Geographical detector-based assessment of multi-level explanatory powers of determinants on China’s medical-service resumption during the COVID-19 epidemic

Abstract

Knowing the multi-level influences of determinants on medical-service resumptions is of great benefits to the policymaking for medical-service recovery at different levels of study units during the post-COVID-19 pandemic era. This article evaluated the hospital- and city-level resumptions of medical services in mainland China based on the data of location-based service (LBS) requests of mobile devices during the two time periods (December 2019 and from February 21 to March 18, 2020). We selected medical-service capacity, human movement, epidemic severity, and socioeconomic factors as the potential determinants on medical-service resumptions and then explicitly assessed their multi-level explanatory powers and the interactive effects of paired determinants using the geographical detector method. The results indicate that various determinants had different individual explanatory powers and interactive relationships/effects at different levels of medical-service resumptions. The current study provides a novel multi-level insight for assessing work resumption and individual/interactive influences of determinants, and considerable implications for regionalized recovery strategies of medical services.

Introduction

The novel coronavirus pneumonia (COVID-19) pandemic is still a global ongoing threat and has caused enormous shocks to the world in terms of health, socioeconomics, and natural environments. Restriction measures due to the pandemic might have some positive effects on global environment (Lal et al., 2020; Muhammad et al., 2020). More attention had been shifted from understanding and mapping COVID-19 cases (Jiang and De Rijke, 2021; Shaw and Sui, 2021) to the impacts of the pandemic on air quality, pollution changes, and others (Li et al., 2020; Sicard et al., 2020). People’s workstyle and social behaviors had also been dramatically changed during the pandemic period (An and Sun, 2021; Liu et al., 2022; Mu et al., 2021; Nathan and Overman, 2020). Global economy (e.g., transportation and tourism-related industries) had been affected by the pandemic to a great extent (Abu-Rayash and Dincer, 2020; Gössling et al., 2021; Hong et al., 2020; Norouzi et al., 2020). Effective strategies and measures were subsequently required for the recovery of various industries (Dube et al., 2021; Li et al., 2021; Zhou et al., 2021).

Many parts of the world have started to come out of the pandemic restrictions and prepared for the post-pandemic recovery. For example, both the United States and European Union have officially claimed the end of the pandemic. The resumption of work and production has become a key policymaking focus in many regions, which is an important guarantee for maintaining the economic and social stability during the post-pandemic phase. After fighting the epidemic for 2–3 months since December 2019, China had made a great success to control the large-scale epidemic spread (Kraemer et al., 2020; Tian et al., 2020; Xu et al., 2020a), and then the focus of strategies shifted to control international importations and recover production and life. Scholars began to focus on the assessments of regional resumptions of work, production, and social life on different geographical scales, such as the entire country (Lai et al., 2022; Tao et al., 2020; Tian et al., 2021; Xu et al., 2020b), provinces/municipalities (He et al., 2021; Zhang et al., 2021), and prefecture-level cities (Bai et al., 2021; Shao et al., 2021). Various strategies and measures for the work resumption and the corresponding risks were explicitly evaluated (Bai et al., 2021; Ge et al., 2021; Wang et al., 2020; Zhang et al., 2021). Moreover, the work resumption of a specific industry (e.g., medical services) and the influences of determinants can be assessed (Hu et al., 2022).

There are multiple data sources which help assess the regional work resumption. Satellite remote sensing data, in particular nighttime light (NTL) images, are beneficial and conducive to evaluating the epidemic impact on human activities, assessing the regional work resumption, and monitoring its spatiotemporal variation in large-scale areas (Liu et al., 2020; Shao et al., 2021; Tao et al., 2020; Tian et al., 2021). Satellite observations can also be combined with other multi-source data (e.g., intracity travel intensity data) to implement the assessment of work resumption (Lai et al., 2022). Nevertheless, the above data sources cannot support the assessment of work resumption in a high-resolution space-time domain. Mobile signaling data of cellphones can be a feasible support to optimize the assessment in resolution and accuracy (Liu et al., 2022). Another better alternative is the location-based service (LBS) data of mobile devices (Huang et al., 2021; Jiang and Yao, 2006), which help indicate explicit trajectories of human movement with high-resolution spatiotemporal information. The LBS data of mobile devices had been successfully applied to explore the spatiotemporal epidemic spread associated with population flow (Hu et al., 2020, 2021a) and to assess the work resumption in hospitals during the epidemic (Hu et al., 2022).

Furthermore, due to the modifiable areal unit problem (MAUP) in most geographical studies (Openshaw, 1984), the assessments of work resumption will vary depending on sampling zones/units under observation. The LBS data of mobile devices are easy to be flexibly aggregated into specific regions (e.g., administrative cities) or locations (e.g., hospitals), providing the potential to explore the work resumption for a specific industry at different levels of study units and the multi-level influences of determinants. On the other hand, multivariate statistical analysis is generally sensitive to zones/units as well (Fotheringham and Wong, 1991). The influences of consistent or similar determinants on work resumption might vary with different levels of units throughout the entire study area. The comparative analysis of the multi-level explanatory powers of determinants on work resumption can help inform more potential characteristics.

