Data sources
Our main analysis was conducted at the provincial level. The MMRs in 31 provinces between 2004 and 2016 were retrieved from the National Health Statistical Yearbooks [
5]. The reported MMRs in the yearbooks were annually collected by the National Maternal Mortality Surveillance System, a well-developed population-based maternal death registry system which was established by the Ministry of Health of China in 1989 [
11]. The accuracy of the data was guaranteed by a rigorous quality control mechanism including the standardized data collection procedures, strict data review and data management [
26].
As the purpose of our study was to evaluate the impact of hospital bed supply as well as its distribution inequality on MMR, three data resources were adopted for extracting hospital bed data. In the first step, provincial-level data on hospital beds between 2004 and 2016 was directly obtained from National Health Statistical Yearbooks [
5]. In order to measure the inequality of hospital bed distribution within each province, we further extracted county/city level data on hospital beds from China County/City Social and Economic Statistical Yearbooks [
27,
28]. According to the divisions of administrative areas in China in 2015 [
29], there was a total of 34 provinces with 2850 county level administrative units. Due to the data availability, we totally obtained hospital bed data on 2347 counties/districts (2082 counties in rural areas and 265 districts in urban areas) among 31 provinces in China, except for Hong Kong Special Administrative Region (SAR), Macao SAR and Taiwan province. We utilized the variations in hospital beds among the 2347 counties/districts to calculate the provincial inequality index of geographic distributions of healthcare resources.
Additionally, facility birth rate was extracted from the National Health Statistical Yearbooks [
5]. Other social-economic characteristics of each province were collected from the China Statistical Yearbook [
30] Detailed information for each variable were presented in Additional file
1: Appendix Table 1.
Independent variables
Hospital bed density and its inequality in geographic distribution were two primary independent variables in our study. According to China Health Statistical Yearbooks, hospital bed density was defined as the total number of beds in healthcare institutions by the end of Dec 31st, per 1000 registered population in each region [
5]. The hospital beds included regular beds, care beds and makeshift beds, but excluded the pre-delivery beds, beds in outpatient observation rooms and beds for newborn babies in obstetric wards [
5]. According to the definition, the reported hospital beds mainly refer to the beds that are available or potentially available for health professionals to provide medical care for patients. In China’s healthcare delivery system, the pre-delivery beds are solely used to help the pregnant women prepare for the deliveries, which are different from obstetric beds where obstetricians actually provide delivery services. Thus, the pre-delivery beds were excluded from the counts of hospital beds.
The inequality in geographic distribution of hospital beds within each province was measured by Gini coefficient, which was one of the most widely used indicators for describing social and economic conditions [
32‐
34]. The Gini coefficient was also recommended as a summarized measure of inequalities in health and has been applied in numerous empirical studies [
16,
20,
35]. In this study, we employed the geometric approach to define this index, with a calculation formula defined as follows:
$$ Gini=\frac{n+1}{n}-\frac{2}{n^2{\mu}_y}\sum \limits_{i=1}^n\left(n+1-i\right){y}_i $$
(1)
where
n is the total number of counties within a province,
μy is the mean of hospital bed density among all counties, and
yi refers to the county
i ’s hospital bed density. When calculated, counties should be ranked in ascending order of
yi at first and corresponding population numbers are usually considered as survey weights in the equation. The value of the Gini coefficient ranges from 0 and 1, with higher values indicating higher inequality in geographic distribution of hospital bed [
36]. According to previous studies, a Gini coefficient smaller than 0.2 means low inequality level, and values higher than 0.4 indicates extreme inequality [
37,
38].
The following variables that may confound the association between hospital bed supply as well as its inequality in geographic distributions and MMR were considered as covariates in our analyses: crude birth rate (‰), female illiteracy (i.e., the proportion of women aged 15 years or older who were illiterate) (%), gross domestic product (GDP) per capita (1000 Yuan), and urbanization rate (percentage of urban population) (%). The GDP per capita was adjusted for inflation based on the Consumer Price Index (CPI) (National Bureau of Statistics, 2017). We also included facility birth rate (%), as a crucial indicator of maternal healthcare utilization, when exploring the casual path from hospital bed density and Gini coefficient to MMR.
