Global, regional, and national comparative risk assessment of 84 behavioural, environmental and occupational, and metabolic risks or clusters of risks, 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 2016 (GBD 2016) remains the most comprehensive effort to conduct a population-level comparative risk assessment across countries and risks. Other sources of population-level estimates of risk include WHO and UNICEF reports as well as independent scientific publications. Notable differences in methods and definitions produce variation in results, although in several cases there is general agreement in regional or global patterns. The GBD study remains the only peer-reviewed, comprehensive, and annual assessment of risk factor burden by age, sex, cause, and location for a long time series that complies with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER). Added value of this study This study builds upon GBD 2015 and provides several important improvements as well as the quantification of five new risks. The innovations and improvements from last year can be summarised as follows. Across all risk factors, there were 7155 additional data sources, according to the GBD 2016 source counting methods. For diet, we included data for dietary recall, household budget, and food frequency questionnaires. We also incorporated sales data from 170 countries as well as national accounting of food available to populations in a given year. In GBD 2016, we are producing estimates for the following five new risks: smokeless tobacco, low birthweight and short gestation, low birthweight for gestation, short gestation for birthweight, and diet low in legumes. We also extended the high body-mass index (BMI) analysis to include childhood obesity. We have also added 93 new risk-outcome pairs. Major revisions to the estimation of the following risk factors were undertaken for GBD 2016. For second-hand smoke, we changed the estimation method to ensure consistency with the estimates for smoking prevalence. For alcohol, we estimated new relative risks (RRs) for all outcomes, we incorporated more data for exposure and new adjustments for tourism and unrecorded consumption, and we redefined the theoretical minimum risk exposure level (TMREL). For diet, we estimated the disease burden of dietary risks based on the absolute level of intake rather than the intake standardised to 2000 kcal per day. We developed an ensemble model of different parametric distributions to generate better fits to the distributions of continuous risk factors. Mediation evidence was reviewed and updated based on an analysis of ten pooled cohorts. We have expanded the analysis of geographic and temporal trends in risk exposure and burden by development, using the Socio-demographic Index (SDI), and have also explored where countries are in the risk transition. We also improved and modified our decomposition methods so that the results shown are additive and can be aggregated to explain trends in all-cause and cause-specific mortality, as well as trends across age groups. The decomposition analysis has been extended to examine how risk factors have contributed to trends in all-cause mortality by age and sex as well as by cause. Implications of all the available evidence Increasingly detailed understanding of the trends in risk exposure and the RRs for each risk-outcome pair provides insights into both the magnitude of health loss attributable to risks and how modification of risk exposure has contributed to health trends. This analysis shows a mismatch between the potential for risk modification to improve health and the relatively modest role that risk modification has played in the past generation in improving global health.

2Evidence before this study

3The Global Burden of Diseases, Injuries, and Risk Factors Study 2016 (GBD 2016) remains the most comprehensive effort to conduct a population-level comparative risk assessment across countries and risks. Other sources of population-level estimates of risk include WHO and UNICEF reports as well as independent scientific publications. Notable differences in methods and definitions produce variation in results, although in several cases there is general agreement in regional or global patterns. The GBD study remains the only peer-reviewed, comprehensive, and annual assessment of risk factor burden by age, sex, cause, and location for a long time series that complies with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER).

4Added value of this study

5This study builds upon GBD 2015 and provides several important improvements as well as the quantification of five new risks. The innovations and improvements from last year can be summarised as follows. Across all risk factors, there were 7155 additional data sources, according to the GBD 2016 source counting methods. For diet, we included data for dietary recall, household budget, and food frequency questionnaires. We also incorporated sales data from 170 countries as well as national accounting of food available to populations in a given year. In GBD 2016, we are producing estimates for the following five new risks: smokeless tobacco, low birthweight and short gestation, low birthweight for gestation, short gestation for birthweight, and diet low in legumes. We also extended the high body-mass index (BMI) analysis to include childhood obesity. We have also added 93 new risk-outcome pairs. Major revisions to the estimation of the following risk factors were undertaken for GBD 2016. For second-hand smoke, we changed the estimation method to ensure consistency with the estimates for smoking prevalence. For alcohol, we estimated new relative risks (RRs) for all outcomes, we incorporated more data for exposure and new adjustments for tourism and unrecorded consumption, and we redefined the theoretical minimum risk exposure level (TMREL). For diet, we estimated the disease burden of dietary risks based on the absolute level of intake rather than the intake standardised to 2000 kcal per day. We developed an ensemble model of different parametric distributions to generate better fits to the distributions of continuous risk factors. Mediation evidence was reviewed and updated based on an analysis of ten pooled cohorts. We have expanded the analysis of geographic and temporal trends in risk exposure and burden by development, using the Socio-demographic Index (SDI), and have also explored where countries are in the risk transition. We also improved and modified our decomposition methods so that the results shown are additive and can be aggregated to explain trends in all-cause and cause-specific mortality, as well as trends across age groups. The decomposition analysis has been extended to examine how risk factors have contributed to trends in all-cause mortality by age and sex as well as by cause.

6Implications of all the available evidence

7Increasingly detailed understanding of the trends in risk exposure and the RRs for each risk-outcome pair provides insights into both the magnitude of health loss attributable to risks and how modification of risk exposure has contributed to health trends. This analysis shows a mismatch between the potential for risk modification to improve health and the relatively modest role that risk modification has played in the past generation in improving global health.

8A core premise of public health is that prevention can be a powerful instrument for improving human health, one that is often cost-effective and minimises harm to individuals from ill health. The core objectives of prevention include the reduction or modification of exposure to risks including metabolic, behavioural, environmental, and occupational factors. Quantifying risks to health and thus the targets of many public health actions is an essential prerequisite for effective public health. The evidence on the relation between risk exposure and health is constantly evolving: new information about the relative risks (RRs) associated with different risks for different outcomes continues to emerge from cohort studies, randomised trials, and case-control studies. These studies can establish evidence for new risks or risk-outcome pairs or reduce the strength of evidence for existing risks. New data are also regularly collected on the levels of exposure in different populations and in different settings. Regularly updated monitoring of the evidence base on risk factors is crucial for public health and for individual risk modification through primary care and self-management.

9Several studies explore risk-attributable burden for individual risks,, at the global, regional, or national level. Other studies provide assessments of exposure for selected risks. However, the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) comparative risk assessment (CRA) is the only comprehensive and comparable approach to risk factor quantification. The most recent of these assessments was GBD 2015.,, With each cycle of GBD, scientific discussions have emerged on various dimensions of risk quantification that have led to improvements and modifications of GBD. Many of these are focused on the strength of evidence supporting a causal connection for specific risk-outcome pairs, while others relate to measurement challenges.,, Further, new risk factors have been added for important health conditions included in GBD, such as neonatal outcomes and Alzheimer's dementia, which have previously not had associated risk factors. The recent trials on blood pressure control at lower levels of systolic blood pressure, including the Systolic Blood Pressure Intervention Trial (SPRINT) and Heart Outcomes Prevention Evaluation-3 (HOPE-3) trial, have also brought attention to the difference between a population health perspective on the quantification of risks and the clinical question of risk reversibility. The CRA framework provides an important insight into the role of different risks in contributing to levels of population health but does not necessarily provide all the information necessary to guide individual clinical decision making.

10The GBD 2016 CRA includes 84 risk factors and an associated 481 risk-outcome pairs. In addition to new data and updated methods, we have included five new risks in the GBD 2016 CRA. The study was undertaken for 195 countries and territories and provides estimates of exposure and attributable deaths and disability-adjusted life-years (DALYs) for 1990 through to 2016. We explored how risks change with development, measured by the Socio-demographic Index (SDI), and also decomposed changes in deaths and DALYs into the contributions of population ageing, population growth, trends in risk exposure, and all other factors combined. As with previous iterations of GBD, the GBD 2016 CRA results presented here supersede all previously published GBD CRA estimates.

11The CRA conceptual framework was developed by Murray and Lopez, who established a causal web of hierarchically organised risks or causes that contribute to health outcomes (method appendix; appendix 1 p 432), which allows quantification of risks or causes at any level in the framework. In GBD 2016, as in previous iterations of GBD, we evaluated a set of behavioural, environmental, and occupational, and metabolic risks, where risk-outcome pairs were included based on evidence rules (appendix 1 p 344). These risks were organised into five hierarchical levels as described in appendix 1 (p 374). At Level 0, the GBD 2016 provides estimates for all risk factors combined, at Level 1 the GBD 2016 provides estimates for three groups: environmental and occupational, metabolic, and behavioral risk factors. At Level 2, there are 17 risks, at Level 3 there are 50 risks, and at Level 4 there are 67 risks, for a total of 84 risks or clusters of risks. To date, we have not quantified the contribution of other classes of risk factors (appendix 1 p 376); however, using an analysis of the relation between risk exposures and socio-demographic development, measured with the use of SDI, we provide some insights into the potential magnitude of distal social, cultural, and economic factors.

