Data sources
We used a dataset from a cross-sectional household health survey conducted in Shaanxi Province in 2013. This survey was a part of the fifth National Health Services Survey (NHSS) and was the largest health survey in Shaanxi Province, an underdeveloped region in western China. Shaanxi Province covers an area of about 205,800 km2 and has a population of 37.6 million in 2013, among which people aged 45 years and older account for about 33.5% [26].
A structured questionnaire was used to conduct face-to-face interviews with all members of selected households. Questionnaires collected data about demographic characteristics, the economic status (annual household expenditure), other socioeconomic information (e.g. educational level, employment status), and health status, including information about chronic conditions and preventive care specific to hypertension and diabetes. In our survey, respondents were required to answer questions on their own and proxy responses by familiar family members were only used when the respondents were unable to express themselves accurately.
A four-stage stratified cluster random sampling approach was used to select the representative survey respondents in Shaanxi Province. In short, 32 districts or counties were stratified, among which 160 sub-districts or townships were randomly selected. Next, 320 communities or villages were randomly selected from these sub-districts or townships. Finally, 20,700 households were randomly selected, and every family member of these households (in accordance with census register information) was interviewed, adding to a total of 57,529 surveyed individuals. In the process of data collection, extensive quality control measures were conducted as previously described [27]. In our study, we only analyzed data from a total of 27,728 individuals aged 45 years and older, focusing on an age group more likely to be affected by chronic diseases.
Variables
The variables used in this paper included: a) the presence of hypertension/diabetes and the behavior of preventive care for those with hypertension/diabetes (these are the outcome variables), b) economic status, c) demographic and other socioeconomic factors.
-
(a)
the presence of hypertension/diabetes and the preventive care behavior
Hypertension and diabetes were selected as model examples of chronic disease in this study because they represent the two most common diseases and the main causes of death and disease burden in China [9, 10]. Furthermore, adequate use of medication and monitoring blood pressure/blood glucose were the two indicators selected as proxies of patient preventive care behavior because these two indicators are the main means of preventing and monitoring complications and advancement of both diseases.
The outcome variables in the study were all binary. The first set of variables was the presence of hypertension and/or diabetes. The presence of hypertension/diabetes was determined from self-reports of physician-diagnosed hypertension/diabetes from answers to questions such as “Have you ever been diagnosed with hypertension?” and “Have you ever been diagnosed with diabetes?” Next, if the respondents had hypertension/diabetes, they were asked whether they made an adequate use of medication according to their physician’s instructions and whether their blood pressure/blood glucose had been monitored in the last three months.
-
(b)
the economic status
The household consumption expenditure per equivalent adult was used as a proxy measure of economic status in this study as proposed by previous studies [28]. It was derived by dividing household consumption expenditure by the equivalent number of adults in the household which has been described in detail before [29, 30]. For the regression analysis, the economic status variable was divided into quintiles according to the household consumption expenditure per equivalent adult as follows: poorest (i.e. the lowest 20%, ≤ 5030 Yuan), poorer (lower 20%, 5031–7704 Yuan), middle (middle 20%, 7705–10,976 Yuan), richer (higher 20%, 10,977–16,651 Yuan) and richest (highest 20%, ≥ 16,652 Yuan).
-
(c)
the demographic and other socioeconomic variables
The demographic characteristics included were gender and age. Age was categorized into three groups: 45–64, 65–79, and 80 or above. Other socioeconomic characteristics included were educational level, employment status, marital status, basic medical insurance (no or yes), commercial medical insurance (no or yes), and urban or rural areas. Educational level was categorized into four groups: primary or below education, middle school, high school, and college or above education. Two employment status categories were employment and unemployment. Two marital status categories were married and single (including unmarried, divorced, separated and widowed). The basic medical insurance refers to China’s public insurance programs including Urban Employee Basic Medical Insurance, New Rural Co-operative Medical Scheme and Urban Residents Basic Medical Insurance. These three insurance programs cover more than 98% of the population in China.
Statistical analysis
The data analyses performed in the study included a descriptive analysis, a simple weighted point estimate, an estimation of the relative index of inequality (RII), a calculation of the economic-related concentration index (C), and a further decomposition analysis for the C. The RII and C are widely used to estimate a magnitude of inequality in health or healthcare [31, 32]. RII compares extremes and C summarizes inequality across the entire socioeconomic spectrum. All analyses were performed independently for hypertension and diabetes.
