Assessing the impact of imperfect immunoassays on HIV prevalence estimates from surveys conducted by the DHS program

This report investigates the impact of potential misclassification of samples on HIV prevalence estimates for 23 surveys conducted from 2010-2014. In addition to visual inspection of laboratory results, we examined how accounting for potential misclassification of HIV status through Bayesian latent class models affected the prevalence estimates. Two types of Bayesian models were specified: a model that only uses the individual dichotomous test results and a continuous model that uses the quantitative information of the EIA (i.e., the signal-to-cutoff values). Overall, we found that adjusted prevalence estimates matched the surveys’ original results, with overlapping uncertainty intervals. This suggested that misclassification of HIV status should not affect the prevalence estimates in most surveys. However, our analyses suggested that two surveys may be problematic. The prevalence could have been overestimated in the Uganda AIDS Indicator Survey 2011 and the Zambia Demographic and Health Survey 2013-14, although the magnitude of overestimation remains difficult to ascertain. Interpreting results from the Uganda survey is difficult because of the lack of internal quality control and potential violation of the multivariate normality assumption of the continuous Bayesian latent class model. In conclusion, despite the limitations of our latent class models, our analyses suggest that prevalence estimates from most of the surveys reviewed are not affected by sample misclassification.