12Two types of risk assessment are possible within the CRA framework: attributable burden and avoidable burden. Attributable burden is the reduction in current disease burden that would have been possible if past population exposure had shifted to an alternative or counterfactual distribution of risk exposure. Avoidable burden is the potential reduction in future disease burden that could be achieved by changing the current distribution of exposure to a counterfactual distribution of exposure. Murray and Lopez identified four types of counterfactual exposure distributions: theoretical, plausible, feasible, and cost-effective minimum risk. In GBD studies, to date and in this study, we focus on attributable burden using the theoretical minimum risk exposure level, which is the distribution of risk comprising the levels of exposure that minimise risk for each individual in the population.

13Overall, this analysis follows the CRA methods used in GBD 2015. The methods described in this study provide a high-level overview of the analytical logic, focusing on areas of notable change from the methods used in GBD 2015, with details provided in appendix 1 (p 10). This study complies with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER) statement (appendix 1 p 377).

14In GBD 2016, locations are arranged as a set of hierarchical categories: seven super-regions, 21 regions nested within the seven super-regions, and 195 countries and territories nested in the 21 regions. Additionally, we present estimates at the subnational level for five countries with a population greater than 200 million in 2016: Brazil, China, India, Indonesia, and the USA. We produced a complete set of age-specific, sex-specific, cause-specific, and location-specific estimates of risk factor exposure and attributable burden for 1990–2016 for all included risk factors.

15Four key components are included in estimation of the burden attributable to a given risk factor: the metric of burden being assessed (number of deaths, years of life lost [YLLs], years lived with disability [YLDs], or DALYs [the sum of YLLs and YLDs]), the exposure levels for a risk factor, the relative risk of a given outcome due to exposure, and the counterfactual level of risk factor exposure. Estimates of attributable DALYs for a risk-outcome pair are equal to DALYs for the outcome multiplied by the population attributable fraction (PAF) for the risk-outcome pair for a given age, sex, location, and year. A similar logic applies for estimation of attributable deaths, YLLs, or YLDs. Risks are categorised on the basis of how exposure was measured: dichotomous, polytomous, or continuous. The PAF represents the proportion of outcome that would be reduced in a given year if the exposure to a risk factor in the past were reduced to the counterfactual level of the theoretical minimum risk exposure level (supplementary results, appendix 2 p 1).

16In this study, as in GBD 2015, we have included risk-outcome pairs that we have assessed as meeting the World Cancer Research Fund grades of convincing or probable evidence (see appendix 1 p 10 for definitions of these grades). Table 1 provides a summary of the evidence supporting a causal relation between a risk and an outcome for each pair included in GBD 2016. For each risk-outcome pair, we used recent systematic reviews to identify independent prospective studies (randomised controlled trials, non-randomised interventions, and cohorts) that evaluated the putative relationship. For risk-outcome pairs with fewer than five prospective studies, we evaluated evidence from case-control studies as well (appendix 1 p 344). Table 1 summarises the evidence using multiple dimensions, which supports our assessment that each included risk-outcome pair meets the criteria of convincing or probable evidence (appendix 1 p 10 contains a justification of the criteria presented to support causality). In this summary of evidence, we have focused on randomised controlled trials and prospective observational studies, along with supporting evidence, like dose–response relationships and biologically plausible mechanisms.

17Information about the data sources, estimation methods, computational tools, and statistical analysis used in the derivation of our estimates are provided in appendix 1 (p 10). The analytical steps for estimation of burden attributable to single or clusters of risk-outcome pairs are summarised in appendix 1 (p 10). Table 2 provides definitions of exposure for each risk factor, the theoretical minimum risk exposure level (TMREL) used, and metrics of data availability. For each risk, we estimated effect size as a function of age and sex and exposure level, mean exposure, the distribution of exposure across individuals, and the TMREL. The approach taken is largely similar to GBD 2015 for each quantity for each risk. Some methodological improvements have been implemented and new data sources incorporated. Appendix 1 (p 34) provides details of each step by risk. Citation information for the data sources used for relative risks are provided in searchable form through an online source tool.

18All point estimates are reported with 95% uncertainty intervals (UIs). UIs include uncertainty from each relevant component, consisting of exposure, relative risks, TMREL, and burden rates. Where percentage change is reported (with 95% UIs), we computed it on the basis of the point estimates being compared.

19In GBD 2015, we produced a summary measure of exposure for each risk, called the summary exposure value (SEV), which is a metric that captures risk-weighted exposure for a population, or risk-weighted prevalence of an exposure. The scale for SEV spans from 0% to 100%, such that an SEV of 0% reflects no risk exposure in a population and 100% indicates that an entire population is exposure to the maximum possible level for that risk. In GBD 2016, we show estimates of SEVs for each risk factor and provide details on how SEVs are computed for categorical and continuous risks in appendix 1 (p 10).

20The most informative data describing the distribution of risk factors within a population come from individual-level data; additional sources of data include reported means and variances. In cases when a risk factor also defines a disease, such as haemoglobin level and anaemia, the prevalence of disease is also frequently reported. To model the distribution of any particular risk factor, we seek a family of probability density functions (PDFs), a fitting method, and a model selection criterion. To make use of the most data describing most populations, we used the method of moments (MoM); the first two empirical moments from a population, the mean and variance, were used to determine the PDF describing the distribution of risk within any population, where exceptions to this rule are justified by context. We used the Kolmogorov-Smirnov test to measure the goodness of fit (GoF), but in some cases, the GoF was based on the prediction error for the prevalence of disease.

21We used an ensemble technique in which a model selection algorithm is used to choose the best model for each risk factor. We drew the initial set of candidate models from commonly used PDF families. We fitted each PDF candidate family to each dataset using the MoM, and used the Kolmogorov-Smirnov test as the measure of GoF. Preliminary analysis showed that the GoF ranking of PDF families varied across datasets for any particular risk factor and that combining the predictions of differently fitted PDF families could dramatically improve the GoF for each dataset. Therefore, we developed a new model for prediction using the ensemble of candidate models, which is a weighted linear combination of all candidate models, {f}, where a set of weights {w} is chosen such that it is the sum of the weights equals to one and the values of the weights were determined by a second GoF criterion with its own validation process. Because of basic differences among risk factors, their distributions, and the risk attribution process, the model selection process was often slightly different for each risk factor. The details can be summarised by (1) the summary statistics for each dataset; (2) a table showing the Kolmogorov-Smirnov statistic for each candidate model and URD; (3) the criterion used for determining the overall GoF; (4) summary results of the validation process; and (5) the weights defining the final ensemble model for each dataset.

22We took several steps to improve the estimation of alcohol use as a risk factor. First, on the exposure side, we added 26 survey series, which contributed 12 195 datapoints in our models. Second, we developed and implemented a method that adjusts total consumption for tourism and unrecorded consumption for each location-year. Third, we calculated the TMREL. We chose TMREL as being the exposure that minimises an individual's risk of suffering burden from any given cause related to alcohol (appendix 1 p 22 for more detail). Fourth, we performed a systematic review of all cohort and case-control studies reporting a RR, hazard ratio, or odds ratio for any risk-outcome pairs studied in GBD 2016 and then modelled a dose-response relationship using DisMod ordinary differential equations (ODE). Fifth, we estimated injury PAFs from cohort studies and adjusted them to account for victims.

23We made several improvements in the process of estimating the burden of disease attributable to dietary risks. To improve the quality and coverage of our dietary estimates, we systematically searched literature for nationally or subnationally representative studies providing information on consumption of each dietary factor. We also made a systematic effort to obtain individual-level data for consumption of dietary factors; re-extracted data from all available sources; and standardised the definition of dietary factors across different sources. To capture recent trends in consumption, we used data on sales of different fresh and packaged foods to inform our estimates. To address the concerns over within-person variation in intake, we estimated usual intake of each dietary factor and used that to estimate the attributable disease burden. To make the current and optimal levels of intake more comparable, we used absolute intake of each dietary factor (rather than intake standardised to 2000 kcal per day). For more detail, see appendix 1 (p 117).

24There were two substantial changes in the estimation of second-hand smoke compared with GBD 2015. First, we estimated the proportion of a population exposed to second-hand smoke using information about household composition and smoking status from household surveys and censuses, rather than using questions that ask directly about exposure to second-hand smoke in surveys. Second, we modelled exposure using spatiotemporal Gaussian process regression (ST-GPR), borrowing strength across sexes and all ages, whereas in GBD 2015 we ran a DisMod model separately by sex and age. Further, we found significant evidence of associations between second-hand smoke exposure and two additional outcomes: breast cancer and diabetes, which were added to the list of risk-outcome pairs for second-hand smoke. More details on the estimation approach are presented in appendix 1 (p 98).

25For the first time in the GBD study, we estimated exposure to and burden attributable to smokeless tobacco, defined as current use of any smokeless tobacco product. RR estimates were derived from prospective cohort studies and case-control studies and vary depending on the type of product used. Based on available evidence, for chewing tobacco RRs were significantly higher than one for oral cancer and oesophageal cancer, while for snus or snuff we did not find sufficient evidence of a RR greater than one for any health outcome. Additional details on the estimation methods and RRs are presented in appendix 1 (p 11, p 181).

