Understanding the determinants of maternal mortality: An observational study using the Indonesian Population Census

Background For countries to contribute to Sustainable Development Goal 3.1 of reducing the global maternal mortality ratio (MMR) to less than 70 per 100,000 live births by 2030, identifying the drivers of maternal mortality is critically important. The ability of countries to identify the key drivers is however hampered by the lack of data sources with sufficient observations of maternal death to allow a rigorous analysis of its determinants. This paper overcomes this problem by utilising census data. In the context of Indonesia, we merge individual-level data on pregnancy-related deaths and households’ socio-economic status from the 2010 Indonesian population census with detailed data on the availability and quality of local health services from the Village Census. We use these data to test the hypothesis that health service access and quality are important determinants of maternal death and explain the differences between high maternal mortality and low maternal mortality provinces. Methods The 2010 Indonesian Population Census identifies 8075 pregnancy-related deaths and 5,866,791 live births. Multilevel logistic regression is used to analyse the impacts of demographic characteristics and the existence of, distance to and quality of health services on the likelihood of maternal death. Decomposition analysis quantifies the extent to which the difference in maternal mortality ratios between high and low performing provinces can be explained by demographic and health service characteristics. Findings Health service access and characteristics account for 23% (CI: 17.2% to 28.5%) of the difference in maternal mortality ratios between high and low-performing provinces. The most important contributors are the number of doctors working at the community health centre (8.6%), the number of doctors in the village (6.9%) and distance to the nearest hospital (5.9%). Distance to health clinics and the number of midwives at community health centres and village health posts are not significant contributors, nor is socio-economic status. If the same level of access to doctors and hospitals in lower maternal mortality Java-Bali was provided to the higher maternal mortality Outer Islands of Indonesia, our model predicts 44 deaths would be averted per 100,000 pregnancies. Conclusion Indonesia has employed a strategy over the past several decades of increasing the supply of midwives as a way of decreasing maternal mortality. While there is evidence of reductions in maternal mortality continuing to accrue from the provision of midwife services at village health posts, our findings suggest that further reductions in maternal mortality in Indonesia may require a change of focus to increasing the supply of doctors and access to hospitals. If data on maternal death is collected in a subsequent census, future research using two waves of census data would prove a useful validation of the results found here. Similar research using census data from other countries is also likely to be fruitful.

There are issues with measurement of MMRs using the DHS data sets that cause us to qualify the message apparent in Fig A. The most crucial step in estimating the MMR is identifying maternal deaths. The DHS relies on the 'sisterhood method'. The DHS women's questionnaire asks each woman in the household about the survivorship of all the live-born children of her mother, that is, her maternal siblings. For any female siblings who died at age 10 or older, further questions determine whether the death was pregnancy-related (pregnant when died, died during childbirth or died within 2 months after the end of a pregnancy). The sisterhood method allows data to be collected on pregnancy-related deaths for a larger sample of womennot just those in the immediate household being surveyed, but also the sisters of those women in the surveyed household -however, even then the number of deaths captures is small. The 2012 DHS figure for Indonesia was based on only 92 reported maternal deaths [1]. As a result, these point estimates of the MMR carry such large sampling errors that it is not possible to determine whether Indonesia's maternal mortality rate has indeed increased. Even the improvement of the MMR from 390 in 1990-94 to 228 in 2003-07 is not a statistically significant decline [2].
An alternative to the DHS is to use a model-based approach which models the MMR drawing on a variety of data sources and combines these in a coherent way to provide estimates of MMRs. This is the approach used by the Maternal Mortality Estimation Inter-Agency Group (MMEIG), a group comprised of representatives from the WHO, UNICEF, UNFPA, UNPD, World Bank, National University of Singapore and the University of California at Berkeley [3]. On this basis Indonesia has been making progress in reducing its MMR from 446 in 1990 to 165 in 2010 to 126 in 2015 (a decrease of 72% since 1990).

Fig B plots
Indonesia's progress using this method relative to other countries in the region. It is important to understand that the MMEIG estimates for Indonesia are at least in part constructed from a multi-level regression model which uses variables such as country GDP, general fertility rates and rates of skilled attendants at birth to predict the proportion of deaths among women of reproductive age that are due to maternal causes. This figure is then converted to a MMR by incorporating data on population estimates and live births so the estimate for 2010 of 165 is best thought of as a model projection of what ought to have occurred, given the values of these inputs. It takes no account of the myriad other factors that would cause variation from the model. For example, with Indonesia's very large and geographically spread population, predicting an improvement in the MMR based on increased numbers of birth attendants will fail to capture the potential variation in spatial distribution of these attendants, or the variable quality of their training, or their access to the required equipment or medical supplies in the health centres where they work. Being a model based on slowly evolving input variables, the MMEIG model is also likely to predict a continuation of the trend evident in past data.
Finally, the 2010 Population Census data can be used to construct an estimate of Indonesia's MMR. There are a number of challenges associated with estimating MMRs from census data [4,5]. As discussed in the main text, the MMR implied from the raw census data of 8075 maternal deaths and 5,866,791 live births, is 137 deaths per 100,000 live births. Adjusting the raw estimate to correct for potential underreporting of pregnancy-related deaths produces an MMR of 263 [6]. Because the critical questions are not used in previous censuses, there is no capacity to compare the same measure across time.
So for 2010 we have estimates of the MMR ranging from 137 using the raw Population Census data, to 165 using the model-based MMEIG method, to 263 using the census data and correcting for potential under-reporting of deaths, to 359 using the DHS sisterhood method. The two series of estimates that allow a comparison across time show very different changes between 1990 and 2010. The DHS suggests a decline of only 8%, whereas the MMEIG figures suggest a decrease of 52% over the same period.

Previous Literature on Determinants of Maternal Mortality in Indonesia
We searched PubMed for studies with the search string "("maternal mortality" OR "maternal death") in the Title AND ("Indonesia") in the Title/Abstract" which were published between 2000 and February 2018. Of the 18 studies that met these criteria, nine examined the determinants of maternal mortality (the others being commentary pieces, focused on measuring maternal mortality rates or evaluations of a specific program's impact on . Two studies used verbal autopsies to investigate causes of maternal death. One investigated causes of maternal deaths in two rural districts in Java and found that delays in seeking and receiving care were often reported by family members to have occurred [7]. The other examined 130 maternal deaths in Kalimantan and identified delays in decision-making and poor quality of care as contributing factors [8]. Of the seven quantitative studies of causes of maternal death, four studies were retrospective case-control studies based on relatively small samples from hospitals in cities in Java [9 -12]. Only one of these studies examined determinants of maternal death beyond direct medical causes, finding that living outside the city and incomplete attendance at ante-natal visits were significant determinants of maternal mortality [9]. Two studies used data from a population-based case-control study in two districts on Java. They found that the odds of dying increased with distance to a health centre and among women who were assisted by a health professional [13], mortality rates were very high among poor women [14]. One study used data from the Demographic and Health Surveys and found a very strong positive association between poverty and maternal death in Indonesia [15].