Area socioeconomic status is independently associated with esophageal cancer mortality in Shandong, China

Esophageal cancer (EC) is a leading cause of cancer death in China. Within Shandong Province, a geographic cluster with high EC mortality has been identified, however little is known about how area-level socioeconomic status (SES) is associated with EC mortality in this province. Multilevel models were applied to EC mortality data in 2011–13 among Shandong residents aged 40+ years. Area-level SES factors consisted of residential type (urban/rural) of the sub-county-level units (n = 262) and SES index (range: 0–10) of the county-level units (n = 142). After adjustment for age and sex, residents living in rural areas had a 22% (95% CI: 13–32%) higher risk of dying from EC than those in urban areas. With each unit increase in the SES index, the average risk of dying from EC reduced by 10% (95% CI: 3–18%). The adjustment of area-level SES variables had little impact on the risk ratio of EC mortality between the high-mortality cluster and the rest of Shandong. In conclusion, rural residence and lower SES index are strongly associated with elevated risks of EC death. However, these factors are independent of the high mortality in the cluster area of Shandong. The underlying causes for this geographic disparity need to be further investigated.

Esophageal cancer (EC) is an important public health issue in China. It is the fourth leading cause of cancer mortality, causing 197,800 deaths in China in 2013, which accounted for nearly half of the total EC deaths worldwide 1,2 . The age-standardized mortality rate (ASMR) ranges almost ten-fold (4.6 to 42.5 per 100,000 population) across different provinces in China 1 . There are also large intra-provincial variations in EC mortality. For example, within the Shandong Province of China, the EC mortality in the eastern region of the province was lower than in the central and western regions, with a consistently high-mortality risk cluster identified in the mid-west region (Fig. 1). Residents living in the cluster area were nearly four times more likely to die from EC than residents in the rest of Shandong (RR: 3.7, 95% CI: 2.8-5.0) 3 .
Studies have shown that lower socioeconomic status (SES) is associated with increased incidence or mortality due to EC [4][5][6][7][8][9][10] . In the Chinese population, the majority of the studies have been cohort or cross-sectional studies focused on an individual's SES [8][9][10] . Population-based studies considering area-level effects are scarce in China. Hence, they are not able to indicate whether, and to what extent, the geographic disparities of EC mortality are due to regional resources.
In Shandong, the economy in the eastern region is more developed compared to the central and western regions 11 . Given the consistency of this economic distribution with the above-mentioned patterns in EC mortality in Shandong, we hypothesize that the area SES is associated with the geographic disparity in EC morality in Shandong population.