26Low birthweight for gestation and short gestation for birthweight are included as new risk factors for GBD 2016. The estimation has been parameterised to be polytomous by 500 g and 2 week categories. Low birthweight and gestational age are highly correlated risks and they are estimated in a completely interdependent manner. For each univariate analysis, identification of TMREL and calculation of PAFs is contingent on the other dimension. In other words, we found the lowest risk birthweight category for each 2 week gestational age band and, correspondingly, the lowest risk gestational age for each 500 g birthweight band. RRs were then estimated for each 500 g per 2 week bin. Exposure for each bin was estimated in three steps. First, we estimated by generating ensemble distribution estimates using modelled mean and categorical prevalence estimates for each of birthweight (mean, % <2500 g) and gestational age (mean, % <37 weeks, % <28 weeks) for each location, year, and sex. Second, we evaluated all microdata where both gestational age and birthweight were available and found a high degree of consistency in the correlation between them. Third, we took the pooled correlation coefficient from step 2 combined with univariate ensemble distributions from step 1 and used a copula linking function to simulate the joint distribution which was then summarised into each 500 g per 2 week category. Joint PAF calculation used a TMREL defined as the lowest overall risk of the entire matrix of birthweight and gestational age (see appendix 1 p 77 for more details).

27In GBD 2016, we updated our approach for estimation of the joint effects of combinations of risk factors (appendix 1 p 23). Using individual-level data from prospective cohort studies, we estimated the proportion of the effect of behavioural risks on cardiometabolic outcomes mediated through metabolic risk factors. We also estimated the proportion of the effect of each metabolic risk factor on cardiometabolic outcomes mediated through other metabolic risks. For each mediation pathway, we only included the mediators for which sufficient evidence existed for their causal relationship with the disease endpoint.

28As in GBD 2015, we undertook a decomposition analysis of changes in DALYs over the time period into four main components, namely, changes in DALYs due to changes in: (1) population growth; (2) population age structure; (3) exposure to all risks for a disease; and (4) all other factors combined, approximated as the risk-deleted death and DALY rates. Risk-deleted rates refers to death and DALY rates that would be observed if we removed all risk factors included in GBD 2016, estimated as DALY rates multiplied by one minus the PAF for the set of risks. We used methods developed by Das Gupta, but as the methods presented there do not result in the decomposition results being linear aggregates over time or risk, we adapted these methods further in GBD 2016. Our decomposition analysis was undertaken for each 5 year time period, at the all-risk level, taking into account risk mediation at the most detailed cause level. The contribution of changes in exposure for the individual risks was scaled to the all-risk effect at the most detailed outcome level. The contribution of risk exposures over longer time periods—eg, 2000–16—or at higher cause aggregates—eg, all cause—were calculated as the linear aggregate of the effect of individual risks at the most detailed cause level and time period.

29We explored how exposure to risks varies across levels of development using the SDI, a composite indicator of development status constructed for GBD 2015 whose components are strongly correlated with health outcomes. It is the geometric mean of 0 to 1 for indices of total fertility rate, mean education for those aged 15 years and older, and lag-distributed income per capita. More details on the estimation of SDI can be found in appendix 1 (p 32).

30The funders of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The authors had full access to all data in the study and had final responsibility for the decision to submit for publication.

31From 1990 to 2016, trends in SEVs varied across the set of risk factors included in GBD 2016. Of note, SEVs decreased by more than 40% for three risks: diet high in transfatty acids (51·3% [95% UI 34·1–70·1]), household air pollution from solid fuels (43·1% [40·7–45·6]), and unsafe sanitation (40·3% [35·5–44·7]; table 3, appendix 2 p 1399). During the same period, SEVs increased by more than 40% for high body-mass index (BMI; 60·2% [45·1–79·1]), diet high in sugar-sweetened beverages (44·7% [36·1–52·7]), occupational exposure to diesel engine exhaust (41·8% [41·3–42·2]), and occupational exposure to trichloroethylene (40·6% [40·2–41·1]).

32Across countries there is substantial variation in risk exposure by level of SDI. Some risk factors, such as high fasting plasma glucose (FPG) and high systolic blood pressure, show similar SEVs across levels of SDI, while others, including household air pollution and unsafe water source, show marked trends with sociodemographic development. Figure 1 shows the relationship between SEVs and SDI for the leading three metabolic, behavioural, and environmental and occupational risk factors and how that changed between 1990 and 2016. Within leading metabolic risks (high BMI, high FPG, and high systolic blood pressure [SBP]), risk-weighted exposure shows an increasing trend with increasing SDI for only high BMI. Overall, the SEV for high BMI has increased during the time period. Looking at the leading three environmental risk factors (ambient air pollution, household air pollution, and unsafe water), figure 2 shows an inverse relationship with SDI for household air pollution and unsafe water, with SEVs approaching zero at high levels of SDI, while the relationship is less consistent with ambient air pollution. Finally, the relationship between SDI and the leading behavioural risk factors is more heterogeneous, with smoking and alcohol use having a positive correlation with SDI, and short gestation for birthweight having a negative correlation with SDI.

33Globally, 59·9% (58·4–61·3) of deaths and 45·2% (43·2–47·3) of DALYs could be attributed to the risk factors assessed in GBD 2016. For deaths, non-communicable diseases (NCDs) show the largest proportion attributable to measured risk factors, at 64·4% (62·6–66·2), with communicable, maternal, neonatal, and nutritional (CMNN) causes at 57·9% (55·4–61·0), and injuries at 25·8% (23·7–27·8). The picture was different for DALYs, however, where we observed that 58·2% (56·4–60·3) of DALYs in CMNN causes are attributable to risk factors, compared with 43·5% (40·7–46·7) in NCDs and 21·0% (19·3–22·7) for injuries. Leading causes of DALYs in CMNN causes, such as diarrhoea and lower respiratory infections (LRI), also showed more than 80% of DALYs can be attributed to risk factors (appendix 2 p 1).

34Within NCDs, three of the leading causes of deaths and DALYs, ischaemic heart disease (IHD; 93·3% [90·3–95·7] of deaths and 94·4% [92·6–95·8] of DALYs), haemorrhagic stroke (88·2% [84·3–91·8] of deaths and 89·5% [87·1–91·6] of DALYS), and chronic obstructive pulmonary disorder (COPD; 76·6% [69·9–82·9] of deaths and 73·8% [67·4–80·2] DALYs) all have high proportions attributable to measured risk factors. Lung cancer, a leading cause of death but not DALYs, also has a large proportion of total deaths and DALYs attributed to measured risk factors (84·1% [78·9–88·3] and 83·2% [78·0–87·6] respectively), while for Alzheimer's disease only 21·4% (11·2–34·0) of total deaths and 22·3% (11·8–35·1) of DALYs can be attributed to measured risk factors. For leading causes of DALYs that do not cause death, such as low back pain and sense organ diseases, less than a third of their total burden can be attributable to measured risk factors (23·0% [20·1–25·9] for low back and neck pain and 13·8% [12·4–15·4] for sense organ diseases). Across all cancers, 42·1% (38·9–45·3) of deaths and 39·8% (36·8–42·8) of DALYs are attributable to measured risk factors; however, there is a very large range within cancers, from cervical cancer at 100% of deaths attributable to risk factors and lung cancer at 84·08% (78·9–88·3) of deaths attributable to risk factors to several cancers at nearly zero, such as brain cancer.

35Across types of risk factors, behavioural risk factors accounted for 32·7% (30·7–34·8) of attributable DALYs, followed by metabolic risk factors at 16·8% (15·7–18·0), and environmental and occupational at 13·1% (12·1–14·2). This pattern was seen in middle SDI, middle-high SDI, and high SDI locations, while in low SDI and low-middle SDI locations environmental risk factors accounted for a larger proportion of attributable DALYs than metabolic risk factors. This is a pattern that has persisted since 1990; notably, however, the importance of metabolic risk factors is growing steadily in low SDI and low-middle SDI locations, while that of environmental and occupational risks has decreased during this time period. More detail can be found in appendix 2 (p 1399).

36Table 4 reports all-cause deaths and DALYs attributable to all risk factors considered here from 2006 to 2016, including detail on attributable deaths and DALYs by risk-outcome pair (appendix 2 p 1865) contains results for every location. Globally, 32·8 million (31·9 million to 33·7 million) deaths were attributable to all risk factors in 2016, a significant increase since 2006 of 2·9% (1·1–4·8); however, age-standardised attributable death rate declined from 2006 to 2016 by 18·7% (17·3–20·0). By contrast, total DALYs attributable to all risks decreased by 8·6% (6·6–10·7) since 2006, and age-standardised DALY rate attributable to all risks decreased by 21·7% (20·0–23·3). Among Level 1 risks, the largest decreases in age-standardised death rates were observed for environmental and occupational risks (24·3% [22·5–26·0]), followed by behavioural risks (21·5% [19·8–23·3]), and metabolic risks (11·9% [9·9–13·5]). Similarly, there were significant decreases in age-standardised DALY rates for all three Level 1 risk factors, although the magnitude of decrease was larger for DALY rates than death rates. In the year 2016, behavioural risk factors accounted for the largest number of deaths (21·8 million [20·5 million to 23·3 million]) and DALYs (781·1 million [737·1 million to 830·1 million]). While there were decreases in both deaths and DALYs attributable to behavioural risk factors since 2006, these decreases were significant for deaths (2·5% [0·1–4·9]) and DALYs (14·3% [11·8–16·6]). There was a significant, 17·9% (15·7–20·6), increase in number of deaths attributable to metabolic risk factors, from 14·8 million deaths (14·0 million to 15·7 million) in 2006 to 17·5 million deaths (16·4 million to 18·5 million) in 2016, with similar increases observed for DALYs. Environmental and occupational risk factors accounted for the fewest number of deaths and DALYs, and there was a significant decline in both measures since 2006.

