Spatial risk for gender-specific adult mortality in an area of southern China
© Ali et al; licensee BioMed Central Ltd. 2007
Received: 08 June 2007
Accepted: 24 July 2007
Published: 24 July 2007
Although economic reforms have brought significant benefits, including improved health care to many Chinese people, accessibility to improved care has not been distributed evenly throughout Chinese society. Also, the effects of the uneven distribution of improved healthcare are not clearly understood. Evidence suggests that mortality is an indicator for evaluating accessibility to improved health care services. We constructed spatially smoothed risk maps for gender-specific adult mortality in an area of southern China comprising both urban and rural areas and identified ecological factors of gender-specific mortality across societies.
The study analyzed the data of the Hechi Prefecture in southern in China. An average of 124,204 people lived in the area during the study period (2002–2004). Individual level data for 2002–2004 were grouped using identical rectangular cells (regular lattice) of 0.25 km2. Poisson regression was fitted to the group level data to identify gender-specific ecological factors of adult (ages 15–<45 years) mortality. Adult male mortality was more than two-fold higher than adult female mortality. Adults were likely to die of injury, poisoning, or trauma. Significantly more deaths were observed in poor areas than in areas with higher incomes. Specifically, higher spatial risk for adult male mortality was clustered in two rural study areas, which did not overlap with neighborhoods with higher risk for adult female mortality. One high-risk neighborhood for adult female mortality was in a poor urban area.
We found a disparity in mortality rates between rural and urban areas in the study area in southern China, especially for adult men. There were also differences in mortality rates between poorer and wealthy populations in both rural and urban areas, which may in part reflect differences in health care quality. Spatial influences upon adult male versus adult female mortality difference underscore the need for more research on gender-related influences on adult mortality in China.
Although economic reforms have brought significant benefits, including improved health care to many Chinese people, the quality of care remains uneven. Higher quality health care is available mainly in cities . There are also signs that for China's women, particularly in the countryside, the reform era has been associated with declining access to quality health care services, suggesting that inequalities in improved health care occur along gender and geographic boundaries . The Chinese government has articulated its commitment to closing the gaps to accessing improved health care services. However, there is no clear understanding of the effects of health care quality by gender or socioeconomic status. It is, therefore, important to identify the variations in accessing improved health care services among different levels of society in order to understand the causes and the magnitude of problem. Some studies suggest that mortality is an important indicator for evaluating variations in accessibility to quality health care services [2, 3].
Twigg et al.  postulated that human health-related behavioral practices are influenced by others within a society. Thus, addressing the impact of social ecology on health is important but greatly limited by available analytical tools. The lack of an effective geocomputational environment and algorithms has hindered the development of spatial analysis techniques  leading to unrealistic assumptions about human behavior. Related analytical problems include the uneven distribution of physical facilities, differences in speed of movement in different areas, and the effect of communications networks. To address these issues, geographical epidemiologists are increasingly using more complex methods of statistical analysis to investigate the spatial distribution of diseases [6, 7].
Disease maps (spatial distribution of disease) provide researchers with visual displays that can suggest, via patterns of physical facilities and the human environment, useful avenues of research into causal processes [8, 9]. However, the use of simple relative risk assessment of a disease (e.g., number of observed cases divided by number of expected cases for each area) may create problems in areas with small populations (usually rural areas), yielding extreme relative risks as the number of expected cases in the denominator is low. Methods for obtaining stable and accurate estimated rates in small populations are critical for effective analysis. Bayesian hierarchical modeling approach that uses Markov chain Monte Carlo (MCMC) procedure deals with complicated data structures and models and has good properties for a broad range of true underlying parameter arrangements .
Here we describe the use of neighborhood level data with identical rectangular 0.25-km2 cells (regular lattice) in a Bayesian hierarchical modeling approach. We used an MCMC computational method to obtain the joint posterior distribution of model parameters, from which we constructed smoothed risk maps of gender-specific adult mortality in an area of southern China with both urban and rural areas. Significant gender imbalance in mortality rates (male mortality more than double that of females of similar ages) induced us to conduct this gender-specific mortality study.
