Menu
Menu
Menu
Menu
Background: Global development goals increasingly rely on country-specific estimates for benchmarking a nation’s progress. To meet this need, the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2016 estimated global, regional, national, and, for selected locations, subnational cause-specific mortality beginning in the year 1980. Here we report an update to that study, making use of newly available data and improved methods. GBD 2017 provides a comprehensive assessment of cause-specific mortality for 282 causes in 195 countries and territories from 1980 to 2017. Methods: The causes of death database is composed of vital registration (VR), verbal autopsy (VA), registry, survey, police, and surveillance data. GBD 2017 added ten VA studies, 127 country-years of VR data, 502 cancer-registry country-years, and an additional surveillance country-year. Expansions of the GBD cause of death hierarchy resulted in 18 additional causes estimated for GBD 2017. Newly available data led to subnational estimates for five additional countries—Ethiopia, Iran, New Zealand, Norway, and Russia. Deaths assigned International Classification of Diseases (ICD) codes for non-specific, implausible, or intermediate causes of death were reassigned to underlying causes by redistribution algorithms that were incorporated into uncertainty estimation. We used statistical modelling tools developed for GBD, including the Cause of Death Ensemble model (CODEm), to generate cause fractions and cause-specific death rates for each location, year, age, and sex. Instead of using UN estimates as in previous versions, GBD 2017 independently estimated population size and fertility rate for all locations. Years of life lost (YLLs) were then calculated as the sum of each death multiplied by the standard life expectancy at each age. All rates reported here are age-standardised. Findings: At the broadest grouping of causes of death (Level 1), non-communicable diseases (NCDs) comprised the greatest fraction of deaths, contributing to 73·4% (95% uncertainty interval [UI] 72·5–74·1) of total deaths in 2017, while communicable, maternal, neonatal, and nutritional (CMNN) causes accounted for 18·6% (17·9–19·6), and injuries 8·0% (7·7–8·2). Total numbers of deaths from NCD causes increased from 2007 to 2017 by 22·7% (21·5–23·9), representing an additional 7·61 million (7·20–8·01) deaths estimated in 2017 versus 2007. The death rate from NCDs decreased globally by 7·9% (7·0–8·8). The number of deaths for CMNN causes decreased by 22·2% (20·0–24·0) and the death rate by 31·8% (30·1–33·3). Total deaths from injuries increased by 2·3% (0·5–4·0) between 2007 and 2017, and the death rate from injuries decreased by 13·7% (12·2–15·1) to 57·9 deaths (55·9–59·2) per 100 000 in 2017. Deaths from substance use disorders also increased, rising from 284 000 deaths (268 000–289 000) globally in 2007 to 352 000 (334 000–363 000) in 2017. Between 2007 and 2017, total deaths from conflict and terrorism increased by 118·0% (88·8–148·6). A greater reduction in total deaths and death rates was observed for some CMNN causes among children younger than 5 years than for older adults, such as a 36·4% (32·2–40·6) reduction in deaths from lower respiratory infections for children younger than 5 years compared with a 33·6% (31·2–36·1) increase in adults older than 70 years. Globally, the number of deaths was greater for men than for women at most ages in 2017, except at ages older than 85 years. Trends in global YLLs reflect an epidemiological transition, with decreases in total YLLs from enteric infections, respiratory infections and tuberculosis, and maternal and neonatal disorders between 1990 and 2017; these were generally greater in magnitude at the lowest levels of the Socio-demographic Index (SDI). At the same time, there were large increases in YLLs from neoplasms and cardiovascular diseases. YLL rates decreased across the five leading Level 2 causes in all SDI quintiles. The leading causes of YLLs in 1990—neonatal disorders, lower respiratory infections, and diarrhoeal diseases—were ranked second, fourth, and fifth, in 2017. Meanwhile, estimated YLLs increased for ischaemic heart disease (ranked first in 2017) and stroke (ranked third), even though YLL rates decreased. Population growth contributed to increased total deaths across the 20 leading Level 2 causes of mortality between 2007 and 2017. Decreases in the cause-specific mortality rate reduced the effect of population growth for all but three causes: substance use disorders, neurological disorders, and skin and subcutaneous diseases. Interpretation: Improvements in global health have been unevenly distributed among populations. Deaths due to injuries, substance use disorders, armed conflict and terrorism, neoplasms, and cardiovascular disease are expanding threats to global health. For causes of death such as lower respiratory and enteric infections, more rapid progress occurred for children than for the oldest adults, and there is continuing disparity in mortality rates by sex across age groups. Reductions in the death rate of some common diseases are themselves slowing or have ceased, primarily for NCDs, and the death rate for selected causes has increased in the past decade. Funding: Bill & Melinda Gates Foundation.