37Figure 2 shows that in 2016, the leading Level 2 risk factors in terms of attributable DALYs at the global level for both sexes combined were malnutrition (11·5% [10·8–12·3] of DALYs), diet (9·6% [8·2–11·1] of DALYs), high blood pressure (8·9% [7·9–9·9] of DALYs), tobacco (7·4% [6·7–8·3] of DALYs), and air pollution (6·8% [6·1–7·6] of DALYs). The list at this level of aggregation is similar for both sexes combined, with a notable difference being that alcohol and drug use is the fifth-leading risk factor for men, with 7·9% (7·2–8·6) of DALYs, but is at eleventh place for women (2·6% [2·3–2·9] of DALYs). More detail can be found in appendix 2 (p 1399). The patterns of risks vary by development, as seen across the panels of figure 3. At the lowest level of SDI, the leading risk is malnutrition with 25·0% (23·2–26·6) of DALYs, followed by air pollution (8·0% [7·1–9·0] of DALYs), unsafe water, sanitation, and handwashing (7·8% [6·6–9·4] of DALYs), and unsafe sex (4·7% [4·3–5·2] of DALYs). While malnutrition remains the leading risk factor at the low-middle level of SDI, diet (7·8% [6·8–9·0] of DALYs), high systolic blood pressure (7·2% [6·8–8·1] of DALYs), and tobacco use (5·9% [5·3–6·6] of DALYs) get included among the leading five causes as well. At the middle SDI level, diet is among the leading five risks with 12·5% (10·6–14·6) of DALYs while high systolic blood pressure and tobacco follow in importance. At the top three levels of SDI, high BMI increases in importance and makes it to the leading five risks, with 7·2% (4·7–10·0) of DALYs in middle SDI locations, with 9·8% (6·5–13·2) of DALYs in high-middle SDI locations, and 8·7% (5·9–11·7) of DALYs in high SDI locations. The panels in figure 3 clearly show the risk transition across levels of development.

38Figure 3 shows the leading 30 risk factors at Level 3 of the hierarchy and the median change in DALYs between 1990, 2006, and 2016. In terms of rates, among the top ten leading risks in 1990, child growth failure, unsafe sanitation, and unsafe water have experienced the largest declines over the period of 1990–2016. While these three risks have remained in the top 30 in 2016 for men, their ranks have fallen by several places to 9th (child growth failure), 21st (unsafe sanitation), and 16th (unsafe water). For women, their ranks have fallen to 5th (child growth failure), 16th (unsafe sanitation), and 13th (unsafe water). Between 1990 and 2006, median age-standardised DALY rates decreased by 46·7% (42·1–51·1) for men and 49·0% (45·0–53·0) for women, and in the most recent period child growth failure demonstrated further declines by 43·8% (36·9–49·8) for men and 48·7% (42·3–54·6) for women.

39The risk factor of low birthweight for gestation and short gestation for birthweight remains among the leading risks (second position in 1990 for both sexes; third position in 2016 for men and fourth position for women), despite declines in both the number of DALYs and the age-standardised DALY rates since 1990. Smoking is another risk where there has been a consistent decline since 1990 in both SEVs and age-standardised DALY rates, yet it has consistently been ranked among the leading three risk factors for men in DALYs since 1990.

40The trend in unsafe sex coincides with the HIV/AIDS epidemic. Figure 3 shows that unsafe sex experienced large increases between 1990 and 2006, by 198·8% (170·45–228·2) for men and 204·0% (170·0–236·4) for women, resulting in a higher rank in 2006, followed by declines of 43·8% (41·7–45·7) for men and 46·7% (44·1–49·0) for women since 2006 resulting in a lower rank in 2016. On the other hand, drug use follows a different trend, and increased for men by 17·6% (13·0–25·5) between 1990 and 2006 and resulted in a higher rank in 2006, and decreased 5·7% (2·2–9·0) since 2006. Despite declines, drug use rose from the 25th leading risk to the 18th leading risk for men between 1990 and 2016.

41Air pollution, both household air pollution and ambient particulate matter, were among the leading ten risk factors for men and women in 1990 and have remained important in 2016. The median percent change in age-standardised DALY rates showed important declines in both time periods for men and women. Specifically, in the most recent time period household air pollution declined by 38·3% (35·3–41·4) for men and 41·1% (37·8–44·2) for women, and ambient air pollution decreased by 14·2% (11·5–17·1) for men and 21·3% (17·8–24·5) for women, in terms of median age-standardised DALY rates.

42The metabolic risk factors have increased in both rank and in the absolute number of DALYs between 1990 and 2016 for both men and women. High blood pressure was the fourth-leading risk factor for both men and women in 1990 and had risen to be the second leading risk factor for men and the leading risk factor for women by 2016. In terms of the number of DALYs, men showed an increase of 16·2% (13·1–19·4) since 2006, while for women the increase was less steep at 7·7% (4·5–11·7). In terms of the median change in age-standardised DALY rates since 2006, both sexes showed a decline, 10·5% (8·2–12·7) for men and 16·8% (13·7–19·3) for women. Other leading metabolic risk factors, including high BMI, high FPG, and high total cholesterol, exhibited similar trends to high blood pressure over this time period. All four of these metabolic risk factors are within the leading ten risk factors globally for men and women in 2016.

43Among the leading risk factors in terms of DALYs, high BMI and high FPG have the fastest increases in SEVs with annualised rates of change of 1·7% (1·5–1·9) and 0·9% (0·6–1·3), respectively, since 1990 (figure 4). On the other hand, other leading risk factors in 2016 such as smoking and household air pollution exhibited significant and fast declines in SEVs, with a −1·3% (–1·6 to −1·1) annualised rate of change for smoking and −2·3% (–2·5 to −2·2) for household air pollution between 1990 and 2016 (figure 4).

44Figure 5 shows the relative contributions to changes in deaths and DALYs of important drivers grouped into four mutually exclusive categories: population growth, population ageing, trends in exposure to all risk factors measured in GBD 2016, and all other factors combined. Globally, trends in exposure to all risk factors combined would have led to a decrease of deaths by 9·3% (6·9–11·6) and DALYs by 10·8% (8·3–13·1). Risk factors play a larger part in CMNN causes, where trends in exposure to risks would have resulted in a decrease of deaths by 14·9% (12·4–17·1) and DALYs by 15·0% (12·7–17·6). Overall, population ageing and population growth are both driving deaths and DALYs to increase significantly. At the global level, across all causes, population growth alone would have resulted in 12·4% (10·1–14·9) more deaths and 12·4% (10·1–14·9) more DALYs, while population ageing would have contributed 14·9% (12·7–17·5) more deaths and 12·4% (10·1–14·9) more DALYs. The contribution of population ageing in NCDs is noteworthy as it is the largest driver of trends in NCDs, and accounts for 19·5% (17·3–22·0) more deaths and 14·0% (11·6–16·3) more DALYs between 2006 and 2016. The residual category, which includes improvements in treatment along with other factors, accounts for a decrease of 15·3% (12·9–17·7) for deaths and 16·5% (14·1–18·8) for DALYs across all causes and is particularly large for CMNN causes accounting for a 30·0% (27·5–32·4) decline in deaths and a 26·8% (24·4–29·2) decline in DALYs since 2006.

45Figure 6 shows the contribution of these drivers across age groups for DALYs. Across age groups, the contributions of the four drivers differ greatly. Changes in risk exposure have played a major part in the declines in DALYs younger than 5 years, accounting for 26·7% (24·3–29·7) of the trend in DALYs in the post-neonatal period and 27·3% (24·9–29·7) among ages 1–4 years. Trends in risks account for a decline of 8·7% (6·3–11·1) of DALYs in older children (ages 5–9 years) and 9·0% (6·5–11·4) of DALYs in young adolescents (ages 10–14 years). As expected, population ageing is a more significant driver among older age groups, accounting for up to 51·4% (49·1–53·9) of the change in DALYs since 2006 among the age group 90–94 years. Finally, the proportion of the change in DALYs that is due to all other factors—ie, not explained by these three major drivers—also shows large variation across age groups, ranging from a decrease of 3·5% (1·1–6·0) in the age group 15–19 years to a decrease of 28·2% (25·8–30·5) in the age group 1–4 years.

