Chapter 1: Background and Objectives
4. Measurement Issues in Maternal Mortality
Maternal mortality is defined as a death of a woman while pregnant or within 42 days of termination of pregnancy, irrespective of the duration and site of pregnancy, from any cause related to or aggravated by the pregnancy or its management, but not from accidental or incidental causes [1].
In addition to the underlying cause of death, antecedent and contributing causes should be recorded. An underlying cause of maternal mortality is a disease or injury that initiates the sequence of morbid events leading directly to the death. A contributing cause is defined as a condition that may exist prior the development of the underlying cause of death and contributes to the death [19, 20].
Recommendations for the categorisation of maternal deaths have distinguished between direct and indirect causes of death. Direct maternal deaths include death due to obstetric causes such as postpartum haemorrhage, puerperal sepsis, or abortion. Indirect maternal deaths are defined as those resulting from previous existing diseases, or diseases that developed during pregnancy, which are not due to direct obstetric causes, but aggravated by the physiological effects of pregnancy, during delivery or in the postpartum period. Examples include deaths due to malaria, anaemia, existing cardiac or renal diseases, or HIV/AIDS.
Incidental causes of maternal death include suicide. Coincidental causes of death include external causes that are thought to have no causal link to pregnancy, for example, car accidents [20].
The importance of consistent inclusion of indirect causes of maternal death, such as deaths due to malaria or HIV in pregnancy, has been stressed, particularly in settings with high HIV prevalence [10]. A recent publication proposes to further separate death due to unanticipated complications of management and unknown causes of death [21, 22]. The category of unanticipated complications of pregnancy was added to allow tracking trends in iatrogenic causes of death, for example, deaths due to Caesarean section. Further, new recommendations suggest to include deaths due to suicide in pregnancy, as a consequence of puerperal psychosis and postpartum depression, as “other direct maternal deaths” [20].
The International Classification of Disease (tenth revision) [19] also includes an alternative definition for maternal death that is more practical: A pregnancy-related death is the death of a woman while pregnant or within 42 days of termination of the pregnancy, irrespective of the cause of death. This definition thus includes accidental and incidental causes of death and
26 makes it possible to have estimates without establishing a cause of death analysis. A pregnancy-related death is thus solely defined by the timing of death relative to pregnancy, childbirth and the post-partum period and is likely to overestimate the true number of maternal deaths. To which extent pregnancy-related mortality overestimates true number of maternal deaths has been a matter of debate. Estimations for over reporting range from 18%
to 41% [23]. Others have argued that it is reasonable to assume that the error of inclusion of non-maternal deaths is counterbalanced by the common omission of deaths due to induced abortion in population-based surveys using siblings methods2 [23].
Main statistical measures of maternal mortality are:
Maternal mortality ratio (MM-ratio): Number of maternal deaths during a given time period per 100,000 livebirths during the same time period
Maternal mortality rate (MM-rate): Number of maternal deaths in a given time period per 100,000 women of reproductive age during the same time period
Lifetime risk of maternal death: the probability of maternal death across a woman’s reproductive life, usually expressed in terms of odds
Proportion of maternal deaths among female deaths (PMDF): Number of maternal deaths during a given time period divided by all causes of death of women of reproductive age, usually defined as 15-49 years [24].
The MM-ratio uses the number of livebirths as a denominator [1]. However, the preferred denominator to describe the risk of deaths in pregnancy and childbirth would be the number of pregnancies, and would thus include also pregnancies resulting in abortion3 and stillbirth4. Only a few demographic surveillance sites such as in Bangladesh [25] document pregnancies.
Historically, there have been some attempts to document stillbirths, but the documentation of livebirths has always been judged to be more accurate and thus became the preferred denominator [26].
Maternal mortality is a part of the all-cause mortality in the group of women of reproductive age. Estimates can be derived from different sources: 1) vital registration (birth and death
2Sibling’s methods are using reports of brother and sisters about living and diseased brothers and sisters to generate mortality rates.
3Abortion is defined as any termination of pregnancy by the removal or expulsion from the uterus of a fetus or embryo prior to viability
4Any death of fetus after 20 (or 24) weeks of gestational age or at least 500g weight (variety of definitions are used)
27 registrations) and sample vital registration with verbal autopsy; 2) population sample surveys where sibling methods are used (direct and indirect sisterhood method); 3) household census with complete birth and deaths ascertainment; 4) reproductive age mortality surveys using a variety of sources to identify all deaths among women of reproductive age; and 5) sentinel sites (demographic sites with documentation of births and deaths and subsequent verbal autopsies.
Another option is to use informant-based methods based on village informant networks. This method requests village heads or volunteers networks to list all deaths of women of reproductive age in their communities and define whether they are pregnancy-related. The listing exercise is followed by a visit to the deceased relatives to confirm or correct the details on age and pregnancy status [27]. This method was used in Indonesia and was able to identify an estimated 85% and 71% of pregnancy-related deaths, respectively, depending whether heads of neighbourhood units or health volunteers were used to identify deaths [27].
