Global, regional, and national incidence, prevalence, and years lived with disability for 301 acute and chronic diseases and injuries in 188 countries, 1990–2013: a systematic analysis for the Global Burden of Disease Study 2013

Overview

Our general approach was similar to that for GBD 2010. The analysis of incidence and prevalence for HIV/AIDS, tuberculosis, and malaria for GBD 2013 have already been reported in detail.⁶¹ Key changes from GBD 2010 were the inclusion of new data through updated systematic reviews and the contribution of unpublished data sources from many collaborators; elaboration of the sequelae list to include asymptomatic states, such as Plasmodium falciparum parasitaemia (without symptoms); use of more detailed nature-of-injury codes; improvements to the Bayesian meta-regression method; increased simulation size for comorbidity; estimation of the prevalence of injuries by cohort; and use of a novel method to estimate the distribution of mild, moderate, and severe anaemia by cause.

Cause and sequelae list changes

Based on feedback about GBD 2010, and input from the GBD 2013 collaborators, we expanded the cause and sequelae list (appendix pp 60–89). There were several key changes. First, we included asymptomatic states as explicit sequelae so that overall disease prevalence estimates were available, which might be useful for disease targeting, health service planning, or mass treatment strategies. Asymptomatic sequelae, by definition, were not associated with ill health and therefore were not assigned disability weights. Second, to deal with the challenge that some of the nature-of-injury categories used in the GBD 2010 were highly heterogeneous, these categories were expanded from 23 to 47. Third, we added several new causes and sequelae. All these additions to the cause list were done to either reduce the size of the large residual categories, such as other injuries, or recognition of substantial epidemiological heterogeneity within a disease category (appendix pp 60–89). With these changes, the cause list was expanded from 289 to 301 causes and from 1160 to 2337 sequelae. Most of the increase in sequelae was due to the expansion of the nature-of-injury sequelae, which applied to each of the external causes of injuries. The appendix pp 90–96 provides a list of the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) and International Classification of Diseases, Ninth Revision (ICD-9) codes for all GBD causes and the nature-of-injury categories.

Data sources

GBD 2010 collaborators undertook systematic reviews for most of the causes and sequelae. For some sequelae, the majority of the data came from household survey microdata reanalysis and administrative data such as hospital discharges. For others, most of the data were extracted from publications. Documentation of the GBD 2010 systematic reviews, however, was not centralised and only some of these reviews have been published. For this study, we updated systematic reviews through Aug 31, 2013. In some cases, studies published after Aug 31, 2013, were identified and included on the basis of GBD collaborator input; no data or studies were extracted after Nov 30, 2014. Household surveys including the demographic and health surveys, multiple indicator cluster surveys, living standards measurement surveys, reproductive health surveys, and various national health surveys included in the Global Health Data Exchange were systematically screened for data relevant to sequelae. For some diseases, case notifications reported to WHO were used as inputs and updated until the end of 2013. The appendix pp 97–653 provides a full list of citations for sources organised by country that were used for this analysis.

We computed an index of the geographical and temporal representativeness of the data sources available for non-fatal health outcomes for each cause or impairment—the data representativeness index (DRI). The overall DRI simply counts the fraction of countries that have any incidence, prevalence, remission, or excess mortality data available for causes that are prevalent in that country. We did not count cause of death data in this measure, even if it was used in the estimation of incidence or prevalence. We computed the same measure for three periods: before 1998, 1998–2005, and 2006 onwards. Table 1 provides the overall DRI and period-specific DRI measures for each cause and table 2 provides the same information for estimation of total impairment prevalence. The DRI was also computed for level 1 and level 2 causes (aggregate causes; see appendix pp 60–89) by counting data availability for any cause within that aggregate. This metric represents the availability of data and does not incorporate any assessment of data quality. The all-cause DRI was 100% overall and for each period, indicating that there was at least data for one cause for all 188 countries in each period. At more detailed levels, however, there was wide variation in the DRI across causes and time. DRI ranged from less than 2% for eight causes, including glucose-6-phosphate dehydrogenase deficiency trait and other mental and substance use disorders, to 100% for Chagas disease, African trypanosomiasis, and food-borne trematodiases. Causes with required infectious disease case reporting had high DRI values. Other disorders, such as cancers, had DRI values above 70% due to the network of population-based cancer registries. Although the time trend varied by disease, many of the highest DRI values were from 1998 to 2005. The lag in data analyses and publications might explain lower DRI values for 2006 to present.

