Data Collection and Estimation Techniques

Early attempts to measure mortality levels of African populations were based on efforts to register all births and deaths, especially in urban areas.

For instance, a vital registration system was implemented as early as 1915 in four cities of Senegal and continues to be maintained. Despite many attempts, the coverage and quality of vital registration systems remain poor at the national level in continental sub-Saharan Africa (De Graft-Johnson, 1988).

Most recent efforts to estimate mortality levels and trends at the national level are based on large-scale surveys and censuses. These surveys employ two basic approaches to estimating infant and child mortality. The first approach involves direct methods, which are based on reported births and deaths of individual children. The most common survey approach is the maternity history, which includes the date of each birth and the date or age at death of each child who has died. Maternity histories have formed the core of both the World Fertility Surveys (WFS) and the Demographic and Health Surveys (DHS). Variations include truncated maternity histories, which include only recent births (i.e., children born during the past five years or the two most recent children born to each mother). Death rates can be calculated directly from these data because they allow tabulations of deaths and person-years of risk at each age. Trends in mortality can be estimated from tabulation of deaths and person-years of risk by both age and year. A second type of direct estimation procedure is based on information about recent deaths (generally deaths in the last year) and the current population size.

The alternative approach is to use indirect methods, which are based on data that do not provide for tabulation of deaths and person-years of risk by age. The most common indirect approach is the Brass child survival method, which uses data on the average numbers of children ever born and the surviving children of women in each five-year age group, along with a simple model. This model requires assumptions about the age pattern of child mortality. It can be used to estimate trends in mortality on the further assumption that the trends in recent years have been smooth. Several variants of this model are described in detail in United Nations Manual X (United Nations Department of International Economic and Social Affairs, 1983).

Most of the data on levels and trends come from comparisons of several surveys in the same country. When a country has a consistent combination of estimates of levels and trends from direct and indirect methods, the statistical evidence can be considered reliable. However, Hill (1992) recently reviewed the available data from 38 sub-Saharan African countries and found as many cases of inconsistent as of consistent data. Common shortcomings of data on child survival include the omission and displacement of births and deaths and misreporting of ages and durations of exposure to the risk of death. Although the general impression of mortality decline is strong, the estimated trends should be considered with caution because of these types of error in data.

In addition to national data, more accurate local studies can provide detailed information and a check on the plausibility of national data. Examples include some of the vital registration systems of cities and some prospective studies of communities. In these research areas with continuous registration of vital events, a small, geographically defined community is followed for a period of time with a multiround survey—the frequency of which can range from weekly to annually—and periodic censuses. These intensive undertakings provide complete and accurate demographic data, including trends, age patterns, seasonal variations, differentials, and sometimes, causes of death, though typically for fairly small numbers of events.

Current Levels of Mortality and Trends Since 1960

The measure we use for tracking levels and trends of child mortality is the probability of dying by age 5 (₅q₀), expressed per thousand live births. This measure captures almost all the mortality risk prior to adulthood, and is less affected by the age profile of child rearing practices such as weaning than the infant mortality rate. Estimates of ₅q₀ are available after 1980 for 15 of 39 sub-Saharan African countries. Table 2-1 presents the probability of dying by age 5 for the years before 1960 to 1985 for all the countries of sub-Saharan Africa with acceptable estimates. The number of estimates for each year ranges from 15 (for 1985) to 27 (for 1970). The recent estimates vary from 100 or fewer deaths by age 5 per 1,000 live births in Botswana, Kenya, and Zimbabwe to 200 or more in Burkina Faso, Liberia, Mali, and Zaire. The remaining countries cluster around 160. The low-mortality countries have mortality rates comparable to those in many countries of Asia, with levels lower than India or Pakistan. The high-mortality countries include the highest values of child mortality recorded in the world during the 1980s.

TABLE 2-1

Child Mortality Rates per 1,000 Live Births (based on the probability of dying by age 5), by Country and Year, Pre-1960 Through 1985.