Population and deaths by gender and age groups, Hechi Prefecture, 2002–2004, Guangxi, China.
Age group, years
Relative risk male vs. female (95% CI)
0.36 (0.09, 1.38)
1.41 (0.40, 4.99)
2.56 (1.94, 3.39)
1.56 (1.40, 1.74)
1.62 (1.46, 1.79)
Overall mortality rates by year and areas (urban vs. rural), Hechi Prefecture, Guangxi, China.
Relative risk (rural vs. urban) (95% CI)
Causes of death for men and women aged 15–<45 years, Hechi Prefecture, Guangxi, China, 2002–2004.
Cause of death
# of deaths
% of total deaths
# of deaths
% of total deaths
Diseases of blood and blood-forming organs and certain immune disorders
Diseases of the nervous system
Diseases of the genitourinary system
Pregnancy, childbirth, puerperium
Mental and behavioral disorders
Diseases of the ear and mastoid process
Diseases of the musculoskeletal system and connective tissue
Certain infectious and parasitic diseases
Endocrine, nutritional and metabolic diseases
Diseases of the respiratory system
Diseases of the circulatory system
Diseases of the digestive system
Abnormal symptoms, signs, clinical and laboratory findings
External causes such as injuries, toxicosis, trauma
Aspatial regression model
Study variables for neighborhoods (0.25-km2 grid cells) for men and women aged 15 to <45 years.
Men (n = 243 neighborhoods)
Women (n = 236 neighborhoods)
Per capita monthly income
Hospital/health facility distance (km)
Distance from river (km)
Per capita health care* expenditure in last month of 2001 census†
Mortality for men aged 15 to <45 years by multiple Poisson regression, 2002–2004, Hechi Prefecture, Guangxi, China.
Wald 95% Confidence Limits
Pr > χ2
Per capita monthly neighborhood income
Neighborhood population density//km2
Hospital/health facility distance (km)
Distance from river (km)
Per capita neighborhood health care expenditure in the last month of 2001 census
Mortality for women aged 15 to <45 years by multiple Poisson regression, 2002–2004, Hechi Prefecture, Guangxi, China.
Wald 95% Confidence Limits
Pr > χ2
Per capita monthly neighborhood income
Neighborhood population density//km2
Hospital/health facility distance (km)
Distance from river (km)
Per capita neighborhood health care expenditure in last month of the 2001 census
Parameter estimates by a multiple-membership multiple classification model for men and women aged 15 to <45 years.
Per capita monthly neighborhood income (in RMB)
Level 2 residual variance
Level 3 residual variance
421.204 (243 of 243 cases in use)
229.999 (236 of 236 cases in use)
In our study, adult mortality was significantly higher for males than that for females. Adults were likely to die of injury, poisoning, or trauma. Overall mortality was higher in households headed by women. The disparity in the mortality rates between adult men and women suggests that if this trend continues, there will be more households with women heads, which could eventually increase the number of deaths in the study area. Our results also illustrate disparity in mortality rates between rural and urban areas, which may in part be due to disparities in health care accessibility.
In our aspatial analysis, the risk for adult mortality was higher in impoverished communities (defined by lower per capita neighborhood monthly income) than in wealthier communities, suggesting that the benefits of improved health care were not evenly distributed throughout the study area. The collapse of China's Cooperative Medical System in 1978 resulted in the lack of an organized financing scheme for health care, adversely affecting access by rural people to health care, especially the poor . This could perhaps account for the high number of "at risk" neighborhoods in the impoverished part of our rural study area. The poor urban neighborhood at high risk for female mortality suggests that even in areas where health care is easier to access because it is closer by, gender influences who receives health care services in poor societies.
