Menu
Menu
Menu
Menu
Background Improving survival and extending the longevity of life for all populations requires timely, robust evidence on local mortality levels and trends. The Global Burden of Disease 2015 Study (GBD 2015) provides a comprehensive assessment of all-cause and cause-specific mortality for 249 causes in 195 countries and territories from 1980 to 2015. These results informed an in-depth investigation of observed and expected mortality patterns based on sociodemographic measures. Methods We estimated all-cause mortality by age, sex, geography, and year using an improved analytical approach originally developed for GBD 2013 and GBD 2010. Improvements included refinements to the estimation of child and adult mortality and corresponding uncertainty, parameter selection for under-5 mortality synthesis by spatiotemporal Gaussian process regression, and sibling history data processing. We also expanded the database of vital registration, survey, and census data to 14 294 geography–year datapoints. For GBD 2015, eight causes, including Ebola virus disease, were added to the previous GBD cause list for mortality. We used six modelling approaches to assess cause-specific mortality, with the Cause of Death Ensemble Model (CODEm) generating estimates for most causes. We used a series of novel analyses to systematically quantify the drivers of trends in mortality across geographies. First, we assessed observed and expected levels and trends of cause-specific mortality as they relate to the Socio-demographic Index (SDI), a summary indicator derived from measures of income per capita, educational attainment, and fertility. Second, we examined factors affecting total mortality patterns through a series of counterfactual scenarios, testing the magnitude by which population growth, population age structures, and epidemiological changes contributed to shifts in mortality. Finally, we attributed changes in life expectancy to changes in cause of death. We documented each step of the GBD 2015 estimation processes, as well as data sources, in accordance with Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER). Findings Globally, life expectancy from birth increased from 61·7 years (95% uncertainty interval 61·4–61·9) in 1980 to 71·8 years (71·5–72·2) in 2015. Several countries in sub-Saharan Africa had very large gains in life expectancy from 2005 to 2015, rebounding from an era of exceedingly high loss of life due to HIV/AIDS. At the same time, many geographies saw life expectancy stagnate or decline, particularly for men and in countries with rising mortality from war or interpersonal violence. From 2005 to 2015, male life expectancy in Syria dropped by 11·3 years (3·7–17·4), to 62·6 years (56·5–70·2). Total deaths increased by 4·1% (2·6–5·6) from 2005 to 2015, rising to 55·8 million (54·9 million to 56·6 million) in 2015, but age-standardised death rates fell by 17·0% (15·8–18·1) during this time, underscoring changes in population growth and shifts in global age structures. The result was similar for non-communicable diseases (NCDs), with total deaths from these causes increasing by 14·1% (12·6–16·0) to 39·8 million (39·2 million to 40·5 million) in 2015, whereas age-standardised rates decreased by 13·1% (11·9–14·3). Globally, this mortality pattern emerged for several NCDs, including several types of cancer, ischaemic heart disease, cirrhosis, and Alzheimer’s disease and other dementias. By contrast, both total deaths and age-standardised death rates due to communicable, maternal, neonatal, and nutritional conditions significantly declined from 2005 to 2015, gains largely attributable to decreases in mortality rates due to HIV/AIDS (42·1%, 39·1–44·6), malaria (43·1%, 34·7–51·8), neonatal preterm birth complications (29·8%, 24·8–34·9), and maternal disorders (29·1%, 19·3–37·1). Progress was slower for several causes, such as lower respiratory infections and nutritional deficiencies, whereas deaths increased for others, including dengue and drug use disorders. Age-standardised death rates due to injuries significantly declined from 2005 to 2015, yet interpersonal violence and war claimed increasingly more lives in some regions, particularly in the Middle East. In 2015, rotaviral enteritis (rotavirus) was the leading cause of under-5 deaths due to diarrhoea (146 000 deaths, 118 000–183 000) and pneumococcal pneumonia was the leading cause of under-5 deaths due to lower respiratory infections (393 000 deaths, 228 000–532 000), although pathogen-specific mortality varied by region. Globally, the effects of population growth, ageing, and changes in age-standardised death rates substantially differed by cause. Our analyses on the expected associations between cause-specific mortality and SDI show the regular shifts in cause of death composition and population age structure with rising SDI. Country patterns of premature mortality (measured as years of life lost [YLLs]) and how they differ from the level expected on the basis of SDI alone revealed distinct but highly heterogeneous patterns by region and country or territory. Ischaemic heart disease, stroke, and diabetes were among the leading causes of YLLs in most regions, but in many cases, intraregional results sharply diverged for ratios of observed and expected YLLs based on SDI. Communicable, maternal, neonatal, and nutritional diseases caused the most YLLs throughout sub-Saharan Africa, with observed YLLs far exceeding expected YLLs for countries in which malaria or HIV/AIDS remained the leading causes of early death. Interpretation At the global scale, age-specific mortality has steadily improved over the past 35 years; this pattern of general progress continued in the past decade. Progress has been faster in most countries than expected on the basis of development measured by the SDI. Against this background of progress, some countries have seen falls in life expectancy, and age-standardised death rates for some causes are increasing. Despite progress in reducing age-standardised death rates, population growth and ageing mean that the number of deaths from most non-communicable causes are increasing in most countries, putting increased demands on health systems. Funding Bill & Melinda Gates Foundation.
GBD employs various analytical tools and a diverse set of data sources to generate comparable estimates of deaths and mortality rates broken down by age, sex, cause, year, and geography. Multiple publications show more detail on the various aspects of the methods.7, 8, 12, 19 Part 1 of the methods appendix (pp 4–51) is a structured and succinct explanation of each step. Figure 1 shows all of the inputs, analytical processes, and outputs from the analysis of all-cause mortality and HIV/AIDS mortality, included because of its important effects on all-cause mortality in countries with large HIV epidemics, and figure 2 does the same for cause-specific mortality. Each input or process is numbered for reference, with part 2 of the methods appendix (pp 52–70) providing explanation for each step. The GBD analytical approach to estimation is guided by standardised solutions to some general analytical problems: inconsistent case definitions or coding over time or across geographies; missing data; conflicting data for the same year and geography; and population groups (eg, the poor, minorities, and vulnerable groups) who are often missed in administrative data sources. In this Article, we provide only a very high-level summary. This analysis adheres to the new Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER) proposed by the World Health Organization (WHO) and others, which includes recommendations on documentation of data sources, estimation methods, and statistical analysis.20 Table 1 shows the precise ways in which we have adhered to each element of the GATHER agreement. Estimation of all-cause mortality by age and sex and HIV/AIDS incidence, prevalence, and mortality for GBD 2015 Data and analyses are indicated by shape and the flow chart is colour coded by major estimation component. The process depicted is performed twice to bring in updated under-5 population estimates and crude death rates due to HIV/AIDS. The inputs that are updated in the second run of the process are shown by patterned boxes in this flow chart. Because of the very large and changing effects of HIV/AIDS on all-cause mortality in several countries with large HIV epidemics and limited data on all-cause mortality, the estimation of HIV/AIDS and all-cause mortality are closely linked and are presented jointly here. GBD=Global Burden of Disease. 