Reliable direct measurement of causes of death in low- and middle-income countries

Mortality and causes of death

There are two components to measuring mortality; the first is capturing the event of death (by at least age and sex, but ideally including geographical residence and education or some measure of social status). The second is ascertaining the underlying COD, which is the disease that contributed most directly to the death (as distinct from immediate cause or co-existing conditions). Death is a concrete, final, and measurable event that households remember well, and it can be captured in household surveys reliably.

Since about 1960, use of population totals from census data and retrospective mortality estimates, paired with indirect demographic methods, has enabled reasonably credible estimates of the age-specific and sex-specific national mortality rates for most countries. These are published and updated regularly by the UN Population Division [6]. Additional methods such as household surveys of birth and death histories of children, their siblings and parents, or similar methods [4, 10, 11], have also improved estimates of age- and sex-specific mortality, or in some cases specific causes (such as maternal deaths).

Even rudimentary data on age- and sex-specific all-cause mortality rates can identify trends for rapidly changing diseases such as those responsible for childhood mortality [12]. A simple analysis of the excess death rates between the 1911 and 1921 censuses in India suggests that the influenza epidemic of 1918/19 might have killed about 20 million of the population [13]. More recently, it has been possible to obtain a crude estimate of the effect of HIV infection on mortality in selected sub-Saharan countries from the observed large increases in all-cause mortality rates in young adults from 1990 to 2000, using simple burial records [14]. In both cases, this was because the ‘signal’ of influenza-attributable or HIV-attributable mortality (the increase in the number of deaths) exceeded the ‘noise’ of misclassification by competing mortality from other causes [15].

Why collect cause of death data?

Over the past century, reliable COD data in high-income countries have led to numerous discoveries and health improvements [1, 16–18]. For example, the dramatic increase in lung cancer deaths in British and American men around the period of World War II spurred research that led to the identification of smoking as a cause of lung cancer and, eventually, a wide range of diseases [19]. In the early 1980s, routine mortality data from San Francisco identified an exceptional increase in immune-related deaths in young men, which signaled the beginning of the American HIV epidemic [20]. Downward trends in infectious disease mortality worldwide have reflected the widespread use of immunizations and antimicrobials [17, 18].

Despite these clear benefits, COD data are exceptionally limited in LMICs because of historical neglect and insufficient resources. Of 115 countries that reported mortality data to the World Health Organization (WHO) around 2005, only 64 had good-quality vital registration that also included COD certification [3]. Fully 75 countries (including about 90% of African countries) did not provide data on COD for any year after 1990 (Figure 1). Similarly, the Child Health Epidemiology Reference Group estimated that COD information was available for only 3% of global child deaths [7]. There is a major knowledge gap for representative information on COD in adults, such as the proportion of adult deaths due to malaria [21]. China and India have nationally representative COD surveys [9, 22, 23], and China and Latin America have expanded civil registration with medical certification. Thus, the major gap in COD information exists in sub-Saharan Africa and in South and East Asian countries [3].

Seven limitations of indirect methods to estimate COD

In 1981, the Ghana Health Assessment Team [24] advanced the concept of measuring burden of disease by combining loss of useful life through premature death with the loss of useful life due to disability. This spurred further work, most notably the influential World Development Report 1993 by the World Bank [25]. The WHO Global Health Estimates (GHE) [26] use similar approaches to estimate mortality levels and COD for the major regions of the world, and for most countries. The 2010 Global Burden of Disease (GBD) Study goes further by claiming it can make reliable estimates for ‘291 diseases and injuries, 67 risk factors, and 1,160 non-fatal health consequences in 21 regions, 20 age groups and 187 countries’ [27]. It is important to understand the strengths and limitations of these indirect methods.

A simplified schema of how GBD estimates are created is shown in Figure 2, with the key inputs of mortality totals (derived from the UN or demographic methods), COD data, estimates of disability weights for specific conditions (based largely on expert opinion), and risk factors (based largely on literature reviews). The GBD uses regression equations (of uncertain error) to predict the level of mortality, or the distribution of several CODs from populations (mostly in high-income countries) for which such data are known. Parameters for independent variables include economic (GDP/capita), educational (literacy rates), and other variables such as generalized HIV epidemic, endemic malaria, and geography, among others. In parallel, simultaneous equations are used to model relations between incidence, prevalence, death rate, duration of disease, and survival, among others, although this does not directly enable estimations of COD distributions.

