Introduction to the analysis of maternal mortality

Background

Maternal mortality has long been a focus of national health services, but its salience has increased over the last two decades with the establishment of quantitative goals. During that time, the international community has periodically established targets for the reduction of maternal mortality, measured as the Maternal Mortality Ratio (MMR), maternal deaths per 100,000 live births. The World Summit for Children in 1990 set the goal of reducing MMR by half between 1990 and 2000. The 1994 International Conference on Population and Development (ICPD) reiterated this goal, but set the additional longer-term target of reducing the rate by a further half by 2015. The Millennium Summit in 2000 adopted the ICPD target for the fifth MDG (the improvement of maternal health). The target was thus to reduce the MMR by three-quarters between 1990 and 2015. The 2011 report of the Commission on Information and Accountability for Women’s and Children’s Health [1], established by the Secretary-General of the United Nations,  reaffirmed the importance of timely reporting on MMR as one of 11 indicators of maternal, newborn, and child health. It is thus clear that the measurement of maternal mortality has a very high priority. This section discusses broad options for such measurement.

Definition

The International Classification of Diseases Revision 10 (ICD-10) [2] defines a maternal death as follows. "A maternal death is defined 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." Measuring maternal deaths thus involves the determination of cause of death, an issue not addressed elsewhere in this manual. Maternal deaths are divided into direct obstetric deaths (of which the major specific causes are haemorrhage, obstructed labour, eclampsia, sepsis and consequences of abortion) and indirect obstetric deaths (pregnancy-related deaths among women with a pre-existing or newly developed health problem exacerbated by the pregnancy or delivery).

The measurement of maternal mortality represents a major problem for countries lacking largely complete birth and death registration (Graham, Ahmed, Stanton et al. 2008) not only because deaths are not recorded but also because of the need to ascertain cause of death (see for example Mathers, Fat, Inoue et al. (2005)). Cause of death is best determined by a physician present close to the time of death, but many deaths occur without the presence of a doctor. Further, even when a doctor does certify the death, deaths that occur outside of the labour ward may be incorrectly ascribed to a non-maternal cause. Some progress has been made in recent years with the development and application of verbal autopsy methods, whereby family members are asked to report signs and symptoms surrounding the death, but there is still considerable controversy about how well such methods work (Chandramohan, Rodrigues, Maude et al. 1998). The description of verbal autopsy instruments and analysis is beyond the scope of this Manual.

In part because of the difficulty of identifying true maternal deaths, ICD-10 also defines a pregnancy-related death as one that occurs during pregnancy, delivery or the 42 days after the end of the pregnancy, regardless of cause of death. The category pregnancy-related death thus includes all maternal deaths plus the accidental or incidental deaths excluded from the category “maternal”. The advantage of the pregnancy-related category is that it appears to be easier to implement; it only requires information on the timing of death relative to a pregnancy, without specific knowledge of true cause of death. The disadvantages are that measures of pregnancy-related mortality are frequently misinterpreted as measures of maternal mortality, and that any trend in pregnancy-related mortality not due to maternal causes will limit the value of the measure for tracking impact of maternal health interventions. Demographic surveys generally measure pregnancy-related deaths and avoid the necessity for cause ascertainment.

There is active debate, and no consensus, as to the relationship that exists in practice between a reported number of pregnancy-related deaths and some unknown true number of maternal deaths. Clearly by definition the true number of pregnancy-related deaths has to be the same as, or larger than, the true number of maternal deaths, since all maternal deaths are pregnancy-related, but not all pregnancy-related deaths are maternal. In practice, however, the situation is less clear-cut because of possible reporting errors. There are those who argue that reported pregnancy-related deaths exceed true maternal deaths (Garenne, McCaa and Nacro 2008; Stecklov 1995) whereas others argue that pregnancy-related deaths are likely to be under-reported because, for example, a respondent may not have known that the deceased was pregnant at the time of death (Shahidullah 1995; Wilmoth 2009). The relationship could, therefore, go either way. This manual does not come down firmly in support of either of these views, but instead emphasizes that good practice requires that measures be labelled correctly. Thus a measure based on reported pregnancy-related deaths should be reported as a measure of pregnancy-related mortality, whereas a measure based on what are thought to be true maternal deaths (identified by a verbal autopsy for example) should be reported as a measure of maternal mortality.

Basic measures of maternal mortality

There are two common measures of maternal mortality (and corresponding measures for pregnancy-related mortality). They are the Maternal Mortality Ratio (MMR), the target for MDG-5, and the Maternal Mortality Rate (here abbreviated as MMRate). There are two other measures that will often be encountered: the proportion of deaths of women of reproductive age that are maternal (often abbreviated as PMDF), and the lifetime risk of dying a maternal death (LTR). The latter measure is used primarily for advocacy purposes.

Maternal Mortality Ratio

 The MMR is the number of maternal deaths in a period per 100,000 live births in the same period. Note the use of live births rather than pregnancies in the denominator. The MMR is primarily a measure of obstetric risk, roughly the risk of dying per 100,000 risky events.

Maternal Mortality Rate

The MMRate is a cause-specific mortality rate. It is the number of maternal deaths in a period per 1,000 person-years lived by the female population of reproductive age (usually ages 15-49).

The MMR and MMRate share a numerator, and have a simple relationship to one another:

$$MMR\text{}=\frac{MD}{LB}\times 100,000=\frac{MD}{FPRA}\times 1,000\times \frac{FPRA}{LB}\times 100=MMRate\times \frac{100,000}{GFR}$$

where, for a given time period, MD is maternal deaths, LB is live births, FPRA is the person-years lived by the female population of reproductive age, and GFR is the General Fertility Rate expressed per 1,000 women of reproductive age.

