Menu
Menu
Menu
Menu
Background: Rigorous analysis of levels and trends in exposure to leading risk factors and quantification of their effect on human health are important to identify where public health is making progress and in which cases current efforts are inadequate. The Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2019 provides a standardised and comprehensive assessment of the magnitude of risk factor exposure, relative risk, and attributable burden of disease. Methods: GBD 2019 estimated attributable mortality, years of life lost (YLLs), years of life lived with disability (YLDs), and disability-adjusted life-years (DALYs) for 87 risk factors and combinations of risk factors, at the global level, regionally, and for 204 countries and territories. GBD uses a hierarchical list of risk factors so that specific risk factors (eg, sodium intake), and related aggregates (eg, diet quality), are both evaluated. This method has six analytical steps. (1) We included 560 risk–outcome pairs that met criteria for convincing or probable evidence on the basis of research studies. 12 risk–outcome pairs included in GBD 2017 no longer met inclusion criteria and 47 risk–outcome pairs for risks already included in GBD 2017 were added based on new evidence. (2) Relative risks were estimated as a function of exposure based on published systematic reviews, 81 systematic reviews done for GBD 2019, and meta-regression. (3) Levels of exposure in each age-sex-location-year included in the study were estimated based on all available data sources using spatiotemporal Gaussian process regression, DisMod-MR 2.1, a Bayesian meta-regression method, or alternative methods. (4) We determined, from published trials or cohort studies, the level of exposure associated with minimum risk, called the theoretical minimum risk exposure level. (5) Attributable deaths, YLLs, YLDs, and DALYs were computed by multiplying population attributable fractions (PAFs) by the relevant outcome quantity for each age-sex-location-year. (6) PAFs and attributable burden for combinations of risk factors were estimated taking into account mediation of different risk factors through other risk factors. Across all six analytical steps, 30 652 distinct data sources were used in the analysis. Uncertainty in each step of the analysis was propagated into the final estimates of attributable burden. Exposure levels for dichotomous, polytomous, and continuous risk factors were summarised with use of the summary exposure value to facilitate comparisons over time, across location, and across risks. Because the entire time series from 1990 to 2019 has been re-estimated with use of consistent data and methods, these results supersede previously published GBD estimates of attributable burden. Findings: The largest declines in risk exposure from 2010 to 2019 were among a set of risks that are strongly linked to social and economic development, including household air pollution; unsafe water, sanitation, and handwashing; and child growth failure. Global declines also occurred for tobacco smoking and lead exposure. The largest increases in risk exposure were for ambient particulate matter pollution, drug use, high fasting plasma glucose, and high body-mass index. In 2019, the leading Level 2 risk factor globally for attributable deaths was high systolic blood pressure, which accounted for 10·8 million (95% uncertainty interval [UI] 9·51–12·1) deaths (19·2% [16·9–21·3] of all deaths in 2019), followed by tobacco (smoked, second-hand, and chewing), which accounted for 8·71 million (8·12–9·31) deaths (15·4% [14·6–16·2] of all deaths in 2019). The leading Level 2 risk factor for attributable DALYs globally in 2019 was child and maternal malnutrition, which largely affects health in the youngest age groups and accounted for 295 million (253–350) DALYs (11·6% [10·3–13·1] of all global DALYs that year). The risk factor burden varied considerably in 2019 between age groups and locations. Among children aged 0–9 years, the three leading detailed risk factors for attributable DALYs were all related to malnutrition. Iron deficiency was the leading risk factor for those aged 10–24 years, alcohol use for those aged 25–49 years, and high systolic blood pressure for those aged 50–74 years and 75 years and older. Interpretation: Overall, the record for reducing exposure to harmful risks over the past three decades is poor. Success with reducing smoking and lead exposure through regulatory policy might point the way for a stronger role for public policy on other risks in addition to continued efforts to provide information on risk factor harm to the general public. Funding: Bill & Melinda Gates Foundation.
