Spatiotemporal Evolution and Drivers of Total Health Expenditure across Mainland China in Recent Years

A substantially growing health expenditure has become an important global issue. Thus, how and why health expenditure is rising should be urgently investigated in systematic research. The Bayesian space-time model and the Bayesian least absolute shrinkage and selection operator (LASSO) model were employed in this study to investigate the spatiotemporal trends and influence patterns of total health expenditure per capita (THEPC) and total health expenditure (THEE) as a share of the gross domestic product (GDP) on the Chinese mainland from 2009 to 2018. The spatial distribution of THEE as a share of GDP in mainland China has shaped a distinct geographical structure with the characteristic of ‘west high/east low’. Its local increasing trends formed a geographical structure that exhibited a ‘north high/south low’ feature. The heterogeneity of the influence patterns of health expenditure was observed from east to west across China. Natural environmental factors, such as air pollution and green coverage, along with changes in dietary structures, have increasingly influenced the growth of health expenditures.


Introduction
Globally, aggregate health expenditure has increased substantially in the past 20 years in developed and developing countries [1][2][3], and it accounts for more than 10% of the global gross domestic product (GDP) [4]. This trend is even more critical in middle-income countries [5]. Global health expenditure in real terms increased by 3.9% annually from 2000 to 2017, whereas the economy grew by 3.0% [5]. Fostering the sustainable growth of health expenditures has become an important issue for all countries [6]. This situation has also put pressure on policymakers and academics to understand how and why health expenditures are rising [7].
In the populous country of China, health expenditure has also risen quickly in recent years; the annual growth rate of the country's total health expenditure (THEE) reached 14.71% in 2009-2017, exceeding that of its GDP (11.30%) [8]. Consequently, an increase in THEE as a share of GDP appeared in the same period [8]. Economic development and population distribution in China are unbalanced [9]. Although countrywide growth in THEE is evident in China, the spatial and temporal disparities in the variation in THEE need to be investigated in depth. A deep understanding of the spatiotemporal trends in health expenditures in China can provide a valuable reference for policymakers. To our knowledge, only a few studies have explored the spatiotemporal patterns of health expenditures at the sub-national level in China.
Understanding the main driving factors of health expenditure can help the government to scientifically plan the future of the healthcare system [8]. The determinants or influential factors of health expenditure have been researched in numerous studies, and the influence of economic factors on the THEE per capita (THEPC) has been widely investigated. Researchers have reached a consensus that increasing income can boost health expenditures. As early as 1977, Newhouse pointed out that national income can explain 90% of the variations in THEPC in developed countries [10]. Further, recent studies have concluded that economic growth and financial development significantly increase health expenditure [1,4,[11][12][13][14][15][16][17]. Beyond economic factors, the influences of non-economic factors (e.g., population ageing, air pollution, medical technology) have been researched [18][19][20][21]; such studies have mainly focused on pooled sample data of Organization for Economic Co-operation and Development (OECD) countries in the twentieth century [22][23][24][25][26]. Indeed, Martin et al. (2011) [27] reviewed 20 relevant primary studies on the determinants of health expenditure, but they indicated that no single pattern of results could be clearly identified.
Comprehensive evidence on the determinants of health expenditures in China is limited. Hou et al. (2020) [8] explored the factors influencing the Chinese THEPC at the sub-national level based on spatiotemporal panel data across 31 provinces from 2009 to 2016. However, the factors considered in the present study-including four category factors (economic factors, population ageing, epidemiology, and number of beds) and natural factors (e.g., air pollution and climate change)-were not considered in Hou et al.'s study; neither were the interactions of the different factors determined.
In light of the lack of previous research, our study has two targets. First, this study explores the spatiotemporal trends of THEPC and THEE as a share of GDP in recent years in mainland China at the provincial level. Second, the influence patterns of the three categories of factors affecting THEPC and THEE as a share of GDP at the national and subnational levels on the Chinese mainland are investigated.

Three Subnational Areas of Chinese Mainland
The study area in this paper is the Chinese mainland. Due to the unavailability of data, Taiwan Province, Hongkong, and Macao were not included. On the basis of provincial scale, the Chinese mainland can be divided into three subnational regions: eastern, middle, and western China ( Figure 1). The locations of the 31 provinces included in this study are shown in Figure 1. The three subnational areas represent developed, moderately developed, and underdeveloped economic regions, respectively.