The obvious limitation of this indirect method is an insufficient number of nationally representative COD data. The 2010 GBD [27] study had to apply these complex methods to a ratio of 1 nationally representative COD to 850 estimated deaths so as to measure the cause of 25 million deaths, across 110 LMICs outside China, India, and Latin America (Figure 3). National COD surveys of reasonable representativeness used by the GBD were available only from Afghanistan, India, Mozambique, Thailand, Vietnam, and Zambia, totaling about 60,000 surveyed annual deaths [27], of which half were from the MDS in India [9]. Other COD data sources in the GBD are described scantily, but appear to comprise hospital archives on deaths and diseases (a patchwork of generally small, non-representative surveys), death and cancer registries, mortuary statistics, and incomplete records from ministries, census departments, and police reports.

Mortality comprises most of the composite of mortality and disability measures, such as disability-adjusted life years [25]. Mortality does not capture all illnesses or specific priorities, such as depression, schizophrenia, or blindness. For most diseases, the GBD estimates of disability are largely expert opinion, and thus there is no easy way to validate the disability estimates or to study the relationship of disability and death. For example, in comparison to the 1993 World Bank report [25] or earlier GBD reports, the 2010 version of the GBD assigns about 2.5 times as many disability adjusted life years to child deaths [39]. However, the correlation between premature mortality (specifically death before the age of 70 years) and morbidity is strong for the majority of other diseases of public health importance, with only some exceptions [27, 28]. This implies that reliable COD data should remain the major priority for most LMICs.

Five options to expand COD reporting at low cost and high impact

For most individual LMICs, the sources of mortality data are too incomplete, too unrepresentative and of too uncertain quality for accurate COD estimation at the national level. Expanding the number of countries with direct COD data is also the best way to improve indirect estimates, and more importantly, to assist national public health action and research [1–5, 28].

The major objective for most LMICs should be to obtain representative data on age-, gender-, and social strata-specific CODs (both total numbers and rates per living population) for the major diseases. Universal civil registration of deaths with medical certification remains the best long-term goal [11], but complete civil registration is achieved over several decades. Even the USA took the better part of a century to increase death certification, and some states did not reach complete coverage until the 1970s [1, 16]. Developing countries have shown little progress in the expansion of civil registration [11], and COD statistics lag even further behind [4]. Slow progress in civil registration in LMICs is mostly a result of limited access to medical care and the fact that most deaths occur at home rather than in hospitals. Incentives for households to register deaths are limited, because pension and insurance schemes or enforceable familial inheritance and property rights are uncommon in LMICs. Strategies to increase civil registration include increasing medical attention at death and training existing healthcare workers to complete death certificates [11]. Requiring burial or cremation grounds to register deaths is practicable in urban settings, but currently less so in rural areas. Efforts to accelerate COD statistics can strengthen civil registration if they work in concert with national census and statistical organizations [4, 9, 11].

There are at least five complementary and practicable approaches to obtain better COD data in the medium term (such as by 2025) [1–5].

The first and most robust option (sometimes called Sample Vital Registration with Verbal Autopsy [8, 9]) is the establishment of a Sample Registration System (SRS). Such a system has worked in representative areas of India since 1971 [9]. China has a similar but less representative system [23]. SRSs are able to continuously and prospectively collect data on the event of death and COD from enrolled households, enabling reliable determination of mortality rates in the population. Mortality rates may require adjustment, as in India, for undercounts that arise from loss to follow-up [40]. SRSs have the disadvantage of requiring more technical capacity to enable continuous enumeration compared with retrospective household surveys.

The second option is to strengthen the ongoing INDEPTH network [41, 42] of about 42 demographic surveillance sites in 19 countries (Figure 4). This would involve efforts to expand the sample size in each country and make the sites representative of the population, and to enable prospective collection of deaths. The Adult Morbidity and Mortality Project in Tanzania was able to strengthen vital statistics and survey capacity, and enabled the use of COD data to monitor and evaluate malaria, HIV/AIDS, and other national programs [42].