5q0=probability of death from birth to age 5 years. 45q15=probability of death from age 15 to 60 years. ART=antiretroviral therapy. CBH=complete birth histories. CDR=crude death rate. COD=causes of death. DSP=disease surveillance points. ED=educational attainment in years per capita above age 15 years and mother’s educational attainment in years per capita for children younger than 5 years. EPP=Estimation and Projection Package. HIV CDR=crude death rate due to HIV/AIDS. LDI=lagged distributed income per capita. SBH=summary birth history. SRS=Sample Registration System. VR=vital registration. Development of the GBD 2015 cause of death database Figure shows (A) different strategies used to model different causes and to (B) combine them into a consistent set of cause-specific deaths for each location, age, sex, and year. Data and analytical processes are indicated by shape and the flow chart is colour coded by major estimation component. GBD=Global Burden of Disease. BTL=basic tabulation list. CDC=Center for Disease Control and Prevention. COD=cause of death. CODEm=Cause Of Death Ensemble model. CR=cancer registry. CRS=civil registration system. DSP=disease surveillance points. ICD=International Classification of Diseases. MI=mortality/incidence ratio. MCCD=medical certification of causes of death. MM=maternal mortality. MMR=maternal mortality ratio. MMS=maternal mortality surveillance. PAF=population-attributable fraction. SCD=survey of causes of death. SEER=Surveillance, Epidemiology, and End Results Program. SRS=Sample Registration System. SR MAD=super-region median average deviation. ST-GPR=spatiotemporal Gaussian process regression. VA=verbal autopsy. VR=vital registration. YLL=years of life lost. GATHER checklist with description of compliance and location of information in the GBD 2015 mortality and causes of death study GBD 2015=Global Burden of Disease 2015 Study. GATHER=Guidelines for Accurate and Transparent Health Estimates Reporting. We have organised geographies into a set of hierarchical categories: seven super-regions; 21 regions nested within the seven super-regions; and 195 countries and territories nested within the 21 regions (table 2). Details on the classification of each geographical unit into each level of this hierarchy are provided in the methods appendix (pp 670–83). Compared with GBD 2013, we have added seven territories—American Samoa, Bermuda, Greenland, Guam, the Northern Mariana Islands, Puerto Rico, and the Virgin Islands—because of the availability of high-quality vital registration data. These territories were not previously included in the national totals of the USA, UK, or Denmark, and were included only in GBD 2013 regional totals. We have further disaggregated data for selected countries or territories into subnational units: 26 states and one district for Brazil, 34 provinces and municipalities for China, 31 states and union territory groupings for India that include 62 rural and urban units, 47 prefectures for Japan, 47 counties for Kenya, 32 states and districts for Mexico, 13 regions for Saudi Arabia, nine provinces for South Africa, two regions for Sweden, 13 regions for the UK (Northern Ireland, Scotland, Wales, England, and nine subregions of England), and 51 states and districts for the USA. At the first subnational unit level, we have 256 geographic units. Subnational level 1 geographies in the GBD 2015 analysis include countries that have been subdivided into the first subnational level, such as states or provinces. The subnational level 2 category applies only to India and England. In this Article we present national, territory, and previously published subnational units in the UK.13 Number of geographies and causes at each hierarchical level for GBD 2015 Nations and territories includes countries, territories, and non-sovereign states. Subnational level 1 includes countries that, in the GBD analysis, have been subdivided into the first subnational level such as states or provinces. Subnational level 2 applies only to India and England. In India, states have been divided into urban and rural units. England has been divided into nine regions. For each level, the number of geographies includes the geographies at that level plus the number of most-detailed geographies at each higher level such that at each level of the hierarchy, all geographies create a collectively exhaustive and mutually exclusive set covering the world. Likewise, the GBD cause list is mutually exclusive and collectively exhaustive. The three Level 1 GBD causes consist of communicable, maternal, neonatal, and nutritional disorders; non-communicable diseases; and injuries. Level 2 causes consist of 21 cause groups, such as neoplasms and cardiovascular diseases. Levels 3 and 4 consist of disaggregated causes, such as liver cancer and cerebrovascular disease (Level 3), and liver cancer due to hepatitis C and ischaemic stroke (Level 4). GBD=Global Burden of Disease. YLD=years lived with disability. YLL=years of life lost. The GBD cause list is the crucial organising framework for the analysis of causes of death and premature mortality, as well as disease incidence and prevalence and years lived with disability.21 The GBD cause list has evolved during the 25 years of the GBD study to become a list of causes that have public health and medical care importance either because they are major causes of lost health or because of policy relevance.7, 21, 22, 23, 24 Because different levels of cause aggregation are appropriate for different purposes and users, the GBD cause list is organised hierarchically (table 2). At each level of the cause hierarchy, the set of causes is mutually exclusive and collectively exhaustive.21 At the first level of the cause list, there are three broad causes: communicable, maternal, neonatal, and nutritional diseases; NCDs; and injuries. At the second level of the hierarchy, these three causes are broken down into 21 cause groups such as neoplasms (cancers) or cardiovascular diseases. Levels 3 and 4 of the cause list provide more disaggregated causes. Based on policy interest and by approval of the GBD Scientific Council, we have added eight causes to the GBD cause list: Ebola virus disease, motor neuron disease, environmental heat and cold exposure, squamous-cell carcinoma, acute lymphoid leukaemia, chronic lymphoid leukaemia, acute myeloid leukaemia, and chronic myeloid leukaemia. Bulimia nervosa has also been added as a cause of death. In total, there are now three causes at Level 1, 21 at Level 2, 166 at Level 3, and 261 at Level 4. Some causes, such as acne, medication overuse headache, and cutaneous leishmaniasis, are not considered causes of death according to the rules of the International Classification of Diseases (ICD), so the number of causes included in this analysis of causes of death is three at Level 1, 21 at Level 2, 144 at Level 3, and 200 at Level 4. The full GBD cause list, including those for which we estimate deaths, is available in the methods appendix (pp 684–90). Because of the greater availability of data on all-cause mortality than cause-specific mortality, the all-cause mortality analysis for GBD 2015 covered 1970 to 2015. The cause of death analysis of GBD 2015 covered 1980 to 2015. A complete set of age-specific, sex-specific, cause-specific, and geography-specific death numbers and rates were generated. We present results covering different periods. However, for the main global and national results, we have focused on trends in the past decade, from 2005 to 2015, and detailed findings in 2015. Data visualisation tools are available online and provide results for each year from 1990 to 2015. Because of the very large and changing effects of HIV/AIDS on all-cause mortality in several countries with large HIV epidemics and scarce data on all-cause