Space-time patterns in maternal and mother mortality in a rural South African population with high HIV prevalence (2000–2014): results from a population-based cohort

We used longitudinal data obtained from the Africa Centre Demographic Information System (ACDIS) in rural KwaZulu-Natal (KZN) Province, South Africa, which was established in 2000 [20] (Fig. 1). The area is 438 km² in size and includes a population of approximately 90,000 people who are members of approximately 11,000 households. The area is typical of many rural areas of South Africa in that while predominantly rural, it contains an urban township and informal peri-urban settlements. It is characterised by large variations in population densities (20–3000 people/km²). In the rural areas homesteads are scattered rather than grouped. Most households are multi- generational and range with an average size of 7.9(SD = 4.7) members [20]. Fieldworkers have diagramed all homesteads and facilities applying differential global position systems [21] to record the geographical locations.

Qualified medical practitioners ascertained the probable cause of death through interviews carried out with caretakers of the deceased or witnesses of deaths. The method used questionnaires to elicit pertinent information on signs, symptoms, and circumstances leading to death, generically described as indicators, which were subsequently interpreted into causes of death. The medical practitioners recorded a narrative of the circumstances leading up to the death based on the standard INDEPTH/WHO verbal autopsy questionnaire [22, 23].The Africa Centre for Demographic Information System used two methods to determine cause of death for each case: physician coding and an automated method using the InterVA probabilistic verbal autopsy interpretation model, before 1 January 2010 and on or after 1 Jan 2010 respectively [24]. In the physician-coded method, two clinicians independently assigned cause of death on the basis of the information collected during the verbal autopsy and their clinical judgement. If consensus could not be reached between the physicians, a third clinician reviewed all cases and codified the causes of death using the International Classification of Diseases, 10th revision (ICD-10) [25]. The InterVA model is based on Bayesian calculations of probabilities that a particular death was due to particular causes, given a set of symptoms and circumstances associated with the death [26]. The probabilistic model involves the building of a defined set of indicators (signs, symptoms, history, and circumstances) as the components of the model. The model produces the likely causes of death for each case, together with respective likelihoods. A definitive grid of conditional prior probabilities was characterized by a trained panel of physicians [26, 27].

We carried out a longitudinal analysis on 38,370 women of child bearing age (15-49 years) registered in the Africa Centre DSA who were born, residing, or in-migrated as a resident in the study area between 2000 and 2014. We determined maternal death using verbal autopsy and defined it, as the death of a woman while pregnant or within 42 days of termination of pregnancy, irrespective of the duration and site of the pregnancy, from any cause related to or aggravated by the pregnancy or its management, but not from accidental or incidental causes. We defined mother death, as the death of the mother while their child was less than or equal to 5 years of age. We included all 1204 deaths (maternal and mother) of women in the age range 15–49 that occurred between 2000 and 2014 in the analysis.

We conducted data analysis using STATA software (version 14) to establish the maternal mortality rates. We calculated the mortality rate using the number of deaths and person-years lived by the women of reproductive age (typically those aged 15 to 49 years) for each year, with a 95% confidence intervals (CI) for mortality rates being computed using the exact CI based on the Poisson distribution.\( MMR=\frac{number\ of\ maternal\ deaths}{number\ of\ live\ births}\times 100,000 \) Person years.

We then used a logistic regression to identify the risk factors for maternal mortality. We employed the Mosley-Chen and Meade models [28] to identify the risk factors affecting maternal mortality based on three levels: community, household and individual.

We utilized an exponential semivariance model for the spatial kriging of maternal mortality in the ACDIS. For the partial sill and range we assumed the following parameters, namely 10 and 3.3, respectively with no nugget effect. The analysis was performed in R software using the geoR package [29]. To identify circular maternal mortality clustering using spatial and spatial–temporal statistics respectively, we adopted the Kulldorff spatial clustering method [30] using the SaTScan software version 9.3 [31]. The Kulldorff Scan statistic identifies clusters with a higher number of observed cases (maternal deaths) relative to expected cases, under the assumption of spatial randomness, and then evaluates their statistical significance by skimming a circular window that covers the study area. A likelihood ratio test analyses the observed maternal deaths within the circle to the expected maternal deaths across the full range to determine significant risk clusters of mortality, giving relative risk and p values for any clusters determined [30]. We ran the model with a maximum cluster size of 50% of the total population, and p values achieved across 999 Monte Carlo replications to assure no loss of power at the alpha = 0.05 level [30]. However, one disadvantage of the Kulldorff scan statistic is that it utilises a circular window to denote the possible cluster areas, and cannot detect irregular shaped clusters. We therefore also used, the Flexible spatial scan statistic (Tango spatial scan statistics implemented in FleXScan [32]), in which the identified cluster is both flexible in shape and restricted to relatively small neighbourhoods of each region. The Flexible Scan statistic sets a practicable limitation of a maximum of 30 nearest neighbours for finding possible clusters due to the heavy computational load [32]. We aggregated the data into 705 grid cell size for FlexScan, given its current limitations of being able to search a maximum of 30 adjacent nodes. To refine the extents of the identified space-time cluster using FleXScan, we then used the Kulldorff space-time statistics for cluster detection of maternal mortality.

During the 15 years, there were 212 (0.55%) maternal deaths from 32,620 live births, with the maternal mortality ratio being 650 per 100,000 live births. Overall, 1204 mothers died while their child was less than or equal to 5 years giving a mortality rate of 370 per 100,000 live births. The trends for maternal mortality are shown in Fig. 2. The trend line in Fig. 2 reveals some evidence for a maternal mortality rate decline over the study period, from approximately 800 to 400 per 100,000 live births. The linear trend above is statistically significant at the p = 0.1 level (IRR = 0.969 [95% CI: 0. 938–1.002], p-value = 0.069). If the current rate of decline is maintained, the Sustainable Development Goal (SDG) 3.1 target for MMR by 2030 will be reached in 2024/2025.

We detected a significant primary cluster of mothers who died when their children were less than 5 years in the south east of the study area, with a relative risk of 1.58 (p = 0.022) (Fig. 3). This cluster is a high density peri-urban area located at the intersection of two major roads. Our study found no clear evidence of spatial clustering of maternal mortality events in the study area (Fig. 4). The summarised location, observed maternal deaths, expected maternal deaths, relative risk per each identified cluster are in Table 1.

We used the Mosley-Chen and Meade models [28] to identify the risk factors affecting child and maternal mortality based on three levels of analysis: community (food distribution, physical infrastructure like railroad, quality of water, electricity, water supply, road networks and political institutions), household (capital, wealth effects (food production, clothing essentials, housing conditions, energy availability, transportation, means to purchase what is necessary for hygienic purposes/preventive care, access to information), and individual (skills, health and time, normally measured by mother’s educational level, whilst father’s education correlates with occupation and household income). Table 2 depicts the baseline characteristics the identified risk factors of maternal mortality.

We applied univariable logistic regression to analyze the relationship between maternal death and risk factors, with crude odds ratios and 95% confidence intervals being estimated for each parameter. We then used multivariable logistic regression to build an overall model from the factors that were significantly associated with maternal mortality in the univariable analysis. In the final model, significant associations with the risk of maternal death were age, HIV status, education and parity (Table 3). Primary education or none, being HIV positive, and higher parity were significant predictors of increased risk of maternal death. We also presented the multivariable logistic regression results for factors associated with mother death clusters (Table 4).