Evolving trends, regional differences, determinants, and disease sources of provincial-level health inequalities in china 1990–2019: a temporal convergence and novel triple decomposition analysis

Abstract

Background

Promoting health equity has been a worldwide goal, but serious challenges remain globally and within China. Multiple decomposition of the sources and determinants of health inequalities has significant implications for narrowing health inequalities and improve health equity.

Methods

Life expectancy (LE), healthy life expectancy (HALE), age-standardized mortality rate (ASMR), and age-standardized disability-adjusted life-year (DALY) rates in 31 provinces of mainland China were selected as health status indicators, obtained from the Global Burden of Disease (GBD) database. Temporal convergence analysis was used to test the evolving trends of health status. Dagum’s Gini coefficient decomposition was used to decompose the overall Gini coefficient based on intraregional and interregional differences. Oaxaca-Blinder decomposition was used to calculate contributions of determinants to interregional differences. The factor-decomposed Gini coefficient was used to analyze the absolute and marginal contribution of each component to overall Gini coefficients.

Results

From 1990-2019, China witnessed notable improvements in health status measured by LE, HALE, ASMR and age-standardized DALY rates.Nevertheless, the three regions (East, Central and West) exhibited significant inter-regional differences in health status, with the differences between the East and West being the largest. The adjusted short-term conditional β-convergence model indicated that the inter-provincial differences in LE, HALE, ASMR, and age-standardized DALY rates significantly converged at annual rates of 0.31%, 0.35%, 0.19%, and 0.28% over 30 years. The overall Gini coefficients of LE, HALE, and age-standardized DALY rates decreased, while the ASMR exhibited an opposite trend. Inter-regional and intra-regional differences accounted for >70% and <30% of overall Gini coefficients, respectively. Attribution analysis showed that socioeconomic determinants explained 85.77% to 91.93% of the eastern-western differences between 2010-2019, followed by health system determinants explaining 7.79% to 11.61%. The source-analysis of Gini coefficients of ASMR and age-standardized DALY rates revealed that noncommunicable diseases (NCDs) made the largest and increasing absolute contribution, while communicable, maternal, neonatal, and nutritional diseases (CMNNDs) had a diminishing and lower impact. However, NCDs exerted a negative marginal effect on the Gini coefficient, whereas CMNNDs exhibited a positive marginal effect, indicating that controlling CMNNDs may be more effective in reducing health inequities.

Conclusions

Regional differences are a major source of health inequities in China. Prioritizing prevention and control of CMNNDs, rather than NCDs, may yield more pronounced impacts on reducing health inequalities from the perspective of marginal effect, although NCDs remain the largest absolute contributor to health inequalities.

Background

Health equity is one of the core dimensions used to evaluate health system performance. Promoting health equity has been a worldwide goal, but serious challenges remain globally and within China [1,2,3]. In the past three decades, China has achieved remarkable success in the context of economic development, and the health of people has greatly improved. Although health gains have continued, significant differences in health status exist among the 31 provinces in mainland China, and government and public concerns about China’s health inequalities are increasing [4, 5]. It is impractical to eliminate health inequalities entirely, but narrowing health inequalities is an integral step in achieving health equity [6, 7]. Due to regional differences in economic and social development, particularly in the realm of healthcare, China has long faced the challenge of narrowing the gap in essential health services and overall health outcomes across different regions and demographic groups [8, 9]. This objective was identified as a top priority in China’s 2009 healthcare system reform plan [10] and is explicitly outlined as a key goal in the Health China 2030 Plan [11].

Previous studies have described persistent differences in health status among or within provinces [12, 13], urban and rural regions [14, 15], income groups [16], sex groups [17], and unemployed and employed individuals [18] in China. However, most of these studies only identified the trends in health inequalities by simple descriptive statistics and did not assess the trends in depth, with no statistically tested temporal convergence analysis. In terms of indicators, many studies have used the self-assessment health (SAH) or self-reported health (SRH) status to reflect health inequalities at the individual level; however, these metrics are not strictly comparable or meaningful because they are subjective ordered responses based on rankings rather than specific values [12, 16]. Other studies used the maternal mortality rate [14, 19], neonatal mortality rate [15], perinatal mortality rate [20], infant mortality rate [9], etc., to measure health inequality at the population level, but these indicators only encompass information from a small part of the population. Moreover, very few studies have noticed that for different types of diseases or factors, such as NCDs, CMNNDs and injuries, their contribution rates to health inequality are usually different, and their marginal contribution rates may even be opposite.

