Burden of disease among the world’s poorest billion people: An expert-informed secondary analysis of Global Burden of Disease estimates

Background The health of populations living in extreme poverty has been a long-standing focus of global development efforts, and continues to be a priority during the Sustainable Development Goal era. However, there has not been a systematic attempt to quantify the magnitude and causes of the burden in this specific population for almost two decades. We estimated disease rates by cause for the world’s poorest billion and compared these rates to those in high-income populations. Methods We defined the population in extreme poverty using a multidimensional poverty index. We used national-level disease burden estimates from the 2017 Global Burden of Disease Study and adjusted these to account for within-country variation in rates. To adjust for within-country variation, we looked to the relationship between rates of extreme poverty and disease rates across countries. In our main modeling approach, we used these relationships when there was consistency with expert opinion from a survey we conducted of disease experts regarding the associations between household poverty and the incidence and fatality of conditions. Otherwise, no within-country variation was assumed. We compared results across multiple approaches for estimating the burden in the poorest billion, including aggregating national-level burden from the countries with the highest poverty rates. We examined the composition of the estimated disease burden among the poorest billion and made comparisons with estimates for high-income countries. Results The composition of disease burden among the poorest billion, as measured by disability-adjusted life years (DALYs), was 65% communicable, maternal, neonatal, and nutritional (CMNN) diseases, 29% non-communicable diseases (NCDs), and 6% injuries. Age-standardized DALY rates from NCDs were 44% higher in the poorest billion (23,583 DALYs per 100,000) compared to high-income regions (16,344 DALYs per 100,000). Age-standardized DALY rates were 2,147% higher for CMNN conditions (32,334 DALYs per 100,000) and 86% higher for injuries (4,182 DALYs per 100,000) in the poorest billion, compared to high-income regions. Conclusion The disease burden among the poorest people globally compared to that in high income countries is highly influenced by demographics as well as large disparities in burden from many conditions. The comparisons show that the largest disparities remain in communicable, maternal, neonatal, and nutritional diseases, though NCDs and injuries are an important part of the “unfinished agenda” of poor health among those living in extreme poverty.