The adult mortality maps (Figures 1 and 2) show many neighborhoods with increased risks of mortality (RR, >1.5). These neighborhoods formed two regional clusters, both in the rural area. In contrast, not many neighborhoods carried increased risk for adult female mortality. These maps may suggest difference in adverse effects of health hazards across subpopulations . Because none of the high-risk neighborhoods for male and female mortality were superimposed, there is a possibility that adult men and women in southern China face different ecological and environmental risks. Future studies should analyze risk factors in greater depth.
We found significantly higher mortality in rural than in urban areas, possibly because of less health care accessibility . Also, health services available in rural areas may not provide adequate treatment for potentially curable diseases. Chinese policymakers are currently trying to narrow the disparity in health care services for rural and urban residents through a five-part reform policy. Specific targets are directed at (1) increasing public funding for primary health care; (2) providing quality and accessible health care; (3) extending coverage of social health insurance schemes; (4) providing government health subsidies to vulnerable portions of the population; and (5) adopting more appropriate health technologies and pharmaceuticals in health care delivery.
A potential limitation of our study is the arbitrary choice of neighborhoods and therefore variations in the size of population across neighborhoods. We believe our selection is an appropriate compromise between loss of resolution and excess dispersion. One may argue that the mass vaccine campaign might have influenced neighborhood level variations in mortality rates. However, the trial was cluster randomized and we assumed that ecological determinants were independent of the cluster effects under the trial design for vaccine assignment. Moreover, the rates of infections targeted by the vaccines were too low to have affected overall mortality. Another limitation of our study design is that our study area has an urban area with a high population density and a sparsely populated rural area. Thus, the population varied greatly across neighborhoods. However, since we incorporated the neighboring area effect into the model, we believe the model adequately addressed heterogeneity of across-neighborhood population. Because we assigned equal weights for neighborhood components without knowing the spatial influences of mortalities and/or anisotropy on the surface, complexities may have been diminished.
Although both the models (MMMC and CAR) can be used to account for the effects of locations, there are differences between the two models. In MMMC model, we consider two sets of random effects: exchangeable area random effects and a multiple membership set of random effects for the neighbors of each neighborhood. This mean the rates in each neighborhood is affected by both the neighborhood and its nearby neighbors. The weight columns contain equal weights for each neighboring neighborhood that sum to 1. In contrast, the CAR prior is a spatial smoothing prior, and individual random effect is not random in the CAR model. This model has only one random effect for each neighborhood, and it is expected value the average of surrounding random effect. Note to make the CAR model identifiable we either need to constrain random effect to sum to 0 or remove the intercept from the model. CAR procedure typically uses weight of 1 for all observations, as these weights will then be divided by the number of neighbors in the model.
We focused on the production of reliable maps for gender-specific adult mortality in an area of southern China. By using the Bayesian hierarchical model, the neighborhood random effects were posterior sampled and the associated relative risk estimates were averaged to produce a posterior-average relative risk , which was then used to produce the gender-specific mortality maps. The approach produces stable and accurate estimates so that data modeling with this approach implies greater reliability in identifying areas at greatest risk for mortality and the underlying reasons.
Study area and data
The study was conducted in the Hechi Prefecture of Guangxi Zhuang Autonomous Region (Guangxi Province) in southern China, which borders Guangdong Province in the east and Vietnam in the south. Hechi Prefecture is in the northwest of the province, approximately 400 km from the provincial capital, Nanning. The catchment area includes two populous areas: Jin Cheng Jiang, an urban area (26 km2), and Don Jiang, a rural area (191 km2). The site was originally set up by the International Vaccine Institute (IVI) in collaboration with the Guangxi Centers for Disease Control and Prevention for a multi-centric Vi polysaccharide vaccine effectiveness evaluation for typhoid fever [13, 14]. All residents of Jin Cheng Jiang and Dong Jiang were enumerated in late 2001 by the vaccine trial project staff and were followed in subsequent years for medical and vital demographic events including deaths. According to the project's census, an average of 124,204 people lived in the area, half male. Beginning in mid-2002 all deaths in the community were recorded in yearly census surveys conducted by project staff.