Proportion Maternal of Deaths of Women of Reproductive Age

The PMDF is MD/FDRA, where FDRA is the number of deaths of women of reproductive age. It is used primarily in modeling exercises (for example, Hill, Thomas, AbouZahr et al. (2007), Hogan, Foreman, Naghavi et al. (2010), Wilmoth, Zureick, Mizoguchi et al. (2010) and Wilmoth, Mizoguchi, Oestergaard et al. (2012)) but is also of some value for data quality assessment (see below).

Lifetime Risk

The LTR is usually implemented as the risk of dying from a maternal cause from age 15 onwards. Wilmoth (2009) suggests calculating the measure per 1,000 women reaching age 15; i.e., as

$$LTR=\frac{(T_{15}-T_{50})}{l_{15}}\times MMRate$$

where T₁₅ and T₅₀ are the person-years lived above ages 15 and 50 respectively, and l₁₅ is the survivors to age 15, in an appropriate life table for the population in question.

 

Each of the four measures above has a pregnancy-related corollary, calculated by replacing maternal deaths by pregnancy-related deaths.

Data sources

Other than civil registration, there are two widely used approaches to the collection of data needed to measure pregnancy-related mortality: the full sibling history (FSH); and a large household survey or census that collects data on recent household deaths (HSHD). The summary sibling history (Graham, Brass and Snow 1989) is now rarely used, partly because it produces estimates that represent averages over very long time frames.

The full sibling history

A full sibling history (FSH) involves complex and detailed data collection, requiring very careful training and supervision of field staff to be executed correctly. It is therefore not an appropriate methodology to include in a census. The FSH has been widely included as the “Maternal Mortality Module” in DHSs since 1991, and has also been included in some other household surveys. The FSH collects information from eligible respondents. In most DHSs, women eligible for the birth history are also those eligible for the FSH, but some surveys have also collected FSHs from eligible male respondents. Information is collected about all brothers and sisters born to the same mother. The FSH can thus be thought of as the respondent’s mother’s full birth history, excluding the respondent herself (or himself). In the DHS, the information collected about each sibling is: name; sex; whether still alive; if still alive, age in years; if dead, how many years ago did the sibling die and how old was he or she at death. For deaths of women of reproductive age, additional questions enquire whether the sister died (i) while pregnant; (ii) during childbirth; or (iii) within 42 days or 2 months of the end of a pregnancy.

It will be clear from the above that measures calculated from an FSH are of pregnancy-related mortality, not maternal mortality. The sibling history does not lend itself to the application of a verbal autopsy (which would be necessary for calculating maternal mortality), because a sister of reproductive age reported to have died may well have died in a different household than the respondent, who thus may have little direct knowledge of signs and symptoms preceding the death. It is generally not feasible to try to identify the household where the death occurred and conduct a verbal autopsy with a member of that household.

The FSH provides information on pregnancy-related deaths and female exposure, and thus a basis for estimating pregnancy-related mortality rates. If pregnancy-related mortality ratios (PRMRs) are to be calculated, information must also be available on live births. A typical DHS collects a full birth history (FBH) as well as an FSH, so this is usually not a problem.

The summary sibling history

The use of information on sibling survival to estimate maternal mortality was first proposed by Graham, Brass and Snow (1989). They proposed using a summary sibling history. Such a summary history collects information by sex on the aggregate number of siblings the respondent had, the number who survived to age 15 (or first marriage), and – for sisters who died after age 15 – whether they were pregnant, in childbirth, or in the 42 days post-partum when they died. This method is not recommended for use. The sisters of a respondent can differ in age from the respondent herself by plus or minus 30 years, with the result that the deaths of sisters can be spread over a very long time period prior to a survey. Reference dates of maternal mortality estimates derived from summary sibling histories are thus located well in the past (on average as much as 12 years before the survey), making them of limited practical value. As a consequence the method will not be described further.

The census or large household survey collecting data on recent household deaths

Censuses in the 1970s and 1980s in countries lacking complete civil registration often collected information on recent household deaths, usually those that occurred in the last 12 months. Concerns about data quality limited the use of such questions in the 1990 and 2000 rounds of censuses, but renewed interest in adult mortality and specifically in maternal mortality led to a sharp increase in their use in the 2010 round. A common format for such questions is to ask whether any usual household member died in the preceding 12 months (sometimes the question refers to a different period such as the time since a fixed date or memorable event). If the answer is yes, the deceased’s name, sex and age at death are recorded. If the death is of a woman of reproductive age, an additional question or questions about the timing of death relative to pregnancy are asked, namely did the deceased die while pregnant, during delivery, or in the 6 weeks (sometimes 2 months) after the end of the pregnancy? The methodology is reviewed by Stanton, Hobcraft, Hill et al. (2001), and experience with it is reviewed by Hill, Queiroz, Stanton et al. (2007) and Hill, Queiroz, Wong et al. (2009).

As generally used, these questions on recent deaths identify pregnancy-related deaths. However, some attempts have been made to follow up reported deaths of women of reproductive age (or a sample of such deaths) with a verbal autopsy to identify true maternal deaths. This has been done in a number of very large household sample surveys (e.g. in the Bangladesh Maternal Morbidity and Mortality Survey 2001 (Hill, El-Arifeen, Koenig et al. 2006), but also following at least two censuses (1986 in Iran and 2007 in Mozambique).

