Introduction to adult mortality analysis

Desired Result

Introduction

Accurate knowledge of adult mortality levels and trends in the developing world is hampered by a widespread lack of complete vital registration systems. Although knowledge of infant and child mortality once faced similar barriers, survey-based techniques – indirect methods and birth histories – have been more successful at measuring child than adult mortality, and we know correspondingly less about the latter than the former.

For the purposes of demographic analysis, adult mortality is usually defined as mortality at ages 15 or more. In some contexts though, the term ‘adult mortality’ is used to refer solely to mortality between exact ages 15 and 60, and is contrasted with older-age mortality, which is used to refer to mortality at ages 60 or more. The probability that a person on their 15th birthday dies before their 60th birthday, (₄₅q₁₅ in the life table) has become a widely used indicator of adult mortality defined in this more restricted way.

In countries that lack complete vital registration systems, the sources of data and methods that are used to study mortality in adulthood usually differ from those used to study mortality in childhood. Some of the methods for adults can be extended to study the mortality of children aged 5 or more, but none of them are reliable sources of information on under-five mortality.

Several general issues make the study of adult mortality inherently more challenging than that of children. First, in broad terms, adult mortality rates for much of the age range are an order of magnitude lower than those of children. Adult deaths are relatively rare events. Obtaining precise measures of adult mortality requires data either on a large sample of people or on events occurring during a long reference period. Second, it is difficult to identify an appropriate informant who can provide reliable information about deceased adults. Data on child mortality can usually be collected from mothers. In addition, the characteristics of parents are among the more important determinants of the risk of dying in childhood. Since there is no single universally-suitable informant to provide data about adult deaths, problems of underreporting and multiple reporting are common. Moreover, it is often unreasonable to use the social and economic characteristics of the respondent as a proxy for those of the dead person in order to investigate mortality differentials.

Age misreporting is another serious problem that affects all sources of adult mortality estimates for low- and middle-income countries. Several factors make it difficult to obtain usable information on adult ages and ages at death. Older people are less likely to have birth certificates or health cards than are the young and, in most developing countries, are likely to have received less schooling. Moreover, even if dead persons knew their own age, the informant who reports their death may not. The reported ages of older adults are often exaggerated and ages at death tend to be exaggerated even more. Thus, ‘raw’ estimates of adult mortality for low- and middle-income countries often require smoothing by fitting a model life table before they can be used to estimate life expectancy or for demographic forecasting, and those on the elderly population may have to be discarded and replaced by data extrapolated from a model life table.

Data for the estimation of adult mortality

A relatively small number – and smaller share by population – of low- and middle-income countries have close to complete registration of adult deaths and population censuses of high quality. A larger number of countries have national or sample vital registration systems that are complete enough to be promising candidates for the methods described in this manual that assess the completeness of registration relative to census counts. In addition, an increasing number of countries have included questions in censuses (or very large sample household surveys) concerning household deaths by age and sex in some period (most often one year) prior to the census. The completeness of reporting of these deaths can be assessed by the same methods that are used to assess the data on adult deaths collected by registration systems.

A number of countries, particularly in sub-Saharan Africa, have conducted sample surveys (most often under the umbrella of the Demographic and Health Surveys programme) that have included sibling histories that ask each respondent about the survival or otherwise of each of their siblings and when their siblings died. Some countries have sought to measure adult mortality by including questions in censuses and surveys concerning the survival or otherwise of each respondent’s mother or father. These data, along with similar summary statistics on siblings, can be tabulated by age of the respondent answering the question and analysed by indirect methods that make use of demographic models to convert them into conventional life table indices of adult mortality.

Description of methods

Those methods that make use of data on deaths and the population at risk by age (and sex) to estimate adult mortality are collectively referred to as Death Distribution Methods. These methods fall into two distinct groups, depending on how the data are used, the Growth Balance methods and Synthetic Extinct Generations methods (Hill, You and Choi 2009). Both groups of methods require data on deaths from either a registration system or a question in the census together with census-based estimates of the population at risk by age.

The first Growth Balance method is the Brass Growth Balance method developed by Brass (1975), which only requires data on the population by age at a single point in time, but is only applicable if the adult population can be considered to be at least approximately stable (i.e. a population with a regular and unchanging age structure over time). The second Growth Balance method is a generalization of the first method to non-stable populations by Hill (1987), referred to as the Generalized Growth Balance method, which requires data on the population by age at two points in time. The first Synthetic Extinct Generations method is a method developed by Preston, Coale et al. (1980), which requires data on the population by age at one point in time and the assumption that the adult population is at least approximately stable. The second Synthetic Extinct Generations method is a generalization of the first approach to non-stable populations by Bennett and Horiuchi (1981, 1984), which requires data on the population at risk at two points in time.

