Injury incidence, prevalence and YLDs

The method for estimating non-fatal injury outcomes including falls in GBD 2017 is described in more detail in related publications.¹⁹ A methodological summary is as follows.

First, we used DisMod-MR 2.1 to measure incidence of falls that lead to any form of medical care (inpatient or outpatient). DisMod-MR 2.1 is a meta-regression tool for epidemiological modelling built on a Bayesian compartmental model framework that solves differential equations that modulate the relationships between a susceptible population becoming injured (incidence) and then either recovering (remission) or dying (excess mortality). For incidence data, we used emergency department records, hospital records, survey data and literature studies to estimate fall incidence by location, year, age and sex, and used the coefficient from outpatient care to split subsequent estimation processes into inpatient and outpatient incidence estimates so that inpatient and outpatient-specific data could be used where possible to preserve differences in incidence and severity. Since survey items for falls can include non-injurious falls, we included an indicator variable for falls that resulted in injury. Since excess mortality is calculated based on locations where there are overlapping incidence and cause-specific mortality data, its computation also allows for estimation of incidence in locations with cause-specific mortality data but no incidence data, requiring an assumption that case fatality rates among falls are affected by income.

Second, we estimated the distribution of nature-of-injury categories among the incidence of all falls. To do this, we created a hierarchy of nature-of-injury categories. We assumed that the disability experienced by an individual who has an injurious fall was determined by the most severe nature-of-injury sustained due to this fall. For example, a fall resulting in a spinal cord injury would determine disability due to the fall instead of a co-occurring wrist sprain. The nature-of-injury hierarchy represents a combination of the likelihood of long-term disability and the corresponding GBD disability weight. To estimate the hierarchy, we used data from pooled follow-up studies in which we translated each individual’s health status measure at 1 year after injury into a disability weight.^20–26

Third, we used a Dirichlet regression method to estimate the proportion of falls that result in each nature-of-injury category being the most severe injury for each fall, since Dirichlet methods enforce coefficient estimates for proportions that must sum to 1.²⁷ These matrices were derived from dual-coded hospital and emergency department data sets from multiple countries and data from the China injury surveillance system where both cause-of-injury and nature-of-injury diagnosis codes are present. The use of these data sources to inform this estimation process is described in more detail elsewhere.^{1 28} Separate cause-nature matrices were created for falls warranting hospital admission versus falls warranting other healthcare, high and low-income countries, male and female, and age category.

Fourth, we estimated short-term disability for falls by nature‐of‐injury category. For each nature-of-injury category and inpatient and outpatient injury, we used the Dutch Injury Surveillance System to derive average duration for treated cases, since for GBD 2017 this was the only available data source that could inform this parameter.^{23 24} These estimates were supplemented by expert-driven estimates of short-term duration for nature-of-injury categories that had insufficient numbers in the Dutch data set and for untreated injuries.

Fifth, we estimated the proportion of falls resulting in permanent disability for each nature-of-injury category by admission status and age. Disability due to falls was assumed to affect all injurious falls in the short term with a proportion having long-term (permanent) outcomes, defined as having persistent disability 1 year after the injury greater than the preinjury health status.

Sixth, we applied the ordinary differential equation solver used as the computational engine in DisMod-MR 2.1 to estimate the long-term prevalence for each fall-related nature-of-injury from incidence and the long-term mortality risk in cases with long-term disability based on meta-analyses of studies providing standardised mortality ratios. For example, since individuals with severe traumatic brain injuries die at a higher rate than the underlying population, we integrated the corresponding standardised mortality ratios to account for decreasing prevalence due to higher mortality risk in this injured population.

Finally, we calculated YLDs as prevalence of each health state multiplied by a disability weight for each nature-of-injury and corrected for comorbidity with other non-fatal diseases using microsimulation methods employed in GBD 2017.