Global, regional, and national incidence, prevalence, and years lived with disability for 328 diseases and injuries for 195 countries, 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016

1Evidence before this study The Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) produces the only assessment of prevalence, incidence, and years lived with disability (YLDs) for a comprehensive list of diseases and injuries, and for all countries from 1990 to the present. The World Health Organization has published YLD estimates for the years 2000–15 largely based on GBD 2015 results apart from ad hoc changes applied to selected disability weights and the prevalence of a small subset of causes. GBD 2016 is a reassessment of the burden of disease due to non-fatal diseases and injuries and updates the GBD 2015 study results. Added value of this study This study adds new knowledge about YLD rates globally and improves upon prior iterations of the GBD study in seven ways. First, new data were incorporated based on 56 356 unique data sources; these were composed mainly of peer-reviewed scientific literature identified by systematic reviews, reports from statistical agencies or ministries of health, household surveys, administrative data systems, claims data, and hospital data. In 2016, we had 14 521 sources from the scientific literature compared with 10 478 such sources used in GBD 2015. Our network of collaborators provided 2598 data sources for GBD 2016 compared with 968 available for GBD 2015; furthermore, 3430 sources of survey data were used in GBD 2016. These counts reflect our updated counting criteria for GBD 2016. Large amounts of new data for the main causes of YLDs were identified through our collaboration with the Indian Council of Medical Research and the Public Health Foundation of India. For particular diseases, the volume of available data increased substantially. Examples include Rapid Assessment of Avoidable Blindness surveys and the detailed studies reported in the Global Atlas of Helminth Infection for schistosomiasis and lymphatic filariasis. Second, we substantially changed the modelling approach for some diseases, such as cancers and tuberculosis. For cancer we improved our analysis of mortality-to-incidence ratios, resulting in considerably higher ratios in lower Socio-demographic Index (SDI) locations and thus lower YLD estimates. We also applied mortality-to-incidence ratios in the analysis of tuberculosis to better predict the gap between true incidence and notified cases—ie, undetected cases. Third, estimation at the subnational level was newly developed for Indonesia, and estimates for England were disaggregated into those for 150 local government areas. Fourth, we have disaggregated several causes to separately estimate drug-sensitive, multidrug-resistant, and extensively drug-resistant tuberculosis; latent tuberculosis infection; alcoholic cardiomyopathy; urogenital, musculoskeletal, and digestive congenital anomalies; and self-harm by firearm, to provide more detail within the GBD cause hierarchy. Additionally, Guinea worm disease was estimated due to the fact that there is policy interest in eradication, which is feasible in the near future. Sexual violence was added as a non-fatal cause of YLDs because of its relevance to several of the new sustainable development goal (SDG) targets. Fifth, we were able to incorporate inpatient hospital data by cause for 222 more locations and a final total of 3557 location-years. A separate analysis of total hospital admissions per capita by country, year, age, and sex allowed the use of hospital data sources that previously had been rejected because of incomplete knowledge about catchment populations. We extended our analyses of US medical claims data to impute a ratio of any health service contact for a diagnosis to inpatient episodes for chronic diseases that we applied to hospital inpatient data from elsewhere to predict prevalence. Sixth, we have extended our analyses of GBD results by SDI with new ways of presenting and visualising changes over time and the relationship with development. Seventh, we extended the terminal age group used in our analyses of older than 80 years into 80–84, 85–89, 90–94, and older than 95 years. Implications of all the available evidence As countries confront the effects of the epidemiological transition, there will be an increased need for up-to-date assessment of non-fatal health outcomes and exploration of the implications of growing numbers of individuals in need of chronic care as populations age. The GBD study provides opportunities to identify important non-fatal health trends across various locations and levels of development, and to assess the strength of available estimates.

2Evidence before this study

3The Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) produces the only assessment of prevalence, incidence, and years lived with disability (YLDs) for a comprehensive list of diseases and injuries, and for all countries from 1990 to the present. The World Health Organization has published YLD estimates for the years 2000–15 largely based on GBD 2015 results apart from ad hoc changes applied to selected disability weights and the prevalence of a small subset of causes. GBD 2016 is a reassessment of the burden of disease due to non-fatal diseases and injuries and updates the GBD 2015 study results.

4Added value of this study

5This study adds new knowledge about YLD rates globally and improves upon prior iterations of the GBD study in seven ways. First, new data were incorporated based on 56 356 unique data sources; these were composed mainly of peer-reviewed scientific literature identified by systematic reviews, reports from statistical agencies or ministries of health, household surveys, administrative data systems, claims data, and hospital data. In 2016, we had 14 521 sources from the scientific literature compared with 10 478 such sources used in GBD 2015. Our network of collaborators provided 2598 data sources for GBD 2016 compared with 968 available for GBD 2015; furthermore, 3430 sources of survey data were used in GBD 2016. These counts reflect our updated counting criteria for GBD 2016. Large amounts of new data for the main causes of YLDs were identified through our collaboration with the Indian Council of Medical Research and the Public Health Foundation of India. For particular diseases, the volume of available data increased substantially. Examples include Rapid Assessment of Avoidable Blindness surveys and the detailed studies reported in the Global Atlas of Helminth Infection for schistosomiasis and lymphatic filariasis. Second, we substantially changed the modelling approach for some diseases, such as cancers and tuberculosis. For cancer we improved our analysis of mortality-to-incidence ratios, resulting in considerably higher ratios in lower Socio-demographic Index (SDI) locations and thus lower YLD estimates. We also applied mortality-to-incidence ratios in the analysis of tuberculosis to better predict the gap between true incidence and notified cases—ie, undetected cases. Third, estimation at the subnational level was newly developed for Indonesia, and estimates for England were disaggregated into those for 150 local government areas. Fourth, we have disaggregated several causes to separately estimate drug-sensitive, multidrug-resistant, and extensively drug-resistant tuberculosis; latent tuberculosis infection; alcoholic cardiomyopathy; urogenital, musculoskeletal, and digestive congenital anomalies; and self-harm by firearm, to provide more detail within the GBD cause hierarchy. Additionally, Guinea worm disease was estimated due to the fact that there is policy interest in eradication, which is feasible in the near future. Sexual violence was added as a non-fatal cause of YLDs because of its relevance to several of the new sustainable development goal (SDG) targets. Fifth, we were able to incorporate inpatient hospital data by cause for 222 more locations and a final total of 3557 location-years. A separate analysis of total hospital admissions per capita by country, year, age, and sex allowed the use of hospital data sources that previously had been rejected because of incomplete knowledge about catchment populations. We extended our analyses of US medical claims data to impute a ratio of any health service contact for a diagnosis to inpatient episodes for chronic diseases that we applied to hospital inpatient data from elsewhere to predict prevalence. Sixth, we have extended our analyses of GBD results by SDI with new ways of presenting and visualising changes over time and the relationship with development. Seventh, we extended the terminal age group used in our analyses of older than 80 years into 80–84, 85–89, 90–94, and older than 95 years.

6Implications of all the available evidence

7As countries confront the effects of the epidemiological transition, there will be an increased need for up-to-date assessment of non-fatal health outcomes and exploration of the implications of growing numbers of individuals in need of chronic care as populations age. The GBD study provides opportunities to identify important non-fatal health trends across various locations and levels of development, and to assess the strength of available estimates.

8Assessment of death rates by cause has been an essential component of tracking progress in global health. The Millennium Development Goals (MDGs) emphasised child, maternal, and infectious disease mortality and spurred investments that contributed to a rapid decline in mortality from these sources, although these have not been universally achieved at the same level., Progress in reducing the disabling outcomes of disease has been much slower. The Global Burden of Diseases, Injuries, and Risk Factors Study 2015 (GBD 2015) estimated a modest 2·1% reduction in the age-standardised rate of years lived with disability (YLDs) for all causes compared with a 22·7% reduction in age-standardised rates of years of life lost (YLLs) for all causes between 2005 and 2015. The slower progress in addressing non-fatal compared with fatal health outcomes and ageing of populations make YLDs an increasingly important component of global disability-adjusted life-years (DALYs). In some high-income countries with advanced ageing, YLDs already make up more than half of the total burden in DALYs. The GBD Study is the only global effort to quantify non-fatal outcomes using a metric that allows comparisons between fatal and non-fatal outcomes of a comprehensive list of diseases and injuries.