Based on the aforementioned background, this study selected a set of provincial indicators that are more objective (than SAH and SRH) and encompass both the entire population’s length and quality of life, namely Life expectancy (LE), healthy life expectancy (HALE), age-standardized mortality rate (ASMR), and age-standardized disability-adjusted life-year (DALY) rates, to measure health inequality. Statistical tested temporal convergence analysis was employed instead of relying solely on intuitive descriptions to analyze the evolution of health inequalities. A novel triple decomposition analysis framework approach was proposed to conduct a comprehensive investigation on health inequality in China from three dimensions: regional differences, determinants, and disease sources. This approach distinguishes itself from previous studies and the findings may shed light on the underlying characteristics of health inequality in China. Such insights are valuable for exploring potential strategies and feasible approaches to mitigate development imbalances within China’s healthcare sector while also providing reference or guidance for other developing countries aiming to narrow down health differences and achieve equitable healthcare.

Methods

Data sources

LE, HALE, ASMR, and age-standardized DALY rates of 31 provinces of mainland China were indicators of health status in this study. The data were obtained from the Global Burden of Disease (GBD) database jointly established by the Chinese Center for Disease Control and Prevention (China CDC) and the Institute for Health Metrics and Evaluation (IHME) at the University of Washington. Since 1990, the database regularly provides updated census data, demographic survey data, disease surveillance system data, national mortality surveillance system data, health survey reports and data from large-scale research studies in China. In 2013, the disease surveillance system and the cause-registration and reporting information system were merged into a comprehensive national cause-of-death surveillance system. The number of monitoring points based on districts and counties has been increased from 161 to 605, covering 31 provinces and regions in mainland China. The GBD world population age standard [21] was used for the calculation of ASMR and age-standardized DALY rates (details are available in Appendix S1b). The years covered in our study are from 1990 to 2019.

In addition, this study incorporated variables pertaining to socioeconomic factors, health system factors, and health risk factors. Socioeconomic factors encompassed per capita gross domestic product (GDP), average years of education (for individuals aged six and above), and the urban registered unemployment rate. Health system factors included the density of health technicians per 1,000 individuals and the proportion of out-of-pocket health expenditure in total health expenditure. Health risk factors comprised population-weighted fine particulate concentration (PM2.5), incidence rate of traffic accident per 10, 000 cars, and incidence rate of Class A and B infectious diseases. These data were mainly obtained from national and international databases and statistical yearbooks (see the Appendix S1a).

Statistical methods

Means were used to describe continuous variables. The independent sample T-test was used to compare health status indicators between any two regions. Bootstrap method was used to calculate 95% confidence intervals (95% CI). Time convergence analysis was employed to examine the temporal trends in health status evolution. Dagum’s Gini coefficient decomposition method was utilized to decompose the overall Gini coefficients of four health status indicators based on intra-regional and inter-regional differences. The Oaxaca-Blinder decomposition method was applied to calculate the contribution of determinants (including socioeconomic factors, health system factors, etc.) to the inter-regional differences. The factor-decomposed Gini coefficient was used to analyze the contribution and marginal contribution of each source of ASMR and age-standardized DALY rates towards the overall Gini coefficient. All statistical analyses were performed using Stata 17 and plots were performed using R (Version 4.1.2). All hypothesis tests were two sided. The primary methodologies are elaborated upon below.

Temporal convergence analysis

Temporal convergence can be analyzed by σ-convergence and β-convergence. σ-convergence focuses on the convergence characteristics of the stock, indicating a gradual decrease in the degree of dispersion among index values over time. β-convergence focuses on the convergence characteristics of increments, suggesting that observation units with lower initial values tend to experience faster growth rates and eventually converge towards a consistent steady state over time. β-convergence is a necessary but not sufficient condition for σ-convergence [22]. β-convergence is divided into absolute β-convergence, which is convergence without considering other factors, and conditional β-convergence, which is the convergence after controlling for other influencing factors.