GBD cause of death estimation incorporates methods to adjust for incomplete or missing vital registration (VR) and verbal autopsy (VA) data, general heterogeneity in data completeness and quality, and the redistribution of so-called garbage codes (insufficiently specific or implausible cause of death codes). A general description of these methods is provided in this section, with further detail presented in appendix 1. GBD 2017 complied with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER)15 statement (appendix 1 section 1.3). Analyses were completed with Python version 2.7.14, Stata version 13.1, and R version 3.3.2. Statistical code used for GBD estimation is publicly available online. The locations included in GBD 2017 have been arranged into a set of hierarchical categories composed of seven super-regions and a further nested set of 21 regions containing 195 countries and territories (appendix 1). Each year, GBD includes subnational analyses for a few new countries and continues to provide subnational estimates for countries that were added in previous cycles. Subnational estimation in GBD 2017 includes five new countries (Ethiopia, Iran, New Zealand, Norway, Russia) and countries previously estimated at subnational levels (GBD 2013: China, Mexico, and the UK [regional level]; GBD 2015: Brazil, India, Japan, Kenya, South Africa, Sweden, and the USA; GBD 2016: Indonesia and the UK [local government authority level]). All analyses are at the first level of administrative organisation within each country except for New Zealand (by Māori ethnicity), Sweden (by Stockholm and non-Stockholm), and the UK (by local government authorities). All subnational estimates for these countries were incorporated into model development and evaluation as part of GBD 2017. To meet data use requirements, in this publication we present all subnational estimates excluding those pending publication (Brazil, India, Japan, Kenya, Mexico, Sweden, the UK, and the USA); because of space constraints these selected subnational results are presented in appendix 2. Subnational estimates for countries with populations larger than 200 million (measured with our most recent year of published estimates) that have not yet been published elsewhere are presented wherever estimates are illustrated with maps but are not included in data tables. The complete cause-specific estimation results include the years 1980 through to 2017, and are available for exploration by an online data visualisation tool. To better support current health policy assessment, we include a subset of analyses in the current study featuring the most recent interval, 2007–17. The GBD study attributes each death to a single underlying cause that began the series of events leading to death, in accordance with ICD principles. The GBD study organises causes of death in a hierarchical list containing four levels (appendix 1 section 7). At the highest level (Level 1), all disease burden is divided among three mutually exclusive and collectively exhaustive categories: communicable, maternal, neonatal, and nutritional (CMNN) diseases; non-communicable diseases (NCDs); and injuries. Level 2 distinguishes these Level 1 categories into 21 cause groups, such as cardiovascular diseases; diarrhoeal diseases, lower respiratory infections (LRIs), and other common infectious diseases; or transport injuries. Level 3 disaggregates these causes further; in most cases this disaggregation represents the finest level of detail by cause, such as stroke, ischaemic heart disease, or road injuries. Where data are sufficiently available or specific policy relevance has been sought, selected causes are further disaggregated at Level 4, such as drug-susceptible tuberculosis, multidrug-resistant tuberculosis without extensive drug resistance, and extensively drug-resistant tuberculosis. For GBD 2017, the cause hierarchy was further refined to separately estimate causes with substantial policy interest or high levels of burden. Specific changes included separate estimation of non-rheumatic calcific aortic and degenerative mitral valve diseases, and myelodysplastic, myeloproliferative, and other haemopoietic neoplasms, resulting in a reduction in the estimates of some residual causes. Disaggregation of residual causes also allowed separate estimation of type 1 and type 2 diabetes, chronic kidney disease due to type 1 and type 2 diabetes, poisoning by carbon monoxide, liver cancer due to non-alcoholic steatohepatitis (NASH), subarachnoid haemorrhage, ectopic pregnancy, and invasive non-typhoidal salmonella. Maternal and neonatal disorders, previously estimated as separate cause groupings at Level 2 of the hierarchy, were estimated for GBD 2017 at Level 3 of the hierarchy, and then aggregated up to Level 2 to better capture the epidemiological connections and linked burden between them. The complete hierarchy of causes included in GBD 2017 and their corresponding ICD9 and ICD10 codes are described in appendix 1 (section 7). The GBD cause of death database consists of VR and VA data; survey and census data for injuries and maternal mortality; surveillance data for maternal mortality and child death; cancer registries; and police records for interpersonal violence and road injuries. Self-harm estimates incorporate VR data and are based on ICD categorisation as described in appendix 1 (section 7). In this iteration of GBD, ten new VA studies and 127 new country-years of VR data were added at the country level. 