Introduction

The coronavirus disease 2019 (COVID-19) pandemic has presented significant challenges in various domains including politics, economy, science, and public health worldwide. Although the World Health Organization declared an end to the global Public Health Emergency for COVID-19, its impacts continue to persist, leaving behind numerous lessons to be learned. These lessons range from the importance of rapid and effective public health responses to understanding the long-term socio-economic repercussions of such a global health crisis.

By the end of June 2023, Japan experienced large-scale COVID-19 cases and the subsequent mortality, pervading all prefectures1,2. The continued study to understand the factors that contributed to controlling COVID-19 infections is critical for better preparedness and response strategies for any future pandemics. In addition to vaccination, social distancing remains a key approach in curbing the virus transmission through human interactions. Concurrently, environmental factors such as the presence of green spaces3, air quality4, and climatic conditions including temperature and humidity5,6, have been indirectly linked to COVID-19 transmission. For example, green spaces may facilitate safer social engagement relative to indoor settings, potentially diminishing transmission risks3. Nevertheless, these environmental factors do not directly measure the intensity of human interactions, which are crucial in understanding the spatial dissemination of the disease7.

In contrast to these environmental factors, high-resolution nighttime light (NTL) imagery, captured from satellites, has emerged as a more direct indicator of human interaction intensity8, significantly influencing COVID-19 transmission dynamics7,9,10,11. Prior studies showed that NTL satellite imagery has been employed to estimate population density12, economic growth13, Gross Domestic Product13,14, urbanization15, and so on, all of which bear relevance to COVID-19 transmission dynamics. Furthermore, exposure to more intense NTL has been associated with elevated obesity prevalence16,17, and obese individuals may be more prone to severe COVID-19 consequences18,19. Some studies have analyzed NTL changes during social distancing periods in several large cities including China and the U.S10,20. Nighttime light exposure potentially facilitates the spread of COVID-19 by extending social activities and disrupting natural sleep patterns, leading to increased human interactions and altered behavior in densely populated areas7,9,11.　However, there has been little spatial analysis concerning how NTL attracts human aggreggation and consquentially propagates infections spatially. More specifically, there is a lack of detailed analysis regarding the “spillover effect” of NTL: that is, how illuminated area attracts individuals, thereby facilitating the transmission of infections to the neighboring areas and impacting health outcome therein. This gap in research highlights the need for a more focused study on the spatial dynamics of NTL’s role in influencing human gatherings and the consequent diffusion of COVID-19.

The aim of this research was to investigate the relationship between COVID-19 mortality and NTL intensity, which is indicative of higher nocturnal activity and human gatherings, with a focus on estimating the spillover effect of NTL on the virus transmission to the neighboring area in Japan. Additionally, this research aimed to examine the variations in this association throughout different pandemic waves, with a particular focus on contrasting earlier waves against subsequent ones dominated by variants such as Delta and Omicron.

Literature review in spatial spillover effect

The spillover effect, which occurs when the impact of a dependent variable or covariates in one area extends beyond its borders and affects neighboring regions, has been studied in various fields such as crime21, air pollution22, human mobility23,24, child health25, obesity26, primary care services27, hospital treatment28 and infectious diseases29,30,31. To assess these effects, several methods have been developed, including spatial Durbin model (SDM) and spatial Durbin error model (SDEM)32. The SDM includes a spatial lag of the dependent variable and covariates and the SDEM extends the SDM with a spatial lag of the error term. From a theoretical perspective, due to the difference in the model specifications, SDM implies global spatial spillovers (i.e., the impact goes to not only neighbors, but neighbors to neighbors, neighbors to neighbors to neighbors, and so on), while SDEM leads to local spatial spillovers (i.e., the impact goes to immediate neighbors)33. In the context of COVID-19, researchers have studied the presence of spillover effects from different perspectives. Some studies focused on the spatial spread of COVID-19 by examining the relationship between cases in different regions34,35. These studies suggest that factors such as deaths, recoveries, or vulnerability in one area could influence the number of confirmed cases in nearby areas. Another approach is to investigate the role of non-pharmaceutical interventions, such as social distancing or lockdown, whch highlighted the significant spillover effects of these interventions in areas with both geographic and social network proximity, as well as considerable differences in the spillover effects produced by various types of places36,37,38.