Variables Selection and Data Sources
The data on THEPC (×1000 Chinese yuan) and THEE as a share of the GDP (%) in this study were collected from the China Health Statistical Yearbook (http://www.nhc.gov.cn/mohwsbwstjxxzx/tjtjnj/new_list.shtml, accessed on 16 August 2022), which covered 31 provincial regions of the Chinese mainland in corresponding

Bayesian Spatiotemporal Hierarchy Model
To investigate the spatiotemporal trends of THEPC and THEE as a share of GDP in China, a Bayesian spatiotemporal hierarchical model (BSTHM) [31]

Bayesian Spatiotemporal Hierarchy Model
To investigate the spatiotemporal trends of THEPC and THEE as a share of GDP in China, a Bayesian spatiotemporal hierarchical model (BSTHM) [31] was used in this study. The two outcome variables can both be regarded as continuous variables. Although the initial value of the THEE share of GDP is at (0, 1), we modelled its percentage multiplied by 100. Consequently, the likelihood distributions of the two dependent variables were all assigned a log-normal distribution, expressed as follows: THEE as a share of GDP : where Y it and y it represent THEPC and THEE as a share of GDP of the i-th provincial region in the t-th year; µ it and γ it are the corresponding expectancy values; and σ 2 and ∂ 2 are the corresponding variances, whose priors were assigned with Gamma distributions. The spatiotemporal process model is expressed as follows: THEE as a share of GDP : where t * and T represent the middle and length of the study period, respectively; T is 10 in this paper. α (r) and α (P) are the basic fixed constants for THEPC and THEE as a share of GDP, respectively; their corresponding priors used non-informative prior distributions. SRM (r) and SRM (p) represent the spatial relative magnitudes (SRM) of THEPC and THEE as a share of GDP, comparing them to the national overall baseline level. β (r) 1i , and β (p) 1i were assigned using the Besag York Mollie (BYM) model [32], expressing the spatial structured and unstructured random effects. The spatial adjacency matrix adopts the first-order 'Queen' adjoining form; ε (r) it and ε (p) it represent the Gaussian random terms, and their priors were assigned with Gaussian distribution with zero mean [33].

Bayesian LASSO Regression Model
The Bayesian LASSO (B-LASSO) model [34] was adopted in our study to overcome the problem of multicollinearity among the variables. The B-LASSO model is the Bayesian version of the ordinary LASSO regression model [35]. In the B-LASSO model, the likelihood distributions of THEPC and THEE as a share of GDP were also all assigned to the log-normal distribution, and the regression parameters were assigned to the independent and identical Laplace priors. Bayesian LASSO estimations differ from ordinary least squares (OLS), penalized by the least squares method, which minimizes the residual sum of squares while controlling the L1 norm of the coefficient vector of regression [34,35]. The B-LASSO model can acquire a more stable estimation and automatically provide interval estimates for all parameters, including the error variance [34]. The B-LASSO model of the relationship among the influence factors, x j (j = 1, . . . , n), n is the number of the influence factors, set as 10 in this paper, and the outcome variables THEPC and THEE as a share of GDP. The mathematical structure can be expressed as follows: THEE as a share of GDP : where ϕ (r) and ϕ (p) are overall intercepts whose priors were assigned to non-informative priors. s it represent the random error whose priors were assigned by Gaussian distribution with zero mean. Y andβ represent the estimate of the dependent variables, Y, and regression parameters, β. X is the matrices of x j,i,t . σ 2 β represents posterior variance of the regression parameters, β. The coefficient τ is greater than 0 and determines the amount of shrinkage.
These Bayesian statistical estimations were implemented using WinBUGS 14 software [36]. The convergence of all the Bayesian estimations was assessed using the Gelman-Rubin statistical coefficient [37], where the closer the value is to 1.0, the better the convergence is. The Gelman-Rubin statistical values of all model parameters ranged from 0.96 to 1.05.