Statistical analysis
First, we performed descriptive analysis to investigate the time trends and regional variations in MMR between 2004 and 2016. Based on GDP per capita [
30], 31 provinces were categorized into five groups: the highest, upper middle, middle, lower middle, and lowest income. Line plot and box plot were used for depicting the time series and regional disparities in MMR in each province by year and income level.
Second, we depicted the time trends and regional variations in hospital bed density as well as its distribution inequality between 2004 and 2016. Two different methods were considered for measuring hospital bed density. The first method was to divide the counties into six categories based on hospital beds per 1000 population, and different shading indicated varied changes in hospital bed density. From the central governments’ definition, China was divided into three regions (eastern, central and western region) based on the disparities embedded in economic progress. Moreover, large variations were found in population quantity and land size at the county level between the eastern and western regions, where a vast territory with a sparse population was typically identified in western counties. Due to small population and large land size, the values of hospital bed density in western counties might conceal the real pattern of geographic distribution in hospital beds. Thus, the second approach further adjusted the land size of each county when calculating the bed density. In addition, we used box plot and bar plot to describe the time trends and variations in inequalities of hospital bed distributions (measured by Gini coefficient) by year and province.
Third, we used a linear mixed model to examine the effect of changes in hospital bed supply as well as its distribution inequality on MMR between 2004 and 2016. Since the correlated structure of yearly reported MMR within each province could be incorporated for analysis into the mixed model, both within- and between-province components of variation in MMR could be distinguished. Moreover, an autoregressive covariance structure was fitted for residual effects to account for serial correlation in MMR across time within the provinces. The model was set as follows:
$$ {\displaystyle \begin{array}{c}\log \left({MMR}_{ij}\right)={\alpha}_{0j}+{\beta}_1{Bed}_{ij}+{\beta}_2{Gini}_{ij}+{\mathbf{X}}_{\boldsymbol{ij}}^{\prime}\lambda +{\varepsilon}_{0 ij}\\ {}{\alpha}_{0j}={\alpha}_0+{\mu}_{0j}\\ {}{u}_{0j}\sim N\left(0,{\sigma}_{\mu_0}^2\right)\\ {}{\varepsilon}_{0 ij}\sim N\left(0,{\sigma}_{\varepsilon_0}^2\right)\end{array}} $$
(2)
where
i denoted the year ranged from 2004 to 2016 and
j indexed the provinces. The
MMRij was the maternal mortality ratio for province
j in year
i.
Bedij and
Giniij were two variables of interest, which indicated the hospital bed density and Gini coefficient for province
j in year
i.
X′
ij denoted a vector of provincial characteristics, including female illiteracy, crude birth rate, GDP per capita and percentage of urban population. Year dummy variables were also included in the regression model. To address the positive skewed distribution of
MMRij, we employed the natural logarithm of this ratio to fulfill the assumption of linearity. Thus, the intercept
α0j indicated the average logarithm of MMR between 31 provinces, equaling
α0 (total mean of log(
MMRij)) plus a random effects’ terms
μ0j.
ε0ij was the error term. The coefficient on
Bedij (
β1) and
Giniij (
β2) were two parameters of primary interest. As we used logarithm of MMR in the model, a one-unit increase in the estimated coefficients
\( \hat{\beta} \) would produce an expected increase in log(
MMRij) of
\( \hat{\beta} \) units. In terms of
MMRij itself, this meant that the expected value of
MMRij was multiplied by
\( {e}^{\hat{\beta}} \). Thus, we exponentiated the coefficient
\( \hat{\beta} \), subtracted one from this number, and multiplied by 100, which gave a corresponding percent increase (or decrease) in the
MMRij for every one-unit increase in the independent variables. So given a negative value of estimated coefficient
\( {\hat{\beta}}_1 \), one-unit increase in hospital bed density was associated with a
\( \left({e}^{{\hat{\beta}}_1}-1\right)\times 100 \) percent decline in the MMR. For a positive value of
\( {\hat{\beta}}_2 \), a 0.1-unit increase in Gini coefficient, indicating a relatively larger inequality in geographic distribution of hospital bed, was associated with a
\( \left({e}^{{\hat{\beta}}_2\times 0.1}-1\right)\times 100 \) percent increase in the MMR.