A census or large household survey that collects information on recent household deaths will always record a household roster by age and sex. This roster provides information on denominators for pregnancy-related mortality rates (PRMRates). Additional information on recent fertility will be needed to calculate PRMRates. This will usually be collected in the form of a question for women of reproductive age as to whether they had a live birth in the year before the survey or a question on the date of each woman’s most recent live birth. Information on life-time fertility should also be collected to permit the evaluation and possible adjustment of the data on fertility (see the section on fertility estimation using a relational Gompertz model [3]).

Data collected at health facilities

A major expense of household surveys is the cost of getting an interviewer to the (correct) household. Much of this expense can be eliminated by taking advantage of respondents coming to the interviewer, such as to give birth in a health facility. Health facilities are also likely to record births and deaths and cause of death that occur at the facility as part of a routine health management information system. However, the problem with such data is selection bias: we can never be sure that the women who give birth in a facility are representative of all mothers. To improve coverage, experiments are being conducted to find out whether health extension workers or the equivalent working in communities can collect adequate data on births and deaths. Such an approach is akin to a sample registration system.

An analysis strategy for facility data has been proposed but not implemented. Starting with the assumption that women who deliver in a health facility (or visit a health facility for some other pregnancy-related condition) are a biased sample of all mothers (it is not clear which way the bias will run, whether women having fewer pregnancy-related health issues or those having more will predominate), if one could estimate the selection probabilities correctly, the statistics collected could be adjusted for bias. For example, women on visiting a health facility could be asked their age, children ever born and children still alive, plus a number of additional questions about their socio-economic condition. The children ever born and children still alive could then be modelled onto the socio-economic structure of the whole population, available for example from a population census. To our knowledge, this approach has never been tested.

 

References

Chandramohan D, LC Rodrigues, GH Maude and RJ Hayes. 1998. "The validity of verbal autopsies for assessing the causes of institutional maternal death", Studies in Family Planning 29(4):414-422. doi: http://dx.doi.org/10.2307/172253 [4]

Garenne M, R McCaa and K Nacro. 2008. "Maternal mortality in South Africa in 2001: From demographic census to epidemiological investigation", Population Health Metrics 6:4. doi: http://dx.doi.org/10.1186/1478-7954-6-4 [5]

Graham W, S Ahmed, C Stanton, C Abou-Zahr and O Campbell. 2008. "Measuring maternal mortality: An overview of opportunities and options for developing countries", BMC Medicine 6:12. doi: http://dx.doi.org/10.1186/1741-7015-6-12 [6]

Graham W, W Brass and RW Snow. 1989. "Estimating maternal mortality: The sisterhood method", Studies in Family Planning 20(3):125-135. doi: http://dx.doi.org/doi:10.2307/1966567 [7]

Hill K, S El-Arifeen, M Koenig, A Al-Sabir, K Jamil and H Raggers. 2006. "How should we measure maternal mortality in the developing world? A comparison of household deaths and sibling history approaches", Bulletin of the World Health Organization 84(3):173-180. doi: http://dx.doi.org/10.2471/BLT.05.027714 [8]

Hill K, BL Queiroz, C Stanton and C AbouZahr. 2007. "Measuring maternal mortality via the population census: Experience from Africa," Paper presented at 5th African Population Conference. Arusha, Tanzania, 10-14 December 2007.

Hill K, BL Queiroz, L Wong, J Plata et al. 2009. "Estimating pregnancy-related mortality from census data: Experience in Latin America", Bulletin of the World Health Organization 87(4):288-295. doi: http://dx.doi.org/10.2471/BLT.08.052233 [9]

Hill K, K Thomas, C AbouZahr, N Walker et al. 2007. "Estimates of maternal mortality worldwide between 1990 and 2005: An assessment of available data", The Lancet 370(9595):1311-1319. doi: http://dx.doi.org/10.1016/S0140-6736(07)61572-4 [10]

Hogan MC, KJ Foreman, M Naghavi, SY Ahn et al. 2010. "Maternal mortality for 181 countries, 1980-2008: A systematic analysis of progress towards Millennium Development Goal 5", The Lancet 375(9726):1609-1623. doi: http://dx.doi.org/10.1016/S0140-6736(10)60518-1 [11]

Mathers CD, DM Fat, M Inoue, C Rao and AD Lopez. 2005. "Counting the dead and what they died from: An assessment of the global status of cause of death data", Bulletin of the World Health Organization 83(3):171-177.

Shahidullah M. 1995. "The sisterhood method of estimating maternal mortality: The Matlab experience", Studies in Family Planning 26(2):101-106. doi: http://dx.doi.org/10.2307/2137935 [12]

Stanton C, J Hobcraft, K Hill, N Kodjogbé et al. 2001. "Every death counts: Measurement of maternal mortality via a census", Bulletin of the World Health Organization 79(7):657-664.

Stecklov G. 1995. "Maternal mortality estimation: Separating pregnancy-related and non-pregnancy-related risks", Studies in Family Planning 26(1):33-38. doi: http://dx.doi.org/10.2307/2138049 [13]

Wilmoth J, S Zureick, N Mizoguchi, M Inoue and M Oestergaard. 2010. Levels and Trends of Maternal Mortality in the World: The Development of New Estimates by the United Nations. Geneva: WHO, UNICEF, UNFPA and the World Bank. http://www.who.int/reproductivehealth/publications/monitoring/MMR_technical_report.pdf [14]

Wilmoth JR. 2009. "The lifetime risk of maternal mortality: Concept and measurement", Bulletin of the World Health Organization 87(4):256-262. doi: http://dx.doi.org/10.2471/BLT.07.048280 [15]