GeoDetector Model
The GeoDetector model [38,39] can detect the explanatory power of individual variable The primary idea of the GeoDetector model is that two variables (linearly or non-linearly) are accordant in strata if one causes or affects the other. The GeoDetector model may capture non-linear association effects, making it different from the ordinary linear regression model. A q-statistic value produced by the GeoDetector model can quantify the explanatory power of a single factor. The q-statistic value can be calculated as follows: where h (h = 1,2, . . . , l) represents the strata of a single variate X j , N j,h and N are the numbers of units in the stratum h of X j and in the total regions, respectively. In addition, σ 2 j,h and σ 2 represent the variances of the THEPC and the THEE's GDP share in the stratum h of X j and in 31 provincial regions of Chinese mainland in 2009-2018, respectively. The magnitude of the q-statistic value, q x j , quantifies the explanatory power of X j . The range of the q-statistic value is from 0% to 100%, and the larger the q-statistic value, the greater its explanatory power on the THEPC and the THEE's GDP share. The significance of the q-statistic value was tested by the noncentral F-test.

Descriptive Statistics
Health expenditures in China showed a trend of continuous growth in our analysis. THEPC and THEE as a share of GDP throughout China increased from 1314.26 Chinese yuan and 5.08% in 2009 to 4236.98 Chinese yuan and 6.43% in 2018. THE's annual growth rate was 14.89%. The annual growth of THEE as a share of GDP was 0.19%. Moreover, the dispersity of the various regions was increasingly obvious, especially for THEPC. Figure 3 illustrates the boxplots of THEPC and THEE as a share of GDP at the provincial level in China from 2009 to 2018. This reveals that not only did the significant increasing trends of the two terms of health expenditure occur throughout China, but the differences between various provincial regions also grew overall. The maximum ranges of THEPC and THEE as a share of GDP increased from 3304.78 Chinese yuan and 5.

Steady Spatial Patterns
The steady spatial pattern, quantified with the SRM to the nationwide overall level estimated by the BSTHM, was explored in this study. Figure 6A shows the SRM of THEPC in China. The SRM of THEPC did not form a particularly obvious geographical structure.  An obvious increasing trend of THEE as a share of GDP also occurred in each of the 31 provincial regions in China. A distinct spatial geographical structure of THEE as a share of GDP has emerged since 2016. Interestingly, the top three THEE as a share of GDP occurred in three provincial regions in western China in 2009, namely, Tibet (8.49%), Gansu (7.79%), and Xinjiang (7.74%). By 2019, three western provinces-Qinghai (11.21%), Tibet

Steady Spatial Patterns
The steady spatial pattern, quantified with the SRM to the nationwide overall level estimated by the BSTHM, was explored in this study. Figure 6A shows the SRM of THEPC in China. The SRM of THEPC did not form a particularly obvious geographical structure. The SRM with values significantly greater than 1.00 emerged in the eight provincial regions of Beijing (3. The SRMs of two regions-Shaanxi and Xinjiang-were equivalent to the overall nationwide level. Figure 6B exhibits the SRM of THEE as a share of GDP in China. The SRMs showed a distinct spatial structure, with a 'west high/east low' feature. The nine provincial regions located in western China-

Local Temporal Trends
This study used the BSTHM to estimate the local trends of THEPC ( Figure 7A) and THEE as a share of GDP ( Figure 7B

National Influence Pattern
The influence factors and their influential effects on THEPC and THEE as a share of GDP in Chinese mainland at the national level were estimated using the B-LASSO model. Nationally, nine factors-GDPPC, UR, PAR, ASY, PMAC, ANHW, NDVI, ASCPC, and AVCPC-significantly affected THEPC and THEE as a share of GDP ( Figure 8). Specifically, the top three dominant influencing factors affecting THEPC are GDPPC, ASY, and UR; the corresponding influence contributions were 27.5%, 24.2%, and 13.2%, respectively. However, the influence pattern of THEE as a share of GDP was different from that of THEPC. The two main influencing factors are ASY and UR, the corresponding influence attributions were 22.8% and 16.7%, and those of three factors, NDVI, AVCPC, and PMAC, were parallel: 11.6%, 11.2%, and 11.2%, respectively. AMCPC did not significantly influence THEPC and THEE as a share of GDP.

