Background

Universal health coverage hopes that everyone has access to high-quality health services without being pushed to impoverishment, owing to the cost of paying for those services. Catastrophic health expenditure (CHE) is an official measurement indicator of financial risk. However, the World Health Organization (WHO) estimated that more than 150 million individuals were facing CHE in 2005 [1]. This number rapidly increased to 808 million in 2015 [2]. Furthermore, half of the global population does not have access to the healthcare they need [3, 4]. There is no doubt that a high CHE has always been a considerable deterrent in reducing the financial risk of disease.

China, the largest develo** country in the world, has a severe financial risk of contracting the disease. The proportion of people aged 60 or over has increased by 5.44% in the last decade and reaching 13.5% in 2020 [5]. According to UN estimates, the number of Chinese elder population may further increase to 365,636,000 (26.1%) by 2050 [6], indicating an increasing disease burden of non-communicable diseases [7] (NCDs) and disabilities [8]. However, current social health insurance mechanisms have been insufficient to reduce the incidence of CHE (ICHE) among the middle-aged and elderly Chinese population [9, 10]. Further, it has become the short board for achieving UHC comprehensively, the determinants of which need to be further clarified.

There are several cohort studies on CHE determinants in middle-aged and elderly Chinese population [9, 11, 12]. The determinants of these studies include poverty, age, disability, NCDs, household size, inpatient or outpatient events, and social health insurance, considering time variation solely at the household level. All these conclusions drawn are based on the assumption that the characteristics of the national population are homogeneous. However, we need to pay special attention to China’s diverse geographical, socioeconomic, and ageing (AG) characteristics. The ICHE varies across regions nationwide as if it varies across countries worldwide [2, 13], with spatial non-stationarity. Unfortunately, despite several studies on the temporal non-stationarity of ICHE [9, 14, 15], the field of spatial non-stationarity is still lacking in all the existing studies related to ICHE. Additionally, whether the differences in factors’ influencing intensity exists between different provinces and whether the actual effect is diluted by the widely used global estimation methods remains to be seen. Therefore, we must further consider several issues. What factors dynamically affect the provincial ICHE among the middle-aged and elderly Chinese population? Is the relationship between these factors and ICHE the same throughout the country and in all provinces?

Our study makes the following contributions. First, we are the first to verify whether the occurrence of CHE is spatiotemporal non-stationary. Second, we constructed a more comprehensive framework for determinant analysis of ICHE by combining spatial and temporal perspectives. The impact of environmental dimension on ICHE was confirmed for the first time. Third, this study used geographical and temporal weighted regression (GTWR) as its primary method to further reduce the bias caused by ignoring the spatiotemporal position of the study sample. It considered strong relevance based on close spatiotemporal distance, while depicting the temporal and spatial evolution characteristics of the influencing factors of CHE for middle-aged and elderly population in China in eight years. Fourth, as the samples evolved from points to areas, the research units changed from households to regions, indicating that these local results are more instructive for regional policymaking without being diluted by global estimation.

To gain a better understanding of local spatiotemporal variation of ICHE and further formulate targeted policies to update health care systems in China, we conducted this retrospective cohort study from 2011 to 2018 by emphasizing the association among ICHE, out-of-pocket payment (OOP), socioeconomic factors, health service access, and health insurance coverage, under the consideration of spatiotemporal non-stationarity.

The remainder of this paper is organized as follows. First, according to the WHO estimation method, we calculated the CHE of every household and then converted the CHE and covariates into regional units. Second, we used global ordinary least squares (OLS) and GTWR with spatiotemporal location. We compared the results of the two models and tested for spatiotemporal non-stationarity. Third, we targeted the change characteristics of the temporal and spatial influencing factors affecting ICHE in households with middle-aged and elderly individuals. Finally, according to the above results, an evidence-based basis for improvement strategies is provided.