In view of the above considerations, we used the data of LBS requests of mobile devices in mainland China during the two time periods (December 2019 and from February 21 to 18 March 2020) to evaluate the hospital- and city-level resumptions of medical services during the epidemic (i.e., the study units were hospitals and cities, respectively). The geographical detector method was conducive to quantifying the explanatory powers of determinants on the explained variable without the assumption of linearity and immune to the collinearity multivariable (Wang et al., 2010, 2016). Besides, it can be applied to explore the interactive relationships and effects of paired determinants of the objective (Wang et al., 2010). Thus, it was selected to further assess and comparatively analyze the individual explanatory powers of determinants on multi-level resumptions of medical services, including fundamental medical-service capacity, human movement, epidemic severity, and socioeconomic factors. Also, the interactive effects of paired determinants at different levels of medical-service resumptions were identified and assessed. This study provides a novel multi-level perspective for the assessment of work resumption and influences of determinants, and the findings may introduce helpful information for the policymaking of recovery strategies of medical services at different levels of study units during the post-pandemic phase.

Data and methodology

LBS data of mobile devices

The data of LBS requests of mobile devices were applied to quantitatively evaluate the hospital- and city-level resumptions of medical services in mainland China during the epidemic. The LBS data used in this study were provided by Wayz Inc., Shanghai, China. The all-day LBS requests from over 80% of mobile devices supported by the three telecommunications operators in China are recorded with high-resolution locations, and the raw data collection is implemented every 2 h (Hu et al., 2020, 2021a, 2022). Note that individual mobile device with multiple LBS requests at a certain location is counted just once for a specific period during the process of data collection. Individual trajectories of mobile devices are recorded in the raw LBS data with high-resolution spatiotemporal information.

The raw LBS data covered the mobile devices which activated their LBS requests in 22,098 general or specialized hospitals, during December 2019 and from February 21 to 18 March 2020 (i.e., the experimental period in this study), respectively. The former data were aggregated into hospitals based on the geofencing technology with the averages of daily medical visits during December 2019, which were expected to be representative of their fundamental medical-service capacities before the epidemic. The latter data were subsequently aggregated into hospitals with daily medical visits during the experimental period. The comparison of medical-service situations during the epidemic with the fundamental capacities before the epidemic can indicate the work resumption in hospitals to a great extent. The cleaning and aggregation of the raw LBS data were implemented by Wayz Inc., and private individual information was deleted from the raw data before data preprocessing.

Calculation of hospital- and city-level resumption rates

A resumption rate, defined in a space-time domain, was used to be representative of the spatiotemporal medical-service resumption situations in hospitals (Hu et al., 2022). It indicates the ratio of the number of daily visits in a specific hospital s at a given date t during the epidemic to that prior to the epidemic. Let v_s,0 and v_s,t be the average of daily medical visits of hospital s during December 2019 and the daily medical visits at a given date t during the epidemic, respectively, and then the corresponding medical-service resumption can be assessed by calculating the following rate

$$y_{\mathit{s},t}=\frac{v_{\mathit{s},t}}{v_{\mathit{s},0}}$$ (1)

where y_s,t is the medical-service resumption rate of hospital s at date t.

It is worth noticing that v_s,t indicates the spatiotemporal distribution of the medical-service visits in hospitals during the epidemic. Thus, the resumption rate in hospitals, y_s,t, varied over space and time (Supplementary Figure S1a), and hospitals with the y_s,t estimates greater than 1 at a certain date were considered to have resumed their normal activities with more visits than those prior to the epidemic.

Based on the above calculation of the medical-service resumption rate in hospitals, we can further assess the medical-service resumption in China’s cities. The estimates of the resumption rates in all hospitals of a specific city at a given date can be averagely aggregated into the corresponding city. Therefore, the resumption rate in cities, which also varied over space and time, can be subsequently generated to indicate the spatiotemporal resumption situations of medical services in cities (Supplementary Figure S1b). In this study, the hospital- and city-level resumption rates of medical services were considered as two explained variables (i.e., multi-level resumption rates), and the multi-level explanatory powers of determinants were subsequently assessed.

Human movement data

Intercity human movement was expected to have potential influence on regional medical-service resumption. Two proxy variables were selected to indicate the human movement, including daily imported visits from Wuhan, China, which was the epidemic source of the large-scale outbreak, and daily imported visits from elsewhere excluding Wuhan, respectively. The human movement data were acquired from the LBS data of mobile devices, which activated their LBS requests in locations away from their attributions. The daily imported visits from Wuhan and from elsewhere to hospitals were generated according to hospital and date, respectively. Next, they were averagely aggregated into cities by date.

