Spatial analysis methods to improve localised estimates of HIV prevalence

Evaluation of geospatial methods to generate subnational HIV prevalence estimates for local level planning.

Anderson SJ, Subnational Estimates Working Group of the HIVMC. AIDS. 2016 Feb 25. [Epub ahead of print]

Objective: There is evidence of substantial subnational variation in the HIV epidemic. However, robust spatial HIV data are often only available at high levels of geographic aggregation and not at the finer resolution needed for decision making. Therefore, spatial analysis methods that leverage available data to provide local estimates of HIV prevalence may be useful. Such methods exist but have not been formally compared when applied to HIV.

Design/methods: Six candidate methods - including those used by UNAIDS to generate maps and a Bayesian geostatistical approach applied to other diseases- were used to generate maps and subnational estimates of HIV prevalence across three countries using cluster level data from household surveys. Two approaches were used to assess the accuracy of predictions: (1) internal validation, whereby a proportion of input data is held back (test dataset) to challenge predictions, (2) comparison with location specific data from household surveys in earlier years.

Results: Each of the methods can generate usefully accurate predictions of prevalence at unsampled locations, with the magnitude of the error in predictions similar across approaches. However, the Bayesian geostatistical approach consistently gave marginally the strongest statistical performance across countries and validation procedures.

Conclusions: Available methods may be able to furnish estimates of HIV prevalence at finer spatial scales than the data currently allow. The subnational variation revealed can be integrated into planning to ensure responsiveness to the spatial features of the epidemic. The Bayesian geostatistical approach is a promising strategy for integrating HIV data to generate robust local estimates.

Abstract access   [1]

Editor’s notes: Data from intensively monitored populations indicates that large differences in HIV prevalence can be seen across small geographic spaces. Understanding these localised spatial variations within a generalised epidemic can enable HIV programme resources to be used most effectively. However the data required for such localised estimation are often lacking. As a result modelling strategies must be used to predict local variation based on the best available data.

This study compares six different geospatial methods of estimating local HIV prevalence. The methods can be categorised by whether or not they use ancillary information such as road networks to improve their predictions and also whether they generated continuously changing prevalence surfaces (like map contours) or gave discrete estimates for geographic sub-regions e.g. districts.

While all methods produced reasonable overall levels of performance, those using a Bayesian geostatistical approach illustrated marginally better predictive accuracies. The levels of accuracy appeared more dependent on the national prevalence than the choice of model used.

The authors conclude by setting out a strategy for improvement of the models, principally through integrating additional data from sources such as antiretroviral therapy and prevention of mother-to-child transmission programmes, antenatal clinic surveys and case based reporting. 

Africa [6]
Kenya [7], Malawi [8], United Republic of Tanzania [9]
  • [10]