MORTALITY AND AIDS DEATHS IN DEVELOPING COUNTRIES Griffith Feeney 1999-05-24 0 INTRODUCTION Prepared for the Meeting of the Reference Group on HIV/AIDS Estimates, Modelling and Projections, Geneva, 10-11, June 1999. Sections 1-6 present some essential background. Section 7 lists some issues for discussion at the meeting, incorporating the suggestions contained in emails of April 9 from Bernhard (from buencaminoc@unaids.org) and May 10 from Geoff Garnett. 1 AIMS 1.1 First, to estimate the life table survivorship function and/or various statistics derived from it, e.g., expectation of life at birth, expectation of life at age 5, or conditional probability of death by age 65 given survival to age 30. 1.2 Second, to estimate the impact of AIDS deaths on adult mortality. Perhaps child mortality as well, but this will involve quite different data and techniques. 1.3 Ideally we would like annual estimates for a series of years extending from the more or less distant past through the most recent completed calendar year. In practice we will generally have to settle for far less. 2 INSTRUMENTS 2.1 Three data collection instruments in general use are: (i) national population censuses; (ii) large scale, nationally representative population surveys; (iii) civil registration systems. Censuses aim to be complete enumerations of all households and persons in a country and are generally taken once every ten years. Population surveys of the type indicated may be taken every 3-5 years. There will often be at least one such survey in every intercensal period. Civil registration aims to record information on all vital events in a country as they occur, with reports for calendar years available annually and perhaps monthly. 2.2 Non-nationally representative surveys may be available and can play an important role. The main liability will usually be that there are few or no alternative sources to compare them to, making it more difficult to assess their results. Surveys to estimate adult mortality are necessarily large because death before old age is a relatively rare event (this is/may be changing in countries most severely hit by the AIDS epidemic). 2.3 Traditionally, information available from these sources has been limited to tabulations and contextual information in published reports. Increasingly, it may now include the computer files and associated documentation from which the published tabulations were produced. New tabulations, tailored to the problem at hand, may then be produced with modest effort. This possibility, a consequence of extremely rapid information technology development, is so new that it may be overlooked. It should not be, though various practical problems will often render it inoperable in the near future. 3 DATA 3 The following data are useful for calculating/estimating adult mortality: (i) population by age and sex; (ii) annual deaths by age, sex and (where available) cause of death; (iii) population by age and survival of mother/father/siblings. Somewhat more precise specifications are necessary in application. Deaths should be classified by time of occurrence, for example, rather than by time of registration, and information on survival of siblings will usually be restricted to siblings living at the time the respondent reached age 15. 4 METHODS 4.1 The classical data sources for calculation of national life tables are one or more population censuses and annual death registration data. If this data is available and accurate, as is generally the case in developed countries, life tables may be calculated with minimal fuss for census years. Obtaining life tables for intercensal years may be more difficult on account of migration. These problems will usually be compounded if life tables are desired for subnational units. 4.2 Most developing countries have taken at least one population census. Many have taken three or more. Completeness of enumeration is generally fairly good, though accuracy of age reporting may be very poor. Many developing countries have some form of civil registration, but completeness of reporting is nearly always far too low for the numbers produced to be used without adjustment. Conditions tending to poor age reporting in censuses will tend to poor reporting in civil registration as well. 4.3 Attempts by demographers to estimate life tables and life table statistics in the absence of (reasonably) complete death registration statistics go back at least 50 years (Mortara 1949), yet there are only a few basic approaches: (i) census survival methods; (ii) use of retrospective reports of deaths from censuses or surveys; (iii) methods that adjust death registration data for under reporting on the basis of supplementary census or survey data; (iv) methods that estimate life table statistics from census or survey reports of survival of various relatives. 4.4 The original literature on these methods is scattered and difficult. The United Nations "Manual X" (1983) provides a unified account of many of the methods but is somewhat dated. A review of methods for adult mortality estimation is underway in the United Nations Population Division. 