In this study, we build on and extend previous studies by using multilevel analysis to simultaneously account for variation in EC mortality across areas and between individuals in Shandong population, accounting for 7.2% of the Chinese population 12 . The analysis aims to elucidate potential associations between area-level SES factors and EC mortality among the Shandong population, thus furthering the understanding of how place matters to person-years. ASMR was higher in rural areas than urban areas (39.3 vs 29.4; p = 0.001); and the highest ASMRs were seen in counties with low SES index (p < 0.001) and low average years of school education (p < 0.001) ( Table 1). A gradient of increasing ASMR was observed as the SES index decreased, or residential type changed from urban to rural (Fig. 2).
In the cluster area, a total of 10,249 deaths were reported during 2011-2013 out of 10,362,408 person-years; In the rest of Shandong, 35,397 deaths were reported out of 125,390,073 person-years. The crude risk ratio for EC mortality between cluster and non-cluster area is 3.39 (95% CI: 2.25-5.11; p < 0.001). After adjusting for age and sex, the risk ratio increased to 3.84 (95% CI: 2.59-5.70; p < 0.001).
Similar to the whole Shandong population, the ASMR increased with age and was higher in males in both cluster (p < 0.001) and non-cluster areas (p < 0.001) ( Table 2). Also, in both areas, rural residents were more likely to die from EC than urban residents (p = 0.026 in cluster areas, p = 0.023 in non-cluster areas); and the ASMR increased with the decrease of county-level SES index in cluster area (p < 0.001) and the trend is close to being significant in non-cluster areas (p = 0.055).
Multilevel regression. The multilevel model revealed the association between area SES and EC mortality in the Shandong population (Table 3, Model 1). After adjustment for age and sex, residents living in rural areas had 22% (95% CI: 13-32%; p < 0.001) higher risk of dying from EC. With each point increase of SES index in www.nature.com/scientificreports www.nature.com/scientificreports/ EC death (N) Person-years (N) ASMR * (95% CI) p §
The interaction between residential type and SES index on EC mortality was not statistically significant (RR: 1.01; 95% CI: 0.95-1.06; p = 0.84), supporting the hypothesis that the influence of residential type and SES index with EC mortality were independent of each other.
Model 2 elucidated to what extent the area-level SES variables explain the identified high-mortality cluster ( Table 3, Model 2). After adjustment for age and sex, the risk ratio of EC mortality between cluster and non-cluster area was 3.84 (95% CI: 2.59-5.70; p < 0.001). After additional adjustment for area-level SES, this reduced slightly to 3.80 (95% CI: 2.56-5.65; p < 0.001).
Interactions between the cluster variable and residential type (RR: 1.11; 95% CI: 0.85-1.44; p = 0.46), and cluster variable and SES index (RR: 0.72; 95% CI: 0.49-1.08; p = 0.11) were not significant, suggesting that the relationship between area-level SES and EC mortality did not differ between cluster and non-cluster areas.
The differences between observed EC mortality and predicted EC mortality (using Model 2) of the 142 counties have been visualized on the Shandong map (Fig. 3). The biggest differences were seen in two counties (Dongping and Ningyang) located in the cluster area. The predicted EC mortality of these two counties (184.0 per 100,000 person-year in Dongping and 155.6 in Ningyang) based on Model 2 is higher than the observed mortality (175.7 and 148.5, respectively). The urban to rural population ratio was 0.53 in Dongping and 0.14 in Ningyang compared to 0.55 in Shandong. The SES index were 1.2 and 1.5, respectively in these two counties compared to the median SES index of 2.1 for the 142 counties.