The σ convergence is usually measured using the coefficient of variation, which is calculated as

$$\:\sigma\:=\sqrt{\sum\:_{i=1}^{\text{N}}\:\left(Y_{i}-\overline Y\right)^{2}/N}/\overline Y$$

(1)

In Eq. (1), N is the total number of observation units, Y_i is the observation value of any observation unit i, and $\overline{Y}$ is the mean value of all observations. β convergence is divided into absolute β convergence, which is the convergence without considering other factors, and conditional β convergence, which is the convergence after controlling for other influencing factors. The econometric model for absolute β convergence is

$$\:\text{ln}\left(Y_{i,t+k}/Y_{i,t}\right)=\alpha\:+\beta\:\text{ln}\;Y_{i,t}+{\mu\:}_i+{\eta\:}_t+{\epsilon\:}_{i,t}$$

(2)

In Eq. (2), Y_{i, t+k} is the observation of observation unit i in period t + k, Y_{i, t} is the observation of observation unit i in period t, ln(Y_{i, t+k}/Y_{i, t}) is the growth rate of the observation of observation unit i in period t-t + k, α is a constant term, β is a parameter to be estimated, µ_i and η_t are individual fixed effects and time fixed effects, and ε_{i, t} is a random perturbation term. At a given significance level, if β < 0 and passes the significance test, it means that there is a negative correlation between the initial observations and the growth rate, and there is absolute convergence at a rate of ν=-ln(1 + β)/k. When k = 1, the model tests for short-term absolute β convergence; when k > 1, the model tests for long-term absolute β convergence.

The econometric model for conditional β convergence is

$$In(Y_{i,t+k}/Y_{i,t})=\alpha+\beta ln\,Y_{i,t}+\lambda Control_{i,t}+\mu_i+\eta_t+\varepsilon_{i,t}$$

(3)

In Eq. (3), Control_{i, t} is a set of control variables affecting ln(Y_{i, t+k}/Y_{i, t}), λ is the parameter to be estimated for the control variables, and the other variables have the same meaning as in Eq. (2).

In this paper, conditional β convergence was used to test whether the interprovincial gap of the four health indicators had a narrowing trend after controlling socioeconomic factors, health system factors, and health risk factors. In other words, this approach can demonstrate whether the short-term and long-term narrowing trend of health inequality among provinces is statistically significant.

Dagum’s gini coefficient decomposition

Dagum’s Gini coefficient decomposition [23] considered the distributions within subgroups and decomposed overall differences into intraregional differences, interregional differences, and the intensity of transvariations. This method not only overcomes the limitation that the traditional Gini coefficient cannot be measured for groups but also solves the overlap problem encountered in Theil index decomposition. It has been widely used to measure regional development gaps [24]. Assuming a total of n provinces divided into k regions, the calculation formula is

$$\:G=\varDelta\:/2\stackrel{-}{Y}=\sum\:_{j=1}^{k}\:\sum\:_{h=1}^{k}\:\sum\:_{i=1}^{{n}_{\text{j}}}\:\sum\:_{r=1}^{{n}_{h}}\:\left|{y}_{ji}-{y}_{hr}\right|/2{n}^{2}\stackrel{-}{Y}$$

(4)

Equation (4) is used to calculate the overall Gini coefficient G for a health indicator, where Δ is the mean difference between the overall Gini coefficients. $\:\stackrel{-}{\text{Y}}\:$ is the overall average of the health indicators for the n provinces, y_ji and y_hr represent the health indicators for province i and r included in region j and region h, respectively. n_j and n_h are the number of provinces included in region j and region h.

$$\:{G}_{jj}={\varDelta\:}_{jj}/2{\stackrel{-}{Y}}_{j}=\sum\:_{i=1}^{{n}_{j}}\:\sum\:_{r=1}^{{n}_{j}}\:\left|{y}_{ji}-{y}_{jr}\right|/{2n}_{j}^{2}{\stackrel{-}{Y}}_{j}$$

(5)

Equation (5) is used to calculate the Gini coefficient G_jj within any region j.

$$\:{G}_{jh}={\varDelta\:}_{jh}/({\stackrel{-}{Y}}_{j}+{\stackrel{-}{Y}}_{h})=\sum\:_{i=1}^{{n}_{j}}\:\sum\:_{r=1}^{{n}_{h}}\:\left|{y}_{ji}-{y}_{hr}\right|/{n}_{j}{n}_{h}({\stackrel{-}{Y}}_{j}+{\stackrel{-}{Y}}_{h})$$

(6)

Equation (6) is used to calculate the Gini coefficient G_jh between any region j and region h.

To further decompose the overall Gini coefficient G, it is necessary to first define Eq. (7a, b, c, d, e, f)

(7)

Equation (7a) is the ranking by the mean value of the health indicators in each region. In Eq. (7b), p_j is the proportion of n_j to the overall number of provinces n in the country, $\:\sum\:{p}_{j}=1$. In Eq. (7c), s_j is the proportion of the sum of health indicators for each province included in region j to the sum of health indicators for the overall provinces in the country, $\:\sum\:{s}_{j}=1$. It can be proved that $\:\:\sum\:_{j=1}^{k}\sum\:_{h=1}^{k}{p}_{j}{s}_{h}=1$.