9There are several challenges in standardising the estimation of YLDs. For example, case definitions vary; there are diverse data sources and study methods; accessible data sources are sparse for many diseases, with large parts of the world lacking adequate data; and data for severity of outcomes are limited and lack a standardised approach. This annual update of the GBD study provides an opportunity to incorporate new data and improved methods within a standardised framework to enhance the precision and accuracy of estimation.

10The extensive GBD network of more than 2518 collaborators from 133 countries and three non-sovereign locations have provided invaluable critiques of methods and helped to identify new data sources. Most of the debates arising from GBD 2015 have come as direct communications from the collaborative network or have been published by collaborators and other researchers. For example, collaborators have addressed omissions in the GBD cause list,,,,,, challenged the GBD hierarchy of causes, or explored the continual debate regarding disability weights. There is also a growing scientific literature on secondary analyses of GBD results—eg, with a focus on chronic kidney disease, oral diseases, or cardiovascular disease among the poorest billion.

11The primary objective of the non-fatal component of GBD 2016 was to estimate prevalence, incidence, and YLDs for 328 GBD causes from 1990 to 2016. For each cycle of GBD, the entire time series is re-estimated to incorporate new data and methods; thus, these results supersede previous GBD results. We explore the patterns of non-fatal disease over time and in comparison to expected levels based on an index of sociodemographic development.

12The GBD study provides a standardised analytical approach for estimating incidence, prevalence, and YLDs by age, sex, cause, year, and location. We aim to use all accessible information on disease occurrence, natural history, and severity that passes minimum inclusion criteria set disease-by-disease (appendix 1, p 33). Our approach is to optimise the comparability of data collected by varying methods or different case definitions; find a consistent set of estimates between data for prevalence, incidence, and causes of death; and predict estimates for locations with sparse or absent data by borrowing information from other locations and using covariates.

13In this study, we use different methods to reflect the available data and specific epidemiology of each disease. Our main approach is to combine all sources of information for a disease using the Bayesian meta-regression tool DisMod-MR 2.1. Subsequently, we use data for severity, the occurrence of particular consequences of diseases, or sequelae, to establish the proportion of prevalent cases experiencing each sequela. Several broad classes of alternative approaches are used within the GBD study. First, for injuries, non-fatal estimates must account for the cause of injury (eg, a fall), the nature of injury (eg, a fracture or head injury), the amount of disability arising in the short term, and permanent disability for a subset of cases. Second, cancers were estimated by assessing the association between mortality and incidence, taking into account the effect on survival of access to, and quality of, treatment for the cancer site. Third, we combined the natural history model strategy for HIV/AIDS with the DisMod-MR 2.1 modelling approach for tuberculosis as HIV infection affects outcomes in patients who also have tuberculosis. Fourth, models for malaria, hepatitis, and varicella relied on data of the presence of circulating antibodies or parasites in the blood to predict the incidence of clinical episodes for which we estimate disability. Fifth, neonatal disorders were estimated from birth prevalence data and cohort studies on the risk of death in the first month and the probability of long-term disabling outcomes. Sixth, incidence of rabies, whooping cough, diphtheria, and tetanus was estimated from cause-specific mortality rates and data on the case fatality of acute episodes (appendix 1, p 33).

14Below we describe these modelling efforts organised into eight sections; the supplementary methods (appendix 1, p 1) presents a single source for additional detail of inputs, analytical processes, outputs, and methods specific to each cause. This study complies with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER) recommendations (appendix 1, p 723).

15The GBD 2016 study was based on a geographic hierarchy that includes 195 countries and territories grouped within 21 regions and seven GBD super-regions (appendix 1, p 726). For this publication, we present subnational estimates in figures and only for Brazil, China, India, and the USA. Details of subnational estimates will be reported in separate publications.

16Cause-specific estimation in GBD 2016 was done for the years 1990, 1995, 2000, 2006, 2010, and 2016 and interpolated to get a full time series. In view of policy priorities, a subset of results focus on change over the time period 2006–16. Results from GBD 2016 by year and location can be explored further in dynamic data visualisations.

17In the GBD Study, causes and their sequelae are collectively exhaustive and mutually exclusive and are organised in a hierarchy with five levels. Level 1 contains three broad cause groups: communicable, maternal, neonatal, and nutritional diseases; non-communicable diseases; and injuries. These are broken down into 21 Level 2 causes with further disaggregation into 163 Level 3 causes and 271 Level 4 causes. Sequelae of these causes are represented at Level 5 of the hierarchy.

18For GBD 2016, we expanded the list of causes of non-fatal outcomes from 310 to 328. This involved the refinement of certain Level 3 causes into new Level 4 causes, including disaggregation of tuberculosis and HIV- tuberculosis into drug-susceptible tuberculosis, multidrug-resistant tuberculosis, extensively drug-resistant tuberculosis, and latent tuberculosis infection. Cardiomyopathy and myocarditis were further refined as alcoholic cardiomyopathy, myocarditis, and other cardiomyopathy. Other leukaemia was added as an additional sub-cause at Level 4. Self-harm was separated into self-harm by firearm and self-harm by other means. The previously named cause grouping “collective violence and legal intervention” was divided into two Level 4 causes: executions and police conflict. New causes of non-fatal outcomes added to the GBD hierarchy for 2016 included Zika virus disease; musculoskeletal, urogenital, and digestive congenital anomalies; Guinea worm disease; and sexual violence. Medication overuse headache was removed as a cause and, instead, characterised as a sequela of migraine and tension-type headache.

19The first step in non-fatal estimation was the compilation of data sources from systematic data and literature searches conducted by cause. This process resulted in 4043 published studies newly included in GBD 2016, leading to a total of 14 521. Our network of collaborators for GBD 2016 provided 2598 data sources and studies. These were systematically screened, together with sources suggested by country-level experts, surveys located in multinational survey data catalogues, and Ministry of Health and Central Statistical Office websites. We analysed 18 792 sources of epidemiological surveillance data (country-years of disease reporting), up from 14 081 in 2015. All counts reflect our updated counting criteria for GBD 2016. The supplementary methods provides details of data adjustments, correction factors, and standardisations employed in incorporating these different data types (appendix 1, p 18).

20The number of location-years of hospital inpatient data by cause increased from 1176 in GBD 2015 to 3557 in GBD 2016. This increase can be attributed to the addition of new years of data for some locations, as well as newly incorporated data for 16 countries where we had previously lacked clear information about the population covered. To allow their use in GBD, we first collated information from surveys and hospital administrative records to estimate hospital admission rates per capita for all GBD locations by age and sex, from 1990 to 2016, using DisMod-MR 2.1 (appendix 1, p 7). We then used inpatient data by cause from locations with unclear denominators as cause fractions of the all-cause inpatient admission rates. Three adjustment factors were derived from USA health insurance claims data on more than 80 million person-years of coverage. The first factor corrected for multiple inpatient episodes for the same cause in an individual. The second adjustment was to include secondary diagnostic fields. The third adjustment was to include any mention of a cause in inpatient or outpatient episodes of care as opposed to inpatient episodes with a primary diagnosis only. This new method of predicting prevalence or incidence from inpatient data allowed us to use these sources for 16 more causes than in 2015. The supplementary methods provides a detailed description of our process for inpatient data (appendix 1, p 11).

21To provide a summary view on data availability, the number of causes at the most detailed level for which we have any prevalence or incidence data from 1980 to 2016 by location is presented in the appendix (appendix 1, p 722). An online search tool is available to view all data sources that were used in the estimation process for each cause.

22Non-fatal diseases were modelled using DisMod-MR 2.1, a statistical method that synthesises sparse and heterogeneous epidemiological data for non-fatal outcomes. Estimation occurred at five levels: global, super-region, region, country, and subnational locations, with results from a higher level providing guidance for the analysis at a lower geographical level (appendix 1, p 18).

