Covariates

We estimated national time series (1980–2013) for a range of covariates with data from surveys (household-level and individual-level data), censuses, official reports, administrative data, and a systematic review. For lagged distributed income and education, we estimated national time series from 1950, to 2013. Details of how we imputed series for GDP, educational attainment, tobacco prevalence, and obesity prevalence have been published previously.^29-32 Appendix (pp 4961–4974) shows the sources and imputation methods used to generate time series for each covariate. Generally, we estimated uncertainty in covariate values when sufficient infor mation was available.

All-cause mortality

We analysed all-cause mortality for 188 countries from 1950 to 2013; we present only results for 1990–2013 to coincide with the period of the overall GBD 2013 assessment. As a result of the split of Sudan, we re-extracted data and made new separately generated estimates for Sudan and South Sudan. We improved how we adjusted for data source bias for the analysis of child mortality in two ways. First, by using the improved functional forms between summary mortality indicators for child (age <5 years) and adult (age 15–59 years) age groups and other covariates, including crude rates of death caused by HIV/AIDS. And second, by modification of the model lifetable system to use a unified standard life-table selection process and improved redistribution of excess mortality rate from HIV/AIDS. We divided the analysis into eight steps. Input data and key indicators for all countries are available online.

First, to estimate under-5 mortality (₅q₀), we analysed all survey, census, sample registration, and vital registration sources. Wherever possible, we analysed microdata from surveys and censuses with updated methods for child mortality.³³ We synthesised all measurements of under-5 mortality with spatiotemporal regression and Gaussian process regression.³³ We corrected for bias in different sources in specific countries.

Second, to estimate adult mortality (₄₅q₁₅), we systematically identified all available vital registration data, sibling history survey data, sample registration data, and household recall of deaths. We assessed vital registration data for completeness by optimised death distribution methods.^2,34 We analysed sibling history data to account for survivor bias, zero-surviving sibships, and recall bias.^2,35 We synthesised sources with a combination of spatiotemporal regression and Gaussian process regression. The mean function for the Gaussian process regression was based on the combination of a non-linear hierarchical model with income per person, mean years of education in age group 15–60 years, mortality caused by HIV/AIDS, and country random effects² as covariates, and a spatiotemporal regression in which we added to the first stage model without country random effects, the smoothed residuals between the first stage model and observed data (appendix pp 66–79). We selected the hyper-parameters for Gaussian process regression through an out-of-sample predictive validity testing process.² We ranked the estimated subnational adult mortality in China, India, Mexico, and the UK to ensure that the sum of subnational estimates for a given age-sex group equalled the national estimates accounting for different population sizes.

Third, we assessed HIV-free under-5 mortality and adult mortality. HIV/AIDS causes more excess mortality in younger people and thus changes the age pattern of mortality that otherwise can be readily described by Gompertz law of mortality or the Kannisto-Thatcher model.^36,37 Where HIV/AIDS is common, this pattern of mortality should be explicitly taken into account. We estimated the HIV-counterfactual under-5 mortality and adult mortality rates using the estimated coefficients of crude death rate from HIV from the non-linear mixed effects models for under-5 mortality and adult mortality respectively, and setting the crude death rates from HIV/AIDS in the respective age groups to zero (appendix pp 90–94).

Fourth, we constructed an HIV-free life-table. The GBD 2010 introduced a model life-table system that used the under-5 death rate and adult mortality rate along with a selected standard mortality schedule to estimate the full age pattern of mortality for country-years of interest.³³ For GBD 2013, we modified how the standard mortality schedule was selected for each country-year so that the same approach was used for all countries. Specifically, we empirically computed a set of space-time weights that relate the observed age pattern of the probability of death in a sex-country-year with other sex-country-year observations. These weights were derived by comparing every empirical life-table that is not affected by the HIV/AIDS epidemic in the GBD database (10 673 life-tables) with every other life-table for the same sex. We estimated space-time weights as a function of the time lag between the paired life-tables and location (ie, within the country, region, or super-region). We estimated these weights as the inverse of the average sum of age-specific differences in the logit of the probability of death (appendix pp 79–91). The key observation from this spatial-temporal analysis of age-specific probabilities of death is that the mortality pattern in a country in a given year was more strongly related to the mortality pattern in the same country within 15 years than to mortality patterns in other countries; however, other countries in the same region or other regions generally are similarly related when the lag-in time was more than 20 years.

Fifth, we assessed the age pattern of HIV/AIDS mortality. Excess mortality from HIV/AIDS as quantified between the estimated ₅q₀ and ₄₅q₁₅ with their HIV counterfactual counterparts leads to increased mortality in specific age groups. This excess HIV/AIDS mortality was assigned by age with the estimated relative risk of death caused by HIV/AIDS in an age group compared with the HIV/AIDS excess death rate in age group 40–44 years. We estimated these relative risks with data from vital registration systems that have International Classification of Diseases 10 coded causes of death from HIV/AIDS, which includes South Africa.³⁸ We used Seemingly Unrelated Regression model³⁹ with only a constant and generated 95% uncertainty intervals (UIs) for the age pattern of relative risks by repeatedly sampling from the mean and covariance matrix of the estimated βs and the error term. Seemingly Unrelated Regression enables the error term of a series of linear regressions to be correlated. We used separate regressions by sex and for the pattern of mortality in concentrated epidemics and generalised epidemics as defined by UNAIDS.^38-40

Sixth, we minimised the difference between demographic estimation of age-specific mortality and HIV models. Murray and colleagues³⁸ used a refined version of the EPP-Spectrum framework to model HIV/AIDS mortality. This analysis yielded very large UIs for HIV/AIDS in many countries. However, in some southern African countries, there remained a large discrepancy between data for all-cause mortality and estimates of HIV/AIDS mortality with demographic sources suggesting smaller epidemics. To minimise the difference between HIV/AIDS mortality and the demographic estimates, which are also uncertain, we computed a loss function that quantifies the extent to which the age-sex-country-year HIV/AIDS estimates exceed all-cause mortality:

For run (r) of a given country, excess mortality (e) is equal to the sum of all non-zero differences between HIV/AIDS mortality (m^HIV) and 0·8 times a randomly selected all-cause mortality draw (m^all-cause) across all year (t), age (a), and sex (s) combinations. 0·8 was the highest observed cause fraction caused by HIV in any age group in any vital registration system. We selected from the uncertainty ranges for HIV and all-cause mortality those that minimised the difference. If no draws had a positive loss function, we sampled randomly from all matched draws.

The CoDCorrect algorithm includes HIV/AIDS cause-specific mortality and can alter the age and time distribution of deaths from HIV/AIDS. To incorporate the overall change in the number of HIV/AIDS deaths over the course of the epidemic in a country implied by the application of the CoDCorrect algorithm, but not to distort the Spectrum estimated time and age pattern, we adjusted the entire HIV/AIDS epidemic up or down on the basis of the cumulative effects of CoDCorrect on HIV/AIDS for all estimated years in each country.

Seventh, we used the same method as Wang and colleagues² to generate child mortality rate and adult mortality rate for natural disasters and armed conflicts. We obtained data for conflict and war, including deaths from one-sided violence, non-state conflict, armed force battles, and other national or international conflicts, from the Uppsala Conflict Data Program⁴¹ and the International Institute for Strategic Studies.⁴² Further data for war were obtained from countries’ vital registration systems and classified as caused by war.⁴³ We included disaster data from the International Disaster Database from the Center for Research on the Epidemiology of Disasters (University of Louvain, Brussels, Belgium).⁴⁴ From this database, we included deaths caused by complex disaster, drought, earthquake, flood, and others. When these databases were not fully up-to-date or did not contain shocks known to exist, we supplemented with case-by-case sources. These with-shock mortality rates were then used as entry parameters to the GBD relational model life-table system to generate age-specific mortality rates with the effect of shocks added.

Eighth, we used age-specific and sex-specific death rates from the life-table to generate numbers of death by multiplying by population estimates from the World Population Prospects 2012 revision⁴⁵ and the Human Mortality Database for people older than age 5 years. For the under-5 age groups, we applied the method of Wang and colleagues.³³ In some cases, assumptions in the UN estimation process led to implausible population numbers for some countries and age groups—for example, low population estimates for older age groups in South Africa, especially for the most recent years. For GBD 2013, we applied a Bayesian population reconstruction model⁴⁶ to re-estimate population for South Africa for 1970–2013.

Cause of death database

Lozano and colleagues³ described the key steps in the development of the GBD cause of death database. The database has been expanded to capture 2233 additional site-years of vital registration data and 52 additional verbal autopsy site-years (table 1); a site-year is defined as data for a specific geographical location (eg, a province of China) in a given year. We included data up to April 15, 2014.

A major new addition was the incorporation of two data systems in China. First, the China National Office for Maternal and Child Health Surveillance provided detailed information for child and maternal mortality by cause from 363 surveillance sites in China for 1996–2013. Second, the Disease Surveillance Points system was the main source of mortality data for 1991–2007, with 145 disease surveillance points used from 1991 to 2003, and 161 points used from 2004 to 2007. From 2008 to 2012, all of the deaths and cause of death information from the Disease Surveillance Points system and other system points throughout China were collected and reported via the Mortality Registration and Reporting System, an online reporting system of the Chinese Center for Disease Control and Prevention, which included 4·0 million deaths in 2012.^47,48 Because of the discrepancy in proportions of deaths in hospital and out of hospital in the Mortality Registration and Reporting System, we divided each province in China into two strata based on the degree of urbanisation from the 2010 China Census. We then applied the proportion of deaths in hospital and out of hospital and degree of urbanisation from the Disease Surveillance Points system to the Mortality Registration and Reporting System to account for biases in the latter. We disaggregated data for both systems by province and urban and rural regions within each province. We obtained new datasets for Russia that provided more detailed causes (appendix pp 180–244). Turkey expanded its vital registration system to cover nearly all the population after 2009 and we incorporated these new data into the analysis.

In total, we identified 538 verbal autopsy site-years, 52 more than in GBD 2010. India, Bangladesh, and Tanzania had the most verbal autopsy site-years available. We re-extracted and re-mapped data from all verbal autopsy studies to ensure consistency with the GBD 2013 cause list. We excluded from the database verbal autopsy studies reporting cause assignment using InterVA because it has very low published validity.⁴⁷ We incorporated 1145 registry-years of new cancer data, including 128 from Cancer Incidence in Five Continents Volume X⁴⁸ and 1017 from supplemental sources. Our analysis of cancer incidence and models of the death:incidence ratio remains unchanged from GBD 2010. Figure 2 shows site-years of data by country for any cause. Of note, Somalia and Equatorial Guinea had no cause of death data for any specific cause.

Open in a separate window

Figure 2

Site-years for all causes of death data by country, 1980–2013

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.