Methods

Sample selection

The China Health and Retirement Longitudinal Study (CHARLS) was conducted by the National Development Research Institute of Peking University. It is a longitudinal survey and is considered to be the most representative nationwide survey of the middle-aged and elderly population in mainland China. CHARLS conducted its national baseline survey from 2011 to 2012, Wave 2 was conducted in 2013, Wave 3 was conducted in 2015, and Wave 4 was conducted in 2018. The baseline survey included two person per household (aged 45 years or older), and in total involved 17,708 individuals living in 10,527 households across 28 provinces in China. However, we only removed households without subsistence food expenses, owing to their prime importance to CHE estimation and differences from common sense. We retained 16,161 individuals in 9224 households from the baseline survey, 15,320 individuals in 8662 households from Wave 2,17,763 individuals in 9977 households from Wave 3, and 17,408 individuals in 10,080 households from Wave 4. This cohort study included 28 provinces from 2011 to 2018 (Additional file 1: Appendix Fig. 1).

Assessment of CHE

We employed CHE as a proxy to describe the financial risk of utilizing health services. CHE was accessed based on the definition of WHO [16], that is, the situation when OOP of a household on health equals or exceeds 40% of its capacity to pay. Household size, OOP, living budget, income, and food expenditure (foodh) were extracted from the CHARLS questionnaire. The maximum value between the living budget and income (exph) was calculated to better assess the consumption ability. food45 and food55 were defined as the food expenditure shares of exph at the 45th and 55th, respectively.

We calculated the food expenditure share by dividing the household’s food expenses by exph, that is, foodexph. The household equivalence scale was used rather than the actual household size due to household consumption size, which is given below:

$${eqsize}_h={hhsize}_h^{\beta }$$
(1)

where hhsizeh is the household size from the CHARLS questionnaire, and parameter β equals 0.56, according to 59 countries’ household survey data [13].

Equivalized food expenditure (eqfoodh) was calculated by dividing household food expenditure by eqsizeh. The poverty line (pl) and subsistence expenditure (seh) were then generated based on the eqfoodh:

$${eqfood}_h=\frac{food_h}{eqsize_h}$$
(2)
$$pl=\frac{\sum {w}_h\ast {eqfood}_h}{\sum {w}_h}, where\ food45<{foodexp}_h< food55$$
(3)
$${se}_h= pl\ast {eqsize}_h$$
(4)

A household’s capacity to pay is the non-subsistence spending of a household. Depending on whether self-production, coupons, food subsidies, and other non-cash means of food consumption are considered, two estimation methods are given below:

$${\displaystyle \begin{array}{c}{ctp}_h={\mathit{\exp}}_h-{se}_h\mathrm{if}\ {se}_h<={food}_h\\ {}{ctp}_h={\mathit{\exp}}_h-{food}_h\ if\ {se}_h>={food}_h\end{array}}$$
(5)

Then, the burden of health OOP (oopctph) was calculated as the health OOP share of ctph.

$${ oop ctp}_h=\frac{oop_h}{ctp_h}$$
(6)

Finally, we considered a household incurring CHE if oopctph equals or exceeds 0.4, according to the WHO definition as follows:

$${\displaystyle \begin{array}{c}{cata}_h=1\ \mathrm{if}\ {oopctp}_h>=0.4\\ {}{cata}_h=0\ \mathrm{if}\ {oopctp}_h<0.4\end{array}}$$
(7)