Wilmoth JR, N Mizoguchi, MZ Oestergaard, L Say et al. 2012. "A new method for deriving global estimates of maternal mortality", Statistics, Politics, and Policy 3(2):Article 3. doi: http://dx.doi.org/10.1515/2151-7509.1038 [16]

Estimation of pregnancy-related mortality from survival of siblings

Description of method

The use of full sibling history data to estimate overall adult mortality, and the evaluation of such data, is described elsewhere [17]. A full sibling history is analogous to a full birth history: a respondent (usually a female of reproductive age) is asked about each of her siblings born of the same mother. For surviving siblings, sex and age in completed years are recorded; for dead siblings, sex, age at death in completed years and calendar year of death are recorded. The resulting history is, with the addition of the respondent herself, the full birth history of the mother. As with a full birth history, the sibling history allows events (deaths) and exposure time to be arranged in calendar time, and hence the calculation of age-period mortality rates. For pregnancy-related mortality, further information is collected concerning deaths of sisters of reproductive age as to whether the sister died during pregnancy, during delivery, or in the 42 days (or sometimes 2 months) post-partum. This section focuses on analyzing information on sisters of reproductive age.

One important issue discussed under full sibling histories will be touched on here. The DHS full sibling history asks respondents for the full birth history of their natural mother, excluding themselves. As a result, there is the potential for multiple responses about the same individual. For example, if two daughters of the same mother are interviewed in the same household, there will be multiple reports about other members of the sibship. The DHS bases events and exposure time entirely on reported siblings, not taking into account the exposure time of the (surviving) respondent hersel. The events and exposure time of siblings are weighted only by the respondent’s sample weight, not taking into account numbers of surviving potential respondents in the sibship. Trussell and Rodriguez (1990) show that if there is no correlation between mortality risks within sibships, this calculation gives an unbiased estimate of overall mortality. Gakidou and King (2006) argue that sibships should include the surviving respondent and should in addition be weighted by the likelihood that they will be reported – that is, by the inverse of the number of potential respondents in the sibship. They also argue that an adjustment should be made for sibships that go unreported because no member remains alive. In a multi-survey analysis of DHS FSH’s, Obermeyer, Rajaratnam, Park et al. (2010) estimate that the effect of not adjusting for likelihood of reporting can bias overall mortality estimates downwards by 20 percent or so. Masquelier (2012) however argues that the Obermeyer et al. analysis took into account all surviving siblings, not just potential respondents, and as a result exaggerated the size of any bias. Masquelier recommends using the DHS calculation approach, for reasons outlined below (see also Sibling History Analysis [17]).

Data requirements and assumptions

Important assumptions

• No correlation exists between mortality risks of women and size of sibship
• There are no selection effects resulting from migration

Tabulations of data required

• Number of women, by five-year age group from household questionnaire.
• Number of sister deaths by time period (typically 0-6 years) before the survey by five-year age group and by whether during pregnancy, delivery or 42 days/2 months post-partum.
• Number of sister-years of exposure by time period by five-year age group of respondent.
• Age-specific fertility rates and General Fertility Rate for the time period.

We will assume here that the DHS approach is followed. The extraction of summary data from the DHS is beyond the scope of this manual. A standard program exists in the free software CSPro to produce this tabulation from the basic data. A decision that has to be made at the outset of the analysis is the time frame to be used. Most DHSs create the basic table for events and exposure in the seven (0 to 6) years prior to the survey, but sometimes the period used is five years, and sometimes ten years. One consideration in choosing one period rather than another is sampling errors; in a small sample with quite low adult mortality, the period may have to be ten years to avoid very large sampling uncertainty, whereas with a large sample and higher mortality, the period may be reduced to five years. This aspect of the method is discussed further in the section on interpretation of results.

Preparatory work and preliminary investigations

Data quality assessment for a full sibling history [17] and for recent age-specific fertility rates [18] are described elsewhere. The only data quality assessment specific to the estimation of Pregnancy-Related Mortality is of the information on the proportion of deaths of women of reproductive age that are pregnancy-related, (PPRD), and the proportions of pregnancy-related deaths that occur during pregnancy, during delivery and in the 42 days (or two months) post-partum.

No formal methods exist for carrying out such assessments. However, the age pattern of the proportion of deaths that are pregnancy-related should resemble the age distribution of age-specific fertility, since it is births that are the risky events. Risks of pregnancy-related death are generally thought to be rather higher at the extremes of the reproductive age period, so the tails of the proportions pregnancy-related may be somewhat “fatter” than for age-specific fertility.

Caveats and warnings

It is widely believed that sibling histories tend to under-report mortality, particularly deaths further in the past. One should thus not attempt to interpret trends over time in pregnancy-related mortality from a single data set. Such attempts across data sets would also need to take into account the wide confidence intervals surrounding estimates even for a seven-year window.

Application of method

Step 1: Extract tabulations of the data

As mentioned earlier, software is readily available to extract the tabulations required relating to reported deaths of sisters, their exposure and pregnancy-related deaths. In addition, estimates of age-specific fertility are required. (If the data on sisters are extracted from a DHS, the approach to estimating fertility rates directly from the data is described elsewhere [19] in this manual). Finally, an estimate of the female population by age group enumerated in the household survey is required. We define the following terms

 

${}_5{D}_x^s$

- the reported number of sisters reported dying between ages x and x+5

 

${}_{5}P{Y}_x^s$

- the number of person-years lived by sisters between ages x and x+5

 

${}_{5}PR{D}_x^s$

- the number of pregnancy-related deaths of sisters between ages x and x+5

 

${}_5{f}_x^{}$

- age-specific fertility rates of women aged x to x+5

 

${}_5{N}_x^f$

- the population of women aged between x and x+5 as enumerated in the survey

Tabulations of each of the above five quantities are required to implement the method.

