Bayesian epidemic models to infer spatio-temporal HIV incidence with applications to Malawi

Abstract

The rate of new human immunodeficiency virus (HIV) infections has decreased since its global peak in the mid-1990s. As the epidemic continues to recede, the efficacy of treatment and prevention programmes will depend on how well they target the right people in the right places. Few existing models of HIV burden offer spatio-temporally resolved estimates of incidence. I propose a novel epidemic model of HIV that bridges the gap between spatially resolved models of prevalence and epidemiologically sound compartmental models of incidence. It relies less heavily on fixed, exogenous data than previous models and fits directly to data from household surveys, antenatal care facilities, and HIV treatment programmes. Here, I present the details of the model, a broad set of specification tests, and descriptive results from an application to Malawi, as well as a comparison of methods for smoothing and interpolating sexual partner age data in preparation for adding age structure. The model comparisons identified a set of specification decisions that consistently led to better out-of-sample fit to district-level data from Malawi. I selected a single specification that fit well to most data and reproduced the programme data nearly perfectly. It estimated that increases in ART coverage resulted in decreases in HIV incidence in Malawi but that spatial heterogeneity in incidence was high. The proposed model offers several improvements on previous models of HIV incidence, and cross-validated model specification experiments provide relative confidence that the specification used here is appropriate for the epidemic in Malawi. The results from the selected specification underscore the continued success of the Malawian HIV treatment programme and highlight the possibility that, even in a high-prevalence setting, the epidemic is becoming increasingly concentrated.