Epidemic data

We collected the spatiotemporal data of daily new confirmed cases from multiple official and publicly available sources and comparatively verified the epidemic data through the public platform of the 2019-nCoV-infected pneumonia epidemic (China CDC, 2021). The epidemic data of daily new confirmed cases included the spatial information of their residential districts and were associated with individual hospitals according to location information based on the geofencing technology. We selected the daily new confirmed cases within the 3 km buffer around hospitals to be a proxy variable of the epidemic severity (Hu et al., 2022). Similarly, the data of daily new cases around hospitals were averagely aggregated into cities by date. Furthermore, an extra proxy variable of epidemic severity, that is, daily cumulative confirmed cases in cities, was selected for the city-level resumption rate of medical services.

Socioeconomic data

Three extra socioeconomic explanatory variables, including population (POP), gross regional product (GRP), and per capita disposable income (PCDI) of cities, were further selected for the city-level resumption rate of medical services. The observations of these socioeconomic variables at the end of 2019 or in 2019 were collected from the 2020 China Statistical Yearbook. A more detailed description of the human movement, epidemic, and socioeconomic data can be found in Supplementary Material.

Geographical detector

Here, we introduce spatially stratified heterogeneity (SSH) to describe the variations of the hospital- and city-level resumption situations of medical services according to various stratifications. SSH refers to ubiquitous phenomena which describe that the within-strata variance is less than the between-strata variance, implies potential distinct mechanisms by stratum, and enforces the applicability of statistical inferences (Wang et al., 2016). Note that the stratification referring to the SSH of an explained variable can be either a geographical division, or determined by a categorical or numerical explanatory variable. The geographical detector q-statistic was firstly developed to quantify the SSH of an explained variable according to a stratification (Wang et al., 2010, 2016) and then was universally applied to assess the explanatory powers of determinants on the explained variable. The fundamental formula of the geographical detector q-statistic is given by

$$\begin{array}{c}q=1-\frac{{\sum }_{h=1}^L{N}_h{\sigma }_h^2}{N{\sigma }^2}\end{array}$$ (2)

where N is the number of the observations of an explained variable throughout the entire study area and σ² denotes the variance of all observations. The explained variable is stratified into L strata, denoted by h = 1, 2, …, L, which is based on a geographical division or determined by an explanatory variable. N_h is the number of observations within stratum h, and ${\sigma }_h^2$ denotes the corresponding variance. The q-statistic, ranging from 0 to 1, indicates the SSH measure of the explained variable depending on a specific stratification, or more specifically, the explanatory power of an explanatory variable on the explained variable, which can be interpreted as it explaining 100 × q% of the SSH of the explained variable.

Furthermore, the geographical detector method also provides an interaction detector for two or more explanatory variables (Wang et al., 2010), which was widely applied to reveal the interactive effects of paired variables (e.g., Hu et al., 2021b; Luo et al., 2016; Xu et al., 2021; Yin et al., 2019). Let u and v be a pair of explanatory variables, and by overlaying u and v, we can calculate the q-statistic of their interactive effect, q_u∩v. Note that the symbol “∩” can indicate either a spatial intersection of two geographical stratifications or a generalized overlaying operation of two variables. Next, by comparing the interactive q_u∩v with q_u and q_v, five types of interactive relationships between variables u and v can be identified (Supplementary Table S1).

More specifically, when the interactive q_u∩v is greater than the sum of q_u and q_v, two variables u and v nonlinearly enhance each other. In order to further assess the variation of nonlinear enhancements from different pairs of variables, we defined the following two indicators to measure the absolute and relative increments, respectively

$$\begin{array}{c}∆q_{u\cap v}=q_{u\cap v}-\left(q_u+q_v\right)\end{array}$$ (3)

$${\xi }_{u\cap v}=\frac{∆q_{u\cap v}}{\left(q_u+q_v\right)}=\frac{q_{u\cap v}-\left(q_u+q_v\right)}{\left(q_u+q_v\right)}$$ (4)

where Δq_(u∩v) denotes the absolute increment of a nonlinear enhancement between u and v, and ξ_(u∩v) is the relative one. A higher value of Δq for paired variables indicates their stronger nonlinearly interactive enhancement. If a variable always receives higher ξ values of nonlinearly interactive enhancements with others, it has the potential to be a dominant interactive determinant to explain the SSH of the explained variable.

Experimental setup

The experimental period was set from February 21 to 18 March 2020, with a time step of 1 day. The daily medical-service resumption rates in hospitals were calculated according to equation (1) and then were aggregated into the daily resumption rates in cities. Various determinants were selected for assessing the explanatory powers on the hospital- and city-level resumption rates of medical services, respectively. As shown in Supplementary Table S2, average daily visits before the epidemic (x₁), imported visits from Wuhan (x₂) and from elsewhere (x₃), and new confirmed cases around hospitals (x₄) were considered as the consistent explanatory variables of multi-level resumption rates. In addition, two geographical divisions were selected to be the extra explanatory variables of hospital-level resumption rate, including province and city stratifications (x₅ and x₆). The former considered provinces, municipalities, and autonomous regions as strata, whereas the latter considered cities as strata. Regarding city-level resumption rate, we further selected daily cumulative confirmed cases (x₇), POP (x₈), GRP (x₉), and PCDI (x₁₀) of cities as its additional explanatory variables.

