Methods for Sub-national Estimates of HIV Prevalence

An interpolated map of HIV prevalence taken from DHS surveys (conducted between 2006-12).

Major decisions in HIV programming planning and the impact evaluation of programs rely on understanding the epidemiology of HIV in a country. However, it is increasingly clear that substantial heterogeneity exists within a country in the extent of the HIV epidemic, its dominant modes of transmission and the implementation of programs. Existing surveillance, estimation and modelling is overwhelming focussed at the national level, but gaining a better understanding of the sub-national variation could be important for improving the efficiency and effectiveness of the response to HIV/AIDS epidemics. Consequently, country programmers and epidemiologists, as well as international funders, increasingly want to routinely use tools to measure, visual and analyse these within-country epidemiological patterns.


In response to these developments, UNAIDS convened a meeting in July 2013 to bring together relevant stakeholders to review the field and to identify ways in which it could move forward. A forthcoming UNAIDS report will emphasize that countries should increasingly seek to leverage available data to identify areas of remaining high HIV transmission (‘epidemic strongholds’) and to fit programs to the local needs of populations. At the meeting, it became clear that many technical questions remain open about how to synthesize available data to provide a robust description of sub-national epidemiology. In particular, how much data is sufficient to make useful inferences on sub-national epidemiological patterns; how should data from different surveys be incorporated together; how fine should the scale of inference be.

Subsequent to that convening, UNAIDS commissioned a task force to advance this agenda and the HIV Modelling Consortium has been asked to investigate these technical challenges. In a related development, USAID has funded a group to review how data is used within countries to form programming decisions, and as part of that maps of key indicators (HIV prevalence, new infections, ART coverage, etc.) will be developed. Any recommendations from this Modelling Consortium project will feed directly into that USAID-funded work, as well as into the UNAIDS Reference Group on Estimates, Modelling and Projections (, which oversees the development of methods to compute the global AIDS statistics.

The Central Question

How should countries develop sub-national estimates relating to their HIV epidemics? 


Researchers within the HIV Modelling Consortium conducted literature reviews to identify experts currently working in mapping infectious diseases. Once an extensive list was compiled, these individuals, from a variety of backgrounds, were invited to participate in a process of development and review of optimal methods for generating sub-national estimates of HIV prevalence.

In total seven modelling groups agreed to participate in the exercise. The Secretariat collected data from a variety of sources, which the modelling groups could use to develop their method (GPS and HIV data from DHS, ANC, land cover, population density etc.), for the following countries:

  • Malawi
  • Tanzania
  • Ghana
  • Uganda
  • Cote d’Ivoire
  • Kenya

The Secretariat convened a meeting in Nairobi, 24 – 25 March 2014 for the modelling groups, representatives from UNAIDS, the Health Policy Project, and representatives from country program teams to discuss the utility of sub-national estimates of HIV epidemiology and review the proposed methods.

The performance of the methods in predicting the pattern of prevalence in a number of countries was assessed in the following ways:

  • Internal validation: (i) Leave one out cross validation (ii) Partitioned data hold back
  • External validation: Compare the estimated surfaces (where applicable) for one year (Y), with that which was measured by different set of clusters in a previous year (Y-T), where T is between 4 and 6 in the countries we have been working with. The comparison aims to demonstrate the acceptable performance of the methods rather than as a direct way to compare the performance of the different methods. This approach is confounded by real changes in the level and spatial distribution of HIV prevalence, the likely extent of which will be assessed post-hoc and potentially with reference to ancillary data sources. 


A summary report in relation to the meeting in Nairobi including recommendations developed by UNAIDS and the HIV Modelling Consortium are now available to download. Further validation exercises are being planned for the models and a paper is currently under development.