In Eq. (7d), d_jh is the weighted average of the differences in health indicators between region j and all provinces with y_ji> y_hr in region h for $\:{\stackrel{-}{Y}}_{j}\ge\:{\stackrel{-}{Y}}_{h}$, F_j and F_h denote the cumulative distribution functions of region j and region h, respectively. In Eq. (7e), p_jh is the weighted average of the differences in health indicators between region j and all provinces in region h with y_ji < y_hr in the case where $\:{\stackrel{-}{Y}}_{j}\ge\:{\stackrel{-}{Y}}_{h}$. In Eq. (7f), D_jh represents the relative impact of health indicators between region j and region h. When $\:{\stackrel{-}{\text{Y}}}_{\text{j}}={\stackrel{-}{\text{Y}}}_{\text{h}}$, D_jh =0; when the probability density functions of y_j and y_h do not overlap at all, D_jh =1. 1-D_jh represents the intensity of transvariation of health indicators between region j and region h.

Based on the above formula, the overall Gini coefficient G can be decomposed into the sum of the contributions of the three components:

$$\:G={G}_{w}+{G}_{nb}+{G}_{t}$$

(8)

In Eq. (8), G_w represents the contribution of intraregional differences, G_nb represents the contribution of interregional differences and G_t represents the contribution of the intensity of transvariation.

$$\:{G}_{w}=\sum\:_{j=1}^{k}\:{G}_{jj}{p}_{j}{s}_{j}$$

(9)

$$\:{G}_{nb}=\sum\:_{j=2}^{k}\:\sum\:_{h=1}^{j-1}\:{G}_{jh}\left({p}_{j}{s}_{h}+{p}_{h}{s}_{j}\right){D}_{jh}$$

(10)

$$\:{G}_{t}=\sum\:_{j=2}^{k}\:\sum\:_{h=1}^{j-1}\:{G}_{jh}\left({p}_{j}{s}_{h}+{p}_{h}{s}_{j}\right)\left(1-{D}_{jh}\right)$$

(11)

In this paper, it is applied to analyse the differences in four health indicators among and within the eastern, central, and western regions of China, considering the intensity of transvariation. Specifically, when calculating $\:{F}_{h}\left(x\right)$ and $\:{F}_{j}\left(x\right)$ in Eq. (7d) and Eq. (7e), we used the empirical cumulative distribution functions (ECDF), which can be expressed as $F\widehat(x)=\#\{X\leq x\}/N$, where $F\widehat(x)$ is the estimated probability less than or equal to $\:x$,$\#\{X\leq x\}$  is the number of observations that less than or equal to $\:x$, and N is total number of observations.

The oaxaca-blinder decomposition

The Oaxaca-Blinder decomposition employs regression analysis to quantify the extent to which differences in observed characteristics between different groups contribute to a disparity in a given characteristic, and identifies significant factors that are associated with the observed disparity [25].

This study’s regression equation is based on the Grossman health production function. The independent variables include a series of health determinants, and the model is presented in linear form. All the independent variables have been logarithmically incorporated into the equation. The regression equations for the eastern and western regions are as follows:

$$\:{Y}_{it}^{eastern}={\beta\:}_{0i}^{eastern}+\sum\:_{j=1}^{J}\:{\beta\:}_{jit}^{eastern}{X}_{jit}^{eastern}+{\mu\:}_{it}^{eastern}$$

(12)

$$\:{Y}_{it}^{western}={\beta\:}_{0i}^{western}+\sum\:_{j=1}^{J}\:{\beta\:}_{jit}^{western}{X}_{jit}^{western}+{\mu\:}_{it}^{western}$$

(13)

Y represents the four health indicators of the dependent variable, X represents the independent variable in the model, i represents the province i, t represents the time, $\:{\beta\:}_{0}$ represents the constant term, $\:{\beta\:}_{j}$ represents the regression coefficient of the j independent variable, and µ represents the random error term.

