Menu
Menu
Menu
Menu
Background: While there is a long history of measuring death and disability from injuries, modern research methods must account for the wide spectrum of disability that can occur in an injury, and must provide estimates with sufficient demographic, geographical and temporal detail to be useful for policy makers. The Global Burden of Disease (GBD) 2017 study used methods to provide highly detailed estimates of global injury burden that meet these criteria. Methods: In this study, we report and discuss the methods used in GBD 2017 for injury morbidity and mortality burden estimation. In summary, these methods included estimating cause-specific mortality for every cause of injury, and then estimating incidence for every cause of injury. Non-fatal disability for each cause is then calculated based on the probabilities of suffering from different types of bodily injury experienced. Results: GBD 2017 produced morbidity and mortality estimates for 38 causes of injury. Estimates were produced in terms of incidence, prevalence, years lived with disability, cause-specific mortality, years of life lost and disability-adjusted life-years for a 28-year period for 22 age groups, 195 countries and both sexes. Conclusions: GBD 2017 demonstrated a complex and sophisticated series of analytical steps using the largest known database of morbidity and mortality data on injuries. GBD 2017 results should be used to help inform injury prevention policy making and resource allocation. We also identify important avenues for improving injury burden estimation in the future.
GBD is predicated on the principle that every case of death and disability in the population should be systematically identified and accounted for in the formulation of global disease and injury burden. On the side of mortality, every death that occurs in the population should have one underlying cause of death which can be assigned to a cause in a mutually exclusive, collectively exhaustive hierarchy of diseases and injuries that can cause death. These data can be used in a method described below to calculate cause-specific mortality rates and years of life lost. For morbidity, every non-fatal case of disease or injury should have an amount of disability assigned for some period of time. These data can be used in a process described below to estimate the incidence, prevalence and years lived with disability. Summing morbidity and mortality from some cause form the burden from that cause, expressed as disability-adjusted life-years (DALY). For causes with known risk factors, some portion of this burden may be explained by exposure to that risk factor. Across causes within some population, it is also a principle of GBD that the sum of all cause-specific deaths should equal all-cause mortality in the population, and that rates of incidence, prevalence, remission and cause-specific mortality can be reconciled with one another such that all death and disability in a population is internally consistent across causes and geographies. As examples, the sum of different types of road injury cases must sum up to overall road injuries, and the sum of deaths from different injuries in a given country must sum up to the estimate of all-injury deaths. The principle of internal consistency extends to populations used in GBD, where every birth, death and net migration must be accounted for in the population estimates which form the denominators of GBD results. While there is immense complexity in the process summarised above, it is important to begin with these core principles which govern the computation processes at the heart of GBD burden estimation. A summarised overview of key GBD 2017 methods is also provided in online supplementary appendix 1. injuryprev-2019-043531supp001.pdf GBD study design, including cause-specific methods, is described in a high level of detail in associated publications.2–7 In addition to the injury-focused methods described in this paper, it is important to define hierarchies used in the GBD study design. In particular, GBD 2017 was built around a location hierarchy where different subnational locations (eg, US states, India states, China provinces) which form a composite of a national location (eg, the USA, India, China). National locations are aggregated to form GBD regions, which are then aggregated to form GBD super regions. These designations affect the modelling structure and utilisation of location random effects, processes which are described in more detail later. The country-level and regional-level GBD location hierarchy used in GBD 2017 is provided in online supplementary appendix table 1. In addition to locations, GBD processes are conducted to produce estimates for every one of 22 age groups, male and female sex and across 28 years from 1990 to 2017 (inclusive). Age-standardised, all-age and combined sex results are also computed for each GBD result. Exceptions exist to the rules above, for example, self-harm is not permitted to occur in the 0–6 days (early neonatal) age group in the GBD age hierarchy. There are no sex restrictions placed on any GBD injury causes, although these restrictions exist for other GBD causes, such as cancers like prostate, cervical and uterine being related to one sex. injuryprev-2019-043531supp002.pdf In the GBD cause hierarchy, injuries are part of the first level of the GBD cause hierarchy, which consists of three broad groups: communicable, maternal, neonatal and nutritional diseases; non-communicable diseases and injuries. Additional levels of the GBD cause hierarchy provide additional detail. The hierarchy of injuries in GBD is provided in table 1. The organisation of the hierarchy has implications both in terms of how results are produced and in terms of analytical and processing steps which are discussed in more detail below. Case definitions including International Classification of Diseases (ICD) codes used to identify injury deaths and cases are provided in table 2. Global Burden of Disease cause-of-injury hierarchy Case definitions for cause of injury in GBD 2017 Injuries definition: damage, defined by cellular death, tissue disruption, loss