Discussion
This multilevel study shows that EC mortality is not only associated with an individual's characteristics, such as age and gender, but also depends on the SES of where people live.
People living in rural areas had a 22% higher risk of dying from EC compared with the urban residents. This pattern was independent from age, sex, and the SES index. There are several possible reasons for the observed association between residential type and EC mortality. First, the prevalence of modifiable risk factors of EC might be higher among rural residents. For example, a recent study reported that people in rural areas of China are more likely to smoke and drink than their urban counterparts 14 . The onset of EC can be associated with Human Papillomavirus, which are more prevalent in rural than urban populations 15,16 . Since the survival rate of EC is relatively low 17 , higher incidence in rural areas is likely to translate to a higher mortality.
Second, disparities in health services between rural and urban areas may result in delayed diagnosis in rural patients which leads to poor survival. In China, the majority of the tertiary referral hospitals are in urban areas. The different health insurance system between urban and rural residents leads to longer wait times for medical investigation and referral for diagnosis for rural patients 18  www.nature.com/scientificreports www.nature.com/scientificreports/ patients from rural areas were significantly more likely than urban patients to present with late-stage diseases 20 . EC patients from rural areas might have the same situation.
Furthermore, financial disadvantage may also lead to lower survival among rural patients and increase mortality. Rural residents on average have less income than urban residents 21 . Because of the different health insurance system, rural patients may need to pay more out-of-pocket expenses for treating their illness even when admitted to the same hospitals 18 . This might lead to delayed or suboptimal treatment among the rural cancer patients 20 .
We developed a summary measure of county-level SES (SES index) which incorporates GDP per capita, average years of school education, and number of hospital bed per capita. The results illustrate that average EC mortality of a county is associated with the county's SES index. The results reflect the results from a Chinese national study, which found the GDP per capita at a county level was inversely associated with EC incidence and mortality 22 .
GDP per capita is a useful measurement when comparing differences in living standards between different regions. It also reflects inequalities in EC risk factors such as smoking, and alcohol use 23 . In addition, local governments in affluent areas with a high GDP level are more likely to allocate a larger proportion of the total expenditure on health care, which may lead to better access to health resources 24 . Our study also found a significant  Table 3. Modelling the association between residential type and cancer mortality. *SES index is continuous variable here to give the best model fit. www.nature.com/scientificreports www.nature.com/scientificreports/ correlation between a county's GDP per capita and number of hospital bed per capita. As a result, a better prognosis and a lower EC mortality could be achieved in areas with higher GDP per capita.
Residents with lower education level were related with suboptimal nutrition and higher prevalence of smoking in the Chinese population 8,25 , which could contribute to a high EC incidence and mortality. A study about breast cancer reported that, in China, lower education level was related to later cancer staging at diagnosis and lower success rates of cancer treatment, therefore leads to worse survival 26 . Our study adds to the existing knowledge by showing a direct link between poor education and increased EC mortality at a population level.
In a previous study we identified a cluster with higher EC risk in Shandong Province. The cluster is located in the mid-west region of Shandong, which has been reported as a less developed region 11 . Our initial hypothesis was the geographic disparity of EC mortality would be associated with the inequitable area-level SES. However, the risk ratio of area-level SES variables was not influenced by including the cluster variable (Model 1 vs Model 2). This suggests these area-level SES factors are highly unlikely to have anything to do with the large difference in EC mortality between the high-and low-risk regions in Shadong, despite their strong associations with the overall mortality.
The biggest differences between observed mortaltiy and predicted mortality were seen in two counties located in the cluster area (Fig. 3). The residential type proportion and SES index of these two counties were not different from other counties in Shandong, suggesting that our current model, which adjusted for age, sex, area-level SES factors, and cluster variable, could not fully explain the EC mortality in these two counties. Interestingly, both of these counties are in the downstream catchment area of the Dawen River. Future investigation especially regarding water influences are needed to elucidate the factors underlying this high EC mortality area in China.
This study has several limitations. First, our analyses were based on the total EC mortality with no information on histological types. EC mainly consists of two different histological types (squamous cell carcinoma (SCC) and adenocarcinoma) with different epidemiologic features. The main risk factors of SCC are alcohol and tobacco use 27 , which are more prevalent in socio-economic disadvantaged areas in China. In contrast, the main risk factors of adenocarcinoma include obesity 27 , which is more common in advantaged areas 28 . Thus, the relationship between EC mortality and SES would likely be different between those two histologic types. Research data have shown that more than 90% of EC cases in China are SCC 29 . Therefore, the small numbers of adenocarcinoma cases might not significantly influence the association between SES and SCC cases in this study. In addition, this study did not include many individual-level characteristics. As the income inequality is relatively high in China 30 , an individual's SES may be different from the county-level SES index. Also, the SES index could be different between urban and rural sub-county units. Adjusting for individual SES variables such as occupation and income in the model or including sub-county-level SES index would help address this limitation.