Data representativeness can also be assessed at the country level. Figure 1 shows a map of the percentages of causes for which there were data available in each of the 188 countries between 1990 and 2013. The DRI values ranged from 6% in South Sudan to 92% in the USA. Many developed countries had data for more than 65% of causes; Brazil, India, and China have similar levels. Low levels of data availability were noted in several sub-Saharan African countries, central Asia, the Caribbean, and the Balkans. There was substantial variation within regions; for example, Kenya had 49%, whereas Djibouti had less than 10%, Laos had 14%, and Thailand had 54%.

Sequelae incidence and prevalence

The appendix pp 654–84 provides a brief description of the modelling strategy used for each sequela and cause. The most extensively used estimation method was the Bayesian meta-regression method DisMod-MR 2.0. For some causes such as HIV or hepatitis B and C, disease-specific natural history models were used in which the underlying three state model in DisMod-MR 2.0 (susceptible, cases, or dead) was insufficient to capture the complexity of the disease process. For some diseases with a range of sequelae differentiated by severity, such as chronic obstructive pulmonary disease (COPD) or diabetes mellitus, DisMod-MR 2.0 was used to meta-analyse the data for overall prevalence. Separate DisMod-MR 2.0 models were then used to analyse data for the proportion of cases with different severity levels or sequelae. Likewise, DisMod-MR 2.0 was used to meta-analyse data for the proportions of liver cancer and cirrhosis due to underlying causes such as hepatitis B, hepatitis C, and alcohol use. For acute sequelae, we report incidence (defined as a duration of 3 months or less) at the cause level in table 3, because incidence is the preferred measure for disorders of short duration.

DisMod-MR 2.0 represents a major advance in the computational speed, geographical disaggregation of full internally consistent posterior estimation, and display of data results compared with DisMod-MR 1.0, which was used in GBD 2010. Through cross-validation tests, Flaxman and colleagues reported⁶² that the log-rates specification of models worked as well or better than the negative binomial specification used in DisMod-MR 1.0. Based on these findings, and the substantial improvements in computational speed for log-rate models, this specification was the default method for DisMod-MR 2.0. The appendix pp 3–6 provides details of the DisMod-MR 2.0 likelihood estimation. The DisMod-MR 1.0 sequence of global estimation, regional estimation, and country prediction, which we call an analytical cascade, is illustrated in the appendix p 734. DisMod-MR 2.0 uses a more complete cascade (appendix p 735). At the global level, a mixed-effects non-linear regression with all available country data was used to generate initial global estimates that are passed to the next level of the DisMod cascade to inform the model for each super region. In turn, a super-region specific mixed-effects non-linear regression was used to estimate for regions. The same regression method was used for estimation of further geographical dis aggregation. The analyst could choose, depending on data density, to branch the cascade in terms of time and sex at different levels. In GBD 2010, DisMod-MR 1.0 was used to generate fits for three periods only: 1990, 2005, and 2010 because of long computational time. For GBD 2013, we generated fits for 1990, 1995, 2000, 2005, 2010, and 2013.

DisMod-MR 2.0 internal validity was assessed by use of R² for adjusted data. Results for all DisMod-MR 2.0 models are provided in the appendix pp 654–84. Adjusted data were the original study data transformed to the reference case definition and measurement method, using the meta-regression component of DisMod-MR 2.0 to make the data from different studies with varying methods comparable. External validity was also evaluated through cross-validation on a small number of sequelae due to the computational time and complexity for this analysis. We selected ten DisMod-MR 2.0 models representing a range of data densities to evaluate. We held out 30% of datapoints for incidence and prevalence at random, refit the model, and compared predictions to the held-out data. We assessed model performance using two metrics: the root-mean squared error of the predictions compared with the data held out, and the coverage of the data prediction with 95% uncertainty intervals. The appendix pp 736–37 provides these metrics for the ten models tested. In all cases, external validity was equal to or only slightly worse than the internal validity.