In order to reduce the subjective influence of various stratifications to the calculation of q-statistics, we implemented the stratification of each numerical explanatory variable by equal-interval division after ordering data, and then equally divided the observations of resumption rates into 10 strata. According to equation (2), the daily q-statistics of explanatory variables to hospital- and city-level resumption rates were calculated, respectively. The statistical significance of q-statistics can be tested based on equations (S1) and (S2) in Supplementary Material. The geographical detector software is publicly accessible at http://geodetector.cn/, and the q-statistics in this study were performed with the use of the R software package. Moreover, the daily interactive q-statistics of paired variables were calculated and compared with two individual q-statistics, and the corresponding interactive relationships were identified based on Supplementary Table S1. While a pair of variables exhibited the interactive relationship of a nonlinear enhancement, their increment indicators, Δq and ξ, can be further calculated according to equations (3) and (4).

Results

Variation of the explanatory powers of determinants on multi-level resumption rates

The daily hospital- and city-level resumption rates of medical services were explicitly evaluated from February 21 to 18 March 2020. Both of them exhibited an obvious ascending trend during the entire period. Areas surrounding Hubei province in central China had a relatively normal resumption of medical services until March 18 (Supplementary Figure S1). Before March 2020, there were approximately 23.3% of hospitals and 18.7% of cities which had the medical-service resumption rates higher than 1, respectively. Nevertheless, the numbers increased to 30.8% and 34.9% after the first week in March and became 41.8% and 59.2% at the experimental end date (March 18). After 2–3 months of fighting the epidemic, nearly half of hospitals had achieved relatively nice resumption of work and over half of cities had resumed their medical-service situations.

We calculated the daily q-statistics of various explanatory variables to hospital- and city-level resumption rates, respectively. As shown in Table 1, imported visits from Wuhan (x₂) received an average q-statistic value of 0.1000 to city-level resumption rate, with a standard deviation (SD) of 0.0539, whereas imported visits from elsewhere (x₃) achieved an average value of 0.1641 with a SD of 0.0440. Note that their q-statistics to hospital-level resumption rate were extremely low (q = 0.0174 and q = 0.0267). Human movement had obviously much stronger influence on city-level resumption rate than on hospital-level resumption rate, and its two proxy variables can averagely explain 10.00% and 16.41% of the SSH of city-level resumption rate, respectively. Average daily visits before the epidemic (x₁) and new cases around hospitals (x₄) had no statistically significant explanatory powers on city-level resumption rate. However, they achieved significant q-statistics to hospital-level resumption rate since the beginning of March; their significant percentages with an alpha level of 0.05 were 62.96% and 66.67%, respectively. Although the explanatory powers were still weak, the medical-service capacity and epidemic severity had started to impact the work resumption in hospitals.

Table 1.

Descriptive statistics of daily q-statistics to multi-level resumption rates.

Resumption rate	Variable	Mean	Min	Max	SD	Significant percentage†
Hospital-level	x1	0.0034	0.0003	0.0203	0.0043	62.96
	x2	0.0174	0.0044	0.0347	0.0093	100
	x3	0.0267	0.0106	0.0583	0.0131	100
	x4	0.0016	0.0001	0.0056	0.0015	66.67
	x5	0.0218	0.0027	0.0706	0.0215	100
	x6	0.0709	0.0281	0.1703	0.0445	100
City-level	x1	0.0264	0.0155	0.0497	0.0101	0
	x2	0.1000	0.0210	0.1857	0.0539	70.37
	x3	0.1641	0.0944	0.2205	0.0440	100
	x4	0.0281	0.0124	0.0630	0.0141	0
	x7	0.1023	0.0763	0.1276	0.0143	100
	x8	0.0604	0.0250	0.1410	0.0289	51.85
	x9	0.1449	0.0524	0.2965	0.0740	96.30
	x10	0.1083	0.0410	0.1979	0.0542	81.48

^†An alpha level of 0.05.

