To what extent do disparities in economic development and healthcare availability explain between‐province health inequalities among older people in China?

Abstract Background Uneven economic development has led to substantial health inequalities between Chinese provinces. The extent of, and factors underlying, between‐province health inequalities have received little attention. Methods Data from 15,278 respondents in Wave 2 (2013) of the China Health and Retirement Longitudinal Study (CHARLS) were used to investigate inequalities among people aged ≥50 years in five health outcomes between 27 Chinese province‐level administrative units. After characterizing the between‐province differences and the relevance of province effects, proportional change in variance between unadjusted and adjusted models was calculated to determine the percentage of between‐province variance in health outcomes explained by province‐level variables including measures of economic development and healthcare availability. Results Although province effects explained <10% of overall variance in health outcomes, they underpinned large between‐province inequalities among people aged ≥50 years. Gross Regional Product per capita was more important than doctor density in explaining between‐province variance in health outcomes, particularly depression symptoms and instrumental activities of daily living impairment. Conclusion Policy efforts, including more equal distribution of healthcare personnel, may be warranted to reduce between‐province health inequalities.

China has experienced rapid economic development since 1978 and a transition from a primarily agricultural to an industrial economy [1].This process has been accompanied by rising inequalities in wealth and income from the household level to the province level, however.Economic disparities between export-oriented coastal provinces and inland provinces in terms of gross regional product (GRP) increased markedly from around 1990 until the mid-2000s [2].Between-province differences in GRP per capita and household income have since started to decline since then, however [3].
Economic changes have been accompanied by a shift in China's population age structure due to declines in both fertility and mortality, and growth in the proportion of people aged ≥50 years among the general population partially as a consequence of the One-Child Policy [4].At the same time, noncommunicable diseases have become a significant policy concern even since the early 1990s [5], driven by changes in diet, physical activity, smoking prevalence, environmental pollutants, and other factors [4,6].
Economic liberalization and uneven development have contributed to increasing inequity in access to public services.Following initial economic reforms, financing of health was devolved from central to provincial governments, and rural health services organized by agricultural collectives were dissolved [7].This has resulted in an increasing concentration of health service expenditure and utilization in more developed provinces [8], greater betweenprovince inequality in health resources such as health workers and beds, and wide disparities in physician pay [3,9].
The relative extent to which inequalities in population health status is explained by economic inequalities and inequalities in healthcare access has yet to be elucidated, and both have contributed to growing between-province health inequalities.This is particularly the case for noncommunicable diseases.Regarding specific outcomes, although prevalence of depression is comparable with other developing countries, there are wide disparities in prevalence between different regions of China [10,11].Prevalence of impairment in instrumental activities of daily living (IADL) has decreased over time but remains higher in rural areas [12].Prevalences of overweight and chronic obstructive pulmonary disease have risen markedly in recent decades and vary significantly across Chinese regions [13,14].
To my knowledge, no systematic attempt has been made to quantify the degree of between-province health inequalities in China, or to identify variables underlying these inequalities.While previous studies have investigated province-level effects in China, these failed to quantify either the magnitude of between-province inequalities or the proportion of variance explained by province-level variables [15,16].Partitioning of variance within a multilevel framework has been employed in studies in other contexts.For example, one study has investigated the proportion of between-country variance in well-being change following exit from paid employment explained by national-level welfare policies across European countries [17].
The objectives of this study were to: (1) Characterize prevalence of five health outcomes (depression symptoms, IADL impairment, limitation in physical functioning, overweight, and lung function impairment), and per capita economic output and doctor density, by province, and betweenprovince differences.(2) Investigate the fixed-effects individual-and provincelevel predictors of each health outcome measure.(3) Estimate and interpret the proportion of overall variance in each health outcome attributable to province effects, and the proportion of province effects attributable to per capita economic output and doctor density.[18].The analytic sample included respondents aged ≥50 years (any outcome, n = 15,278) 1 .CHARLS has been approved by the ethics committee of Peking University Health Science Center and all participants gave written informed consent before participation.No further ethical approval or participant consent was required for this study as it was based on a secondary analysis of existing data.

