3. 3www.healthdata.org
Institute for Health Metrics and Evaluation (IHME)
• Independent research center at the University of Washington
• Core funding by Bill & Melinda Gates Foundation and state of Washington
• 190 faculty, researchers, and staff
• Providing independent, rigorous, and scientific measurement and evaluations
– What are the world’s major health problems?
– How well is society addressing these problems?
– How do we best dedicate resources to get the maximum impact in improving
population health in the future?
• “Our goal is to improve the health of the world’s
populations by providing the best information
on population health”
5. 5www.healthdata.org
The Global Burden of Disease Study
• A systematic, scientific effort
to quantify the comparative magnitude of
health loss due to diseases, injuries & risk factors
• GBD 2010 published in The Lancet in 2012
• GBD 2013 published in 2014
– 323 diseases and injuries, 1,501 sequelae, 69 risk factors
– 188 countries, 1990 to 2013
– Findings published in major medical journals, policy
reports, data visualizations
7. 7www.healthdata.org
Measuring burden of diseases and injuries
DALYs (Disability-Adjusted Life Years)
Health
AgeDeath
Deaths
Best
life
expectancy
YLLs
YLLs (Years of Life Lost)
YLDs YLDs
YLDs (Years Lived with Disability)
Disability Weight
8. 8www.healthdata.org
GBD data inputs
•Vital registration
•Censuses
•Surveys
•Verbal autopsy
•Disease registries
•Surveillance systems
Population-based Encounter-level Other
•Hospital records
•Ambulatory records
•Primary care records
•Claims data
•Literature reviews
•Sensor data
•Mortuaries/burial sites
•Police records
12. 12www.healthdata.org
Data & Model Flow
Mortality
2
Causes
of death
3
Nonfatal
health
outcomes
4
Risk
factors
5
Co-
variates
1
YLLs/
YLDs/
DALYs
6
13. 13www.healthdata.org
Vital records in GBD
• Mortality
• Preparing data for Causes of Death analysis
• Causes of Death Ensemble Modeling (CODEm)
• CodCorrect
• Results
15. 15www.healthdata.org
Causes of death data: 600M deaths back to 1980
Type Site
years
Coun-
tries
Vital
registration
2,798 130
Verbal
autopsy
486 66
Cancer
registries
2,715 93
Police reports 1,129 122
Surveys/
census
1,564 82
Maternal
mortality
surveillance
83 8
Deaths in
health
facilities
21 9
Burial and
mortuary
32 11
24. 24www.healthdata.org
Garbage codes: summary
• US is doing very well in international comparison
• Active role in discouraging use of garbage codes
• Consistency: maternal mortality increase in US
(pregnancy check-box on some states’ death certificates)
• Methods available to correct for garbage codes;
working on software to provide to others
25. 25www.healthdata.org
Cause of Death Ensemble Modeling (CODEm)
1. Identify and prep all available data
2. Develop a diverse set of plausible models for each cause
– Different types: negative binomial, fixed proportion, natural history, etc.
– Different (sets of) covariates
3. Assess predictive validity of each individual model and each ensemble
of models via out-of-sample test
4. Use best performing model/ensemble for analysis
26. 26www.healthdata.org
CodCorrect
• Ensure that cause-specific deaths fit all-cause mortality envelopes
• Key advantage of looking at all causes at once in GBD
• Implemented taking into account uncertainty in every cause of
death model
• Applied at all hierarchical levels
28. 28www.healthdata.org
Visualizing results
• Vetting input data
• Reviewing results
• Collaborating with experts
• Communicating results
Simple
visualizations
Google
Motion Charts
Viz platforms
Custom
coding
Static graphs
29. 29www.healthdata.org
Communicating Data for Impact
• Audiences and characteristics
– Casual user
– Data actor
– Data analyst
– Researcher
• Granularity of data
• Type of tool or visual
http://bit.ly/1mogRom
32. 32www.healthdata.org
Strengths of the GBD approach
• Synthesis of all available data
• Innovative, peer reviewed methods
• Consistent methods make results comparable
• Uncertainty bounds for all metrics
• Coverage of all causes prevents
double-counting, e.g., mortality,
anemia
• Fully imputed dataset
33. 33www.healthdata.org
Looking ahead: US burden by county
• Successful collaborations with UK,
China, Mexico
• Extend US burden to subnational level
– All counties
– Sub-county for large counties
– Objective: entities smaller than 100K
people
• Starting with Causes of Death by
county
• Funding discussions for proof of
concept with RWJF (10-20 counties)
34. 34www.healthdata.org
US burden by county: access to data
• Issues with some data at the county/sub-county level
– Access only at state or county level
– Masking at county level
– Access via RDC
• IHME data security
– Servers owned and operated, not shared
– Access control by individual for Limited Use folders
– Secure room
– Data use agreements
35. 35www.healthdata.org
US burden by county: collaboration
• Expert collaboration like GBD Global
– Discussion of input data
– Review of preliminary results
– Joint outreach
– Collaboration at state and county level
• Visualizations
• Trainings
36. 36www.healthdata.org
Summary
• Fantastic data work in the US at the county, state, and national levels
• Great progress over the past 30 years in quality of VR
• There can never be enough data
• Looking forward to collaborations on US burden and more
Contact me:
Peter Speyer
speyer@uw.edu
@peterspeyer