4.5 Census survival methods rely on two censuses of a population closed to migration to estimate adult mortality for the intercensal period, typically 10 years. Infant and child mortality cannot be estimated by this method (though models may be used to infer infant and child mortality from adult mortality). These methods provide excellent results under optimal conditions, but good results in practical applications are rare. The migration assumption generally renders them useless for subnational areas. Problems with the closure to migration assumption at the national level may sometimes be attenuated by suitable redefinition of the population, e.g., by considering the native born population rather than the total population. Estimates are provided only for a long period in the more or less distant past. For Zimbabwe at this writing, to take a single example, the most recent estimates would be for the 1982-1992 intercensal period. In practice these methods will most often be useful as a means of assessing the quality of census data. 4.6 Population censuses or surveys may include questions on deaths occurring in enumerated households during the year (or other period) prior to enumeration. Under optimal conditions these provide a tabulation of deaths by age and sex of decedent that may be used as a proxy for death registration data. Estimates are available only for the year (rarely, longer periods) preceding the census or survey. 4.7 Numbers of deaths in data of this kind way may be sharply under or over reported. In the 1989 census of Vietnam, for example, approximately half of all deaths were omitted (General Statistical Office 1991:105). Under reporting may result from reluctance of respondents to report deaths, memory failure, or misunderstanding of the reference period to which the question refers. The last may also result in over reporting. Obviously poor results are less likely to be published, complicating the assessment of how well such questions work. 4.8 Several methods are available for estimating completeness of death registration by combining data on registered deaths classified by age and sex with census or survey information on population by age and sex. As with census survival, it is necessary to assume that the population is closed to migration. This will often be the only data source capable of providing, in some circumstances and with some extra effort, a time series of annual estimates. 4.9 The original "growth balance" method of Brass requires only a population age distribution and a corresponding distribution of deaths by age, which may be taken either from death registration or from retrospective reports in a census or survey. Generalizations of this method utilize data from two successive censuses and intercensal deaths and are capable of estimating differential completeness of census enumeration (Brass 1979, Brass and Swamy 1980, Hill 1987). 4.10 Another approach to estimates from the same data is provided by "extinct generations" methods, in which population at each age enumerated in a census or survey is compared with population at the same age estimated by cumulating estimates of deaths at this and later ages to the cohort in question (Bennett and Horiuchi 1981). 4.11 Reports on the survival of relatives may contain information sufficient to estimate various life table statistics. Most famous and widely applied is the "Brass method" for estimating child mortality from census or survey reports on numbers of children ever born and surviving to women in each age group. Adult female mortality may be estimated from reports on whether or not the mothers of persons enumerated are surviving, and also from reports on whether female siblings of persons enumerated are surviving. Adult male mortality may be similarly estimated from reports on survival of fathers and/or brothers. 4.12 Survival of relatives methods may provide information on mortality trends using the dating procedure of Brass and Bamgboye (1981). This is likely to be the only source of data on mortality trends (excepting very long trends, as in successive intercensal periods) other than civil registration. 5 OBSERVATIONS 5.1 None of the data may be taken at face value. Age is often misreported, sometimes severely, in censuses, surveys, and civil registration. Censuses commonly miss a few to several percent or more of the target population. Civil registration rarely covers more than a fraction of deaths occurring in the population, perhaps as little as 25 percent. Coverage is likely to be selective, e.g., higher in urban than in rural areas. Survival of relatives may be misreported in various ways, e.g., adopted children of deceased mothers may report that their (adoptive) mother is surviving. 5.2 All of the methods make assumptions that will never be perfectly satisfied and will produce imperfect results (even with perfect input data) to the extent that actual demographic conditions depart from assumed conditions. We have much to learn about the robustness of different methods. 5.3 Different methods applied to the same data, or to different data representing the same population and time period, may produce very different results. This is good in that it alerts us that something is wrong; if different methods gave consistently wrong results we would be likely to draw wrong conclusions without realizing it. It does pose the question of which of the available estimates is closest to the truth. Arriving at answers will in some cases be more difficult than producing the initial estimates. 