The project's mortality surveillance team collected information from sentinel posts (i.e., hospitals, funeral houses, family planning offices and local government, village, and police registrars). A modified (shorter) verbal autopsy (based on procedures developed by the UK Department for International Development, 1997; the World Health Organization; The Johns Hopkins School of Hygiene and Public Health; and The London School of Hygiene and Tropical Medicine, 1999) was used by a trained physician to determine cause of death. The verbal autopsy form containing data on the cause of death (using locally adapted classifications of diseases) was entered into the project database. The verbal autopsy was conducted only for the vaccine trial target population (ages 5–60 years).
We created a household geographic information system (GIS) in the study area using handheld global positioning system receivers. Among the geographic features included in the GIS were Long Jiang River and its branches, roads, and mountains plus the 35 hospitals/health facility from which disease surveillance was carried out. By linking demographic and mortality data, the GIS allowed us to pursue spatial modeling of adult mortality in the study area.
We divided the study area into grid cells of 500 × 500 m, which we called neighborhoods. Grid cells less than 0.25 km2 resulted in overdispersion of the mortality data (that is greater variation in the data than expected while larger cells obscured details of the ecological status) . We removed parts of cells that fell outside the study area boundary. In total, we obtained 267 neighborhoods with at least one person living in the neighborhood. We subsequently excluded neighborhoods with fewer than four persons or no neighboring cell with at least four persons because neighborhood level ecological data derived from few observations could bias the outcome. Ultimately we obtained 243 neighborhoods with at least four adult males (ages 15–<45 years) in each and 236 neighborhoods with at least four adult females (ages 15–<45 years). Mortality data for individuals and several socioeconomic covariates were aggregated by neighborhood. Linear distances to the nearest hospital/health facility and to the nearest river side were computed from the neighborhood center.
Poisson regression analysis
a generalized linear model with log link function and Poisson distributed errors where E(Y) is the expectation of observation, log(exp) is the logarithm of expected number of cases, x1 is monthly per capita neighborhood income, x2 is population density in the neighborhood, x3 is hospital/health facility distance, x4 is distance from the river, and x5 is neighborhood per capita expenditure on health care in a month (considered the last month of the census 1 survey conducted in late 2001). βi is the coefficient corresponding to x i . The term log(exp) was an offset with the parameter estimate constrained to 1 since we were interested in (relative) rates rather than counts. Neighborhood income is a surrogate for the neighborhood economic status, population density and distance from river are the surrogates of environmental differences among neighborhoods, distance to hospital/health facility describes access to health care, and health care expenditure is a surrogate for health care utilization. We used Stata/SE 9.0 for Windows (StataCorp LP, College Station, TX 77845 USA) to analyze the data using Poisson regression.
Univariate Moran scatter plot
To assess spatial patterns we analyzed the residuals of the Poisson multiple regression model using univariate Moran scatter plot (GeoDA™ software; Luc Anselin and the Regents of the University of Illinois). The spatial weight was determined using first order Queen Contiguity (i.e., all common points including boundaries and vertices were included in the neighbor definition). The method produces four quadrants within the graph that provides a classification of two types of positive spatial autocorrelation: high-high (upper right), low-low (lower left); and two types for negative spatial autocorrelation: high-low (lower right) and low-high (upper right). Inference for Moran's I was based on a permutation approach, in which a reference distribution is calculated for spatially random layouts with the same data as observed. The randomization uses an algorithm to generate spatially random simulated data sets as outlined by Anselin . We used 9999 random permutations in constructing the reference distribution.
