Geographical Inequalities and Social and Environmental Risk Factors for Under-Five Mortality in Ghana in 2000 and 2010: Bayesian Spatial Analysis of Census Data

Background: Under-five mortality is declining in Ghana and many other countries. Very few studies have measured under-five mortality—and its social and environmental risk factors—at fine spatial resolutions, which is relevant for policy purposes. Our aim was to estimate under-five mortality and its social and environmental risk factors at the district level in Ghana. Methods and Findings: We used 10% random samples of Ghana’s 2000 and 2010 National Population and Housing Censuses. We applied indirect demographic methods and a Bayesian spatial model to the information on total number of children ever born and children surviving to estimate under-five mortality (probability of dying by 5 y of age, 5q0) for each of Ghana’s 110 districts. We also used the census data to estimate the distributions of households or persons in each district in terms of fuel used for cooking, sanitation facility, drinking water source, and parental education. Median district 5q0 declined from 99 deaths per 1,000 live births in 2000 to 70 in 2010. The decline ranged from <5% in some northern districts, where 5q0 had been higher in 2000, to >40% in southern districts, where it had been lower in 2000, exacerbating existing inequalities. Primary education increased in men and women, and more households had access to improved water and sanitation and cleaner cooking fuels. Higher use of liquefied petroleum gas for cooking was associated with lower 5q0 in multivariate analysis. Conclusions: Under-five mortality has declined in all of Ghana’s districts, but the cross-district inequality in mortality has increased. There is a need for additional data, including on healthcare, and additional environmental and socioeconomic measurements, to understand the reasons for the variations in mortality levels and trends. We used 10% random samples of Ghana’s 2000 and 2010 National Population and Housing Censuses. Using a set of predefined questions, both censuses gathered information on a number of individual- and household-level variables related to socioeconomic factors (e.g., literacy and educational attainment for persons 11 y or older), living environment (e.g., household’s main water supply source, sanitation facility type, and cooking fuel type, as a commonly used metric for household air pollution [4,22]), and children’s births and deaths for females 12 y or older. Each record in the census data had information on the census enumeration area (EA), the smallest geographical unit, with an average population of 750. The 2000 and 2010 censuses had nearly 26,000 and 38,000 EAs, respectively. Our analysis was conducted at the district level, the country’s second smallest level of subnational administrative divisions, with EAs mapped to the district of residence. We used districts as units of analysis because they are administrative units for resource allocation and for program implementation. Further, EAs were defined separately in each census and could not be mapped from one census to the other. There were 110 and 170 districts in the 2000 and 2010 censuses, respectively. We merged the 2010 districts, linking them to their original districts that had split since 2000, to create 110 common districts for our analyses. The 110 districts are administratively assembled into ten regions: Ashanti, Brong-Ahafo, Greater Accra, Central, Eastern, Northern, Western, Upper East, Upper West, and Volta. We used the data to calculate the distribution of households or persons in each district for the following variables that are associated with child survival: household’s main source of cooking fuel (wood, charcoal, other biomass, kerosene, liquefied petroleum gas [LPG], electricity), type of sanitation (toilet) facility usually used by households (improved, unimproved), household’s main source of drinking water (improved, unimproved), maternal education (highest educational grade completed: none, primary, secondary or higher), paternal education (highest educational grade completed: none, primary, secondary or higher), and urban versus rural place of residence. We classified census responses on drinking water source and sanitation facility as improved versus unimproved based on WHO/UNICEF joint monitoring program categories for water supply and sanitation (http://www.wssinfo.org). Our measure of under-five mortality was the probability of dying by 5 y of age (5q0). The censuses asked all women of childbearing age to report on the total number of children ever born and children surviving. This information was the basis for the application of indirect demographic models to estimate under-five mortality, an approach commonly used by researchers and by national and international statistical and health agencies. The method converts the proportion of deaths among children