5.4 A high incidence of AIDS deaths poses at least three special problems: (i) rising rather than falling mortality, (ii) nonlinear change, and (iii) distortions in the age pattern of mortality. In the past, one was usually justified in assuming that mortality risks were falling, or at least not rising. A rise in retrospectively reported deaths in the months prior to a census or survey was therefore evidence, for example, that deaths in the more distant past were under reported--an invalid inference if deaths have been increasing. Assuming linear change is often reasonable and very useful as a first approximation, but this is unacceptable in the face of sharp rises in AIDS deaths. 5.5 Distortions in the age pattern of mortality influence the validity of inferences based on model age patterns of mortality, an essential tool in nearly all of adult mortality estimation methods. To illustrate briefly, census survival and methods based on incomplete death registration provide estimates of expectation of life at age 5. Methods based on survival of parents give estimates conditional probability of death by age 65 given survival to age 30. These statistics are not directly comparable, but it is common practice to effect comparison by using the relation between them that exists in a family of model life tables. In doing so one assumes that the age pattern in the population conforms to the age pattern in the model. Distortions in the age pattern due to unusually high death rates in adult ages may not wholly invalidate this approach, but they certainly complicate it. 5.6 Census and survey questions that work well in one country may not work well elsewhere. Sharp differences are observed even in the same country over time and between population subgroups. Children born and surviving questions worked well in the early censuses of Western Samoa, for example, but broke down in later censuses (Banister 1979). The same questions worked well for Malays and Indians in the 1970 census of Malaysia but very poorly for Chinese. Assumptions made by the various methods will be valid in some countries and invalid in others. 5.7 Producing useful estimates of adult mortality risks for developing countries and of the role played by AIDS is a non-trivial task. For most developing countries, however, it is probably less difficult than producing good estimates of AIDS prevalence (to say nothing of incidence). We are likely to have more data to work with, from a wider variety of sources, with more historical depth, a wider variety of estimation methods, and more experience in application. 6 RECOMMENDATIONS 6.1 Given that we want to produce the best possible estimates of adult mortality and of AIDS impact for a country, how should we proceed? 6.2 Inventory available data. Use primary sources to the greatest extent possible. Errors tend to slip into secondary sources and important contextual information (e.g., sample design) will usually be missing. Inventory digital as well as print sources. 6.3 Apply all (within reason) methods applicable to the available data and compare results. Be prepared to find discrepancies, sometimes severe. In some cases the hardest work begins at this point. 6.4 Attempt to assess errors in available data and the extent to which the various assumptions of methods are satisfied. Attempt to assess how different estimates are impacted by errors in the data and departures of actuality from assumptions. Try to arrive at a defensible conclusion about best estimates. 6.5 Begin the analysis at the national level. Proceed to subnational levels if data and resources permit and if the results promise to justify the effort. Studying data for subnational levels will often provide insights valuable even if estimates are desired only for the national population. 6.6 Expect to adapt the analysis and innovate in response to particular opportunities or problems thrown up by circumstance, not merely to apply available methods and record their results. The most important evidence may not result directly from the application of any off-the-shelf method. 6.7 In Zimbabwe, for example, important evidence of the demographic impact of the AIDS epidemic comes from death rates calculated from death registration data, available for 1982, 1986, 1990-92 and 1995. The rise in death rates throughout the adult ages is far too rapid to be explained by improving registration completeness. On the other hand, rises in death rates for the 10-14 age group suggest that improving completeness has played a role and provide a basis for estimating changing completeness. 7 ISSUES 7.1 Why precisely is the Reference Group concerned with mortality estimates? What is the importance of mortality estimates in the work as a whole? Perhaps the same answers will be obvious to all, but it may be prudent to raise the question. 7.2 Given that it is impossible to exhaustively analyze available data for all (or even a large number of) developing countries, how should effort be allocated? Intensive analysis of a small number of countries is an obvious strategy. What countries? How to select them? How much diversity or focus? 7.3 How much emphasis should be put on death registration data? Incomplete and selective reporting of deaths pose serious problems, but civil registration alone is capable of providing information annually (and indeed, monthly) for subnational as well as national units. How can we assess changing completeness of registration? 