Based on the equations estimating health indicators for the eastern and western provinces separately, this study allowed for the estimated coefficients of the independent variables to be derived. Since under OLS estimation, there exists $\:E\left({\mu\:}_{mean}^{eastern}\right)=E\left({\mu\:}_{mean}^{western}\right)=0$, the difference between the means of the two groups of health indicators can be expressed as follows:

$$\:{Y}_{mean}^{eastern}-{Y}_{mean}^{western}=\left({\beta\:}_{0}^{eastern}-{\beta\:}_{0}^{western}\right)+{\sum\:}_{j=1}^{J}\left({\beta\:}_{j}^{eastern}{X}_{j\:mean}^{eastern}-{\beta\:}_{j}^{western}{X}_{j\:mean}^{western}\right)=[{\sum\:}_{j=1}^{J}\left({X}_{j\:mean}^{eastern}{-X}_{j\:mean}^{western}\right){\beta\:}_{j}^{eastern}]+[\left({\beta\:}_{0}^{eastern}-{\beta\:}_{0}^{western}\right)+{\sum\:}_{j=1}^{J}({\beta\:}_{j}^{eastern}-{\beta\:}_{j}^{western}){X}_{j\:mean}^{western}]$$

(14)

In this study, the Oaxaca-Blinder decomposition technique was applied to decompose the differences in the overall health indicators into two parts, one for the differences in the means of the independent variables (the explainable part) and the other for the differences in the intercept terms and regression coefficients (the unexplainable part). The explainable part is $\:[{\sum\:}_{j=1}^{J}\left({X}_{j\:mean}^{eastern}-{X}_{j\:mean}^{western}\right){\beta\:}_{j}^{eastern}]$, the unexplainable part is $\:[\left({\beta\:}_{0}^{eastern}-{\beta\:}_{0}^{western}\right)+{\sum\:}_{j=1}^{J}({\beta\:}_{j}^{eastern}-{\beta\:}_{j}^{western}){X}_{j\:mean}^{western}]$.

The factor-decomposed gini coefficient

The factor-decomposed Gini coefficient was originally used in the field of income distribution to calculate the contributions of different sources of income to total income inequality [26].

The overall Gini coefficient G was decomposed into three components by factor (source).

$$\:G=\sum\:_{k=1}^{K}\:{S}_{k}{G}_{k}{R}_{k}$$

(15)

Assuming that the indicator Y used to calculate the overall Gini coefficient has K factors (sources), which is $Y=\:\sum\:_{k=1}^{K}{Y}_{k}$, then in Eq. (15), S_k denotes the weight of an element Y_k in the sum of all factors Y, G_k denotes the Gini coefficient of element Y_k, and R_k denotes the Gini correlation coefficient between element Y_k and the distribution of the sum of all factors Y, defined as $\:{\text{R}}_{\text{k}}=\text{C}\text{o}\text{v}\left[{\text{Y}}_{\text{k}},\:\text{F}\left(\text{Y}\right)\right]/\text{C}\text{o}\text{v}\left[{\text{Y}}_{\text{k}},\:\text{F}\left({\text{Y}}_{\text{k}}\right)\right]$. Cov represents the covariance and F is the cumulative distribution function. The value interval of R_k is [-1,1], R_k=1 if Y is exactly the same as the ranking of Y_k, or R_k=-1 if it is exactly the opposite. The marginal contribution M_k of any factor Y_k to the overall Gini coefficient G can be further calculated according to this as follows

$$\:{M}_{k}=\frac{{S}_{k}{G}_{k}{R}_{k}}{G}-{S}_{k}$$

(16)

Equation (16) implies that there are two scenarios for alleviating inequality (reducing the overall Gini coefficient) by increasing the factor Y_k: the first is when R_k ≤ 0; the second is when R_k > 0 and G_kR_k ≥ G.

In this study, it is token to analyse the proportional and marginal contributions of the components of the ASMR and age-standardized DALY rates to the corresponding inequality levels. The two indicators were both categorized into three groups according to the attributable factors, namely chronic noncommunicable diseases(NCDs); communicable, maternal, neonatal, and nutritional diseases (CMNNDs); and injuries.

Results

Basic characteristics of provincial health status and health inequalities in China

As shown in Table 1, from 1990 to 2019, the mean LE of all the 31 provinces in mainland China increased by 9.68 years (14.37%), HALE increased by 8.23 years (13.75%), the ASMR decreased by 5.03 kilopoints (42.41%) and the age-standardized DALY rates decreased by 19,676.68 person-years per 100,000 people (45.68%). The data from various regions indicated that the eastern region displayed the highest level of health, while the western region exhibited the lowest level. In 2019, there were statistically significant differences in almost all four health indicators between any two regions, although these gaps have narrowed compared to those observed in 1990. In Fig. 1, LE and HALE demonstrated an overall upward trend in the past 30 years for across both the entire country and different regions, while ASMR and age-standardized DALY rate exhibited downward trend. Detailed data are available in Appendix S2 and S3.