The q-statistics of province and city stratifications (x₅ and x₆) to hospital-level resumption rate received low average values (q = 0.0218 and q = 0.0709) but were always significant during the entire period (Table 1). Nevertheless, both of them exhibited a consistent temporally increasing tendency, and the q-statistics of x₆ were always higher than those of x₅ (Supplementary Figure S2a). The maximum q-statistic of city stratification reached 0.1703. The variation of hospital-level resumptions between cities gradually increased by date. As an extra proxy variable of epidemic severity on city-level resumption rate, cumulative cases in cities (x₇) achieved significant q-statistics with an average of 0.1023 and a SD of 0.0143. Its daily q-statistics exhibited no obvious increasing or decreasing tendency, and the explanatory power was relatively steady over time (Supplementary Figure S2b). Note that new cases around hospitals (x₄) had low-value insignificant q-statistics to city-level resumption rate. The entire epidemic situation of cities played a great role in affecting city-level resumptions. Regarding three socioeconomic variables of city-level resumption rate, as shown in Table 1, the q-statistics of POP (x₈) began significant since the first week in March (the significant percentage was 51.85%), whereas those of GRP (x₉) and PCDI (x₁₀) were nearly significant during the entire period (the percentages were 96.30% and 81.48%). GRP (x₉) and PCDI (x₁₀) can explain averagely 14.49% and 10.83% of city-level resumption rate, respectively. Both of them had a consistent increasing trend of explanatory powers, especially when entering the first week in March (Supplementary Figure S2b); their increasing trends began rather rapid and then they reached the maximum q-statistics of approximately 0.3 and 0.2, respectively.

In general, various determinants had relatively weak explanatory powers on hospital-level resumption rate, whereas the explanatory powers of two geographical divisions started to increase since the beginning of March, indicating the regional disparity of hospital-level resumptions gradually increased by date. Meanwhile, human movement, epidemic severity, and socioeconomic factors played a great role in explaining the SSH of city-level resumption rate. Nevertheless, their explanatory powers exhibited different temporal variations. For instance, GRP (x₉) and PCDI (x₁₀) had the explanatory powers with an increasing tendency by date, whereas cumulative cases (x₇) had a relatively steady explanatory power.

Interactive effects of paired determinants on hospital-level resumption rate

We further calculated the interactive q-statistics of each pair of explanatory variables and revealed their interactive relationships and effects on multi-level resumption rates. As shown in Table 2, the interactive relationships of a bivariate enhancement appeared in the majority of interactions on hospital-level resumption rate (i.e., the interactive q-statistic is greater than either of two individual ones but smaller than their sum). In particular, the interactive q-statistic of province and city stratifications (x₅ and x₆) was smaller than either of two individual ones, and thus, they were nonlinearly weakened by one another. Nevertheless, they always introduced a nonlinear enhancement when interacting with others (i.e., the interactive q-statistic is greater than the sum of two individual ones). More specifically, x₆ played a much more important role in interactive effects than x₅. It provided the dominant interactive powers on hospital-level resumption rate, especially when interacting with x₁, x₂, and x₃. As shown in Table 2, those interactions received interactive q-statistics with the averages of 0.2130 (SD = 0.0409), 0.2710 (SD = 0.0823), and 0.2911 (SD = 0.0807), respectively, which were much greater than the sums of two corresponding individual ones.

Table 2.

Descriptive statistics of daily interactive q-statistics to hospital-level resumption rate.^†

Variable	x1	x2	x3	x4	x5	x6
x1	0.0034 (0.0043)
x2	0.0314 (0.0165)	0.0174 (0.0093)
x3	0.1158 (0.0614)	0.046 (0.0175)	0.0267 (0.0131)
x4	0.0059 (0.0053)	0.0232 (0.0126)	0.0305 (0.0148)	0.0016 (0.0015)
x5	0.0381 (0.0252)	0.0585 (0.0273)	0.0608 (0.0281)	0.0309 (0.0269)	0.0218 (0.0215)
x6	0.2130 (0.0409)	0.2710 (0.0823)	0.2911 (0.0807)	0.0887 (0.0512)	0.0709 (0.0445)	0.0709 (0.0445)

^†Standard deviations are listed in parentheses.

The interactive matrix of daily q-statistics of paired determinants to hospital-level resumption rate is demonstrated in Figure 1. Note that the diagonal subplots show the boxplots of daily q-statistics for each explanatory variable, whereas the upper and lower triangular ones show the boxplots and temporal curves of daily interactive q-statistics for each pair of variables, respectively. Several lower and narrower “boxes” can be found in the upper triangular area (e.g., x₁∩x₄), indicating relatively weak and stable interactive enhancements. Contrarily, higher and wider “boxes” can indicate relatively strong and unstable interactive enhancements (e.g., x₂∩x₆). Obviously, x₆ interacting with x₁, x₂, and x₃ introduced higher and wider “boxes” in the boxplots of interactive q-statistics. Their interactions derived strongly nonlinear enhancements which were unstable during the entire period. Besides, their temporal curves in the lower triangular area show a consistent temporally increasing tendency with fluctuation, and the interactions of x₂∩x₆ and x₃∩x₆ fluctuated more intensely than that of x₁∩x₆ (Figure 1).

Figure 1.

Interactive matrix of daily q-statistics to hospital-level resumption rate.