Step 2: Derive measures of mortality

The age-specific mortality rate is derived by dividing the reported deaths of sisters by the number of person years lived by those sisters in each age group,

$${}_5{M}_x=\frac{{}_5{D}_x^s}{{}_{5}P{Y}_x^s}$$

Equation 1

The age-specific pregnancy-related mortality rate in each age group is given by

$${}_{5}PRMRat{e}_x=\frac{{}_{5}PR{D}_x^s}{{}_{5}P{Y}_x^s}$$

Equation 2

The proportion of deaths that are pregnancy-related is

$${}_{5}PPR{D}_x=\frac{{}_{5}PR{D}_x^s}{{}_5{D}_x^s}$$

Equation 3

And the age-specific pregnancy-related mortality ratio is

$${}_{5}PRMRati{o}_x=100,000.\left(\frac{{}_{5}PR{D}_x^s}{{}_{5}P{Y}_x^s{.}_5{f}_x}\right)$$

Equation 4

Step 3: Estimate population-level measures

Estimates of the pregnancy-related mortality rate, and the proportion of deaths that are pregnancy-related in the population of women of reproductive age (taken here as those aged 15-49) are derived by weighting the age-specific rates derived above by the number of women aged 15-49 in the households surveyed. Thus

$${}_{35}PRMRat{e}_{15}=\frac{{\displaystyle \sum _{x=15,5}^{45}{}_{5}PRMRat{e}_x{.}_5{N}_x^f}}{{\displaystyle \sum _{x=15,5}^{45}{}_5{N}_x^f}}$$

Equation 5

and

$${}_{35}PPR{D}_{15}=\frac{{\displaystyle \sum _{x=15,5}^{45}{}_{5}PPR{D}_x{.}_5{N}_x^f}}{{\displaystyle \sum _{x=15,5}^{45}{}_5{N}_x^f}}$$

Equation 6

and

${}_{35}PRMRati{o}_{15}=100,000.\frac{{\displaystyle \sum _{x=15,5}^{45}{}_{5}PRMRat{e}_x.{}_{5}N{}_x^f}}{{\displaystyle \sum _{x=15,5}^{45}{}_{5}f{}_x.{}_{5}N{}_x^f}}$

Equation 7

 

Worked example

The application of the method is illustrated using data collected in the 2004 Malawi DHS. Women were asked about the survival, and - if dead - timing of death relative to pregnancy, of sisters.

Step 1: Extract tabulations of the data

Tabulations of the required input data are presented in Table 1. The tabulations of reports of sisters’ deaths and exposure are for the seven years before the survey.

Table 1 Input data used to estimate pregnancy-related mortality, Malawi 2004 DHS

Age group	Sister deaths	Sister exposure	Pregnancy-related deaths	Age-specific fertility	Household female population
	(i)	(ii)	(iii)	(iv)	(v)
15-19	117	27,622	8	0.162	2,570
20-24	227	29,331	52	0.293	3,036
25-29	299	23,763	64	0.254	2,247
30-34	245	17,228	50	0.222	1,516
35-39	230	12,206	33	0.163	1,122
40-44	177	7,892	22	0.080	970
45-49	82	4,574	12	0.035	743
Total	1376	122,616	240	0.204*	12,204
* General Fertility Rate calculated as age-specific fertility rates weighted by age distribution of the female household population
Source: Malawi DHS 2004 Tables 13.2 (p.245) and 13.3 (p.247)

Step 2: Derive measures of mortality

Application of the method is shown in Table 2, using the data presented in Table 1. The first column of Table 2 shows age-specific mortality rates, calculated by dividing female deaths by age (col. i of Table 1) by sisters’ exposure in years (col. ii of Table 1) as described by Equation 1. The second column shows age-specific pregnancy-related mortality rates, calculated as for the all-cause age-specific rates but using pregnancy-related deaths only (col. iii of Table 1) in the numerator (Equation 2). Column 3 shows the age-specific proportions of female deaths that were reported to be pregnancy-related (col. iii of Table 1 divided by col. i of Table 1 - Equation 3). Column 4 shows age-specific pregnancy-related mortality ratios, calculated by dividing pregnancy-related deaths (col. iii of Table 1) by the product of sisters' exposure in years and the age-specific fertility rates (col. ii and col. iv) and multiplying by 100,000 (Equation 4).

Table 2 Adult female mortality rates and pregnancy-related mortality rates, Malawi, 2004 DHS

Age Group	Age-Specific Mortality Rate	Age-Specific Pregnancy-Related Mortality Rate	Proportion of Deaths Pregnancy-Related	Age-Specific Pregnancy-Related Mortality Ratio
	= 1000*(i)/(ii)	= 1000*(iii)/(ii)	= (iii)/(i)	= 100000* iii /(ii*iv)
15-19	4.24	0.29	0.0684	178.8
20-24	7.74	1.77	0.2291	605.1
25-29	12.58	2.69	0.2140	1060.3
30-34	14.22	2.90	0.2041	1307.3
35-39	18.84	2.70	0.1435	1658.6
40-44	22.43	2.79	0.1243	3484.5
45-49	17.93	2.62	0.1463	7495.8
Total*	11.51	1.99	0.1681	970.7
* Note: all the totals in this table are standardized onto the age distribution of the female household population (see text).

