Objectives: The aim of the study was to propose and demonstrate an approach to allow
additional nonsampling uncertainty about HIV prevalence measured at antenatal clinic
sentinel surveillance (ANC-SS) in model-based inferences about trends in HIV incidence
and prevalence.
Design: Mathematical model fitted to surveillance data with Bayesian inference.
Methods: We introduce a variance inflation parameter s2
in fl that accounts for the
uncertainty of nonsampling errors in ANC-SS prevalence. It is additive to the sampling
error variance. Three approaches are tested for estimating s2
in fl using ANC-SS and
household survey data from 40 subnational regions in nine countries in sub-Saharan, as
defined in UNAIDS 2016 estimates. Methods were compared using in-sample fit and
out-of-sample prediction of ANC-SS data, fit to household survey prevalence data, and
the computational implications.
Results: Introducing the additional variance parameter s2
in fl increased the error variance
around ANC-SS prevalence observations by a median of 2.7 times (interquartile
range 1.9–3.8). Using only sampling error in ANC-SS prevalence (s2
in fl ¼ 0), coverage
of 95% prediction intervals was 69% in out-of-sample prediction tests. This increased to
90% after introducing the additional variance parameter s2
in fl. The revised probabilistic
model improved model fit to household survey prevalence and increased epidemic
uncertainty intervals most during the early epidemic period before 2005. Estimating
s2
in fl did not increase the computational cost of model fitting.
Conclusions: We recommend estimating nonsampling error in ANC-SS as an
additional parameter in Bayesian inference using the Estimation and Projection Package
model. This approach may prove useful for incorporating other data sources such as
routine prevalence from Prevention of mother-to-child transmission testing into future
epidemic estimates.