The data for almost all African countries show declining mortality since 1960. Figure 2-1 shows data from Table 2-1, with different symbols used for data points from the broad subregions—western, middle, eastern, and southern. Three points stand out from Figure 2-1. First, there is a great deal of variability in measures of child mortality among countries of sub-Saharan Africa. In 1960, probabilities of dying by age 5 varied from 140 to 400 among the 19 countries with estimates; in 1985, the range was from 60 to 250 across the 15 countries with estimates. Second, there has been a substantial and rather steady decline in child mortality risks from 1960 to 1985; the median ₅q₀ in 1960 of 225 fell to 180 by 1985. Third, there is a substantial differential in child mortality between the countries of eastern and southern Africa and those of middle and western Africa; although affected by the countries included in the data set for specific years, the median ₅q₀ values for eastern and southern Africa are about 60 per thousand lower than those for western and middle Africa from 1960 to 1985.

FIGURE 2-1

Estimates of the probability of dying by age 5 (₅q₀) by subregion and time period. NOTE: Western Africa (shown as a plus) includes Benin, Burkina Faso, The Gambia, Ghana, Guinea, Guinea-Bissau, Liberia, Mali, Niger, Nigeria, Senegal, Sierra Leone, and (more...)

In countries where long-term series are available, there is consistent evidence that the mortality decline started well before 1960. A number of estimates of child mortality suggest that current levels reflect long-term, substantial changes. The earliest estimate of ₅q₀ for Kenya, for example, goes back to 1947, when it was 262 (Hill, 1991). By 1970, the probability of dying before age 5 had decreased to 165, and it was estimated to be 100 by 1985. The estimate of ₅q₀ for Ghana in 1935 was 371 (Hill, 1991), had declined to 185 by 1970, and continued to decline, although more slowly to 160 in 1985. Senegal also shows a large decrease in mortality, although current levels remain fairly high. The probability of dying by age 5 was estimated to be 373 in 1946 (Hill, 1991); it fell to 285 in 1970, and was estimated to be 190 in 1985. Vital statistics for Dakar, Senegal, suggest that mortality has been declining for at least 70 years.

In general the decline in mortality has been steady since 1960 in most countries. However, exceptions exist. One exception is Ethiopia, with no evidence of decline between 1960 and 1975, and no data thereafter. In Ghana, the decline seems to have stopped after 1980 and there may actually have been some increase. The same seems to be true in Uganda, where mortality appears to have been rising slightly since 1975. The scanty data from Mozambique do not show any evidence of decline. It also appears that mortality did not change in Rwanda between 1960 and 1980. Several of these countries went through periods of major political turmoil during the 1970s and the 1980s.

In several countries, periods for which data are available are too short to permit any serious conclusion about trends. In Nigeria, data are so poor before the 1990 DHS that mortality trends cannot be determined. Other countries—Angola, Chad, and Mauritania—do not have nationally representative data that allow any estimation of mortality.

In addition to variations between countries, there are major variations in mortality levels within countries. These striking subnational differences were noted especially in Kenya, Angola, Niger, and Mozambique by Coale and Lorimer (1967). Large differences were also found by Ewbank et al. (1986) in Kenya. Hill (1992) found variations by district ranging from 140 to 380 in the 1955-1957 survey in Zaire and variations from 130 to 290 in the 1969 census of Zambia. Cantrelle et al. (1986) found large variations in Senegal. Farah and Preston (1982) showed major differences between northern and southern Sudan. Extreme regional differences seem to be a distinctive feature of African mortality, probably reflecting major differences in local epidemiological and socioeconomic environments.

In summary, there is clear evidence that mortality of children under 5 has been declining rapidly in sub-Saharan Africa, in some parts at least since the 1940s and possibly since the 1910s. However, the data are not detailed or precise enough to show the specific effects of health systems and health interventions. Such a conclusion is limited by both the imprecision of the estimated trends in mortality and the difficulty of establishing changes in the quantity and quality of medical services actually provided during each time period, as discussed in Chapter 6. Nonetheless, the mortality decline accompanied the emergence of modern medicine, and the arrival of antibiotics and antimalarial drugs, as well as socioeconomic development, characterized by modern transportation, increase in agricultural and industrial production, urbanization, and modern education.