-
(a)
the relative index of inequality
The RII, mainly used in public health and epidemiology, is a regression-based index of inequality used to compare rates of disease prevalence between those with the lowest and the highest socioeconomic status [31]. RII can be interpreted as the ratio of the estimated prevalence of disease between the poorest and the richest. Therefore, a RII value greater than one signifies a higher prevalence among those with the lowest socioeconomic status and vice versa. Given that binary outcomes, Poisson regressions with robust error variance were used to generate the prevalence rate ratio estimates across the economic groups adjusted for confounding variables (i.e., RIIs) in a cross-sectional study, as suggested in previous reports [33]. RIIs were reported in two stages: first RIIs were adjusted for age and gender, while second RIIs were additionally adjusted for educational level, occupational status, and other socioeconomic factors.
-
(b)
the concentration index
The C captures the socioeconomic-related inequalities in health or health care and gives a measure of the magnitude of inequality across the entire socioeconomic spectrum [32]. The C ranges from − 1 to 1, with an index of 0 equivalent to perfect equality. A positive C signifies that a health or health care variable is more concentrated among the richer population and vice versa. The C formula is as follows:
$$ \mathrm{C}=\frac{2}{\mu}\mathit{\operatorname{cov}}\left({y}_i,{r}_i\right) $$
(1)
Where y is the health or health care variable (e.g. hypertension/diabetes prevalence and preventive care in this study), μ is the mean of the health or health care variable, ri is the fractional rank of the i th individual in the economic distribution, ranging from 0 to 1.
-
(c)
the decomposition analysis for C
Wagstaff et al. proved that the C can be decomposed into its contributing demographic and socioeconomic factors, where the contribution of each factor is the product of the degree of economic-related inequality in that factor and the sensitivity of the health or health care outcome variable with respect to that factor [34].
Since the outcome variables analyzed in our study were binary variables, probit regressions were used to calculate the partial effects of each explanatory variable and the results should not be used to infer a direction of causality [35]. All explanatory variables in regressions were categorical dummy variables. Health (or a health care outcome variable) (y) is modelled as follows:
$$ {y}_i={\sum}_k{\beta}_k{x}_{ki}+{\varepsilon}_i $$
(2)
where βk are the partial effects, dy/dx of each regressor is evaluated at the sample mean; and ε is the error term. xk are a set of explanatory variables.
The concentration index C(y) can be decomposed as:
$$ \mathrm{C}={\sum}_k\left(\frac{\beta_k{\overline{x}}_k}{\mu}\right){C}_k+\frac{GC_{\varepsilon }}{\mu } $$
(3)
where βk are the partial effects of the k regressors (i.e. explanatory variables), taken from Eq. (2). \( {\overline{x}}_k \) are the means of each regressor and μ is the mean of the health or health care variable. Ck is the concentration index of each regressor and GCε is the generalized concentration index of ε. The residual component (\( \frac{GC_{\varepsilon }}{\mu } \)) represents the inequality that is not explained by the regressors. The deterministic component (\( {\sum}_k\left(\frac{\beta_k{\overline{x}}_k}{\mu}\right){C}_k \)) focuses on two elements. These are the degree of unequal distribution of each regressor across the economic spectrum (Ck) and the elasticity of health with respect to the regressor\( , \kern0.5em \left({\eta}_k={\beta}_k\frac{{\overline{x}}_k}{\mu}\right) \). We calculated the absolute contribution of each regressor (Qk = ηkCk). The contribution of each regressor can take both positive and negative values. According to Eq. (3), even if an explanatory variable has a massive effect on the health or health care variable, if the variable is equally distributed between the rich and the poor, then the explanatory variable is not a key source of inequality. Then we calculated the percentage contribution of each regressor (100Qk /C). Of note, the negative and positive contributions may cancel out in the aggregate and the percentage contribution of the regressors and error term sum would be 100%, so the percentage contribution of several regressors may represent large positive and negative contributions, even over 100%.
Inequalities in health or health care are associated with demographic factors, e.g. age, and socioeconomic-related factors, e.g. economic resources and urban-rural indicators. Policy makers may be more focused on inequalities arising from socioeconomic-related factors, because some of the demographic factors are inevitable [35]. In this study, age-sex adjusted C was calculated by subtracting the contributions of age and gender from the total C based on the decomposition results [36, 37].
Sampling weights were used to account for the sampling design and to ensure the results represented the population of Shaanxi Province as previously described [38]. Svyset command in STATA version 14.0 was used to specify the design for all analytical models. Furthermore, we adjusted standard errors for clustering at the family level for all models.