As in GBD 2010, DisMod-MR was not used to model estimates for a shortlist of causes; custom models were created for many of these. For some of these causes, important improvements in the modelling strategy were implemented. Changes for HIV and malaria have been described elsewhere.⁶¹ For dengue, the model was modified to use the first component of a principal components’ analysis of Bhatt and colleagues’ dengue transmission probability to improve estimation of case rates.⁶³ For lymphatic filariasis, precontrol levels were estimated from data reported in the lymphatic filariasis atlas.⁶⁴ Last, based on crucial input from GBD collaborators, we chose to model rheumatic heart disease in low-income and middle-income countries separately from high-income countries in view of potential differences in long-term cohort effects of treatment.

Estimation for cancer in GBD 2013 largely followed a similar analytical strategy to GBD 2010, which used a combination of incidence data, survival data, and sequelae durations to estimate cancer prevalence and YLDs.^65,66 The analysis benefited from the inclusion of both the latest edition of Cancer Incidence in Five Continents and a larger number of other cancer registries particularly in China. In GBD 2013, we also incorporated new data from the US National Cancer Institute’s Surveillance, Epidemiology, and End Results Program (SEER)⁶⁷ and WHO’s International Agency for Research on Cancer’s Cancer Survival in Africa, Asia, the Caribbean, and Central America to update best and worst case survival, yearly survival trends, and sequelae durations for all cancers.⁶⁸ Based on evidence that individuals with most cancers continue to have higher mortality beyond 5 years than do the general population, we estimated the burden of cancer for up to 10 years after incidence. Estimates for cancer sequelae now represent the burden for all cancer patients by contrast with estimation of the burden just for cancer survivors (see appendix pp 7–8 for more detail on aspects of estimating non-fatal cancer outcomes that were different from the methods used in GBD 2010).

Injuries

We followed a similar strategy to GBD 2010 for estimating the burden of injuries, except for an expanded list of 26 external cause-of-injury categories (from 15) and 47 nature-of-injury categories (from 23) for both short-term outcomes and lasting disability (see appendix pp 90–96 for ICD codes). More detail was added to both external causes and nature-of-injury categories to reduce epidemiological heterogeneity within each combination of cause and nature-of-injury category. The key analytical steps are explained in greater detail in the appendix pp 9–14. Here we provide a summary of the methods.

First, for each external cause, DisMod-MR 2.0 was used to analyse incidence based on hospital, emergency department, and survey data. Second, we estimated the distribution of nature of injury for each external cause using data that had both types of code available. When individuals were coded with more than one nature-of-injury code, we used the most severe. Third, we analysed seven studies that provided at least 1 year of follow-up for various natures of injury to estimate long-term disability.^69–75 Fourth, we estimated cohort prevalence of long-term disability from the incident cases of injury for each external cause and nature-of-injury combination while accounting for excess mortality for the more severe post-injury sequelae. For some injuries, treatment modifies the disability weight. In these cases, we approximated the fraction of injuries receiving treatment as a function of an indicator of health system access.⁷⁶

Short-term disability was estimated for all natures of injury by cause-of-injury categories as the product of prevalence (estimated by multiplying incidence by mean duration) and the appropriate disability weight. The duration for treated cases of injuries was determined by information in the Dutch Injury Surveillance System follow-up studies of 2001–04 and 2007–10.^71,73 We used expert opinion to estimate a multiplier for the duration of short-term disability from untreated injuries and used the estimates of access to care by country and year as we have described for the long-term disability.