23Custom models were created where DisMod-MR 2.1 does not capture the complexity of the disease, or if incidence and prevalence needed to be calculated from other data. Further details of these custom models can be found in the supplementary methods (appendix 1, p 18). Prevalence was estimated for nine impairments, disorders that are sequelae of multiple diseases and for which there were better data available to estimate the overall occurrence than for each underlying cause: anaemia, intellectual disability, epilepsy, hearing loss, vision loss, heart failure, infertility, pelvic inflammatory disease, and Guillain-Barré syndrome.

24The methods for estimating YLDs from a number of diseases changed substantially for GBD 2016. We improved our estimation of mortality-to-incidence ratios for cancers to better reflect lower survival probabilities in low-income and middle-income locations based on each location's Socio-demographic Index (SDI) value. As a consequence, our prevalence and YLD estimates were lower in those locations but did not change much for higher-SDI locations. We made major changes to our modelling of tuberculosis. First, we made explicit estimates of latent tuberculosis infection from tuberculin skin testing data and the risk of developing active tuberculosis by induration size. Second, we predicted mortality-to-incidence ratios in locations with high data-quality ratings (4-star or 5-star using a system developed for the GBD 2016 causes of death estimation) and SDI as a covariate. We anchored the lower end of the SDI scale with a datapoint from an untreated cohort of pulmonary tuberculosis cases in the 1960s, half of whom had died after five years., Third, we estimated incidence from these mortality-to-incidence ratios in all locations except those with higher reported notifications. Fourth, we modelled these incidence estimates as well as the prevalence data from surveys in low-income and middle-income countries and cause-specific mortality rates among the proportion of the population with latent infection in DisMod-MR 2.1. Fifth, we estimated the proportions of tuberculosis cases with multidrug-resistant tuberculosis or extensively drug-resistant tuberculosis from notification and survey data and included an increased risk of multidrug-resistant tuberculosis in HIV/AIDS-infected patients with tuberculosis from a meta-analysis.

25In our measles estimation strategy, we included the coverage of measles-containing vaccine second-dose (MCV2) rather than just the coverage of the primary vaccine as a covariate. As relatively few countries in sub-Saharan Africa have introduced MCV2, the estimated incidence for those locations is notably higher compared with previous estimates.

26For 214 causes at Level 4 of the GBD hierarchy, sequelae were defined in terms of severity, usually graded as mild, moderate, or severe outcomes. We followed the same approach as in GBD 2015. For Zika virus disease, we included sequelae for those with symptomatic acute infection, a small proportion with Guillain-Barré syndrome, and the number of neonates with congenital Zika virus disease as reported to the Pan American Health Organization (PAHO). For sexual violence, we estimated YLDs associated with concurrent physical injuries and the short-term psychological outcomes following sexual violence.

27A more substantial change in estimating severity was applied to stroke. A systematic review was done to collect data on modified Rankin scores, a measure of neurological disability. Levels of Rankin score were analysed in DisMod-MR 2.1 and mapped to the existing GBD health state lay descriptions for mild, moderate, and severe motor impairment from stroke, and, separately, the proportion of stroke patients with moderate-to-severe motor impairment who also experienced cognitive impairment. For GBD 2016 we used the same disability weights as in GBD 2013 and GBD 2015; the supplementary methods provides a complete listing of lay descriptions of all 235 health states used in GBD 2016 (appendix 1, p 799).

28We estimated comorbidity by simulating 40 000 individuals in every location-age-sex-year combination as exposed to the independent probability, based on the prevalence of the sequelae included in GBD 2016. In simulants with two or more sequelae, we assumed a multiplicative function to combine disability weights and then distributed the reduced combined weight proportionately among all comorbid sequelae. Averaging these adjusted values across all simulants with a particular sequela gave the adjusted value of YLDs. There was no change in the approach compared with GBD 2015.

29All computations in GBD were done 1000 times, every time drawing from the distribution of the sampling error of data inputs, the uncertainty of data corrections for measurement errors, the uncertainty in coefficients from model fit (eg, in DisMod-MR 2.1), and the uncertainty of severity distributions and disability weights. Uncertainty bounds for a quantity of interest were defined by the 25th and 975th value of the ordered 1000 estimate values. If there was a change in disease estimates between locations or over time that was in the same direction in more than 950 of the 1000 samples we report it as significant. Age-standardised prevalence YLD rates were calculated based on the GBD reference population.

30The GBD cause hierarchy is comprehensive and includes 35 residual disease categories to capture YLDs from conditions for which we do not currently make separate estimates. For 22 of these residual categories, we made explicit epidemiological estimates of prevalence and incidence, and define sequelae based on the most common diseases in the Level 2 or 3 cause group and severity distributions from the Medical Expenditure Panel Survey (MEPS). For 13 residual categories, we had no epidemiological data and estimated YLDs from the ratio of YLDs to YLLs from explicitly modelled diseases in the cause category, assuming that relative to each death, the number of YLDs was similar to that of other diseases at the same level of the GBD hierarchy (appendix 1, p 29).

31SDI is a summary measure that places all GBD locations on a spectrum of socioeconomic development. The SDI was developed for GBD 2015 to provide a comparable metric of overall development. This was achieved by using an equal weighting of lag-distributed income per capita, average years of education in the population over age 15 years, and total fertility rate. For GBD 2016, we modified the estimation of SDI by taking into consideration that SDI scales were subject to change based on increasing geographic units and an extended time period of analysis, affecting the interpretability across GBD iterations. We redefined the values of zero and one for each component of the index: zero now represents the level below which we have not observed GDP per capita or educational attainment or above which we have not observed the total fertility rate in known datasets. Maximum scores for educational attainment and Lagged Distributed Income represent a plateau in the relationship between each of the two components and life expectancy or under-5 mortality rates, suggesting no additional benefit. Analogously, the maximum score for total fertility rate represents the minimum level at which the relationship with the selected health outcomes plateaued. An SDI value was generated for each location and year as the geometric mean of each component score. Five SDI quintiles, high, high-middle, middle, low-middle, and low, were selected based on 2016 values of SDI; additional details are available in the supplementary methods (appendix 1).

32A Gaussian process regression was used to evaluate the average relationship for each age-sex-cause group, for cause-specific YLD rates on SDI at Levels 1, 2, and 3 of the GBD cause hierarchy using data from 1990 to 2016. These rates were used as the expected values for cause-specific YLD rates at a given level of SDI. Additional detail on this analysis is available in the supplementary methods (appendix 1, p 30) and in previous GBD publications.

33The funder of the study had no role in study design, data collection, data analysis, data interpretation, or the writing of the report. All authors had full access to the data in the study and had final responsibility for the decision to submit for publication.

34Global prevalence, incidence, and YLDs for 328 causes and nine impairments, as well as percent change of YLDs and percent change of age-standardised YLD rates from 2006 to 2016 are listed in the table. Cause-specific estimates for each year of the GBD estimation period 1990–2016 by location, age, and sex are available through an online results tool.

35In 2016, the ten causes with the greatest prevalence were caries of permanent teeth (2·44 billion, 95% UI 2·29 billion to 2·59 billion), latent tuberculosis infection (1·91 billion, 1·79 billion to 2·03 billion), tension-type headache (1·89 billion, 1·71 billion to 2·10 billion), age-related and other hearing loss (1·27 billion, 1·21 billion to 1·34 billion), iron-deficiency anaemia (1·24 billion cases, 1·21 billion to 1·28 billion), migraine (1·04 billion, 1·00 billion to 1·09 billion), glucose-6-phosphate dehydrogenase deficiency (G6PD) trait (866 million, 852 million to 882 million), genital herpes (860 million, 748 million to 992 million), refraction and accommodation disorders (854 million, 822 million to 886 million), and ascariasis (800 million, 738 million to 872 million; table). The vast majority of cases for two of these causes are asymptomatic sequelae that have no YLDs associated with them: genital herpes with no active lesions, and G6PD trait that does not result in anaemia. Similarly, latent tuberculosis infection is a highly prevalent cause but without any associated disability. The leading ten causes of prevalence accounted for 17·5% (95% uncertainty interval [UI] 15·8–19·3) of YLDs globally in 2016.