Siblings’ methods include the indirect and direct sisterhood method. The indirect method asks sisters about all ever-married sisters (born to the same mother), whether they are still alive or have died and whether the dead sisters died while they were pregnant, or during childbirth, or six weeks post-partum. The MM-ratio is approximated by the lifetime-risk of death calculated from age specific sister units of risk exposure and the reported pregnancy-related deaths and the total fertility rate of the population [28]5.
The data requirement for the direct sisterhood method are more demanding and respondents are asked to provide a birth history of her mother, including the current age of all living siblings and the age at death and years since death for all deceased siblings. The information allows the calculation of the MM-rate from which the MM-ratio can be calculated when the general fertility rate of the same population is available [23].
Each of these methodologies has inherent weaknesses in estimating the true level of maternal mortality. Civil registration with death certification by a medical professional is seen as the best method to ascertain maternal deaths [29, 30]. Misclassification and underreporting of maternal deaths is common; even in countries with good quality and complete vital registration. Often documentation of indirect, associated, and late maternal deaths is incomplete [31-33].
5 (1-Lifetime risk) =(1-MM-ratio)Total fertility rate
28 In low and middle income countries, maternal mortality estimates are often derived from household surveys, predominantly DHS and more recently also from Multiple Cluster Information Surveys [34]. Both household surveys use the direct sisterhood methods. A major disadvantage of the direct sisterhood method as used in the DHS (but also the indirect method) is that estimates are produced for a time in the past and cover commonly a period of five to ten years prior to the survey.
Although estimates based on the indirect and direct sisterhood methods are constrained by the fact that they i) estimate pregnancy-related mortality ratio but not the true MM-ratio, ii) suffer from large confidence intervals because a limited numbers of households are generally included and iii) produce estimates in the past [23], they nevertheless produce population-based estimates. This is in contrast to the 1980s where maternal mortality data were largely only available from health facility and hospital statistics [35]. Facility statistics are prone to be incomplete in places where most women deliver and die at home. The selection of the population utilizing the services affect the representativeness of the levels and pattern of maternal mortality of facility-based data [12].
Country based estimates are regularly published since a decade by WHO and partners [3, 24, 36, 37]. In 2010 and 2011 IHME also published estimates [2, 4]. The estimates are based on available country data such as vital registration where judged complete, the DHS estimates on maternal mortality or other national sources such as special surveys. Data are adjusted for underreporting and misclassification. While the latest estimates published by the WHO and partners and the IHME have been relatively consistent for mortality estimates at the global and regional level, this is not so at the country level [1, 2]. Figure 2 compares the estimates from WHO and partners and IHME for 2008 and the revised estimates from the same groups for 2010/2011 [1-4].
29 Figure 2: Comparison of MM-ratios from WHO and IHME for selected countries
The differences seen between the estimates provided by the four publications highlight the problems in getting reliable figures on maternal mortality from countries where estimates rely on surveys using the sibling methods.
Reasons discussed for the variations between estimates provided by the two groups are:
differences in adjustment for misclassification and under-reporting; different adjustments for HIV; and the use of sub-national data by IHME. Moreover, differences in life tables used to calculate the proportion of maternal deaths occurring in women of reproductive age due to maternal causes (PMDF) explains some of the variations [29].
Some countries have used census data to obtain information on pregnancy-related mortality [38, 39]. Although this approach decreases the uncertainty around the maternal mortality estimates introduced through sampling, incompleteness of data on both births and deaths is a concern and major adjustments might be needed [8, 40, 41]. Another approach increasingly used is sample vital registration with verbal autopsy method. India and China have established such systems to monitor maternal mortality [42, 43].
Burkina Cameroon
0 200 400 600 800 1000 1200
MM-ratio WHO & partner
MM-ratio IHME
Comparison of data from WHO and partners and IHME
for 2008
WHO et al. 2010 & Hogan et al. Lancet 2010
Comparison of data from WHO and partners and IHME
for 2010/2011
WHO et al. 2012 & Lozano et al. Lancet 2011 Burkina
0 200 400 600 800 1000 1200
MM-ratio WHO & partner
MM-ratio IHME
30 Methodological aspects of ascertainment of causes of death including maternal
deaths
Ideally, a medical professional captures maternal mortality and causes of death through vital registration systems and death certification. However, in the majority of countries where maternal mortality is high, vital registration does not exist or is incomplete [1]. In some of these countries, verbal autopsies of deaths ascertained from demographic sites are used.
Relatives or caretakers of the deceased person are interviewed to investigate symptoms, signs, and circumstances of the death. The information is then interpreted and analysed to assign a probable cause of death. This procedure is based on the assumption that most causes of death can be distinguished by their signs and symptoms, and that these can be accurately recognised, recalled, and reported by lay respondents. However, the validity of the approach is influenced by several factors including the type of illness leading to the death, the design and content of questionnaires, and how collected information is analysed [44].