Assessment and enhancement of quality and comparability of cause of death data

Using the general approach of the GBD 2010, we followed six steps to assess the quality of data and enhance comparability. First, we adjusted cause of death data from vital registration systems for incompleteness. The analysis of all-cause mortality yields a separate estimate of completeness for deaths of children younger than 5 years and deaths of people older than 5 years, which we used to correct the data for cause of death. When correcting for incomplete registration, we assumed that for each age-sex-country group, the cause of death composition of registered deaths and non-registered deaths were the same. 77% of datapoints were from registration or sample registration that were more than 85% complete, 17% from systems that were 70–84% complete, and 6% were from systems less than 70% complete. Of the 6% of observations less than 70% complete, most (62%) were for children younger than 5 years. In sensitivity tests in the GBD 2010, exclusion of data below a fixed threshold of completeness for child causes of death did not substantially change the results; thus, we have used all the data in our analysis for GBD 2013.³

Second, we developed 103 maps (excluding verbal autopsy studies) to translate causes found in the data to the GBD 2013 cause list. The expanded cause list of this study required us to adjust the maps used for data included in GBD 2010. Appendix pp 245–251 show GBD 2013 cause maps for International Classification of Diseases 9 and 10. The appendix (pp 252–253) includes more detail about changes made to the handling of various shorter tabulation lists used by some countries for reporting, such as the International Classification of Diseases 9 Basic Tabulation List.

Third, a crucial aspect of enhancing the comparability of data for cause of death is to deal with uninformative, so-called garbage codes. Garbage codes are codes for which deaths are assigned that cannot or should not be considered as the underlying cause of death—for example, heart failure, ill-defined cancer site, senility, ill-defined external causes of injuries, and septicaemia. Figure 3 shows the number of deaths in the database for each calendar year with the number assigned to garbage codes. Because of lags in national reporting of cause of death data, the number of deaths available after 2010 fell. Important changes for the GBD 2013 in our approach to redistributing garbage codes included the statistical estimation of the fraction of deaths following the methods outlined by Ahern and colleagues⁴⁹ for deaths assigned to ill-defined cancer site, ill-defined external causes of injury, heart failure, unspecified stroke, hypertension, and atherosclerosis by region, age, and sex. Because of variation in the coding of International Classification of Diseases 10 code X59 (exposure to unspecified factor)⁵⁰ and its subcauses in high-income countries, we redistributed these garbage codes with country-specific estimates for high-income countries derived from our statistical analysis. Additionally, we did not use malaria as a target for any garbage code redistribution in adults.⁵¹ We also implemented geographical restrictions on garbage code redistribution for Chagas disease based on endemicity so that Chagas disease was not assigned deaths in countries outside Latin America.

Open in a separate window

Figure 3

Total garbage and non-garbage coded deaths from vital registration and verbal autopsy sources, 1990–2013

Fourth, for some datasets, particularly some verbal autopsy studies, deaths were reported for broad age groups or with both sexes combined. With the addition of new data for GBD 2013, we identified 30 new age formats, totalling 112 unique age tabulations in the database. We used the algorithms described in the GBD 2010 to split these aggregated categories into estimates for specific age-sex groups.

Fifth, because few overall deaths were included in verbal autopsy studies or reported in small countries, the number of deaths by cause can fluctuate substantially from year to year. For example, in Iceland, no maternal deaths were recorded from 1991 to 2000, then one maternal death in 2001. We modified our approach to smoothing these stochastic fluctuations used in the GBD 2010 by use of a simple Bayesian algorithm. We assume a normally distributed prior and a normal data likelihood, such that:

Where X is the mean of the data and μ is the mean of the prior. We estimated the prior for vital registration series with a negative binomial regression with fixed effects for year and age estimated separately for each country. When the data are based on a large sample size the variance is small and the prior has little effect on the posterior. When the data have a large variance because of a small sample size, the prior has more effect, effectively borrowing strength on the age pattern from data within the same country but allowing for different levels in each year. For verbal autopsy studies, we modified this approach because many published reports are for a single site in a single year. The prior for each cause was based on a negative binomial with fixed effects for age groups and random effects for study-year; the regression was estimated independently for each region. For malaria, we did not group studies by region but by super-region and level of endemicity. To avoid very large negative values for log death rates or logit cause fractions, we limited the minimum non-zero posterior values to 1 per 10 000 000.

Sixth, we excluded outliers based on four criteria. (1) Studies with biologically implausible values, such as 100% of mortality from a single rare cause. (2) Studies with results that were greatly inconsistent with other studies for the same country. (3) Studies that were greatly inconsistent with studies from other countries with similar sociodemographic profiles within the same region. (4) Studies that, if included, led to abrupt changes in model-estimated time trends that could not be explained by contextual changes or policy initiatives. Outliers (0·89% of database entries) are shown in the online data visualisation of the cause of death database.

Modelling individual causes of death

As in the GBD 2010, we used six modelling strategies for causes of death depending on the strength of the available data. Where extensive data were available, we used cause of death ensemble modelling (CODEm), where fewer data were available we used simpler statistical models, and where available cause of death data might be substantially biased or not available we used natural history models (appendix pp 278–282). We generated 95% UIs from all the modelling strategies. Uncertainty in the number of deaths for an age-sex-country-year was propagated into the computation of years of life lost (YLLs) for the same category.

We used CODEm for 155 causes of death. CODEm has been extensively used for global health estimation including the GBD 2010.^3,52,53 An advantage is that a wide range of different models are tested; only models meeting predetermined criteria for statistical significance and direction of regression coefficients are retained. We excluded 30% of the data from the initial analysis so that the performance of different models could be assessed in terms of how well they predict the omitted data. Through multiple iterations of this process (cross-validation), we obtained stable objective information about the model’s performance. The best performing models in terms of root-mean squared error for level and trend were combined into a model ensemble. For some causes, we developed separate ensemble models for GBD developed and GBD developing regions;⁵⁴ the main advantage of this approach is that uncertainty in model estimation from heterogeneous data in low-income regions does not inflate the UI for high-income countries. We used this approach for all cancers and transport injuries.

For 13 causes, the number of deaths in the cause of death database was too low to generate stable estimates. For these causes, we developed negative binomial regressions with either a constant or constant multiplied by the mean assumption for the dispersion parameter, using reverse step-wise model building. We selected between the two model dispersion assumptions on the basis of best fit to the data. Compared with the GBD 2010, we modified how we generated uncertainty from these regressions by including in the uncertainty sampling draws from the γ distribution with shape equal to the expected rate (μ) divided by expected dispersion, and scale equal to the expected dispersion if the dispersion was assumed to be constant. For models in which dispersion was assumed to equal a constant multiplied by the mean, the scale parameter included μ as a multiplicative term (instead of the shape parameter).

As in the GBD 2010, for 14 causes for which death is rare, we first modelled the parent cause in the GBD hierarchy with CODEm and then allocated deaths to specific causes using proportions of the parent cause for each sub-cause. For these causes, we identified no significant predictors in negative binomial regressions. We estimated proportions by simple averaging based on available vital registration data. Depending on the availability of data, we averaged the data across age, sex, region, and year.

We used DisMod-MR⁵⁵ to estimate detailed cause fractions for several causes of death that had sufficient data to estimate proportions of a parent cause resulting from subcauses that vary across regions and countries but insufficient data to run CODEm. The source code for estimation is available online. DisMod-MR uses data for subcause fractions gathered from systematic review and from International Classification of Diseases-coded vital registration and sample registration systems. It uses two types of fixed effects (study characteristics and country covariates) with hierarchical random effects for super-region, region, and country to generate estimates for each country, age group, both sexes, and six discrete time points: 1990, 1995, 2000, 2005, 2010, and 2013. We calculated predictions for intervening years—and back to 1980—assuming an exponential rate of change. We used this approach for eight causes of maternal death, four causes of meningitis, one cause of chronic kidney disease, four causes of cirrhosis, four causes of liver cancer, and three causes of haemoglobinopathies.

For 14 causes in the GBD 2010, we used natural history models because data systems for cause of death did not capture sufficient information. The natural history model for African trypanosomiasis was updated to include the most recent case notification data from WHO (up to 2012). We made substantial changes to the HIV natural history model.²⁵ Our natural history model for congenital syphilis was estimated as in the GBD 2010, with updated data for antenatal care coverage to inform the number of births at risk and additional vital registration data sources to inform age and sex distribution of deaths. We also used simple natural history models for typhoid and paratyphoid fever, whooping cough, measles, visceral leishmaniasis, and yellow fever. Additionally, because vital registration data recording the specific type of hepatitis were very sparse, we used natural history models for all the detailed causes of hepatitis. The natural history model takes into account the extensive serological data for the prevalence of antibodies or antigens for hepatitis A, B, and C, and more limited data for case-fatality rates.

Alzheimer’s disease and other dementias were analysed with CODEm in GBD 2010. Because of the large inconsistency between the data for prevalence and mortality, we used a natural history model in the GBD 2013. Prevalences have not changed substantially over time, whereas age-standardised mortality rates in high-income countries have increased, ranging from about 25% (Denmark, Switzerland, Norway) to 46% (Germany). The prevalence of dementia varies between countries by a factor of three, whereas dementia mortality recorded in vital registration data and verbal autopsy studies varies by more than 20-fold. On the basis of these findings, we believe that the variation in dementia mortality rates between countries and over time was probably affected by changes in coding practices with increased propensity to assign dementia as an underlying cause of death. To correct for this, we assessed data from 23 high-income countries with high-quality vital registration systems to estimate the ratio of registered dementia deaths:prevalent cases. In DisMod-MR, we used the mean of these ratios as an estimate of excess mortality to estimate age-specific and sex-specific mortality from dementia consistent with the meta-regression of prevalence.

In GBD 2010, because single-cause models were developed for each cause, the final step was to combine these models into estimates that are consistent with all-cause mortality for each age-sex-country-year group. For each cause-specific model and for all-cause mortality, we had 1000 draws from the posterior distribution for each age-sex-country-year group. We combined causes by taking a random draw without replacement from the posterior distribution of each cause and all-cause mortality. Each cause was rescaled by a scalar equal to the draw of all-cause mortality divided by the sum of the draws of individual causes. The GBD 2010 induced a correlation of 1·0 between the sum of cause-specific and all-cause mortality. CoDCorrect was applied in a hierarchical fashion: first to level 1 causes and then to level 2 and level 3 causes. Level 2 causes were constrained to sum to the level 1 parent cause. Levels of this cascade were largely the same as those used in the GBD 2010 and were chosen on the basis of the amount and quality of available data for cause of death.

For GBD 2013, we made slight modifications to this approach. Because tests showed no substantial effect of the 1·0 correlation between draws of all-cause mortality and the sum of individual causes and because each cause is modelled independently such that the ordering of draws across causes were unrelated, we have removed this assumption. Furthermore, because the modelling of HIV through the GBD version of Spectrum uses relationships between incidence, CD4 progression, and death that are age-dependent and antiretroviral therapy scale-up over time has had major effects, we modified the way in which HIV deaths are handled in CoDCorrect. We ran CoDCorrect for all causes and then computed the pre-CoDCorrect cumulative deaths over time and age and compared with the cumulative deaths post-CoDCorrect. This provided an overall scalar, which we used to adjust the entire HIV epidemic. To avoid in any age-sex-country-year the sum of individual deaths exceeding all-cause mortality, we computed the difference between the cumulatively scaled HIV deaths and the CoDCorrect HIV deaths and added this difference to the estimate of all-cause mortality at the draw level.