| Variable definitions
We defined six health outcome measures, operationalized as binary variables.These included depression symptoms based on the CES-D-10 instrument [19], which consists of 10 Likert-type items and yields scores ranging from 0 to 30. 2 A cutoff score of 10, which has shown high specificity in samples of older people, was used to define probable depressive cases [20,21].Limitation in physical functioning 3 and IADL impairment 4 were defined using self-reports of difficulties in performing physical and functional tasks.Overweight was defined using a body mass index (BMI) cutoff of ≥24, which is the typical cutoff employed in China [22], based on recordings of height and weight taken during the nurse visit.Lung function impairment was defined as having a peak expiratory flow (PEF) of <70% of an individual's expected value based on a formula including age, sex, and height [23,24]  5 .Two province-level variables for 2013 were extracted from the National Bureau of Statistics (NBS) China Statistical Yearbook to measure economic development and healthcare availability [28].GRP per capita, a provincelevel measure of economic output equivalent to Gross Domestic Product per capita, was denominated in 2015 Yuan (¥) per inhabitant by increments of ¥1,000 (or $282 in 2015 Purchasing Power Parity adjusted United States Dollars) [29].Doctor density was measured in doctors per 10,000 inhabitants.Doctors are defined as those who have passed a licencing examination and registered at a county or higher level as physicians or assistant physicians [9].
Individual-level covariates included sex (male or female), age (years), partnership status (married and cohabiting, married but living apart, never married, or divorced, separated or widowed), ethnicity (Han or other), residence (urban or rural), health insurance (uninsured, enhanced government-sponsored insurance, basic government-sponsored insurance or private insurance) 6 , employment status (working or not working), quintile of gross square root equivalized household income, level of education (less than elementary school, elementary school, middle school, high school or vocational college and university), smoking status (nonsmoker or current smoker) and alcohol intake 1 This this age cutoff was selected as it corresponds to the minumim age for inclusion in comparable panel studies (such as the Health and Retirement Study (HRS) and the English Longitudinal Study of Ageing (ELSA), is consistent with typical definitions of an "older person" employed in previous studies, and, in China, is the minimum statutory retirement age (for women). 2 Depression symptom outcomes were measured using the 10-item Center for Epidemiological Studies Depression Scale (CES-D-10).This instrument was based on a 10-item scale comprising symptoms includeing feeling "bothered," "trouble concentrating," "depressed," "effort," "fearful," "restless sleep," "hard to get going," "hopeful," and "happy" within the last week.Each item was scored on a four-point scale ranging from 0 ("rarely or none of the time") to 3 ("most or all of the time"), and the former two positively worded items were reversecoded before calculating summary scores.Summary scores ranged from 0 to 30 and cutoff score of 10 was used to define probable depressive cases. 3Respondents were categorized as having a limitation in physical functioning if they self-reported difficulties in performing any of the following tasks: sitting, climbing stairs, kneeling, reaching or extending, lifting objects, picking up objects.This variable was not based on a validated scale but has been employed in previous studies on functional abilities of older people.While some studies 25 refer to these as "limitation in physical functioning," other studies have also used the term "limitation in mobility" or "limitation in mobility activities". 26,27 4 Respondents were defined as having an IADL impairment if they reported problems carrying out any of the following: managing finances, managing transportation, shopping, (hot) meal preparation, household chores, communication, and managing medications.Respondents' peak expiratory flow (PEF), defined using the highest of three measures taken with a peak flow meter during the nurse visit, was compared with their expected value based on their gender, age, and height calculated using the equations below (see Nunn and Gregg [22]): Respondents whose actual PEF was <70% of their predicted value were categorized as having lung function impairment.This cutoff represents a highly sensitive measure of severe COPD.CHARLS Wave 2 respondents were invited to self-report their type of health insurance coverage from a list of 10 options.In this study, respondents were categorized by their type of health insurance coverage using the four categories below.Categorization was hierarchical, and respondents who reported having more than one type of health insurance coverage were assigned to the highest (lowest-numbered) category: 1. Government scheme, enhanced coverage: Government employee medical insurance (公费医疗), urban employee medical insurance (职工医保) 2. Government scheme, basic coverage: New cooperative medical insurance (新型农村合作医疗), rural and urban resident medical insurance (城乡居民基本医疗保险) 3. Private or other: Any other insurance scheme not included in the questionnaire 4. Uninsured: No membership of any health insurance scheme reported.
(none, once per month or >once per month).A 0-7 point index of housing quality was specified as a proxy for household wealth and based on the presence of electricity, an indoor toilet, central heating, internet, running water, sufficient living space (<2.5 people per room), and brick or concrete as primary building materials (Cronbach's α = 0.81).