We further calculated the increment indicators, Δq and ξ, to measure the variations of nonlinear enhancements for the interactions of x₁∩x₆, x₂∩x₆, x₃∩x₆, and x₁∩x₃, respectively. As shown in Figure 2(a) and (b), ranked by Δq, it was found that Δq(x₃∩x₆) = 0.1935>Δq(x₂∩x₆) = 0.1828>Δq(x₁∩x₆) = 0.1387, whereas ranked by ξ, it was found that ξ(x₂∩x₆) = 3.4026>ξ(x₃∩x₆) = 3.24>ξ(x₁∩x₆) = 2.784. Note that their relative increments were approximately 3, indicating that the explanatory powers introduced by their nonlinear enhancements were three times more than the sum of two individual powers. City stratification (x₆) received both absolute and relative strong increments when interacting with others, and was identified as a dominant interactive determinant to explain hospital-level resumption rate.

Figure 2.

Nonlinearly interactive enhancements of determinants to hospital-level resumption rate: (a, b) absolute and relative increments of city stratification interacting with others and (c) temporal increments of medical-service capacity interacting with human movement.

Moreover, as shown in Table 2, it is worth noticing that the pair of x₁ and x₃ received interactive q-statistics with an average value of 0.1158 (SD = 0.0614), which was much greater than the sum of two individual ones (q = 0.0034 for x₁ and q = 0.0267 for x₃). Medical-service capacity and human movement nonlinearly enhanced each other to a very great extent, and their interaction can averagely explain 11.58% of the SSH of hospital-level resumption rate. Nevertheless, although the interaction of x₁∩x₃ derived a strongly nonlinear enhancement, this enhancement was still not steady (the “box” is wide) and exhibited a gradually increasing trend during the entire period (Figure 1). Besides, their absolute increment, Δq(x₁∩x₃), exhibited a decreasing trend by date, whereas their relative increment, ξ(x₁∩x₃), showed a gradually increasing trend (Figure 2c). Although both x₁ and x₃ had extremely weak individual explanatory powers, their pair introduced a nice interactive effect with a relatively large explanatory power and a gradually increasing increment; thus, the interaction of x₁∩x₃ can be identified to be an abnormal one.

Interactive effects of paired determinants on city-level resumption rate

The interactive q-statistics of paired determinants to city-level resumption rate are listed in Table 3. All of them introduced strongly nonlinear enhancements in explaining city-level resumption rate, while one interacting with another. The interactions of the selected determinants were conducive to explaining the SSH of city-level resumption rate to a great extent. The most dominant interactive effect was x₃ and x₁₀, with an average q-statistic of 0.5656 (SD = 0.1242). Their interaction of x₃∩x₁₀ can averagely explain 56.56% of the SSH of city-level resumption rate. Note that the least interactive effect still achieved an average q-statistic of 0.1899 (SD = 0.0536), which was x₁ interacting with x₄, with two extremely small individual ones of 0.0264 and 0.0281.

Table 3.

Descriptive statistics of daily interactive q-statistics to city-level resumption rate.^†

Variable	x1	x2	x3	x4	x7	x8	x9	x10
x1	0.0264 (0.0101)
x2	0.3534 (0.1026)	0.1000 (0.0539)
x3	0.4004 (0.0356)	0.4392 (0.0658)	0.1641 (0.044)
x4	0.1899 (0.0536)	0.2431 (0.0851)	0.3429 (0.0452)	0.0281 (0.0141)
x7	0.3295 (0.0709)	0.4295 (0.0947)	0.4480 (0.0433)	0.1932 (0.0628)	0.1023 (0.0143)
x8	0.3099 (0.0448)	0.3778 (0.0797)	0.3836 (0.0373)	0.2763 (0.0493)	0.2677 (0.0289)	0.0604 (0.0289)
x9	0.3769 (0.0920)	0.3855 (0.1069)	0.4805 (0.0682)	0.3553 (0.0685)	0.3490 (0.0836)	0.2792 (0.1002)	0.1449 (0.0740)
x10	0.4411 (0.0586)	0.3939 (0.0669)	0.5656 (0.1242)	0.2370 (0.0897)	0.4348 (0.0286)	0.3427 (0.092)	0.3985 (0.0450)	0.1083 (0.0542)

^†Standard deviations are listed in parentheses.

Human movement, epidemic severity, and socioeconomic factors had nice individual exploratory powers on city-level resumption rate and still provided dominant interactive effects while interacting with others (Table 3). x₂ and x₃ received interactive exploratory powers of averagely 37.46% and 42.49%, respectively, whereas the number of x₇ was 33.51%. The averages of interactive effects provided by x₈, x₉, and x₁₀ were 32.76%, 37.50%, and 40.19%, respectively. They were the dominant interactive determinants to explain the SSH of city-level resumption rate. Moreover, attention should be paid to the pairs of two human movement variables (x₂∩x₃) and two epidemic severity ones (x₄∩x₇). The former pair derived an interactive q-statistic with an average of 0.4392, whereas the number of the latter pair was 0.1932.