Methods

Data and outcome variable

This study used national mortality (vital statistical) data obtained from the Ministry of Health, Labour and Welfare between January 2020 and October 2022. It is a population-based database that includes all deaths in Japan, governed by the Japanese law, “Japanese Family Registration Law”. This dataset includes demographic information including the date of death, age at death, and place of residence, covering all deceased individuals in Japan, irrespective of nationality, as well as those who had a residence card. COVID-19 deaths were defined on the ICD10 classification basis (U071). The primary outcome variable was the number of COVID-19 deaths per 10,000 persons in the following analytical model.

Environmental from satellite image

NTL data was derived from the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) at a spatial resolution of 463 m (source: https://ngdc.noaa.go). It is ultra-sensitive to low-light conditions, offering superior spatial and temporal resolution for nighttime lights compared to previously provided nighttime lights data. To ensure accuracy, the data is filtered to remove effect from stray light, lightning, lunar illumination, and cloud-cover using the VIIRS Cloud Mask product, although anomalies such as aurora, fires, boats, and other temporal lights39,40. Then, averaging was conducted on a monthly basis. Figure 1 shows the average NTL over the study period. Additionally, we extracted daily Absorbing Aerosol Index (AAI) data from Sentinel-5 Precursor satellite launched in October 2017 by the European Space Agency to monitor air pollution. The AAI is calculated using measurements at 354 nm and 388 nm wavelengths and is sensitive to changes in Rayleigh scattering in the UV range, allowing for the monitoring of aerosol plumes from dust, volcanic ash, and biomass burning events41. These environmental covariates, which are shown to be associated to the COVID-19 mortality4,7, were then aggregated at the municipal level. The time trend of NTL during the study period is shown in Supplementary file 2.

Environmental covariates

Environmental factors such as average daily temperature and humidity42,43 were extracted from the Japan Meteorological Agency. The data were sourced from one weather station within each prefectural capital city with hourly measurements across 24 h being averaged to derive daily temperature values.

Socio-demographic covariates

Furthermore, we obtained prefecture- or municipality-level census variables, identified as potential confounders in the relationship between NTL and COVID-19 outcomes, from the Regional Statistics Database provided by Ministry of Internal Affairs and Communications in Japan3,7. These variables included the proportion of residents aged < 15 and ≥ 65 as of 2015, average household taxable income in 2020, the number of restaurants in 2006, the proportion of single-person households in 2020, the proportion of employees in tertiary industry in 2009, and population density (per 1km2) in 2015.

Data processing on outcome and covariates

A directed acyclic graph (DAG) elucidating the assumed causal pathway is presented in the supplementary file. The spatial distributions of these covariate values are shown in the Supplementary file 3. All data, including both the outcome variable and environmental covariates, were aggregated according to the following time periods corresponding to five distinct COVID-19 waves in Japan44: Wave3 (December, 2020 – February, 2021), Wave5 (July – September, 2021, dominated by Delta variant), Wave6 (January – March 2022, dominated by the Omicron variant (BA.1/BA.2)), Wave7 (July – September 2022, dominated by the Omicron variant (BA.5)), and the Reference period (WaveRef, April – June 2022). These Waves were selected for primary two reasons: they had a significantly large number of deaths, and they overlapped with periods when nation-wide prevention measures were implemented in Japan. In particular, during the study period, Wave7 was the largest wave with around 1.5 million COVID-19 cases. WaveRef is set between Wave 6 and Wave 7, a period with relatively few COVID-19 cases, corresponding to the normal-life period. In addition, Wave3 overlapped with Japan’s second state of emergency (January 8 to March 21, 2021), Wave4 overlapped with the fourth state of emergency (July 12 to September 30, 2021), and Wave6 overlapped with the second implementation of the Priority Measures to Prevent the Spread, commonly known as “Manbo” (January 9 to March 21, 2022).

Spatial analysis

Japan was divided into 1902 municipalities at the time of the study, organized within 47 prefectures. Geographical analysis in this study was conducted at the municipality level. Geographical coordinates defined as the centroid of the area (measured by latitude and longitude) were extracted from map data for each municipality. Baseline data were reported as mean (standard deviation, SD) and linear regression was used to calculate the p-for-trend by the strength of NTL.

Firstly, empirical Bayes estimates of mortality per 10,000 persons were calculated to visually inspect how the COVID-19 death were geographically distributed or concentrated in Japan during the study period. The advantage of use of empirical Bayesian estimates is that it can incorporate the information from the spatial neighborhood areas to smooth the mortality toward the local neighborhood mean, stabilizing estimates for municipalities with small numbers of deaths32,45. The spatial neighborhood was defined as the k-nearest neighborhood method with k = 8. Then, to test whether the neighbors were spatially associated with one another, global Moran’s I statistics and the p-values were calculated by each wave.

