variables were selected in the regression model, which their VIF value were less than 5. Except for walk trips (t6), the rest of the coefficients are statistically significant at 99% level. Regarding the direction of the relationship, all of the variables have a positive coefficient. Telecommuting rate (s9) has the highest coefficient value, while life-related services and entertainment workers’ density has the lowest coefficient.

Table 3.

The results of stepwise regression model.

	Dependent variables:
	log(r)
s5: Share of single families (VIF: 3.81)	0.013∗∗∗
	(0.003)
s9: Telecommuting rate (VIF: 2.62)	2.090∗∗∗
	(0.487)
lu3: Share of industrial area (VIF: 1.05)	0.003
	(0.001)
d22: Life related services, entertainment workers density (VIF: 3.35)	0.0003∗∗∗
	(0.0001)
t6: Share of walk trips (VIF: 2.13)	0.017∗∗∗
	(0.007)
Constant	−0.784∗∗∗
	(0.146)
Observations	51
R2	0.921
Adjusted R2	0.911
Residual Std. Error	0.117 (df = 44)
F Statistic	85.841∗∗∗ (df = 6; 44)

Note:^∗p<0.1,^∗∗p<0.05. ^∗∗∗p<0.01

Based on the step-wise regression model results, R² is very high (0.92), which shows how well the selected variables could explain the variability of the infection rate in Tokyo. However, this result could not be generalized because the accuracy is not investigated. Global Moran's I is employed to analyse the spatial autocorrelation of residuals. The result of this method showed some level of spatial dependency on the distribution of residuals, and it is statistically significant (p-value < 0.01). It means that some explanatory variables are missed out. To eliminate the spatial dependency of residuals, which means the model has a problem (over-prediction or under-prediction), we have employed both spatial lag and spatial error model.

The results of the spatial lag and error model are presented in Table 4 ., The same variables as the global model are also selected in both the spatial error and spatial lag models. The results for both spatial lag and spatial error models are more or less the same. They all have the same direction, but the value of the coefficients slightly changed. As shown in Table 4, σ ² is very low (0.01), which means there is relatively little unexplained spatial variation in the dependent variable. Regarding the coefficients, except for the share of industrial area (lu3), the rest are statistically significant at 95% level. To investigate the best model which explains the dynamics of infection rate, we employed BIC. The results show a relatively small difference between the global step-wise regression model (−52.19), spatial lag (-51.96)and the spatial error model (−51.32). It means that neither model is a good fit for the selected data. It is also worth noting that we changed the number of input variables in both models, and they were not stable. In other words, the explanatory variables selected in the model were sensitive to the total number of variables. As previously stated, the infection rate is based on the cumulative number of cases that were originally reported on a monthly basis. Therefore, we employed a time-fixed model to consider the effect of time in our estimation.

Table 4.

The results of spatial lag and spatial error models.

	Spatial lag	Spatial error
	log(r)	log(r)
s5: Share of single families	0.009∗∗∗	0.013∗∗∗
	(0.003)	(0.003)
s9: Telecommuting rate	1.353∗∗∗	1.968∗∗∗
	(0.513)	(0.436)
lu3: Share of industrial area	0.001	0.002∗
	(0.001)	(0.001)
d22: Life related services, entertainment workers density	0.0003∗∗∗	0.0003∗∗∗
	(0.0001)	(0.0001)
t6: Share of walk trips	0.016∗∗∗	0.023∗∗∗
	(0.006)	(0.006)
Constant	−0.654∗∗∗	−0.883∗∗∗
	(0.147)	(0.53)
Observations	51	51
Log Likelihood	43.677	43.356
σ2	0.010	0.010
Akaike Inf. Crit.	−69.353	−68.712
Wald Test	7.248∗∗∗ (df = 1)	10.047∗∗∗ (df = 1)
LR Test	5.727∗∗ (df = 1)	3.062∗ (df = 1)
BIC	−51.96	−51.32

Note:^∗p<0.1

^∗∗p<0.05

^∗∗∗p<0.01

Table 5 details the results of the regression model with time-fixed effects. In this model, all 19 time periods were included as input variables, alongside 53 other explanatory variables. The results showed that the time periods’ coefficient and significance levels vary in different months. For certain periods (i.e., period 2, period 3, and period 19) the coefficients are negative, whereas, for the others, they are positive. Moreover, some time periods, such as period 2, period 6, and period 7, do not demonstrate statistical significance. The R ² for this model is the same as step-wise regression model (0.922). But the time-fixed effects model is more stable than the previous step-wise regression model.