National Influence Pattern
The influence factors and their influential effects on THEPC and THEE as a share of GDP in Chinese mainland at the national level were estimated using the B-LASSO model. Nationally, nine factors-GDPPC, UR, PAR, ASY, PMAC, ANHW, NDVI, ASCPC, and AVCPC-significantly affected THEPC and THEE as a share of GDP ( Figure 8). Specifically, the top three dominant influencing factors affecting THEPC are GDPPC, ASY, and UR; the corresponding influence contributions were 27.5%, 24.2%, and 13.2%, respectively. However, the influence pattern of THEE as a share of GDP was different from that of THEPC. The two main influencing factors are ASY and UR, the corresponding influence attributions were 22.8% and 16.7%, and those of three factors, NDVI, AVCPC, and PMAC, were parallel: 11.6%, 11.2%, and 11.2%, respectively. AMCPC did not significantly influence THEPC and THEE as a share of GDP. Furthermore, the specific influence effects of the significant factors on THEPC and THEE as a share of GDP at the national level were estimated by the B-LASSO model and are listed in Table 1. The results show that the increase of six factors-GDPPC, PAR, ASY, ANHW, PMAC, and ASCPC-may lead to an increase in THEPC. The concrete increments of THEPC can be found in Table 1, when the six factors, GDPPC, PAR, ASY, ANHW, PMAC, and ASCPC, increased by 1000 Chinese yuan, one percent, one year, one day, one microgram per stere, and one kilogram, respectively, if other variables were invariant. Inversely, an increase in the other three factors, UR, NDVI, and AVCPC, may contribute to a decrease in THEPC. The corresponding decrements of THEPC are listed in Furthermore, the specific influence effects of the significant factors on THEPC and THEE as a share of GDP at the national level were estimated by the B-LASSO model and are listed in Table 1. The results show that the increase of six factors-GDPPC, PAR, ASY, ANHW, PMAC, and ASCPC-may lead to an increase in THEPC. The concrete increments of THEPC can be found in Table 1, when the six factors, GDPPC, PAR, ASY, ANHW, PMAC, and ASCPC, increased by 1000 Chinese yuan, one percent, one year, one day, one microgram per stere, and one kilogram, respectively, if other variables were invariant. Inversely, an increase in the other three factors, UR, NDVI, and AVCPC, may contribute to a decrease in THEPC. The corresponding decrements of THEPC are listed in Table 1. Except for GDPPC, the other nine factors maintained the same relationship between THEE as a share of GDP and THEPC ( Table 1). The estimations showed that it may lead to an increase of THEE as a share of GDP when the five factors PAR, ASY, ANHW, PMAC, and ASCPC increased. Conversely, increase of the four factors GDPPC, UR, NDVI, and AVCPC may affect a decrease of THEE as a share of GDP. The corresponding concrete increments and decrements of THEE as a share of GDP are listed in Table 1.

Model Validation
To validate the results estimated from the B-LASSO model, the GeoDetector model was employed to detect the national influence pattern. As previously mentioned, the GeoDetector model can estimate the explanatory powers of the 10 factors on the THEPC and the THEE's GDP share. Table 2 lists the specific results. It should be pointed out that the estimated values of the explanatory powers by the GeoDetector model is non-negative. Compared with the results of the national influence pattern estimated from the B-LASSO model (Figure 8), it can be seen that the explanatory powers of the selected 10 factors are almost equivalent to the absolute normalized regression coefficients estimated by the B-LASSO model. This indicates that the results of the B-LASSO model are robust. The ordinary least squares (OLS) regression model was also used to validate the estimated results by the B-LASSO model. The corresponding mathematic expression is as follows: where y it represents THEPC or THEE's GDP share; α is intercept; S i and T t represent spatial and temporal fixed effect; β j and X j represent the j-th OLS regression coefficient and influencing factor; ε it is error term. The estimated results of the OLS regression coefficients list in the Table 3. Generally, the influencing directions of the 10 variables is in accordance with the estimations of the B-LASSO model. Furthermore, it can be seen that the regression coefficients values of the OLS regression model are not significantly different from that of the B-LASSO model. The maximal relative differences of all the regression coefficients are less than 15%, some are even very close. This validation demonstrates again the robustness of the B-LASSO model's estimation.