502 new cancer-registry country-years were added, as was one additional new surveillance country-year. Data sources comprising the GBD cause of death database can be reviewed on the Global Health Data Exchange website. Multiple factors can influence changes between GBD studies in estimates for a given cause-location-year, including the quality of a country’s data system (as represented by the GBD star rating system) and the addition of more recent data. Figure 1 shows the relative stability of GBD estimates between study iterations. Variation between GBD 2016 and GBD 2017 estimates was greater in countries with both low star ratings and no new VR data updates occurring between these iterations of the study. Changes to estimates can be seen even in high star rating locations because of changes in modelling strategy or model covariates even when no new VR data were available between cycles. Effect of new VR data on Level 1 cause estimates from GBD 2016 to GBD 2017, based on national locations with varying quality of VR data, 2008–16 The figure shows the degree of consistency between GBD 2016 and GBD 2017 estimates for Level 1 causes at the national level from 2008 to 2016. The diagonal line represents no change from GBD 2016 to GBD 2017. Each point represents one country-year, with colours indicating the Level 1 cause grouping (communicable, maternal, neonatal, and nutritional diseases; non-communicable diseases; and injuries). Panels indicate whether or not any new VR data between 2008 and 2016 were added for that location for GBD 2017, and whether or not a location has 4-star or 5-star VR quality. Points that are outside of the standard 95% prediction interval for a linear regression of 2017 values on 2016 values are annotated (if the same location-cause had multiple points in a time series, only the furthest-most point was annotated). The Spearman’s correlation coefficient is noted in the lower right-hand corner of each panel. CSMR=cause-specific mortality rate. GBD=Global Burden of Diseases, Injuries, and Risk Factors Study. VR=vital registration. To standardise cause of death data, we used protocols to address the minor proportion of deaths that were assigned to age groups broader than the GBD five-year age groups or were not assigned an age or sex, and to address differences in ICD codes due to national variation or revision, as described in appendix 1 (section 2). Garbage codes, deaths with non-specific codes (eg, unspecified stroke), deaths assigned to ICD codes that could not be underlying causes of death (eg, senility), or deaths assigned to intermediate but not underlying causes of death (eg, heart failure), were redistributed by age, sex, location, and year to the most likely causes of death. Methods used for this redistribution included regression models, redistribution based on fixed proportions, proportional reassignment, and fractional assignment of a death assigned to multiple causes, as developed by Naghavi and colleagues16 and detailed in appendix 1 (section 2.7). We excluded all data sources with more than 50% of deaths assigned to major garbage codes (those at Level 1 or Level 2 of the GBD hierarchy) in any location-year to mitigate the potential for bias from these sources. The proportion of VR data assigned to major garbage code categories for each location-year is shown, with supporting detail, in appendix 1 (section 7). New to GBD 2017, the uncertainty around redistribution methods was also estimated. Additional details for this process are provided in appendix 1 (section 2.7). Because mortality due to HIV/AIDS is sometimes coded to other causes of death such as tuberculosis, meningitis, or toxoplasmosis, we also corrected the cause of death assignment to HIV/AIDS for peak epidemic times. Tuberculosis deaths can be misclassified as pneumonia deaths in children in locations with a high tuberculosis burden. Methods to adjust for this potential misclassification are described in detail in appendix 1 (section 3.3). Mortality rates from dementia and Parkinson’s disease reported in VR systems cannot be reconciled with observed trends in prevalence and excess mortality—a disparity that can be attributed to variation in death certification practices for these causes across countries and over time.17 For GBD 2017, we sought to address this known bias by using details from multiple cause of death data. For GBD 2017, multiple cause of death data were available to investigators only for the USA, where recent years show improved use of previously under-utilised codes such as dementia. Statistical models of these USA data were used to reclassify deaths from other GBD causes and garbage codes to dementia and Parkinson’s disease according to the pattern of intermediate and immediate causes observed in the most recent years. Model results were applied to all countries. A similar reallocation process was used for atrial fibrillation deaths misclassified as deaths due to heart failure or thromboembolic events. A detailed description of these redistribution procedures and the manner in which they were applied to all countries is available in section 2 of appendix 1. This reallocation is illustrated in appendix 1 (section 7). For the first time in GBD 2017, we separately estimated deaths from diabetes by type. Deaths due to diabetes can be reported in VR and VA data as type 1, type 2, or unspecified. Two data manipulation steps were necessary. First, we assumed all deaths reported in individuals younger than 15 years were type 1 regardless of the original code assignment. Second, we redistributed unspecified diabetes deaths on the basis of a regression in which the true proportions of type 1 and type 2 deaths by age-sex-location-year are a function of the proportion of unspecified deaths, age, the age-standardised prevalence of obesity, and an interaction