| Analytic methods
All analyses were performed in Stata 14 [30].We used logistic random-effects models for binary outcomes using the xtmelogit command to investigate the odds of respondents meeting the criteria for each of the six health outcome measures.Observations were nested within province-level units and considered nonindependent due to spatial dependence.All models fitted random intercepts for each province.For models with covariate adjustment, covariates were fitted as fixed effects only.Analysis of variance components involved estimations of three random-effects statistics: intraclass-correlation coefficient (ICC), median odds ratio (MOR), and proportional change in variance (PCV).
ICC is defined as "the proportion of the variance explained by the grouping structure" [31].Calculations of ICC are based on both individual-level and area-level variance.In multilevel logistic regression, area-level variance is calculated on a logistic scale while individual-level variance is on a probability scale.This study employed the latent variable method to convert individual-level variance to the logistic scale [31].
In logistic random-effects models for binary outcomes, ICC is calculated using the latent variable method by the following equation [32]: where V A represents residual area-level variance.The unobserved individual variable follows a logistic distribution with variance equal to π 2 /3 (3.29).The latent variable method assumes that ICC is a function only of the area-level variance and is independent of the prevalence of the outcome.The simulation method, proposed by Merlo et al. [33], was not used as it was considered unecessary.First, although estimation of random-effects parameters may be biased when there are fewer than 20 area-level units, this study employed data from 27 province-level administrative units of China.Second, the assumption of the latent variable method that an individual's propensity to be positive for a given outcome is a continuous latent variable underlying the binary response variable was likely to have been met as all outcomes were derived from continuous measures and coded using cutoffs.
MOR, a measure of residual heterogeneity between areas, translates area-level variance into an odds ratio (OR).It is defined as the median OR for a given outcome between the area at highest risk and the area at lowest risk when randomly selecting two areas, and is statistically independent of the prevalence of the phenomenon in question [33].MOR can be directly compared with fixed-effects OR estimates from the same model when considering the relevance of residual between-area heterogeneity.The MOR is based on estimates of area-level variance (V A ) and was calculated using the following formula: In this context, 0.6745 is the 75th centile of the cumulative distribution function of the normal distribution with mean 0 and variance 1.
PCV is calculated by comparing variance estimates from an "unconditional" or "empty" intercept-only model (without adjustment) with those from a "conditional" model (adjusted for individual-and/or provincelevel covariates) [34].It is defined as the percentage difference in area-level variance between the empty and conditional models, and describes the proportion of arealevel variance explained after adjusting for (individual or groups of) individual or area-level variables in the conditional model.
Between-area differences can be attributable to compositional (i.e., differences in their population characteristics) and contextual factors (i.e., betweenarea differences).PCV, expressed as a percentage of level-2 variance explained by compositional and/or contextual factors, was estimated using the following formula: where V A represents an estimate of area-level variance from the "empty" (unconditional) intercept-only model without adjustment for either individual-or provincelevel variables (Model 2 in this study) and V B residual area-level variance from the "conditional" model (Models 3-6) with adjustment for individual-and provincelevel covariates.

| Descriptive analysis
We estimated the prevalence of each of the six health outcomes in 2013, both overall and by province, using cross-sectional survey weights for individual respondents.Provinces were categorized by quartile of prevalence of each health outcome, per capita GRP, and doctor density.Individual-level characteristics of the analytic sample were then described for each outcome.

| Statistical analysis
Six models were fitted for each health outcome.Model  The meresc package was used to rescale the results of xtmelogit models to the same scale as unconditional intercept-only models and allow comparison of variance components across models [35].We analyzed complete cases only as meresc does not support missing data techniques such as multiple imputation.Model 3 was fitted as a conditional model with individual-level variables (compositional factors), and fixed-effects OR and 95% confidence intervals (95% CI) were estimated to show associations between individuallevel variables and each health outcome.Model 4 was fitted as a conditional model with adjustment for individual-level variables and province-level GRP per capita while Model 5 was adjusted for individual-level variables and doctor density.Model 6 fitted all individual level covariates in addition to both province-level GRP per capita and doctor density.Fixed-effects associations between province-level variables and health outcomes, in addition to PCV compared with Model 2, were estimated for Models 4-6.