46In 2016, for Level 3 risks factors, more DALYs were attributable to increased SBP than any other risk factor. Increased SBP was the second leading risk factor for men and leading risk factor for women globally, accounting for 89·9 million (80·9 million to 98·2 million) DALYs among women and 124·1 million (111·2 million to 138·0 million) DALYs among men. IHD was the largest source of DALYs attributable to increased SBP, followed by haemorrhagic stroke and ischaemic stroke. Since 1990, the SEV for increased SBP rose for men (22·9 [21·5–24·6] in 1990 to 24·6 [23·0–26·6] in 2016, a 7·5% increase [7·0–8·0]), and increased for women (24·2 [22·7–25·8] in 1990 to 24·2 [22·7–25·8] in 2016, a 0·7% increase [0·2–1·2]).

47In 2016, 7·1 million (6·5 million to 7·8 million) deaths and 177·3 million (162·3 million to 194·3 million) DALYs were attributable to tobacco, most of which is attributable to smoking tobacco. Smoking-attributable deaths have increased by 20·1% (15·3–25·2) since 1990, with most deaths occurring in China, India, the USA, and Russia. Smoking is the second-leading risk factor for men for deaths and leading for DALYs, accounting for 16·3% (14·6–17·9) of deaths and 9·5% (8·5–10·7) of DALYs, and the sixth for women for deaths and ninth for DALYs, with 5·8% (5·0–6·7) of deaths and 2·9% (2·5–2·94) of DALYs. In 2016, there were 177·3 million (162·3 million to 194·3 million) smoking-attributable DALYs globally. Overall, in 2016 chronic respiratory diseases (30·3% [25·2–36·0]), neoplasms (19·2% [16·0–22·8]), and cardiovascular diseases (18·0% [16·0–20·0] were the three leading causes of smoking-attributable age-standardised DALYs across both sexes. For women, the leading cause of DALYs was COPD, whereas the leading cause for men was IHD.

48Second-hand smoke exposure is highest in eastern Asia and Oceania and higher among women and children compared with men. The distribution of DALYs attributable to second-hand smoke exposure is bimodal, with peaks in the post-neonatal period and again in older age groups. Globally 0·9 million (0·7 million to 1·1 million) deaths were attributable to second-hand smoke exposure, of which 56 340 (28 951–89 043) occurred among children younger than age 10 years.

49In estimating the burden attributable to smokeless tobacco, we found that the risk varies by the toxicity of the type used; there is sufficient evidence that chewing tobacco and other products of similar toxicity cause excess risk of oral and oesophageal cancer while, at this time, existing evidence does not support attributing burden to snus or similar smokeless tobacco products. Globally, smoking tobacco causes far more burden than smokeless tobacco; nonetheless, smokeless tobacco is an important risk factor for oral and oesophageal cancer in India, where more than half of the 32 141 (24 930–39 243) global deaths attributable to smokeless tobacco occur.

50Low birthweight and short gestation, new risk factors in GBD 2016, were the third-ranked Level 3 risk factor globally for all-ages DALYs in 2016, which reflects a 61·6% (59·3–64·0) decrease in all-ages DALY rates from 5112·8 (4934·2–5389·6) DALYs per 100 000 in 1990 to 1960·8 (1862·0–2060·3) DALYs in 2016. In 1990, this risk factor was the second-ranked Level 3 risk factor globally for all-age DALYs; most of the decrease from 1990 to 2016 is due to a lower mortality burden in the causes attributable to low birthweight and short gestation rather than changes in exposure itself. Increasing SDI was associated with decreasing exposure, but the exposure gradient between SDI quintiles was not as large as the differential between high and low SDI in attributable burden. Exposure was highest in South Asia, eastern sub-Saharan Africa, and parts of the western Sahel zone, while attributable burden was highest in South Asia and parts of the western Sahel zone. The trend in exposure to low birthweight for gestation decreased at the global level from 1990 to 2016, reflective of the overall decrease in DALYs burden during the same time period. The biggest improvements were seen in Colombia, Brunei, and Zimbabwe, with broad improvements also seen across much of eastern sub-Saharan Africa.

51In 2016, high FPG was the third-leading risk factor for deaths and the fourth-leading risk factor for DALYs globally among Level 3 risk factors, accounting for more than 5·6 million deaths (4·5 million to 7·0 million) and 144·1 million DALYs (119·9 million to 171·6 million). Since 1990, the age-standardised percent of deaths and DALYs attributable to high FPG has increased globally from 7·8% (6·0–10·1) to 10·5% (8·3–13·1) and 4·4% (3·7–5·3) to 6·2% (5·3–7·3), respectively. Diabetes was the largest source of DALYs attributable to increased FPG, followed by ischaemic heart disease and chronic kidney disease. We re-evaluated epidemiological evidence supporting the causal relationship between high FPG and disease endpoints and found sufficient evidence to include ten new outcomes for high FPG. These new outcomes included glaucoma, cataracts, dementia, liver cancer, lung cancer, ovarian cancer, breast cancer, bladder cancer, colorectal cancer, and pancreatic cancer. The new outcomes together contributed to 174 352 (37 297–388 039) additional deaths and 2·6 million (0·6 million to 5·7 million) additional DALYs beyond the causes that were included in GBD 2015.

52In 2016, BMI was the fifth-ranked Level 3 risk factor for death globally, accounting for more than 4·5 million (2·9 million to 6·4 million) deaths and 135·4 million (88·6 million to 187·4 million) DALYs. Among Level 3 risk factors with more than 10 million attributable DALYs, high BMI had the fastest annualised rate of increase in SEV since 1990 (appendix 2 p 1399). Despite this significant increase in risk exposure, increases in attributable burden were attenuated by significant decreases in risk-deleted DALY rates, mainly due to reductions in cardiovascular disease mortality rates. We find that the burden attributable to high BMI increases with increasing development, with the lowest rates of disease attributable to high BMI found in sub-Saharan Africa, yet development is not the only predictor. We conducted a systematic search of health outcomes caused by excess bodyweight and added eight new causes for GBD 2016, which together contributed to 442 750 (191 407–796 350) additional deaths beyond the causes that were included in GBD 2015. Additionally, we included childhood overweight and childhood obesity as new risk factors, allowing us to better capture the health effects of excess bodyweight across the life course. Within the CRA framework, the only childhood overweight and obesity outcome eligible for inclusion was asthma. We found that 10·4% (3·1–21·2) of asthma can be attributed to childhood excess bodyweight globally, a total of 1128 (311–2354) deaths and 642 532·1 (180 916·3 to 1 456 342·7) DALYs. While childhood burden is much smaller compared with adult burden, estimating exposure for children is crucially important in view of the well described effects of childhood overweight and obesity on adult health outcomes.

53Air pollution was ranked sixth in terms of attributable DALYs in 2016. We found that 7·5% (6·6–8·4) of deaths globally were attributable to ambient air pollution in 2016 (4·1 million [3·6 million to 4·6 million] deaths, 1·3 million [1·1 million to 1·5 million] in South Asia). Countries with notably high levels of attributable deaths include China (11·1% [9·7–12·7] of all deaths attributable to ambient particulate matter) and India (10·6% [9·2–11·9] of all deaths). The diseases with the largest proportion of burden attributable to air pollution are LRI and COPD; ambient particulate matter is responsible for 27·5% (21·4–34·4) of all LRI and 26·8% (16·1–38·6) of COPD deaths and 33·3% (26·3–40·5) of LRI deaths in children younger than 5 years. In terms of overall ranking, ambient particulate matter has increased from seventh in 1990 with 115·2 million (99·1 million to 132·9 million) DALYs to sixth in 2016 with 105·7 million (94·2 million to 117·8 million) DALYs. For deaths, it is among the top ten ranked risk factors in 195 countries and territories, including India and China, where it was in third and fourth place, respectively. Also of note is that updated satellite data indicate increased ambient air pollution in 2015–16 in West Africa that is driven by wind-blown dust from the Sahara. This effect has profound effect on disease burden in this region, as intense particulate matter with an aerodynamic diameter smaller than 2·5 μm (PM2·5) events affect Africa's densest region.

54Globally, alcohol is estimated to be the seventh-leading risk factor in 2016 in terms of DALYs. In the same year, alcohol use was estimated to have caused 99·2 million DALYs (88·3 million to 111·2 million), accounting for 4·2% (3·7–4·6) of total DALYs. This is a larger share of total burden than previously reported, driven primarily by changes made to both the exposure and RR models. This burden is distributed unequally among the sexes and regions. When decomposed by sex, alcohol use accounts for 6·2% (5·6–6·9) of total DALYs among men and 1·7% (1·4–2·0) of total DALYs among women. When decomposed by region in 2016, alcohol use accounts for 13·9% (11·5–16·8) of age-standardised DALYs in eastern Europe, 4·0% (3·4–4·6) of age-standardised DALYs in Southeast Asia, but only 0·8% (0·6–1·0) of age-standardised DALYs in the Middle East. Alcohol use attributable DALYs have also increased by more than 25% over the years 1990–2016, driven primarily by increased consumption in South Asia, Southeast Asia, and Central Asia, among both men and women. Globally, alcohol use exposure has increased by 15·2% (8·7–22·6) over that time frame among men and decreased by 3·2% (–9·1 to 3·1) among women. However, the largest increases in exposure have been in countries in the low-middle quintile of SDI. Globally, alcohol use is the leading risk factor in DALYS between the ages of 15 years and 49 years in 2016. However, unlike tobacco or drugs, governments have been discouraged from efforts to limit alcohol's availability by trade agreements and disputes. Given alcohol's health burden within these age groups, an increased focus on alcohol control policies is needed to effectively address this risk factor.