In GBD 2010, diarrhoea deaths and lower respiratory infection deaths were reported for pathogen-specific causes in tabulations that summed to 100% of each parent cause. Since the GBD 2010, the GEMS study⁵⁶ has been published, which provided data for the relative risk of diarrhoea being related to different pathogens. This relative risk approach used a different conceptual framework than did the International Classification of Diseases approach for underlying cause. Underlying cause follows the more than 200-year history of health statistics of assigning each death uniquely to a single underlying cause. The relative risk approach follows the approach used more generally for risk factors, where cause is assigned based on comparison to a counterfactual. Counterfactual attribution to specific risks or in this case pathogens, can sum to more or less than 100%. On the basis of the GEMS study and consultations among experts in both diarrhoea and lower respiratory infection, we report results for counterfactual causes in GBD 2013. To estimate diarrhoea mortality attributable to different pathogens, we calculated the population attributable fraction for pathogens including rotavirus, Shigella, enteropathogenic Escherichia coli, enterotoxigenic E coli, adenovirus (enteric adenovirus), norovirus, Aeromonas, other Salmonella (non-typhoidal Salmonella), Cryptosporidium, Campylobacter, and Entamoeba. We used the Miettinen formula, which uses the distribution of pathogens in patients and relative risks of pathogens for diarrhoea, to provide a population attributable fraction for each pathogen:^57,58

Where PAF_i is the population attributable fraction of diarrhoea caused by pathogen i, p_i is the prevalence of pathogen i in patients with diarrhoea, and odds ratio_i is the odds ratio of diarrhoea in people with the pathogen. We used DisMod-MR to estimate the proportion of patients in each age-sex-country-year with each pathogen with data from studies of inpatients and community samples. By use of study-level covariates in the meta-regression, we obtained different estimates for inpatients and community samples. We assumed inpatients to be a proxy for severe diarrhoea and death. We reanalysed GEMS⁵⁹ to estimate the odds ratio for each pathogen in a multipathogen model by conditional logistic regression. Regression models included fixed effects for a specific pathogen with interaction terms for three age groups (0–1 years, 1–2 years, and 2–6 years) to allow different odds ratios by age and interaction terms for different GEMS field sites (Bangladesh, India, Kenya, Mali, Mozambique, Pakistan, and The Gambia) to estimate site-specific odds ratios. For other countries in the region, the odds ratio we used was the average of the odds ratios (in logarithm scale) of the countries with GEMS sites in that region. For countries in central and southern sub-Saharan Africa, we used the average of GEMS sites located in eastern and western sub-Saharan Africa. For all other countries in regions without GEMS sites, we used the average of all odds ratio. To produce odds ratio uncertainty while averaging odds ratio, we generated 1000 draws of joint normal distribution using a covariance matrix from the conditional logistic regression for each of the GEMS countries in a region. To produce the draws for non-GEMS countries, we selected draws from each of the GEMS countries until we had a full set of 1000 draws. For example, to generate 1000 draws for countries in eastern sub-Saharan Africa, we used draws from Kenya and Mozambique—the two GEMS countries within that region. We pulled 500 of the Kenya draws and 500 of the Mozambique draws to produce our full set of 1000 draws, which were used for all the other countries in this region. Because GEMS included only diarrhoea in children younger than 5 years, we applied the odds ratio of pathogens calculated for children aged 2–5 years when calculating for adults. We did not assign diarrhoea cases or deaths to a pathogen for an age-country-year if more than 95% of draws were greater than 1.

For cholera, we used data from previous studies (appendix p 310) and compared them with WHO case notification data to estimate under-reporting for cholera and then the number of cases (appendix p 310). To estimate cholera deaths, we modelled cholera case fatality in DisMod-MR with data from previous studies.

Clostridium difficile as a cause of diarrhoea in children is rarely studied; we could not estimate the epidemiological population attributable fraction as we did for other pathogens because C difficile was not included in GEMS. Because C difficile is related to hospital and health-care use, we used hospital data as the primary source for estimation. We modelled the incidence and case fatality of C difficile and assumed a 1-month risk of death⁶⁰ in DisMod-MR to estimate the number of deaths.

For GBD 2013, we split lower respiratory infection mortality into four categories: Streptococcus pneumoniae, Haemophilus influenzae type B pneumonia, respiratory syncytial virus pneumonia, and influenza. The counterfactual approach captures the complex interactions between the these causes⁶¹ and also excludes the “other lower respiratory infection” category included in GBD 2010. Moreover, we did not attribute lower respiratory infection to any cause for children younger than age 1 month. We adopted a different approach to estimate bacterial and viral causes on the basis of available data. For pneumococcal and H influenzae type B pneumonia, we estimated the causal fraction from vaccine efficacy studies.^62-64 For pneumococcal pneumonia, we included data from controlled trials and observational studies, such as before-after population analyses of the introduction of pneumococcal vaccine.^65-76 For H influenzae type B, we excluded case-control studies because of implausibly large estimates of vaccine efficacy. Furthermore, unlike for pneumococcal vaccine, little data were available from vaccine efficacy studies on the effect outside of child ages. As a result, we did not estimate the causal fraction of H influenzae type B for lower respiratory infection in people aged 5 years or older. We adjusted data for efficacy, using invasive disease as a marker as well as serotype coverage for pneumococcal vaccine.⁶⁴ We calculated pooled estimates of causal fractions by age with DisMod-MR for pneumococcal vaccine and random-effects meta-analysis for H influenzae type B, adjusted post-hoc for national-level coverage of pneumococcal vaccine and H influenzae type B vaccine. For respiratory syncytial virus and influenza, we relied on observational studies that measured causal fractions among hospital admissions for lower respiratory infection. We estimated the causal fractions among cases by country, age, and sex with DisMod-MR and the odds ratio of exposure from case-control studies. To account for the higher case-fatality of bacterial versus viral lower respiratory infections, we applied a relative case-fatality differential based on in-hospital case-fatality using hospital admissions that included cases coded to the specific pneumonia causes.

Convergence measures

To test whether all-cause and cause-specific mortality converged in the 188 countries since 1990, we computed two measures: the average relative difference and the average absolute difference between any pair of countries included in the GBD 2013 study. The average relative difference is known as the Gini coefficient and is the most commonly used measure of inequality. For international comparisons, we used the population-weighted version of the Gini coefficient in age-specific mortality rates so that small populations do not have an undue influence on the assessment of global mortality convergence (appendix pp 556–557).⁷⁷ For the Gini coefficient to fall, the percent decrease in mortality for countries with higher mortality must in general be faster than that for countries with lower mortality.

We also computed the mean absolute difference for all-cause mortality for each age group for 1990–2013 and for age-standardised rates for each cause (appendix pp 556–557). Average absolute difference can fall while average relative difference (the Gini coefficient) rises. The two measures provide different perspectives on convergence.

Multiple decrement lifetables

We used age-specific cause of death and all-cause mortality life-tables to compute the conditional probability of death for three summary intervals: childhood and adolescence (0 to exact age 15 years), reproductive-age adults (15 years to exact age 50 years), middle-aged adults (50 years to exact age 75 years), and the cause-specific contributions to each of these summary indicators. For each conditional probability of death, we used the multiple decrement life-table method⁷⁸ to compute the probability of death from each cause and the overall contribution of each cause of death to the summary probability of death indicators for the three broad age groups (appendix pp 556–557). We calculated the decomposition of changes in life expectancy by age and cause of death as detailed by Beltran-Sanchez, Preston, and Canudas-Romo.⁷⁹

Age-standardised rates and YLLs

For GBD 2010, we computed age-standardised mortality rates and YLL rates from the world population age standard issued by WHO in 2001.⁸⁰ To account for the substantial change in global demographics since 2001, we updated this standard. We used the same method as WHO and computed a standard population structure with population estimates for 2010–35 from the most recent World Population Prospects by the United Nations Population Division. Appendix pp 95–96 provides details of the GBD world population age-standard. We computed YLLs by multiplying numbers of deaths from each cause in each age group by the reference life expectancy at the average age of death for those who die in the age group following the standard GBD 2010 methods.³ The appendix (pp 121–40) shows key indicators from the new GBD standard life-table.

Ranking lists and decomposition analysis

We used the GBD 2010 approach to create ranked lists of specific diseases and injuries. We modified GBD 2010 ranking list to incorporate newly estimated causes with the same overall assignment of rank causes as GBD 2010: typhoid and paratyphoid separately, haemolytic disease in fetus and newborn and other neonatal jaundice, mesothelioma, unintentional suffocation, pulmonary aspiration and foreign body in trachea or lung, and foreign body in other part of body. Following the methods developed by Lozano and colleagues,³ we decomposed changes in the number of global deaths and global YLLs into the contributions from population growth, population aging, and age-specific death rates.

Global causes of death

We decomposed change in global and regional life expectancy by cause (level 2 of the GBD cause hierarchy; figure 7). Increased life expectancy since 1990 was mainly caused by a fall in mortality from lower respiratory infections and diarrhoeal diseases (contributing 2·2 years), cardiovascular and circulatory diseases (contributing 1·1 years), neonatal conditions (contributing 0·7 years), cancers (contributing 0·4 years), and chronic respiratory diseases (contributing 0·5 years). Decreases in mortality from unintentional injuries added another 0·3 years to life expectancy, while female life expectancy increased by about 0·2 years because of reductions in maternal mortality. These gains were offset by increased mortality from diabetes, chronic kidney diseases, and related conditions, as well as musculoskeletal disorders, although the net effect of these increases was small, reducing life expectancy, on average, by about 0·1 years. Five main causes reduced life expectancy: HIV/AIDS was a major cause of death in southern sub-Saharan Africa and to a smaller extent in western and eastern sub-Saharan Africa; diabetes, chronic kidney disease, and other endocrine disorders decreased life expectancy across many regions, most notably in Oceania and central Latin America; mental disorders made a negative contribution in multiple regions, especially high-income north America; intentional injuries reduced life expectancy in south Asia, high-income Asia Pacific, and southern sub-Saharan Africa; and cirrhosis made a negative impact in eastern Europe and central Asia (figure 7). Large gains in life expectancy in sub-Saharan Africa were mainly driven by reductions of diarrhoea and lower respiratory infections and of neonatal disorders. Gains in high-income regions were driven by reductions in cardiovascular disease, some cancers, transport injuries, and chronic respiratory conditions (figure 7).

Open in a separate window

Figure 7

Change in life expectancy at birth by GBD region and cause group from 1990 to 2013

An interactive figure with these data is available at http://vizhub.healthdata.org/le. Changes in life expectancy as a result of specific causes were decomposed from the difference between all-cause lifetables and cause-deleted lifetables (mortality set to zero for a specific cause). Because all changes in life expectancy are based on cross-sectional lifetables, the cause-specific changes add up to the total change in life-expectancy. NTDs=neglected tropical diseases.