Conclusion
In this multilevel study of EC mortality with a large Chinese population, we found that area-level SES factors were independently associated with EC mortality. However, these factors seem to be independent of the large difference in EC mortality between the low-and high-risk regions. The underlying causes for the observed geographic disparity need to be further investigated to inform public health policies and programs to lower the high disease burden and to address inequalities.

Methods study site and units of analysis.
Shandong Province is a coastal province in North-East China, covering an area of 157,100 square kilometres. It is the second most populous province in China, with about 99% of its 98.5 million people belonging to the Han ethnic group 31 .
Variation in EC mortality was analyzed according to the hierarchical structure where the numbers of death by age and sex (individual-level) were nested within sub-county-level units (sub-county-level) within county-level units (county-level). For this study, we divided the total Shandong population (without gaps or overlaps 32 ) into 142 county-level units according to gazetted administrative boundaries. Counties were further subdivided into 262 sub-county-level units according to residential type.
Cancer data sources. Deaths caused by EC (ICD-10: C15, malignant neoplasm of esophagus) in the population aged 40+ years residing in Shandong Province between 1 st January 2011 and 31 st December 2013 were extracted from the Shandong Death Registration System (SDRS) database. EC-specific death information was collected based on a national standard protocol 33 and were double checked by physicians and their direct managers. More details about the quality of the data has been discussed elsewhere 3,11,32 . The extracted de-identified EC death dataset represents the number of EC deaths in each of the 262 sub-county-level units in Shandong by sex, and 5-year age group. The age groups above 85 years were combined into an 85+ age group.
The population for each sub-county-level unit by the same age-sex categories were obtained from Shandong Centre for Disease Control and Prevention (Shandong CDC), who generated their estimates based on data from local police departments and statistical bureaus 31 .

Measurement of sub-county-level SES.
Residential type is the sub-county-level SES variable and consists of two levels: urban and rural. Sub-county-level units consisting of townships are classified as rural areas. A township is an administrative unit consisting of a town center and dozens of surrounding villages. Sub-county-level units made up of subdistricts are classified as urban areas. A subdistrict consists of urban communities or neighborhoods clustering around the center of a county/city/district. More information about the residential type classification in Shandong has been published previously 32 . (2019) 9:6388 | https://doi.org/10.1038/s41598-019-42774-x www.nature.com/scientificreports www.nature.com/scientificreports/ Residential type has been reported as a strong indicator for an individual's income and socioeconomic class in China, as rural residents on average have less household income per head, less mean years of education, less bank savings, and a higher percentage of blue collar workers than urban residents 21,34,35 . Measurement of county-level SES. The county-level (n = 142) SES variables considered were average gross domestic product (GDP) per capita, average years of school education for adults, and number of hospital beds per capita. The 2011-2013 data for these measurement were extracted from the Shandong Statistical Yearbooks 31 .
Pearson's correlation suggested significant correlations between the three county-level SES variables, with the coefficient between GDP and education being 0.5 (p < 0.001); GDP and number of hospital beds being 0.3 (p < 0.001); and number of hospital beds and education years being 0.7 (p < 0.001). Therefore, rather than treating these measures separately, principal component analysis (PCA) was used to generate an SES index. The benefits and applications of constructing an SES index with PCA has been published before [36][37][38] . In this study, components with an eigenvalue of 1.0 or higher was chosen as SES index.
Three components were created from the county-level SES variables. Only one of the components had an eigenvalue more than 1 (eigenvalue = 2.00), explaining 66.7% of variation in the original SES variables. This component was transformed to scale from 0 to 10 using the min-max normalization method:  Tables 1 and 2, the continuous SES index was categorized into three groups based on its value (low SES index from 0.0 to 3.3, middle from 3.4 to 6.7, and high from 6.8 to 10.0). In multilevel regression model (Table 3), we used the continuous SES index variable directly.
Multilevel regression. Multilevel mixed effect negative binomial models were used to examine the association between area-level SES (i.e. residential type and SES index) and EC mortality (Model 1), and to estimate how area-level SES factors contribute to the high-mortality cluster in Shandong (Model 2).
Specifically, Model 1 concurrently included age, sex, residential type, and SES index to estimate the independent contribution of the two area-level SES factors on EC mortality. Interaction between residential type and SES index was also considered to ascertain if the relationship between county-level SES index and EC mortality differed by residential type.
Model 2 added a cluster variable (cluster/non-cluster) to Model 1, to elucidate the contribution of area SES factors on the mortality ratio between cluster and non-cluster areas. Interactions between the cluster variable and residential type, cluster variable and SES index were considered to clarify if the contribution of area SES on EC mortality differed between cluster and non-cluster areas.
Data were analyzed and visualized using R software (Version 3.4.3) and Stata (version 15.0). Parameter estimates are presented as risk ratios (RR) with their 95% confidence intervals (CI).

Data Availability
The data that support the findings of this study are available from Shandong Center for Disease Control and Prevention, but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Shandong Center for Disease Control and Prevention.