ever born to women in each 5-y age group of the reproductive period into estimates of the probability of dying by exact ages of childhood, and calculates the number of years before the survey date to which the estimates refer [23,24]. Estimates derived from the two youngest age groups of women (15–19 and 20–24 y) tend to be overestimated compared to the population average because children of women in these age groups tend to have a higher risk of dying than children of older women [23]. Therefore, and following other analyses (including those by the UN Inter-agency Group for Child Mortality Estimation [IGME]) [23], we excluded 5q0 estimates based on reports of women aged 15–24 y. The remaining five estimates (one for each 5-y age group between 25 and 49 y), each with a reference date in years prior to the survey date, were used in our analysis. The reference dates for 5q0 estimates for the 2000 census covered the period 1987–1996; for the 2010 census, the period was 1997–2007. To obtain a single 5q0 estimate for each district for the index years 2000 and 2010, we fitted a Bayesian space-time model to the five estimates per district from each census. The model included a linear time trend for estimates from each census in each district. The district intercepts and slopes were modeled using the Besag, York, and Mollié (BYM) model [25]. In the BYM model, information is shared both locally (amongst neighboring districts), through spatially structured random effects with a conditional autoregressive (CAR) prior, and globally, through spatially unstructured Gaussian random effects. District-specific intercept and slope values are estimated by the sum of their respective spatially structured and spatially unstructured random effects. The prior distributions in the Bayesian framework allow district-specific parameters to be estimated on the basis of a district’s own data and those of its neighbors. This approach balances between overly unstable within-district estimates and overly simplified aggregate national estimates. Samples from the posterior distributions of the intercepts and slopes were used to estimate 5q0 for the years 2000 and 2010. The national 5q0 estimates based on the Ghana censuses alone were different from the UN IGME estimates, which use a larger number of data sources [1], especially for 2000 (the additional sources used in the UN IGME estimates are not available at the district level). We adjusted our estimates to match the UN IGME estimates in 2000 and 2010, because the additional data sources likely help provide more reliable estimates. As detailed above, the linear trends fitted to the census-based national estimates were used to obtain 5q0 in 2000 and 2010. We then applied a correction factor, calculated as the ratio of the national estimate of 5q0 from the UN IGME to that from the census alone, to the 5q0 estimates for each district. The means of the corrected district 5q0 estimates for the 2000 and 2010 censuses were 104.7 and 77.5, respectively, compared to 102.1 and 75.1 deaths per 1,000 live births as estimated by the UN IGME. We estimated the associations of under-five mortality with its social and environmental determinants at the district level. We analyzed the associations in 2000 and 2010, as well as the associations of the change in under-five mortality between 2000 and 2010 with changes in these factors. The model for associations in each census year was where 5q0 is the district-level under-five mortality (per 1,000 live births); X is a vector of district-level risk factors (each as percent of households or persons), including cooking fuel type (wood, charcoal, other biomass, kerosene, LPG, electricity), sanitation facility (improved, unimproved), drinking water source (improved, unimproved), maternal education (none, primary, secondary or higher), paternal education (none, primary, secondary or higher), and place of residence (urban, rural); U is a district-specific spatially structured random effect; V is a district-specific unstructured random effect; and α and β are regression coefficients. When analyzing change in under-five mortality, Ln(5q0) and X were replaced with their 2000–2010 differences. Noninformative normal priors with mean 0 and variance 10,000 were placed on all fixed effects parameters; gamma priors, Gamma(0.5,0.0005), were specified for the precision parameters of all random effects. For each analysis, two chains were run, and convergence was monitored using Brooks-Gelman-Rubin diagnostics [26] and visual inspection of the chains. Following convergence, which was before 10,000 iterations in all analyses, a further 200,000 iterations were run, with thinning to every tenth iteration, yielding final samples of 20,000 iterations for inference. All analyses were implemented using R2WinBUGS and sqldf libraries in the open-source statistical package R version 3.1.0 (R Project for Statistical Computing) and WinBUGS version 1.4 [27].