55It is worth noting some key results for dietary risks as well. In 2016, suboptimal diet was the second-leading risk factor for deaths and DALYs globally, accounting for 18·8% (16·0–21·7) of all deaths and 9·6% (8·2–11·1) of all DALYs. Comparing men and women, suboptimal diet accounts for the greatest percentage of total deaths in men (19·0% [16·3–21·8]) and the second largest in women (18·6% [15·7–21·7]). Meanwhile, suboptimal diet accounts for the second-largest percent of total DALYs in both men (10·6% [9·1–12·2]) and women (8·4% [7·0–9·9]). More than 50% of deaths (51·5% [44·2–59·2]) and DALYs (54·1% [47·1–61·5]) attributable to suboptimal diet were due to cardiovascular diseases. Among the individual dietary risks, a diet low in whole grains accounted for the largest number of deaths (4·6% [3·0–6·4]), followed by a diet low in fruits (4·3% [2·7–6·3]) and a diet high in sodium (4·2% [1·2–8·3]). Leading dietary risks for DALYs were low intakes of whole grains (2·6% [1·8–3·6]), fruits (2·6% [1·6–3·7]), and nuts and seeds (2·1% [1·4–2·8]). The greatest increase in attributable deaths and DALYs between 1990 and 2016 occurred for a diet high in red meat, followed by a diet high in sugar-sweetened beverages and a diet low in milk, respectively.

56Based on the analysis of 22717 sources, we estimated disease burden attributable to 84 metabolic, environmental, occupational, and behavioural risk factors or clusters of risks from 1990 to 2016 in 195 countries and territories. In 2016, all risks combined contributed to 59·9% (58·4–61·3) of deaths and 45·2% (43·2–47·3) of DALYs worldwide, compared with 60·3% (59·0–61·6) of deaths and 49·6% (47·6–51·7) of DALYs in 1990. The role of changes in risk factors in explaining changes in deaths and DALYs varies considerably across causes and ages, with the largest effects noted in children due to infectious diseases. Since 1990, exposure increased significantly for 30 risks, did not change significantly for four risks, and decreased significantly for 31 risks. The risks with the highest increases in SEVs include high body-mass index, occupational exposure to diesel engine exhaust, and occupational exposure to trichloroethylene, while the risks with the largest decreases in exposure are diet high in transfatty acids, household air pollution from solid fuels, and unsafe sanitation.

57We found substantial heterogeneity across countries in the leading risk factors. Some notable patterns are the role of unsafe sexual practices as a driver of the HIV epidemic in Eastern and Southern Africa and the role of alcohol consumption in Eastern Europe and Central Asia. There are also marked spatial patterns for other risks such as high BMI in Central America, North Africa and the Middle East, and Oceania. Interpreting spatial patterns needs to take into account the fact that some risks have a strong relationship with socioeconomic development. Several environmental and behavioural risks, including water, sanitation, handwashing, household air pollution, and childhood growth failure decline profoundly with development. Another cluster of risks tends to increase with socioeconomic development, including high BMI, high SBP, red meat consumption, sugar-sweetened beverages, alcohol, and high FPG.

58Many factors should determine government priorities for action including the size of the problem, inequalities related to the problem, likely future trends, the availability of effective policy options, and the opportunity cost of tackling a particular problem. In this analysis, we provided information about the size of the problem, trends in exposure in the last 27 years, and the range of exposure at given levels of socioeconomic development. Problems that are large, increasing, and variable across countries at the same level of development likely warrant particular policy attention. Our analysis showed that components of diet, obesity, FPG, and SBP are the most prominent global risks fulfilling these criteria. Because of the strong interrelationships between these risks, the true driver of this cluster is likely diet, the risk in BMI, or both, with knock-on consequences for FPG and SBP. The rise of obesity and the associated increases in FPG and SBP warrant considerable global policy attention. Other major risks that should continue to receive attention—even intensified attention in some locations—such as smoking, are nevertheless declining at the global level. The unique combination of large current effect and increasing exposure puts obesity in a special category of risks. Obesity is likely to not only influence future population health in many locations, but will have considerable financial implications for health systems, given what we know about treatment costs for the associated diseases. Since important drivers of obesity such as physical activity and diet patterns are adopted in childhood and adolescence, more work is needed to proactively address the adoption of these risks in these younger age groups.

59For the first time, we assess the contribution of changes of risk exposures to the overall global trend for deaths and DALYs; for example, in the past 10 years, changes in all risk exposures contributed to an 10·8% (8·3–13·1) decline in DALYs, while other factors contributed to a 16·5% (14·1–18·8) decrease in DALYs. More detailed assessments show large declines in CMNN causes and increases in injuries and non-communicable DALYs. In each case, the contribution of other factors was substantially larger than the contribution of risk reduction. Our findings of the relatively small contribution of risk reduction to the declines in NCDs are not at odds with published studies for the UK and the USA,,, because we are reporting at the global level; our results at the national level suggest a larger role for risk reduction in some high-SDI locations. These observations lead to two directions for further analysis. First, what is the explanation for the declines driven by other factors? Some of this effect might be social policy working through various causal channels, and some is likely due to improvements in access to high-quality health care. This is particularly true for conditions such as selected cancers, ischaemic heart disease, cerebrovascular disease, chronic kidney diseases, HIV/AIDS, tuberculosis, and maternal mortality, for which health care is known to have large effects. Second, in view of the enormous potential of risk reduction to change health outcomes as documented in this and many other studies, why has progress on many risks been comparatively slow? For example, even though global tobacco consumption is declining in terms of rates, the pace of decline has been remarkably slow on average, despite more than 50 years of good evidence on the harms of tobacco. The relatively poor track record for global risk reduction might in part reflect the low rate of investment in risk reduction compared with curative health care. It might also reflect the continuing challenge of changing many risky behaviours. Relatively little funding for research on changing behaviours compared with new diagnostics and therapeutics might also be part of the explanation of the prevention paradox., Changing behavioural risks could also require more than government action; harnessing the private sector to facilitate behavioural change might also be crucial.

60Increased SBP remains the leading global risk at Level 3 in the GBD risk hierarchy. Highly effective interventions exist to manage blood pressure at the primary care level, as do a range of public health interventions, so it is quite remarkable that global exposure to increased SBP is increasing. Part of this increase might be tied to the global rise in high BMI, but the increase in SBP represents significant missed opportunity for the world's health systems. In 54 countries high SBP is actually declining, while its increase in China is now well documented in a series of population-based surveys.,, Tackling rising SBP is a global concern, but this is particularly important in those locations where rates are increasing. In view of the effect of the risk and the large array of available, effective interventions, health systems and the global health community need to mobilise increased resources and policy attention to tackle this problem. It might be necessary to design a variety of public policies including food reformulation to reduce sodium content and efforts to incentivise primary care providers to give priority to the management of SBP.,,

61In moving toward developing a comprehensive picture of tobacco use globally, in GBD 2016, we have for the first time included smokeless tobacco use as a risk factor. While the burden of smokeless tobacco is minimal in the majority of countries, it is of huge importance in south Asia, where the highest risk-weighted exposure is observed in Bangladesh (risk-weighted exposure of 0·75 [0·61–0·87]), Bhutan (0·53 [0·44–0·62]), Myanmar (0·50 [0·42–0·59]), Nepal (0·50 [0·42–0·58]), and India (0·45 [0·43–0·47]). In these countries more women use smokeless tobacco products than smoked tobacco products, and we find that use of any tobacco products, smoked or smokeless, continuously increases with age, a regional age pattern that differs from the global and male regional age pattern. The combination of high exposure and large population results in a majority of global deaths attributable to smokeless tobacco in 2016 occurring in India, where it is also the leading risk factor for oral cancer.

62In GBD 2016, we also improved the estimation of burden attributable to second-hand smoke. At the global level, while the burden of second-hand smoke remains substantial, exposure to second-hand smoke has been declining significantly at an annualised rate of change of 1·9% (1·5–2·4). These reductions are likely attributable to a wide range of public health measures to control tobacco, which have accelerated in a large number of countries since the implementation of the Framework Convention on Tobacco Control (FCTC).

63Progress combatting the tobacco epidemic has resulted in global declines in prevalence of tobacco use and second-hand smoke exposure, yet the number of deaths and DALYs attributable to tobacco has increased since 1990. Increases in burden were driven by a combination of population growth and population ageing, along with persistently high smoking prevalence in some of the most populous countries of the world. Taken together, we can expect the burden of tobacco to remain high in years to come, unless the rate of progress is significantly accelerated. Many countries with persistently high levels of daily smoking recorded marginal progress in the past decade, and smoking remains a leading risk factor in most countries. The fact that tobacco use patterns diverge by location, level of development, and sex highlights the need for more tailored approaches to change smoking behaviours in the future. Particularly worrisome are the trends among young men and women. For example, in Indonesia, a country that has not yet ratified the FCTC, more than half of men aged 20–24 years are daily smokers. Understanding what works—and what does not—for tobacco control across contexts and within subpopulations (ie, men and women, younger and older individuals, various socioeconomic groups) is of growing priority. To significantly and permanently change the toll of tobacco, a renewed and sustained focus is needed on comprehensive tobacco control policies around the world.

