Quantifying within-city inequalities in child mortality across neighbourhoods in Accra, Ghana: A Bayesian spatial analysis

Objective Countries in sub-Saharan Africa suffer the highest rates of child mortality worldwide. Urban areas tend to have lower mortality than rural areas, but these comparisons likely mask large within-city inequalities. We aimed to estimate rates of under-five mortality (U5M) at the neighbourhood level for Ghana’s Greater Accra Metropolitan Area (GAMA) and measure the extent of intraurban inequalities. Methods We accessed data on >700 000 women aged 25-49 years living in GAMA using the most recent Ghana census (2010). We summarised counts of child births and deaths by five-year age group of women and neighbourhood (n=406) and applied indirect demographic methods to convert the summaries to yearly probabilities of death before age five years. We fitted a Bayesian spatiotemporal model to the neighbourhood U5M probabilities to obtain estimates for the year 2010 and examined their correlations with indicators of neighbourhood living and socioeconomic conditions. Results U5M varied almost five-fold across neighbourhoods in GAMA in 2010, ranging from 28 (95% credible interval (CrI) 8 to 63) to 138 (95% CrI 111 to 167) deaths per 1000 live births. U5M was highest in neighbourhoods of the central urban core and industrial areas, with an average of 95 deaths per 1000 live births across these neighbourhoods. Peri-urban neighbourhoods performed better, on average, but rates varied more across neighbourhoods compared with neighbourhoods in the central urban areas. U5M was negatively correlated with multiple indicators of improved living and socioeconomic conditions among peri-urban neighbourhoods. Among urban neighbourhoods, correlations with these factors were weaker or, in some cases, reversed, including with median household consumption and women’s schooling. Conclusion Reducing child mortality in high-burden urban neighbourhoods in GAMA, where a substantial portion of the urban population resides, should be prioritised as part of continued efforts to meet the Sustainable Development Goal national target of less than 25 deaths per 1000 live births. Ghana is among the most urbanised countries in SSA with an estimated urban population of over 18 million in 2021 (58% of the total population), that is growing by ~3% each year.7 GAMA is Ghana’s administrative and economic capital and accounted for 29% of the country’s urban population in 2010.35 It covers ~1500 km2 on the southern coast of the Greater Accra region. According to the 2010 census, GAMA comprises 5019 enumeration areas (EAs)—the smallest administrative geographical unit in Ghana—nested within 406 localities and 12 districts or ‘municipalities’ (figure 1). The Greater Accra Metropolitan Area (GAMA) with neighbourhood boundaries shown in grey and district boundaries shown in black (grey) (source: Ghana Statistical Service). Urban neighbourhoods are shown in purple and peri-urban neighbourhoods shown in green. The inset shows the location of GAMA in Ghana and Western Africa. The centrally located Accra Metropolitan Area (AMA)—together with the more heavily industrialised Tema and Ashaiman municipalities to the southeast—contain the most densely populated neighbourhoods. AMA contains the central business district and functions as the city’s commercial, industrial and administrative centre.36 37 Rapid development since Ghana’s independence in 1957 has contributed to increased congestion in AMA’s residential areas. Planned residential neighbourhoods in AMA remain as legacies of the colonial era, while migrants and low-income individuals have been pushed into slums and other low-income neighbourhoods that can lack basic services and infrastructure.17 Tema is GAMA’s planned industrial hub with structured housing developments and services, and was the fastest-growing municipality following independence. A small fraction (~5%) of GAMA’s population lives in areas classified in the 2010 census as rural, mostly in northern GAMA and predominantly in the Ga West and Ga South districts. These districts are characterised by sprawling urban development with high rates of population growth since the 1970s due to congestion of the city centre.36 38 Overall, the U5M rate in Ghana almost halved from 1990 to 2010, though considerable subnational inequality persisted.1 31 32 During this period, the government implemented several national health policies and programmes to improve the use and delivery of maternal and child healthcare services.39 The National Health Insurance Scheme (NHIS) provides free healthcare to participants40 and, in 2008, enrolment was made free of charge for pregnant women and children under 18 years of age.41 Antenatal and postnatal visits, facility delivery (including emergency obstetric care) and neonatal care are all included under the