Interpersonal Violence

The GBD cause of death database consists of VR and VA data; survey and census data for injuries and maternal mortality; surveillance data for maternal mortality and child death; cancer registries; and police records for interpersonal violence and road injuries. Self-harm estimates incorporate VR data and are based on ICD categorisation as described in appendix 1 (section 7). In this iteration of GBD, ten new VA studies and 127 new country-years of VR data were added at the country level. 502 new cancer-registry country-years were added, as was one additional new surveillance country-year. Data sources comprising the GBD cause of death database can be reviewed on the Global Health Data Exchange website. Multiple factors can influence changes between GBD studies in estimates for a given cause-location-year, including the quality of a country's data system (as represented by the GBD star rating system) and the addition of more recent data. Figure 1 shows the relative stability of GBD estimates between study iterations. Variation between GBD 2016 and GBD 2017 estimates was greater in countries with both low star ratings and no new VR data updates occurring between these iterations of the study. Changes to estimates can be seen even in high star rating locations because of changes in modelling strategy or model covariates even when no new VR data were available between cycles.Figure 1

Effect of new VR data on Level 1 cause estimates from GBD 2016 to GBD 2017, based on national locations with varying quality of VR data, 2008–16

The figure shows the degree of consistency between GBD 2016 and GBD 2017 estimates for Level 1 causes at the national level from 2008 to 2016. The diagonal line represents no change from GBD 2016 to GBD 2017. Each point represents one country-year, with colours indicating the Level 1 cause grouping (communicable, maternal, neonatal, and nutritional diseases; non-communicable diseases; and injuries). Panels indicate whether or not any new VR data between 2008 and 2016 were added for that location for GBD 2017, and whether or not a location has 4-star or 5-star VR quality. Points that are outside of the standard 95% prediction interval for a linear regression of 2017 values on 2016 values are annotated (if the same location-cause had multiple points in a time series, only the furthest-most point was annotated). The Spearman's correlation coefficient is noted in the lower right-hand corner of each panel. CSMR=cause-specific mortality rate. GBD=Global Burden of Diseases, Injuries, and Risk Factors Study. VR=vital registration.

For GBD, each death is attributed to a single underlying cause—the cause that initiated the series of events leading to death—in accordance with ICD principles. This categorical attribution of causes of death differs from the counterfactual approach, which calculates how many deaths would not have occurred in the absence of disease. GBD also differs from approaches involving excess mortality in people with disease monitored through cohort or other studies. Deaths in such studies might be assigned as the underlying cause, be causally related to the disease, or include deaths with confounding diagnoses.³ (link)
The GBD cause list is organised as a hierarchy (appendix 1 p 477), with each level composed of causes of death that are mutually exclusive and collectively exhaustive. The GBD cause hierarchy, with corresponding ICD9 and ICD10 codes, is detailed in appendix 1 (p 300). GBD Level 1 causes are grouped as three broad categories: communicable, maternal, neonatal, and nutritional (CMNN) diseases; NCDs; and injuries. Level 2 causes contain 21 cause groups, including subsets of CMNN causes, cancers, cardiovascular diseases, and types of injuries (eg, transport injuries, self-harm, and interpersonal violence). Individual causes are primarily recorded at Level 3 (eg, malaria, asthma, and road injuries), while a subset of Level 3 causes are disaggregated further to Level 4 causes (eg, four sub-causes within chronic kidney disease).
For GBD 2016, we disaggregated some Level 3 causes to expand the cause hierarchy used for GBD 2015 by 18 causes of death. GBD cause list expansion was motivated by two main factors: inclusion of causes that result in substantial burden and inclusion of causes that are of high policy relevance. New causes for GBD 2016 included Zika virus disease, congenital musculoskeletal anomalies, urogenital congenital anomalies, and digestive congenital anomalies. Other leukaemia was added as a Level 4 subcause to leukaemia rather than being estimated in the Level 3 residual category of other neoplasms. The Level 3 cause of collective violence and legal intervention was separated into “executions and police conflict” and “conflict and terrorism”. Disaggregation of existing Level 3 causes resulted in the addition of 11 detailed causes at Level 4 of the cause hierarchy: drug-susceptible tuberculosis, multidrug-resistant tuberculosis, and extensively drug-resistant tuberculosis; drug-susceptible HIV–tuberculosis, multidrug-resistant HIV–tuberculosis, and extensively drug-resistant HIV–tuberculosis; alcoholic cardiomyopathy, myocarditis, and other cardiomyopathy; and self-harm by firearm, and self-harm by other means. Within each level of the hierarchy the number of collectively exhaustive and mutually exclusive causes for which the GBD study estimates fatal outcomes is three at Level 1, 21 at Level 2, 145 at Level 3, and 212 at Level 4. For GBD 2016, separate estimates were developed for a total of 264 unique causes and cause aggregates.