36In 2016, the ten causes with the highest incidence were upper respiratory infections (17·78 billion, 15·85 billion to 19·90 billion), caries of permanent teeth (7·26 billion, 6·72 billion to 7·84 billion), diarrhoeal diseases (4·48 billion, 4·25 billion to 4·74 billion), fungal skin diseases (2·10 billion, 1·88 billion to 2·34 billion), caries of deciduous teeth (1·76 billion, 1·26 billion to 2·39 billion), tension-type headache (1·06 billion, 899 million to 1·24 billion), other sense organ diseases (708 million, 668 million to 754 million), other skin and subcutaneous diseases (634 million, 617 million to 652 million), vitamin A deficiency (552 million, 469 million to 671 million), and pyoderma (474 million, 461 million to 489 million; table). These ten causes of incidence accounted for 4·4% (3·7–5·6) of all YLDs globally in 2016.

37Globally, communicable, maternal, neonatal, and nutritional deficiency diseases accounted for 12·6% (95% UI 11·3–14·0) of YLDs in 2016 (101·5 million, 72·3 million to 136·9 million), while NCDs accounted for 80·6% (78·2–82·5) or 648·6 million (481·3 million to 836·5 million) of YLDs and injuries accounted for 6·9% (5·6–8·4) or 55·4 million (37·2 million to 78·7 million) of YLDs (table). In 2016, the age-standardised YLD rate for all causes was lowest in China at 9201 YLDs per 100 000 (95% UI 6862–11 943 per 100 000); Yemen had the highest age-standardised YLD rate at 14 774 YLDs per 100 000 (11 018–19 228 per 100 000).

38Figure 1C shows the change in global YLDs over time in three ways: mean percentage change in number of YLDs, mean percentage change in all-age rates, and mean percentage change in age-standardised YLD rates from 1990 to 2006 and from 2006 to 2016. The mean percentage change in the number of YLDs reflects the combined effects of population growth, population ageing, and epidemiological change. Population ageing and epidemiological change explain the mean percentage change in all-age rates. The mean percent change in age-standardised YLD rates reflects epidemiological change that is not due to ageing or population growth. All top-30 YLD causes increased in the number of YLDs between 1990 and 2016. Alcohol use disorder, major depressive disorder, and refraction and accommodation errors had the largest declines in age-standardised YLD rates from 2006 to 2016 but the change did not exceed 5%. Preterm birth complications had the largest increase in age-standardised YLD rates, 8·5%. Among the top 30 causes of YLDs, the difference between the change in all-age and age-standardised rates between 2006 and 2016 was 10% or greater for age-related and other hearing loss, diabetes, chronic obstructive pulmonary disease (COPD), osteoarthritis, ischaemic stroke, edentulism, and ischaemic heart disease, indicating that these diseases largely affect the elderly and therefore become more prominent causes of YLDs in an ageing global population.

39Low back pain and migraine were the leading causes of YLDs in high-income, high-middle-income, and middle-SDI quintile countries, but iron-deficiency anaemia was the leading cause in low-middle-SDI and low-SDI quintiles in 2016 (appendix 2, p 14). Low back pain, migraine, and age-related and other hearing loss were in the top five causes of YLDs in all SDI quintiles. Major depressive disorders appeared in the top five in all quintiles but the middle, where it was displaced by diabetes. Neck pain was a top five cause of YLDs in high-income, high-middle-income, and middle-SDI quintiles. In 2016, 28 of the 30 leading causes of age-standardised YLD rates for high-SDI countries were NCDs, versus 23 out of 30 in low-SDI countries. Between 2006 and 2016, there were large drops in age-standardised rates of YLDs for malaria, HIV/AIDS, onchocerciasis, and schistosomiasis in the low-SDI quintile.

40For men, the most common leading detailed cause of YLDs in 2016 was low back pain, resulting in the highest age-standardised rates of YLDs in 133 of 195 countries and territories, including every country in the high-income regions, central and eastern Europe, central Asia, Andean and Tropical Latin America, and eastern and central sub-Saharan Africa, as well as most countries in southeast Asia, north Africa and the Middle East, and western sub-Saharan Africa. Diabetes, the second most common leading cause of YLD rates for men, ranked first in 38 countries in Central Latin America, the Caribbean, Oceania, and North Africa and the Middle East. Iron-deficiency anaemia was the leading cause in India, Bhutan, Sudan, Yemen, and Mali. HIV/AIDS was the leading cause in South Africa, Lesotho, Swaziland, Namibia, and Botswana.

41In 2016, low back pain was the leading cause of age-standardised YLD rates for women in 104 of the 195 countries and territories (figure 2B). It was the main cause of YLDs in almost all high-income, central Europe, eastern Europe, North Africa and the Middle East, and Andean and Tropical Latin American countries. Iron-deficiency anaemia was the leading cause for women (35 countries), followed by migraine and diabetes in 24 and 17 countries, respectively. HIV/AIDS was the leading cause in southern sub-Saharan Africa, Zambia, and Malawi.

42Trends for the number of YLDs (in millions) and age-standardised YLD rates from 1990 to 2016 at Level 1 of the GBD cause hierarchy are shown by SDI quintile in figure 3. Generally, there was little or no change in the age-standardised rates of YLDs for NCDs, communicable, maternal, neonatal, and nutritional causes, or injuries, apart from a decrease in communicable, maternal, neonatal, and nutritional YLD rates between 2000 and 2016 in the low-SDI quintile. The large increase in total YLD numbers for NCDs reflects the combined effect of population growth, ageing, and epidemiological change. NCDs contributed the most YLDs at each SDI quintile, followed by communicable, maternal, neonatal, and nutritional causes and injuries, with exceptions for high and high-middle quintiles, where injuries contributed more YLDs than communicable, maternal, neonatal, and nutritional causes in 2016. This was particularly the case at the highest level of SDI, where YLDs from communicable, maternal, neonatal, and nutritional causes and injuries represented only a small fraction of overall YLDs.

43Between 1990 and 2016 the number of YLDs for 21 Level 2 causes increased, particularly in the 40–69 year age range (figure 4). In childhood, other NCDs (skin diseases being the largest contributor in this category), nutritional deficiencies, infectious diseases, and mental and substance use disorders were the main causes of YLDs. After childhood, NCDs were the dominant source of YLDs. Mental and substance use disorders were the largest contributors to disability in young adults, while at older ages, other NCDs (with hearing loss and vision loss being the largest contributors to this category), musculoskeletal disorders, and cardiovascular diseases were the most important causes. YLDs from injuries were largest between ages 20 and 69 years.

44YLD rates were higher in males than in females at ages under 10 years and between ages 75 and 94 years (figure 5). At all other ages, all-cause YLD rates were higher in females. Boys had higher YLD rates for mental disorders (particularly conduct disorder and autism), while girls had higher YLD rates for other non-communicable diseases (most of this difference is explained by higher YLD rates from dermatitis). At ages over 10 years, musculoskeletal disorders contributed most to higher YLD rates in females. The next two cause groups with higher YLD rates in females were neurological disorders (particularly migraine and, at older ages, Alzheimer's disease and other dementias) and nutritional deficiencies (mostly iron-deficiency anaemia). The sex differences in mental and substance use disorders were small, with slightly elevated rates in younger males, and somewhat higher rates in older females. The small overall sex difference masks much higher YLD rates from depressive disorder and anxiety in females and higher YLD rates from injuries, substance use disorders, and autism spectrum disorders in males. Adult males older than 55 years had higher YLD rates from diabetes, urogenital, blood, and endocrine disorders (diabetes and benign prostatic hyperplasia being the main drivers of differences), injuries, chronic respiratory diseases, and cancers.