term for age and obesity prevalence. These methods are described in detail in appendix 1 (section 3.3). Completeness of VR data was assessed by location-year, and sources with less than 50% completeness were excluded. We multiplied the estimated all-cause mortality for each age-sex-location-year by the cause fraction for the corresponding age-sex-location-year to adjust all included sources to 100% completeness. VA and VR data availability and completeness are shown for each location-year in appendix 1 (section 7). To further characterise the quality of data available in each country, the GBD study rated each location-year from 1980 to 2017 on a level of 0 to 5 stars according to methods previously described.18 Ratings convey an overall measure of the reliability of cause of death estimates for each location-year but do not directly affect the estimation process. The GBD Cause of Death Ensemble model (CODEm) systematically tested and combined results from different statistical models according to their out-of-sample predictive validity. Results are incorporated into a weighted ensemble model as detailed in appendix 1 (section 3.1) and below. For GBD 2017, CODEm was used to estimate 192 causes of death (appendix 1 section 7). To predict the level for each cause of death, we used CODEm to systematically test a large number of functional forms and permutations of covariates.18 Each resulting model that met the predetermined requirements for regression coefficient significance and direction was fit on 70% of the data, holding out 30% for cross-validation (appendix 1 section 3.1). Out-of-sample predictive validity of these models was assessed by use of repeated cross-validation tests on the first 15% of the held-out data. Various ensemble models with different weighting parameters were created from the combination of these models, with the highest weights assigned to models with the best out-of-sample prediction error for trends and levels, as detailed in appendix 1 (section 7). Model performance of these ensembles was assessed against the root-mean squared error (RMSE) of the ensemble model predictions of the log of the age-specific death rates for a cause, assessed with the same 15% of the data. The ensemble model performing best was subsequently selected and assessed against the other 15% of the data withheld from the statistical model building. CODEm was run independently by sex for each cause of death. A separate model was run for countries with 4-star or greater VR systems to avert uncertainty inflation from more heterogeneous data. The distribution of RMSE relative to cause-specific mortality rates (CSMRs) at Level 2 of the GBD hierarchy shows that model performance was weakest for causes of death with comparatively low mortality rates (figure 2; appendix 2), while models for more common causes of death such as stroke, chronic obstructive pulmonary disease, and self-harm and interpersonal violence generally had low RMSE. Out-of-sample model performance for CODEm models and age-standardised cause-specific mortality rate by Level 1 causes Model performance was defined by the root-mean squared error of the ensemble model predictions of the log of the age-specific death rates for a cause with 15% of the data held out from the statistical model building. The figure shows the association between the root-mean squared error and the log of the CSMR, aggregated over 1980–2017. Each point represents one CODEm model specific for model-specific age ranges and sex. Circles denote models run with all locations. Triangles denote models run on only data-rich locations. Colours denote the Level 1 cause categories. Open circles and triangles denote models that were run with restricted age groups of less than 30 years. CODEm=Cause of Death Ensemble model. CSMR=cause-specific mortality rate. Alternative estimation strategies were used to model a subset of causes of death with unique epidemiology, large changes in reporting over time, or particularly limited data availability, including HIV/AIDS, malaria, chronic kidney disease, cirrhosis, liver cancer, meningitis, dementia, and atrial fibrillation. Alternative strategies included prevalence-based models, incidence and case fatality models, and sub-cause proportion models as described in appendix 1 (section 7). Mortality-incidence ratio models based on registry data were used to estimate mortality from 32 cancers (appendix 1 section 3.3). Negative-binomial models were used for eight causes of death with typically low death counts or causes that typically have no deaths in countries with a high Socio-demographic Index (SDI), including ascariasis, cystic echinococcosis, cysticercosis, diphtheria, iodine deficiency, other intestinal infectious diseases, schistosomiasis, and varicella and herpes zoster virus. Once underlying cause of death estimates and accompanying uncertainty were generated, these models were combined with the cause of death correction procedure (CoDCorrect) to establish estimates consistent with all-cause mortality levels for each age-sex-year location. Fatal discontinuities are large changes in deaths due to unexpected spikes in injuries or epidemics—defined by GBD as more than one per million or more than 25 deaths—in a specific location-year. We classified fatal discontinuities as conflict and terrorism, major transportation accidents, natural disasters, other forms of disaster such as large fires or the collapse of large buildings, or major outbreaks of infectious diseases. Data on fatal discontinuities came from VR data in the 75 countries with a 4-star or 5-star data quality rating for the interval of 1980–2017. For the remaining 