Statistical analysis

From the perspective of previous multi-country analyses [2, 13], the main determinants of CHE can be represented by OOP, economic level, health service demand and use, and failure of social mechanisms to pool financial risks. We used regional gross domestic product (GDP) to reflect the economic development, particulate matter (PM2.5) to reflect the air pollution, and used the proportion of population aged 65 or over, incidence of NCDs, incidence of disability, and number of nurses per thousand persons to reflect the health service demand and provision. Health insurance coverage was adopted to reflect the success or failure of social mechanisms in pooling financial risks. In addition, regional development strategies undertaken by the central government may have impacts on ICHE via economic development, health service provision, and health insurance coverage and therefore need to be added as a control variable [17]. Some studies [18] reported the association between geographical location and regional development strategies [19, 20]. We therefore employed the geographical subdivision as the proxy variable to replace this unobservable variable, further dealing with endogeneity caused by omitted regional development strategies and unobserved heterogeneity caused by population migration. Finally, we introduced a seven-category variable based on geographical area as a control variable to distinguish between regional development strategies [17, 21]. All individual and household indicators were converted into regional indicators for further spatiotemporal analysis (Additional file 1: Appendix Fig. 1). Descriptions of all the indicators are shown in Table 1 and Additional file 1: Appendix Table 1. In future work, we will develop global OLS and GTWR models to measure the potential effects caused by spatiotemporal non-stationarity.

Table 1 Description of Independent Variables
Full size table

Global OLS model

First, the global OLS model was used for the analysis without any spatial and temporal considerations. The equation is as follows:

$${Y}_i={\beta}_0+\sum_k{\beta}_k{X}_{ik}+{\varepsilon}_ii=1,\dots, n$$
(8)

GTWR model

GTWR is a spatiotemporal statistical method and generally fits local regression models in space-time based on distance-decay effects, i.e., data of their own and surrounding units are considered. GTWR further considers timestamps and spatial locations based on the OLS regression model, allowing for specific parameter estimation for each spatiotemporal unit. It considers both temporal and spatial heteroskedasticity and outperforms OLS in the model accuracy for the sample data. Further, it has been used widely in house pricing and air pollution studies [22,23,24,25].

Based on previous ICHE studies, we found that ICHE is a disease financial risk indicator that varies by spatiotemporal location. With the increasing accuracy requirements of ICHE studies, its influencing factors should be analyzed with full consideration of spatiotemporal variation. Making full use of spatial panel data, GTWR helps provide direct evidence for regional policymaking. Especially in the case of health policy, regional guidelines need to take full account of their characteristics and those of neighboring regions. GTWR is expected to be one of the effective tools for regional policymaking.

Similar to OLS, GTWR takes timestamps into the origin coordinate framework and constructs a weight matrix to account for spatiotemporal non-stationarity in the parameters. The timestamp was addressed by setting it to a certain year. Therefore, the GTWR model and its parameter estimation can be expressed as:

$${Y}_i={\beta}_0\left({u}_i,{v}_i,{t}_i\right)+{\sum}_k{\beta}_k\left({u}_i,{v}_i,{t}_i\right){X}_{ik}+{\varepsilon}_i\ i=1,\dots, n$$
(9)
$$\hat{\beta}\left({u}_i,{v}_i,{t}_i\right)={\left[{X}^TW\left({u}_i,{v}_i,{t}_i\right)X\right]}^{-1}{X}^TW\left({u}_i,{v}_i,{t}_i\right)Y$$
(10)

where W(ui, vi, ti) denotes a new matrix whose diagonal elements are the spatiotemporal distance functions of (ui, vi, ti) and weights when calibrating a weighted regression adjacent to spatiotemporal unit i. An adaptive kernel and AICc [26] were used in our GTWR model.

Test for spatiotemporal non-stationarity

Many non-stationarity assessment techniques have been further employed to test its significance. However, Fortheringham [23, 27], as the frontier in applying the geographical weighted regression model, provided a technique to twice compare the standard errors of OLS estimates with the interquartile range of GTWR to assess the non-stationarity of coefficient estimated. Larger values of the latter are considered to indicate significant spatial non-stationarity.

Result

Range of ICHE variation from 2011 to 2018

We calculated the ICHE nationwide across 28 provinces in mainland China using Ke Xu’s method. Figure 1 shows the distribution of provincial ICHE and its temporal variations. According to the results, temporal variations in ICHE were found. The ICHE shows a gradual increase from 2011 to 2015 but suddenly decreased from 2015 to 2018. Individually, the ICHE of Bei**g has decreased from 9.804 to 3.226% from 2011 to 2018, with the most drastic decline, whereas Qinghai reached its highest value of ICHE (27.5%) in 2015. In addition, the southwestern area (including Qinghai, Sichuan, and Chongqing) and Hebei in China shows persistently high ICHE. The ICHE in 2015 shows great differences among provinces, whereas in other years, the difference is relatively small.