The GBD 2019 estimation of attributable burden followed the general framework established for comparative risk assessment (CRA)14, 15 used in GBD since 2002. Here, we provide a general overview and details on major innovations since GBD 2017. More detailed methods are available in appendix 1. CRA can be divided into six key steps: inclusion of risk–outcome pairs in the analysis; estimation of relative risk as a function of exposure; estimation of exposure levels and distributions; determination of the counterfactual level of exposure, the level of exposure with minimum risk called the theoretical minimum risk exposure level (TMREL); computation of population attributable fractions and attributable burden; and estimation of mediation of different risk factors through other risk factors such as high body-mass index (BMI) and ischaemic heart disease, mediated through elevated systolic blood pressure (SBP), elevated fasting plasma glucose (FPG), and elevated LDL cholesterol, to compute the burden attributable to various combinations of risk factors.10 GBD 2019 estimated prevalence of exposure and attributable deaths, YLLs, YLDs, and DALYs for 23 age groups; males, females, and both sexes combined; and 204 countries and territories that were grouped into 21 regions and seven super-regions. GBD 2019 includes subnational analyses for Italy, Nigeria, Pakistan, the Philippines, and Poland, and 16 countries previously estimated at subnational levels (Brazil, China, Ethiopia, India, Indonesia, Iran, Japan, Kenya, Mexico, New Zealand, Norway, Russia, South Africa, Sweden, the UK, and the USA). All subnational analyses are at the first level of administrative organisation within each country except for New Zealand (by Māori ethnicity), Sweden (by Stockholm and non-Stockholm), the UK (by local government authorities), and the Philippines (by province). In this publication, we present subnational estimates for Brazil, India, Indonesia, Japan, Kenya, Mexico, Sweden, the UK, and the USA; given space constraints, these results are presented in appendix 2. For this cycle, nine countries and territories (Cook Islands, Monaco, San Marino, Nauru, Niue, Palau, Saint Kitts and Nevis, Tokelau, and Tuvalu) were added, such that the GBD location hierarchy now includes all WHO member states. These new locations were previously included in regional totals by assuming that age-specific rates were equal to the regional rates. At the most detailed level, we generated estimates for 990 locations. The GBD diseases and injuries analytical framework generated estimates for every year from 1990 to 2019. Individual risk factors such as low birthweight or ambient ozone pollution are evaluated in the GBD CRA. In addition, there has been policy interest in groups of risk factors such as household air pollution combined with ambient particulate matter. To accommodate these diverse interests, the GBD CRA has a risk factor hierarchy. Level 1 risk factors are behavioural, environmental and occupational, and metabolic; Level 2 risk factors include 20 risks or clusters of risks; Level 3 includes 52 risk factors or clusters of risks; and Level 4 includes 69 specific risk factors. Counting all specific risk factors and aggregates computed in GBD 2019 yields 87 risks or clusters of risks. For a full list of risk factors by level, see appendix 1 (section 5, table S2). Since GBD 2010, we have used the World Cancer Research Fund criteria for convincing or probable evidence of risk–outcome pairs.16 For GBD 2019, we completely updated our systematic reviews for 81 risk–outcome pairs. Preferred Reporting Items for Systematic Reviews and Meta-Analyses flowcharts on these reviews are available in appendix 1 (section 4). Convincing evidence requires more than one study type, at least two cohorts, no substantial unexplained heterogeneity across studies, good-quality studies to exclude the risk of confounding and selection bias, and biologically plausible dose–response gradients. For GBD, for a newly proposed or evaluated risk–outcome pair, we additionally required that there was a significant association (p<0·05) after taking into account sources of potential bias. To avoid risk–outcome pairs repetitively entering and leaving the analysis with each cycle of GBD, the criteria for exclusion requires that with the available studies the association has a p value greater than 0·1. On the basis of these reviews and meta-regressions, 12 risk–outcome pairs included in GBD 2017 were excluded from GBD 2019: vitamin A deficiency and lower respiratory infections; zinc deficiency and lower respiratory infections; diet low in fruits and four outcomes: lip and oral cavity cancer, nasopharynx cancer, other pharynx cancer, and larynx cancer; diet low in whole grains and two outcomes: intracerebral haemorrhage and subarachnoid haemorrhage; intimate partner violence and maternal abortion and miscarriage; and high FPG and three outcomes: chronic kidney disease due to hypertension, chronic kidney disease due to glomerulonephritis, and chronic kidney disease due to other and unspecified causes. In addition, on the basis of multiple requests to