64The global increase in FPG is likely tied to the increase in BMI. While exposure is increasing, age-standardised attributable mortaliy rate is not; a related pattern is that the prevalence of diabetes is increasing, but deaths from diabetes have been declining, likely because clinical management of the macrovascular complications of diabetes has improved in many (but not all) locations. Prevention trials show that with intensive resources devoted to weight loss and physical activity, reductions in FPG can be achieved; however, these interventions have not been implemented at a national scale and adherence in the long run is challenging. Systematic efforts to screen for high FPG implemented in some countries may increase awareness and action in more patients but can be resource-intensive. Clinical interventions to reduce FPG can be effective, although there are more recent debates on the appropriate targets for treatment in some cases. With FPG increasing in many settings, it is difficult to determine the population effect of treatment of blood sugar on population FPG. FPG remains one of the risk factors that is most likely influenced at the primary health-care level, emphasising the role of universal coverage for primary care in a multipronged response to this increasing problem.

65One of the most alarming risks in the analysis is increased BMI, because its burden is large and increasing, and it is prevalent across all levels of SDI., The potential drivers of this global epidemic include changes in food industries and systems, which increase availability, accessibility, and affordability of energy-dense foods, along with intense marketing of such foods, as well as reduced opportunities for physical activity. A range of interventions have been proposed to reduce obesity, including restricting the advertisement of unhealthy foods to children, improving school meals, taxation of sugar-sweetened beverages, and taxation to reduce consumption of other unhealthy foods and subsidies to increase intake of healthy foods, and using supply-chain incentives to increase production of healthy foods. However, the evidence base that many of these interventions can affect trends in obesity at scale is currently weak. What we know without a doubt is that obesity rates continue to increase in almost all locations. Low-SDI and middle-SDI countries generally have little financial resources for nutrition programs and mostly rely on external donors whose programmes often preferentially target undernutrition. The increase in exposure to high BMI is greater than the increase in attributable burden largely because cardiovascular disease death rates continue to decline because of other changes, particularly improvements in treatment and declines in smoking and high cholesterol. Proposed policies, even if fully implemented, are unlikely to rapidly reduce the prevalence of obesity. While not a solution to the rise of overweight and obesity, clinical interventions that control high SBP, cholesterol, and FPG (the major risk factors for cardiovascular disease) can be used to mitigate some of the cardiovascular ill-effects. Expanded use of such interventions among obese people could effectively reduce the disease burden of high BMI. Sustained progress, however, will require policies that effectively control weight in childhood and in young and middle-aged adults.

66In GBD 2016, poor dietary habits were the second leading risk factor at Level 2 of the hierarchy for mortality globally, accounting for nearly one in every five deaths. The overall burden of dietary risks at the global level was 14·8% (11·7–18·5) lower than in GBD 2015. Additionally, important differences were observed in the attributable burden and the ranking of individual dietary risks. Multiple factors have contributed to these differences, including using more data sources, as well as improving the method of estimation of the mean and distribution of intake for each dietary factor. In GBD 2016, for the first time, we used sales data to inform our estimates of consumption for most dietary factors. Using sales data, in addition to improving our overall data coverage, allowed us to capture recent trends in consumption. This was particularly important for specific dietary factors, such as sugar-sweetened beverages, which have been the target of dietary policies in several countries.,,,,, Additionally, to improve the consistency of definitions of dietary risk factors across surveys, we made a systematic effort to obtain and re-extract individual-level data from nutrition surveys. To make the current level of intake and optimal level of intake more comparable, we used the absolute level of intake (rather than the intake standardised to 2000 kcal per day) as the primary exposure in GBD 2016. We also corrected our estimated daily intake of each individual dietary factor for within-person variation and characterised the usual intake at the population level. Finally, given the differences in the health effects and patterns of intake for legumes and vegetables, we estimated the burden of disease attributable to low intake of legumes and low intake of vegetables separately.

67The decade of 2016–25 has been declared as the Decade of Action on Nutrition by the United Nations General Assembly. GBD 2016 provides a comprehensive picture of various forms of malnutrition (ie, undernutrition, overweight or obesity, and poor dietary habits) across all countries at the start of the Decade of Action on Nutrition and can inform priorities for evidence-based interventions in each country. GBD also provides an independent avenue to annually monitor the progress of countries toward achieving their nutrition-related goals in a comparable and consistent manner. Our results show that among all forms of malnutrition, poor dietary habits, particularly low intake of healthy foods, is the leading risk factor for mortality. This finding has important implications for national governments and international organisations aiming at ending malnutrition over the next decade, highlighting the need for comprehensive food system interventions to promote the production, distribution, and consumption of healthy foods across nations.

68Low birthweight and short gestation have been added for GBD 2016; they are the third-leading global risk at Level 3 in the GBD risk hierarchy. Improvements in burden attributable to low birthweight and short gestation have been largely driven by other factors influencing neonatal death rates, given that exposure to low birthweight and short gestation have not improved much over the past 27 years. Little progress in exposure suggests suboptimal coverage of interventions and programmes that can prevent low birthweight and short gestation. These include women-centred services for optimising nutrition (including minimising obesity), infection control, smoking cessation, and preventive care for pregnant women or those contemplating pregnancy.,, Efforts should also focus on maximising the quality of antenatal care services to identify and appropriately manage at-risk and high-risk pregnancies, including avoidance of provider-initiated preterm delivery. If evidence-based interventions are employed, it should be possible even in resource-limited settings to shift the risk curve for those babies who will be born early, small, or both, despite best efforts. Before birth, this includes potentially antenatal steroid administration to promote lung development; at birth, this requires presence of adequately trained and equipped neonatal resuscitation services;, post-delivery, it should include physicians with neonatal specialisation and availability of supportive equipment such as continuous positive airway pressure. Facility-based infection control measures are crucial to prevent nosocomial transmission, as such events are highly lethal in low birthweight or short gestation neonates. The inclusion of this risk for a major cause of DALYs—namely, neonatal mortality—also expands the share of overall burden that can be attributed to risks in general. More work remains, however, to understand the relationship between low birthweight and short gestation and childhood growth failure after 1 month. Our analysis to date may actually underestimate the importance of this risk if the share of childhood growth failure that can be traced to low birthweight and gestational age is fully established.

69Globally, alcohol is estimated to be the seventh-leading risk factor in 2016 in both DALYs (4·2% [3·7–4·6]) and deaths (5·2% [4·4–6·0]). Previous studies have noted the possibility that the preventive effects of alcohol might have been overstated due to selection bias and choice of the reference population.,,, Our findings lend further credence to these hypotheses; with the exception of IHD, our results show either a minor or non-significant preventive effect for causes previously estimated to have large preventive effects. Further, our analysis noted a much larger risk of neoplasms due to alcohol use than previously reported. Combined with our new data for alcohol use exposure, alcohol use is ranked as one of the leading risk factors, surpassing cholesterol as a share of total DALYs, comparied with previous iterations of GBD.,,

70In GBD 2016 we have introduced a more accurate method for developing the distributions of exposure for many risk factors. Our work on distributions and the shift to ensemble distributions shows that the assessment of attributable burden is sensitive to distributional assumptions. Given that a number of risks, such as BMI, SBP, cholesterol, and FPG, rise exponentially as a function of exposure, the estimation of the tail of the distribution has an important effect on the results. The ensemble modelling approach can provide more accurate estimation of the full distribution, including the tails of the distribution. In general, we believe that the assessment of the distribution of the risks deserves more careful attention in future research.

71The GBD study is the most comprehensive effort to conduct a population-level CRA across countries and risks. Differences between GBD 2016 estimates and other global estimates are generally related to approaches to data processing, access to data sources, and analysis decisions. For several risks, including smoking, ambient ozone pollution, household air pollution from solid fuels, lead exposure, intimate partner violence, unsafe water source, and breastfeeding, GBD estimates were lower than published WHO estimates.,,, These discrepancies can be attributed to different definitions, methodological decisions, granularity, and input data. For some findings, annual estimates might disagree, but regional patterns were consistent between WHO and GBD. UNICEF produces estimates for child stunting that are lower than GBD estimates with some disagreement where progress has been made globally. There is more consistency in estimates between UNICEF and GBD for child wasting and child underweight. GBD estimates for the prevalence of low birthweight and short gestation are slightly lower when compared with WHO estimates, but show similar geographical patterns. Scientific literature reveals similar results to GBD for impaired kidney function and low birthweight and short gestation;, research analysing ambient air pollution differed from GBD estimates due to older methods and less granularity. Research published on iron-deficiency anaemia differs from GBD in methods and definitions, resulting in generally higher GBD estimates. GBD estimates were much lower than published research on occupational estimates,,, largely due to different cause-outcome pairs and GBD's application of the CRA approach (see appendix 1 p 10).