Table 5.

Results of regression model with time fixed effects.

	Dependent variable:
	log(r + 0.5)
period2	−0.020 (0.013)
period3	−0.027∗∗ (0.013)
period4	0.032∗∗ (0.013)
period5	0.053∗∗∗ (0.013)
period6	0.018 (0.013)
period7	0.020 (0.013)
period8	0.078∗∗∗	(0.013)
period9	0.173∗∗∗	(0.013)
period10	0.371∗∗∗	(0.013)
period11	0.097∗∗∗	(0.013)
period12	0.069∗∗∗	(0.013)
period13	0.154∗∗∗	(0.013)
period14	0.196∗∗∗	(0.013)
period15	0.104∗∗∗	(0.013)
period16	0.381∗∗∗	(0.013)
period17	0.884∗∗∗	(0.013)
period18	0.291∗∗∗	(0.013)
period19	−0.013 (0.013)
h3: Household crowding	−0.146∗∗∗
	(0.019)
s9: Telecommuting rate	0.441∗∗∗
	(0.065)
d22: Life related services, entertainment workers density	0.0001∗∗∗ (0.00001)
s7: Ratio of university degree	−0.001∗∗∗ (0.0003)
lu2: Share of commercial land use	−0.001∗∗∗ (0.0002)
Constant	−0.407∗∗∗ (0.051)
Observations	969
R2	0.922
Adjusted R2	0.920
Residual Std. Error	0.065 (df = 944)
F Statistic	465.886∗∗∗ (df = 24; 944)

Note:^∗p<0.1

^∗∗p<0.05

^∗∗∗p<0.01

As given in Table 5, six explanatory variables were selected in the model. Interestingly, household crowding, ratio of university people with university degree and share of commercial land use were negatively associated with infection rates. On the other hand, life-related services, entertainment workers density, construction workers density and telecommuting rate were positive predictors of infection rate. All of these variables were statistically significant (p value < 0.01) predictors of COVID-19 infection rate in Tokyo. The variables with the highest coefficient values were telecommuting rate (0.441) and household crowding (−0.143).

4. Discussion

As mentioned in the previous section, the COVID-19 infection rate was clustered in the central municipalities. The longitudinal representation of global spatial autocorrelation showed that the level of clustering changes over time. Spatial autocorrelation in Tokyo was very strong (0.76). However, this level of clustering might be sensitive to the scale of analysis, and conclusive judgments should be avoided. Mena et al. (2021) and Xu et al. (2022) found clusters of COVID-19 cases in areas with high population and employment density. The same results were tangible in the case of Tokyo, as central municipalities are where the clusters of high infection rates are located. The most affected municipalities during all stages of the pandemic are Shinagawa-Ku, Meguro-Ku, Shibuya-Ku, Nakano-Ku, Suginami-Ku, Toshima-Ku, Kita, and Itabashi-Ku. These are characterized by high population and employment density. In addition, compared to other municipalities, they feature a higher density of urban amenities and services (e.g., retail stores, medical facilities, bars, etc.) and a lower share of open and green spaces. These may have contributed to their vulnerability as the results of our statistical analysis show.

However, as the COVID-19 data does not represent the place where the person has been infected, finding causality between density and COVID-19 is not straightforward. In the case of Tokyo, when we used infection rate, the clustering included municipalities with high population and employment . But when we used the cumulative number of cases, population density seemed more important than employment density.

During the pandemic, the Japanese government introduced several initiatives to preserve its economic growth and boost/recover struggling sectors, like tourism (Fukuchi et al., 2022). Key initiatives included Go-To-Travel and Olympics, which were followed by a surge in cases in Tokyo. The highest levels of clustering were found in these months. Other studies also have found that certain locations and age groups have played key roles during different waves of the outbreak. Young people in bars and schools were identified as the main sources of transmission in at least two outbreaks in Japan (Lertsakornsiri et al., 2022).