The third option is to carry out retrospective surveys of CODs to supplement the globally standardized Demographic and Health Surveys (DHS). These surveys were initiated by the US government about three decades ago, and now cover about 90 countries (Figure 4, [43]). The addition of COD information for children to selected DHS has reduced much of the substantial uncertainty that existed in the 1960s [7]. DHS sample frames could be used to conduct follow-up retrospective COD surveys for children and adults.

The fourth option involves carrying out large, well-planned retrospective surveys of mortality using census sample frames. The marked variation in disease in China became apparent about 3 decades ago, following a complete national survey of the causes of 20 million deaths, carried out in the period 1973 to 1975 [44]. This survey yielded the first assessment of age-specific and cause-specific mortality rates for each province and county, and for China as a whole. The study showed a large variation in disease rates across the country, which in turn brought about health interventions and further research. A similar survey covered about 1 million deaths in 24 urban and 74 rural areas of China during 1986 to 1998 [45]. More recently, Mozambique conducted a national COD survey of over 11,000 deaths, based on deaths identified in the preceding census [46].

Finally, in small countries, health departments may record significantly more deaths and COD data (through hospitals and community health centers) than is recorded through civil vital registration, as has been documented recently for Tonga and Fiji [47–49]. Similarly, a small number of countries which have reasonable levels of civil registration of death but more limited COD data, should consider expanding medically certified COD while concurrently instituting household-based verbal autopsies (VAs) for deaths that occur without medical attention [1, 4].

Lessons learned from the Million Death Study in India

Various countries, notably China and India, have introduced large-scale VA studies [9, 23]. The common lesson from these is the usefulness of even crude COD data that report on the most important diseases of public health importance. The MDS in India is a recent and one of the largest of these VA studies, so lessons from it are particularly relevant for other countries.

In the absence of reliable COD data, nationally representative VA surveys are a viable alternative. VAs involve a structured investigation of the circumstances and symptoms leading to a death, through interviewing an associate of the deceased during a household survey. VAs have inherent limitations, particularly the ability to identify only broad categories of diseases, and a high proportion of undetermined (or ‘ill-defined’) causes in people aged over 70 years [8, 50]. Even in high-income countries, COD at older ages is difficult to document [51]. However, because the mean age of death in most LMICs is far lower than in developed countries, VA remains very relevant for public health and disease control priorities. With properly standardized training and data collection, double-coding by physicians, and other quality controls, the aggregation of VA data can be extremely powerful at the population level.

Prior to 2004, it was difficult in India to accurately assess local health conditions, explore key risk factors, and evaluate the effects of investments and policies. To address this fundamental gap, the MDS was launched in 2002 [8, 9]. In collaboration with the Registrar General of India, the MDS applied simple, low-cost methods (less than US$2 per household) to record COD data from about 1 million homes throughout the country, which were randomly selected from the preceding census. The MDS focused on a simple process using local field staff who obtained key questions on all deaths, including a half-page narrative in the local language. These were converted into electronic records and coded independently by at least two physicians [8, 9]. Household participation rates have averaged close to 100%. Despite the obvious limitations of VA, the COD information in India today is many times better than the little or no data available previously. Even crude COD data applied widely to various diseases and risk factors across an entire country have yielded unexpected results and influenced disease control priorities. Some selected results from the MDS and their implications for disease control are shown below:

Faster, cheaper, and better nationally representative COD surveys

Information technology, most notably electronic capture of field data with global positioning satellite tracking and electronic coding by physicians, biological co-sampling of deaths, and use of computer algorithms and geospatial sampling, could substantially improve the validity, reliability, and feasibility of VA, while simultaneously lowering costs. Computer-assisted coding is a promising complement to physician coding, which remains the preferred standard [64]. Future combinations of algorithms might well improve or, less certainly, even replace physician coding.

The large-scale demonstration of the technical and political feasibility of the MDS, paired with the aforementioned innovations in COD surveys, suggests that it is possible for LMICs to adopt at low cost a number of methods to obtain nationally representative COD data by age and gender. A key emerging theme is the need for all such data to be fully and openly available (with protection for individual data and privacy) to ensure that these may be used imaginatively and without restriction [65]. A reasonable goal would be to ensure that all LMICs, but particularly low-income African and Asian countries, have in place large, simple, nationally representative surveys of COD by 2025 [66, 67] in time to monitor the proposed replacement for 2030 of the UN global development goals [39, 68].