The interactive effects of paired determinants on city-level resumption rate exhibited different varying temporal trends during the entire period. As shown in Figure 3, the interactions between x₇, x₈, x₉, and x₁₀ derived strongly nonlinear enhancements, which were relatively stable by date or exhibited a gradually increasing tendency with less fluctuation (narrower “boxes” in the boxplots and increasing temporal curves). Nevertheless, some relatively intense fluctuations appeared in the interactions while x₂ or x₃ interacting with others (wider “boxes” in the boxplots and fluctuating temporal curves without obviously increasing or decreasing trends). In other words, the influence of human movement on city-level resumption rate was unsteady during the entire period, whereas the interactive influences of epidemic severity and socioeconomic factors were always steady by date or gradually getting strengthened.

Figure 3.

Interactive matrix of daily q-statistics to city-level resumption rate.

Subsequently, we focused on the nonlinearly interactive increments of x₇ interacting with other variables, and calculated the corresponding increment indicators. As shown in Figure 4, ranked by Δq, it was found that Δq(x₂∩x₇) = 0.2272>Δq(x₇∩x₁₀) = 0.2242>Δq(x₁∩x₇) = 0.2008>Δq (x₃∩x₇) = 0.1816 (the rest were less than 0.11), whereas ranked by ξ, it was found that ξ(x₁∩x₇) = 1.5650>ξ(x₇∩x₁₀) = 1.2351>ξ(x₂∩x₇) = 1.2054 (the rest were less than 0.72). By controlling the influence of cumulative cases in cities, city-level resumption rate was primarily affected by medical-service capacity, PCDI, and human movement. Note that Δq(x₄∩x₇) and ξ(x₄∩x₇) were both the minimums of all Δq and ξ values; the limited nonlinearly interactive increment of two epidemic severity variables might be caused by the collinearity existing between them.

Figure 4.

Nonlinearly interactive enhancements of cumulative cases interacting with others to city-level resumption rate: (a) absolute increments and (b) relative increments.

At last, the individual explanatory powers of x₁ and x₄ were both extremely small, but the nonlinearly interactive increments of their interaction were large (Δq(x₁∩x₄) = 0.1354 and ξ(x₁∩x₄) = 2.7311); the enhancement was nearly three times more than the sum of individual effects, and thus, the interaction of x₁∩x₄ can be identified to be an abnormal one. Also, the increments caused by x₁ or x₄ interacting with others can be assessed, for example, Δq(x₁∩x₂) = 0.2270 and ξ(x₁∩x₂) = 2.2041, and the enhancement was over two times more than the sum of individual effects.

Discussion on the implications of this study

The LBS data used in this study contain substantial spatiotemporal information and cover the majority of mobile devices in China. High spatiotemporal resolution and large-scale regional coverage of the LBS data provide the potential to implement the assessments of multi-level work resumptions with different study units in a high-resolution space-time domain. This study applied the LBS data to calculate the hospital- and city-level resumption rates of medical services from February 21 to 18 March 2020, during the pandemic. Note that several “outliers” appeared in both hospital- and city-level resumptions, which had the rates greater than 1.5 or even 2 (Supplementary Figure S1). The substantial increase of medical-service visits than those prior to the epidemic may not represent the resumption of medical-service situations but had the potential to be caused by other abnormal factors. For instance, several hospitals were designated as the quarantine and isolation centers, and the increase of confirmed cases intensified the medical-service burdens of several hospitals during the epidemic.

We subsequently used the geographical detector method to quantify the explanatory powers of determinants and the interactive effects of paired determinants on multi-level resumption rates. That is to say, the multi-level insight was provided in this study for assessing both medical-service resumptions and explanatory powers of determinants. It is worth noticing that the explanatory powers and the interactions of same determinants varied from level to level on medical-service resumptions, which can be explained by the MAUP effect and sensitivity of the multivariate statistical analysis depending on study units. Moreover, the findings help us inform different strategies for the recovery measures of medical services at different levels. For instance, more attention should be paid to the interaction of medical-service capacity and human movement when informing the hospital-level strategies of medical-service recovery. When dealing with city-level recovery strategies, the interactions of human movement, epidemic severity, and socioeconomic factors could be the primary criteria leading to varying degrees of recovery measures. Since there was a specific outbreak center during the early epidemic era in China, human movement played a great role in explaining the spatiotemporal epidemic spread and multi-level resumptions of medical services. Other regions which had a similar characteristic of the epidemic spread should focus on human movement and its interactions with other key factors, when informing the recovery strategies.

The geographical detector method used in this study was conducive to measuring the explanatory powers of individual determinants, identifying the interactive relationships of paired determinants, and assessing the interactive effects on multi-level resumptions of medical services. It had been applied in plentiful previous studies to explore the interactive relationships and effects of paired determinants of the objective (e.g., Hu et al., 2021b; Luo et al., 2016; Wang et al., 2010; Xu et al., 2021; Yin et al., 2019; Zhang et al., 2019). In this study, however, they were identified and comparatively explored at different levels of medical-service resumptions. The interactive relationship of a bivariate enhancement was found in the majority of hospital-level interactions, whereas all city-level interactions were identified as the interactive relationship of a nonlinear enhancement. The most dominant hospital-level interaction just reached an explanatory power of 29.11%, but the majority of city-level interactions achieved an explanatory power of over 30%. More specifically, the most dominant interactive effect can explain 56.56% of the SSH of city-level resumption rate.