YLDs from 29 residual causes

Despite expanding our list of causes and sequelae in GBD 2013, many diseases remain for which we do not explicitly model disease prevalence and YLDs. The GBD cause list is collectively exhaustive such that all sequelae with an ICD code are mapped to a cause group (appendix pp 90–96). Many less common sequelae are included in 29 of the residual categories. For 14 of these cause groupings, epidemiological data for incidence or prevalence are available so that they can be modelled as other causes have been modelled—this set includes meningitis, cirrhosis, liver cancer, pneumoconiosis, and chronic kidney disease due to other causes, other neoplasms, other cardiovascular and circulatory diseases, other drug use disorders, other mental and substance use disorders, other gynaecological diseases, other musculoskeletal disorders, other skin and subcutaneous diseases, age-related and other hearing loss, other vision loss, other sense organ diseases, and other oral disorders. For 12 residual categories (other intestinal infectious diseases, other neglected tropical diseases, other maternal disorders, other neonatal disorders, other nutritional deficiencies, other infectious diseases, other chronic respiratory diseases, other digestive diseases, other neurological disorders, other urinary diseases, other haemoglobinopathies and haemolytic anaemias, and other congenital anomalies), epidemiological data for incidence and prevalence were not available for the entire residual cause groupings but sufficient cause of death data allowed for cause of death estimates. For each category, we identified causes within the larger cause group that had both estimates of years of life lost (YLLs) and YLDs, which we expected to have similar ratios of mortality to morbidity. We then computed the ratio of YLLs to YLDs for these specific causes (on a country–sex–year basis) and applied them to the residual category’s YLLs to estimate its YLDs. This approach makes the simplified assumption that on average within a level 2 disease grouping the disability is proportionate to mortality within a country–sex–year. For an additional three residual categories (other sexually transmitted diseases, other drug use disorders, and other mental and substance use disorders), there were no overall epidemiological data or sufficient deaths to generate cause of death estimates. For the last two, we used US outpatient data or prevalence data from the Medical Expenditure Panel Survey (MEPS), National Epidemiologic Survey on Alcohol and Related Conditions (NESARC), or the 1997 Australian mental health survey⁷⁷ and applied a severity distribution from these surveys in all countries and periods. These two causes for which US and Australian data were applied worldwide account for 1·6% of global YLDs.

Impairments

As in GBD 2010, we estimated the country–age–sex–year prevalence of nine impairments: anaemia, epilepsy, hearing loss, heart failure, intellectual disability, infertility, vision loss, Guillain-Barré, and pelvic inflammatory disease. These impairments were selected because they are sequelae of more than one disease and data are available to estimate prevalence for the overall impairment. Generally, overall impairment prevalence was estimated using DisMod-MR 2.0. Cause-specific estimates of impairments, such as the 19 causes of blindness, are required to provide the total prevalence estimated for that impairment. Anaemia, epilepsy, hearing loss, heart failure, intellectual disability, and pelvic inflammatory disease are estimated for different levels of severity. Separate estimates were made for primary infertility (in couples who have not been able to conceive) and secondary infertility (in couples having trouble conceiving again) and, for each, if the impairment is affecting men or women, or both. The severity distribution of cause-specific prevalence of each impairment was estimated as explained above or, in the absence of severity-specific data, assumed to be proportionate across all levels of severity. In the case of epilepsy, severity levels were determined by mixed-effect models for the proportions of primary, severe, and treated epilepsy, and a meta-analysis for seizure-free treated epilepsy, and thus had values that were specific for country, age, sex, and year. DisMod-MR 2.0 models produced country-specific, age-specific, sex-specific, and year-specific levels of hearing loss and vision loss. Due to little information about the severity levels of intellectual disability, we assumed a similar distribution of severity worldwide based on meta-analysis of intelligence quotient (IQ)-specific data for the overall impairment. This was supplemented with cause-specific distributions for chromosomal causes and iodine deficiency, whereas the severity of intellectual disability included in the long-term sequelae of causes such as meningitis, neonatal tetanus, and malaria was combined with several impairments such as motor impairment, blindness, or seizures. The severity of heart failure is derived from our MEPS analysis and therefore is not specific for country, year, age, or sex.

Our method for estimating overall anaemia was largely the same as in GBD 2010 but with the addition of new data sources, specifically subnational data for the UK, China, and Mexico.⁴³ We adopted different thresholds for defining anaemia during the neonatal period, because the GBD 2010 thresholds did not account for haematological realities of early life. The GBD 2013 thresholds match the WHO recommendations⁷⁸ with the exception of thresholds of less than 1 month because there is no international cutoff for diagnosis at that age.^43,79 To disaggregate marginal estimates of anaemia severity and cause into a complete set of prevalence estimates for cause and severity pairs, we developed a new method for GBD 2013 that used techniques from Bayesian contingency table modelling.^80,81