72Interpretation of our results and prioritisation at the national level might also need to take into account the variable strength of evidence supporting the causal connection for each risk-outcome pair. In GBD 2016, we have continued to use the World Cancer Research Fund criteria of convincing or probable evidence to select risk-outcome pairs for inclusion. Some aspects of these definitions are subjective. Not all researchers would agree on the interpretation of the available evidence as fulfilling these criteria. For example, there are six studies on non-exclusive breastfeeding and LRI; there are two studies on discontinued breastfeeding and diarrhoeal diseases. We have sought to quantify the number of studies of different kinds that are available to support these judgements in table 1, but not all studies support causality to the same extent. Randomised trials, if well conducted, provide the strongest evidence of causality, because they are likely not affected by confounding. But even randomised trials can have biases when there are missing observations, as is often the case. Randomised trials are also not feasible in many cases, or if feasible, not representative for many risks, including environmental risks. Cohort studies can provide compelling evidence, but many cohorts do not adequately control for socioeconomic confounders and can suffer from many other issues related to the quality of exposure measurement or outcome ascertainment. To go beyond, the quantification of the number of studies of each type we have provided here will necessitate a deeper analysis of the potential limitations of all 2579 studies used across the risk-outcome pairs. In future work, we plan to evaluate the quality of each of these studies with a standardised approach and work toward an overall evidence summary. There is also a more fundamental philosophical question about the presentation of risk information. Should decision makers only pay attention to risk factor quantification for those risks supported by the strongest causal evidence such as randomised trials? Or do notions such as the precautionary principle suggest that we should pay attention to risk quantification even for risk-outcome pairs where the evidence is less definitive.,, Because the social response to risks, particularly risks that might be emerging, can take considerable time, ignoring risks for which the evidence is less definitive might actually lead to worse outcomes for society. Conversely, in a world of scarce political and financial resources, devoting attention to risks that might turn out not to be causal might lead to less action on more well documented risks.

73As part of future iterations of GBD, we plan to quantify the burden attributable to some distal social risks. We have embarked on this work, but it proves to have challenges that are qualitatively different than many of the risks included here. For nearly all risk-outcome pairs, we assume in the absence of other evidence that the RRs by age and sex are generalisable across populations (the exception is for BMI in Asian and non-Asian populations for breast cancer). In principle, if there is evidence of statistically significant RRs for different population groups, we would incorporate these into the CRA. For distal social risks, the pathways to outcomes can be modified in many ways by other risks or by health-system interventions. We expect that the RR due to low education for 40-year-old men would be different in Norway than in Kenya. Given the greater potential for variation in RRs for distal risks, inclusion in GBD will require more local quantification of RRs and then a further modelling step to estimate RRs for these determinants for all locations. Our first planned target for this quantification is educational attainment.

74Given the global policy focus on the potential health effects of climate change driven by rising levels of greenhouse gases, and consequently temperature, we will add temperature and precipitation as risk factors that are quantified on an annual basis in future iterations of GBD. Even though most of the potential harm that might come from rising temperatures or extreme weather events will occur in the future, in some locations, we might already find significant attributable burden. This analysis will need to examine the relationship between disease and mortality risk and temperature for each relevant outcome. For some outcomes, these relationships are likely to be U shaped, with an optimal temperature for minimum risk. These U-shaped relationships could mean that for some outcomes in some locations, rising temperature might reduce harm, even if in most locations it will increase burden. Likewise, a major issue in understanding the temperature and health outcome relationships is that we would expect these to be attenuated in high-SDI settings, where many individuals can protect themselves from some of the consequences. In other words, generalising from studies in high-SDI locations to other locations might underestimate the risk relationships.

75In the GBD CRA approach, the TMREL is the level of risk exposure that leads to minimum risk for individuals. In principle, the TMREL could vary by location, age, and sex. To date, the TMREL in the GBD work has been selected to be universal. For more detail on TMREL, see appendix 1 (p 22). The analysis of alcohol, where for IHD there is a protective effect at mild to moderate consumption but a harmful effect for neoplasms and injuries, is a good example of where it would be desirable to vary TMREL by age. In younger ages, injuries will be more important than cardiovascular diseases, pushing the TMREL toward zero consumption of alcohol, whereas at older ages, the TMREL might be higher. Letting the TMREL vary by age and sex and even location will add an extra analytical step to GBD; like all other estimation steps, this can have estimation error. To date, we have thought the estimation error associated with a TMREL that varies may not make the effort worthwhile. As evidence accumulates on some risks like alcohol, we will carefully evaluate this position.

76A study of this scope has many limitations. Here we discuss the limitations that apply to the overall risk factor analytical framework and limitations in the estimation approach for new risks and risks that have undertaken significant revisions from GBD 2015. More details and limitations of the analytical approach for each risk factor are presented in appendix 1 (p 43). First, we continue to include risk-outcome pairs that meet the World Cancer Research Fund criteria of convincing or probable evidence for causality. While these criteria have proven a useful bar for inclusion, there is an important subjective element to their interpretation. Some risk-outcome pairs included in this study might not meet these criteria or alternative criteria that are developed as new randomised trials, cohort studies, or case-control studies are published. Second, we used published cohort studies to evaluate the degree to which different risks are mediated through other risks. Estimates of pathways of mediation are used to compute the burden attributable to aggregates of risk factors such as all behavioural risks or all risks combined. While we have conducted pooled cohort analyses to strengthen the assessment of mediation, this work was not yet ready for inclusion in this assessment. Pooled cohort studies have the advantage of providing a more standardised framework for assessing mediation across multiple risks. A related issue is the validation of the aggregation of risks in GBD. Pooled cohort studies will allow (in some circumstances) the opportunity to estimate if the aggregation of GBD RRs is as predictive of outcomes as suggested by the risk-by-risk analysis with mediation. Third, we have used the Das Gupta formula applied for each 5-year interval and for GBD Level 3 causes. Aggregations at higher levels of causes and for longer periods of time are based on these more granular analyses to guarantee consistency. Given the non-linear nature of the Das Gupta decomposition formula, however, alternative results are possible using different time periods and causes in the formula. Fourth, we have introduced the use of ensemble distributions to improve the empirical fitting of distributions of risk exposure in settings where only mean and standard deviation are known or where we use models to predict the mean and standard deviation of exposure. Ensemble models provide more accurate fits as assessed out of sample for settings with microdata. The underlying assumption is that the same ensemble weights are applicable across all settings. It is possible that the shape of distributions of risk exposure might vary across locations, for example because of the effects of access to treatment.

77Limitations that apply to new risks in GBD 2016 or risks with significant estimation updates are presented here. For low birthweight and short gestation, we have included the effect of low birthweight and short gestation only on neonatal outcomes; we have not found the evidence to meet our inclusion criteria for the link between low birthweight and short gestation and NCDs in adult age groups. Our analysis of RRs has used a very large US-linked birth cohort dataset and much more limited data from middle-SDI and low-SDI populations. Given the large number of observations from the USA, our results are heavily influenced by the pattern of RRs across birthweight and gestational age in that population. The microdata used to develop the ensemble distributions for birthweight and gestational age are largely from middle-SDI and high-SDI locations. The estimation of alcohol use relies heavily on sales data, which are limited and whose quality we cannot easily assess. Also, the estimation of unrecorded consumption of alcohol is based on limited data and has significant uncertainty; nevertheless, we feel it is important to include it and plan to continue to look for additional sources of information to improve the estimation of unrecorded consumption in future iterations of GBD. Lastly, methods for calculating TMREL rely on observed DALYs for a given time rather than on the expected share of DALYs estimated from alcohol use alone. Future iterations of GBD will likely need to test this assumption further and determine if separate TMREL by age and sex should be calculated.

78Understanding the levels and trends of major risks for human health is essential to prioritise public health action and evaluate the success of different programmes and policies. This study provides a comprehensive and comparable assessment of 84 metabolic, environmental, occupational, and behavioural risks across locations and time. Our findings show that risk modification has been an important contributor to reductions in communicable, maternal, neonatal, and nutritional causes, but has played a relatively small part in trends in NCDs. Conflicting trends in risks for NCDs at the global level, such as the decline in smoking prevalence coupled with the rise in obesity, FPG, and SBP, account for this finding. By contrast with trends in diseases and injuries at the global level and even at the national level, there is much greater heterogeneity of global trends across risks and considerable geographical variation in leading risks as well. Public health action in each country and region needs to focus on the major risks in that community. Our findings reinforce the crucial need for robust monitoring of the exposure to risks to health and assessment of the evidence supporting causal effects for each risk-outcome pair; GBD provides the main global mechanism for this monitoring function.

79Correspondence to: Prof Emmanuela Gakidou, Institute for Health Metrics and Evaluation, Seattle, WA 98121, USA gakidou@uw.edu

80Correspondence to: Prof Emmanuela Gakidou, Institute for Health Metrics and Evaluation, Seattle, WA 98121, USA gakidou@uw.edu

81For SHARE Project see www.share-project.orgFor HBSC see http://www.hbsc.org

82For SHARE Project see www.share-project.org

83For HBSC see http://www.hbsc.org

84This online publication has been corrected. The corrected version first appeared at thelancet.com on September 18, 2017

85This online publication has been corrected. The corrected version first appeared at thelancet.com on September 18, 2017