begin capturing important dimensions of climate change into GBD, we evaluated the direct relationship between high and low non-optimal temperatures on all GBD disease and injury outcomes. Rather than rely on a heterogeneous literature with a small number of studies examining relationships with specific diseases and injuries, we analysed individual-level cause of death data for all locations with available information on daily temperature, location, and International Classification of Diseases-coded cause of death. These data totalled 58·9 million deaths covering eight countries. On the basis of this analysis, 27 GBD cause Level 3 outcomes met the inclusion criteria for each non-optimal risk factor (appendix 1 section 2.2.1) and were included in this analysis. Other climate-related relationships, such as between precipitation or humidity and health outcomes, have not yet been evaluated. In GBD, we use published systematic reviews and for GBD 2019, we updated these where necessary to include any new studies that became available before Dec 31, 2019. We did meta-analyses of relative risks from these studies as a function of exposure (appendix 1 sections 2.2.2, 4). For GBD 2019, 81 new systematic reviews were done, including for 44 diet risk–outcome pairs. To allow for risk functions that might not be log-linear, we relaxed the meta-regression assumptions to allow for monotonically increasing or decreasing but potentially non-linear functions for 147 risk–outcome pairs. Appendix 1 (section 2) provides the mathematical and computational details for how we implemented this approach for meta-regression. 218 risk–outcome pairs were estimated assuming log-linear relationships. For 126 risk–outcome pairs, exposure was dichotomous or polytomous. For 37 risk–outcome pairs, the population attributable fractions were assumed by definition to be 100% (eg, 100% of diabetes is assumed to be, by definition, related to elevated FPG). For 32 risk–outcome pairs, other approaches were used that reflected the nature of the evidence that has been collected for those risks (appendix 1 section 4). For risks that affect cardiovascular outcomes, we adjusted relative risks by age such that they follow the empirical pattern of attenuation seen in published studies for elevated SBP, FPG, and LDL cholesterol. For each risk factor, we systematically searched for published studies, household surveys, censuses, administrative data, ground monitor data, or remote sensing data that could inform estimates of risk exposure. To estimate mean levels of exposure by age-sex-location-year, specific methods varied across risk factors (appendix 1 sections 2.1, 4). For many risk factors, exposure data were modelled using either spatiotemporal Gaussian process regression or DisMod-MR 2.1,17, 18 which are Bayesian statistical models developed over the past 12 years for GBD analyses. For most risk factors, the distribution of exposure across individuals was estimated by modelling a measure of dispersion, usually the SD, and fitting an ensemble of parametric distributions to the predicted mean and SD. Ensemble distributions for each risk were estimated based on individual-level data. Details for each risk factor modelling for mean, SD, and ensemble distribution are available in appendix 1 (section 4). Because of the strong dependency between birthweight and gestational age, exposure for these risks was modelled as a joint distribution using the copula method.19 In many cases, exposure data were available for the reference method of ascertainment and for alternative methods, such as tobacco surveys reporting daily smoking versus total smoking; in these cases, we estimated the statistical relationship between the reference and alternative methods of ascertainment using network meta-regression and corrected the alternative data using this relationship. For harmful risk factors with monotonically increasing risk functions, the theoretical minimum risk level was set to 0. For risk factors with J-shaped or U-shaped risk functions, such as for sodium and ischaemic heart disease or BMI and ischaemic heart disease, the TMREL was determined as the low point of the risk function. When the bottom of the risk function was flat or poorly determined, the TMREL uncertainty interval (UI) captured the range over which risks are indistinguishable. For protective risks with monotonically declining risk functions with exposure, namely risk factors where exposure lowers the risk of an outcome, the challenge is selecting the level of exposure with the lowest level of risk strongly supported by the available data. Projecting beyond the level of exposure supported by the available studies could exaggerate the attributable burden for a risk factor. In these cases, for each risk–outcome pair, we determined the exposure level at the 85th percentile of exposure in the cohorts or trials used in the risk meta-regression. We then generated the TMREL by weighting each risk–outcome pair by the relative global magnitude of each outcome. Appendix 1 (section 2.4 and 4) provides details on the TMREL estimation for each risk. Global age-standardised SEVs