In GBD 2010, hearing loss of more than or equal to 35 dB in DisMod-MR 1.0 was estimated and then broken down into six severity levels based on a series of regressions on the proportionate distribution across categories. In GBD 2013, we first estimated the prevalence of normal hearing, hearing loss of 20–34 dB (mild), and greater than 35 dB (moderate and above); these three categories were fixed to add up to 100%. We then ran separate DisMod-MR 2.0 models for five severity levels (ie, moderate 35–49 dB, moderately severe 50–64 dB, severe 65–79 dB, profound 80–94 dB, and complete ≥95 dB), which were then proportionally rescaled to fit in the 35 dB or greater envelope. In GBD 2010, the same severity distribution was assumed for each cause of hearing loss. In GBD 2013, we customised the prevalence estimation for each cause. Hearing loss due to otitis media and age-related hearing loss were estimated by DisMod-MR 2.0 using prevalence data. Hearing loss due to meningitis was estimated as a proportion of meningitis cases from a meta-analysis.⁸² Congenital hearing loss was estimated using birth prevalence data in DisMod-MR 2.0, assuming a constant prevalence for all ages because there was no evidence of an increased mortality risk. We assumed all hearing loss from otitis media was mild or moderate on the basis of reported distribution of hearing loss.^83,84 To account for hearing aids, we assumed that the use of a hearing aid reduces the severity of hearing loss by one severity level. The other causes were assumed to cover the full range of severities. More details about impairments are provided in the appendix pp 15–32.

Severity distributions

For 213 causes, a range of sequelae are defined in terms of severity. Important changes to the sequelae list with regards to severity include low back pain, alcohol and drug dependence categories, uterine prolapse, and epilepsy. Milder states for low back pain and alcohol and drug dependence categories were added because these disorders had a large gap between asymptomatic cases and the high value of the disability weight for the least severe symptomatic categories, whereas the epidemiological data for severity indicates a sizeable proportion of cases with milder disability. Stress incontinence was added as a sequela of uterine prolapse with a new disability weight that is distinct from full incontinence. Also, epilepsy health states are now better aligned with epidemiological data based on seizure frequency. In cases in which severity is related to a particular impairment, such as mild, moderate, and severe anaemia due to malaria, the analysis is driven by the impairment estimation described above. For some outcomes such as COPD or asthma, data have been gathered in different locations around the world and these have been modelled using DisMod-MR 2.0 (see appendix pp 694–733 for details). In other cases, published meta-analyses have been used to estimate the allocation of cases by severity. For the remaining causes, we used the same approach for estimating the distribution of severity as in the GBD 2010; empirical analysis of this model was updated through the addition of 2 years from the US MEPS. The appendix pp 685–87 lists the GBD causes that can be identified in MEPS and the corresponding ICD-9 CM codes. In total, 203 960 observations, covering 119 676 individuals, were used. In the cases of dementia, Parkinson’s disease, multiple sclerosis, osteoarthritis, schizophrenia, and bipolar disorder, data identified through literature reviews were used to inform the severity distribution. The introduction of a mild health state for four drug dependence categories required identification of epidemiological data to estimate the proportion of cases with mild versus more severe disability. For cannabis dependence, we used the NESARC survey in the USA and the Australian National Survey of Mental Health and Wellbeing. For the remaining three drug dependence categories, we only had access to one study on polydrug users in Australia, which led to about half of dependent cases being assigned to the more severe and mild health states. Although this information is derived from a non-representative cohort of drug users, it was thought to be more appropriate than deriving a severity distribution from a household survey like NESARC in which only a small proportion of individuals dependent on opioids, cocaine, or amphetamines would be represented.

Revisions to disability weights

The GBD 2010 disability weights measurement study introduced a new method of pairwise comparisons as a means of eliciting weightings for health states in population surveys.^85,86 Data were gathered in five countries (Bangladesh, Indonesia, Peru, Tanzania, and the USA) and supplemented with a web survey. In total, responses were gathered from 30 230 people in 167 countries. Respondents were presented with a series of randomly selected pairwise comparisons of lay descriptions of health states and asked to state which health state is healthier than the other. Salomon and colleagues⁸⁵ developed a statistical model that yields from these pairwise comparisons disability weights on a scale from 0 (no health loss) to 1 (equivalent to death).

Based on important commentary and review of the GBD 2013 collaborators, we have revised the lay descriptions of 32 states and added 16 new states. The revised lay descriptions were based on identifying inconsistency in the way progression across levels of severity had been handled for some outcomes and the addition of social isolation to the descriptions for complete, profound, and severe hearing loss. New states included five milder health states for alcohol and drug dependence; two health states for the alignment of epilepsy with the epidemiological data defining severe epilepsy in individuals who had on average one or more fits per month and less severe epilepsy in those with between one and 11 fits in the past year; two milder health states for low back pain; and one each for stress incontinence, concussion, hypothyroidism, hyperthyroidism, thrombocytopenic purpura, vertigo, and amputation of one arm without treatment. The appendix pp 688–93 provides a complete list of the lay descriptions of all 235 GBD 2013 health states.