It is important to note that the entries in the Total row in Table 2 are not calculated by summing events and exposure across age groups. The reason for this is that the age pattern of sister exposure is not the same as the age pattern of the female population of reproductive age. To obtain valid population-level estimates of totals, it is necessary to re-weight the age-specific rates and ratios in Table 2 by the proportional female age distribution in col. v of Table 1, as described by Equations 5, 6 and 7. The denominator in Equation 7 is also not the General Fertility Rate as normally calculated (births divided by female population aged 15 to 49) but rather the age-distribution weighted sum of age-specific fertility rates.

Step 3: Estimate population-level measures

Table 3 compares the proportion of births in each 5-year age group (calculated as the product of the household female population and the age-specific fertility) to the proportion of PRD (calculated as the product of the household female population and the age-specific PRMRate) in each age group.

Table 3 Comparison of proportions of births and pregnancy-related deaths by age group: Malawi DHS 2004

Age	Female household population	Age-specific fertility	Pregnancy-related deaths	Births	Proportion of :
					Pregnancy-related deaths	Births
15-19	2,570	162	8	416	0.033	0.167
20-24	3,036	293	52	890	0.217	0.356
25-29	2,247	254	64	571	0.267	0.228
30-34	1,516	222	50	337	0.208	0.135
35-39	1,122	163	33	183	0.138	0.073
40-44	970	80	22	78	0.092	0.031
45-49	743	35	12	26	0.050	0.010
Total	12,204		240	2,500	1.000	1.000

Diagnostics, analysis and interpretation

Checks and validation

For checking and validating overall estimates of female mortality, see the section of the manual on the analysis of sibling histories [17]. Checking and validating the extra information provided about pregnancy-related mortality depends on plausibility checks. Is the overall proportion of deaths of sisters of reproductive age reported as pregnancy-related plausible, given the estimated pregnancy-related mortality ratio? Is the distribution of pregnancy-related deaths by age plausible given the age pattern of births (the risky events)?

No generally accepted way exists to assess the plausibility of the overall proportion of deaths pregnancy-related. In general, there is a positive association between the proportion of deaths pregnancy-related and the PRMR, but the association hinges on the level of non-pregnancy-related mortality and provides no useful basis for evaluation. The plausibility of the age distribution of pregnancy-related deaths is assessed in comparison to the distribution of births by age, as shown in Table 3. In the case of the 2004 Malawi DHS, the proportions of pregnancy-related deaths contributed by the age groups 15-19 and 20-24 are much lower than the corresponding proportions of births, and the reverse is true over the age of 35. The latter can be plausibly explained by the increasing riskiness of pregnancy and childbirth for women over age 35, but no plausible explanation exists for the large differences under age 25. The suspicion is strong that deaths of sisters under the age of 25 that were actually pregnancy-related have not been reported as such.

Interpretation

Interpretation of estimates of pregnancy-related mortality from sibling histories needs to pay close attention to sampling uncertainty and typical data errors. Sampling uncertainty is very large by comparison with DHS estimates of under-5 mortality. Figure 1 plots coefficients of variation (standard error divided by the estimate) for DHS estimates of PRMRs by number of siblings reported on. The coefficients of variation are above 0.08 even for very large samples, and above 0.10 for all but a handful of surveys.

Figure 1 Coefficients of variation for PRMRs by DHS sample size [20]

 

Sources: Stanton, Abderrahim and Hill (2000) for estimates prior to the year 2000, and DHS country reports for years 2004 and later

Given the large sampling uncertainty, interpretation of sub-national differences or other sub-group differences such as by age is hazardous. Typical data errors, particularly the omission of deaths for time periods more distant from the survey, make any interpretation of trends within data sources questionable. Any conclusions about trends should be based on estimates from two or more surveys for comparable time periods before each survey and taking into account sampling uncertainty.

References

Gakidou E and G King. 2006. "Death by survey: estimating adult mortality without selection bias from sibling survival data", Demography 43(3):569-585. doi: http://dx.doi.org/10.1353/dem.2006.0024 [21]

Masquelier B. 2013. “Adult mortality from sibling survival data: A reappraisal of selection biases?”, Demography 50(1):207–228. doi: http://dx.doi.org/10.1007/s13524-012-0149-1 [22]

Obermeyer Z, JK Rajaratnam, CH Park, E Gakidou et al. 2010. "Measuring adult mortality using sibling survival: a new analytical method and new results for 44 countries, 1974-2006", PLoS Medicine 7(4):e1000260. doi: http://dx.doi.org/10.1371/journal.pmed.1000260 [23]

Stanton C, N Abderrahim and K Hill. 2000. "An assessment of DHS maternal mortality indicators", Studies in Family Planning 31(2):111-123. doi: http://dx.doi.org/10.1111/j.1728-4465.2000.00111.x [24]

Trussell J and G Rodriguez. 1990. "A note on the sisterhood estimator of maternal mortality", Studies in Family Planning 21(6):344-346. doi: http://dx.doi.org/10.2307/1966923 [25]

Estimation of pregnancy-related mortality from deaths reported by households

Description of method

If questions are asked in a census or large household survey about deaths in the household in a reference period, and further questions are asked about the timing relative to pregnancy of deaths of women of childbearing age, it is possible to derive estimates of pregnancy-related mortality. If additional information on cause of death is available, for instance from a verbal autopsy enquiry, it is possible to estimate maternal mortality, but this is quite unusual and will not be covered explicitly here.

The use of census or large survey data on recent household deaths to estimate overall adult mortality, and the evaluation of such data, are described elsewhere [26]. Any error in reporting on deaths is likely to have a proportionately similar effect on estimates of pregnancy-related mortality, so evaluation of data quality, and adjustment if needed, are essential parts of the analysis.