Between 1990 and 2013, numbers of deaths from non-communicable diseases and injuries steadily increased while deaths from communicable, maternal, neonatal, and nutritional causes decreased (table 2). However, age-standardised rates decreased in these three broad categories. The shift to non-communicable diseases, at least at globally, was driven by faster rates of decline for communicable, maternal, neonatal and nutritional causes and an ageing world population.

In 2013, 11·8 million (11·3–12·3) deaths were caused by communicable, maternal, neonatal, and nutritional disorders: 2·7 million (2·4–2·8) by lower respiratory infections, 1·3 million (1·3–1·5) by HIV/AIDS, 1·3 million (1·2–1·4) by tuberculosis, and 1·3 million (1·2–1·4) by diarrhoeal diseases, 2·0 million (1·9–2·2) by neonatal conditions, 854 600 (702 924–1 032 497) by malaria, and 293 336 (261 322–328 200) by maternal causes (about 20% less than in 2000). Between 2000 and 2013, deaths from diarrhoeal diseases fell by 31·1% (from 1·8 million [1·7–2·0] to 1·3 million [1·2–1·4]), tuberculosis and meningitis each by about 20% (1·6 million [1·4–1·7] to 1·3 million [1·2–1·4] for tuberculosis, 377 300 [331 400–438 000] to 303 500 [261 400–346 300] for meningitis), while mortality from tetanus fell by about 60% (142 400 [108 800–163 100] to 58 900 [39 800–77 300]), from measles by about 80% (494 500 [279·4–763·8] to 95 600 [48 500–172 900]), from diphtheria by about 40% (5400 [2800–10 700] to 3300 [1700–6600]), and from whooping cough by about 40% (111 800 [42 300–242 400] to 60 600 [22 300–136 800]). Deaths from neonatal causes fell by a quarter since 2000 (2·8 million to 2·0 million), and by about one-fifth for maternal causes (364 900 to 293 300). Comparing 2013 to 1990, malaria deaths decreased by 4·4% and HIV/AIDS increased by 368% (table 2). HIV/AIDS mortality and malaria mortality both peaked in 2005 (1·7 million [1·6–1·9] for HIV/AIDS, 1·2 million [1·1–1·4] for malaria); HIV/AIDS mortality fell by 21% (20·4–21·5) from 2005 to 2013, and malaria mortality fell by 30% (24·8–35·4).³⁸ The risk of death from various leading causes of communicable, maternal, neonatal, and nutritional disorders as measured by the age-standardised death rate (table 2), has generally declined by an even greater amount than the risk for HIV/AIDS and malaria. Age-standardised death rates decreased by about 40% since 1990 for the category as a whole, as well as most notably, for lower respiratory infections, maternal disorders, neonatal disorders, and asthma, and by 50–60% for tuberculosis, diarrhoeal diseases, pneumoconiosis and several neglected tropical diseases (table 2). Despite a small decrease in numbers of deaths, age-standardised malaria mortality have fallen by 19% since 1990, with much of that decline occurring in the past 5 years or so (data not shown).³³

For most of the leading non-communicable diseases, the number of deaths has increased, by 42% between 1990 and 2013 (from 27·0 million [UI 26·3–27·6] in 1990, to 38·3 million [37·2–39·4] in 2013), but age-standardised mortality rates have fallen. Allowing for changes in the age structure of the world’s population between 1990 and 2013, age-standardised death rates from non-communicable diseases fell by 18·6%; by 22% for cardiovascular and circulatory diseases, 13·7–14·7% for cirrhosis of the liver and cancer, and 21·9–30·4% for other digestive diseases and chronic respiratory diseases (table 2). For many disorders, including stomach cancer, Hodgkin lymphoma, rheumatic heart disease, peptic ulcer disease, appendicitis, and schizophrenia, age-standardised death rates have fallen by more than one-third since 1990 (table 2). Age-standardised death rates for some cancers have fallen (lung by 9%, breast by 18%, and leukaemia by 20%), but have remained unchanged for others (table 2). Global age-standardised death rates have fallen by more than one-fifth for ischaemic heart disease and stroke (table 2).

Global age-standardised mortality rates increased significantly for very few disease between 1990 and 2013. The largest increase was for HIV/AIDS, which peaked in 2005 and then fell by 31·0% (UI 25·7% to 35·9%) from 2005 to 2013 (from 26·9 to 18·5 per 100 000). Among the cancers, only liver cancer caused by hepatitis C increased substantially (table 2). Although age-standardised mortality for cardiovascular and circulatory diseases decreased by 22%, significant increases occurred for atrial fibrillation and flutter and peripheral vascular disease (table 2). Mortality rates for Alzheimer’s disease and other dementias increased by only 3·2% (UI −3·01 to 11·61) and Parkinson’s disease by 28·2% (−6·42 to 37·83; table 2). Important worldwide increases occurred for diabetes (9·0%) and an even larger increase for chronic kidney disease (36·9%; table 2). The age-standardised death rate for sickle-cell disease increased while that of other haemoglobinopathies fell (table 2). The residual category of endocrine, metabolic, blood, and immune disorders increased by 11·6% (table 3).

Age-standardised death rates for injuries and most specific causes of injury also fell between 1990 and 2013, but typically much less than for diseases (table 2). Age-standardised death rates for transport injury decreased since 1990, with most deaths from road injuries. The number of deaths from unintentional injuries as a whole remained essentially unchanged since 1990, although age-standardised death rates fell by more than a quarter (table 2). Especially large falls occurred for drowning, fires, unintentional suffocation, and venomous animal contact (table 2). Age-standardised death rates also fell for self-harm and interpersonal violence (table 2). Among injuries, only falls, foreign body in other body part, adverse effects of medical treatment, and pedestrian road injuries had increased age-standardised death rates, but these were not statistically significant.

For only three level 2 causes did the age-standardised death rates increase: neurological disorders; diabetes, urogenital, blood, and endocrine diseases; and musculoskeletal disorders (figure 8). Increases in the musculoskeletal disorders were driven by the category other musculoskeletal disorders; causes with high number of death in this category were systemic lupus erythematosus, systemic sclerosis (scleroderma), pyogenic arthritis, and chronic osteomyelitis (data not shown). The average relative difference between countries (the inter-country Gini coefficient) ranged from 0·31 for non-communicable diseases to 0·90 for forces of nature, war, and legal intervention. Mean differences in the age-standardised rates between countries ranged from 0·32 for musculoskeletal disorders to 104·87 for communicable, maternal, neonatal, and nutritional diseases. Generally, inequality was much greater for communicable, maternal, neonatal, and nutritional causes than for non-communicable causes or injuries. An important exception to that general pattern was war and disaster, which were extraordinarily unequal across countries. For neoplasms, chronic respiratory diseases, and forces of nature, war, and legal intervention the age-standardised death rate had fallen and the two convergence metrics improved significantly from 1990 to 2013. For many communicable, maternal, neonatal, and nutritional causes, age-standardised death rates and mean absolute differences decreased but relative differences increased. For digestive diseases, unintentional injuries, and other communicable, maternal, neonatal, and nutritional diseases, death rates and mean absolute differences were falling and relative difference was not significantly different than zero.

Open in a separate window

Figure 8

Measures of convergence for causes of death in 188 countries

Global causes of child death

We divided child causes of death into those occurring in children younger than age 1 month and those aged 1–59 months (table 3). The number of neonatal deaths decreased from 4·5 [UI 4·4–4·6] million in 1990, to 2·6 [2·5–2·7] million in 2013, a 42% (40–45) decrease. The most important cause of neonatal death in 2013 was neonatal encephalopathy, followed by neonatal sepsis, congenital anomalies, and lower respiratory infections (table 3). Causes with more than a 50% reduction in the number of neonatal deaths include tetanus, diarrhoeal diseases, lower respiratory infections, other neonatal disorders, and other communicable diseases.

For children aged 1–59 months, the global number of deaths fell by 52·0% from 1990 (7·6 [UI 7·4–7·8] million) to 2013 (3·7 [3·4–3·9] million). Communicable, neonatal, and nutritional causes accounted for three-quarters of deaths in 2013, the remainder from non-communicable diseases and injuries (table 3). For this age group, two causes each accounted for more than half a million deaths and collectively accounted for more than a third of deaths: lower respiratory infections, and malaria (table 3). Four causes accounted for 100 000–500 000 deaths: diarrhoeal disease, meningitis, congenital anomalies, and nutritional deficiencies (table 3). Another seven causes each caused 50 000–100 000 deaths: drowning, syphilis, measles, whooping cough, intestinal infectious diseases, HIV/AIDS, and road injuries (table 3). Deaths fell by more than 50% between 1990, and 2013, for diarrhoeal diseases, lower respiratory infections, meningitis, whooping cough, tetanus, measles, preterm birth complications, drowning, other neonatal disorders, and other injuries. Of the causes detailed in table 3, only HIV/AIDS increased significantly from 1990, to 2013, although from 2005, to 2013, deaths fell from 25·95 (UI 24·54–27·50) per 100 000 to 9·83 (9·01–11·04) per 100 000. In high-income countries, cancers accounted for 5·86% (5·06–6·66) of deaths for children younger than 5 years compared with only 1·02% (0·89–1·16%) in low-income countries.

Figure 9 shows death rates of children for 19 major cause groups for the 21 GBD regions. Because infants younger than age 1 month are exposed to, at most, 1 person-month, rates in that group were high compared with children aged 1–59 months. Across regions, death rates varied widely for preterm birth complications, neonatal encephalopathy (birth asphyxia and birth trauma), other neonatal disorders and jaundice, sepsis, and lower respiratory infections (figure 9). In addition, lower respiratory infections, HIV/AIDS, congenital syphilis, malaria, diarrhoeal diseases and congenital abnormalities had an important contribution. Congenital anomalies varied by more than five-times from a high in central sub-Saharan Africa to a low in high-income Asia Pacific. Eastern, central, and western sub-Saharan Africa had substantially high death rates for children aged 1–59 months compared with other regions including south Asia (figure 9B). These higher rates were largely related to malaria, diarrhoeal diseases, measles, and nutritional disorders.

Open in a separate window

Figure 9

Child death rates by region and cause groups in 2013

(A) Of children younger than age 1 month per person-year of exposure. (B) Of children aged 1–59 months per person-year of exposure. The set of causes is mutually exclusive and collectively exhaustive. SIDS=sudden infant death syndrome.

Global YLLs

Between 1990 and 2013, large falls occurred for measles, meningitis, tetanus, syphilis, and whooping cough (figure 10). Increases of 50% or more are evident for diabetes, HIV/AIDS, hypertensive heart disease, chronic kidney disease, Alzheimer’s disease and other dementias, interstitial lung disease, and pancreatic cancer. Among the top ten causes in 1990, nine remain in the top ten in 2013, with HIV/AIDS moving in and tuberculosis moving to 11th. The largest percentage increases in YLLs were for HIV/AIDS (343·97%, 95% UI 245·48–444·17), atrial fibrillation and flutter (211·89%, 182·55–242·63), peripheral vascular disease (119·79%, 101·04–136·78), and drug use disorders (119·22%, 83·77–140·02).

