Discussion_Covariates_of_Injury

Please note that the pages on this website do not reflect the methods and definitions that were used to generate the final set of results for the GBD-2010 study, which was published in December 2012.

-- If you are interested in the methods or results of GBD-2010 and subsequent revisions, please visit the official website of IHME's GBD project (http://www.healthdata.org/gbd). The materials described on the website are now obsolete but these webpages have been retained as an internet archive of the work of the group.

-- Please visit the website www.globalburdenofinjuries.org to find out more about other closely related collaborations of our group members.

Discussion Covariates of Injury Mortality

Status of discussion: Closed

Last Updated: Feb 7 2011

The GBD-2005 project is now in the final stages of estimating cause-specific mortality for all countries. At present, researchers at IHME-Seattle are using statistical models to predict the distribution of causes-of-death based on observed historical data and covariates of mortality, such as per-capita income.

At this stage, the GBD Injury Expert Group is advising the project on the appropriate choice of additional covariates that should be considered for modeling injury mortality. The purpose of this document is to summarize the recommendations for various injury causes, and invite recommendations from members of the Expert Group.

Characteristics of a good covariate

A covariate that is suitable for these statistical models has a set of important characteristics:

    • Strong predictor: There should be a theoretical reason to expect a strong relationship between the covariate and the injury being modeled.
    • High geographical coverage: Measurements of the covariate at the national level should be available for most countries.
    • Near complete time history: Data for the covariate should be available for most country-years from 1990 to 2005.

Interpolation/extrapolation models are being used to handle small amounts of missing data in fairly covariate. However, the requirement for a near complete time-history for most countries is quite stringent and usually rules out most candidate covariates. Thus, availability of sufficient data should be a primary consideration in the choice of covariate.

Please also note that in addition to the variables suggested by the Injury Expert Group, the mortality models will also use a generic set of variables that usually affect mortality. This list of generic variables (e.g. socio economic status, population density, urbanization) is still being finalized based on data availability. Thus the primary role of the Injury Expert Group is in identifying covariates that are important/strong predictors of specific injuries.

Example: Road Injuries

We have just gone through a process of identifying a suitable covariate for mortality from road injuries. This example should prove instructive for selecting covariates for other injury causes. The process was as follows:

    • We started by identifying country-level covariates that could be expected to have a relationship with road injury mortality and for which there was a possibility that complete country-year data may be available. The following variables were identified: per capita vehicle ownership (two-wheelers and four-wheelers), roadway length, transport fuel use, alcohol consumption, population density, and income.
    • Next, we evaluated data completeness. We constructed a panel dataset for these variables using various international databases (notably including the IRF, IRTAD, and World Development Indicators). We concluded that we had nearly complete data for income and population density. The completeness of vehicle information was patchy but the variable was considered very important. The completeness of the data for the remaining covariates was too low.
    • The missing data for vehicle ownership was interpolated using a statistical model of income and vehicles.
    • The final list of covariates for predicting road injury mortality in GBD-2005 is: per capita 4-wheeler ownership, per capita 2-wheeler ownership, per capita income and population density.

How can you contribute to this discussion

You can contribute by emailing suggestions on appropriate covariates to Kavi Bhalla (kavi_bhalla@harvard.edu). When suggesting covariates, it is important that you also suggest where we can access cross-country time history data for the covariate. Please send comments by Feb 6 2011.

Compilation of advice on covariates for injuries

Covariates that should be used as default for all injury causes:

The following set of co-variates should be used for all injury causes by default unless there is reason to exclude them:

  • Population age structure: These are two covariates
    • % population below 5 years; and
    • % population above 65 years
  • Income
  • Education
  • Alcohol consumption
  • Population density:
    • or % of population living above a certain density threshold as a proxy for urbanization

Acknowledgements

This document was developed by Kavi Bhalla and James Harrison by compiling suggestions from several members of the GBD Injury Expert Group. Members who have contributed at this stage include:

  • Fred Rivara
  • Members of the Homicide Advisory Group (including Richard Matzopoulos, Elisa Gilgen, Anna Alvazzi, and Lisa Knowlton)
  • Members of the Suicide Advisory Group (including Matt Miller, Catherine Barber, and Stephanie Burrows)
  • Richard Gosselin
  • Dan Judkins
  • David Schwebel
  • Researchers at IHME (Mohsen Naghavi, Jed Blore, Ruru Wang, Katrina Ortbald, and Rebecca Cooley)
  • Ken Winkel