of homeostasis, pain limiting activities of daily living or short-term psychological harm (for cases of sexual violence), inflicted on the body as the direct or indirect result of a physical force, immersion or exposure, which may include interpersonal or self-inflicted forces. GBD, Global Burden of Disease; ICD, International Classification of Diseases. GBD separates the concept of cause of injury from nature of injury. Cause of injury (eg, road injuries, falls, drowning) have historically been used for assigning cause of death as opposed to the ‘nature’ of injury, which more directly specifies the pathology that resulted in death. For example, an individual who falls, fractures his or her hip, undergoes surgery and then develops hospital-acquired pneumonia and dies while hospitalised would still have a fall as the underlying cause of death, regardless of whether sepsis or some other disease process leads to death more proximally in the chain of events. In this individual, the ‘nature’ of injury would have been specified as a hip fracture, since it is the bodily injury that would dictate the disability this person experiences. Since it is evident that a hip fracture is more disabling than a mild skin abrasion, it is important for measuring non-fatal burden to consider both the cause and the nature in the formulation of complete injury burden. A full list of nature of injury is provided in table 3. GBD nature of injury GBD, Global Burden of Disease; GI, gastrointestinal; TBI, traumatic brain injury. As described above, cause-specific mortality is measured for every cause of injury in the GBD cause hierarchy with the exception of foreign body in the ear and sexual violence, which undergo only non-fatal burden estimation (described in more detail below). GBD adheres to five general principles for measuring cause-specific mortality, which are described in more detail elsewhere but are summarised as follows.12 First, GBD 2017 identifies all available data. For injuries, this includes vital registration (VR), vital registration samples, verbal autopsy (VA), police records and mortuary/hospital data. VR is the preferred data source but is not available in every location in the GBD location hierarchy. Prior VA research has demonstrated that VA is more accurate for certain injury causes than it is for certain diseases.13 Police data undergo additional validity checks to ensure that systematic under-reporting does not occur in comparison to VR data, which is described in more detail in a related publication.6 The second general principle relevant to injury mortality estimation is maximising comparability and quality of the dataset. For the purposes of injury mortality estimation, this process is largely focused on (1) ensuring appropriate accounting for different ICD code versions used for cause of death data classification over time, (2) redistribution of ill-defined causes of death (described in more detail elsewhere) and (3) processing VA studies into usable data that map to the GBD cause hierarchy.8 9 12 The third general principle for injury cause of death models in GBD 2017 is to develop a diverse set of plausible models. This process is conducted via the Cause of Death Ensemble model (CODEm) framework, which is the standard, peer-reviewed cause of death estimation process used extensively in the GBD study. CODEm generates a large set of possible models based on covariates suggested by the modeller based on expert input and literature review (eg, alcohol for road injuries) and then runs every plausible model, which can range into the thousands per cause. These models can be conducted in both rate space and cause fraction space and use an assortment of combinations among the user-selected covariates (table 4). Fourth, the predictive validity of each one of these submodels is tested using test-train holdouts, whereby a specific model is trained on a portion of data and tested on a separate portion to determine out-of-sample predictive validity. Once the submodels are conducted and predictive validity is measured, then an ensemble model is developed out of the submodels. The submodels and the ensemble model are then subject to the fifth principle, which is to choose the best-performing models based on out-of-sample predictive validity. The chosen models may be a single cause model or an ensemble of models. Beyond these processes, which have become automated with expert review in the GBD processing architecture, there is also considerable time required by the analysts, modellers, collaborators and principal investigators who are involved in the GBD study. Such processes also come under expert scrutiny via the GBD Scientific Council and the peer-review process in the annual GBD capstone publications.2–7 Covariates used in GBD cause of death models BMI, body mass index. Once submodels and ensemble models have been conducted for each cause in the GBD cause hierarchy, a process to correct for cause of death rates to ensure internal consistency is conducted. Specifically, each subcause within some overall cause is rescaled such that, for example, every subtype of road injuries sums to road injuries deaths overall, and then road injuries and other transport injuries sum to equal the overall transport injuries cause. As this cascades to the overall cause hierarchy and the overall all-cause mortality rates, cause-specific mortality across all causes ultimately equals the overall mortality in the population. An example of an injuries cause of death model with vital registration data (Colombia, females) is shown in figure 1. A similar model with relatively less data is shown in figure 2 (Honduras, females). While data are absent in more recent years in Honduras, the model is still able to follow temporal trends, age patterns and broader geographical patterns by harnessing signals from covariate-based fixed effects (eg, alcohol consumption per capita) and location-based random effects (eg, the regional trends in Central Latin America and patterns in neighbouring countries). All cause of death models from GBD 2017 are publicly available for review (https://vizhub.healthdata.org/cod/). Cause-specific deaths are converted to cause-specific mortality rates (CSMRs) using GBD populations. Once CSMRs are established, years of life lost (YLLs) are computed as the product of CSMRs and residual life expectancy at the age of death. The residual life expectancy is based on the lowest observed mortality rate for each age across all populations over 5 million. For example, if a death from road injuries occurs at age 25 and the residual life expectancy is 60 years, then there are 60 YLLs attributed to that death. If the death had occurred at age 50 with a residual life expectancy of 38 years, then 38 YLLs would be attributed. Life tables used for GBD 2017 are provided in related publications. 