We downloaded publicly available data from the GBD Study 2017. The available data contain population, incidence, prevalence, deaths, years of life lost (YLLs), years lived with disability (YLDs), and disability-adjusted life years (DALYs) by causes of morbidity and mortality for 23 age groups, by sex, and in 195 countries and territories [1, 28, 29]. We used the estimates for the year 2017. For poverty estimates, we used a set of eight household-level indicators from the Multidimensional Poverty Index (MPI) created by the Oxford Poverty and Human Development Initiative (OPHI). Typically, the OPHI MPI is estimated by aggregating ten indicators in three dimensions including education, living standards, and health [30]. We excluded the health indicators because our analysis examined health as an outcome. We defined the poorest billion as those living in households deprived in five or more of the eight categories in our poverty index, which include child school attendance, highest educational attainment, electricity, sanitation, safe water, floor material, cooking fuel, and a set of assets (details in S1 Appendix, p 3). The data to classify populations according to these indicators came from representative country surveys (S1 Appendix, pp 4–11), and more detail on this multidimensional definition of poverty can be found elsewhere [31]. The prevalence of people in the poorest billion was estimated by sex and five-year age groups, ending with and including a group for those 80 years and older. This non-monetary approach had advantages in terms of capturing multiple dimensions of poverty, being derived from microdata that allowed for age- and sex-specific estimates, and being available across a broad set of countries [20]. The poverty indicators have clear theoretical links by which to influence disease—use of biomass fuels is associated with household air pollution, lack of sanitation and safe drinking water is linked to diarrhea and malnutrition, dirt floors provide environments for particular pathogens, maternal and childhood education have well established links to mortality, and certain household assets are linked with general resources and wealth, access to information, and mobility to access health care [20]. Prevalence of people in the poorest billion could be estimated in 105 countries from surveys since 2005. Ninety-one of these surveys (87%) were from 2010 or after. For countries with surveys prior to 2010, only Madagascar, Somalia, and Bolivia had more than one million people estimated in the poorest billion, based on 2017 populations. In addition, there were two low-income, seven lower-middle-income countries, and 22 upper-middle-income countries, according to the World Bank list of economies (calendar year 2017), in which surveys to create the poverty index were not available and in which disease burden data existed (S1 Appendix, pp 4–11) [32]. From these countries without surveys, we assumed that the prevalence of people with five or more deprivations was equal to the age- and sex-specific average by country income group among countries with surveys. This constituted less than 5% of the total population in the poorest billion. We defined the populations in high-income countries to be entirely outside the poorest billion. To define populations consistent with the GBD demographic and disease burden estimates, we applied the proportions of people in the poorest billion from the survey data in each age, sex, and location to the corresponding age-, sex-, and location-specific populations from the GBD study for the year 2017. We used the proportion of under-5-year-olds in the poorest billion as the proportion for the early neonatal (0–6 days), late neonatal (7–27 days), post neonatal (28 days to under 1 year), and 1–4 year old age groups from GBD, as well as the proportion of 80 year-olds and over in the poorest billion for the 80–84, 85–89, 90–94, 95 and older age groups. We found 838 million people with 5 or more deprivations on our poverty index across these countries. Including the low- and middle-income countries without surveys, we added 34 million additional people. In total, our “poorest billion” population contained 873 million people. Though this population was not precisely one billion, we refer to this population of interest as the poorest billion to remain consistent with literature describing this population. To create disease burden estimates for the poorest billion, we utilized GBD 2017 estimates. We employed five strategies to aggregate the burden for the world’s poorest populations. We report results from one method and the range across methods in the main text, and we show comparisons in the S1 Appendix (pp 37–119). We called the main approach presented in the results the Selective Ecological approach. The population in the poorest billion was defined as described using the household survey poverty data. Rather than using country-level burden estimates for both the poorest and non-poorest within each country (an approach we also took, S1 Appendix pp 12–13), we sought to inform estimates using ecological relationships between disease burden and poverty across countries. By age, sex, and cause, we conducted mixed-effects linear regressions predicting rates of death and YLDs across countries, using the prevalence of poorest billion population as a covariate and including a random effect for region to better isolate associations with poverty prevalence from other regional differences. We then made predictions from these models for hypothetical groups in which either 100% of the population was in the poorest billion or 0% was in the poorest billion. We scaled these estimates to the national-level GBD estimates in each country, such that the population-weighted average rate for the poorest and non-poorest in each country was consistent with the national-level estimate. We called this the Full Ecological approach (presented as one of the five approaches, S1 Appendix pp 12–15). Cross-country associations between disease rates and poverty are not necessarily consistent with the within-country associations. To understand how rates of disease may vary by socioeconomic status within LMICs to the best of our knowledge, we conducted a survey of perceived relationships between poverty and disease in LMICs among 97 health practitioners and researchers with a broad range of disease expertise and experience working in or researching health in LMICs. The S1 Appendix (pp 16–17) characterizes the participants in greater detail. Participants answered questions about their perception of relationships between rates of diseases and poverty within LMICs, indicating whether they thought occurrence (defined in survey instructions as incidence) rates and fatality (defined as case fatality) rates were (1) much higher in the non-poorest, (2) higher in the non-poorest, (3) not different, (4) higher in the poorest, or (5) much higher in the poorest. The survey also asked respondents how confident they were in their selection. To choose conditions to model using this Selective Ecological approach, we found the set of conditions with differences in rates between poorest billion and non-poorest from the expert perception survey and which showed agreement with the direction of the association in the ecological model in over half of age groups (S1 Appendix pp 16–19). To determine conditions that experts thought varied between the poorest billion and non-poorest, we treated the answers on the expert survey as ordinal and used a non-parametric one-sample two-sided sign test to test for a difference in the median from the “No Difference” response. We modeled each of these selected conditions except for those in which the number of countries with non-zero rates in a given age and sex group was too small (fewer than 10 countries) to create stable results. We used the national-level death and YLD rates for both the poorest and non-poorest within countries for the remaining conditions that did not have alignment between the expert survey and the observed ecological relationships. We conducted analysis at the lowest levels of the GBD hierarchies (most specific conditions and ages, and by sex) and aggregated to create internally consistent results. We report results for the poorest billion from the Selective Ecological approach, though we show results for other approaches in the S1 Appendix (pp 37–119) and discuss differences. These other approaches primarily differed in relation to assumptions about within-country burden gradients (assuming none or using ecological relationships to assume a gradient) and the definition of the poorest billion population (population in poorest countries or the household survey population definition). Uncertainty bounds are available on the online GBD results tool; however, random draws from the underlying distributions are used to propagate uncertainty. Without these draws, we were unable to propagate uncertainty when aggregating our results. In many cases, we present the resulting point estimate from the Selective Ecological approach and the range for all 5 approaches. While we conducted our analysis for the most specific age groups within the GBD up to the age group 95 and older, we group the ages above 80 when we present age-specific results because such a small proportion of the population in the poorest billion is above that age. We report rates, rate differences, and rate ratios of YLLs, YLDs, and DALYs to describe the overall levels as well as absolute and relative differences in morbidity and mortality between populations. We used the GBD 2017 age standard for age-standardized results [28]. Analyses were conducted using R version 3.3.1, 3.3.3, and 3.6.1 [33]. The study was submitted to the Harvard Faculty of Medicine IRB office and found exempt from further review (IRB17-0615).

• coates, matthew m.
• ezzati, majid
• aguilar, gisela robles
• kwan, gene f.
• vigo, daniel v.
• mocumbi, ana olga
• becker, anne e.
• makani, julie
• hyder, adnan a.
• jain, yogesh
• stefan, cristina
• gupta, neil
• marx, andrew
• bukhman, gene