Mapping HIV prevalence using population and antenatal sentinel-based HIV surveys: a multi-stage approach

Sound public health policy on HIV/AIDS depends on accurate prevalence and incidence statistics for the epidemic at both local and national levels. However, HIV statistics derived from epidemiological extrapolation models and data sources have a number of limitations that may lead to under- or overestimation of the epidemic. Thus, adjustment techniques need to be employed to correctly estimate the size of the HIV burden.

A multi-stage methodological approach is proposed to obtain HIV statistics at subnational levels by combining nationally population-based and antenatal clinic HIV data.

The stages range from computing inverse probability weighting (IPW) for consenting to HIV testing, to HIV status prediction modelling, to the recently developed Bayesian multivariate spatial models to jointly model and map multiple HIV risks.

The 2010 Malawi Demographic and Health Survey (MDHS 2010) and the 2010 Malawi Antenatal Clinic (ANC 2010) Sentinel HIV data were used for analyses. Gender, residence, employment, marital status, ethnicity, condom use, and multiple sex partners were considered when estimating HIV prevalence.

The observed MDHS 2010 HIV prevalence among people aged 15–49 years was 10.15 %, with 95 % confidence interval (CI) of (9.66, 10.67 %). The ANC 2010 site HIV prevalence had a median of 10.63 %, with 95 % CI ranging from 1.85–24.09 %. The MDHS 2010 prevalence was 10.61 % (9.9, 11.33 %) and 10.19 % (9.69, 10.71 %) using the HIV weight and IPW, respectively. After predicting the HIV status for the non-tested subjects, the overall MDHS 2010 HIV prevalence was 11.05 % (10.80, 11.30 %). Higher HIV prevalence rates were observed in the mostly Southern districts, where poverty and population density levels are also comparatively high. The excess risk attributable to ANC HIV was much larger in the central-eastern and northern parts of the country.

Inverse Probability Weighting combined with an appropriate HIV prediction model can be a useful tool to correct for non-response to HIV testing, especially if the number of tested individuals is very minimal at subnational levels. In populations where most know their HIV status, population-based HIV prevalence estimates can be heavily biased. High-coverage antenatal clinics’ surveillance HIV data would then be the only important HIV data information sources.