Open in a separate window

Figure 10

Top 50 causes of global years of life lost in 1990 and 2013

An interactive version of this figure is available at http://vizhub.healthdata.org/gbd-compare/. COPD=chronic obstructive pulmonary disease.

Identification of key transition points in the comparative importance of different causes of premature mortality will help to better inform programme evaluation. From 1990 to 2013, worldwide crude YLLs fell by about 16% (from 2005·5 million to 1685·4 million), more so for communicable, maternal, neonatal, and nutritional diseases (39% decrease, 1098·3 million to 667·8 million) compared with non-communicable diseases (20% increase, 674·6 million to 806·5 million), and injuries (9% decrease, 232·6 million to 210·8 million). Recent progress with disease control programmes for HIV/AIDS and malaria is clear, as is the substantial and steady progress to prevent child deaths from neonatal disorders, diarrhoeal diseases, and lower respiratory infections; YLLs from these diseases fell by 40–65% since 1990. The success of vaccination programmes since 1990 is also evident (figure 11), with YLLs from measles and tetanus, in particular, at very low levels in 2013. Specific trends for major non-communicable diseases are much less evident, with incremental decreases for several leading causes of cancer, as well as from major vascular and chronic respiratory diseases, contributing to the 30% reduction in YLLs since 1990 (37 021·0 to 24 493·4 per 100 000). With the exception of drowning, only modest progress was made in reducing premature mortality from other leading causes of injury, with the effect of the 1994 genocide in Rwanda clearly visible (figure 11).

Open in a separate window

Figure 11

Global years of life lost by large cause groupings for 1990 to 2013

COPD=chronic obstructive pulmonary disease.

Causes of diarrhoea and lower respiratory infection

Deaths caused by diarrhoea fell by 51% (46–56) between 1990 and 2013 (table 2). Rotavirus was the main cause of diarrhoea in children younger than 5 years. It was also the most common cause of diarrhoea deaths in this age group in 2013, with a slight decrease in the population attributable fraction since 1990, followed by cholera, Cryptosporidium, and shigellosis (table 4). At least 55·6% of diarrhoea in 2013 was unexplained by these pathogens in all ages, an increase from 48·1% in 1990 (table 4).

The distribution of Shigella and Aeromonas in patients had a significant ecological association with sanitation (data not shown) and along with non-typhoid Salmonella, deaths from these pathogens fell by 5·4% (28 062 deaths) since 1990. Rotavirus was the most important pathogen for children younger than age 5 years in east and southeast Asia and eastern Europe, with a population attributable fraction of 35–41% (104–6390); although it had the lowest population attributable fraction in high-income north America, central sub-Saharan Africa, and Caribbean. Shigella was an important pathogen in north Africa and Middle East and Oceania (causing 19·4% [11·8–29·4] of deaths, 3790 [1976–6381] deaths and 13·9% [9·5–19·3] of deaths, 118·4 [45·4–249·6] deaths, respectively). Cryptosporidium in sub-Saharan Africa, cholera in central sub-Saharan Africa, Andean Latin America, and Oceania, and enterotoxigenic E coli were important causes of diarrhoea death. Campylobacter did not have a significant epidemiological relationship with diarrhoea in most countries and was an important cause only in some age groups in the GEMS data in India, Bangladesh, and Mozambique (data not shown). Clostridium difficile caused more than 45% of diarrhoea deaths in western Europe, high-income north America, high-income Asia Pacific, and Australasia in 2013, ranging from eight deaths in Australasia to 92 in high-income North America, and to a lesser extent in central Europe (20·1% [14·0–27·1] or 16 [14–19] deaths, and eastern Europe (19·0% [12·6–26·6] or 47 [38–57] deaths).

Cholera caused 45 000 (23 000–68 000) deaths of children younger than age 5 years, with most deaths (34 000 [76%]) in central and eastern sub-Saharan Africa and south and southeast Asia. Cholera was the third leading cause of diarrhoea deaths for all ages in 2013, behind rotavirus and shigellosis (table 4). Clostridium difficile was particularly important in adults in high-income countries, where it caused as many as 95% of diarrhoea deaths in elderly people.

Globally in 2013, pneumococcus was responsible for the largest number of lower respiratory infection deaths in children younger than age 5 years—followed by H influenzae type B, respiratory syncytial virus, and influenza—and in people of all ages (table 4). The fraction of deaths caused by lower respiratory infection in children younger than age 5 years attributable to H influenzae type B has decreased substantially since 1990, as a result of the global scale-up of H influenzae type B vaccine, with the largest decreases in high-income regions and Latin America (data not shown), where vaccine coverage is highest. The fraction of lower respiratory infection deaths in children younger than 5 years attributable to pneumococcus also fell in high-income regions such as western Europe, north America and Australasia, caused by the scale-up of pneumococcal conjugate vaccine, but continued to account for a large proportion of such deaths in eastern Europe and elsewhere. Many lower respiratory infection deaths attributable to the four pathogens occurred in older populations.

Country-specific probabilities of death during childhood and young adolescence

We computed conditional probabilities of death for three phases of life (children and young adolescents, reproductive age, and middle age) by country and cause; conditional probabilities are a useful summary because the values are readily interpretable. The probability of death in children and adolescents (age 0–14 years) varied greatly between and within regions, from a low of three per 1000 girls in Iceland to a high of 179 per 1000 boys in Guinea-Bissau (figure appendix 1). In the more demographically advanced regions (measured by mean age of death, fertility, and mortality change) the probability of death was below ten per 1000 people for both sexes in all countries except Albania, Bulgaria, Belarus, Brunei, Moldova, Macedonia, Romania, Russia, and Ukraine. Causes of death were dominated by congenital anomalies and neonatal causes. In southeast Asia, Malaysia had the lowest probability of death and Laos the highest, with substantial mortality from lower respiratory infections, diarrhoea, and neonatal disorders. Drowning was an important cause of death for children in this region. In central Asia, probabilities of death ranged from 17 per 1000 people in Armenia to 65 per 1000 people in Turkmenistan, with unusually high contributions from lower respiratory infections; drowning was also an important cause. Bahrain, Oman, and United Arab Emirates had death rates below ten per 1000 for girls, whereas in Sudan and Yemen, they were more than 40 per 1000. Mortality in these countries was dominated by congenital and neonatal causes. Pulmonary aspiration and foreign body in trachea and lung was an unusually high cause of child death in Bolivia based on the available vital registration data. In south Asia, the contribution of drowning in Bangladesh was notable but mortality was still dominated by neonatal causes, diarrhoea, and pneumonia in all the countries of the region. Throughout sub-Saharan Africa, patterns of causes of death varied but malaria, HIV/AIDS, tuberculosis, nutritional deficiencies, and haemoglobinopathies played an important part in many countries. Mauritius, Seychelles, and Cape Verde all had much lower probabilities of death during childhood than did other countries in this region. Countries with the highest probabilities of death in central and western sub-Saharan Africa had important contributions from malaria. Interpersonal violence was an important contributor to the probability of death in children and adolescents in many countries of Latin America and several in southern Africa.

Country-specific probabilities of death during reproductive age

The probability of death in reproductive-age adults (exact age 15 years to exact age 50 [₃₅q₁₅]) ranged from 1·2% for women in Andorra to 52% for men in Lesotho (figure appendix 2). In high-income regions, the probability of death was generally twice as high for men as for women. Across all countries, transport injuries made an important contribution, especially in men.

Among men in low-income countries, suicide and transport accidents made important contributions. In some countries, such as Norway and USA, drug and alcohol use disorders account for more than 8% of the total probability of dying in this age interval for either sex. In central Europe, interpersonal violence was high in Albania, drug and alcohol disorders were notable in Poland, Croatia, Slovakia, and Montenegro, and cirrhosis was a common cause in Bulgaria, Croatia, Hungary, Poland, Romania, Slovakia, and Slovenia. In east Asia, liver cancer was an important cause of death. In southeast Asia, maternal mortality in Laos, Cambodia, Myanmar, and Timor-Leste were important contributors for women. For men in the same region, interpersonal violence in Philippines, Sri Lanka, and Thailand were notable; suicide in Sri Lanka also accounted for a large probability of death (2·2%). Other studies have suggested that interpersonal violence in Philippines is concentrated in Mindanao.⁸²

Risk of death from transport injuries was greater than 2% for men in Afghanistan, Côte d’Ivoire, Cameroon, Ecuador, Gabon, Guinea-Bissau, Equatorial Guinea, Kazakhstan, Mauritania, Nigeria, Oman, Sierra Leone, El Salvador, Thailand, and Uganda. Probability of death from cirrhosis in Myanmar was 1·9% for men. Tuberculosis stood out in Cambodia, Indonesia, Laos, Myanmar, and Philippines for men. The gap between the probability of death for men and women was particularly large in eastern Europe and central Asia. Mongolia had the highest probability of death in reproductive age for men in these two regions, as a result of unusually high probabilities of interpersonal violence, self-harm, alcohol and drug use, cirrhosis, liver cancer, and tuberculosis. More generally for men, in eastern Europe and central Asia, there were major contributions from ischaemic heart disease, self-harm, alcohol and drug use, cirrhosis, and tuberculosis; HIV played an important part in Ukraine and Russia. In north Africa and Middle East, transport injuries and ischaemic heart disease were predominant. For women in the region, breast cancer in all countries and maternal mortality in Sudan and Yemen were also major factors.

In Latin America and Caribbean, there was a large contribution of interpersonal violence in men, with the probability of death exceeding 2% in Brazil, Colombia, El Salvador, Guatemala, Honduras, and Venezuela. Despite generally high violence in the region, Cuba, Costa Rica, Bolivia, and Peru had low probabilities of death from violence. In men, the probability of death because of HIV/AIDS exceeded 1% in Belize, Haiti, Saint Vincent and the Grenadines, The Bahamas, Grenada, Guyana, Suriname, and Trinidad and Tobago. Cirrhosis contributed more than 1% to the probability of death of men in Guatemala, Mexico, and Guyana. Cervical cancer was a larger contributor to the probability of death than was breast cancer in eight countries of Latin America and Caribbean (Bolivia, Ecuador, Guatemala, Nicaragua, Peru, Paraguay, El Salvador, and Venezuela). Probabilities of death for men and women in Afghanistan were more than twice that of other countries in the region; for women, maternal mortality was the largest cause. In Oceania, Samoa and Tonga had much lower probabilities of death than did other countries in the region. Throughout sub-Saharan Africa, there were major contributions for women from maternal mortality, HIV/AIDS, and tuberculosis. For men in the region, HIV/AIDS, tuberculosis, and transport injuries dominated in most countries. Liver cancer was also a major factor, particularly for men in western sub-Saharan Africa.