mortality, especially in eastern and southern Africa,11 the estimation of HIV/AIDS mortality and all-cause mortality are closely linked and presented jointly in figure 1. We divided the estimation effort into five distinct components: estimation of under-5 mortality rate (5q0); estimation of the adult mortality rate (45q15); age-specific mortality estimation; HIV/AIDS mortality estimation; and addition of the effects of events such as wars, pandemics, and disasters, which can cause abrupt discontinuities in death numbers (fatal discontinuities). Because of the interdependencies in the estimation of HIV/AIDS incidence, prevalence, and mortality and all-cause mortality, the estimation steps shown in figure 1 were repeated, with the HIV/AIDS crude death rates produced in step 4.7 used as covariates in steps 1.5, 1.11, 2.4, and 3.3 in the flow diagram. Seven types of primary data contributed to the estimation process (oval shapes in figure 1). The most important set of inputs were the data for estimating the overall level of under-5 mortality (5q0) that were obtained from vital registration systems, surveys, and censuses. Figure 3A provides information about the proportion of the 519 geographies included in the analysis for which data were available in each year from 1980 to 2015. Because of lags in reporting of both vital registration data and the release of household survey or census data, the availability of data was much lower for 2014 and 2015 than for previous years. Different data types, such as summary or complete birth histories, were processed to yield estimates for each year of the under-5 death rate; country-specific and year-specific details of the measurements are provided in the methods appendix (pp 4–19). Figure 3B shows the nature of the data and estimation process for under-5 mortality using the example of Zambia, as well as the uncorrected and bias-adjusted datapoints for each source. We used spatiotemporal Gaussian process regression to synthesise the sources and simultaneously correct for biases in specific source types.8 Bias corrections were made by comparison to reference sources, which for Zambia were the Demographic and Health Surveys. Further details of this estimation process are provided in the methods appendix (pp 4–19). Examples of under-5 mortality data availability and estimation (A) Percentage of global under-5 population covered by under-5 mortality data for each year, 1980–2015. The percentage of under-5 population covered was calculated by dividing the population of children aged 0–4 years in locations covered by available under-5 mortality data by the total global under-5 population. Because of lags in reporting of both vital registration data and the release of household survey or census data, the availability of data was much lower for 2014 and 2015 than for previous years. (B) Country-specific example of data and under-5 mortality estimates in Zambia, 1980–2015. The black line shows Gaussian process regression fit with 95% uncertainty interval shown in grey. Black circles denote reference data. Triangles denote complete birth history data. Inverted triangles denote summary birth history data. Transparent symbols are the data post-adjustment for non-sampling error. Hollow shapes represent data identified as outliers. (C) Percentage of global under-5 population covered by under-5 age-specific and sex-specific data for each year, 1980–2015. The percentage of under-5 population covered was calculated by dividing the population of children aged 0–4 years in locations covered by available under-5 age-specific and sex-specific data by the total global under-5 population. Because of lags in reporting both vital registration data and the release of household survey or census data, the availability of data was much lower for 2014 than for previous years, and no data existed for 2015. (D) Country-specific example of probability of female post-neonatal mortality in Bangladesh, 1980–2015. Standard Demographic and Health Surveys generally include large population samples and standard sets of questions. Special Demographic and Health Surveys can survey smaller, more targeted populations, such as women who have given birth. The black line shows probability of death, with 95% uncertainty interval shown in grey. Solid circles represent data sources. Hollow circles represent outliers. The post-neonatal period is 28–364 days. Because there are many sources for measuring under-5 mortality, such as summary birth histories from censuses and surveys, that do not provide sex and specific age group detail, we first estimated under-5 mortality and then split it into mortality for four age groups: early neonatal (0–6 days), late neonatal (7–28 days), post-neonatal (29–364 days), and ages 1–4 years. Splitting into these age groups was based on a statistical model using the analysis of available data that provide breakdowns by age and sex. Figure 3C shows the availability by country–year of data used to build the model to estimate mortality for specific age–sex groups younger than age 5 years. We modelled the ratio of male-to-female probability of death from birth to age 5 years as a function of both sexes’ combined under-5 mortality rate and country and regional random effects. We further disaggregated sex-specific probability of death between birth and age 5 by modelling the ratio between age-and-sex-specific probability of deaths in the early neonatal, late neonatal, post-neonatal, and 1–4 year age groups and sex-specific probability of death between birth and age 5 years. This model allowed for the association between these age-and-sex-specific probabilities and the under-5 death rate to be non-linear, and included other covariates consisting of the death rate due to HIV/AIDS in children younger than 5 years, average years of schooling among females of reproductive age, and country and regional random effects. More details, including the equations are provided in the methods appendix (p 18). Figure 3D shows an example of the empirical fit for the post-neonatal period for Bangladesh. This model was applied to all countries to generate the under-5 estimates for each geography–year. With the estimated mortality by detailed age group, we generated both deaths and population estimates for the respective age groups for each location, sex, and year. Measurements of adult mortality (45q15) were mainly derived from vital registration data and household surveys that ask about the birth and death of siblings.25 In a smaller set of cases, information was obtained from censuses or surveys about household deaths in a defined interval before the interview. Figure 4A shows the number of geographies for which data in each year were available for adult mortality estimation. Vital registration data were assessed for completeness with death distribution methods optimised for performance.26, 27 We generated a best estimate of the completeness of vital registration in each geography over time by combining estimated completeness of registration for under-5 deaths with the results for different intercensal periods of the application of three death distribution methods. These sources were combined by use of spatiotemporal Gaussian process regression—details are provided in the methods appendix (p 21). Data from sibling histories were corrected for known biases, including selection bias, zero reporter bias, and recall bias.7, 25 Our sibling history method can also deal with data sparsity in many sibling survival modules (ie, sibling history questions and variables from surveys). The predictive validity of the sibling history analytical methods has been assessed with simulated data and shown to be unbiased.25 Additionally, we compared estimates of adult mortality rates from sibling survival data with completeness-adjusted vital registration data in countries from which both sources are available and found no systematic biases from sibling survival method (methods appendix p 292).7, 25, 26 We synthesised vital registration data corrected for completeness and adjusted sibling history data into a best time series estimate of adult 45q15 using spatiotemporal Gaussian process regression. Examples of the