Table 1 Characteristics of four health indicators across China and different regions in 1990 and 2019

Fig. 1

Trends in four provincial health indicators across China and its eastern, central, and western regions (1990–2019)

Temporal convergence trends in provincial health inequalities in China

To obtain more precise insights into the evolutionary trends, time convergence analysis was employed to examine the σ-convergence and β-convergence of the four health indicators. As is illustrated in Fig. 2, the variation coefficient of three indicators LE, HALE, and the age-standardized DALY rates exhibited a general decreasing trend from 1990 to 2019, whereas that of ASMR demonstrated a general increasing trend. This implies that there was a σ-convergence in LE, HALE, and age-standardized DALY rates from 1990 to 2019; however, ASMR did not exhibit σ-convergence. Detailed data are provided in Appendix S4.

Fig. 2

σ-convergence for four provincial health indicators in China (1990–2019)

As depicted in Table 2, after controlling for other factors, results of the short-term conditional β-convergence model showed that the adjusted regression coefficients of LE, HALE, ASMR, and the age-standardized DALY rates were − 0.3137 (p = 0.018), -0.3450 (p = 0.026), -0.1936 (p < 0.01) and − 0.2832 (p = 0.018), respectively. This means that during the period from 2010 to 2019, the provincial differences of these indicators converged significantly at an annual rate of 0.31%, 0.35%, 0.19% and 0.28%, respectively, and the gaps narrowed year by year. The conditional β-convergence model with the period k = 5 was used to test the long-term conditional β-convergence, and the results were shown in Appendix S5. The results are similar to the observed short-term β-convergence feature, suggesting a “catch-up” feature in health levels between provinces.

Table 2 Short-term conditional β-convergence test results for four health indicators in China (2010–2019)^a

The first decomposition: regional differences in provincial health inequalities and the associated contributions in China

Fig. 3

Regional differences in Gini coefficients for LE and the ASMR and the contribution sources in China (1990–2019)

As shown in Fig. 3A, the Gini coefficient of LE across all regions exhibited a declining trend from 1990 to 2019. Specifically, the decline was particularly pronounced during the period from 1990 to 2007, followed by fluctuations between 2008 and 2010 before ultimately stabilizing. When considering intraregional differences, the central region exhibited the lowest Gini coefficient and has demonstrated a consistent downward trend from 1990 to 2019. The western region previously held the highest Gini coefficient until 2007, also experiencing an overall decline over the subsequent three decades. Conversely, the eastern region displayed an upward trend in its Gini coefficient. Prior to 2007, it fell between the western and central regions but gradually surpassed the western region thereafter to claim the highest position.

In terms of interregional differences (Fig. 3C), the largest gap in LE Gini coefficient existed between the eastern and western regions. However, this gap has decreased significantly since 1990, particularly after a rapid decline from 1990 to 2007. On the other hand, the central-western difference and the central-eastern difference were relatively small and have shown an overall downward trend. Based on Fig. 3E, interregional differences were the primary contributor to the overall LE Gini coefficient, although the relative impact decreased from 77.53% in 1990 to 74.38% in 2019. Intra-regional variation was the next significant factor, increasing from 20.15 to 22.45%, while the contribution of intensity of transvariation remained less than 4%. Detailed data are available in Appendix S6.

As shown in Fig. 3B, the trend of ASMR Gini coefficient among Chinese provinces exhibited significant divergence from that of LE. Prior to 2000, a gradual decline was observed, followed by a rapid increase between 2000 and 2015, eventually stabilizing thereafter. Over the three decades, China’s overall ASMR Gini coefficient has risen from 0.0949 to 0.1230. In 1990, the western region displayed the highest ASMR Gini coefficient while the central region had the lowest; meanwhile, the eastern region fell somewhere in between. However, during this thirty-year period, there was a remarkable surge in the central region’s ASMR Gini coefficient which surpassed both the eastern and western regions in 1994 and 2003 respectively.

In terms of interregional differences (Fig. 3D), the eastern-western difference in the Gini coefficient of the ASMR was the highest, and the central-western difference was the lowest. The differences in both the eastern-western and eastern-central regions steadily increased beginning in 2003 and started to decline in 2015. Similar to LE, the largest contribution to the overall Gini coefficient of the ASMR came from interregional differences, accounting for 71–77% (Fig. 3F). Detailed data are available in Appendix S7.