Fig. 1
figure 1

Spatiotemporal distribution for incidence of catastrophic health expenditure from 2011 to 2018 among middle-aged and elderly Chinese population

Full size image

Results of global estimation and the multi-collinearity test

OLS was first used for global estimation, and multi-collinearity tests of all independent variables were then performed. The results are presented in Table 2. The F-test result indicates this model is significant at 99% level, while the global estimation can solely explain 38.8% of the variation in ICHE based on R2.

Table 2 Global OLS regression result
Full size table

Furthermore, OOP and nurses are considered significant at 99% level, disability is deemed significant at 95% level, and PM2.5 and AG are considered marginally significant at 90% level. In addition, all the independent variables passes the multi-collinearity test and are considered for local estimation.

Results of the GTWR model and the spatiotemporal non-stationarity test

The GTWR model was tested further (Table 3). It should be noted that the residual sum of squares decreases from 1293.203 to 492.273, and R2 increases from 0.388 to 0.769 largely between OLS and GTWR, although the AICc increases slightly. This demonstrates that the local estimation of GTWR is better than the global estimation of OLS for this spatial cohort study, indicating that spatiotemporal non-stationarity helps to explain the variety of data. Introducing geographical subdivisions as control variable also makes the further improvement of the goodness of fit of the model (Additional file 1: Appendix Table 2).

Table 3 Comparison of Global OLS and GTWR models
Full size table

Furthermore, this study applied GTWR as the final estimation model. The coefficient variation distribution of the eight independent variables (the control variable is not listed) is shown in a parallel coordinate plot [28] (Fig. 2). It is worth noting that the number of nurses per thousand persons and health insurance coverage have almost always been protective factors against ICHE.

Fig. 2
figure 2

Variation trend of coefficients of all variables among spatiotemporal units

Full size image

Then spatiotemporal non-stationarity tests were adopted to test whether differences exist in eight independent variables among spatiotemporal units (Table 4). The coefficients of OOP, GDP, PM2.5, AG, NCDs, disability, and nurses exhibited extra local variation beyond purely sampling expectations. Therefore, health insurance coverage was considered a global variable without spatiotemporal non-stationarity.

Table 4 Spatiotemporal non-stationarity tests of independent variables
Full size table

Description of the variety spatiotemporal trends

As shown in Fig. 2, the distribution of the coefficients of the seven independent variables with spatiotemporal non-stationarity shows a significant difference between regions.

OOP

It is estimated that the positive impact of OOP decreases gradually from west to east spatially and temporally from 2011 to 2018. However, some spatiotemporal units differs. For instance, the impact of OOP on the ICHE in North China increases from 2011 to 2015 and decreases after 2015.

GDP and PM2.5

For GDP, which describes economic development, fascinating conclusions can be drawn. In relatively economically developed East China, GDP is a protective factor for ICHE, while it is a weakening risk factor in some regions, such as Central China and Northeast China. The impact of GDP on ICHE may demonstrate the complexity of the health effects of economic development.

Air pollution is a topic of interest. Spatially, its positive impact on the ICHE decreases from east to west. Moreover, its positive impact on ICHE gradually increases, but decreases in parts of East China and North China after 2015.

AG, NCDs and disability

For AG, most of its positive impact is concentrated in regions other than North China and Northeast China. Temporally, the positive impact of AG increases from 2011 to 2018 in East China, Central China, and South China but decreases after 2013 in Northwest China and Southwest China.

Most of the positive impacts of NCDs on ICHE are concentrated in regions similar to AG. Temporally, the positive impact of AG increases from 2011 to 2018 in East China but decreases after 2013 in Central China, South China, Northwest China, and Southwest China.