Another potential contribution of this study is the measure of the absolute and relative increments for the interactions of a nonlinear enhancement. It can help identify the dominant interactive factors (e.g., city stratification was found to be a dominant interactive determinant to explain hospital-level resumption rate). Moreover, it provides the potential to detect abnormal interactions. For instance, x₁ interacting with x₃ on hospital-level resumption rate and interacting with x₄ on city-level rate were identified as abnormal interactions, in which two individual explanatory powers were extremely small but the interactive increment of the nonlinear enhancement was very large. See Supplementary Material for the discussion of limitations.

Conclusions

This article provides a novel multi-level perspective for evaluating medical-service resumptions during the post-epidemic era based on the LBS data of mobile devices and assessing the explanatory powers of determinants and interactive effects of paired determinants on multi-level resumption rates using the geographical detector method. Various determinants had different explanatory powers on multi-level resumption rates. Human movement, epidemic severity, and socioeconomic factors played a great role in explaining city-level resumption rate, whereas city stratification had an increasing explanatory power on hospital-level resumption rate. The individual explanatory powers of same determinants varied on multi-level resumption rates. Human movement and epidemic severity can explain the SSH of city-level resumption rate to some extent, but their explanatory powers on hospital-level resumption rate were extremely weak.

Different types of interactive relationships of paired determinants appeared in explaining multi-level resumption rates. The majority of interactions derived bivariate enhancement effects on hospital-level resumption rate, whereas all interactions introduced strongly nonlinear enhancements on city-level resumption rate. The interactive effects of paired determinants varied on multi-level resumption rates as well. City stratification was the most dominant interactive determinant on hospital-level resumption rate, with a maximum interactive effect of 29.11%. Human movement, epidemic severity, and socioeconomic factors provided dominant interactive effects on city-level resumption rate. Their interactive effects were always higher than 30% and even reached a maximum of 56.56%. The interactive effects on multi-level resumption rates also exhibited different temporal variations. The nonlinearly interactive enhancements of epidemic severity and socioeconomic factors on city-level resumption rate were relatively stable by date or exhibited an increasing tendency with less fluctuation. Nevertheless, the influences of human movement on city-level resumption rate and city stratification on hospital-level resumption rate were unstable or fluctuating during the entire period.

Attention should be paid to several abnormal interactions of paired determinants on multi-level resumption rates, which had two extremely small individual explanatory powers but a very large interactive increment of the nonlinear enhancement. Medical-service capacity of hospitals produced abnormal interactions on both hospital- and city-level resumption rates with another determinant.

Biographies

Prof. Bisong Hu is a full professor in School of Geography and Environment, Jiangxi Normal University, Nanchang, China. His research interests include spatial statistics, spatial epidemiology, epidemic spread simulation, and spatial-temporal big data analysis. Email: hubisong@jxnu.edu.cn

Miss Sumeng Fu is currently a master candidate in Jiangxi Normal University, Nanchang, China. Her research interests include spatial epidemiology and human geography. Email: fusumeng@jxnu.edu.cn

Dr. Qian Yin is an associate professor in the Institute of Geographic Sciences and Nature Resources Research, Chinese Academy of Sciences, Beijing, China. Her research interests include spatial statistics and spatial epidemiology. Email: yinq@lreis.ac.cn

Prof. Hui Lin is a full professor and the dean of School of Geography and Environment, Jiangxi Normal University, Nanchang, China. His research interests include spatial database and data mining, microwave remote sensing image processing and analysis, and virtual geographical environments. Email: huilin@cuhk.edu.hk

Dr. Vincent Tao works at Wayz AI Technology Company Limited, Shanghai, China, as the founder and chairman. He specializes in spatiotemporal big data analysis. Email: vincent.tao@wayz.ai

Prof. Bin Jiang is a full professor of computational geography at Faculty of Engineering and Sustainable Development (Division of GIScience) of the University of Gävle, Sweden. His research interests center on geospatial analysis of urban structure and dynamics, or geospatial big data in general. Email: bin.jiang@hig.se

Dr. Jin Luo is an associate professor and the execute dean of School of Geography and Environment, Jiangxi Normal University, Nanchang, China. He specializes in spatial database and data mining. Email: luojin@jxnu.edu.cn

Dr. Lijun Zuo is an associate professor in Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. Her research interests include land-use change monitoring, and remote sensing for the sustainable use of cropland. Email: zuolj@radi.ac.cn

Prof. Yu Meng is a Chair Professor in Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. She specializes in spatial-temporal big data analysis. Email: mengyu@radi.ac.cn

ORCID iDs

Bisong Hu https://orcid.org/0000-0003-3875-8792

Bin Jiang https://orcid.org/0000-0002-2337-2486