120 countries with a rating of 3 stars or lower, we used alternative databases (appendix 1 section 7). Cholera and meningitis were estimated as fatal discontinuities to reduce the risk of underestimation for small-magnitude outbreaks caused by the smoothing of VR or VA data over time in CODEm. To address lags in reporting and publishing of data, we included news reports and other supplemental data sources when known gaps existed. Further detail about fatal discontinuity estimation is presented in appendix 1 (section 3.3). Aetiology-specific mortality was estimated for LRIs and diarrhoeal diseases by use of a counterfactual approach that relates the frequency of each aetiology in a population and the association with that aetiology and either LRI or diarrhoea. LRI and diarrhoea were selected as initial candidates for this counterfactual analysis approach given the large disease burden they represent and the broad interest in interventions, mostly vaccine-based, to reduce their burden.19 We attributed LRI deaths to four aetiologies: Haemophilus influenzae type B pneumonia, Streptococcus pneumoniae pneumococcal pneumonia, influenza, and respiratory syncytial virus pneumonia. Diarrhoeal deaths were attributed to 13 aetiologies: adenovirus, Aeromonas spp, Campylobacter spp, Clostridium difficile, cryptosporidiosis (Cryptosporidium spp), amoebiasis (Entamoeba histolytica), typical enteropathogenic Escherichia coli, enterotoxigenic E coli, norovirus, rotavirus, non-typhoidal Salmonella spp, shigellosis (Shigella spp), and cholera (Vibrio cholerae). The mortality attributable to each aetiology is the product of the attributable fraction and the mortality due to LRI or diarrhoea. The current counterfactual analysis is an extension of work begun in GBD 2010, based on the most common pathogens and available data. This method allows for less common aetiologies to be added in the future. Years of life lost (YLLs) are a measure of premature death calculated as the sum of each death multiplied by the standard life expectancy at each age. The standard life expectancy was taken from the lowest observed risk of death for each five-year age group in all populations greater than 5 million. In 2017, GBD 2017 included a new demographic assessment of population, fertility, migration, and all-cause mortality.14 We used these components to generate single calendar-year and single age-year estimates of the population using transparent and replicable methods.14 This independent assessment of the population was subsequently used in the calculation of YLL rates and age-standardised mortality rates. Details of these calculations are available in appendix 1 (section 4.3). Using methods adapted from demographic research from Das Gupta,20 we decomposed change in numbers of deaths by cause from 2007 to 2017, using three explanatory components: as change occurring from growth in the total population; as shifts in population structure by age; or as changes in cause-specific mortality rates. We calculated the fraction of change in deaths by cause from each component using counterfactual scenarios, changing the level of one factor from 2007 to 2017, with all other factors held constant. Since the effect depends on the order of entry of the factor, we calculated the average of all combinations of the three factors. Thus, the change in global deaths due to shifts in population age structure could be calculated by comparing the number of deaths in 2007 to the number of deaths in 2017, using the population age structure from 2017 and holding both population size and cause-specific mortality rates at 2007 levels (appendix 1 section 7). Uncertainty in our estimates was attributable to cause-specific model specifications; varied availability of data by age, sex, location, or year; and variability of sample size within data sources. We quantified and propagated uncertainty into final estimates by calculating uncertainty intervals (UIs) for cause-specific estimation components based on 1000 draws from the posterior distribution of cause-specific mortality by age, sex, location, and year.21 95% UIs were calculated with the 2·5th and 97·5th percentiles, and point estimates were calculated from the mean of the draws. Changes over time were considered statistically significant when the uncertainty interval of the percentage change over time did not cross zero. The SDI is a value between 0·0 and 1·0 calculated from the geometric mean of three rescaled components: total fertility rate under 25 years (TFRU25), lag-distributed income per capita (LDI), and average educational attainment in the population older than 15 years.22 Because the total fertility rate—used in the calculation of SDI for GBD 2016—has a U-shaped association at the highest levels of development, for GBD 2017 we recomputed the SDI using TFRU25 only, an age range for which the association with development is clearest.14 We used a generalised additive model with a Loess smoother on SDI to estimate the association between SDI and each age-sex-cause death rate using GBD estimates from all national locations across all years from 1980 to 2017 (appendix 1 section 7). Expected cause-specific death rates were scaled to the expected all-cause death rate to ensure internal consistency. We then computed the number of YLLs and deaths expected for each age-sex-location-year based on SDI alone and compared these estimates to observed rates. Additional details of the development and calculation of SDI for GBD 2017 are described in appendix 1 (section 5). The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and had final responsibility for the decision to submit for publication.