Secondly, Spatial Durbin Error Model (SDEM), which includes the lagged error term and the lagged dependent variable as well as the independent variables, was used to model the spatial direct association between NTL and COVID-19 mortality and spillover effect of NTL to neighboring municipalities. The SDEM can be formulated by:

$$\:\text{log}\left(\varvec{Y}\right)=\rho\:\varvec{W}\text{l}\text{o}\text{g}\left(\varvec{Y}\right)+\varvec{X}\varvec{\beta\:}+\varvec{\theta\:}\varvec{W}\varvec{X}+\varvec{u},\:\:\:\:$$
$$\:\varvec{u}=\lambda\:\varvec{W}\varvec{u}+\varvec{\epsilon\:},\:\:\:$$

where Y represents a \(\:N\times\:1\) vector consisting of one observation of the outcome variable (i.e., the number of COVID-19 deaths per 10,000 person) for every municipality (\(\:i=1,\dots\:,N\)), W is the \(\:N\times\:N\) row-standardized adjacency matrix defined by the Gaussian kernel, \(\:\rho\:\) is a (scaler) parameter to control the spatial association of Y, X denotes an \(\:N\times\:P\) matrix of covariates with the \(\:P\times\:1\) vector of parameter \(\:\varvec{\beta\:}\), \(\:\varvec{\theta\:}\) is a \(\:P\times\:1\) vector of parameter indicating the spatial spillover effect of X, \(\:\varvec{u}\) is a \(\:N\times\:1\) vector of error term with a parameter \(\:\lambda\:\) to control the spatial association of the error term, and \(\:\varvec{\epsilon\:}\) is a \(\:N\times\:1\) vector of independently and identically distributed error term with zero mean and variance \(\:{\sigma\:}^{2}\). The covariate vector X include all variables explained above. Note that \(\:\varvec{\beta\:}\) and \(\:\varvec{\theta\:}\) are also known as direct and indirect (= spillover) effects, respectively. The direct effect, \(\:\varvec{\beta\:}\), represents the extent to which the outcome in a given area changes when a covariate in that same area increases by one unit. The indirect effect, \(\:\varvec{\theta\:}\), represents the extent to which the outcome in neighboring areas changes when a covariate in the given region increases by one unit. More detailed estimation procedure can be found in elsewhere32. The statistical significance of direct and indirect effects are evaluated with Monte Calro simulations with 1000 iterations. In addition, to validate the use of SDEM by checking the residual distribution, we conducted Lagrange Multiplier (LM) lag and LM error tests for comparison with a simple linear model. Lastly, to check the robustness of the results against spatial heterogeneity, we divided Japan into several subzones and estimated the regression parameters within each subzone. In these subzone analyses, we conducted “SubAnalysis 1”: three subzones, including Kita-Nihon (Prefectures #1–7), Higashi-Nihon (Prefectures # 8–23) and Nishi-Nihon (Prefectures # 24–47) and “SubAnalysis 2”: seven subzones, including Hokkaido/Tohoku (Prefectures #1–7), Kanto (Prefectures #8–14), Chubu (Prefectures #15–23), Kinki (Prefectures #24–30), Chugoku (Prefectures #31–35), Shikoku (Prefectures #36–39) and Kyusyu/Okinawa (Prefectures #40–47). Note that due to the small sample size, Wave3 was not analyzed in the subzone analysis.

Statistical analysis was performed with R version 4.3.1 using the spdep package. For the statistical two-sided tests, p-values less than 0.05 were considered statistically significant. This study was conducted in accordance with the guidelines laid down in the Declaration of Helsinki. All experimental protocols were approved by the Ethical Committee of the National Institute of Infectious Diseases (authorization no. 1747). The need for informed consent as waived by the Ethical Committee of the National Institute of Infectious Diseases.

Results

Basic characteristics of data

A total of 43,199 COVID-19 deaths were included in this study: i.e., 6,038, 3,751, 12,948, 16,883, 3,579 deaths during Wave3, Wave5, Wave6, Wave7 and WaveRef, respectively. After removing missing data, 1895 municipalities were included. Table 1 indicates the basic characteristics of included municipalities, stratified by the quantile values of average NTL. As NTL move from Q1 to Q5, that is, as a city became brighter at night, an increase was observed in COVID-19 mortality per 10,000, the number of restaurants, taxable income, the proportion of inhabitants aged ≥ 65, and population density (p-for-trend < 0.01 for all variables), while the proportion of people aged < 15 decreased (p-for-trend < 0.01). Supplementary file 2 indicates the time trend of NTL, showing that it varies significantly across prefectures and tends to increase during the winter, but does not show substantial fluctuations across different periods.