Subnational Influence Pattern
To consider the spatial heterogeneity, the influence patterns of THEPC and THEE as a share of GDP at the subnational level of three economic areas-east, middle, and west ( Figure 1)-were also explored in this study. As previously mentioned, the three subnational areas represent three different levels of the economic development: developed, moderately developed, and underdeveloped, respectively. Table 4 lists the B-LASSO model's estimations of the influential effects of the independent variables affecting THEPC in the three subnational areas of Chinese mainland. The results showed that the influencing patterns of the three subnational areas are different. GDPPC significantly and positively affected THEPC in the three subnational regions, and the corresponding influence effects gradually increased from the east, middle, to the west. PAR was a significant influence factor on THEPC in the middle and west; however, this influence was not significant in the east. The ASY significantly affected THEPC only in the east. The environmental factor PMAC significantly and positively affected THEPC in the east and middle but not significantly in the west. The dietary factor, AVCPC, negatively affected THEPC in the east, but this factor was not significantly associated with THEPC in the middle and west. The natural factor, NDVI, was a significant influence factor only in the west. Other factors, UR, ANHW, ASCPC, and AMCPC, were not significantly influential factors in the three subnational regions. The results of the B-LASSO's regression coefficients of the 10 influence factors on THEE as a share of GDP (%) at subnational scale, east, middle, and west of the Chinese mainland are listed in Table 5. Similarly, the heterogeneity of the influence patterns of the 10 factors on THEE as a share of GDP existed in the three subnational regions. For socioeconomic factors, ASY was associated significantly with THEE as a share of GDP in the east, PAR was significant in the middle, and UR and PAR were significant in the west. In terms of natural environmental factors, ANHW and NDVI influenced significantly THEE as a share of GDP only in the west, whereas PMAC influenced significantly THEE as a share of GDP in the east and middle. The different dietary factors had different influence effects in the three subnational regions. An increase of one kilogram of ASCPC and AVCPC in the east may lead to THEE as a share of GDP decreasing by 1.159% (95% CI: 0.333%, 2.913%) and 0.040% (95% CI: 0.005%, 0.074%), respectively. AVCPC and AMCPC correlated negatively with THEE as a share of GDP in the middle and the west, respectively.