The GBD Cause of Death Ensemble model (CODEm) systematically tested and combined results from different statistical models according to their out-of-sample predictive validity. Results are incorporated into a weighted ensemble model as detailed in appendix 1 (section 3.1) and below. For GBD 2017, CODEm was used to estimate 192 causes of death (appendix 1 section 7). To predict the level for each cause of death, we used CODEm to systematically test a large number of functional forms and permutations of covariates.¹⁸ (link) Each resulting model that met the predetermined requirements for regression coefficient significance and direction was fit on 70% of the data, holding out 30% for cross-validation (appendix 1 section 3.1). Out-of-sample predictive validity of these models was assessed by use of repeated cross-validation tests on the first 15% of the held-out data. Various ensemble models with different weighting parameters were created from the combination of these models, with the highest weights assigned to models with the best out-of-sample prediction error for trends and levels, as detailed in appendix 1 (section 7). Model performance of these ensembles was assessed against the root-mean squared error (RMSE) of the ensemble model predictions of the log of the age-specific death rates for a cause, assessed with the same 15% of the data. The ensemble model performing best was subsequently selected and assessed against the other 15% of the data withheld from the statistical model building. CODEm was run independently by sex for each cause of death. A separate model was run for countries with 4-star or greater VR systems to avert uncertainty inflation from more heterogeneous data. The distribution of RMSE relative to cause-specific mortality rates (CSMRs) at Level 2 of the GBD hierarchy shows that model performance was weakest for causes of death with comparatively low mortality rates (figure 2; appendix 2), while models for more common causes of death such as stroke, chronic obstructive pulmonary disease, and self-harm and interpersonal violence generally had low RMSE.Figure 2

Out-of-sample model performance for CODEm models and age-standardised cause-specific mortality rate by Level 1 causes

Model performance was defined by the root-mean squared error of the ensemble model predictions of the log of the age-specific death rates for a cause with 15% of the data held out from the statistical model building. The figure shows the association between the root-mean squared error and the log of the CSMR, aggregated over 1980–2017. Each point represents one CODEm model specific for model-specific age ranges and sex. Circles denote models run with all locations. Triangles denote models run on only data-rich locations. Colours denote the Level 1 cause categories. Open circles and triangles denote models that were run with restricted age groups of less than 30 years. CODEm=Cause of Death Ensemble model. CSMR=cause-specific mortality rate.

All interviews were audio-recorded and afterwards transcribed by the lead researcher to increase familiarity with the data. After checking the transcripts for transcription errors, meaning was constructed into the data using a grounded theory approach. Transcripts were explored by the lead researcher for pertinent themes and discussed with the research team.The data were labelled line by line using open coding as we asked, “what is this an example of?” Open coding was used to identify basic themes relating to women’s experiences (see Table 2) and decision-making processes in relation to homelessness across the 20 interviews. Concepts such as trauma, interpersonal violence, and child removal. Were grouped under category labels. Each category was considered in terms of its characteristics and as differences and similarities emerged, we collapsed our initial collection of concepts into a code list of important concepts. It was at this stage that the centrality of networks and resources, and thus social capital, in the discussion became apparent. Interview transcripts were then re-examined to identify broader themes (or forms) of agency under which certain basic themes could be grouped.Table 2

Number of participants who identified different experiences (n = 16)

	identified	disagreed	not addressed
Habitus of instability
Early trauma	88% (14/16)		12% (2/16)
Care experienced	69% (11/16)	31% (5/16)
Homeless before age 21	63% (10/16)	31% (5/16)	6% (1/16)
Hidden homelessness
Significant trauma whilst homeless	56% (9/16)		44% (7/16)
Domestic abuse
Experienced physical abuse	88% (14/16)	6% (1/16)	6% (1/16)
Experienced emotional abuse	88% (14/16)		12% (2/16)