However, the spatial distribution of the positive impacts of disability differs from that of the first two and exists in regions other than East and Northeast China. However, it is noteworthy that it was significantly enhanced in Southwest China and Shaanxi Province. Temporally, most regions show a weakened positive impact of disability on ICHE after 2015, but this does not apply to Southwest China and Shaanxi Province. In these two regions, the positive impact of disability on ICHE maintains a possible upward trend.

Nurses

For nurses, we found a clear downward trend in all regions nationwide except Northeast China, especially until 2018. Furthermore, the widespread strong negative impact in Northwest China should not be ignored.

Discussion

In this study, the spatiotemporal relationship between ICHE and influencing factors was locally estimated for each province in China using the GTWR model. As far as we are concerned, this study is the first application of the GTWR model in health policy and made up the major blank area of spatial non-stationarity in the existing time-series studies. Based on the GTWR model, the results in this study confirmed the existence of spatiotemporal non-stationarity of OOP, economic factors, health service demand, provision and provided evidence for the formulation of related region-targeted policies.

Existence of the spatiotemporal heterogeneity of ICHE among the middle-aged and elderly Chinese population in recent years

The ICHE among the middle-aged and elderly Chinese population is spatiotemporally heterogeneous. Temporally, the ICHE in this study increases from 2011 to 2015 and decreased from 2015 to 2018, showing similar trends in the previous literature [29]. In fact, with the frequent occurrence of NCDs and physical multimorbidity in the elderly in recent years [9], the economic burden of diseases among middle-aged and elderly residents in China has gradually increased. Under the background of such high demands and utilization of diseases, China’s ICHE could still show a downward trend in time series, with the implementation of a series of health reform measures in China, such as the establishment and deepening of multi-level health insurance systems and remarkable achievements in the essential drug system [30, 31]. The timeline of China’s series of health reform measures almost matches that of ICHE reduction, which proves that health policies make outstanding contributions to the control of ICHE in China [32]. Spatially, the distribution of ICHE in this study is also markedly heterogeneous, which echoes much of the literature [33]. Southwest China and Hebei Province exhibits persistently high ICHE. NCDs is a major cause of CHE in households with the middle-aged and elderly population and its causes in China vary greatly between regions [34]. The burden of NCDs is high in several provinces in the Southwest China, where poor lifestyle habits due to ethnic customs, urbanization, and westernization [35, 36] have raised the risk of develo** NCDs. As for Hebei Province, it is facing serious air pollution problems which may further cause NCDs. In addition, the relatively low economic level [

Conclusion

Based on the GTWR method, this study breaks the hypotheses of fixed coefficients and global stationarity. It explores the spatiotemporal non-stationarity of the relationship between various factors and ICHE, which can be used as an essential basis for regional policy reform.

First, this study called for extensive consideration of spatiotemporal non-stationarity in health. Based on the results of the model comparison in this study, we further expanded the interpretation of the global model by considering the coefficient of variation with spatiotemporal coordinates. Significant progress has been made after considering spatiotemporal non-stationarity, indicating that extensive related studies should consider it to further improve their quality.

Second, the existence of spatiotemporal non-stationarity in ICHE has far-reaching implications for subsequent related policy formulation.

Third, after the completion of universal health insurance coverage, the spatiotemporal non-stationarity of the prevalence of NCDs and disability and AG should be the focus in the future. Based on this, improvements to compensation coverage and benefit packages are possible policy instruments. For example, consider outpatient services for patients with NCDs, rehabilitation services for people with disabilities, and the financial risk of disease for the elderly in the subsequent design of health insurance. Considering the spatiotemporal non-stationarity of these factors together can help improve the propensity of health insurance policies.

Forth, economic development shows complex impacts on ICHE. Especially, air pollution, as one of its consequences, leads to higher ICHE. In contrast, the allocation of nursing resources helps to reduce the ICHE. This further suggests that the governance and causes of CHE should not be only at the individual/household level and need to be laid out from a macro perspective.