45Figure 6 shows the annualised rates of change in age-standardised YLD rates between 2006 and 2016, with Level 3 causes against age-standardised YLD rates presented on a log-scale from right to left. 15 causes significantly increased by more than 1%: six cancers (brain and nervous system, liver, non-Hodgkin lymphoma, testicular, thyroid, and other neoplasms), dengue, gastritis and duodenitis, leishmaniasis, neonatal encephalopathy due to birth asphyxia and trauma, neonatal sepsis and other neonatal infections, other chronic respiratory diseases, other neurological disorders, peptic ulcer disease, and Zika virus disease. Of 31 causes with declines of greater than 1·0% per year, 25 were communicable, maternal, neonatal, and nutritional causes and five were NCD causes. Road injuries was the only cause out of 30 causes with age-standardised rates of YLDs greater than 100 per 100 000 that significantly increased by more than 0·5% annually. There were no large causes with a greater than 0·5% decrease in age-standardised YLD rates.

46Age-standardised YLD rates for the top ten most detailed causes are plotted by region and year against SDI in the supplementary results (appendix 2, p 4). The black solid line represents the expected value based on SDI. Two of the top ten causes, iron-deficiency anaemia and age-related and other hearing loss show a large decline in expected values with increasing SDI, while the regional estimates are relatively close to the expected line. The interpretation is that much of the variation in YLD rates for these two causes is linked to SDI. The expected lines for the other top ten conditions show a less clear pattern with SDI while regional estimates are dispersed widely around the line. These graphs show that most of the variation in YLD rates for these leading causes of YLDs depends on factors other than sociodemographic development, unlike the much clearer relationship between SDI and rates of YLLs for the leading causes of death.

47Low back pain and migraine ranked in the top ten of YLDs in all 195 countries and territories in 2016. Adult-onset hearing loss was in the top ten in 193 countries and territories. Major depressive disorder ranked in the top ten in all but four countries. Anxiety disorders, iron-deficiency anaemia, neck pain, diabetes, and other musculoskeletal disorders ranked in the top ten of more than half of the countries and territories (figure 7).

48Apart from showing the top ten causes of YLDs by country, figure 7 also shows for each top ten cause the variation from expected values based on SDI. The cells in dark red have at least twice the expected values and those in dark blue at least 50% lower values. In the high-income super-region, the USA had much higher than expected YLDs from opioid dependence. YLDs due to Alzheimer's disease and other dementia were more than double what was expected in Italy and Japan. Greece had much higher than expected rates of COPD. YLDs from other musculoskeletal disorders were more than twice the expected values in Australia, Canada, Chile, Greenland, and the USA. Many eastern and central European countries had much higher than expected YLDs from falls, alcohol use disorders, and ischaemic stroke. Four countries in central Asia had more than twice the expected rates of iron-deficiency anaemia. In Latin America, YLDs from diabetes were more than twice the expected rate in Barbados, Jamaica, Mexico, Puerto Rico, Saint Vincent and the Grenadines, Trinidad and Tobago, and the Virgin Islands. Venezuela had less than half the expected rate of YLDs from low back pain but more than double the expected rate of YLDs from diarrhoea. Haiti had higher than expected YLDs from asthma. YLDs from iron-deficiency anaemia were more than double the expected rate in five Caribbean nations. China had more than twice the expected YLD rate of ischaemic stroke, while North Korea had more than double the expected YLD rate for neck pain. In the southeast Asia region, YLD rates of iron-deficiency anaemia were more than twice the expected values in Maldives and Mauritius. Mauritius also had more than twice the expected YLD rates for diabetes and COPD. Thailand and the Seychelles had more than twice the expected rate of COPD. Timor-Leste was the only country in the region where lymphatic filariasis and long-term disability from conflict ranked among the top ten causes of YLDs. Diabetes ranked highly in all island nations of Oceania with much higher YLDs than expected based on SDI. In North Africa and the Middle East, more than twice the expected YLD rates from opioid dependence were found in Afghanistan, Bahrain, Iran, Iraq, Jordan, Kuwait, Lebanon, Libya, Palestine, Oman, Qatar, and the United Arab Emirates. Afghanistan and Bahrain had more than twice the expected YLD rates of diabetes. Long-term disability from injuries sustained during conflict ranked in the top ten for Afghanistan, Lebanon, Syria, and Iraq.

49Iron-deficiency anaemia was much higher than expected in India and Bhutan. India had considerably higher than expected YLDs from COPD. Bangladesh was the only country in 2016 with YLDs from the residual category of other congenital anomalies in the top ten. Bhutan had more than double the expected YLDs from diarrhoea. HIV/AIDS was the leading cause of YLDs in southern sub-Saharan Africa. Malaria ranked in the top ten for YLDs in 14 sub-Saharan African countries. Three neglected tropical diseases, schistosomiasis, lymphatic filariasis, and onchocerciasis, ranked in the top ten of YLDs in six, three, and seven sub-Saharan African countries, respectively. In Madagascar, protein-energy malnutrition ranked in the top ten. Epilepsy was a top ten cause of YLDs in Senegal and São Tomé and Príncipe. Much lower than expected YLDs from iron-deficiency anaemia were found in Burundi, Ethiopia, and Rwanda. The long-term consequences of injuries sustained during the 1994 genocide in Rwanda still ranked among the top ten causes of YLDs in 2016.

50The largest Level 2 disease groups contributing to non-fatal burden in 2016 were mental and substance use disorders (18·7%, 95% UI 15·9–21·0), other NCDs (18·6%, 15·6–22·5), and musculoskeletal disorders (17·1%, 15·3–18·9), together covering more than half of global YLDs. Despite mostly stagnant age-standardised rates, the number of YLDs from NCDs grew rapidly in all the SDI quintiles, due to population growth and ageing. Across all causes of YLDs, age-standardised rates of YLDs decreased between 1990 and 2016 by 2·7% (95% UI 2·3–3·1) in contrast to the 39·1% (37·8–40·2) decrease in YLL rates over the same time period. Thus, the relative contribution of YLD to the overall burden of disease in DALYs increased from 21·7% (17·2–26·6) in 1990 to 33·5% (27·4–39·7) in 2016.

51Age-standardised YLD rates for all conditions combined were 10·4% (9·0–11·8) higher in women compared to men. Iron-deficiency anaemia, migraine, Alzheimer's disease and other dementias, major depressive disorder, anxiety, and all musculoskeletal disorders apart from gout were the main conditions contributing to higher YLD rates in women. Men had higher age-standardised rates of substance use disorders, diabetes, cardiovascular diseases, cancers, and all injuries apart from sexual violence.

52Globally, we found much less geographical variation in disability than has been documented for premature mortality. In 2016, there was a less than two-fold difference in age-standardised YLD rates for all causes between the location with the lowest (China, 9201 YLDs per 100 000, 95% UI 6862–11 943 per 100 000) and highest rates (Yemen, 14 774 YLDs per 100 000, 11 018–19 228 per 100 000). By contrast, there was a greater than tenfold range in age-standardised rates of YLLs between countries in GBD 2016. Two other countries with mainly Chinese populations, Singapore and Taiwan (Province of China), were ranked second and fourth lowest in terms of age-standardised YLD rates, respectively. The low YLD rates in China were largely determined by much lower prevalence of headaches, musculoskeletal disorders (in particular low back pain), major depressive disorder, iron-deficiency anaemia, falls, and anxiety. Low rates of reporting chronic pain and low prevalence of depressive disorder and anxiety in China and among Chinese immigrants elsewhere have frequently been reported, but these findings may partly be explained by standard diagnostic tools inadequately detecting cases in the Chinese and cultural differences in the perception and communication of pain.,,,,

53Trends across the causes of YLD were also different. For example, age-standardised YLD rates for NCDs were stagnant between 1990 and 2016 and showed little variation by SDI quintile; YLD rates for communicable, maternal, neonatal, and nutritional diseases were highest at lower SDI quintiles but with a steady drop; and injury YLD rates showed modest declines since 1990. Over the same period, age-standardised YLL rates of communicable, maternal, neonatal, and nutritional diseases rapidly declined, particularly in lower SDI quintiles. The decline in YLL rates from NCDs was slower than the decline in communicable, maternal, neonatal, and nutritional diseases but still substantial. The main causes of YLLs and YLDs were also notably different. Of the top 30 causes of YLLs, only seven appeared in the top 30 causes of YLDs (diabetes, falls, COPD, ischaemic stroke, neonatal preterm birth complication, diarrhoeal diseases, and ischaemic heart disease). At Level 4, 17 of the top 30 causes of YLDs were diseases for which we estimated no YLLs as they are not considered underlying causes of death.