scheme. The User Fees Exemption for Delivery Care was scaled up in 2005 and exempts pregnant women who are not enrolled in NHIS from paying delivery fees.42 Ghana’s 2007–2015 Child Health Policy aimed to unify fragmented programme delivery under a recommended continuum of care for mothers and children, scaling up interventions with proven efficacy to prevent child deaths, including, for example, oral rehydration therapy and zinc for treatment of diarrhoea, vitamin A supplementation and antibiotic treatment for pneumonia.43 Together, these efforts have contributed to reductions in overall child mortality and in inequalities between subregions, though wider concerns have persisted over the quality of care.44 In Accra, women enrolled in NHIS were found more likely to seek formal care and visit clinics, but enrolment rates were lower (under 35%) among women of childbearing age compared with women over the age of 50.41 Although most births (>90%) in Greater Accra take place in a health facility in the presence of a skilled health professional, the coverage rate of age-appropriate vaccines in children drops with age (from 76% to 48% comparing children aged 1 vs children aged 2–3 years, respectively). This indicates considerable variability in continued access to care for children.45 We accessed the full microdata of the most recent Ghana Population and Housing Census, conducted in 2010, via the Ghana Statistical Service. The census collected information on the number of children born and surviving at the time of survey, known as their summary birth history, for all women aged 12 years and older. Other individual and household characteristics captured in the census include employment status, occupational industry, schooling level, literacy, household amenities (including type of cooking fuel, drinking and non-drinking water source, sanitation facilities, lighting source and waste disposal method in use), dwelling type and structural features (including roof, floor and wall materials) and EA of residence. Together, these provide information on the socioeconomic and living environments. We obtained a shapefile from the Ghana Statistical Service with all GAMA EAs, localities and districts geocoded according to the 2010 census geographies. Localities within GAMA were the neighbourhood units used in our analysis, each containing between 1 and 95 EAs. We linked the census data and the shapefile using codes that uniquely identified EAs to determine each individual’s neighbourhood of residence. Neighbourhoods were defined by the Ghana Statistical Service and are the administrative units at which urban versus rural classification is defined in Ghana; those with 5000 inhabitant or more are considered urban, and rural otherwise. Some neighbourhoods were recently subdivided due to population growth and, thus, did not meet the urban population threshold despite the urban designation of their constituent EAs in the census. We, therefore, classified GAMA neighbourhoods as urban or periurban according to the historic census-derived urban–rural designation of their constituent EAs. Most neighbourhoods (98%) comprised exclusively urban or exclusively rural EAs. We classified neighbourhoods that contained both urban and rural EAs as urban if over 50% of the population lived in urban EAs and periurban otherwise (figure 1). Ghana does not have an official definition that distinguishes periurban from rural neighbourhoods; however, we used the term periurban to better describe ‘rural’ neighbourhoods that are located within the administrative border of GAMA on the periphery of the densely populated inner-city and industrial areas (figure 1). To assess U5M, we summarised the birth history data of women of reproductive age (15–49 years) by five-year age group and neighbourhood. The Maternal Age Cohort (MAC) method was used to estimate the neighbourhood probability of death for children before the age of five (5q0) for each five-year age group, based on the number of children ever born, proportion of children who have died and average parity.46 The MAC method outperforms alternative methods for estimating U5M from summary birth history data for subnational populations.47 Each 5q0 was assigned to a reference year prior to the census, using maternal age as a proxy measure for duration of exposure to risk of death for a child. The assigned 5q0 reference years covered the period from 1990 to 2005. We excluded 5q0 estimates derived from women aged 15–19 and 20–24 years, owing to the low numbers of births recorded for these age groups in many neighbourhoods, which could lead to spurious fluctuations in the 5q0 estimates, especially in the more sparsely populated periurban areas. Notably, this is common practice when using demographic methods to estimate population U5M rates.46–48 This left five 5q0 estimates for each neighbourhood (one derived from each