| Descriptive analysis
Table 1 shows GRP per capita and doctor density by province and categorizes them by quartile.Province-level GRP per capita ranged from ¥23,151 (Guizhou) to ¥100,105 (Tianjin).Mean GRP across the 27 provinces was ¥49,248.Mean doctor density was 2.2 doctors per 10,000 inhabitants, with a range of 1.6 (Anhui and Jiangxi) to 4.3 (Beijing).Table 2 shows the prevalence of each health outcome, in addition to GRP per capita and doctor density, by province, and categorizes provinces by quartile.There was an east-to-west gradient in prevalence of depression symptoms, limitation in physical functioning and IADL impairment, and a south-tonorth gradient in overweight.Table 3 shows the individual-level characteristics of the analytic sample for each outcome measure.

| Fixed-effects
Tables 4 and 5 show the fixed-effects estimates for the associations between individual-level variables and each health outcome (Model 3).Age and high alcohol intake (>once per month) were positively and significantly associated with higher odds of all health outcomes, except for overweight for which p > 0.05.Higher housing quality was negatively associated with all health outcomes except lung function impairment and overweight.Higher household income was protective against all outcomes except lung function impairment but predictive of overweight.Smoking status was associated with lung function impairment.Age was associated with lower odds of depression symptoms and overweight but positively associated with higher odds of other outcomes.Associations between other individual-level variables and each of the six health outcomes varied.Table 6 shows fixed-effects associations between province-level variables and each outcome (Models 4-6).GRP per capita was significantly associated with lower odds of reported depression symptoms (OR: 0.86, 95% CI: 0.81, 0.92, p < 0.001) and IADL impairment (OR: 0.94, 95% CI: 0.87, 1.00, p = 0.049), and higher odds of overweight (OR: 1.01, 95% CI: 1.00, 1.02, p= 0.039), but not other outcomes (Model 4).Doctor density was associated with lower odds of depression symptoms (OR: 0.54, 95% CI: 0.40, 0.75, p = 0.002) (Model 5).After mutual adjustment, only GRP per capita was found to have a significant association with the health outcomes: there was a borderline significant association between GRP per capita and lower odds of depression symptoms, and a significant association between GRP per capita and lower odds of IADL impairment (Model 6).

| Random-effects
Random-effects area-level variance parameter estimates for each health outcome are shown in Table 6.Variance parameter estimates were similar between Models 1 and 2 for all outcomes; this implies that estimates of province effects were unaffected by loss of respondents with missing covariate observations from the analytic sample.
ICC estimates for Model 2 showed that province effects accounted for 5.8% of variance in depression   The results of Model 3 show random-effects statistics after adjustment for individual-level variables (compositional effects).PCV estimates show that these accounted for 46.5% and 50.% of between-province variance in depression symptoms and overweight but only 9.4% and 3.3% for limitation in physical functioning and IADL impairment.After adjustment for GRP per capita, Model 4 explained the largest proportions of between-province variance in depression symptoms, IADL impairment, and overweight (72.3%, 28.2% and 55.4%).After adjustment for doctor density, meanwhile, Model 5 explained 67.9% and 50.8% of between-province variance in depression symptoms and overweight, but only 9.6% and 5.5% of variance for limitation in physical functioning and IADL impairment.When models mutually adjusted for GRP per capita and doctor density (Model 6), PCV estimates for depression symptoms, limitation in physical functioning and IADL impairment were 71.7%, 21.5%, and 27.4%.