Country-specific probabilities of death during middle age

The probability of death in middle age (exact age 50 years to exact age 75 years [₂₅q₅₀]) ranged from 10·3% for women in Andorra to 76·3% for men in Lesotho (figure appendix 3). In all countries, ischaemic heart disease and stroke were important contributors to the risk of death in middle age, and were greater for men than for women. Probabilities of death from ischaemic heart disease ranged from 0·8% in Japan for women to more than 24% in Belarus for men. In high-income regions, lung cancer was as a major contributor to the risk of death for men. Breast cancer for women and prostate cancer for men also made substantial contributions. Probability of death from liver cancer was greater than 2% in China, Mongolia, Myanmar, North Korea, South Korea, Taiwan (province of China), Thailand, and Vietnam.

In central Europe, chronic respiratory diseases and cirrhosis made clear contributions in Bulgaria, Croatia, Hungary, Poland, Romania, and Slovenia. In some high-income countries, including Singapore, Argentina, and Uruguay, lower respiratory infections were important causes of death for this age group, more so for men than for women. In southeast Asia, liver cancer, diabetes, and tuberculosis made larger contributions than in many other regions, particularly in Myanmar, Philippines, Laos, Indonesia, and Cambodia. Elsewhere in the region, stroke, ischaemic heart disease, other cardiovascular and circulatory diseases, and chronic respiratory diseases were predominant. In all countries of eastern Europe and central Asia, stroke and ischaemic heart disease were particularly prominent for both sexes. Liver cancer in Mongolia had the highest probability of causing death in the world for this age group.

Egypt had extraordinarily high cirrhosis mortality, particularly from hepatitis C, in middle aged men and women. Deaths caused by diabetes were particularly high in Morocco, Bahrain, Oman, and Qatar. In central Latin America and Caribbean, diabetes made large contributions to causes of death in men and women; the highest probability of death in these regions from diabetes for males was 9·2% in Trinidad and Tobago and 8·4% for women in Guyana. Chronic kidney disease was particularly high in El Salvador, Mexico, and Nicaragua; more so for men than for women.^75,76 In the Caribbean, diabetes, stroke, and ischaemic heart disease accounted for 33·4% of the probability of death in this age group in Haitian men, and 54·8% in Guyanese women.

In all the countries of Oceania, diabetes accounted for an extremely large fraction of mortality in middle-aged women. For nearly all countries in sub-Saharan Africa, stroke and other cardiovascular diseases (including cardiomyopathies) were important. HIV/AIDS and tuberculosis, diarrhoea, and lower respiratory infections were also estimated to be important causes in almost every country in the region. The probability of death from liver cancer was high in most countries of western sub-Saharan Africa.

Convergence or divergence?

Ambitious goals have been set for maternal and child mortality,^84-86 such as the end of preventable maternal and child death in a generation. The Lancet Commission Global health 2035: a world converging within a generation has argued that a grand convergence in health is possible between high-income, middle-income, and low-income countries.²⁴ Trends in the past 23 years provide an important starting point for framing how great a challenge achieving these aspirations will be and the political will and financial resources required. Part of the answer depends on how the goals are framed—for example, what does convergence mean? In the development literature on economic convergence,^87-89 convergence has been framed in terms of poverty rates or in terms of income inequality measured by the Gini coefficient or other measures of inequality. Work on convergence in life expectancy has tended to focus on measures of absolute difference^90-92 rather than relative difference.⁹³ We found unequivocal divergence in mortality rates for women aged 25–39 years and older than 80 years and for men aged 20–44 years and 65 years and older, similar to previous estimates of divergence of life expectancy at birth since the 1980s.⁹⁴ In these age groups, both the Gini coefficient and the mean absolute difference in death rates are rising. In all other age groups, except girls aged 10–14 years, relative inequality is increasing but the absolute gap is narrowing. Framing a grand convergence as simply achieving a reduction in the differences in mortality rates across countries might not be sufficiently ambitious to meet the goals of many national policy makers. If mortality decreases in all countries by the same percent per year, absolute difference will decrease and relative differences will stay constant. For age groups in which global relative and absolute differences in death rates are diverging, extraordinary efforts will be needed to achieve laudable goals such as a grand convergence. If convergence includes reducing the ratio of the highest to lowest death rates, even for under-5 mortality, major new efforts will be needed to have faster percent decreases in countries with higher mortality.

Arguments that convergence is technically and financially feasible are grounded on the rapid improvements of some countries.²⁴ For example, from 1990 to 2013, 13 countries (all low-income), achieved increases in life expectancy greater than 10 years (appendix pp 141–151). The real challenge is whether the strategies to decrease mortality used by these countries are generalisable or transferable to those countries who are making the least progress. The Lancet Commission on global health 2035 drew attention to the four Cs (Cuba, Costa Rica, Chile, and China). Life expectancy has improved faster than the global aggregate trend in China and Chile in the past 23 years.

The good news is that some countries that were low-income in 1990 have achieved remarkable progress in the past 23 years—for example, in Nepal, life expectancy has increased by 12·16 years since 1990, reaching 70·64 years in 2013 for both sexes combined (appendix pp 141–151). Other examples of improvements greater than 12 years for both sexes combined include Rwanda, Ethiopia, Niger, Maldives, Timor-Leste, and Iran. Because the Rwandan genocide occurred after 1990, the progress from the peak of mortality during the genocide until 2013 is even larger, 49·63 years. Studies have already assessed progress in Bangladesh, Ethiopia, and Niger, particularly in reducing child mortality.⁹⁵ Further study of these countries might provide insights about how to achieve low mortality, including the role of development assistance for health, rapid economic growth, and addressing chronic challenges such as famines. Simple assessments built up from individual technology analyses, such as the Disease Control Priorities-2,⁹⁶ assume a high-level of health system efficiency and contextual factors that enable technology to be delivered such as levels of maternal education. Plans to achieve a grand convergence in the face of diverging mortality will need to take into account low levels of health system efficiency and low levels of health system resourcing in some countries and the greater efforts needed to achieve high intervention coverage in low-income countries with inadequate primary health-care systems and low levels of educational attainment. The challenge of improving health system management, particularly locally, is a crucial component of the future plans.

The analysis of average relative difference between countries and average absolute difference between countries by cause (data not shown) shows the general pattern that many communicable, maternal, and neonatal causes, along with war and natural disasters, are highly unequal across countries; almost all have average relative differences of more than 50%. Among the non-communicable disease categories, mental and substance use disorders is the only cause with a mean relative difference greater than 40%. Following the more stringent criteria for convergence—in which global rates and the Gini coefficient are both falling—only neoplasms and chronic respiratory diseases are converging. As more countries go through the epidemiological transition, it seems likely that cross-country inequalities or relative differences for communicable causes will rise and inequalities for non-communicable causes will narrow. Narrowing inequalities across countries will not necessarily narrow inequalities for non-communicable disease within countries. Because mortality exponentially rises with age, at least after age 50 years, relative differences at older ages, when mortality becomes concentrated, tend to be small. Causes such as diabetes, chronic kidney disease, and alcohol and drug use disorders—for which global death rates are rising and inequality is increasing—are exceptions to this general pattern.

Non-communicable diseases

Age-standardised death rates for cardiovascular and circulatory diseases have fallen in high-income and many middle-income countries since 1990. Rapid falls have occurred in some countries. For example, five countries (Israel, Denmark, Norway, South Korea, and UK), had at least a 65% decrease in age-standardised death rates for ischaemic heart disease. Many other countries have had decreases of 40–65%. Age-adjusted death rates caused by haemorrhagic stroke fell by three-quarters in South Korea.

The ageing and growth of populations has led to an increase in the total number of cardiovascular deaths, accounting for almost a third of all deaths globally in 2013. Ischaemic heart disease, ischaemic stroke, and haemorrhagic stroke continue to cause most cardiovascular and circulatory deaths in almost all countries. Some Balkan countries are an exception; cardiomyopathy was a leading cause of death, possibly as a result of alcohol exposure or local patterns of garbage codes.⁹⁷ Additional studies are needed to establish whether this finding is driven by medical certification practices or is related to alcohol or some other factor.⁹⁷ Age-standardised death rates for atrial fibrillation and flutter and peripheral vascular disease have increased, possibly because of increased awareness of these conditions or better survival from cardiovascular diseases that share the same risk factors. Much uncertainty remains for trends in mortality caused by rheumatic heart disease, partly because endemic populations are concentrated within poorer subnational regions where data collection is limited and rheumatic heart disease might not always be coded as the underlying cause of death.⁹⁸ Efforts to benchmark changes in cardiovascular and circulatory diseases will benefit from increasing access to verbal autopsy in India and sub-Saharan Africa, household surveys focused on chronic diseases, and improvements in electronic health records.

Generally, cancer deaths are increasing but age-standardised cancer death rates are falling. Some cancer-related risk factors, such as tobacco consumption, have decreased, but others, such as obesity, have increased. The substantial general fall in cancers require further explanation. Death rates for five cancers increased (non-Hodgkin lymphoma, mesothelioma, kidney cancer, pancreatic cancer, and multiple myeloma); some explanations, such as the potential link between the rise of diabetes and pancreatic cancer might account for some of these reversals. Because of different rates of decrease for other sites, the mix of cancers is steadily changing, particularly in low-income regions, such as the relative importance of breast cancer compared with cervical cancer. These local changes have important implications for the development of cancer care programmes and training. Because of the strong relation between cancer mortality and age, ageing of the world’s population is the most important driver of the rising number of cancer deaths in most countries. Most countries can expect to have to deal with more patients who need diagnosis, treatment, and palliation in coming years.

Alzheimer’s disease and other dementias

We used a substantially different approach to estimate Alzheimer’s disease and other dementia mortality in the GBD 2013 by focusing on studies of prevalence and using data from countries with the highest death to prevalence ratios in 2013 to estimate mortality in other regions and back in time. This change greatly lowers the increase compared with GBD 2010 in the age-standardised death rate for dementia although the numbers of dementia deaths nevertheless increased. Lower increases in the age-standardised rate were because the meta-regression of prevalence studies did not show a rapidly rising trend; one study, reported decreases in age-specific rates, although our overall assessment suggests a slight increase in age-specific rates.⁹⁹ Even in high-income countries with complete medical certification of causes of death, we argue that dementia was systematically underestimated as a cause of death in earlier periods. Other studies, such as the National Mortality Followback Study in USA, support this idea.¹⁰⁰ Dementia deaths might have been misclassified into categories such as senility. Our garbage code redistribution algorithms for this broad category might have under-allocated dementia deaths in earlier periods. For future research, we may want to more carefully trace to which garbage codes dementia deaths might have been assigned using hospital linkage or other approaches.