Data requirements and assumptions

Tabulations of data required

• Number of women, by five-year age group from household questionnaire.
• Number of household deaths in the previous 12 months (or similar period) by age and sex.
• For deaths of women of reproductive age (usually 15 to 49), whether the death occurred during pregnancy, delivery or 42 days/2 months post-partum.
• Age-specific fertility rates and General Fertility Rate for the time period.

If sample or design weights have been provided with the data, they must be applied in the appropriate manner in producing input tables.

Important assumptions

• That any effect of household dissolution arising from death of a member is trivial.
• (If adjustment of deaths or births is to be made), that errors in reporting deaths and births are proportional by age.
• That any omission of deaths does not vary with whether or not the deaths are pregnancy-related.
• That reporting of deaths as pregnancy-related is approximately accurate.

Preparatory work and preliminary investigations

Data quality assessment for household deaths and for recent age-specific fertility rates [27] are described elsewhere. The only data quality assessments specific to the estimation of Pregnancy-Related Mortality are of the information on the proportion of deaths of women of reproductive age that are pregnancy-related, PPRD, and the proportions of pregnancy-related deaths that occur during pregnancy, during delivery and in the 42 days (or two months) post-partum.

Potential for data quality assessment for issues other than recording of overall numbers of births and deaths is limited to a plausibility check for reporting of deaths as pregnancy-related. The age pattern of the proportion of deaths that are pregnancy-related should resemble the age distribution of age-specific fertility.

Caveats and warnings

It is widely believed that questions on household deaths and on births in the year before a census or survey often under-report true numbers of events. Careful evaluation of coverage of both types of event is essential. In the unlikely event that under-reporting of both types of event is approximately equal, the pregnancy-related mortality ratio will not be seriously biased, although the pregnancy-related mortality rate will still be biased. However, if data evaluation indicates omission of deaths and births, the data will need to be corrected before a final estimate of pregnancy-related mortality is arrived at.

Application of method

The method is applied in the following stages:

Step 1: Extract tabulations of the data

Instructions as to how to extract the data are outside the scope of this manual. It is usually a simple matter to produce cross-tabulations of the number of women by age group, and the number of deaths of women by age group and timing of death relative to pregnancy (during pregnancy, during delivery, or during the 42 days/2 months post-partum). Where appropriate, the tabulations should be weighted to compensate for under-enumeration (following a post-enumeration survey) and whether the data being analysed are a micro-sample or not. In addition, estimates of age-specific fertility are required. The process of estimating fertility is described elsewhere.

We define the following terms:

 

${}_{5}D{}_x$

- the reported number of female household members dying between ages x and x+5

 

${}_{5}PR{D}_x$

- the number of pregnancy-related deaths of female household members between ages x and x+5

 

${}_5{f}_x^{}$

- age-specific fertility rates of women aged between x and x+5

 

${}_5{N}_x^f$

- the population of women aged between x and x+5 as enumerated in the census or large survey

 

Tabulations of all four of the above variables are required.

Step 2: Derive measures of mortality

The age-specific mortality rate is derived by dividing the reported deaths of women in the household by the number of person years lived by the population in each age group,

$${}_5{M}_x=\frac{{}_5{D}_x^{}}{{}_5{N}_x^f}$$

Equation 1

The age-specific pregnancy-related mortality rate in each age group is given by

$${}_{5}PRMRat{e}_x=\frac{{}_{5}PR{D}_x^{}}{{}_5{N}_x^f}$$

Equation 2

And the age-specific pregnancy-related mortality ratio is

$${}_{5}PRM{R}_x=100,000.\left(\frac{{}_{5}PR{D}_x^{}}{{}_5{N}_x^f{.}_5{f}_x}\right)$$

Equation 3

Finally, the proportion of deaths that are pregnancy-related is

$${}_{5}PPR{D}_x=\frac{{}_{5}PR{D}_x^{}}{{}_5{D}_x^f}.$$

Equation 4

The proportionate distribution of the PRD over ages 15 to 49 is given by

 

$${}_{5}DPR{D}_x=\frac{{}_{5}PR{D}_x^{}}{{\displaystyle \sum _{x=15,5}^{45}{}_{5}PR{D}_x^{}}}$$

Equation 5

Worked example

We use as an example the data from the 2008 Malawi Census, specifically the data from questions concerning deaths in the 12 months before the survey/census.

Step 1: Extract tabulations of the data

Table 1 is tabulated from individual-level data from a 10 per cent IPUMS sample from the Malawi Census [28]. The table shows the female population of reproductive age by five-year age groups, female deaths reported as occurring in the 12 months before the census, and, for deaths of females aged 15 to 49, whether the death occurred during pregnancy, during delivery, or in the 42 days post-partum.