| DISCUSSION
After mutual adjustment for GRP per capita and doctor density, only GRP per capita both independently predicted depression symptom outcomes and limitation in physical functioning.Within-province differences (Model 2) still explained the majority of variance across all outcomes (93.6%-97.4%).The scale of province-level differences, as shown by MORs, was still large across all outcomes.The impact of residual between-province residual heterogeneity on depression symptom outcomes (MOR: 1.36) is comparable to that of being divorced, separated, or widowed (OR: 1.33, 95% CI: 1.15, 1.53) (Model 3).Although 46.5% and 50.0% of betweenprovince variance in depression symptoms and overweight outcomes were explained by compositional effects (Model 3), PCVs were ≤10.6% for all other outcomes.GRP per capita was important in explaining between-province variation in depression symptom outcomes (Model 4, PCV: 72.3%).
To my knowledge, this study is the first to quantify the relevance of between-province heterogeneity in health outcomes in China, and to estimate the degree to which these are explained by province-level variables.Its methods present a new approach to investigating health inequalities at the national and subnational levels.The results highlight the wide inequalities among people aged ≥50 years in multiple health outcomes between Chinese provinces as found in previous studies [10,[13][14][15].As Zhou et al. [6] conclude: "China in epidemiological terms is not one nation, but five: rapid transitions are occurring in all of them, but the most important health problems and the challenges imposed on the health system by demographic and epidemiological change are different." The results underscore that there is relatively limited potential for reduction in between-province health inequalities through ensuring more equal distribution of healthcare personnel, with GRP per capita being more important in explaining between-province differences in health outcomes (particularly for depression and IADL impairment outcomes).However, availability and distribution of qualified healthcare personnel can be improved through expansion of training, standardization of credentials and specific initiatives such as exchange programmes to increase equity between provinces [36,37].

| Strengths and limitations
Strengths of the study include its large sample size and CHARLS' representativeness of the Chinese population.The available data provide near-complete coverage of Chinese provinces, in addition to a wide range of health outcomes encompassing psychosocial and physical functioning, and objective measures of BMI and lung function.
CHARLS only samples official residents with valid Hukou status and excludes individuals residing in institutions such as nursing homes, (the latter represent only a small proportion of older people in China) [18].Non-Hukou residents are more likely to reside in provinces with higher GRP per capita, and may experience disparities in health and access to healthcare compared with permanent legal (Hukou-holding) residents [3].This may have undermined representativeness of province-level samples, and inflated GRP per capita estimates for wealthier provinces [38].Doctor density does not account for distribution of personnel across primary and secondary care facilities or numbers of specialists.
The results of this study do not imply causation as to the province-level determinants of the health outcomes analyzed but provides a description of the degree to which they are explained by the former.If causal associations exist, these may be mediated by other factors; for example, the association between provincelevel GRP and overweight may be mediated by factors such as nutrition and sedentary lifestyle.

| CONCLUSION
Ongoing reforms are targeted at equalizing investment and public services across provinces [7], in addition to achieving universal and equitable coverage of basic healthcare for all Chinese citizens [35].As all provinces of China complete their epidemiological transition, tackling the noncommunicable disease burden will be key to reducing both social and health inequalities [4].Success in reducing between-province health inequalities requires a coordinated health policy approach across national and province-level governments.Localized approaches to healthcare delivery are needed to address different provinces' diverse challenges, however [6].

T A B L E 3 4
Characteristics of analytic samples for six health outcome measures, China Health and Retirement Longitudinal Study (CHARLS) Wave 2 (2013) Results of conditional logistic random-effects models for depression caseness, functional disability and instrumental activities of daily living (IADL) impairment with adjustment for individual-level variables (Model 3)

F
I G U R E 1 Random-intercepts residual plots of province-level effects for depression symptoms (CES-D-10), (a) limitation in physical functioning, (b) instrumental activities of daily living (IADL) impairment, (c) overweight (d) and lung function impairment (E) derived from unconditional models (Model 2) expressed as odds ratios with 95% confidence intervals China National Bureau of Statistics (NBS) estimates of gross regional product (GRP) per capita and doctor density by province in 2013 T A B L E 1 T A B L E 2 Percentage prevalence of six health outcomes by province for individuals aged 50 years and over in China Health and Retirement Longitudinal Study (CHARLS) Wave 2 (2013) Results of conditional logistic random-effects models for overweight and lung function impairment with adjustment for individual-level variables (Model 3) Summary of analyses of between-province variance and area level effects for depression, functional disability, instrumental activities of daily living (IADL) impairment, 4% in overweight, and 4.0% in lung function impairment, but only 2.3% and 2.6% of variance in limitation in physical functioning and IADL impairment in the analytic sample (Model 2).Figure1(a-e) shows random-effects residuals for each health outcome by province transformed into ORs with 95% CIs relative to the grand mean for the 27 province-level units (Model 2).MORs ranged from 1.53 for depression symptoms and 1.57 for overweight to 1.30 and 1.33 for limitation in physical functioning and IADL impairment.
T A B L E 6