The other effect of using this approach is that we estimated considerably more dementia deaths in middle-income countries than in the GBD 2010. Prevalence studies suggest dementia occurs in these countries although it is rarely recorded on a death certificate as a cause of death. Our overall conclusion is that dementia is more common worldwide and that numbers are increasing because of population ageing with only a small component of the increase caused by rising age-specific rates. The analysis of dementia will benefit from further population-based prevalence surveys, especially with repeated measurement over time using standardised definitions and methods. As further studies of this type become available and incorporated into the GBD, our estimates of dementia burden might be substantially revised. Trends in the category Alzheimer’s disease and other dementias might mask upward trends in Alzheimer’s and downward trends in vascular dementia; however, these disorders are difficult to tell apart in population-based prevalence studies and cause of death data. Nevertheless, our finding that the number of dementia deaths is increasing implies that governments should remain concerned about the rising demands for care that will come with population ageing even if future rates do not increase substantially.

Diarrhoea and lower respiratory infections

We report that the distribution of the causes of diarrhoea is different around the world. The distribution of pathogens has also changed significantly since 1990—for example, almost 50% (20 343 [9054–41 216] deaths) of all cholera deaths in children occur in sub-Saharan Africa. Because we used GEMS data to estimate relative risks, it is perhaps not surprising that our results are comparable to their findings.⁵⁷ The population attributable fraction for pathogens such as Campylobacter, Shigella, and Salmonella were not significant in some countries and some ages. Because of the nature of case notification data, we had to estimate all types of cholera and were unable to breakdown cholera into O1, non-O1, and O-139.

In high-income countries, C difficile is an important threat that has increased during the past two decades. 65% (744 413 deaths) of unexplained diarrhoea in people older than 5 years is an important knowledge gap. Although the new counterfactual approach is successful for estimating attributable death empirically and adjusts for the overall pathogen load in the country, it still suffers from limitations such as the potential low sensitivity of diagnostic tests. Some pathogens are more prevalent in controls than in cases, which might present a distorted causal picture because of continuous shedding of pathogen long after the acute phase.^101-105 These findings could also suggest a protective effect of infection from one or more pathogens against other pathogens or could be simply caused by a differential decrease in the sensitivity of diagnostic tests (for other pathogens) where diarrhoea presents assuming a single pathogen caused the diarrhoea. More sensitive diagnostic tests help to improve sensitivity but at the price of decreased specificity because of contamination and post-diarrhoea pathogen excretion. Follow-up studies with multiple measurements of pathogens in children during healthy and diarrhoea periods could help to elucidate the true causal associations. Better case definition and more strict criteria for pathogens such as excluding recent cases of diarrhoea could decrease exposure misclassifications.

Our estimates of the fraction of under-5 lower respiratory infection deaths attributable to the four causes of pneumonia (pneumococcus, H influenzae B, respiratory syncytial virus, and influenza) are much the same as previous estimates, with pneumococcus and H influenzae B the predominant causes.^{63,64,106-108} The large fraction of lower respiratory infection attributable to pneumococcus and H influenzae B, particularly in low-income regions where the absolute burden is highest, shows the potential benefit of continuing to scale up pneumococcal conjugate and H influenzae B vaccination. We calculated the contribution of each cause with a counterfactual approach. This approach means that they do not add up to 100% but also that there might be overlap; for example, death from lower respiratory infection might involve viral and bacterial co-infection. These results should also be interpreted with caution because of the data used to generate these estimates. Data for cause are sparse and prone to several biases, which is shown in the large UIs.

Estimates of the mortality burden of pneumococcal pneumonia in children rely on data from vaccine probe studies, which showed that disease in infants fell after pneumococcal conjugate vaccination; there are no sensitive diagnostic tests to detect non-bacteraemic pneumococcal pneumonia in children. To calculate burden in the absence of a diagnostic assay, the pneumococcal conjugate vaccination probe studies assumed a vaccine efficacy against non-bacteraemic pneumonia caused by vaccine types equal to that of protection from vaccine type bacteraemia (75%). Data from a large randomised trial of pneumococcal conjugate vaccination in adults confirmed efficacy against bacteraemic pneumonia of 75%, but efficacy against non-bacteraemic pneumonia was only 45%.¹⁰⁹ If similar efficacy estimates are applied to infants, then the contribution of the pneumococcus to pneumonia mortality in infants could be as high as 63% (166 324 deaths). Because data were sparse, we did not estimate the fraction of deaths caused by H influenzae B among people aged 5 years and older. The before-and-after vaccine efficacy studies used to estimate the burden of pneumococcus were limited to high-income settings. These types of studies might also be biased because of underlying temporal trends in hospital admissions for lower respiratory infection. Furthermore, the only variation included for pneumococcus and H influenzae B is a result of differences in vaccination coverage.

The observational studies used for respiratory syncytial virus and influenza were based on case series data from predominantly tertiary-level hospitals, which might not be representative of the underlying population and are prone to varying case-definitions and diagnostic methods. Finally, hospital discharge data for the relative differences in case-fatality for respiratory syncytial virus and influenza compared with pneumococcus and H influenzae B were limited to high-income and middle-income countries. Several of these shortcomings are being addressed by the Pneumonia Etiological Research for Child Health project.¹¹⁰

Injuries

Most global road traffic deaths occur in low-income and middle-income countries and are rapidly increasing because of the growth in motorisation. Mortality rates caused by traffic-related injuries are increasing in low-income and middle-income countries. Pedestrians are most often affected, followed by car occupants and motorcyclists. Conversely, traffic deaths are decreasing in high-income countries. We noted a similar divergence between low-income and high-income countries for occupational injuries: they generally fell in high-income countries (with the exception of deaths resulting from asbestos-related mesotheliomas), whereas occupational injury deaths have increased in low-income countries (data not shown).

Suicide continues to be a major public health problem in many regions. Half of all suicide deaths occur in China and India alone. However, the trends are in opposite directions, decreasing rapidly in China but rising in India between 1990 and 2013. Both countries have undergone economic growth and urbanisation, a key factor in limiting access to lethal pesticides, a common method of suicide by poisoning in both countries.¹¹¹ Therefore, as yet unexplained reasons must exist for the divergence between the two countries.

We recorded several sharp increases in mortality caused by war and disaster. Particularly, the 2010 Haiti earthquake, conflict in Syria over the past several years, the 2011 Tōhoku earthquake and tsunami in Japan, and conflict in Libya in 2011 have caused considerable loss of life. The war in Syria led to an estimated 29 947 deaths (19 392–54 903) in 2013, and about 10 504 deaths and 21 422 deaths in each of the preceding 2 years. Uncertainty around these estimates is large because several different estimates exist. These estimates are of the direct deaths attributable to armed conflicts and natural disasters and do not account for the full effects of mechanisms such as the breakdown of health systems or critical infrastructure. For example, the conflict in Syria has had a substantial effect on routine immunisation for polio, with coverage now as low as 50% in some areas.¹¹² The estimation of direct deaths caused by war and natural disasters is one of the most challenging components of the GBD measurement. We depend on the work of various groups to collate combatant reports, newspaper reports, humanitarian agency assessments, and other direct accounts to approximate the number of deaths. Vital registration systems often do not function in war or conflict but might be more useful in countries with natural disasters as a way of measuring the number of deaths. More work is needed to better measure shock mortality.

India

India accounts for 19% of the world’s deaths in 2013. Estimations of cause of death for India are important both for health policy in India and for global understanding of causes of death. India has had remarkable progress in reducing both child and adult mortality over the past 23 years. Average yearly rates of decline were 1·3% per year for adults and 3·7% for children.

Unfortunately, less cause of death data were available for 2013 than for 1990 or 2000. The Medical Certification of Causes of Death system provided ongoing information about patterns of urban mortality with better completeness in some states than in others. In rural areas, the Survey of Causes of Death (Rural) routinely reported causes of death from verbal autopsy from 1980 to 1998. This survey was replaced with a verbal autopsy sample collected by the Registrar-General of India based on the ongoing Sample Registration Scheme. Data for 2002–04 have been reported but not in full detail—results were released in a series of articles spanning 2008–14 but even these have not provided the standard tabulation of deaths by International Classification of Diseases cause, age, and sex used by most countries. Verbal autopsies were collected after 2004 but no data have been analysed or released. Attempts to add verbal autopsy to other major data collection efforts of the Government of India, such as the Annual Health Survey and the latest round of the District Level Household Survey, have so far been unsuccessful. Small community studies continue to be published but there is a major gap in knowledge of rural cause of death.

In view of the rapid change in India, including decreases in child mortality and adult mortality, simple predictions based on the 2002–04 data are inadequate. Our modelling strategy takes into account trends for key covariates that explain some changes in age-specific rates for many causes; nevertheless, more recent national data would be helpful to develop more precise estimates of causes of death for India. Epidemics such as Chikungunya, dengue, and H1N1 influenza also point to the need for better ongoing surveillance of causes of death in India that does not suffer from long time lags.^113-116

Comparing different global health estimates

Comparison of the GBD 2013 results with GBD 2010 for 1990 or 2010 shows some important differences. The overall correlation coefficient of age-sex-country-cause rates was 0·998 in both 1990 and 2010 but some causes have changed substantially at the global level. The ten causes in terms of the largest change in the number of global deaths were Alzheimer’s disease and other dementias, ischaemic heart disease, interstitial lung disease and pulmonary sarcoidosis, cerebrovascular disease, neonatal encephalopathy caused by birth asphyxia and trauma, lower respiratory infections, other cardiovascular and circulatory diseases, cirrhosis, malaria, and chronic kidney disease. These changes might be because of new data, modifications of garbage coding algorithms, and revised modelling strategies (appendix). Generally, the data used has substantially increased: from 8967 site-years to 14 244 site-years.

Some specific changes are worth noting. First, data for China has greatly increased. Given China’s population, the incorporation of large amounts of new data for cause of death led to large changes in China and these affected even global estimates. The five largest changes for China in 2013 compared with the GBD 2010 were ischaemic heart disease, Alzheimer’s disease and other dementias, cerebrovascular disease, interstitial lung disease and pulmonary sarcoidosis, and chronic obstructive pulmonary disease. Second, more detailed cause of death data covering 189 causes instead of 98 causes were available for Russia for the GBD 2013. This affected several smaller causes, such as those related to alcohol. Third, we included new vital registration data for Turkey for 2010–12. Fourth, we modelled India in two components, urban and rural, which enabled us to make much more use of some data sources such as the Survey of Causes of Death (Rural) for rural India. Because India is large, these changes have a global effect. Fifth, for cancers, we incorporated 1145 registry-years of new data, including 128 from the Cancer Incidence in Five Continents Volume X.⁴⁸ Sixth, the change to use of a Bayesian noise reduction algorithm for smoothing has reduced the number of outliers, particularly in small verbal autopsy studies, some of which were included in the GBD 2010. Seventh, changes to garbage code redistribution algorithms, particularly the use of statistically derived algorithms that vary by region and country, has had effects on injuries, cancers, and cardiovascular diseases. Other changes included treating unspecified anaemia as a garbage code whereas in the GBD 2010 it was mapped to iron-deficiency anaemia, moving abdominal hernia from other digestive diseases to hernia, as well as moving deaths related to specific procedures to the category of adverse effects of medical treatment. In the GBD 2010, we included abdominal hernia, including umbilical hernia, ventral hernia, and diaphragmatic hernia in the category “other digestive diseases”. In the GBD 2013, we combined these with inguinal hernia and femoral hernia into one cause named “hernia”. Additionally, we moved some ill-defined causes from the other digestive diseases category to more specific causes, thereby reducing the number of deaths in other digestive and changing the distribution of all digestive deaths among its more disaggregated causes. Seventh, the assessment of all-cause mortality in the GBD 2013 benefited from both new data and improved approaches for assessment of the age pattern of mortality in the model life-table system. Finally, the more detailed analysis of HIV/AIDS led to major changes both for HIV/AIDS (particularly in countries with concentrated epidemics) and for other causes, particularly in the people of reproductive age and in countries with moderate-to-large epidemics.