Findings from this study suggest that built environment and human factors have, in some cases, very strong correlation with COVID-19 spread in Tokyo. Among others, density was one of the most critical factors that was positively associated with the number of cases. We employed various types of density to see to what extent they are associated with the infection rate. Most previous studies have found population density to be the most correlated factor (Kamble and Bahadure, 2021; Vaz, 2021; Kaufman et al., 2021). Our results confirm that population density is positively correlated with the COVID-19 infection rate. Nonetheless, other density variables, like the density of schools, retail stores, medical facilities, and the density of healthcare workers, demonstrated stronger correlations. These results are consistent with those reported for Tehran (Lak et al., 2021), Hong Kong (Kan et al., 2021), Lombardy (De Angelis et al., 2021), and Wuhan (You et al., 2020), but different from results found for Huangzhou (Li et al., 2021a). Interestingly, construction workers' density is strongly and positively correlated with the number of cases, which is not (to our best knowledge) reported in previous studies. Based on these results, it seems that other representations of density, like the concentration level (i.e., crowding) of people in specific places like hospitals, retail stores, schools, bars, and restaurants, increase the risk for locals and workers in these services.

Regarding the socio-economic variables, correlation analysis showed that people who are not in the working age group (<15 and 65<) and the unemployment rate are negatively correlated with the infection rate. This result is in contrast with the findings of Wheaton and Kinsella Thompson (2020) and Baser (2021). Despite some previous studies (Kashem et al., 2021; Ali et al., 2021; Wang et al., 2021), mean income, telecommuting rate, and people with a university degree are positively associated with the COVID-19 infection rate in Tokyo. Arguably, the share of single families in Tokyo municipalities has the highest level of correlation with the COVID-19 infection rate. Liao et al. (2021) also used single-parent families as a variable and found it strongly and positively correlated with infection rate.

Findings on household crowding did not align with existing literature.. This variable has been tested in different ways by the number of people per room or the number of rooms per household (Consolazio et al., 2021; Wheaton and Kinsella Thompson, 2020; Nguyen et al., 2021). All of them were positively associated with infection rate, which means people who live in smaller places face a higher risk of infection. However, in the case of Tokyo, number of people per room or rooms per household was not significant, but the household size was negatively associated with the infection rate. Regarding physical factors, our results confirmed that the share of travel by public transport is strongly correlated with the infection rate. Similar results have been reported for Tehran (Lak et al., 2021; RazaviTermeh et al., 2021), New York (Cordes and Castro, 2020; Yang et al., 2021), and Chilean cities (Villalobos Dintrans et al., 2021). The shares of open and green spaces were not highly correlated with the COVID-19 infection rate. However, they were negatively associated with infection rate, which corroborates studies by Nguyen et al. (2020) and Spotswood et al. (2021). As discussed earlier, in this study, we analysed the effectiveness of different models, specifically to understand how spatiality and temporality affect the prediction model. Our global regression model was not stable enough, and the residuals were spatially correlated. Like previous studies (Tchicaya et al., 2021; You et al., 2020) we tested SEM and SLM. But in our case, the validity of the SEM, SLM and OLS was roughly the same. However, the time fixed effects model was more stable and provided a better explanation of the COVID-19 spread in Tokyo. It has been thoroughly discussed that the impact of time and space should be considered in modeling COVID-19 spread (Wu and Zhang, 2021; Urban and Nakada, 2021; Rahman et al., 2021; Chen et al., 2021). However, most studies just emphasized the spatial dimension of COVID-19 spread (Basellini and Camarda, 2021). In our case study, the impact of time was more important than spatial settings.

Regarding the predictors of COVID-19 infection rate, telecommuting work had the greatest coefficient. According to the findings of (Yang et al., 2021), the relationship between the percentage of remote workers and the infection rate was negative in New York City, but positive in Tokyo. In other words, while a higher proportion of remote workers was associated with lower infection rates in New York, the opposite relationship was observed in Tokyo. One of the main explanations for this contrasting result could be that some socio-economic or spatial characteristics may lead to a higher rate of infection, which is not presented by the remote workers variable. Wang et al. (2021) discussed that remote workers are usually people who work in professional services, education and research, and technical services. However, in Japan, the emphasis was on prevention protocol rather than remote work policies, in contrast to other developed nations.

We found that household crowding was a negative predictor of COVID-19. This differs from previous studies, which linked overcrowding with a heightened risk for infection. This relationship has been consistently reported in studies conducted at different levels, including state, county, city, and urban districts (Chan et al., 2021; Hu et al., 2021; Liu et al., 2021). The primary reason for this contradiction could be that central municipality households in Tokyo are typically small and have access to necessary amenities and services, allowing for effective compliance with protection measures. Additionally, there is a chance that in the areas with higher rate of infection have lower household size, but it is not a causal relationship.