The standard hierarchical model structure does not explicitly incorporate spatial structure, although through the use of higher levels of geography as additional levels in the model, we can indirectly incorporate spatial clustering effects. One extension to the standard hierarchical model is the multiple-membership model that addresses the effect of spatial correlation between neighboring areas . Browne et al.  consider Bayesian extensions of this model as a member of the family models they call a multiple-membership multiple classification (MMMC) model. Alternatively, we can use a conditional autoregressive (CAR) model that also assesses spatial correlation between neighboring areas. We used MLwiN Version 2.0 (©Multilevel Models Project, Center for Multilevel Modelling, University of Bristol, UK) to fit MMMC models; for CAR models we used through likelihood-based estimation methods and MCMC estimations.
Here y is an N (number of lowest level unit) vector, β is a vector of fixed effect parameters, and , are the vectors of residuals for the random effects for classifications 2 and 3, respectively. and are the vectors of predictor values and is a scalar weight for the classification 3 unit j for lowest level unit i.
To fit the data into the model, we employed neighbors of the neighborhoods in level 3, and the neighborhood (group) in levels 1 and 2. We included a covariate "per capita neighborhood income," which had a significant relationship with mortality in the aspatial regression model, in the fixed part of the model (Xβ), in addition to the intercept term (CONS). The use of a covariate in a Bayesian spatial model is important for investigating environmental variations . The CONS was a vector of 1s, which allowed for a variance component for each neighborhood to be estimated. The weight was employed based on the first-order neighborhood using Queen Contiguity (i.e., both boundaries and vertices are included in the definition). Hence we had eight neighbor columns and eight weight columns. For modeling, we fitted the variance component using a burn-in period of 500 and a chain of 50,000.
Spatially smoothed relative risk of mortality
To evaluate the status of each neighborhood with respect to adult mortality, we obtained spatially smoothed relative risks of adult mortality within neighborhoods by means of the Bayesian approach described above. In general, the relative risk in disease/mortality mapping measures whether an area has a higher occurrence of disease incidence/mortality than that expected from the reference rate. In Bayesian disease/mortality models, the relative risk decomposes into two parts that are fixed terms consisting of overall level of relative risk and due to covariates and random terms. The random terms are spatial correlation structure, which introduces estimates of the risk in any area depending on neighboring areas, and uncorrelated heterogeneity, which pertains to the random sampling variability of the observed counts about the local mean.
Financial support was provided by the Bill and Melinda Gates Foundation through the Diseases of Most Impoverished (DOMI) Program administered by the International Vaccine Institute (IVI), Seoul, Korea. Current donors providing unrestricted support to IVI include Republic of Korea, Swedish International Development Cooperation Agency, and Kuwait. We are grateful for the contributions of other IVI staff members and the staff members of the Guangxi Centers for Disease Control and Prevention, Guangxi, China.
- Smith CJ: Modernization and health care in contemporary China. Health Place. 1998, 4: 125-139. 10.1016/S1353-8292(98)00005-7.PubMedView ArticleGoogle Scholar
- Rosero-Bixby L: Socioeconomic development, health interventions and mortality decline in Costa Rica. Scand J Soc Med Suppl. 1991, 46: 33-42.PubMedGoogle Scholar