application of these steps in three types of settings are shown in figure 4. Examples of adult mortality data availability and estimation (A) Percentage of global adult population covered by adult mortality data from vital registration systems, sibling survival surveys, sample registration systems, or censuses, 1980–2015. The percentage of available data was calculated by dividing the population of adults aged 15–59 years in locations covered by available adult mortality data by the total global population aged 15–49 years. Because of lags in reporting both vital registration data and the release of household survey or census data, the availability of data was much lower for 2014 than for previous years, and no data existed for 2015. Country-specific examples of (B) vital registration data and adult male mortality (45q15) estimation for a country with complete vital registration and large population (USA), 1980–2015; (C) vital registration data and adult male mortality (45q15) estimation for a country with complete vital registration and small population (Iceland), 1980–2015; and (D) sibling survival data and adult female mortality (45q15) estimation (Malawi), 1980–2015. Black line shows final estimates of adult mortality among males or females in each country, with 95% uncertainty interval shown in grey. Squares show sibling survival histories. 45q15=probability of death from age 15 years to 60 years. The spatiotemporal Gaussian process regression method used to fit the model to the available data included lag distributed income per capita, educational attainment, and the estimated HIV/AIDS death rate as covariates. Because the estimation of the HIV/AIDS death rate used the estimate of HIV-free mortality rate by age and sex as an input, the entire estimation loop was repeated once, which dealt with this interconnection. Step 2.9 in figure 1 deals with situations in which an inconsistency exists between the spatiotemporal Gaussian process regression-estimated adult mortality rate and the separately estimated crude death rate due to HIV/AIDS. When the HIV/AIDS death rate as estimated from the natural history model is too high compared with demographic sources, there is a risk that HIV-free death rates are depressed to implausibly low levels. In step 2.9, we scaled the HIV/AIDS crude death rate by imposing a maximum proportion of deaths that can be attributed to HIV/AIDS, as shown in our version of UNAIDS’ Spectrum model, which estimates HIV/AIDS prevalence and deaths by age and sex. Our adult mortality estimation is for ages 15–60 years (45q15), but other adult age groups that can be calculated for other purposes include 35q15 (ages 15–50 years, corresponding to the reproductive age period), and 20q50 (ages 20–70 years). In demographic estimation, measures of child mortality, adult mortality, or both are used alongside a model life table system to predict age-specific mortality.27, 28, 29, 30 The UN mostly still uses the Coale-Demeny model life tables, which were based on 192 empirical tables gathered before 1963, and in a few cases they use the 33-year-old UN Model Life Table for Developing Countries.31, 32 Murray and colleagues33 developed the Modified Logit Model Life Table system that is used by WHO to estimate age-specific mortality, which captures a much wider range of age patterns of mortality through the year 2000. The GBD approach uses three inputs to generate age-specific mortality: 5q0, 45q15, and a relevant empirical reference pattern of mortality by age.7 The reference in the GBD system was selected on the basis of empirical age patterns that are closest to the population in space and time.7 The reference was developed with a database of 16 507 age patterns of mortality from settings that meet explicit inclusion criteria as described in the methods appendix (pp 34–42). Table 3 shows a summary of the availability of empirical age–sex patterns of mortality in the GBD database. Distribution of empirical life tables by GBD super-region and decade, 1950–2015 Numbers show available empirical life tables in the GBD 2015 database. All life tables included in the database meet quality inclusion criteria whereby the observed age-specific mortality rate in an empirical life table conforms to the age pattern of mortality as described by the Gompertz–Makeham law of mortality and that observed in countries with high-quality vital and civil registration systems. GBD=Global Burden of Disease. To account for the effect of HIV/AIDS on the age pattern of mortality, the GBD model life table system for locations affected by HIV/AIDS and without high-quality vital registration data used a two-step process whereby we first estimated an HIV-free age pattern of mortality assuming that deaths due to HIV/AIDS were removed. This was accomplished by use of the HIV-free and without-fatal-discontinuity 5q0 and 45q15 estimates, crude death rates due to HIV/AIDS in age groups 0–4 and 15–59 years, and the methods detailed in the methods appendix (p 44), which reconcile the potential disconnect between HIV/AIDS mortality implied in the spatiotemporal Gaussian process regression estimates of all-cause mortality and those estimated by the Estimation and Project Package (EPP)-Spectrum. We then added the excess mortality due to HIV/AIDS to specific age groups to match the with-HIV/AIDS 5q0 and 45q15 by using the estimated age pattern of excess mortality due to HIV/AIDS for generalised and concentrated epidemics. These age patterns of excess mortality were based on ICD-10-coded vital registration data from various countries, including high-income countries with good-quality vital registration data and other middle-income nations that are affected by HIV/AIDS such as South Africa, Thailand, and Trinidad and Tobago. A list of country–years for which we obtained the empirical age pattern of HIV/AIDS excess mortality rate is shown in the methods appendix (p 313). Figure 5 shows examples of the life table system estimates of age-specific mortality compared with observed patterns for males and females in France in 2011. There was a very close fit between the estimated age-specific mortality and the observed mortality. Age-specific mortality estimation with GBD life table method versus observed data excluded from the model Country-specific examples for (A) males and (B) females in France, 2011. The red line shows the GBD 2015 life table system estimates of age-specific mortality rates from birth through age 75 years in log space, compared with observed age-specific mortality (blue line). Year 2011 selected for illustration purposes. GBD=Global Burden of Disease. Because HIV/AIDS estimation is so closely connected to all-cause mortality estimation, we discuss HIV/AIDS estimation separately here rather than in the later section about estimating other causes of death. We divided geographies into two broad groups: countries with larger epidemics and incomplete or non-existent vital registration systems and the remaining geographies. For the first group of geographies for which we had necessary information about the transmission of HIV/AIDS among adults and children and other programme information, we fitted a modified version of EPP-Spectrum11, 34 to the data on prevalence collated by UNAIDS from antenatal clinic serosurveillance and household surveys. EPP-Spectrum is a natural history model of the HIV/AIDS epidemic that has two distinct components. In the EPP component, data on the prevalence of HIV are used to back-estimate incidence of HIV. In the Spectrum component, the estimated incidence and a set of assumptions are used to estimate prevalence and deaths by age and sex. These assumptions are informed by published or unpublished cohort studies on the initial CD4 distribution of new HIV infections, rates of decline in CD4 counts, death rates on and off antiretroviral therapy (ART) differentiated by age, sex, and CD4 count, and prevention of mother-to-child transmission (PMTCT) coverage data, as well as other demographic assumptions, such as the HIV-free death rate. We have modified EPP-Spectrum to enhance the internal consistency between EPP and Spectrum and to more accurately reflect published cohort data on CD4 progression