Table 1 Female population of reproductive age 15-49 and deaths by whether pregnancy-related, Malawi 2008 Census

		Female Deaths
Age Group	Female Population	During Pregnancy	During Delivery	Post-Partum	Total Pregnancy-Related	All deaths
	(i)	(ii)	(iii)	(iv)	(v)	(vi)
15-19	67,918	43	25	26	94	235
20-24	69,069	68	40	36	144	389
25-29	57,478	84	31	32	147	442
30-34	41,073	92	24	37	153	471
35-39	29,993	56	15	23	94	346
40-44	22,294	42	4	14	60	238
45-49	17,564	38	3	4	45	185
Total 15-49	310,748	423	142	172	737	2,306
Source: Malawi 2008 Census, 10% sample

Step 2: Derive measures of mortality

Application of the method for data available in the form given in Table 1 is shown in Table 2. All cause age-specific mortality is calculated (col. i) by dividing deaths (col. vi of Table 1) by female population (col. i of Table 1) as per Equation 1. Note that strictly speaking the deaths pertain to a population on average half a year earlier than that recorded, but the error involved in ignoring this complication is trivial and will be included in an adjustment if the deaths are corrected using one of the appropriate death distribution methods [26]. Age-specific pregnancy-related mortality rates are then calculated by dividing pregnancy-related deaths (col. v of Table 1) by female population (col. i of Table 1) as in Equation 2. Age-specific pregnancy-related mortality ratios in col. iv of Table 2 are then obtained by dividing the age-specific pregnancy-related mortality rates (col. ii) by age-specific fertility rates (col. iii, obtained from other sources) – as in Equation 3. Age-specific proportions of deaths pregnancy-related are then calculated, dividing pregnancy-related deaths (col. v of Table 1) by all-cause deaths (col. vi of Table 1) (Equation 4). Finally, the proportional contribution of each age group to overall pregnancy-related deaths is calculated by dividing the number of pregnancy-related deaths in each age group by the total number of pregnancy-related deaths (col. v of Table 1) (Equation 5).

Table 2 Adult female mortality rates and pregnancy-related mortality rates: Malawi 2008 Census

Age Group	Age-Specific				Proportion of Deaths Pregnancy-Related	Proportion of Pregnancy-Related Deaths	Proportion of Births in Previous Year
	Mortality Rate	Pregnancy-Related Mortality Rate	Fertility Rate	Pregnancy-related Mortality Ratio
	Table 1 (vi)/(i)	Table 1 (v)/(i)		(ii)/(iii)	Table 1 (v)/(vi)	Table 1 (v)/Sum(v)
15-19	0.00346	0.00138	0.1108	1249.12	0.40000	0.127544	0.14408
20-24	0.00563	0.00208	0.2464	846.13	0.37018	0.195387	0.32584
25-29	0.00769	0.00256	0.2296	1113.89	0.33258	0.199457	0.25267
30-34	0.01147	0.00373	0.1941	1919.15	0.32484	0.207598	0.15264
35-39	0.01154	0.00313	0.1457	2151.04	0.27168	0.127544	0.08367
40-44	0.01068	0.00269	0.0718	3748.34	0.25210	0.081411	0.03065
45-49	0.01053	0.00256	0.0311	8238.13	0.24324	0.061058	0.01046
Total 15-49	0.00755	0.00241	0.1713	1895.63	0.33804	1	1
Source: Table 1 and Malawi 2008 Census, 10% sample

Diagnostics, analysis and interpretation

Checks and validation

The key checks for this methodology are the assessment of coverage of adult female deaths and of births (see Death Distribution Methods [26] and Assessment of recent fertility data [27]). The only checks specific to this method are of the distribution of pregnancy-related deaths by age and a very weak check, available in many instances, on the distribution of pregnancy-related deaths by whether they occurred during pregnancy, during delivery, or in the 6 weeks/2 months post-partum. For the assessment of the distribution of pregnancy-related deaths by age, the key comparison is with the distribution of births by age. These two distributions are shown in the last two columns of Table 2. The proportions of pregnancy-related deaths through age group 25-29 are lower than the corresponding proportions of births, but above age 30 the reverse is true. This pattern is plausible given accumulating evidence that pregnancy-related mortality risks are broadly similar by age below age 30 but then rise steeply above age 30. It is instructive to compare the patterns with those based on the sibling histories of the 2004 Malawi DHS (described elsewhere [29]). In the sibling history example, the proportions of pregnancy-related deaths contributed by mothers under age 25 were much smaller than the corresponding proportions of births. Here, the proportions are smaller, but not so much smaller as to lead to concerns about data accuracy.

It is noted above that the distribution of pregnancy-related deaths by whether they occurred during pregnancy, during delivery, or in the 6 weeks/2 months post-partum is a weak check. It is weak because there is no strong prior as to what this distribution should look like across different settings, and because in practice the distribution is found to vary wildly by data source.

Interpretation

The importance of evaluating coverage of adult female deaths and births cannot be over-emphasized. Household deaths in some settings seem to be under-reported by as much as 50 per cent, and such an error would translate into a bias in the pregnancy-related mortality ratio of a similar magnitude. Recent births may also be under-reported, an error that may partially compensate for omission of deaths in the PRMR. Death distribution methods suggest that female deaths were under-reported in the 2008 Malawi Census by somewhere between 40 and 50 per cent, whereas application of the relational Gompertz method to the same data [3] indicates births in the year before the census were under-reported by about 18 per cent. The net effect of the two compensating errors would be an under-estimate of the pregnancy-related mortality ratio of about two-fifths.

In interpreting information on pregnancy-related mortality for Malawi, it is also essential to remember that Malawi was affected by a substantial HIV epidemic in the late 1990s and early 2000s. Associated HIV-related mortality will affect the results of death distribution methods, as described elsewhere, so the adjustment factor derived above should be used with caution.

One advantage of census data or large census samples is the lack of sampling uncertainty in the results. Thus whereas it is hazardous to draw conclusions from sibling histories about differentials, similar reservations do not apply to the same extent to estimates derived from census data (though even with a census numbers may be small for sub-groups, introducing stochastic error). Also, given the need to evaluate, and often adjust, data from census questions, any estimates will still be subject to considerable uncertainty.

References

Hill, K, C. Stanton, M. Levin et al. (2011) Measuring Maternal Mortality from a Census: Guidelines for Potential Users. Geneva: World Health Organization.