7.4 How much emphasis should be put on estimates based on reports of survival of relatives? There may be serious problems with accuracy of reports, but these methods do not require any assumption about migration, may be applied (with some caveats) at the subnational as well as national level, and may provide information on trends. In the absence of civil registration data they may be the only source of data on trends. 7.5 What can/should be done to obtain more data for estimating adult mortality? There are numerous options for exploiting existing data. (i) Most obviously, ensuring that all available published data is available for analysis (many reports are hard to obtain); this is less trivial than it ought to be. (ii) Securing the use of tabular data produced but unpublished by national statistical offices (there usually is some, and it may be very valuable, e.g., unpublished death registration data for Zimbabwe). (iii) Urging/supporting the the production of special tabulations from unit record data files available in national statistical offices or elsewhere (e.g., tabulation of data available in DHS sibling histories). (iv) Urge inclusion of pertinent questions in upcoming surveys and 2000 round population census schedules (with follow through/support for tabulation and analysis). 7.6 How do we measure impact of HIV/AIDS on adult mortality? What is the role of cause of death data, traditional or other? What about inferences based on changing age pattern or trend of mortality? 7.7 How to overcome the special problems caused by large and rapidly changing numbers of AIDS deaths? Cf. 5.5 above. 7.8 Under what conditions can we use mortality data to adjust estimates of prevalence? I should think the answer is "under much the same conditions as epimodel allows us to estimate (however imperfectly) infections and incidence from prevalence data, and by similar methods", but elaboration is required. 7.9 What can be done to learn more about the way in which available methods respond to data errors and departures of actual from assumed conditions? Do particular data/methods perform better than others in the presence of large numbers of AIDS deaths. We need to know more about sensitivity and robustness to move from discrepant estimates from various data sources and methods to best estimates, in general as well as with respect to the special problems raised by AIDS mortality. 7.10 What can be done to incorporate AIDS mortality into existing model life tables? One simple approach would be to express age-specific death rates as the sum of rates in an existing model and a factor k (say) times a standard pattern of adult AIDS deaths, k a parameter that may be increased to capture increased AIDS deaths. 7.11 How to approach the issue of AIDS mortality among children? The nature of available data and methods suggests that this should be approached separately from the question of AIDS mortality among adults. 7.12 What about progression rates for adults and children? This is certainly a mortality issue, but not one that the data sources and methods noted above help us answer. 8 REFERENCES Banister, Judith. 1979. Census questions on fertility and child mortality: Problems with questionnaire design. Asian and Pacific Census Forum 6(1):5-8. Bennett, N.G., and Shiro Horiuchi. 1981. Estimating the completeness of death registration in a closed population. Population Studies 47(2):207-221. Brass, William. 1979. A procedure for comparing mortality measures calculated from intercensal survival with the corresponding estimates from registered deaths. Asian and Pacific Census Forum 6(2):5-7. Brass, William and E.A. Bamgboye. 1981. The time location of reports of survivorship estimates for maternal and paternal orphanhood and the ever-widowed. Working Paper No. 81-1, London School of Hygiene and Tropical Medicine, University of London. Brass, William, and Subramania Swamy. 1980. Measurement of death registration completeness using the growth balance procedure applied to data from India. Asian and Pacific Census Forum 7(1):5-8. General Statistical Office [Vietnam]. 1991. Vietnam Population Census - 1989: Detail Analysis and Sample Results. General Statistical Office, Hanoi. Hill, K. 1987 Estimating census and death registration completeness. Asian and Pacific Census Forum 1(3):8-13. Mortara, G. 1949. Methods of using census statistics for the calculation of life tables and other demographic measures. United Nations, New York. United Nations. 1983. Manual X: Indirect Techniques for Demographic Estimation. United Nations, New York. 9 DISTRIBUTION Please advise of any errors or omissions! MORTALITY ISSUES GROUP Feeney_Griffith (sender) Gray_Ron
Gregson_Simon Hill_Ken Timeaus_Ian Zaba_Basia Zlotnik_Hania FOCAL POINTS Boerma_Ties Feeney_Griffith (sender) King_Gary Staneck_Karen Yan_Ping COORDINATION Buencamino_Carmen Garnett_Geoff Schwartlander_Bernhard Walker_Neff rgray@jhsph.edu, simon.gregson@zoology.ox.ac.uk, khill@jhsph.edu, ian.timaeus@lshtm.ac.uk, b.zaba@lshtm.ac.uk, zlotnik@un.org, ties_boerma@unc.edu, king@harvard.edu, kstaneck@census.gov, Ping_Yan@hc-sc.gc.ca, buencaminoc@unaids.org, geoff.garnett@zoology.oxford.ac.uk, schwartlanderb@unaids.org, walkern@unaids.org