and death rates on and off ART. At this point in the estimation process for the first group of geographies, we generated two estimates of HIV/AIDS. One is informed by available data on all-cause mortality and the statistically related association between all-cause mortality and the HIV/AIDS crude death rate; the other is the EPP-Spectrum natural history model. In some locations, these estimates can be quite different. Given the inherent uncertainties in both methods, for GBD 2015 we have adopted an ensemble model which is the average of HIV/AIDS deaths for each age–sex–year from the two approaches. Figure 6 shows the results of this process using Zimbabwe as an example for incidence, prevalence, and deaths from HIV/AIDS. For comparison, we provide prevalence data from surveys and the UNAIDS estimates from their 2014 round of estimation.35, 36 Comparisons of GBD 2015 estimates and UNAIDS 2014 estimates for Zimbabwe Country-specific example comparing estimates of deaths due to HIV/AIDS (A), new HIV infections (B), and people living with HIV/AIDS (C) in Zimbabwe from GBD 2015 and UNAIDS 2014. Curves show the estimation process of a particular country and highlight the differences in results from the GBD and UNAIDS analysis of the same prevalence data. Numbers are reported in thousands. Uncertainty intervals are shown in red and blue shading. GBD=Global Burden of Disease. UNAIDS=The Joint United Nations Programme on HIV and AIDS. For the second group of geographies, we estimated mortality due to HIV/AIDS on the basis of vital registration data if available. Estimates of incidence and prevalence for HIV were based on calibrating Spectrum to match the observed numbers of HIV/AIDS deaths after accounting for under-registration of the vital registration system. This calibration method is based on tracking incidence cohorts through Spectrum and adjusting incidence to fit the observed deaths for that cohort in each year in a specific age group (methods appendix p 44). Using the methods discussed in the age-specific mortality section, with-HIV/AIDS all-cause mortality was estimated for these countries. Depending on the subgroup categorisation within the second group of geographies (methods appendix pp 44–45), we generated the HIV/AIDS-specific mortality either by applying spatiotemporal Gaussian process regression to HIV/AIDS cause-specific data from the vital registration system if the quality of vital registration was deemed high (methods appendix pp 21–26) or by using cohort incidence bias-adjusted mortality estimates from Spectrum. The fifth stage of estimation (figure 1) re-estimates all-cause mortality by incorporating the effects of HIV/AIDS and fatal discontinuities. To incorporate fatal discontinuities from natural disasters (eg, the 2011 Japan earthquake and tsunami), wars, pandemics, wildfires (eg, the Australian bushfires in 2009), or major transportation accidents (eg, the Al Ayyat train accident in Egypt in 2002), we used death counts reported in a wide range of international databases such as the International Disaster Database, the Uppsala Conflict Data Program, the International Institute for Strategic Studies Armed Conflict Databases, the Robert S Strauss Center, and various internet sources for more recent events such as the ongoing Syrian and Yemeni conflicts (databases are listed in the methods appendix [p 270], and additional sources are downloadable from the online source tool).37, 38, 39, 40 When multiple sources for the same fatal discontinuity event exist, we prioritised data from vital registration systems if it had the highest estimate, and gave least priority to data from internet searches. We constructed uncertainty on the basis of high and low estimates when available. We generated regional and cause-specific uncertainty intervals in relative terms and applied them to fatal discontinuities when only the mean estimate was provided by a specific source. The fatal discontinuity section of the methods appendix (pp 270–73) provides more detail on how we assigned fatal discontinuity deaths to different GBD causes as appropriate and how we applied a cause-specific age–sex splitting model to arrive at age-and-sex-specific deaths due to specific fatal discontinuity events. Given that all-cause mortality analysis requires estimates of crude death rates from HIV/AIDS as initial inputs and that the ensemble model changes the HIV/AIDS-specific and with-HIV/AIDS age-specific mortality rates, the all-cause mortality and HIV/AIDS estimation processes were performed twice. Crude death rates due to HIV/AIDS and under-5 population estimates were generated in the first run of the processes and propagated the second run of the processes to make the HIV/AIDS-specific and all-cause mortality processes more consistent. The GBD cause list relies on categorical attribution of deaths to a single underlying cause in accordance with the principles outlined in the ICD. The core principle of the ICD is to assign each death to only the underlying cause of death; ie, the cause that initiated the series of events leading to death. We used the ICD principle of underlying cause of death for the primary tabulations in this Article. Data from vital registration sources, verbal autopsy studies, and other sources all adhere to the same principle that one death can only have one cause. The categorical attribution of causes of death differs from a counterfactual approach, which answers the question “in the absence of the disease of interest how many deaths would not have occurred?”, similar to how we estimate burden due to risk factors in GBD. The categorical attribution of causes of death also differs from excess mortality in people with a disease followed up over time in a cohort study or through linkage of a disease registry to vital registration data. The excess mortality in such studies might contain deaths that are assigned as the underlying cause, those that are causally related to the disease, and those that are due to confounding, such as by a common underlying risk that predisposes to the disease but where there are also additional pathways to death. These counterfactual and excess mortality relationships are important and need to be quantified by considering the underlying risk, such as elevated fasting plasma glucose. Figure 2 shows the steps in the estimation of causes of death, which are divided into seven categories: cause of death database development (figure 2A), Cause of Death Ensemble modelling (CODEm), negative binomial models for rare causes, natural history models, subcause proportion models, prevalence-based models, and CodCorrect (figure 2B). For each component, we discuss the steps, with more extensive detail provided in the methods appendix (pp 52–283). Details about the modelling of HIV/AIDS and fatal discontinuities are also described in detail in the methods appendix (pp 255–64, 270–73). Figure 2A shows the detailed steps from data inputs and processing to the finalisation of the cause of death database. The methods appendix (pp 52–70) includes details on each step. Cause of death data collected through vital registration systems are available from governments and coded to different variants of the ICD including various national ICD variants. Multiple sources were used in addition to vital registration data, including verbal autopsy data, cancer registries, maternal mortality surveillance, census and survey data on maternal death, census and survey data on selected injuries, and police records for some injuries. Figure 2A shows how each type of data were processed to deal with the challenges of different coding schemes, different age group reporting, variation in certification, misclassification of HIV/AIDS deaths, misclassification of maternal HIV/AIDS deaths, and incorporation of population-based cancer registry data. The first and second steps in the cause of death database development were standardisation of multiple data formats to a single GBD standard, then the mapping of each ICD or verbal autopsy variant to the GBD cause map. Figure 7A shows the number of deaths captured for each year in the GBD causes of death database by coding version. In step 3, we split a small subset