Standardized questionnaires are used for verbal autopsies that mostly include a combination of an open-ended and a checklist-based section to investigate the cause of death. WHO suggests the use of three different questionnaires for three age groups: 1) 0-4 weeks (the neonatal period); 2) 4 weeks to 14 years; and 3) 15 years and above [44-48]. Most verbal autopsy classification systems are based on the organ-based classification system of the tenth revision of the International Classification of Diseases[19].
There has been much debate around the methods for assigning the cause of death. The verbal autopsy questionnaires can be interpreted: through physician review; by using a pre-defined or expert algorithm; or through a data-derived algorithm based on logistic regression, neural networks, decision trees, or probability density [49]. In addition, Murray et al. described another method called the “Symptom Pattern Method” [50].
In many sites, physician review is used. Commonly, two physicians independently assign a cause of death using the information collected during the interviews and a third physician might be consulted if the causes identified by the physicians differ [44]. Inter-observer reliability is often moderate. Also physician review is criticised because of high costs and use of scarce resources of highly trained professional [44].
Several validation studies have been done to cross-validate physician review based on verbal autopsy questionnaires with standard diagnosis obtained from hospital records. Reasonable sensitivity and specificity with in a positive predictive value of over 50% has been shown for most diseases including direct maternal causes of death [45, 51-53]. Diseases with distinct
31 features like injuries or tetanus, and also direct maternal causes of death, have higher validity than some infectious diseases, particularly in areas where malaria is very common [54]. A multisite validation study based on 12,542 verbal autopsies that employed defined clinical diagnostic criteria as a gold standard estimated that around 50% of individual cause of death assignments using the physician review were in concordance with the cause of deaths based on clinical records [55]. A problem associated with physician review in some places is the high number of unresolved cases when more than one physician is assigning the cause of death [56]
and the low repeatability [57].
Expert-based algorithms coding is based on algorithms using signs and symptoms elaborated by health experts. Byass developed this method further using a Bayesian approach to define the probability of a given cause of death based on the presence of a particular symptom or sign (also called InterVA) [58]. Relatively high agreement of 40-80% comparing InterVA with physician-review-based cause of deaths assignment was reported using this method in several studies [59-62]. A recent large multi-country study with 12,542 verbal autopsy cases comparing the cause of death ascertainment using the InterVA method with a defined clinical diagnostic gold standard suggested that InterVA is inferior to physician review. On the individual level, the InterVA achieved concordance in only about 25% of cases, less than the physician review where concordance of 50% was reported [63]. However, this study also had weaknesses which could be responsible for the low performance of InterVA. Firstly the study only provided information on the relationship between cause of deaths assigned using InterVA and hospital diagnosis on deaths for which a diagnosis based on strict clinical diagnostic criteria from tertiary health facilities was available. These deaths were thus a highly selected set of causes of deaths with a large proportion of deaths due to non-communicable diseases.
Further, InterVA was not conceived to identify many of the relatively specific causes of deaths but only a limited number of causes of deaths.
The latest version, the InterVA-4 tool has taken several shortcomings of the previous version into account and includes now 62 different cause of death groups and is based on the revised 1012 WHO verbal autopsy tool [64]. Certain apparent over- and under-diagnosis and lack of differentiations between various cancers have been taken into consideration.
All validation studies face the difficulty that the reference standard or the gold standard used is unlikely to reflect the causes of deaths in the overall population, as they rely on death certificates from hospitals. Hospital deaths do not represent deaths in populations where most deaths occur at home. Also, information bias might be introduced by using hospital audits as a
32 reference standard. The caregivers interviewed might recall important signs and symptoms, or even a diagnostic test and a diagnosis given by health providers. In contrast, if the death was at home, no such information might be available. Although the afore mentioned large validation study tried to overcome some of these obstacles [55, 63], limitations have to be considered, particularly regarding the extent to which the hospital reference data reflect the true variety of signs, symptoms, and causes of disease in the population [44, 51, 65].
Studies on maternal mortality using verbal autopsies
Few studies have used verbal autopsy to investigate the causes of maternal deaths [57, 66-73].
Setel et al estimated a positive predictive value of direct and indirect maternal death of 60%
and 50% respectively, comparing 1,912 deaths in Tanzania using hospital notes and diagnosis as a reference [53]. Another multi-centre validation study indicated a positive predictive value of 70% and 67% for institutional direct and indirect maternal deaths, respectively, using hospital notes as a standard [57]. The Bayesian approach, as described above, has also been used to assign causes of maternal death. Results from 258 verbal autopsies of women of reproductive age from Burkina Faso indicated more cases of pregnancy-related sepsis compared to physician review, but fewer non-pregnancy-related infections and HIV-related deaths. For other causes, the results were broadly comparable (no figures available)[72]. In another setting in Burkina Faso, only 7% of pregnancy-related deaths due to haemorrhage were reported, whereas pregnancy-related sepsis was the most predominant cause of mortality (30%) when using the Bayesian approach [73]. However, the study lacked a comparison with health-facility-based assignment of cause of death for further validation.
33