The extent of political and economic organization in a country seems to play a major role in mortality reduction. Countries where political unrest, war, and famine disrupt the social organization tend to have a stagnant mortality and, in some cases, rising mortality rates. Most likely, future trends in mortality will also continue to be shaped by the combination of improved health services, activities related to economic development, and unexpected social and political problems.

Verbal Autopsies

A number of researchers have used a technique called "verbal autopsy" to assess the causes of deaths that were not attended by trained medical personnel (e.g., Garenne and Fontaine, 1986; Gray et al., 1990). The concept behind this technique is to use all the available evidence—clinical, epidemiological, and demographic—to determine the probable causes of death. Although verbal autopsies provide guesses of the causes of deaths that might otherwise have been included in a miscellaneous category of ''other deaths," these gains can be illusory. In particular, verbal autopsies, like all diagnostic techniques, can lead to false attribution. In some cases, these false positives may outnumber the true cases.

Little is known about the precision of algorithms (i.e., the series of questions asked to determine the cause of death) used in verbal autopsies. For example, what is the ability of algorithms for malaria to pick up all cases of malaria (i.e., the sensitivity of the algorithm, denoted here by ) or their ability to identify deaths due to other causes as "nonmalarial" (i.e., the specificity, )? Given the limited amount and types of information that can be collected in a verbal autopsy, it is difficult to achieve high rates of both sensitivity and specificity. If the specificity is low and the condition is rare, then the estimated proportion of deaths due to a cause is likely to be exaggerated. On the other hand if the sensitivity is low and the condition is common, the estimated rate might be understated.

Kalter et al. (1990) studied some algorithms for tetanus, measles, diarrhea, and acute lower respiratory infections (ALRI) in the Philippines. Table 2-2 presents their results for four sets of criteria for diagnosing ALRI. As expected, the specificity of algorithms increased with the number of items included in the algorithm, but the sensitivity declined accordingly. Thus, with more symptoms, fewer cases of ALRI were picked up, but those that were detected were more likely to be real cases (i.e., an increased positive predictive value).

TABLE 2-2

Sensitivity and Specificity of Various Criteria for Diagnosing Acute Lower Respiratory Infections in the Philippines.

The implications of this trade-off between sensitivity and specificity for the estimated proportion of deaths due to ALRI are shown in Figure 2-2. If the real proportion of deaths due to this cause is , and the sensitivity and specificity of the diagnostic criteria are and , then the estimated proportion of deaths due to the cause is + (1 - )

FIGURE 2-2

Relationship between true proportion of deaths to acute lower respiratory infection and proportion estimated from verbal autopsies using four sets of criteria for diagnosis.

(1 - ). For example, if ALRI was not responsible for any deaths in a population, criteria C (cough of four or more days and breathing difficulty of one or more days) would still estimate that 23 percent were due to ALRI because the specificity of these criteria is only 77 percent. If 35 percent of the deaths are truly caused by ALRI, then the false diagnoses of ALRI by using criteria C are counterbalanced by missed cases of ALRI. In this case the estimated proportion of deaths due to ALRI is 36 percent, which is very close to the true 35 percent.

Criteria D (cough of four or more days and breathing difficulty of two or more days) have a higher specificity (and lower sensitivity), and produce an accurate estimate if ALRI is actually responsible for 10 percent of deaths. If ALRI is responsible for more than 10 percent, criteria D underestimate the true rate. If ALRI is responsible for less than 10 percent, criteria D lead to an overestimate. Criteria A (cough and breathing difficulty) and B (cough and breathing difficulty of one or more days) have such low specificities (47 and 60 percent, respectively) that they do not produce reasonable estimates of the proportion of deaths due to ALRI at any plausible true level.