The International Agency for Research on Cancer produces cancer estimates by country, age, sex, and cancer site for 2008 and 2012 (GLOBOCAN). Our definitions and the GLOBOCAN definitions are compatible for 25 sites. For these cancer sites, the total estimated prevalence from GLOBOCAN was 6 848 204 cases in 2008 and 7 483 018 cases in 2012. By comparison, the GBD estimates were 6 930 377 for 2008, and 7 437 018 for 2012. Worldwide, the largest variation in estimates occurs for thyroid cancer, testicular cancer, and other pharynx cancers, with differences of 20–30%. The rough similarity of results at worldwide masks substantial national variation. Comparing age-standardised death rates for 2012, the correlation ranges from 0·94 for tracheal, bronchus, and lung cancer, to 0·20 for thyroid cancer. Five cancers have correlations below 0·5 (ovarian, non-Hodgkin lymphoma, testicular, Hodgkin lymphoma, and thyroid). A further six cancers have correlations of 0·5–0·7 (uterine, nasopharynx, lip and oral cavity, breast, leukaemia, and multiple myeloma).

Because both GLOBOCAN and our estimates used population-based cancer registry data and vital registration data as inputs, the wide variation in results requires explanation. As with all comparisons of global health estimates, the differences stem from data, data processing, and model development. We included a wider range of registries than did GLOBOCAN, particularly in China, and we used of a broader database of vital registration data. Our redistribution of cancer of unknown primary was based on a statistical model. The most important differences, however, probably stem from the modelling strategy. For all cancer sites in all countries, we used CODEm. GLOBOCAN used nine different methods to estimate cancer mortality depending on the country.^83,117 The choice of method can lead to surprising differences in estimated rates for neighbouring countries without data. For example, the age-standardised death rate for male thyroid cancer in Timor Leste is 250% higher than that for Indonesia; age-standardised death rates for testicular cancer differ by 1300% between Mali and Mauritania. The GLOBOCAN estimates have a substantial subjective component in the choice of which modelling strategy to use and do not provide any estimate of uncertainty. Empirical assessment of the validity of the GLOBOCAN methods—for example, through cross-validation—would help to understand the strength of the approach.

Understanding causes of death begins with assessment of all-cause mortality. There are some notable differences between our assessment of global age-specific deaths and those produced by the United Nations Population Division in their World Population Prospects 2012 revision (WPP2012). For the three periods (1995–2000, 2000–05, and 2005–10) as defined in WPP2012, the total numbers of deaths were 2·4–3·6% higher (6·1 million–9·1 million deaths) than estimated by us. These differentials translate into a difference of 7·8 million deaths for the 5-year period between 2005 and 2010. The difference is greatest for younger age groups. For 2005–10, estimated under-5 deaths from WPP2012 are 10·7% higher (3·9 million more deaths) than for the GBD 2013.

The WPP2012 global under-5 death estimates were also higher than those of UNICEF; part of this difference might be a result of the agencies releasing their estimates at different times. The biggest relative difference was for the adolescent age group (age 5–14 years). For 2005–10, the estimated deaths in adolescents from WPP2012 were 45·1% higher than in the GBD 2013, even though the absolute difference was about 2·2 million for a 5-year period, less than 1% (2·17 million of 264·7 million) of the total deaths for the same period. The differences are even greater at the GBD regional level. For 2005–10, the relative difference between WPP2012 and GBD 2013 ranged from 26·7% (122 800) lower in WPP2012 in Oceania, to 36·0% (1·9 million) higher in WPP2012 in central sub-Saharan Africa. WPP2012 tends to have high estimates of adolescent mortality compared with the GBD 2013 for all regions in sub-Saharan Africa, Andean Latin America, north Africa and Middle East, and southeast Asia. Overall, we find more differences in estimates for sub-Saharan Africa across all age group in both relative and absolute terms.

Such discrepancy originates from different assessments of child mortality rates and the difference in model life-table systems, both of which used child mortality rate to generate age-specific mortality rates. Estimating mortality for the adolescent age group is important.^118-120 As part of the background research for the GBD 2013, we assessed the Demographic Health Surveys complete birth history data for age groups 5–9 years and 10–14 years and compared this data in countries with almost complete vital registration or sample registration systems, such as India. We also systematically assessed estimates of adolescent mortality from sites of the health and demographic surveillance systems, a network known as INDEPTH. When we assessed the ratio of ₅q₅ (probability of death from age 5 years to age 10 years) and ₅q₁₀ (probability of death from age 10 years to age 15 years) to under-5 mortality, conflicting pictures arise: our GBD 2013 estimates are sometimes higher than one source and lower than the other. Further analysis is warranted to validate our approaches for estimating adolescent mortality in low-income and middle-income countries without working vital registration systems. In addition, efforts are needed to improve both data collection and method development to better estimate mortality for adolescents.

As in the GBD 2010, we noted differences for causes of child death compared with those produced by the Child Health Epidemiology Reference Group (CHERG; table 5). Given the complexity of both approaches, it is difficult to isolate the reason for the differences. One reason might be the different studies used: we included 5039 site-years of vital registration and 358 of verbal autopsy data compared with 578 and 192 for CHERG. Our modelling strategy was founded on modelling each individual cause separately, using the most appropriate method for each cause, and then combining the different cause estimates into an overall assessment consistent with all-cause mortality using CoDCorrect. CHERG used separate modelling strategies for HIV/AIDS, measles, pertussis, and malaria outside of Africa, and four different models for the remainder of the child causes. Separate logistic models, each with subtly incomparable cause lists, were used for neonates and children older than 1 month, for low mortality countries excluding China, for high mortality countries excluding India, for India alone, and for China alone. This partition of the world into separate models was not justified statistically—for example, they have not shown statistically different relationships with covariates for their four sets of models. Additionally, post-estimation adjustments were applied to pneumonia, meningitis, and malaria to account for intervention effectiveness; pneumonia, sepsis, meningitis, and tetanus to account for the reliance on a combined severe infection cause in the primary model; and diarrhoea, neonatal sepsis, and sudden infant death syndrome in China to account for studies that report few causes.

We used a more empirical approach. We quantified both the root-mean squared error and validity of the UIs through cross-validation; CHERG has not to date reported any cross-validation results. Given the possibility that different relationships might exist between covariates such as access to clean water or sanitation and diarrhoeal mortality in different parts of the world, we undertook a sensitivity analysis in which we excluded vital registration data from high-income regions from the models for lower respiratory infections and diarrhoea. We detected no substantial differences for estimated global cause of death patterns in these cases. Furthermore, in CHERG, neonatal causes were assumed to not cause deaths after 1 month although high quality vital registration systems routinely report deaths from these causes that extend into the second month of life.

We thank the countless individuals who have contributed to the Global Burden of Disease Study 2013 in various capacities. We acknowledge the extensive support from all staffmembers at the Institute for Health Metrics and Evaluation and specifically thank: Kelsey Pierce for her valuable guidance; James Bullard, Serkan Yalcin, Evan Laurie, and Andrew Ernst for their tireless support of the computational infrastructure required to produce the results; Linda A Ettinger for her expert administrative support; and Peter Speyer and Eden Stork for their persistent and invaluable work to gain access to and catalogue as much data as possible to inform the estimates. We also acknowledge the support of the Rwandan Ministry of Health’s GBD Team, led by Agnes Binagwaho, for their collaboration and for reviewing the manuscript: Uwaliraye Parfait, Karema Corine, Jean Pierre Nyemazi, Sabin Nsanzimana, Yvonne Kayiteshonga, Marie Aimee Muhimpundu, Jean de Dieu Ngirabega, Ida Kankindi, Sayinzoga Felix, and Gasana Evariste. The following individuals acknowledge various forms of institutional support. RA-SS was funded by a UK MRC senior clinical fellowship. SB acknowledges additional funding or institutional support from International Development Research Center of Canada, Stanford University, and Rosenkranz Price for health-care research in developing countries. AR was supported by research grants from Brazilian research agencies CNPq and FAPEMIG. MK was supported by a NIDDK T32 grant through June 2014. RGN acknowledges that this work was supported in part by the Intramural Research Program of the National Institute of Diabetes and Digestive and Kidney Diseases. KK acknowledges the Government of India for giving him a University Grant Commission Junior Research Fellowship. GDT acknowledges support from NYU’s US National Institute of Environmental Health Sciences Center grant (number ES00260). HC and SJL are supported by the intramural programme of NIH, the National Institute of Environmental Health Sciences. KD is supported by a Wellcome Trust Fellowship in Public Health and Tropical Medicine (grant number 099876). HW, AF, HE, and FC are affiliated with the Queensland Centre for Mental Health Research, which receives funding from the Queensland Department of Health. LAR acknowledges the support of Qatar National Research Fund (04-924-3-251). TF is grateful to the Swiss National Science Foundation for an Early and an Advanced Postdoc Mobility fellowship (project number PBBSP3-146869 and P300P3-154634). IA acknowledges the UK National Institute for Health Research and the Medical Research Council for funding. HWH acknowledges support from Parnassia Psychiatric Institute, The Hague, Netherlands; the Department of Psychiatry, University of Groningen, University Medical Center Groningen, Netherlands; and the Department of Epidemiology, Columbia University, New York, NY, USA. JM has received support from the National Health and Medical Research Council John Cade Fellowship APP1056929. UM acknowledges funding from the German National Cohort Consortium. BOA acknowledges a Susan G. Komen for the Cure Research Program – Leadership Grant (number SAC110001). AJC’s work on GBD was funded by Health Effects Institute and the William and Flora Hewlett Foundation. RD acknowledges that funding from the US Deptartment of Veterans Affairs supports his salary. RD acknowledges funding from the American Parkinson’s Disease Association for support of this work. MK receives research support from the Academy of Finland, the Swedish Research Council, Alzheimer Association, and AXA Research Fund. SS receives postdoctoral funding from the Fonds de la recherche en santé du Québec. GA-C acknowledges funding and support from Health Sciences and Neurosciences (CISNEURO) Research Group, Cartagena de Indias, Colombia. No authors received additional compensation for their efforts.

Funding Bill & Melinda Gates Foundation.