five-year age group of women aged 25–49 years). To obtain neighbourhood estimates of 5q0 for 2010, the year of the census, we fitted a Bayesian spatiotemporal model to the MAC-derived 5q0 estimates across all 2030 neighbourhood-reference year units, transformed to the probit scale. The model included a linear time trend that could vary by neighbourhood. The time trend allowed data from different reference years, each of which is associated with a different age group, to inform the 5q0 in 2010. The neighbourhood intercepts and slopes were modelled using the Besag, York and Mollié model,49 where information is shared locally (ie, among adjacent neighbourhoods) through spatially structured random effects with a conditional autoregressive prior and globally through spatially unstructured Gaussian random effects. Neighbourhood-specific intercept and slope values were estimated by the sum of their respective spatially structured and spatially unstructured random effects. The prior distributions in the Bayesian framework allow the neighbourhood-specific parameters to be estimated by a neighbourhood’s own data and data of contiguous neighbourhoods. This approach balances overly unstable within-neighbourhood estimates and overly simplified aggregate estimates for all of GAMA. The reported estimate for the wealthy Ringway neighbourhood in AMA is informed entirely by data in bordering neighbourhoods as the data on child deaths were considered implausible (see online supplemental appendix 1). To account for excess variability resulting from small numbers of children born to women in a given age group and neighbourhood, we included a weighted variance term that gave more weight to estimates derived from a higher number of births. Samples from the posterior distributions of the intercepts and slopes were used to estimate 5q0 for the year 2010. bmjopen-2021-054030supp001.pdf To avoid infinite values on the probit scale, we adjusted all MAC-derived 5q0 estimates of zero (n=146; 7%) to half the minimum estimated non-zero value across all units (0.00316). We conducted sensitivity analyses to ensure that our results were robust to this choice by replacing zero estimates with a lower value of 0.0001 and with the minimum estimated non-zero 5q0. We monitored convergence using trace plots and obtained 5000 post burn-in samples from the posterior distributions of model parameters. We summarised the distributions of neighbourhood-specific parameters to report neighbourhood U5M estimates and mean U5M across neighbourhoods within districts for 2010, with 95% credible intervals (CrI) that represent the mean and the 2.5th and 97.5th percentiles of the posterior samples, respectively. We present neighbourhood U5M estimates as deaths per 1000 live births. We calculated neighbourhood-level summary statistics of individual and household characteristics to provide context for our mortality results. We used the within-neighbourhood median household consumption as a measure of neighbourhood socioeconomic level. Household consumption is considered a better indicator of living standards than household income in low-income and middle-income settings.50 51 The census did not include consumption data, so we used small-area estimation methods to indirectly calculate consumption based on household characteristics described in detail elsewhere.52 Briefly, we used the 2012 Ghana Living Standards Survey to develop a statistical relationship between household characteristics and consumption. Then using those same household characteristics, we predicted consumption for households in the census. We additionally calculated population density; the proportion of the women of reproductive age (15–49 years) who were literate, had schooling to at least primary, middle, secondary and postsecondary levels; the proportion of the working age (15–64 years) population in any employment and in primary, secondary and tertiary sector occupations; and the proportion of households with indicators of improved living conditions (including dwelling type, materials of flooring, roofing and walls, methods of solid and liquid disposal, type of toilet facility, type of cooking fuel and type of drinking and other water source). Details of our classification of the census responses into indicators of ‘improved’ versus ‘unimproved’ living conditions are provided in online supplemental appendix 2. bmjopen-2021-054030supp002.pdf We measured the correlations between neighbourhood U5M and neighbourhood socioeconomic and living environment indicators using the non-parametric Spearman’s rank method. We measured the correlations across all GAMA neighbourhoods and separately across urban and periurban neighbourhoods. All analyses were implemented in the open-source statistical software R V.3.6.1. The Bayesian model was implemented using the NIMBLE package V.0.9.1. The study used secondary data only.