7 Cause of Death Ensemble model with data points for road injuries in Colombia for females. Cause of Death Ensemble model with data points for road injuries in Honduras for females After cause-specific models for each cause of injury in the GBD cause hierarchy are conducted, the non-fatal estimation process is conducted. An overview of this process is depicted in figure 3. In the first stage, we estimate the incidence of injuries warranting medical care using DisMod-MR 2.1 (abbreviated DisMod). DisMod is a meta-regression tool for epidemiological estimation that uses a compartmental model structure whereby a healthy population may become diseased or injured, at which point the individual either remains a prevalent case, goes into remission or dies. DisMod essentially fits differential equations to reconcile the transitions between these different compartments, so that the final posterior estimate for each epidemiological parameter can be explained in the context of the other parameters. Similar to the principles described in CODEm, DisMod uses all available data, ranging from incidence data to cause-specific mortality rates from the corrected CODEm results, to produce estimates for every age, sex, year and location. For the purposes of injuries, we established our case definition for non-fatal injuries as injuries that require medical care. This is a necessary case definition as we do not want to consider minor stumbles and falls, for example, that led to no actual bodily harm as injuries for GBD, since they would not have any associated disability. These models are conducted only for injury causes as opposed to the nature of injuries references above. Each data input is designated based on type of data—specifically, inpatient data, outpatient data, surveillance data, survey data and literature studies that are population-representative. We model incidence rates for hospital admissions for injuries, so the non-inpatient data sources get adjusted according to their classification so that the model inputs are consistent as injuries that warranted or received inpatient medical care. The coefficients measured by DisMod that were used for adjustment are provided in table 5. Input data for injury cause incidence models included sources identified as part of systematic reviews conducted in past GBD cycles, new sources identified by the GBD collaborator network and new sources of clinical data and other injuries data obtained by the core injuries burden estimation team at the Institute for Health Metrics and Evaluation at the University of Washington. In addition, CSMRs from the corrected CODEm models described above are used in this stage of DisMod modelling. The list of non-fatal injury sources used in GBD 2017 is provided in online supplementary appendix table 2. The completed DisMod models for inpatient incidence for each cause of injury are publicly available at https://vizhub.healthdata.org/epi/. Injuries non-fatal estimation flow chart. Covariates and coefficients used in Global Burden of Disease incidence cause models injuryprev-2019-043531supp003.pdf Once an incidence cause model is constructed for each cause of injury, an extensive analytical ‘pipeline’ follows which converts injury cause incidence into years lived with disability. First, inpatient incidence is split into inpatient and outpatient incidence using coefficients empirically measured by DisMod. The outpatient coefficients for each injury cause are also included in table 5. Separate pipelines are then conducted for inpatient and outpatient injury incidence—each step below can be considered to have been run for both streams of data, for each cause of injury. After the coefficient is applied, incidence is adjusted by the excess mortality rate measured by DisMod to essentially remove injury cases that died after the injury occurred. Once these deaths are removed from the incidence pool, the resulting steps are applied to these surviving cases of injury. First, each new case of injury is considered to have 47 possible ‘natures’ of injury that can result. These are the types of bodily injury that are considered to be possible outcomes from a given injury cause. The proportion of new cases of injury that would have some nature of injury as the most disabling outcome is determined based on dual-coded clinical data sources where both the cause and nature of injury were included as ICD codes.10 Of note, one limitation of this process is that due to computational demands, it is currently only possible to apportion the most disabling nature of injury for each new case of injury. As such, the probability that each nature of injury is the most disabling nature of injury for some cause of injury is modelled in a Dirichlet regression such that the probabilities sum to 1. In other words, each nature of injury has some probability of being the most disabling injury suffered by the victim of some cause of injury, but if multiple natures of injury occurred, then the less disabling injuries are not captured as part of that injury cause’s disability. This limitation has been recognised as a limitation of GBD injury burden estimation in various peer-reviewed articles and will likely be addressed in future GBD updates as computational efficiency improves.3 10 The probability distributions of each cause-nature are computed separately for each age, sex, year and location. At this point, the analytical stage has the age-specific, sex-specific, year-specific, location-specific