Reasonable estimates of the proportion of deaths due to ALRI (or any other cause that is responsible for less than 30 percent of deaths in an age group) are obtained only if the specificity of the criteria is quite high. For example, with a sensitivity of 65 percent and a specificity of 80 percent, the prevalence of a cause that is actually responsible for 15 percent of deaths will be overestimated by 78 percent (i.e., it will be estimated to be 27 percent of deaths). In some cases, it may not be possible to achieve levels of specificity high enough for important causes of child deaths in Africa.

Snow et al. (1992) used verbal autopsies for 303 infants who died at Kilifi hospital in Kenya. However, Becker and Gray (personal communication) have raised several questions about the questionnaire used by Snow et al. They note in particular that measles, accidents, and malnutrition are the only causes that were addressed by specific items in the questionnaire. All other diagnoses apparently were based on responses to a checklist of 29 symptoms and signs. It is not clear how the interviewers asked the questions on the checklist, and it is now known what algorithms were used to derive the verbal autopsy diagnoses. Because of these problems, the results obtained by Snow et al. do not necessarily reflect the sensitivities and specificities that could be achieved by the type of detailed questionnaires and algorithms used in other studies. Therefore, the study by Snow et al. does not provide a test of the true potential of verbal autopsies.

Mobley et al. (personal communication, 1993) examined the sensitivity and specificity of carefully defined algorithms for five major causes of death in Namibia. Their sample included deaths in a hospital of children who were diagnosed with malnutrition, diarrhea, pneumonia, malaria or measles either at the time of admission or before the time of death. Sixty-one percent of the cases had two or more of these diagnoses.

Mobley et al. (personal communication, 1993) reviewed the reliability of several algorithms for each disease. For diarrhea deaths, the presence of loose or liquid stool combined with either thirst or sunken eyes gave a sensitivity of 75 percent and a specificity of 71 percent. Raising the specificity to 90 percent by using only the presence of 6 or more loose or liquid stools per day leads to a sharp reduction in the sensitivity to 36 percent. For pneumonia, the specificity never rose above 71 percent using the presence of cough and fever combined with difficult or rapid breathing. This algorithm was associated with a sensitivity of only 61 percent. A diagnosis of measles based on age greater than 120 days and rash led to a sensitivity of 71 percent and a specificity of 85 percent. Adding fever for at least three days raised the specificity to 90 percent but dropped the sensitivity to 67 percent. Their algorithms for malaria reached relatively high levels of specificity (several in the range of 87 to 97 percent) based on fever plus convulsions and/or loss of consciousness. However, these were associated with low sensitivities (41 to 45 percent). These algorithms performed better for children diagnosed in the hospital with cerebral malaria.

The above study provides very useful information about the reliability of verbal autopsy estimates for major causes of child mortality in Africa. However, by limiting the sample to deaths ascribed to these five causes in the hospital, the estimated specificities might be different than they would be in a completely random sample of deaths. For example, the specificities for cerebral malaria might have been lower if they had included deaths to meningitis and other infectious diseases that can be associated with high fever and convulsions. However, the fact that even the most stringent algorithms rarely lead to specificities as high as 90 percent suggests that verbal autopsies will rarely be useful for measuring mortality rates for causes of death that are responsible for less than 15 percent of all deaths in an age group.

Studies of Causes of Death Among Infants and Children

The appendix to this volume reviews the results of studies of causes of deaths among infants and children in 9 areas of Africa. These studies are summarized in Table 2-3. Given the problems discussed above, we have to exercise caution in interpreting the results of these studies and in trying to generalize to the whole of sub-Saharan Africa.

TABLE 2-3

Summary of Cause-of-Death Studies by Rank of Importance of Cause.

Low birthweight, birth trauma, and congenital defects are responsible for the largest number of neonatal deaths in most studies. The few exceptions (western Sierra Leone and Niakhar, Senegal) have very high rates due to neonatal tetanus. During the postneonatal period, diarrheal diseases are generally ranked first, followed by ALRI-pneumonia, malaria, and measles in varying orders.