of data reported in non-GBD-standard age formats into GBD age categories using the global relative age pattern of mortality for each cause as estimated from the pooled data that provide full age detail. In step 4, based on expert judgment, some causes were not allowed for certain age–sex groups, for example, male uterine cancer. Availability and quality of cause of death data in the GBD 2015 database (A) Total deaths with a WHO-standard death certificate available in the GBD 2015 cause of death database classified by the variant of the International Classification of Diseases used for reporting. Cause of death data have been reported in national variants of ICD-8, ICD-9, and ICD-10 during the interval 1980–2015. Because of lags in reporting of both vital registration data and the release of household survey or census data, the availability of data was much lower for 2014 than for previous years and no data existed for 2015. (B) Percentage of global population covered by cause-specific data in the cause of death database for GBD 2015, 1980–2015; the percentage of available data was calculated by dividing the population of locations covered by available cause-specific data by the total global population. This figure is computed using vital registration, verbal autopsy, maternal, cancer, and injury sources. (C) Overall classification of each GBD subnational level 1 geography by availability and quality of cause of death data for the period 1980 to 2015. Countries have been assigned on the basis of the available time series of data into one of six categories. The figure uses GBD subnational level 1 geographies because subnational level 2 cannot be easily seen on a map. Extensive complete representative vital registration was defined as 25 total years or more of vital registration data with an estimated 95% completeness or above. All geographies that do not meet the threshold for extensive complete representative vital registration are classified as one of the following: limited years of complete representative vital registration, defined as 5 years or more of vital registration data with an estimated 95% completeness or above; incomplete representative vital registration, defined as at least 1 year of vital registration data with an estimated 70% completeness or above; more than 200 cause-years VA or non-representative VR, defined as more than 200 cause-years of verbal autopsy or at least 1 year of vital registration with an estimated 50% completeness or above; less than 200 cause-years of VA; or no data. Cause-years are defined as the number of years for each cause for which data are available. GBD=Global Burden of Disease. ICD=International Classification of Diseases. BTL=basic tabulation list. VA=verbal autopsy. VR=vital registration. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. FSM=Federated States of Micronesia. In step 5, deaths assigned to causes that cannot be underlying causes of death (ie, garbage coded) were reassigned to their likely underlying cause of death.4, 7 These redistribution algorithms are based on three approaches. For some garbage codes, such as senility or old age, deaths were proportionately reassigned to all causes that are not garbage codes for a country–age–sex–year. For HIV/AIDS in many countries, deaths from HIV/AIDS have been misclassified as opportunistic infections, tuberculosis, cancer, digestive diseases, and immune deficiencies. In step 6, using methods developed by Birnbaum and colleagues,41 these deaths were identified and reclassified as HIV/AIDS in select countries with evidence of misclassification. In step 7, data from the China Center for Disease Control and Prevention (CDC) vital registration system were re-weighted to take into account potential selection bias caused by a larger fraction of deaths being captured in hospital than out of hospital in some locations.14 Step 8 ensured that the process of redistributing garbage codes or identifying misclassified HIV/AIDS deaths would not assign deaths to causes in an age–sex–country–year that violated age–sex or other restrictions. Step 9 excluded vital registration sources that were less than 50% complete in a given geography from the database, because of the potential for selection bias in highly incomplete sources. Sources estimated to be 50–70% complete were identified as non-representative, which was information that we used in the building of the cause of death statistical model to increase the estimated data variance for these datapoints. All included sources were corrected to be 100% complete by multiplying the cause fraction in a source for a country–age–sex–year by the estimate of all-cause mortality for that country–age–sex–year. Step 10 aggregated causes of death from most to least detailed levels of the GBD hierarchy, ensuring deaths for a given cause were representative of all branches of the hierarchy that fall beneath it. In step 11, deaths due to HIV/AIDS and various types of fatal discontinuities were removed before cause fractions were computed. Because of the very large effects of fatal discontinuities, such as wars and natural disasters in some cases, and the impact of HIV/AIDS in countries with large epidemics, we converted cause fractions to be cause fractions excluding HIV/AIDS and fatal discontinuities in the denominator. Deaths from HIV/AIDS and the fatal discontinuities were added back during the final stages of the modelling process. Because many sources on maternal mortality identify deaths during pregnancy and the post-partum period and not maternal deaths, the separation of HIV/AIDS deaths during pregnancy and HIV/AIDS deaths aggravated by pregnancy was more complicated (methods appendix p 66). Figure 7B provides information about the fraction of the 519 geographies in the analysis for which cause of death data were available in each year from 1980 to 2015 for any cause, including maternal death and injuries. Data availability by geography–year by cause is shown in the methods appendix (pp 318–401). To facilitate understanding of the range of quality and availability of data for each geography, we classified geographies into six categories: extensive complete representative vital registration (vital registration data that are 95% complete and cover more than 25 years); moderate data (vital registration data that are 95% complete but cover fewer than 25 years); incomplete representative vital registration (all other geographies with some representative vital registration data); extensive verbal autopsy and other sources (covering more than 20% of cause-years); limited verbal autopsy or other data (all others with some data available); and no data for any cause (methods appendix pp 691–710). Figure 7C shows this designation for each class of country. Figure 2B shows the analytical flow chart for modelling different causes of death and combining them into internally consistent estimates of cause-specific mortality that sum to all-cause mortality with uncertainty levels. 