54The availability and quality of epidemiological data the GBD Study draws upon to make non-fatal estimates vary enormously among diseases and by location. A multitude of study methods and preferred case definitions pose challenges to making comparable estimates. A large part of the effort in the GBD study in making non-fatal estimates is to identify and correct for known sources of measurement error. The main classes of data sources each have their own limitations. Surveys can be biased by low response rates or by exclusion of individuals who do not reside in a traditional household, and may be limited in the capacity to properly identify cases within the limited diagnostic means available during a survey. Administrative data on health service encounters include only those seeking and receiving health care and are likely biased toward more severe disease, but may offer more accurate diagnostic variables than available in surveys. With the development of electronic medical records and advancements in linkage between data sources, non-fatal population health assessments can benefit tremendously if survey and administrative records can be linked. This would help to overcome many of the limitations of individual data sources: administrative records can provide rich diagnostic information, while surveys can provide better information on cases that are not in contact with health services, exposure to risk factors, and health status measures that can provide information on severity. The challenges in low-income and middle-income countries will be to fund the data systems and infrastructure and development of technical expertise, and to provide access to generic rather than proprietary software to manage electronic medical record data and produce the linkages while preserving confidentiality.

55In the GBD 2016 companion paper on causes of death we reported a faster decline in age-standardised YLL rates for the top ten causes of YLL (ischaemic heart disease, cerebrovascular disease, lower respiratory infections, diarrhoeal diseases, road injuries, neonatal preterm birth complications, malaria, COPD, HIV/AIDS, and neonatal encephalopathy) than for the rest of causes, globally and in each of the SDI quintiles. We hypothesised that a greater investment in large causes has led to greater improvements. We did not observe an equivalent rapid decline in the age-standardised YLD rates for these top ten causes of YLLs. YLD rates for ischaemic heart disease, stroke, LRI, COPD, and road injuries increased by less than 10% over the last ten years. YLDs for diarrhoeal diseases declined by less than 10%. The only top ten YLL causes with a greater decline in YLD rates were HIV/AIDS and malaria, though at a lesser pace than the change in YLL rates. Antiretroviral treatment greatly reduces the death rate in persons living with HIV but also improves the immune status of survivors. Thus, while there was a small increase in prevalence, the age-standardised YLD rate dropped by almost a fifth since 2006 as a larger proportion of cases were estimated to be in higher CD4 categories, for which we apply a lower disability weight. YLD rates from preterm birth complications and neonatal encephalopathy increased, while YLL rates dropped by almost a quarter. Improved survival through neonatal intensive care interventions predisposes survivors to the risk of long-term disabling outcomes. In terms of how we measure the burden of disease in DALYs, this means that large gains by prevention of YLLs can be accompanied by a smaller amount of health loss in terms of YLDs in new survivors. Similarly, health interventions that address mortality by reducing the case fatality among causes of disease or injury may in turn lead to an increase in non-fatal outcomes. For example, reducing cardiovascular risks in people with diabetes will increase the prevalence of those with sequelae such as neuropathy or vision loss. Similarly, preventing death from a myocardial infarction by revascularisation improves survival and thus exposes more people to heart failure as a complication.

56The MDGs focused health policy on predominant infectious, neonatal, and maternal causes of death. The health-related SDGs have expanded into a number of non-communicable diseases and injuries. However, the focus remains on reducing mortality rather than the main causes of disability. The only goals that correspond with non-fatal outcomes are those related to substance use (goal 3.5) and physical and sexual violence (goals 5.2, 11.7, 16.1, and 16.2). Furthermore, the targets formulated on incidence of HIV, tuberculosis, malaria, and hepatitis were included because of the large number of deaths associated with these infectious diseases rather than presumably out of concern for their non-fatal outcomes. Diabetes and opioid use disorder are the only causes in the top 20 of age-standardised YLD rates at Level 4, globally, that are mentioned in the SDGs. The diabetes target, however, is formulated as an indicator of mortality only. The mental, musculoskeletal, sense organ, and neurological disorders that contribute to more than half of all YLDs are not considered. While this may partially be a bias toward assessing global health in terms of mortality, it may also reflect a perception that there is inadequate knowledge to address these major causes of disability. Indeed, the much smaller gap in YLD rates compared to the large differences in YLL rates between high-SDI and low-SDI countries suggests that the capacity to intervene in non-fatal outcomes is more limited.

57Several threats exist that could lead to reversals in global health gains, such as widespread antimicrobial resistance,, conflict, climate change,, and obesity. Of these, obesity has the more immediately apparent health consequences and has a large effect on non-fatal outcomes such as diabetes, low back pain, and osteoarthritis. The increase in diabetes incidence due to obesity and the improved survival in people with diabetes, largely through prevention of deaths from cardiovascular complications, both contribute to increases in prevalence and a need for intensive long-term management to prevent fatal and disabling complications. Recent conflicts such as those in Libya, South Sudan, Syria, and Yemen are major health threats not only in terms of casualties but also because they lead to long-term physical and mental consequences directly related to the violence as well as the broader health consequences of disrupted health services and declining economic status.

58The intensive collaborations with countries for which we present subnational estimates and the expanding network of GBD collaborators have contributed to a 36·2% increase in data sources for non-fatal outcome estimates compared to GBD 2015, as well as greater scrutiny of our methods and results. In particular, the close collaboration with the Indian Council for Medical Research and the Public Health Foundation of India, and 14 disease expert groups convened by these partners, has greatly enhanced the quality of the estimates for India. There was a 49·9% increase in non-fatal data sources in India, and disease experts provided access to state-level data for causes, while in GBD 2015 we had only access to national aggregates.

59Increasingly, the GBD Study is making use of administrative data based on health care encounters. Public access to these data sources at the level of detail required for GBD analyses remains limited, and enormous scope remains to expand access to these valuable sources of health information in many more countries. By estimating total admission rates from both administrative and survey sources, we were able to increase the number of countries with available inpatient data from 26 in 2015 to 41 in 2016. To predict the level of disease under any type of care for other countries with inpatient data available, we applied the ratio of people with a diagnosis during inpatient and outpatient episodes to that of inpatient diagnoses only in US medical claims data. In our analyses for most conditions, the administrative data sources were adjusted to the level of a reference case set for the disease of interest. Thus, we make use of the detailed information on age and sex in these data sources while adjusting for a potential systematic bias compared to data sources with measurements according to the reference case. For instance, our reference case for diabetes prevalence was the proportion of a population with fasting plasma glucose greater than 7 mmol/L or on diabetes treatment in a representative survey. The adjustment of administrative data reflects the proportion of people with diabetes who may be unaware of their condition or who for other reasons have not contacted health services in the year of interest.

60To estimate indicators of sexual violence for the SDGs, we added sexual violence as a subcause of interpersonal violence in the GBD hierarchy. The focus of the effort in GBD 2016 has been to estimate the prevalence of people who report sexual violence by intimate partners and other perpetrators in the last 12 months. For non-fatal outcomes, we quantified the immediate concurrent physical injuries and the more immediate mental health consequences. We realise this is a limited scope for estimating all relevant health loss from sexual violence. In future iterations of GBD we intend to include sexual violence as a risk factor for longer-term outcomes such as major depressive disorder, anxiety disorder, and substance use disorders. Going forward, we aim to become more comprehensive in our estimation of violence to include the wider health consequences of physical violence to children and adults as well as bullying.

61Although efforts were undertaken to update data for all causes of YLDs in GBD 2016, a number of sources of new data were especially influential in improving estimates for selected causes. A large number of rapid-assessment vision loss surveys contributed to an increase in the number of sources to measure vision loss and its underlying causes from 397 in 2015 to 768 in 2016, with data available in 75 low-SDI and middle-SDI countries. Likewise, the Global Atlas for Helminth Infections Project provided a large amount of new data for schistosomiasis and lymphatic filariasis.