incidence of a cause-nature combination, for example, the incidence of road injuries that led to a cervical-level spinal cord injury in males aged 20–24 years in 2017 in Stockholm, Sweden. The next step converts these incidence estimates into short-term and long-term injury incidence estimates, where long-term disability is defined as having a lower functional status 1 year postinjury than at the time of injury. These probabilities were measured using long-term follow-up studies.14–20 For some natures of injury, such as lower extremity amputation, the probability of being a long-term injury is 1. The probabilities of short-term versus long-term injury for each cause-nature combination are used to split the incidence values into short-term and long-term pipelines. The long-term incidence is then converted to prevalence using the ordinary differential equation solver used in DisMod, which also uses as an input excess mortality estimated for certain natures of injury such as traumatic brain injury and spinal cord injury conducted in a previous systematic review and meta-analysis. The short-term incidence is converted to prevalence by multiplying incidence and duration of injury, where duration of injury was either computed directly from follow-up studies or, in the case of unavailable data, estimated by an expert clinical panel involved in previous iterations of the GBD study. Since access to medical treatment is assumed to affect duration of injury and disability, the GBD Healthcare Access and Quality Index is used to estimate the proportion with and without access to medical treatment on a location-specific basis.21 The average duration for short-term injury is therefore calculated as the percentage treated multiplied by treated duration added to the percentage untreated multiplied by the untreated duration. The output from this step is the short-term prevalence of each cause-nature combination. Short-term prevalence is subtracted from long-term prevalence at this stage to avoid double counting the same case of injury. Once short-term and long-term prevalence estimates for each cause-nature are computed, then disability weights as derived by the Salomon et al process are assigned to each injury nature.22 Short-term disability weights by injury nature are shown in table 6, which does not include amputations since we assume they cause only long-term disability. The full list of long-term disability weights by injury nature, location and year are provided in online supplementary appendix table 3, which does not include foreign body in respiratory system, foreign body in gastrointestinal and urogenital system, foreign body in ear and superficial injury of any part of body, since we assume these natures of injury do not cause long-term disability. After disability weights are assigned to each injury case, years lived with disability for each cause of injury are calculated as the prevalence of each health state multiplied by the corresponding disability weight and then summed across natures of injury for each cause to compute years lived with disability (YLDs) for each age, sex, year and location for that injury cause. YLDs then undergo comorbidity adjustment used across the GBD study whereby comorbid cases of disease and injury in the population are simulated and adjusted disability weights are computed. These processes are described in more detail in GBD literature.3 GBD 2017 provided an important methodological update whereby nature of injury results, regardless of cause of injury, could be reviewed in the results from this process; this has enabled more advanced GBD research such as measuring the burden of traumatic brain injury and spinal cord injury, measuring the burden of facial fractures and measuring the burden of hand and finger fractures.10 Short-term disability weights for each nature of injury GI, gastrointestinal; TBI, traumatic brain injury. injuryprev-2019-043531supp004.pdf Sexual violence follows a different analytical pathway than the other causes of injury. This process is shown in figure 4. We used the same study framework as was developed for other injury rates in the GBD 2017 study to estimate the yearly proportion of the population that experienced at least one episode of sexual violence in the past year, using a case definition of any sexual assault including penetrative sexual violence (rape) and non-penetrative sexual violence (other forms of unwanted sexual touching). To inform the sexual violence estimates, we identified data in 93 countries that met the case definition above. This resulted in 263 site-years of data, which mainly were derived from surveys such as Demographic and Health Surveys and Reproductive Health Surveys. Similar to our other injury models, we used DisMod 2.1 to model prevalence. The sexual violence prevalence model used study-level covariates for each type of survey question, for example, we used a study-level covariate to identify surveys that identify penetrative sexual violence only to account for how the overall incidence of sexual violence is greater than this value. This model also used a covariate on alcohol use in litres per capita for each location to help fit the model in data-sparse locations. Once yearly prevalence was measured, sexual violence cases undergo a process by which short-term disability from the physical and psychological harm of sexual violence cases is assigned to each prevalent case; however, long-term sequelae of sexual violence are currently not captured in this process, which has been a known limitation of sexual violence estimation in the GBD framework. Sexual violence estimation flow chart. HAQI, Healthcare Access and Quality Index. After estimation of cause-specific mortality and YLLs as well as non-fatal health outcomes estimation including YLDs, DALYs are calculated as the sum of YLLs and YLDs for each cause of injury. YLDs are also calculated for each nature of injury category. GBD 2017 adheres to the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER). GATHER is described in more detail in online supplementary appendix 2. injuryprev-2019-043531supp005.pdf
N/A