In 2013, we had the opportunity to collaborate with the European Centre for Disease Prevention and Control to gather new data for disability weights in four population-based national surveys (Hungary, Italy, Sweden, and the Netherlands) using the Salomon and colleagues’ protocol.^85,87 Because of funding and questionnaire length, the surveys included 140 of 220 GBD 2010 health states for which the lay descriptions had not been revised, 32 health states with revised lay descriptions, and 42 new health states, 16 of which were included in GBD 2013. These nationally representative samples were comprised of 30 660 respondents. For GBD 2013, the data of GBD 2010 disability weights measurement study and the European disability weights measurement study⁸⁸ were pooled in a single analysis of individual responses, thus doubling the number of respondents to 60 890 in both studies. For states where the lay description was not previously included, revised, or new, only the European disability weights measurement study data were used. This means that all disability weights in GBD 2013 differ from the GBD 2010 disability weights. Most disability weights changed slightly, but some differ more widely (appendix pp 688–93). Some of the more substantial changes were due to the inclusion of incontinence in the lay descriptions for spinal cord injury and the inclusion of the psychological consequences of social isolation in people with more severe hearing loss, leading to much higher disability weights. The statistical analysis generates uncertainty distributions for each disability weight that are propagated into the uncertainty distributions of the estimates of YLDs.

Comorbidity

Many individuals have more than one disease or injury sequela at the same time. To accurately account for comorbidity and its effect on disability for individuals, we used the GBD 2010 microsimulation approach. In the microsimulations, a set of individuals are exposed to the probability of having all the different sequelae included in the GBD to estimate a distribution of the combinations that might be seen in each country–age–sex–year. We modelled the probabilities within each country–age–sex–year of different sequelae as independent. Although there are clear examples of the probability of one sequela changing the probability of other sequelae, such as diabetes and ischaemic heart disease, testing reported by Vos and colleagues¹ suggested that modelling assuming independence was a reasonable approximation. However, for less common sequelae the microsimulation tends to increase the estimated uncertainty in the number of YLDs substantially because, for example, a sequela that is estimated to have a prevalence of less than one in 10 000 will not appear randomly in many microsimulations of 20 000. Two steps have been taken to reduce the inflation of uncertainty for uncommon sequelae. First, the number of simulants in each country–age–sex–year was increased to 40 000; the main limiting factor for the number of simulants is computational resources needed to run each of the 62 880 country–age–sex–year simulations 1000 times to account for uncertainty in each of the input prevalence rates. Second, we excluded sequelae in a country–age–sex–year with a prevalence of less than one in 20 000 from the microsimulation. The combined disability weight for individuals with several sequelae was computed as in the GBD 2010 using a multiplicative model; namely, the individual’s disability weight is equal to one minus the cross product of one minus the disability weight for each sequela that the individual has. An output from the comorbidity microsimulation is counts of the number of sequelae for each simulant in the population. The numbers of simulants with different comorbidities in a country–age–sex–year was adjusted from 40 000 to equal the estimated population in each country–age–sex–year to produce the estimated distribution of individuals in each country with comorbidities. Sequelae with a prevalence of less than one in 20 000 that were not included in the microsimulation, are also not included in the population pyramids showing individuals by numbers of sequelae (figure 2A–C). A technical description of the comorbidity simulation is given in the appendix p 2.

We have reported 95% uncertainty intervals for each quantity in this analysis. For disease or sequelae incidence or prevalence rates, age-standardised rates or counts, the models such as DisMod-MR 2.0 provide posterior distributions for each quantity from which 95% uncertainty intervals are computed. For YLDs, we incorporated uncertainty in prevalence and uncertainty in the disability weight into the posterior distribution of YLDs. In practice, we estimated the posterior distribution of YLDs by taking 1000 samples from the posterior distribution of prevalence and 1000 samples of the disability weight to generate 1000 samples of the YLD distribution. We estimated the 95% uncertainty interval by reporting the 25th and 975th values of the distribution. Uncertainty intervals for YLDs at different timepoints (1990, 1995, 2000, 2005, 2010, and 2013) for a particular disease or sequela are correlated because of the shared uncertainty in the disability weight.

Role of the funding source

The funder 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.