Discussion
The spatiotemporal trends and influencing patterns of THEPC and THEE as a share of GDP in China from 2009 to 2018 were explored in this study. The BSTHM and B-LASSO models were employed to explore these two terms. The two outcome variables represent health expenditure from absolute and relative angles.
Generally, China's THEPC is much less than that of the global average and developed countries, for example, the United States and Germany, according to data from the World Bank (https://data.worldbank.org/indicator/, accessed on 16 August 2022). China's THEE as a share of GDP is of an above-average level; however, it is still below that of Western nations, such as the United Kingdom and the United States. The sustainable increasing trend of THEE as a share of GDP in China is different from the global trend, which has shown a decreasing trend since 2016.
Although THEPC and THEE as a share of GDP at the provincial level throughout China both increased from 2009 to 2018, the spatial diversities of both of the rising tendencies are distinct. The steady geographical structures of the SRMs and local trends of THEE as a share of the GDP are more apparent than those of THEPC. We argue that the relative index-THEE as a share of the GDP-has less uncertainty compared with the absolute index-THEPC. We assume that THEE as a share of GDP in the regions in western China is higher because these areas are underdeveloped, and the corresponding GDP is not yet high. Consequently, the THEPC of the developed provincial regions was higher than that of the undeveloped areas. Although the THEPC in western provincial areas are not high, the local areas show increasing trends. Surprisingly, not only is THEE as a share of the GDP in the western provincial regions higher than the overall nationwide level but the corresponding local trends are also higher.
Nationally, the increase in PAR and ASY may contribute to an increase in health expenditures. The results showed that climate change (heatwave) and air pollution should increase health expenditure; however, an increase in vegetative cover may reduce health expenditures. This discovery can provide evidence for policy making to reduce health expenditure. The statistical results also provided information that a decrease in sugar intake and an increase in vegetable intake may lower health expenditures. From the perspective of regional heterogeneity, the influence patterns of health expenditure exhibited differences on the subnational scale. The residents' education level affects health expenditure in the east, but not in the middle or the west. We assume that the baseline education level of the east is higher than that of the middle and the west; therefore, the residents' education level's sensibility to health expenditure in the east is consequently higher. The influential effect of the heat wave is not detected at the subnational scale, except for THEE as a share of GDP in the west. Air pollution significantly impacted health expenditure in the east and middle but not in the west. One possible reason is that the pollution level of the west is lower than that of the east and the middle, and consequently the influencing effects of PM 2.5 in health expenditures is not significant in the west. Some researchers have concluded that exposure to air pollution is associated with an increase in health expenditure, especially PM 2.5 pollution [39]. Some scholars have also pointed out that the growth of CO 2 and Nitrous oxide emissions can increase health expenditure [40]. According to the results of our study and the previous research, air pollution and climate change are the major negative factors for an increase of health expenditure. It is interesting that the influence of the effect of NDVI on health expenditures only occurred in the west. A known fact is that the vegetation cover of the west is lower than that of the middle and the east. Hence, an increase in the NDVI can reduce health expenditures. Dietary factors have a more significant influence on THEE as a share of GDP, although the specific factor is not the same in the three areas. The dietary structure needs to adjust in the east and the middle, but nutrition needs to improve in the west.
In general, spatial and temporal heterogeneity should be considered during the process of making relevant policies about decreasing health expenditures in China and other countries around the world. The results of the influence patterns of THEPC and THEE as a share of GDP can also provide some revelations for predicting and controlling health expenditures in a precise manner. Environmental factors, such as PM 2.5 pollution and green coverage, have been growing increasingly important for the rise in health expenditure. The results of this study indicate that the decrease of PM 2.5 concentrations can reduce THEE in eastern and middle China; nevertheless, the increase in green cover can reduce the THEE in the western China. Simultaneously, a change in dietary structure is also a vital factor influencing health expenditures across China. Besides environmental factors, improving access to healthcare may also reduce THEE. Easier healthcare accessibility can enable patients, especially the elderly and children, to receive timely medical care [41][42][43][44] and then reduce medical expenses. Therefore, it is important for decreasing the THEE of China to maximize the effectiveness of the healthcare system. This suggestion works elsewhere, too. Although we hope that this study has deeply and comprehensively investigated the problem of health expenditure in China as far as possible, there are some limitations to this study. First, the provincial area was taken as the spatial statistical unit in our study, but this delineation was not sufficiently fine in terms of spatial granularity. If the spatial units were to be further subdivided, the problem could be studied in greater detail. Second, this study only discussed the factors associated with THEPC and THEE as a share of GDP; it did not investigate the causal mechanisms. Future studies should consider this aspect by employing causal inference methods. It should also be noted that it is interesting to see how COVID-19 affects health expenditure in mainland China. Unfortunately, the related data covering any period of COVID-19 in China are not available. However, this is a potential topic to be addressed in the future if the relevant data can be collected.

Conclusions
Based on our findings, this study offers some conclusions. First, the spatial distribution of THEPC in China did not form a particularly obvious geographical structure, whereas the THEE as a share of GDP in China displayed a distinct spatial structure, with a 'west high/east low' feature. Second, THEPC and THEE as a share of GDP in the 31 studied provincial regions of China all increased during 2009-2018, and the local trends of THEE as a share of GDP throughout China demonstrated a geographical structure with a 'north high/south low' feature. Third, the influence patterns of THEPC and THEE as a share of GDP are distinct. The dominant influencing factors of THEPC are the degree of development, represented by GDPPC and ASY; the dominant influencing factors of THEE as a share of GDP are ASY and UR. Fourth, the heterogeneity of the influence patterns of health expenditures exists from the east to the west in Chinese mainland. Fifth, natural environmental factors, such as air pollution and green coverage have been growing increasingly important for the rise of health expenditure; simultaneously, the change of dietary structure is also a vital factor influencing health expenditure.