Table 1 Basic characteristics of 1,895 Japanese municipalities in average (SD), stratified by the quantile values of average nighttime light (NTL).
Full size table

Geographical distribution of COVID-19 mortality during study period

Figure 2 presents the empirical Bayes estimates of COVID-19 mortality rate per 10,000 persons and Table 2 indicates the global Moran’s I values. Overall, the global Moran’s Index of spatial autocorrelation ranged from 0.084 to 0.216 (p < 0.01 across all waves), reflecting a moderate level of spatial association among contiguous municipalities of Japan. Notably, the global Moran’s I has increased in each successive wave relative to the reference period (i.e., WaveRef), indicating a strengthening of positive spatial correlation. This trend suggests that municipalities with high COVID-19 mortality rate were often spatially surrounded by other municipalities with similar high COVID-19 mortality rates, indicating spatial clustering. Particularly during Wave6, the high COVID-19 mortality rate and high Moran’s I value suggest a nationwide increase in COVID-19 deaths in Japan, characterized by a significant spatial concentration of COVID-19 mortality rate. This finding indicates the presence of spatiotemporal autocorrelations within the outcome variable, suggesting that the SDEM could theoretically achieve superior results compared to the conventional Ordinary Least Square technique, which may not produce consistent results and thus may lead to inaccurate conclusions.

Table 2 Global Moran’s I for estimating spatial autocorrelation.
Full size table

Results from SDEM by waves

Both the LM lag test and LM error test provided significant results (p < 0.01), supporting the use of SDEM over a simple linear model. The SDEM estimation results are also presented in Table 3. We observed that the total effect (i.e., the sum of direct and indirect effects) was gradually degraded across Wave3-Wave7 (degraded from 0.079 to -0.101) with positively significance during Wave3 (p = 0.013), Wave5 (p = 0.045) and Wave6 (p = 0.050), while it was negatively significant during Wave7 (p = 0.013) and non-significant during WaveRef (p = 0.844). Interestingly, focusing on the estimation of indirect (spillover) effect of NTL, the results suggest that the NTL levels in neighboring municipalities had a positively significant effect on the COVID-19 mortality rate during Wave3 (Effect estimate: 0.079 (95% CI: 0.016–0.141), p = 0.013) and a negatively significant effect during Wave7 (Effect estimate: -0.089 (95% CI: -0.169 - -0.009), p = 0.029), while no significant effects were observed during other periods. In more formal terms, these results suggest that for one unit (nanoWatts/sr/cm2) increase in NTL in the neighboring municipalities on average, it increases COVID-19 deaths by 7.9% during Wave3, and decreases by 8.9% during Wave7. The results of the subzone analysis are included in Supplementary File 4. In SubAnalysis 1, a similar tendency was observed, particularly in the Nishi-Nihon area. Specifically, a positive spillover effect was observed (though non-significant, p = 0.163) in the early stage of the pandemic (Wave 5), which shifted to a negative and significant spillover effect in the later stage of the pandemic (p = 0.014, Wave 7). In SubAnalysis 2, a similar tendency was also observed: the spillover effect changed from positive to negative over time, especially in Shikoku and Kyushu/Okinawa, which are classified as Nishi-Nihon. However, the results were not significant because the subzones were further divided into smaller subzones, reducing the sample size.

Table 3 Estimates and 95% confidence interval (CI) of direct, indirect and total effects in spatial Durbin Error Model, stratified by waves.
Full size table

Discussion

Numerous prior studies have elucidated the influence of environmental factors, such as temperature, air pollution, humidity, and the presence of green outdoor spaces, on the transmission dynamics of COVID-193,4,6,10,43. This effect is attributed to the modulation of factors such as viral stability, host susceptibility, and interpersonal contact rates3,4,7,42,43. In addition, there is emerging evidence linking NTL with both the incidence and mortality rates of COVID-197. However, to the best of our knowledge, few studies have statistically examined how the propagation effect of NTL on COVID-19 deaths in neighboring areas. In other words, it was still unclear how the illuminated area at night attracts people and spatially transmit COVID-19 to the neighboring areas.