- Timæus IM: Commentary: Mortality decline in Senegal. Int J Epidemiol. 2001, 30: 1294-1295. 10.1093/ije/30.6.1294.View ArticleGoogle Scholar
- Twigg L, Moon G, Jones K: Predicting small-area health-related behaviour: a comparison of smoking and drinking indicators. Social Sci Med. 2000, 50: 1109-1120. 10.1016/S0277-9536(99)00359-7.View ArticleGoogle Scholar
- Kwan M: Analysis of human behavior in a GIS environment: recent development and future prospects. J Geograph Syst. 2000, 2: 82-90. 10.1007/s101090050034.View ArticleGoogle Scholar
- Elliot P, Cuzick J, English D: Geographical and Environmental Epidemiology: Methods forSmallAarea Studies. 1992, New York: Oxford University PressGoogle Scholar
- Elliot P, Martuzzi M, Shadick G: Spatial statistical methods in environmental epidemiology: a critique. Stat Methods Med Res. 1995, 4: 137-159.View ArticleGoogle Scholar
- Langford I, Leyland A, Rashbas J, Goldstein H: Multilevel modelling of the geographic distribution of diseases. J R Stat Soc Ser C Appl Stat. 1999, 48: 253-268. 10.1111/1467-9876.00153.PubMedView ArticleGoogle Scholar
- Xia H, Carlin BP, Waller LA: Hierarchical models for mapping Ohio lung cancer rates. Environmetrics. 1997, 8: 107-120. 10.1002/(SICI)1099-095X(199703)8:2<107::AID-ENV241>3.0.CO;2-E.View ArticleGoogle Scholar
- Spiegelhalter DJ, Best NG, Carlin BP, van der Linde A: Bayesian deviance, the effective number of parameters and the comparison of arbitrarily complex model. J R Stat Soc Ser B. 2002, 64: 583-640. 10.1111/1467-9868.00353.View ArticleGoogle Scholar
- Wang H, Yip W, Zhang L, Wang L, Hsiao W: Community-based health insurance in poor rural China: the distribution of net benefits. Health Policy Plan. 2005, 20: 366-374. 10.1093/heapol/czi045.PubMedView ArticleGoogle Scholar
- Lawson AB: Disease map reconstruction. Stat Med. 2001, 20: 2183-2204. 10.1002/sim.933.PubMedView ArticleGoogle Scholar
- Acosta CJ, Galindo CM, Ali M, Elyazeed RA, Ochiai RL, Danovaro-Holliday MC, Page AL, Thiem VD, Jin Y, Park JK, Lee H, Puri MK, Ivanoff B, Agtini MD, Soeharno R, Simanjuntak CH, Punjabi N, Canh DG, Sur D, Nizami Q, Manna B, Bai-qing D, Anh DD, Honghui Y, Bhattacharya SK, Bhutta Z, Trach DD, Xu ZY, Pang T, Donner A, Clemens JD: A multi-country cluster randomized controlled effectiveness evaluation to accelerate the introduction of Vi polysaccharide typhoid vaccine in developing countries in Asia: rationale and design. Trop Med Int Health. 2005, 10: 1219-1228.PubMedView ArticleGoogle Scholar
- Yang J, Acosta CJ, Si G, Zeng J, Li C-Y, Liang D-B, Ochiai RL, Page A-L, Danovaro-Holliday MC, Zhang J, Zhou B-D, Liao H-Z, Wang M-L, Tan D-M, Tang Z-Z, Gong J, Park JK, Ali M, Ivanoff B, Liang G-C, Yang H-H, Tikki P, Xu Z-Y, Donner A, Galindo CM, Dong B-Q, Clemens JD: A mass vaccination campaign targeting adults and children to prevent typhoid fever in Hechi; expanding the use of Vi polysaccharide vaccine in Southeast China: cluster-randomized trial. BMC Public Health. 2005, 5: 49-10.1186/1471-2458-5-49.PubMedPubMed CentralView ArticleGoogle Scholar
- Talbot TO, Kulldroff M, Teven PF, Haley VB: Evaluation of spatial filters to create smoothed maps of health data. Stat Med. 2000, 18: 2399-2408. 10.1002/1097-0258(20000915/30)19:17/18<2399::AID-SIM577>3.0.CO;2-R.View ArticleGoogle Scholar
- Anselin L: MicroQAP: a microcomputer implementation of generalized measures of spatial association. Working paper. 1986, University of California, Santa Barbara, Department of GeographyGoogle Scholar
- Lawson AB, Browne WJ, Rodeiro CLV: Disease mapping with WinBUGS and MLwiN. 2003, Chichester, UK: John Wiley & Sons;View ArticleGoogle Scholar
- Browne W, Goldstein H, Rasbash J: Multiple-membership multiple classification (MMMC) models. Statistical Modelling. 2001, 1: 103-124. 10.1191/147108201128113.View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.