At ages 1 to 4 years, measles is often the most common cause of death. In some parts of the continent, measles vaccination rates may have increased sufficiently to reduce its importance; however, it certainly continues to be a major factor in child mortality and must remain a top priority for health programs. Diarrheal diseases are generally the second cause of deaths in this age group, with malaria, malnutrition, and ALRI-pneumonia generally ranked third through fifth in varying orders. Of course, it is often difficult to determine a specific cause of death because of the common simultaneous occurrence of measles, ALRI, and diarrhea.

Although it is not possible to produce reliable estimates of the proportions of infant and child deaths in Africa that are due to each major cause, the primary importance of diarrheal diseases, measles, malaria, ARI, and malnutrition is clear. The list can be rounded out with the addition of a few neonatal causes (prematurity, birth trauma, congenital defects, and neonatal tetanus), and the possible addition of a few causes that are more difficult to diagnose (e.g., pertussis and meningitis). The relative importance of these main causes will vary across populations because of differences in ecology (e.g., malaria), behavior (e.g., neonatal tetanus), and vaccination patterns (e.g., measles). Given the difficulty in determining specific causes of death, it may be unrealistic to determine whether or not the disease-specific reductions in mortality, such as those stated in the goals of the World Summit for Children, are actually being achieved.

Effects of AIDS on Child Mortality

The acquired immune deficiency syndrome epidemic is a new factor influencing infant and child mortality in sub-Saharan Africa. Currently, we do not have solid data on the incidence of AIDS deaths among children in the region. However, we do know that the AIDS epidemic will have both direct and indirect effects on infant and child mortality. The direct transmission of the human immunodeficiency virus (HIV), which causes AIDS, from mother to child occurs primarily during delivery, but may occur also through breast milk. It is estimated that at least 80 percent of children infected with HIV will die by age 5. The indirect effect is the additional mortality risk arising from orphanhood faced by an uninfected child whose mother is ill from or dies of AIDS.

How large are these effects likely to be? Stoto (1993) reviews simulation models of AIDS and finds that infant mortality rates might increase by as much as 20 percent in urban areas. However, these estimates have wide margins of uncertainty. First, the estimates depend critically on perinatal HIV transmission rates. The models assume that about 30 percent of children born to HIV-positive mothers will themselves be HIV positive, although some European studies have shown rates as low as 13.9 percent (European Collaborative Study, 1991). Second, it is not known if HIV-positive women have higher or lower fertility rates than women in general because of social behavior or health status. Third, the estimates consider only the direct effect described above; they do not include the additional risks to AIDS orphans. Such risks may be large for very young children, but probably fall rapidly as age of the child increases.

Simple calculations provide an estimate of the order of magnitude of the effect of AIDS on child mortality in Africa. It is estimated that approximately 2.5 percent of adults in sub-Saharan Africa are HIV positive (U.S. Agency for International Development, 1991). If HIV-positive women have the same fertility rates as other women, then about 2.5 percent of all births in sub-Saharan Africa are to HIV-positive mothers. If the transmission rate from mother to child is 30 percent, and if 80 percent of HIV-infected children die by age 5, then .006 (i.e., .025 • .3 • .8) of children born in Africa will die by age 5 of AIDS. However, about 18 percent of these children would have died by age 5 in the absence of AIDS. Thus, the direct effect of AIDS is to increase mortality by age 5 by about .006 (1 - .18), or 5 per 1,000. Therefore, AIDS is probably responsible for only about 3 percent of all deaths to infants and children in Africa. Adding to this the extra child deaths resulting from orphanhood among HIV-negative children born to HIV-positive mothers would increase this percentage only slightly.

Estimates of the recent effects of AIDS on child mortality in sub-Saharan Africa are uncertain enough; forecasts of its effects for the rest of the decade are more hazardous still. Perinatal transmission rates and case-fatality rates are not likely to change much in the next decade. However, seroprevalence might continue to increase or might start to fall as behavioral modifications occur and high-risk groups are reduced by excess mortality.