Analysis of study-specific variables shows a very clear cut-off between the different study designs, primarily surveys and vital registration (see Figure 3). Most of the estimates from developing countries come from surveys, and the inherent methodology entails certain study characteristics that are consistently different from estimates derived from vital registration, the established method in developed countries (e.g. sampling method, information on non-respondents or completeness of records, definition of maternal death). For example, whereas vital registration systems tend to use a definition of maternal death on the basis of cause of death, surveys tend to use a 'time-of-death' definition which includes all deaths that occur during pregnancy, childbirth or within a determined period of time after delivery regardless of the cause. Examination of the effect of different definitions on the estimates invariably groups countries according to their development status again, indicating close to complete confounding in this data set between study characteristics and country-level determinants of maternal mortality. For this reason, the effect of development status on the estimates is difficult to disentangle from the independent effect that other variables (especially study-design related) may have, and further analysis to explain variability at this level is not possible with this data set.
Interestingly, within all subregions except developed countries (Europe, Northern America, Japan, Australia and New Zealand) there is substantial variability between nations indicating the need for careful attention to country-level variables in understanding the specific issues of maternal mortality. In this analysis, inter-country comparison possibilities are very limited due to serious methodological differences between estimates. However, outliers in Figure 2 and 4 deserve special attention. There are six outliers in the less developed group in Figure 2 and four of these countries have MMRs greater than the median of the least developed countries. Although these countries do not fulfil the criteria to be classified as least developed, their high MMRs are more comparable to countries with the lowest development status. In-depth study of other possible economic, health, demographic and social differences could provide some light as to how to reduce maternal mortality in those settings.
Similarly, we can consider outliers at subregional level in Figure 4. For example, in Western Asia, a subregion with little variation between countries, Yemen (MMR = 351) and Iraq (MMR = 294) are clear outliers. Further research on other indicators at national level shows that these two countries present the lowest rates in the region for skilled birth attendant (21.6% and 72,1%, respectively), the lowest contraceptive-use prevalence rates (20.8% and 13.7%, respectively) and the highest infant mortality rates (80 and 95 per 1000, respectively).
The analysis of country-specific variables that might influence maternal mortality provides the opportunity to examine key factors in understanding the variability in maternal mortality. From the large number of country-specific variables that may be associated with maternal mortality we selected a few that represented a broad array of factors. These included variables proximal to maternal health including skilled birth attendant and contraceptive-use prevalence rate; intermediate variables such as health expenditure per capita; and distal measures such as population growth rate, proportion of urban population, probability of death between 15 and 59 (male) and female net primary school enrolment ratio. We reviewed the association between variables to identify independent indicators of the different factors.
A small number of these variables explained a substantial proportion of the variance observed in MMRs. These variables were all health related: skilled birth attendant, health expenditure per capita and infant mortality rate, and the first two have already been reported as determinants of maternal mortality using a smaller data set from sub-Saharan African countries [26]. Contraceptive-use prevalence rate, an indicator commonly used to describe access to health care for women was not independently associated with MMR in the multi-variate model. Although these indicators are all associated with development status, development status was not included in the analysis because it measures a wide range of issues and is potentially difficult to interpret and translate into programmatic recommendations. The findings here are supportive of the potential positive impact of skilled attendant at birth to reduce maternal mortality. Independent of this, increased health expenditure is also an important indicator supporting that not only skilled care but also general health infrastructure has an impact on maternal mortality.
Except for skilled birth attendant, the variables proposed here to predict maternal mortality are different than those used in the WHO model in 2000 (i.e. general fertility rate, skilled birth attendant, gross domestic product per capita, a regional variable, and a measure of completeness of death registration) [3], suggesting that further work towards understanding the covariates of maternal mortality would be useful. While the two exercises are not directly comparable (different dataset and modelling a different variable), it is useful to highlight certain contrasting findings in the belief that different approaches to the problem may each individually shed light on data gaps and on strategies to address them.
There are many country-level variables not included in this analysis that may also explain variability in MMR. In our initial analysis we considered a large number of potential indicators covering a broad range of factors. Many of these variables were highly correlated (r > 0.70). As part of our analysis plan we attempted to identify a sub-set of variables that covered different areas that were not highly inter-correlated but also variables most plausibly related to maternal mortality and that could be incorporated into public health programmes. While other variables may also explain variability, it is likely that they would point to similar set of programmatic activities.
Although this analysis presents national estimates from a large proportion of the world's live births, the lack of national estimates from China is the primary factor limiting near complete coverage of global information on maternal mortality. Data included in this analysis represent the most recent available estimate for each country. However, in some cases, especially in developing countries, most recent information could date from the 1980s given the retrospective aspect of sisterhood method estimates (even though the reporting date is after 1997). Changes in maternal mortality that may have occurred since the time of the study could influence both the variability and the identified associations. Normally, changes in maternal mortality, especially at national level, are slow, thus this would not likely impact the findings presented here. Furthermore, sisterhood methods, which include the oldest information presented here, essentially estimate an average across the time period so impact of rapid changes would be minimal.
In addition, data used for these analyses are derived from a variety of sources and methodologies, each presenting different pitfalls, constraints, precision and degree of reliability in identifying maternal deaths. For this reason, we do not attempt to provide here a summary estimate for maternal mortality. Previous attempts have been made to adjust for discrepancies, incompleteness or under-reporting in order to obtain estimates that may reflect the real situation better [3]. The data used here, however, are taken as reported, without adjustment, because we think that underreporting of maternal deaths must be estimated carefully through specific surveys tailored for each country [4, 5]. This implies that MMRs presented here are likely to be underestimates, especially when we take into account that more than half of the estimates did not involve any efforts to capture maternal deaths by, for example, searching multiple sources, and the same proportion did not attempt to confirm maternal deaths through, for example, confidential enquiries or verbal autopsy [27].
The results of this analysis of maternal mortality at national level show significant variation for developing regions (from 127 to 1289 in least developed and 2 to 695 in less developed). Although due to co-linearity, an in-depth exploration of study-specific variables could not be done to explain the variability, the distribution and outliers suggests clearly specific countries as priority targets. On the basis of the analysis of correlation of MMRs with other country-specific variables, it seems prudent that a model to predict maternal mortality at national level takes into account infant mortality rate, skilled birth attendant and health expenditure per capita, as well as possibly other variables.