167 individual causes of death were modelled using CODEm. Developed for GBD 2010,5 CODEm tests a large number of model specifications, comparing different functional forms and permutations of relevant covariates for each cause of death. Models that met requirements for direction and significance of the regression coefficients were then evaluated for out-of-sample predictive validity through multiple iterations of cross-validation testing. We then combined these models into an ensemble, weighting them such that top performing models (in terms of out-of-sample prediction error on levels and trends) contributed the most to the final prediction. Out-of-sample predictive validity testing was also used to select the psi parameter that determines the number of models and their weight in the final ensemble (figure 8). Distribution of out-of-sample model performance for CODEm models used for GBD 2015 Model performance was assessed by use of the root mean square error of the ensemble model predictions of the log of the age-specific death rates for a cause assessed with 15% of the data held out from the statistical model building. The figure shows the distribution of root mean square error across the set all models for all causes. Model performance varies substantially across causes. GBD=Global Burden of Disease. CODEm=cause of death ensemble modelling. For each cause of death, we ran independent CODEm models by sex and for countries with extensive complete vital registration representation and all other countries. We included all datapoints for the other categories of geographies, whereas for countries with extensive complete vital registration representation, we included only datapoints from those countries, so that heterogeneous data from other countries did not inflate the uncertainty interval. For ten causes of death, the number of events are so low, including many zero counts in countries with high income per capita or high educational attainment, that CODEm out-of-sample predictive validity testing was unstable. For these rare causes of death, which included other intestinal infectious diseases, upper respiratory infections, diphtheria, varicella and herpes zoster, malaria, schistosomiasis, cysticercosis, cystic echinococcosis, ascariasis, and iodine deficiency, we used negative binomial regression to develop simple models to predict deaths. More details are available in the methods appendix (pp 185–200; negative binomial models). For some causes, deaths are rarely recorded in either vital registration data or verbal autopsy data. Partly, this is because of the geographical location of the deaths or because of the potential for systematic bias in vital registration data or verbal autopsy data. For 14 causes, we have developed natural history models in which incidence and case-fatality rates are modelled separately and combined to yield estimates of cause-specific mortality. We developed natural history models for typhoid fever, paratyphoid fever, whooping cough, measles, visceral leishmaniasis, African trypanosomiasis, yellow fever, syphilis (congenital), and acute hepatitis A, B, C, and E. Additionally, for malaria in sub-Saharan Africa, we have used a natural history model based on the incidence estimated by the Malaria Atlas Project and age–sex-specific case-fatality rates estimated from available data. Further details on the development of these natural history models are available in the methods appendix (pp 201–26; natural history models). For meningitis, maternal disorders, liver cancer, cirrhosis, and chronic kidney disease, we estimated detailed causes for each of these cause groupings by modelling the proportion of the cause grouping (parent cause) due to each of the component causes. We used this approach because the available data on the specific causes can come from sources other than vital registration, such as end-stage renal disease registries, or from too few places to model the death rates directly. For these causes, the parent cause was first estimated with CODEm and the fraction of the parent due to each component cause for each age–sex–geography–year was generally estimated with DisMod-MR 2.1, a Bayesian meta-regression method developed for the GBD studies.42, 43 Details for each cluster of causes analysed in this way are shown in the methods appendix (pp 233–52; subcause proportion models). For Alzheimer’s disease and other dementias and atrial fibrillation and flutter, there is evidence of marked changes over time in the propensity of individuals who completed death certificates to list these causes as underlying causes of death.44, 45 These changes created increases in the reported death rates. Conversely, prevalence surveys do not show a matching increase in age-specific disease prevalence. Garbage code redistribution algorithms used in the development of the cause of death database have so far not accurately captured this shift over time in the certification of underlying causes of death. For these two causes, we based our estimates on prevalence surveys and estimates of excess mortality based on deaths certified in countries with the greatest proportion of deaths allocated to the correct underlying cause of death in recent years. In both cases, more detail is available in the methods appendix (pp 227–32; prevalence-based models). We developed models for prevalence and excess mortality using DisMod-MR 2.1. Depending on the specific data availability and details of individual causes, we adopted different modelling strategies for each cause. We generated a set of underlying cause of death estimates, with uncertainty intervals, that equalled all-cause mortality, with uncertainty intervals, for each age–sex–year–geography and cause and all-cause mortality at the individual draw level.24 In CodCorrect, for each draw from the posterior distribution of each cause, the sum of cause-specific estimates is rescaled to equal the draw from the all-cause distribution (methods appendix p 285). We used a counterfactual analysis approach to estimate aetiology-specific population attributable fraction for mortality due to lower respiratory infections and diarrhoeal diseases. This approach involved analysing changes in mortality on the basis of the estimated prevalence of each pathogen and relative risk of developing disease given pathogen exposure. The prevalence of each pathogen in diarrhoeal cases was extracted from a systematic literature review and modelled with DisMod-MR 2.1. The odds ratios of an episode of diarrhoea given exposure to the pathogen were estimated from a reanalysis of the Global Enteric Multicentre Study (GEMS) that used the TaqMan Array Card (TAC), which is based on a quantitative polymerase chain reaction diagnostic (qPCR).46, 47 We attributed mortality to all pathogens, even if the odds ratio was not significant in all age groups. We corrected the estimated prevalence for each pathogen on the basis of conventional laboratory techniques, such as bacterial culture or enzyme-linked immunosorbent assay (ELISA), to be consistent with the new qPCR method. Cholera mortality was estimated by modelling the under-reporting to the WHO cholera case notification system and applying this correction factor to estimate the number of cholera cases and deaths (methods appendix p 281). The incidence and mortality of Clostridium difficile was modelled with natural history and incidence data in DisMod-MR 2.1. We estimated attributable mortality due to respiratory syncytial virus and influenza with a similar approach to that for diarrhoea. We used a counterfactual approach whereby the prevalence in patients with lower respiratory infection was extracted from a systematic literature review and modelled with DisMod-MR 2.1. The odds ratios of lower respiratory infections given pathogen presence were obtained from a meta-analysis by Shi and colleagues.48 We adjusted the population attributable fraction for lower respiratory infection mortality due to respiratory syncytial virus and influenza for the relative case-fatality rate of viral to bacterial pneumonia episodes by age. Haemophilus influenzae type b and pneumococcal pneumonia (Streptococcus pneumonia) were estimated with a vaccine probe approach whereby the attributable fraction was calculated as the ratio of vaccine efficacy against non-specific pneumonia to vaccine efficacy against pathogen-specific and serotype-specific pneumonia. Studies that report vaccine efficacy against vaccine-type invasive pneumococcal disease were adjusted for the relative efficacy against vaccine-type clinical pneumococcal pneumonia using a uniform distribution of uncertainty around this ratio.49, 50 In this Article, we built on GBD 201351 concepts by improving the interpretability of sociodemographic status and characterising and describing this relationship in more detail for years of life lost due to premature mortality (YLLs), as well as highlighting changes in age-standardised death rates, population age structure, and YLL rates. We have made two important changes to the GBD 2013 computation. First, we have used only lag-dependent income per capita, average educational attainment in the population over age 15 years, and the total fertility rate. We excluded the mean age of the population because it is directly affected by death rates. Second, we have applied the methods used to compute the Human Development Index to generate an interpretable