Menu
Menu
Menu
Menu
Background Maternal mortality remains a public health problem despite several global efforts. Globally, about 830 women die of pregnancy-related death per day, with more than two-third of these cases occurring in Africa. We examined the spatial distribution of maternal mortality in Africa and explored the influence of SDoH on the spatial distribution. Methods We used country-level secondary data of 54 African countries collected between 2006 and 2018 from three databases namely, World Development Indicator, WHO’s Global Health Observatory Data and Human Development Report. We performed descriptive analyses, presented in tables and maps. The spatial analysis involved local indicator of spatial autocorrelation maps and spatial regression. Finally, we built Bayesian networks to determine and show the strength of social determinants associated with maternal mortality. Results We found that the average prevalence of maternal mortality ratio (MMR) in Africa was 415 per 100 000 live births. Findings from the spatial analyses showed clusters (hotspots) of MMR with seven countries (Guinea-Bissau, Guinea, Sierra Leone, Cote d’Ivoire, Chad and Cameroon, Mauritania), all within the Middle and West Africa. On the other hand, the cold spot clusters were formed by two countries; South Africa and Namibia; eight countries (Algeria, Tunisia, Libya, Ghana, Gabon and Congo, Equatorial Guinea and Cape Verde) formed low-high clusters; thus, indicating that these countries have significantly low MMR but within the neighbourhood of countries with significantly high MMR. The findings from the regression and Bayesian network analysis showed that gender inequities and the proportion of skilled birth attendant are strongest social determinants that drive the variations in maternal mortality across Africa. Conclusion Maternal mortality is very high in Africa especially in countries in the middle and western African subregions. To achieve the target 3.1 of the sustainable development goal on maternal health, there is a need to design effective strategies that will address gender inequalities and the shortage of health professionals.
This is an ecological study based on aggregated country-level data extracted from three publicly available databases: World Development Indicator, WHO’s Global Health Observatory Data and Human Development Report (see online supplemental file 1). Data from 54 African countries were included in the study, the data were majorly based on nationally representative cross-sectional surveys such as Demographic and Health Surveys, Household Income and Expenditure Surveys, House Living Standard and Socioeconomic Surveys. The study year for the variables differs, ranging from 2006 to 2018; detailed description of data collection and survey designed are published elsewhere.16–18 bmjgh-2020-004233supp001.pdf MMR was defined and computed as the number of women per 100 000 live births who died from pregnancy-related causes during pregnancy or within 42 days of pregnancy termination. This is computed by dividing recorded or estimated maternal death (non-AIDS women aged 15–49 years) by the total record or estimated live birth within the same period. The data were extracted from the World Bank Development Indicator, which is based on estimations from raw data collection in national representative cross-sectional surveys such as household income and expenditure surveys, house living standard and socioeconomic surveys. Twenty-three independent variables measured at country-level were involved in the study. They were selected based on previous research on maternal mortality in Africa, and to cover the different aspects of SDH framework (see figure 1). As demonstrated in figure 1, we operationalised the selected independent variables using the social determinant models to highlight the possible relationships between MMR and the selected variables. The country-level demographic, socioeconomic and sociocultural factors examined include income inequality, Gross National Income (GNI) per capita, Gender Inequality Index (GII), urban residence, poverty rate, crude birth rate and female educational status. Behavioural factors involved are the prevalence of adult female currently smoking and total alcohol per capita consumption in female adult. We included prevalence of anaemia during pregnancy, the prevalence of obesity among female adults (age-standardised); underweight among female adults (as a proxy for maternal nutrition), the prevalence of hypertension among adult females (age-adjusted) and prevalence of diabetes mellitus in female adults (age-adjusted). Finally, we included factors associated with healthcare service coverage; the percentage of deliveries by caesarean section (CS, as a proxy for access to emergency healthcare during childbirth), antenatal care (ANC) coverage—at least four visits, skilled birth attendants during delivery, and proportion of women of reproductive age who have their need satisfied for family planning, adolescent birth rate and prevalence of child marriage—proxies for sexual and reproductive health. We hypothesised that the overarching determinants in the outermost box of the framework adapted from the WHO will influence factors in inner boxes, and this will further modify MMR in Africa. We conducted multicollinearity test and removed highly correlated variable by picking most relevant one maternal mortality research. Social determinants framework for maternal death. We conducted a descriptive analysis to explore the distribution of the dependent and independent factors examined in the social determinants of health (SDoH) framework. We adapted the WHO framework to build our models in three main domains (socioeconomic/cultural, healthcare resources and maternal conditions) for easy interpretations and policy implications. The values were expressed as mean with SD or median with IQR. We used GeoDa v. 1.14 software to perform the spatial analysis, it is more suitable because it explicitly handles spatial data and allows statistical test findings as desired in this study.19 Country was used as a unit scale for spatial analysis, the distance-based threshold of 4272 km (Arc distance) was generated with X and Y coordinates from shapefile downloaded from Intergovernmental Action on Development Climate Prediction and Application Centre (ICPAC) GeoPortal (http://geoportal.icpac.net/). The appropriateness of this spatial weight was confirmed by assessing the symmetry of the connectivity histogram. The connectivity map also showed that all the 54 countries interlinked which is necessary to ascertain spatial dependency. We generated quantile cluster map to show the spatial distribution of the MMR in Africa descriptively. We also performed Global Moran’s I analysis to examine if spatial autocorrelation occurs at local level. Local indicator of spatial autocorrelation (LISA) cluster maps was also generated to statistically show the hot and cold spots spatial clusters of neighbouring countries with high and low MMR, respectively. Local spatial autocorrelation was measured with Local Moran’s I index which ranges from −1 to +1; with positive (+) values indicating strong clustering and negative (–) values indicating dispersion. Finally, we conducted spatial regression based on approaches developed by Anselin20 to investigate the association between MMR and independent variables. Six models were built with the first five covering a different aspect of SDoH framework and the final model only contained significant variables from each model. A pair-wise correlation was used to deal with missing data which were <2%. Ordinary least square (OLS) diagnostic was examined for each model, where spatial dependency was indicated, the model was fitted with spatial error or spatial lag regression as appropriate; best-fit model was determined using R-squared, Log likelihood and Akaike information criterion. We also tested for multicollinearity of independent variables, value <30 was used;20 999 Monte Carlo permutation was used for randomisation to ensure p-value<0.05. We used the Bayesian network analysis to examine the strength of the relationship between MMR and the determinants. A machine learning R package called ‘bnlearn’ was used for the Bayesian network modelling;21 22 it is based on GeNie software. By using the variables that showed significant association in the final model of the spatial regression, we built the structure of the network from the domain knowledge of the authors on temporal precedence of the variables, using score-based structure algorithms. We introduced geographical regions based on the United Nation’s categorisation as a variable to deal with possible residual cofounding and accounted for missing values by using the expectation–maximisation algorithm. All the variables were continuous data, we discretised them by their median values to show the diagnostic conditional probability distribution. We used supervised machine learning techniques to predict the joint conditional probability of significant independent variables from the Bayesian network when countries with higher MMR are reduced to 70 MMR per 100 000; the target for the 2030 SDG for MMR. Finally, we used k-fold cross-validation (at k=10) to compare and examine the Bayesian model’s goodness of fit; log-likelihood loss was used as the loss function, therefore, the lower the value, the better the fit.23 No patients or the public were directly involved in the design, conduct, reporting or dissemination plans of this research.
N/A