Provided that the assumptions of constant completeness of coverage of the censuses and reporting of deaths by age are reasonably valid, net migration over the period has been small in scale, and there are no major distortions of age reporting between five-year age groups, Death Distribution Methods are the preferred methods for estimating adult mortality both because they provide age-period specific estimates of mortality rates and because they are capable of producing reasonably timely estimates (Hill 2001). However, deciding if these conditions have been met in practice requires a great deal of experience, which means that these methods are amongst the most subjective of the indirect techniques.

One can calculate age-specific death rates directly from counts of deaths and person-years of exposure by age and year derived from sibling histories collected in surveys. As the sample size in surveys such as those conducted by the Demographic and Health Surveys programme is rather small for the estimation of adult mortality, the data should be aggregated into periods of several years. Moreover, as it is common for increasing numbers of dead siblings to be omitted from the histories as the time since their death increases, only mortality estimates for the recent past should be produced from these data.

Data collection instruments such as sibling histories, in which the respondent generally does not live in the same household as the deceased, do not provide a suitable starting point for the collection of data on causes of death. The respondent is unlikely to know the medically-certified cause of death with any precision, especially if the dead person received little or no medical care; alternative approaches, using verbal autopsy methods, which enquire about the signs and symptoms preceding death, will also not work well because the respondent will generally have little first-hand knowledge of such indicators. Some of the same factors apply to information about household deaths collected by censuses, in that household members may well not know the true cause of death, and that the training of interviewers and time available for interviewing each household do not permit detailed probing. However, the census approach can provide a frame for a follow-up verbal autopsy enquiry, on a sample of households that reported deaths, using carefully-trained interviewers, but such surveys are expensive and complex undertakings.

Two other exceptions to this general rule are that it may be possible to distinguish injury deaths from deaths from natural causes and to identify pregnancy-related deaths, defined as deaths occurring while a woman was pregnant, during childbirth, or during the six weeks after the end of pregnancy. The estimation of pregnancy-related mortality from questions asked both in censuses and during the collection of sibling histories is described in the section on maternal mortality.

The alternative to trying to collect accurate data on deaths and the population at risk by age in order to estimate adult mortality is to use indirect methods of estimation. These methods do not require detailed information on the ages and dates at which people died. Instead, the proportion of individuals remaining alive among some specific category of relative of the respondents answering the question is tabulated according to the age of those respondents. Then, conventional life table measures of survivorship are predicted from these proportions using a regression model fitted to model data in which the relationship between the two quantities is known.

The most successful of the techniques that analyse data on the survival of relatives, estimates the mortality of adult women and men from data on the survival of respondents’ mothers and fathers by means of the orphanhood method first developed by Brass and Hill (1973). Contemporary applications of the method usually use the regression coefficients proposed by Timæus (1992) to estimate life table survivorship, rather than the weighting factors proposed initially, as the revised method generates more precise estimates for men. Variants of the method are also discussed here that are intended for use in populations with a high prevalence of HIV infection or when respondents have been asked whether their parents died when the respondent was a child or an adult as indexed, for example, by whether the respondent had married.

If successive sets of data have been collected on maternal or paternal orphanhood in multiple inquiries conducted in the same population, they can be used to estimate adult mortality during the intervening period from synthetic cohort data on orphanhood (Zlotnik and Hill 1981). Such estimates can be made from data on orphans of all ages using the regression coefficients developed for the basic method. This manual, however, focuses on the analysis of synthetic cohort data on orphanhood in adulthood as proposed by Timæus (1991), as this variant of the method is less vulnerable to underreporting of orphanhood by respondents whose natural parent(s) died when the respondent was a young child.

Lastly, the manual describes methods developed by Timæus et al. (2001), that make it possible to estimate adult mortality from data on siblings indirectly if respondents are asked how many of their brothers and sisters survived to adulthood and how many of them have since died.

Other methods for estimating adult mortality indirectly from data on the survival of relatives have been proposed such as asking about the survival of respondents’ first husbands and first wives. Experimentation with these questions has shown that respondents often fail to report that they have been widowed. Thus, the method commonly produces severe underestimates of adult mortality. The widowhood method and further methods based on other questions about the survival of relatives that have proved to be unsuccessful are not described in this manual.

A final approach that has been used to estimate adult mortality is the analysis of changes in population size in between two censuses. In a population closed to migration with accurate data, anybody who was present at the first census, but not at the second one, must have died. In practice, except at older ages, the net number of international migrants each year in most countries amounts to a significant fraction of the number of adult deaths. Few countries measure international migration flows accurately enough to adjust for their impact on intercensal population change before estimating mortality. In addition, even small changes in the completeness of the census enumerations can produce severe biases in estimates of adult mortality produced in this way. Thus, in general, the approach cannot be recommended as a method for estimating adult mortality and a detailed account of it is not offered in this manual. Any reader who wishes nevertheless to learn more about this way of estimating mortality is advised to look first at the variant of the approach proposed by Preston and Bennett (1983).

Introduction

Data for the estimation of adult mortality

Description of methods

Further reading and references