scale, resulting in the Socio-demographic Index (SDI).52 The Human Development Index method weights each component equally and rescales each component on a zero-to-one scale with zero being the lowest value observed in the time period 1980 to 2015 and 1 being the highest value observed. The final composite SDI value is the geometric mean of each of the components. The SDI ranges from 0·060 in Mozambique in 1987 to 0·978 in Washington, DC, USA, in 2015. The correlation of the SDI with the sociodemographic status principal component analysis used in GBD 2013 was 0·982. The very high correlation is because the principal component analysis yields weights that are nearly equal across components. The advantage of the index is that 1 can be interpreted as the level of SDI at which a geography has the highest observed log income per capita and educational attainment and lowest fertility rate. We tested whether alternative lags of the components of SDI would provide a better predictor of outcomes such as life expectancy and age-specific probabilities of death. Using lag distributed income per capita, educational attainment, and the total fertility rate in the current year was the most predictive of these mortality outcomes (methods appendix p 286). To report on aggregate results, we divided geographies into SDI quintiles in 2015. Quintile cutoffs were based on the entire distribution of geography–years from 1980 to 2015, excluding populations smaller than 1 million. Figure 9 shows a map of the SDI level in 2015 categorised into five groups including subnational geographies. Because SDI includes educational attainment and the total fertility rate, some countries which have very high income, such as Saudi Arabia, are classified in the second quintile of SDI because of lower educational attainment and higher fertility rates.53 SDI quintiles by GBD subnational level 1 geography, 2015 SDI is calculated for each geography as a function of lag-dependent income per capita, average educational attainment in the population older than age 15 years, and the total fertility rate. SDI units are interpretable; a zero represents the lowest level of income per capita and educational attainment and highest total fertility rate observed during 1980–2015, whereas a one represents the highest income per capita and educational attainment and lowest total fertility rate observed in the same period. Cutoffs on the SDI scale for the quintiles have been selected on the basis of examination of the entire distribution of geographies 1980–2015. GBD=Global Burden of Disease. SDI=Socio-demographic Index. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. FSM=Federated States of Micronesia. To capture the average relationships for each age–sex–cause group, we used spline regression of death rates on SDI (methods appendix pp 285–86). To ensure a coherent set of estimated death rates for Levels 1, 2, and 3 in the GBD cause hierarchy for each level of SDI, the Level 2 death rates were rescaled such that for each age–sex–cause bin, the sum of Level 2 death rates equalled the Level 1 death rate. This procedure was repeated for Level 3 and Level 2 causes. These rates were used as the expected death rates by age–sex–cause and SDI. Various summary measures have been computed on the basis of the age–sex–cause-specific predictions based on SDI, including age-standardised death rates, age-standardised YLL rates, and life expectancy at birth. To further characterise how patterns of crude death rates and death numbers change with SDI, we have computed the average population age structure associated with each SDI level. These population age structures have then been used to estimate how crude death rates and death numbers by cause are expected to change with rising SDI. To analyse the drivers of change in the numbers of deaths by cause or geography, we decomposed change from 2005 to 2015 into three explanatory components: change due to growth of the total population; change in the population structure by age or sex; and change in age-specific, sex-specific, and cause-specific rates. We refer to all changes in age-specific, sex-specific, and cause-specific death rates not explained by demographic change (population growth and ageing) as the epidemiological change. The observed change in the total number of deaths equals the net change of these three components. Decomposition analyses for 1980 to 2015 and 2000 to 2015 are shown in the results appendix (pp 6–7). The decomposition analysis uses methods developed in demographic research by Das Gupta.54 As an example, we describe our approach to decomposition for the 2005 to 2015 period. We used counterfactual scenarios to calculate two different sets of numbers for death. In the first scenario, for population growth, the number of deaths in 2015 was the number expected if the total population increased from 2005 as observed, but the age–sex-specific population structure and rates of death were the same in 2015 as in 2005. In the second scenario, for population growth and ageing, the number of deaths in 2015 was the number expected according to the 2015 age–sex-specific population structure, but with the age–sex-specific rates of death held constant to 2005. The difference between the number of deaths observed in 2005 and those estimated for 2015 with the population growth scenario is the change in the number of deaths exclusively from population growth. The difference between the scenario for population growth alone and the scenario for population growth and ageing is the change in the number of deaths exclusively attributable to population ageing. When considering the estimated levels and changes in all-cause and cause-specific mortality rates for each geographical area covered by GBD 2015, it is important to understand the relative contribution of changes in mortality due to each cause to the overall changes in life expectancy at birth during the same period. To examine the changes in life expectancy at birth between 2005 and 2015, we have applied the state-of-the-art life expectancy cause-specific decomposition method developed by Beltran-Sanchez, Preston, and Canudas-Romo.55 We computed YLLs using the standard GBD methods whereby each death is multiplied by the normative standard life expectancy at each age. The normative standard life expectancy at birth is 86·59 years, which is based on the lowest observed death rates for each 5-year age group in populations larger than 5 million. For GBD 2015, we computed age-standardised mortality rates and YLL rates from the updated world population age standard developed for GBD 2013.7 Details of the GBD world population age standard are available in the methods appendix (pp 286–287 and 314). To account for uncertainties that arise from sample sizes of data, adjustments to sources of all-cause mortality, model specifications in spatiotemporal Gaussian process regression and model life table systems, and cause-specific model specifications and estimation, we have estimated uncertainty intervals in key steps of the all-cause mortality and cause-specific mortality estimation processes. We have produced 1000 draws of all mortality metrics, including under-5 mortality rate, adult mortality rate, age-specific mortality rate and envelope, and cause-specific mortality rates and death numbers for each location by sex for all years covered by each analytical step from the posterior distribution in the estimation process. This allowed the quantification and propagation of uncertainty into the final quantities of interest. Because of computational time limitations, we have not propagated uncertainty in covariates used in cause of death models, nor have we been able to propagate uncertainty in garbage code redistribution algorithms into the final results. Our tests on the estimation of under-5 mortality rates show that the incorporation of uncertainty for included first stage model covariates such as crude death rate due to HIV/AIDS does not have a significant impact on the final estimates (data not shown). The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to the data in the study and had final responsibility for the decision to submit for publication.
