Transcript Slides - View the full AIDS 2016 programme

Development of methods to produce
spatial estimates of HIV epidemics
Samir Bhatt
Department of Infectious Disease Epidemiology, Imperial College London
Aims
• To produce reliable estimates at any spatial scale.
• To model with robust uncertainty
• To learn new HIV dynamics
17th May 2016
Conceptual framework
Thembisa prevalence
(national/subnational, year,
age, sex)
Prevalence
Prevalence (surveys)
(x,y, year, age, sex)
Geospatial model for
prevalence
Prevalence
(pixel, year, age, sex)
Population
(pixel, year, age, sex)
×
N infected
(pixel, year, age, sex)
Facilities, transport network etc
Facility ‘catchment’
model
N infected per catchment
(catchment, year, age, sex)
Geospatial model for
ART coverage
ART coverage
(catchment, year, age, sex)
Prevalence (PMTCT facilities)
(x,y, year, age)
Prevalence (ANC sentinels)
( x,y, year, age)
Incidence
ART Coverage
Spatial covariates
Facility-level numbers on ART
(x,y, year, age, sex)
Thembisa ART coverage
(national/subnational, year,
age, sex)
Thembisa incidence
(national/subnational, year,
age, sex)
Geospatial model for
incidence
Incidence
(pixel, year, age, sex)
South Africa
Types of Data: HSRC
17th May 2016
South Africa
Types of Data: HSRC
17th May 2016
South Africa
Types of Data: HSRC
17th May 2016
South Africa
Map of prevalence 2012 – all ages
17th May 2016
Women Aged 18-25
17th May 2016
Women Aged 18-25
17th May 2016
Mozambique
Types of Data: Survey
17th May 2016
Mozambique
Types of Data: PMCT/Routine data
2009
2014
17th May 2016
Mozambique
Types of Data: ANC/Sentinel data
17th May 2016
Mozambique
ANC – Great Temporal resolution (old years), bad spatial
PMTCT – Great Spatial and partial temporal (recent years)
17th May 2016
Mozambique
Time series: Survey + PMTCT + ANC
17th May 2016
Mozambique
Bayesian Uncertainty
95% Credible Interval
Ability to aggregate to any
resolution >5km while
preserving uncertainty
bounds
17th May 2016
Mozambique
Unpicking the black box
PMTCT+ANC: 45%
Night Lights: 15%
Age: 15%
Accessibility: 10%
Land Aridity: 5%
Day/Night temperature flux: 5%
Population: 5%
Sex: 0%
17th May 2016
Mozambique
Partial Dependence Marginal Plots
17th May 2016
Mozambique
Partial Dependence Marginal Plots
17th May 2016
Conclusions
• Modelling methods can synthesize multiple sources of data
• The “big data” approach can yield results at finer scales
• Model performance has been very good
• Routine and surveillance data are co-informative
• Next steps – ART coverage and Incidence
17th May 2016
Acknowledgments
• UNAIDS
• Pete Gething, Jeff Eaton and Tim Hallett (HIVE MAP)
• Mozambique Ministry of Health
• GAWG (Geospatial Analysis Working Group), HSRC
• Mary Mahy and Kim Marsh
17th May 2016