62A new approach to fitting ensembles of distributions improved the estimation of anaemia by better representing the tail end of the distributions of haemoglobin that most concerns our estimation of disability than can be derived from a single parametric distribution. This is important because for many surveys for which we rely on academic papers and survey reports we only have data for means and standard deviations.

63The change in modelling strategy for tuberculosis was substantial. The new estimates of the prevalence of latent tuberculosis infection is of policy interest, as there are intervention strategies that target treatment of latent tuberculosis in high-risk individuals. It allowed us to model active disease among those with latent infection to improve the search for a consistent fit between cause of death rates, prevalence surveys, and notification of incident disease, the three main sources of epidemiological data on tuberculosis. Next, we estimated the association between SDI and the ratio of mortality to incidence, assuming this measure of case fatality is dependent on development as a proxy for quality of case finding and management. This provided greater alignment of mortality and non-fatal estimates of tuberculosis than we have been able to achieve in past GBD studies.

64A substantial new effort on the severity of stroke was the analysis of 18 studies providing Rankin score data on the severity of disability in stroke survivors. Thus, we were able to make estimates by age, sex, year, and location to capture greater nuances in outcomes – for example, reflecting differences in access to rehabilitative services.

65The GBD study is the only source of comprehensive quantification of the disabling outcomes of diseases and injuries. For a number of diseases there are efforts to estimate global disease prevalence or incidence. Where appropriate, in the following sections on major sources of YLDs, comparisons with other global estimates are made; further detail can also be found in the supplementary methods (appendix 1, p 30).

66GBD 2016 confirms that mental and substance use disorders, led by depressive disorders, are a major cause of non-fatal burden. With age-standardised prevalence and YLD rates for mental and substance use disorders showing less than 10% change between 1990 and 2016, apart from bulimia YLD rates which increased by almost 20%, our findings show no sizeable improvement in population mental health over time at the global level. Interventions can decrease the severity or increase remission for certain disorders and can be effectively administered in low-income, middle-income, and high-income settings. However, these have yet to be brought up to scale in most countries. Because there are no data measuring the severity of mental disorders in a consistent manner, we have been unable to capture treatment effects that alter severity of disease. By imposing severity distributions from two high-income countries only, we are probably underestimating the burden of mental disorders in countries with less access to quality care.

67Low back pain and neck pain are the two largest causes of musculoskeletal disability; their measurement is fully dependent on self-report measures. In GBD, we adjust for variations in recall period, anatomical location, minimum duration of episodes, and whether activity-limiting or not. Although less than 40% of low back pain can be attributed to occupational risks or increased BMI, there are no risks in GBD that are linked to neck pain. The lack of predictive covariates raises concern that we might be assigning measurement error as spatial variation and that the two to two-and-a-half-times difference in age-standardised prevalence between locations is overestimating true differences in disease occurrence of these two conditions.

68Several factors have contributed to the increase in the prevalence of diabetes. Ageing of populations along with greater exposure to lifestyle-related risk factors, most importantly high BMI, has increased the incidence of diabetes in almost all countries. At the same time, improvements in treatment of diabetes have increased the life expectancy among people with diabetes. The large increase in prevalence imposes a substantial economic burden on health-care systems. In the USA, diabetes was responsible for the largest health-care spending and the greatest increase over the past two decades among 155 health conditions. This highlights the importance of development and implementation of more effective population-level strategies to prevent diabetes.

69Age-standardised prevalence of dementia varied four-times between countries. Dementia surveys generally have a two-step approach to identifying cases, with an initial screening phase followed by more intensive diagnostic procedures in those who screened positive. We scrutinised all sources of dementia and identified a wide range of screening and diagnostic tools, varying thresholds on these methods, and diagnoses made according to different classification systems. This made it impossible to identify a reference study method or case definition. Thus, it is likely that we are overestimating the variation in prevalence.

70We changed our strategy to include medication overuse headache as a sequela of migraine and tension-type headache. This led to an increase in their prevalence and YLD estimates, particularly for migraine as more than two-thirds of medication overuse headache occurs in people with migraine as the primary headache. Thus, migraine has become the second largest cause of disability in 2016. As limited surveys are available that report on the frequency and duration of headache episodes, we have not been able to quantify any effect of treatment.

71Globally, malaria case incidence trends estimated in GBD 2016 followed a very similar pattern to GBD 2015, where there was a gradual rise in cases until 2005 followed by a steady decline. For all years, however, the updated estimates are approximately one-quarter lower than those from GBD 2015, reflecting mainly lower estimates outside Africa and in particular for India. Outside Africa, and for lower-burden countries within Africa, 2016 estimates were produced using a spatiotemporal geostatistical model. Globally, this improved approach led to a reduction in estimated cases of around 75 million (about 26% reduction) of which most (roughly 60 million fewer cases) were in India, with other notable reductions in Myanmar (roughly 11 million fewer), Indonesia (roughly 3 million fewer), and Pakistan (roughly 1 million fewer). In high-burden countries in sub-Saharan Africa, where the methodology remained similar to GBD 2015, changes were relatively modest and reflected the inclusion of newly available cross-sectional parasite rate surveys or updates to data on malaria intervention coverage in recent years.

72For the first time, we have estimated the prevalence of latent tuberculosis infection. Globally, this disease affects 1·91 billion people (95% UI 1·79 billion to 2·03 billion) who are at risk of developing active disease when the immune system weakens through old age, HIV infection, diabetes, malnutrition, or excessive alcohol use. Globally, we estimate there were 10·4 million (9·37 million to 11·7 million) incident cases of tuberculosis in 2016, which is the same as the 10·4 million cases estimated by WHO for 2015. The separate estimation of 331 000 cases (95% UI 293 000–373 000) of multidrug-resistant and 19 800 incident cases (17 300–22 600) of extensively drug-resistant tuberculosis with or without HIV infections has great policy relevance, as the resources required for treatment of resistant cases are many times greater given that treatment needs to be applied for a longer period and with more expensive drugs. Drug resistance also requires more expensive detection and monitoring methods.

73A major change in methods for estimating burden of HIV/AIDS for GBD 2016 was the distribution of antiretroviral therapy coverage by age, sex, and CD4 count. We used two AIDS Indicator Surveys, to predict the age-sex-CD4 distribution of antiretroviral therapy coverage. This shifted the coverage distribution to groups with higher CD4 counts and gave a better fit to the data. UNAIDS produces periodic updates to global, regional, and national estimates of HIV/AIDS incidence and prevalence. In their latest assessment there were 36·7 million (30·8 million to 42·9 million) people living with HIV/AIDS in 2016, compared with 36·4 million (34·2 million to 39·1 million) estimated by GBD 2016. Comparisons of prevalence estimates at the country level in 2005 (the estimated peak of the HIV/AIDS epidemic globally) and 2016 show a high level of concordance between GBD 2016 and UNAIDS, with an average intra-class correlation coefficient of 0·992. For estimates of annual new infections, UNAIDS and GBD follow similar patterns, with UNAIDS having slightly lower estimates for years between 1996 and 2002. The estimates are similar between the two for most of the 2000s, with UNAIDS estimates showing a slightly faster rate of decline in annual new infections from 2008 to 2016 at the global level. GBD 2016 estimates about 1·9 million new infections globally in 2016, while UNAIDS estimates about 1.8 million for the same year.

74Country-specific cancer statistics are published by the International Agency for Research on Cancer for their GLOBOCAN project. Their most recent estimate of incidence for all cancers combined is 14·1 million cases in 2012, compared with GBD estimates of 14·8 million cases in 2010 and 17·2 million cases in 2016. The appendix provides a comparison of incidence by cancer type between GBD and GLOBOCAN (appendix 1, p 255).