Another finding which could be critical is the positive and significant relationship between the density of life-related services, and entertainment workers and COVID-19 spread patterns. Previous studies have found that people working in entertainment and life-related services are more likely to get infected (De Angelis et al., 2021; Olmo and Sanso-Navarro, 2021; Credit, 2020). The same results were reported for Japan, as bars were one of the city's key spaces that contributed to different outbreaks (Lertsakornsiri et al., 2022). Based on previous studies, there have been some reports that unskilled workers are more likely to be infected due to their social and economic insecurities (Tamrakar et al., 2021; Fielding-Miller et al., 2020). Our results also showed that municipalities with higher rate of university degrees have lower rate of infection.

An important issue that needs to be acknowledged is that we have tried to consider different variables that may affect the dynamics of COVID-19 spread in Tokyo. However, due to the complex nature of these dynamics and data availability issues, there could still be other variables that have not been considered. For instance, differential measures and conditions related to quarantine policies, hospital facilities, and vaccination may affect the dynamics across different municipalities. Due to data limitations, we have not integrated these measures into our model. We hope this issue will be addressed in future research.

5. Conclusion and policy implications

Cities have been at the forefront of the battle against COVID-19 (Carozzi et al., 2022). Non-pharmacological short-term initiatives and long-term solutions are critical for dealing with current and future pandemics. This study aimed to understand the associations between urban features and COVID-19 in Tokyo. There were also three primary objectives. Firstly, to understand the spatial and temporal characteristics of COVID-19 in Tokyo. Secondly, to scrutinize the association between urban features and COVID-19 infection in the city. Thirdly, to find the predictors of COVID-19 infection in Tokyo.

We employed cluster analysis, correlation analysis, and different regression models to find the predictors of COVID-19 infection in Tokyo. Results showed that COVID-19 infection rate is clustered during the peaks, and the level of clustering was reduced when some mobility restrictions were introduced. The clusters were mainly found in the central to north and western municipalities. Regarding correlation analysis, we found that our explanatory variables, commonly analysed in the literature, correlate with one another. Among others, the density of restaurants and health facilities, health, entertainment and life-related workers’ density, and public transport use were positively correlated with the COVID-19 infection rate. In contrast, household crowding and private car use were negatively associated with infection rate. Additionally, we found that global and spatial models had the same validity, while when the effect of time was considered, the model was more stable. Household crowding and telecommuting were among the most significant factors affecting the spread of COVID-19 in Tokyo. It is worth mentioning that the density of entertainment and life-related factors were also among the predictors of COVID-19 spread in Tokyo.

From a policy perspective, urban factors are very complex and interdependent. To make cities more pandemic-resilient, this complexity of urban sub-systems should be taken into consideration. While some factors like density might be positively associated with COVID-19 cases, sprawl, on the other hand, increases travel time, which could increase the vulnerability to pandemics. Additionally, the concentration of health services and facilities in specific areas of the city is a negative point. Decentralisation and distribution of services and facilities could enhance safe service accessibility and reduce the risk of infection. Finally, some occupations like health, food services, and entertainment and life-related jobs are essential for everyday life and cannot be restricted even during lockdowns. Therefore, workers in these sectors need more attention from policy makers to maintain their livelihood security and reduce pressure on them during times of crisis.

This research uses COVID-19 data that indicates the residential locations of infected individuals. However, to be more accurate, the data should report the place where the person got COVID-19. In future research, it is advised to utilize data indicating the COVID-19 transmission location rather than the residence of infected individuals. This could, however, be challenging. Second, the scale of analysis in this research was municipalities (53), which is not stable enough to use geographically weighted regression. We recommend that future studies focus on lower scales, like neighbourhoods or blocks to represent a more reliable analysis. A more granular analysis across different neighborhoods will also be desirable as it better explains each neighborhood's unique characteristics that have contributed to the spread and/or control of the pandemic. Third, we employed step-wise regression model to find the most logically important and statistically significant predictors of COVID-19. However, it is recommended to also use pricnipel component analysis (PCA) as another suitable tool to reduce the number of factors. Fourth, Japan initiated several restrictions to control the pandemic, but we did not consider these policies in our model. It is suggested that future studies consider this factor in modeling the dynamics of COVID-19 in cities. And finally, our analysis was based on 19 months of the pandemic in Japan. Follow-up studies are recommended to extend the time period in future studies.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

• Data will be made available on request.