Fourth, we explored whether the effect of hospital bed density and Gini coefficient on MMR was mediated by facility birth rate as well as the extent to which this effect was mediated. The three steps approach outlined by Baron and Kenny [
39] was used to assess the mediating effect. Three mixed-effects regression models were estimated: (1) a model examining the effect of hospital bed density (or Gini coefficient) on MMR; (2) a model examining the effect of hospital bed density (or Gini coefficient) on facility birth rate; (3) a model for MMR conditioning on hospital bed density (or Gini coefficient) and facility birth rate. The mediating influence of facility birth rate on the effect of hospital bed density (or Gini coefficient) on MMR was tested using the Sobel test [
40]. Moreover, we performed the mediation analysis for two periods respectively (2004 to 2009 and 2010 to 2016). Since the maternal mortality reduction and neonatal tetanus elimination programme was launched in 1999 [
10], facility birth rate in China has been steadily increased. In 2009, this program was implemented in a nationwide range in order to provide free hospital delivery for all women in China [
41], which achieved universal facility births by 2009, with an average rate of 94.51% reported in 31 provinces (Table
1). Thus, we conducted the mediation analysis for two different periods to examine the mediating role of facility birth rate. In addition, it should be noted that there are some limitations in this traditional mediation analysis implemented within the framework of linear structural equation models [
42]. A new approach, namely causal mediation analysis, has been developed under the counterfactual framework [
42]. But in this analysis, we primarily aims to identify the role of facility birth rate on the association between hospital bed density (or Gini coefficient) and MMR, instead of identifying the causality which needs more assumptions to be clarified. In this regards, the traditional mediation method was adopted in the analysis.
Table 1
Characteristics of 31 provinces in China, 2004–2016
Independent Variables |
Hospital beds per 1000 population | 2.76 (2.44, 3.08) | 3.48 (3.22, 3.73) | 5.36 (5.11, 5.61) | 2.41 (0.92) | 3.31 (0.84) | 5.39 (1.27) | 2.60 (2.22, 2.97) |
Gini coefficient | 0.32 (0.28, 0.35) | 0.30 (0.26, 0.33) | 0.25 (0.22, 0.28) | 0.34 (0.09) | 0.31 (0.09) | 0.26 (0.07) | −0.07 (−0.09, − 0.05) |
Birth rate, ‰ | 11.45 (10.26, 12.65) | 11.38 (10.40, 12.36) | 11.80 (10.72, 12.88) | 11.67 (4.56) | 11.70 (4.17) | 12.18 (3.93) | 0.35 (−0.36, 1.06) |
Female illiteracy, % | 16.18 (12.67, 19.69) | 11.84 (8.75, 14.93) | 9.17 (6.11, 12.22) | 14.22 (12.03) | 9.90 (7.17) | 7.32 (5.12) | −7.01 (−8.25, −5.77) |
GDP per capita, 1000 yuan | 19.39 (14.01, 24.76) | 34.75 (27.25, 42.24) | 56.77 (47.33, 66.20) | 13.23 (12.45) | 26.33 (22.44) | 46.38 (32.06) | 37.38 (32.13, 42.63) |
Urbanization rate, % | 29.94 (24.02, 35.86) | 49.27 (43.89, 54.65) | 57.85 (53.28, 62.42) | 25.99 (12.47) | 46.00 (15.65) | 56.21 (14.39) | 27.91 (24.23, 31.59) |
Facility birth rate, % | 81.68 (75.40, 87.97) | 94.91 (91.38, 98.44) | 99.47 (98.90, 100.04) | 85.40 (21.40) | 98.50 (3.80) | 100 (0.40) | 17.79 (11.89, 23.69) |
Dependent Variable |
Maternal mortality ratio, per 100,000 live births | 55.00 (34.56, 75.44) | 28.84 (14.33, 43.35) | 16.63 (9.80, 23.46) | 42.80 (43.70) | 19.20 (16.40) | 12.70 (8.10) | −38.37 (−52.46, −24.28) |
All analyses were performed using ArcGIS 10.5, SAS 9.4, and R 3.4.3.