Results of regional differences for HALE and age-standardized DALY rates are shown in Appendix S8-S10. It indicated a reduction in provincial-level HALE inequality between 1990 and 2019, with the overall Gini coefficient declined from 0.0322 to 0.0178. Furthermore, there was also a decrease in provincial-level inequality for the age-standardized DALY rates, as evidenced by a decline in the Gini coefficient from 0.1277 to 0.0944. The trends and contributions of regional differences to the overall Gini coefficient in HALE and the age-standardized DALY rates mimicked those observed for LE.

The second decomposition: determinants of interregional health differences and their contributions in china

As is proved above, the interregional differences, specially the eastern-western difference, contributed the most to the provincial health inequalities. Socioeconomic factors, health system factors, and health risk factors were explored in this study as potential determinants of inter-regional variation. Further attribution analysis based on the Grossman health production function [27] using the Oaxaca-Blinder decomposition showed that from 2010 to 2019¹, the explainable parts explained 59.91–66.81% of the differences in health levels between eastern and western regions. And within the explainable parts, socioeconomic determinants² explained 85.77–91.93%, following by health system determinants³ which explained 7.79–11.61%. Health risk determinants⁴ had little effect (see Table 3 and Appendix S11).

Table 3 Determinants and their contributions to differences in health levels between eastern and western regions in China (2010–2019)

The third decomposition: sources of provincial health inequalities and their absolute and marginal contributions in China

To investigate the sources of health inequality, we analyzed the absolute and marginal contribution pertaining to the three primary sources (NCDs, CMNNDs and Injuries) of ASMR and age-standardized DALY rates.

As illustrated in Table 4, NCDs contributed the largest proportion among the three primary sources and this contribution experienced a significant upward trajectory (from 54.27 to 84.53%), particularly after 2000. CMNNDs ranked second in terms of their contribution, with a declining trend (from 37.22 to 10.17%). Contributions from injuries have remained consistently low (around 5-10%). The ASMR associated with NCDs consistently displayed a negative marginal contribution, with a 1 unit increase associated with a corresponding decrease in the overall Gini coefficient of 0.04–0.25%. In contrast, ASMR associated with CMNNDs consistently showed a marginal positive contribution, with a 1 unit increase associated with a corresponding increase in the overall Gini coefficient of 0.05–0.24%. The marginal contribution of ASMR associated with injuries was generally minimal, with absolute values generally less than 0.02%.

The results for age-standardized DALY rates were generally consistent with the ASMR findings. In terms of absolute contribution rate, NCDs accounted for the largest proportion, exhibiting a significant upward trend (from 32.90 to 69.21%), while the contribution of CMNNDs showed a subsequent downward trend (from 54.46 to 20.78%). The age-standardized DALY rates associated with NCDs consistently demonstrated a negative marginal contribution, whereas the age-standardized DALY rates linked to CMNNDs consistently exhibited a positive marginal contribution.

Table 4 Gini coefficient of China’s ASMR and age-standardized DALY rates decomposed by NCDs, CMNNDs, and injuries from 1990 to 2019

Discussion

This study, using data on both the length and quality of life in 31 provinces of mainland China, employed advanced statistical methodologies in examining health inequality trends, and introduced a novel framework of triple decomposition analysis for comprehensive investigation of regional differences, determinants, and sources of health inequality at the provincial level in China.

Our results indicated that health status improved, and health inequalities decreased within China from 1990 to 2019. The health inequalities, as measured by the Gini coefficient, were primarily driven by inter-regional differences rather than intra-regional differences, with the most significant differences observed between the Eastern and Western regions. This finding was supported by another research [9], in which significant differences were also found between the western, central and eastern rural regions in health outcomes during 2000–2010.

Socioeconomic factors and health system factors were the primary determinants of interregional differences, which is overall similar to another study by Jiang et al. [28], focusing on determinants of life expectancy in China. The difference is that Jiang et al. found that education had the greatest impact on life expectancy, but they did not include economic factors as independent variables, while in our study, economic factors had a greater impact than education.

Although NCDs had increasingly become the main effectors of Gini coefficients of the ASMR and age-standardized DALY rates, the CMNNDs were most relevant for improving health inequality given their positive marginal contributions. Almost no previous research has found this issue.