for both sexes combined in 1990, 2010, and 2019, and annualised rate of change between 1990 and 2019 and 2010 and 2019 Data in parentheses are 95% uncertainty intervals. SEVs are measured on a 0 to 100 scale, in which 100 is when the entire population is exposed to maximum risk and 0 is when the entire population is at minimum risk. SEVs are shown for all levels of the risk factor hierarchy. ARC=annualised rate of change. SEVs=summary exposure values. For each risk factor j, we computed the population attributable fraction (PAF) by age-sex-location-year using the following general formula for a continuous risk: where PAFjoasgt is the PAF for cause o, for age group a, sex s, location g, and year t; RRjoasg(x) is the relative risk as a function of exposure level x for risk factor j, for cause o controlled for confounding, age group a, sex s, and location g with the lowest level of observed exposure as l and the highest as u; Pjasgt(x) is the distribution of exposure at x for age group a, sex s, location g, and year t; and TMRELjas is the TMREL for risk factor j, age group a, and sex s. Where risk exposure is dichotomous or polytomous, this formula simplifies to the discrete form of the equation. Estimation of the PAF took into account the risk function and the distribution of exposure across individuals in each age-sex-location-year. By drawing 1000 samples from the risk function, 1000 distributions of exposure for each age-sex-location-year, and 1000 samples from the TMREL, we propagated all of these sources of uncertainty into the PAF distributions. PAFs were also applied at the draw level to the uncertainty distributions of each associated outcome for that age-sex-location-year. For the estimation of each specific risk factor, the counterfactual distribution of exposure is the TMREL for that specific risk with no change in other risk factors. Thus, the sum of these risk-specific estimates of attributable burden can exceed 100% for some causes, such as cardiovascular diseases. It is also useful to assess the PAF and attributable burden for combinations of risk factors, such as all diet components together or household air and ambient particulate matter pollution. To estimate the combined effects of risk factors, we should take into account how one risk factor might be mediated through another (eg, the effect of fruit intake might be partly mediated through fibre intake). We used the mediation matrix as developed in GBD 201712 to try to correct for overestimation of the PAF and the attributable burden for combinations of risks if we were to simply assume independence without any mediation. Appendix 1 (section 5, table S6) provides the estimated mediation matrix. As in previous rounds of GBD, we summarised exposure distributions for dichotomous, polytomous, and continuous risk factors using the SEV. The SEV compares the distribution of excess risk times exposure level to a population where everyone is at maximum risk. For a given risk r and outcome c pair where RRmax is the relative risk at the 99th percentile of the global distribution of exposure. We then averaged across outcomes to compute the SEV for a given risk as where N(c) is the total number of outcomes for a risk. The SEV is effectively excess risk-weighted prevalence, which allows for comparisons across different types of exposures. Maximum risk in the denominator of the SEV is determined by the relative risk at the 99th percentile of the global distribution of exposure. The SEV is on a 0–100 scale where 100 means the entire population is at maximum risk and 0 means everyone in the population is at minimum risk. We computed age-standardised SEVs by age-standardising age-specific SEVs across the age groups in which that risk factor was evaluated; this method is a change from GBD 2017 in which age-standardisation included age groups in which the risk was not evaluated. For example, the SEV for low birthweight is now age-standardised across age groups 0–6 days to 7–27 days. To estimate SEVs for groups of risk factors, we first estimated the value of RR2 without mediation through risk 1 (RR2/1). where RR2 is the relative risk of risk factor 2 and MF2/1 is the mediation factor, or the proportion of the risk of risk factor 2 that is mediated through risk factor 1. We then computed the PAF using the non-mediated relative risk (RR1/2) and computed the joint PAF as We cannot simply multiply RRmax values used for the SEV of each component risk as this would exaggerate the joint RRmax. We approximated the 99th percentile of risk for the combination of risk factors by taking the geometric mean of the ratio between the individual risk maximum risk and the individual risk global mean risk and multiplied that by the global mean joint risk. Formally, where N(r) is the total number of risks. We computed risk-deleted death rates as the death rates that would be observed if all risk factors were set to their respective TMRELs. This was calculated as the death rate in each age-sex group multiplied by 1 minus the all-risk PAF for that age-sex group in each location. The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.