This study has demonstrated two important findings regarding COVID-19 mortality. Initial findings revealed moderate spatial inequalities, including clusters of high COVID-19 mortality rate across Japan, closely correlated with NTL intensity (Table 1). This finding is consistent with the prior research: Zhang et al. (2022) found that compare to the first quintile of NTL value, the fifth quintile was associated with 23% higher mortality rates of COVID-19 in the U.S. during 2019 and 20207. Furthermore, our study also identified a significant increase in COVID-19 mortality rate in brighter areas after adjusting sociodemographic and environmental factors. Notably, the spillover effect of NTL showed a positive sign with COVID-19 deaths in neighboring areas during the early waves, Wave3. Conversely, during Wave7, a negative sign of the spillover effect was observed, indicating a reduction in neighboring area deaths as cities became brighter. The flip of the sign is related to differences in the backgrounds and behavioral patterns of those who contributed to the spread of the infection. The positive spillover effect during Wave3 may be explained by the “nightlife effect”, where people congregating in illuminated night-time area become infected and subsequently disseminate the virus upon returning to their localities, leading to increased mortality46,47. In contrast, the negative spillover effect during Wave7 could be attributed to high-risk individuals, such as frequenters of social gathering spots, such as nightclubs, karaoke or bars46,48, having been previously infected or vaccinated, thereby mitigating virus spread to neighboring areas despite gatherings in these bright night-time areas. Additionally, the majority of deaths during Wave7 occurred mainly among the elderly44, and limited interaction with such high-risk groups can also be considered a contributing factor. Another possible explanation is that these illuminated night-time areas are typically densely populated with huge human mobility. Therefore, stringent infection prevention measures such as hand-washing and cough-etiquette are more likely to be implemented49,50,51,52. Especially during the large-scale wave of infection Wave7 including 16,883 COVID-19 deaths, the continuation of these rigorous preventive measures might have played a key role in keeping the transmission of the virus from congregated individuals to neighboring areas at a lower level. The identification of local spillovers underscores the value of epidemiological studies focused on individual movement patterns in understanding virus transmission. More importantly, our results imply that, even in situations where it is challenging to conduct detailed contact survey during pandemic, satellite-derived NTL data can serve as a useful proxy for estimating the level of nocturnal human contact and the subsequent COVID-19 mortality rates.

This research presents several significant limitations, which are further nuanced when considering the context of the Japanese COVID-19 situation and NTL intensity. Firstly, non-human light sources (e.g., aurora, wildfires) are not easily distinguishable from human light sources. Even if we assume that the human light source is well identified, the use of NTL data obtained from satellite imagery, while a useful proxy for outdoor NTL levels, may not accurately represent indoor nightlight exposure7,53. This potential discrepancy could lead to misclassification in assessing NTL exposure54,55 and would be critical in densely populated Japanese urban areas where indoor and outdoor lighting environments differ markedly7. Secondly, employing an ecological study design at the municipality level can capture the association between NTL and COVID-19 mortality in population level. However, due to the ecological fallacy, this approach may obscure individual risk factors and personal activity in very small mesh areas that cannot be captured by satellites, thereby precluding the derivation of individual-level associations. For example, individual-level covariates, such as compliance with preventive measures (ex. mask-wearing and hand-washing) and proactivity in accessing health information were not adjusted for. The impact of these unobserved covariates may have changed as the pandemic progressed. Additionally, the prevalence of asymptotic or un-tested COVID-19 cases and related death, a notable concern in Japan during various waves during the pandemic, might affect the analytical outcomes of this study. Lastly, we used COVID-19 mortality as an outcome instead of COVID-19 incidence. While mortality data provided a robust and reliable measure, it does not capture the full spectrum of the COVID-19 pandemic’s impact, particularly in terms of infection dynamics. Future research could benefit from utilizing more granular COVID-19 incidence data, should it become available, to explore the relationship between NTL spillover and COVID-19 transmission in greater depth.

Conclusion

These findings underscore the importance of understanding spatial dynamics and human behavior in pandemic responses. Furthermore, the utilization of satellite-derived NTL data as a proxy for human interactions offers a valuable tool for epidemiological studies, especially in situations where direct data collection is challenging. This study not only contributes to the knowledge on COVID-19 transmission dynamics but also highlights the potential of innovative data sources in enhancing our response to public health emergencies.

Fig. 1
Fig. 1
Full size image

Average nighttime light (NTL) intensity (nanoWatts/sr/cm2) during the study span. This map was created using R version 3.4.1 (https://cran.r-project.org/) and zipangu package.

Fig. 2
Fig. 2
Full size image

Empirical Bayes estimates of COVID-19 mortality rate per 10,000 persons, colored by percentiles. This map was created using R version 3.4.1 (https://cran.r-project.org/) and zipangu package.

14