The improvements in provincial health indicators in China, namely, LE, HALE, and the age-standardized DALY rates, slowed between 2010 and 2019, differing from the trends from 1990 to 1999 and 2000 to 2009. However, the improvement in the ASMR gradually accelerated, potentially due to the utilization of the Life Table for determining LE, and the younger the population group was, the greater the relative contribution to LE [29]. As a result, changes in child mortality had a relatively large impact on LE, and changes in adult mortality had a comparatively small impact. However, the contribution of each age group in the calculation of the ASMR mainly depended on the proportion of the age group population in the total population. Thus, the impact of changes in child mortality on the ASMR was fairly limited, and the impact of changes in adult mortality was relatively large. When economic and social development levels became relatively high, the child mortality rate was low [30], and further decreases were difficult to achieve; therefore, the rate of decline was slow. However, the adult mortality rate declined relatively quickly due to the prevention and control of environmental pollution, advancements in medical technology and other factors. In this case, a paradoxical slowing of the increase in life expectancy and an increase in the rate of decline of standard mortality may be occurring. This implies that the ASMR is more sensitive than the three other indicators and may not be a robust indicator for analyzing health inequality, which has not been revealed by previous studies as we know.

The health status among provinces generally displayed an obvious temporal convergence phenomenon over the past 30 years, regardless of whether σ-convergence, β-convergence, or club convergence occurred. This conclusion still holds after controlling for variations in economic development, public education, health services, health security, and other influential factors, thus providing a solid empirical basis for mitigating health inequalities in China. Oppositely, temporal divergence on health status has been observed in some other countries, which can be explained by a combination of socioeconomic, race/ethnicity, behavioral risk, and health care factors [31].

The evolution of the Gini coefficients of LE, HALE, and the age-standardized DALY rates can be roughly divided into two stages: the first was a significant decrease stage (1990–2007), and the second was a nearly stable stage (2008–2019). According to the detailed analysis using Dagum’s Gini coefficients, the overall Gini coefficients declined dramatically during the first stage, mainly because the differences between the western region and the eastern region and within the western region distinctly decreased. However, the declining trend of the Gini coefficients slowed and stagnated during the second stage. When further exploring the determinants of the health differences between eastern and western regions using the Oaxaca-Blinder decomposition, socioeconomic determinants such as per capita GDP contribute the most, followed by health system determinants such as health technicians per 1,000 population. This may have important implications for explaining the changing trends in health inequalities.

As indicated by the factor-decomposed Gini coefficient results, NCDs had the largest impact on the ASMR and age-standardized DALY rates, which is similar to other studies [31]. However, CMNNDs had the most important impact in the context of public policy and improving health inequality in terms of the marginal contribution rate.

The main contribution of our study may be that we introduced a novel framework of triple decomposition analysis for a comprehensive investigation of regional differences, determinants, and sources of health inequality at the provincial level in China. Compared with other similar decomposition methods, our decomposition has several advantages. Firstly, Dagum’s Gini coefficient decomposition in our study not only overcomes the limitation that the traditional Gini coefficient cannot be measured for groups but also solves the overlap problem encountered in Theil index decomposition. Secondly, when decomposing regional differences in health status, the Oaxaca-Blinder decomposition in our study can decompose the differences into two parts, one for the differences in the means of the independent variables (the explainable part) and the other for the differences in the intercept terms and regression coefficients (the unexplainable part), while the classical decomposition highlighted by Horiuchi et al. [32] can not distinguish between explainable and unexplainable parts. Thirdly, when decomposing diseases sources of health inequalities, the factor-decomposed Gini coefficient in our study can calculate both proportional contributions and marginal contributions from each sources, which may not be achieved by the classical decomposition highlighted by Horiuchi et al. [32].

Our study also has several limitations. Firstly, the analysis in this study was based on panel data obtained at the provincial level in China, and health inequalities couldn’t be explored in regard to factors such as race and gender. Secondly, the disease classification method only included three major categories and didn’t provide further subdivision of diseases. Thirdly, the lack of corresponding data prevented the conduct of rural / urban decomposition. Finally, health outcome inequalities only account for a portion of all health-related inequalities; notably, unequal health financing, resource allocation, and service utilization are important and require further in-depth analysis in future research.

Conclusions

In conclusion, the health levels in Chinese provinces improved significantly over the study period and health inequalities within China distinctly decreased. The importance of the effects of geographical factors, socioeconomic factors, health system factors, and the prevention and control of CMNNDs is discussed. These findings should aid in developing targeted strategies and policies to reduce health inequalities in China and provide inspiration for other developing countries to achieve better health equity.