75The large outbreak of Zika virus disease in Latin America in 2016 led WHO to declare it a public health emergency of international concern. There were 7·6 million new infections in 2016, 7·4 million of which occurred in Latin America and the Caribbean. Most of the health loss associated with Zika virus disease was from non-fatal outcomes, including fever and malaise during acute infection and congenital anomalies and Guillain-Barré syndrome as longer-term consequences.

76Guinea worm disease was included as a cause of non-fatal health loss because eradication of the disease is achievable in the near future.lthough not having received much attention in general global health policy debates, a remarkable reduction from an estimated 3·5 million cases in the mid-1980s to just 15 cases in 2016 has been achieved. Guinea worm eradication interventions include the distribution of filter cloths and pipe filters for drinking water, treatment of water sources with larvacide, case detection, containment to prevent further transmission, and health education.

77The annual updates of GBD allow incremental improvements to methods. However, several measurement challenges in non-fatal estimation remain. The foremost concern is how best to disentangle measurement error from true variation in disease occurrence. We correct for known bias from non-reference methods or case definitions, but often have to rely on sparse information to make those adjustments. Lack of consensus among researchers on how to measure a disease can make it difficult to define a manageable set of alternative methods for which we can make adjustments. An extreme example is dementia: among 234 data sources we identified 228 different methods of screening, diagnostic tools, and diagnostic criteria. When possible, we use predictive covariates in our non-fatal models. For diseases with a lot of information on what drives differences in disease, covariates help to estimate true variation and minimise residual measurement error. However, many of the leading causes of YLDs lack strong predictors as indicated by our risk assessment, which assigned less than 20% of YLDs from mental, musculoskeletal, sense organ, and neurological disorders to the 79 risks evaluated in the GBD study.

78Second, many non-fatal models continue to rely on sparse data in some regions or super-regions despite the steady increase in data sources. With the addition of some new data sources, this can lead to large variation with previous GBD estimates. Alcohol dependence is a good example. Our estimate of the global prevalence of alcohol dependence increased by 70% after adding nine new data sources for Brazil, 12 for India, seven for sub-Saharan Africa, and four in eastern and central Europe. Until there is better coverage of data, non-fatal disease models will continue to be prone to estimates varying between GBD iterations.

79Third, the adjustments to our hospital admission data for multiple admissions and secondary diagnosis fields reflect all health-care episodes rather than inpatient episodes only. Adjustments derived from a large, though non-representative, source of medical claims data in the USA are assumed to be generalisable. The generalisability of claims data, the validity of assuming diagnoses based on primary diagnosis alone or all diagnostic fields, and the validity of trends in claims data have been challenged.,,,, Also, there may be considerable inter-country variation in how diseases are treated between inpatient and outpatient settings. Our models attempt to adjust for such potential biases by using a covariate on claims and hospital admission data to correct for systematic error, although in a relatively crude manner, this assumes no geographical variation in the error. Gaining access to claims data from other countries, particularly in low-income and middle-income countries, is a crucial endeavour for future GBD iterations.

80Fourth, our estimates of severity distributions for most diseases are based on sparse studies from the literature for diseases with a more established consensus on a measure of severity, or our analyses of three high-income surveys which combine diagnostic information with a general health status measure, the Short Form-12 (SF-12).,, Disease-specific researchers can provide better data with more routine use of a single established measure of severity in surveys and patient populations. Another potential improvement can be expected from countries that are able to link survey data with a general health assessment instrument like SF-12 and administrative diagnostic information. The latter approach has the advantage that it would allow adjustments for comorbidity, which can affect measures aiming to capture the severity of a single disease.

81Fifth, apart from COPD, stroke, and epilepsy, where we have epidemiological data to estimate the geographical variation in the distribution of sequelae, for most diseases we are unable to capture differences in severity that occur due to treatment. As almost all our data sources on severity are from high-income countries, this would underestimate YLDs in low-income and middle-income countries with poorer access to and quality of treatment. If greater geographical information on severity of disease becomes available, we can incorporate a more explicit gradient based on our health-system access and quality index.

82Sixth, we have not yet been able to incorporate dependent comorbidity into our simulation methods. We aim for this to take place in GBD 2017 and will require substantial developmental work to make the computational code more efficient.

83Seventh, most of the uncertainty in our YLD estimates comes from the disability weights. We have not conducted any new surveys since GBD 2013. With further data collection, we may be able to reduce some of the uncertainty by increasing the volume of data and by removing some of the ambiguities in lay descriptions, but large uncertainty is inherent to health state valuations that rely on perceptions by respondents of brief descriptions of complex health problems.

84Eighth, the separate estimation of non-fatal models in DisMod-MR 2.1 for 1990, 1995, 2000, 2006, 2010, and 2016 implies uncertainty of estimates over time is independent. Also, compositional bias can lead to spurious time trends being estimated. DisMod-AT will allow for the more appealing simultaneous estimation over age and time and thus resolve the current limitations. However, as it is a more complex modelling tool, extensive testing and understanding what sensible settings are will need to take place before we can replace the current DisMod-MR approach.

85Ninth, scrutiny of the multitude of results generated by GBD is a challenge. Increasingly, we generate visualisation tools to query results and intermediate steps in our analyses to better vet our results. We share our data and modelling tools with our expanding network of over 2500 collaborators who frequently query data sources and our measurements for particular diseases or locations. Also, by publishing all code and describing the methods used disease-by-disease we encourage scrutiny of every aspect of the GBD study.

86Tenth, as our GBD collaborators have raised, the argument of including Guinea worm disease to document pending eradication can be used to justify the inclusion of poliomyelitis in the next round of GBD. There are also a number of large residual categories, particularly other musculoskeletal disorders and other cardiovascular disease, where we could make more detailed estimates in future rounds of GBD.

87Eleventh, despite adding estimates of the prevalence of several asymptomatic states of disease that are precursors of disease that can be targeted with preventive interventions (such as latent tuberculosis infection, chronic hepatitis, or stage III chronic kidney disease without anaemia), we are not comprehensive in doing so. As diagnostic capacity increases, more asymptomatic precursor states can be identified and, in GBD, we will aim to include these if they can be measured at the population level and become relevant to policy.

88We have introduced a quality grading system for cause of death sources for GBD 2016. We want to introduce a similar grading system of non-fatal data sources for GBD 2017. Important elements to capture in a grading system of non-fatal data sources will include the proportion of data sources using the reference method and case definition for a cause, the proportion of data sources representative of a location, and the amount of detail in reporting of age groups, sex, locations, and time periods covered. Unlike the location-specific grading of cause of death data sources, a grading system of non-fatal data sources will need to be location-specific and cause-specific, making it a much larger endeavour that will likely take several years to complete.

89A prototype of a new version of a new analytical tool, DisMod-AT, is currently being tested. It has the advantage over DisMod-MR 2.1 in that it solves for the differential equations that govern the relationship between incidence, prevalence, remission, and mortality simultaneously over age and time. This is of great relevance for causes that vary substantially over time, and overcomes the assumption of a steady disease state that underlies previous versions of DisMod. We expect for GBD 2017 to run both the current tool and DisMod-AT in parallel for a number of time-sensitive causes with a view of DisMod-AT becoming the tool of choice from GBD 2018 onward.

90We plan to reinstate a web survey facility that will allow derivation of disability weights for new health states; examination of the effect of changes in the wording of lay descriptions of health states on valuations provided by respondents; and creation of health states for combined impairments for which we currently estimate disability weights multiplicatively.

91Diverging trends in YLLs and YLDs raise the importance of non-fatal health outcomes as contributing to the burden of disease. Even if changes in trends for many causes are not large, more attention should be paid to assessing the burden of non-fatal health outcomes, as populations age and countries progress through the epidemiological transition, causing growing numbers of people to be in need of chronic care. More research is needed to identify risks of major disabling diseases, better methods of intervening by prevention, and treatment or improving severity of disability. As countries progress through the epidemiological transition, more attention should be paid to assessing the burden of non-fatal health outcomes, particularly in light of growing numbers of people in need of chronic care as populations age. More research is needed to identify risks of major disabling diseases, better methods of intervening by prevention, and treatment or improving severity of disability.